Latent Space: The AI Engineer Podcast — CodeGen, Agents, Computer Vision, Data Science, AI UX and all things Software 3.0

Materials science is the unsung hero of the science world. Behind every physical product you interact was decades of research into getting the properties of materials just right. Your gym clothes contain synthetic fibers developed over decades. The glass screen, diodes, and chip substrate technology needed to read this blog post were only viable due to many teams of material scientists.Our guest Prof. Heather Kulik was one of the first material scientists to realize that there was alpha in combining computational tools with data driven modeling — she did AI for science before it was cool. She has a hard-fought perspective for how to succeed in this field. Yes, she believes the wins are real. To get there you must work hard to deeply integrate domain expertise with AI techniques, and also maintain a discriminating mind. Ultimately what matters is you succeed in the lab, and nature doesn't care about how hyped a model is. These lessons personally resonated with the Latent.Space Science team and our own experience.This episode is a must watch for all aspiring AI for science practitioners. A few highlights:Designing new polymers with AI: Heather's group recently used AI to design new polymers that are significantly stronger. These materials were created and tested in the lab, and the scientists who built them were surprised by the designs. The AI had figured out certain building blocks could break in a novel way. The AI discovered a purely quantum mechanical effect, and after convincing their lab collaborators to actually synthesize it, the material turned out to be four times tougher!The twenty-two-atom ligand challenge: When asked about the role and need of human scientists, Heather points out that AI has a strong understanding of academic chemistry, but is still lacking intuition. Every time an LLM is updated, Heather asks it to design a ligand that contains exactly twenty-two heavy atoms. She has yet to find one that can succeed at this seemingly simple task that any expert could do in a second! Is this the chemistry counterpart to counting ‘r's in strawberry?Side note: Heather joked that this comment would date itself immediately, so we decided to see if this was still true three months after recording. We found some interesting results! We asked both Claude and ChatGPT to design a 22 atom ligand for both a metal-organic framework (MOF) and a Kinase protein. * For the Kinase, both models got it right: Claude pulled out RDKit in a python script and iterated on several designs, whereas ChatGPT just one-shotted it. * For MOFs, both models got it wrong, generating ligands with 21, 23, or 24 atoms, yet stubbornly not getting 22 atoms. Is there something different about how LLMs reason in the materials and bio domains?Materials vs biology: The two biggest domains of AI in science have been biology and materials. We asked Heather if there could be an AlphaFold moment for materials. Her answer reframes how we should think about the field:* First, the datasets in material science are woefully lacking in comparison to the bio world. The closest to ground truth in most cases are noisy DFT datasets. These are just approximations to the real world! The datasets that are accurate are all boring, as Heather quipped “We have really good datasets for really boring chemistry.” Furthermore, good experimental structures are hard to come by and require interpretation. So generating generating high-quality, novel datasets at scale would really drive the field forward.* More philosophically, AlphaFold is making predictions in a fairly limited space: there are just twenty amino acids. Sure, even here AlphaFold doesn't get everything right, but it seems plausible that one could learn the entire design space. For materials, each element is a new set of interactions and chemistry, with little to no transferability. This is a massive open problem in material science that we hope some of the smartest AI scientists will want to work on!The difficulties of trusting the literature: Heather's team has spent the last few years using NLP and later LLMs to extract data from literature. Even a few thousand data points from these papers can be valuable for guiding her group's work. One surprising result: sometimes the reported values for a property (say temperature) do not match up with the graphs in the papers! So there's lots of potential in using LLMs to mine data from the literature, just do it with care.The role of academia in an ever-changing world: One theme that has been running through many of our conversations has been the changing role of the academic — and the scientist — in science. When startups are raising $100s of millions and hyperscalers and Big Pharma are all ramping up AI-for-science efforts, the academic researcher needs both resources and judgement about problems to chase more than ever.Resources include data that is organized for machine learning, access to high throughput experimentation labs, and compute resources. These are all things that academics can build together. More importantly, Heather emphasizes curiosity about problems that haven't hit the radar of the heavily capitalized AI companies. After so many years on the forefront of AI for Science, Heather's judgement that Chemical Engineering and Material Science still need curious people asking questions with no clear path to money is a welcome beacon in the AI fog.Full Video podcast Is on Youtube! This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe

For a limited time, Latent Spacenauts can skip the waitline to join Dreamer and also compete for a $10,000 cash prize for most useful tools for Dreamer! Thanks @dps!In 2024, David Singleton left Stripe and joined forces with Hugo Barra for a buzzy stealth startup named /dev/agents. This month they emerged out as Dreamer, a consumer-first platform to discover, build, and use AI agents and agentic apps, centered on a personal “Sidekick” that helps users customize experiences via natural language. Sidekick is nothing less than an “agent that builds agents”, with all the complexity that that entails:You've seen many many website builder, app builder, and even agent builder startups by now, but our favorite detail is the sheer amount of work that has gone into the “full stack” nature of the platform, including shipping their own SDK, logging, database, prompt management, serverless functions, and so on. Most platforms restrict the tech stack you can use just to get off the ground — Dreamer does it “right” by letting you push whatever arbitrary code you want to their VMs.Paying the BuildersOf course former leaders of Stripe and Android would not stop at just building the tools, but also building the ecosystem. Dreamer is deeply aware of the 4 sided network effect it has going on and is ready to fund all of it - from hiring Builders in Residence to awarding $10,000 cash prizes to the best tool builders for the Dreamer ecosystem.It's time to Dream!Full Video Episodeon youtube.Transcript[00:00:00] Meet Dreamer Purple[00:00:00] swyx: Okay, we're here in the studio with David Singleton. Welcome.[00:00:08] David Singleton: Hey, Wix. It's great to be here.[00:00:09] swyx: It's great to have you. Uh, we have very sympa that your company color is the same as Lean Spaces color.[00:00:15] David Singleton: That's right. Dreamer Purple.[00:00:17] swyx: It used to be Devrel agents, which I thought was very cool. It's like you call back to Devrel Payments.[00:00:22] David Singleton: Yeah.[00:00:22] swyx: And you were obviously CTO Stripe. And talk to me about just the origin or thinking process behind Dreamer. Yeah. And maybe, maybe start with like, what, what is Dreamer?[00:00:31] David Singleton: Yeah.[00:00:31] What Is Dreamer[00:00:31] David Singleton: So Dreamer is a new product, uh, which everyone can come and play with today. Um, it's a place where everyone, literally, everyone can discover, build, and enjoy and use AI agents and agenda apps.[00:00:45] And we really did design it for consumers, for folks who are not necessarily. Uh, have any kind of technical background. It's really aimed at everyone. I think often of my sister, she's very smart. She's not in the slightest bit technical. She has lots of problems in her life that [00:01:00] she would like to be able to have great software and intelligent software to solve.[00:01:04] But you know, even with the rise of tools like Cloud Code and so forth, she's got no way to get started. And Dreamer is a place where she can come in, grab some intelligent apps that other people in the community have built, start using them right away, and solve real problems in her life.[00:01:19] Sidekick And Waitlist[00:01:19] David Singleton: And at the core, we have a personal agent called the Sidekick.[00:01:24] Um, you can give your sidekick a name, you can give it its own personality, and it really helps you across your entire day, your life. It helps you use all of the agents on the platform, and it also helps you build anything you want. And we've been working in this for a little while. We recently launched in beta.[00:01:41] So anyone can go to dreamer.com, join the wait list. Um, and we have many, many, many people in the community now who are building really fun, really powerful, really useful. Agents and the agentic apps for themselves.[00:01:54] swyx: I think we're gonna go right into a demo. Yeah. I just wanna make an observation that, uh, you, you, [00:02:00] you put discover first before build.[00:02:02] Mm-hmm. But actually, at least for the engineers in the audience. ‘cause we are primarily engineers and you're primarily targeting consumers, right?[00:02:08] David Singleton: Yeah.[00:02:08] swyx: For engineers. Like, there's a huge full stack of stuff, which we're gonna dive into. Let's write. It's so impressive. I'm like, holy s**t, this, this is what I've always wanted.[00:02:16] Cool. Uh, so, so I think that's really good and I've, in some ways, I think given your background given, uh, Hugo's, is it Hugo? Hugo.[00:02:24] David Singleton: Hugo. Hugo Bar. Yeah.[00:02:25] swyx: Hugo, it's not surprising that you can basically kind of build an app store Yeah. For agents.[00:02:30] David Singleton: Yeah. So Hugo was my co-founder. Yeah. Um, Hugo and I met with our other co-founder Nicholas Checkoff in the very early days of Android at Google, where we were building Google's first mobile apps.[00:02:41] Uh, we then contributed to very core pieces of Android itself. And you're right, we were really excited about building two things. One, solving a bunch of problems. That this breakthrough technology here I'm talking about mobile needed to have solved in order to make it work for real people at scale. And then secondly, building this ecosystem, um, [00:03:00] of third party developers using the Play Store, um, and able to deliver way more value on the platform than we could have delivered on our own.[00:03:08] And we think about Dreamer in exactly the same way. So I was working at Stripe, as you mentioned, and we had the opportunity to put some of the very first AI agent systems in the world into production. And from the moment we did the first of those, I was just struck with a strong sense of conviction that this is breakthrough technology that's gonna change how all of us work with computers and phones and so forth, all of the, the technology in our lives, but.[00:03:34] There's a lot of problems to be solved, for real people to be able to make this approachable. Um, and it really is kind of a direct analog for what we were solving back in the early days of mobile apps at Google and, and Android. So it's, it's been fun to bring that to life.[00:03:47] swyx: Yeah. Uh, let's look at it.[00:03:48] David Singleton: Yeah, let's take a look.[00:03:49] Dashboard And Daily Briefing[00:03:49] David Singleton: So, uh, dreamer.com, this is our homepage. This is where you can come and, uh, watch some videos about what is here and sign up for the wait list. Once[00:03:57] swyx: you, I, I just wanna say for those listening, ‘cause we have a lot, you [00:04:00] know, switch to YouTube, look at the animations. So much care.[00:04:03] David Singleton: We, we really care about, uh, this product being fun.[00:04:07] Uh, and, and interesting to use. Obviously a lot of people are using it to do real important stuff. You can do real work, uh, here, uh, but also you can build fun things too. Once you get off of our wait list, you'll come into the product. The first thing that happens is you'll have a conversation with your side cake, which is this little friendly, uh, character here.[00:04:27] And psychic will seek to get to know you and understand you. What do you care about? And will help you discover and build your first AI agents or agentic apps. After that, you're, you're gonna have a dashboard. This is my dashboard. Everyone's is different. Um, you can see I have a few things here. I have a feed.[00:04:42] So a lot of our agents do things in the background when you're not looking and the feed is how they let you know what they've been up to. I have, uh, some widgets, uh, from apps that I have built. Uh, this one is called Calendar Hero. Uh, this is something that I installed from the gallery. Uh, so built by someone in our community.[00:04:59] It's a [00:05:00] really powerful calendar app because for each of my meetings, if it's with someone I don't already know, well it'll actually go off and research it, um, and give me both a history of my interactions with those people and also a bunch of, you know, public useful information to, to get started. One of the things I love about this particular app is that every day it generates a podcast, um, a daily briefing.[00:05:24] And one of the things that we've done with the platform is we've made it possible for all the things that agents do to show up in places that you care about. So if you look over here, this is the screen in my phone, and if I go ahead and open my Apple Podcasts, you can see right here. Your Daily briefing podcast is ready.[00:05:39] This was produced by an agent running in my Dreamer account, and it was very easy by scanning a QR code to connect it to my Apple podcast. That's what I listened to in the car now every morning. Yeah. On my way to work.[00:05:50] swyx: It, it[00:05:50] David Singleton: preps me for, for my day.[00:05:52] swyx: So one additional bit of context. I asked you immediately after seeing this was like, what, what about, I wanna talk back to my agent and you said you actually started with voice and then you went to [00:06:00] podcasts.[00:06:00] ‘cause it's nice to have it pre downloaded[00:06:02] David Singleton: that, right? That's right. Um, yeah, we, you, you can talk to your sidekick. So, you know, on mobile we have, uh, a dreamer app and you can talk to the sidekick right here. Um, but we've actually found that making things, uh, show up in the other apps that you already use in your life is incredibly powerful.[00:06:19] So let's take a look at what's kind of under the hood here.[00:06:21] Gallery Tools And Payouts[00:06:21] David Singleton: So I already mentioned that we have a gallery, so this is where you'll find a lot of agents from our community. Uh, there's. Many at this point, hundreds. And they are solving all kinds of, uh, use cases. I'd say the the top use cases are on personal productivity, but also a lot of information management that can range from personal information like docs and so forth, managing your emails.[00:06:42] It also ranges out to public information that you might be interested in, but you need something to help manage the, the kind of fire hose of stuff that's coming at you. For instance, I have, um, an agent which looks at all the AI news, um, all the time. There's a lot of it and it finds the stuff that I would actually be [00:07:00] interested in, um, and I find it incredibly useful.[00:07:03] So these are agents that you can install that other people have built. Anything that you install on Dreamer, you can actually just say, I wanna start making some changes, and we'll look at that in a second. But in natural language, with the sidekicks help, you can change any of these experiences to work just the way you want them.[00:07:18] But the base layer of the system are tools. So you know, as well as anyone swyx, that any AI system is only as good as the quality of data that it can pull in and the quality of action it can take. So before we launched our beta, we worked very hard to make sure that we seeded our tools with a bunch of very high quality and powerful integrations.[00:07:39] So, you know, for instance, this is real Google search, this is actual Gmail. Um, and you can do very useful things with those. But also this is a platform for everyone. And as we got started talking to people in our alpha community, a whole bunch of sports use cases popped out and we realized if you want to build something cool for sports with ai, you need really high quality live data.[00:07:58] So look at these [00:08:00] Formula one M-L-B-N-F-L, uh, these are tools, uh, that we've built. We've done a, these are not data scraped off the web. This is a, a direct data feed integration. And because it's live and ‘cause it's high quality, you can build really powerful stuff. But tools is not something that we are just going to kind of control ourselves.[00:08:19] The platform is open for tool Builders to contribute tools that anyone on Dreamer can use. So, um, this is actually the place in the platform where I think software engineers, um, well number one, would love for you to come and play with it. Uh, but software engineers are really gonna build, um, a lot of powerful stuff into the system.[00:08:38] And we are actually sharing something for the first time on this podcast, which there is, uh, tool builders on Dreamer get paid. So if you publish a tool to the platform and a lot of agents use it, you'll actually get paid, uh, in proportion to their usage. And we'd love for folks to come and give this a try.[00:08:54] We've got good docs that help you get started and you can build things that, you know, scratch your own itch. For instance, someone built this [00:09:00] Ski Bum tool, which provides live snow conditions for a bunch of, uh, ski resorts. I'd love to show you how I've used that in a second. And also we have some tools, partners where the tools themselves are paper use.[00:09:12] So for instance, parallel web systems is a premium tool. Uh, you can do really cool stuff with it. Um, it's a a, an agentic web research tool. And that one, because it's expensive to operate, is paid on a, on a per usage basis. But if you're coming in to build agents on the platform, even the premium tools, you get a free trial.[00:09:29] So you get a chance to actually try them out, make sure that the use case is good for you before you decide to, to to sign up. So that's tools. So we have the gallery, we have tools, and then the sidekick helps us put all of this together to build agents. We do that in the agents studio. You can also do this on your phone, but if I open up Agent Studio here on Desktop psychic's, just gonna start a conversation about what you want to build together.[00:09:51] I'd love to show you one that I made recently.[00:09:53] swyx: Let's do[00:09:53] David Singleton: it.[00:09:53] Building A Conference App[00:09:53] David Singleton: Um, let's look at something that hopefully is kind of near and dear to your heart. So one of the things I love about Dreamer and this kind of moment in technology is that if you think about it. There are all these things in your life where, have you ever gone to a conference?[00:10:09] I know you have. Right? And, uh, big conferences have apps. Um, and these apps are usually built by agencies and they're, they're usually actually quite expensive to build. I've been involved in running some of these myself. And how many conferences have you been to where the app was good? Zero. Honestly.[00:10:23] swyx: Exactly. Zero,[00:10:24] David Singleton: maybe one. I, I've, I've been to one conference. That was pretty good. Wait, wait session sessions. Um, but, but the point is, they're rarely great pieces of software. Right. And they're also expensive to build, but they're, they're interesting ‘cause they're episodic, they last for this one thing. Um, and then they're, they're not relevant anymore.[00:10:43] Um,[00:10:43] swyx: and so it's the worst feeling to invest in them because, you know, it's like, it's got a limited. Date?[00:10:48] David Singleton: Absolutely. So I decided to build, uh, a conference app for your AI engineer conference. Amazing. Uh, on Dreamer. One of the things that Swix has done, uh, which I [00:11:00] thought was very forward-looking, is actually put a whole bunch of data about the conference on the webpage in an LLM readable way.[00:11:06] There's an LLMs txt file, there's a feed of all of the sessions in js, ON. So I used the data from your conference last year and built this intelligent app, uh, just by talking to our sidekick, uh, in Dreamer. So just to give you a quick tour, this is my Dream Conference app. What I always wanna do for conferences is I wanna be able to search for speakers.[00:11:28] I'm usually there because, uh, there, uh, is a speaker I care about. So, you know, SWIX, you're the speaker I care about. I can actually see here who you're on stage with. So here's, here's Greg Brockman. You've read even ai, uh, and this is his session. And look Greg and Swix for the speaker. So let's add that to my schedule.[00:11:45] Great. And then maybe there's a couple others I might see here. Like on day two, I remember there were some keynotes. So, uh, building the open agenda web, that sounds fun. So I add that to my schedule.[00:11:55] swyx: She's now CEO of Xbox.[00:11:56] David Singleton: Awesome.[00:11:57] swyx: Which is interesting. So cool. So,[00:11:59] David Singleton: so I've [00:12:00] gone through and picked out a couple of sessions that I cared about.[00:12:03] That's as far as I usually get with any conference app. But of course you've got the whole of the rest of the conference to figure out what to do. So here is where the native intelligence of, of these things you build on Dreamer can come in. So I'm gonna click guide me. So Dreamers sidekick actually parsed out the whole schedule and figured out what some of the themes are and I can choose what I'm interested in here.[00:12:23] I'm definitely interested in agents. Uh, I'm definitely interested in code generation and also reasoning in rl. So now I'm gonna say build my schedule. So what this is doing is. It's going across every time slot for the conference. And it's choosing among the things I could go to, which one it thinks is best for me based on my interests.[00:12:41] It also uses its own memory of me that's part of Dreamer, uh, to understand what I might like best. And you know, there's an LLM prompt running for each one of these time slots. So this is, it's not super fast, but it'll be done in about 30 or 40 seconds. And I'm gonna have a special custom schedule for the conference.[00:12:57] This, like I said, is my [00:13:00] dream conference app is exactly what I've always wanted and I was able to build this yesterday morning. Um, I did it between some meetings. I think I spent a total of 25 minutes of wall clock time on it. I did it over the course of a couple of hours. And, uh, here is my schedule for the conference.[00:13:15] I can see it in a calendar view. This is what I should do on Tuesday, this is what I should do on Wednesday. Oof, no conflicts, but, you know, I may not go to every single thing. And there you have it built in, you know, dreamer. So let's take a look at what the building experience actually looks like. So this is the, the actual account that I made it on.[00:13:32] Oh, of course I should say anything you build on Dreamer also works on your phone. So, uh, here is my AI engineer conference app right here on my phone. Got all the same functionality, and of course this is the best place to jump into my schedule.[00:13:46] swyx: Yeah.[00:13:46] David Singleton: Um,[00:13:46] swyx: so you could generate a podcast about it just completely multimodal, absolute thing, right?[00:13:51] To me, I mean, this is why I outsource, I mean, well, I, I posted the L-M-T-X-T, the JSON because you cannot run an engineer conference in 2025 [00:14:00] and not let engineers. Do whatever they want.[00:14:02] David Singleton: Yeah.[00:14:03] swyx: And since all conference apps suck, I'm just gonna put up a ba minimum viable app and just let people do whatever they want.[00:14:09] David Singleton: Totally. And the cool thing about this on Bremer is I published this to the gallery and you can use it so you've got one that's built to my taste of conference apps. I think it's pretty cool. But you might want something different. Yeah. In which case you just start telling the sidekick how to change it.[00:14:23] So let's just very quickly look[00:14:24] swyx: at our, what sports grid is also, you can fork it, right? That I can publish. That's right. I can publish your one and go, this is the base starter. It's, it's got good defaults, but go customize, whatever.[00:14:32] David Singleton: That's right. That's right.[00:14:33] swyx: Yeah.[00:14:33] Agent Studio Under The Hood[00:14:33] David Singleton: So let's take a look at how I actually built this.[00:14:34] This is real. So I'm gonna say make changes. This experience we're looking at now is our, uh, agent development studio. Um, like I said, you can do this on your phone as well. And in fact, this one I started out on desktop. Let's look at my actual prompts. I said, let's make an agent called AI Engineer Schedule Planner should be a custom schedule planner for the AI engineer conference.[00:14:53] I'm not gonna read this all up. You get, you get the point and it told it where to get the data from. So that was the first prompt. And actually after I gave it that [00:15:00] prompt, I actually had a simple version of this app working, um, after the sidekick took one turn. So the Sidekick is a, like a professional software engineer, and we've worked very hard to make this work and build functional apps for folks that might not have any engineering experience whatsoever.[00:15:14] So, you know, done here we have build logs that are technical, but you can hide those away. And sidekick, as it is building, will actually translate everything that is coming out of, uh, of the, the harness into English that you can actually read. And by the way, this English is in the personality of your sidekick, which is fun.[00:15:32] Um. And the way that we build agents and agent apps, it's a little different to what you might have seen in some other platforms for a couple of reasons. One, just the build process. The very first thing that Sidekick does, it understands all the agents you've got set up. It understands all the tools and it will come up with a plan for how to realize your goal, how to make sure it actually has the data and the capabilities to complete it.[00:15:54] It will occasionally refuse. If it can't do what you're asking, it will tell you I can't do that. It needs another tool. And that's a good [00:16:00] jumping off point for any of the tool builders out there to build a new tool. So it'll fi first figure out how, then it will build it, and then it will actually test it.[00:16:07] So it will actually make sure that the thing that it has generated is realizing your goal. And you probably know as well as anybody that anytime you can get any. Modern state-of-the-art coding model into a loop where it can make changes and perceive its own output and then fix bugs. Magic happens. So these builds, the first build will often take 10 to 15 minutes on Dreamer, which is a little bit longer than you might've seen on some other platforms.[00:16:31] But the first thing that it creates will work most of the time. And then of course, as you start making smaller changes, you can like ask it to tweak the UI in any way that you like. Those are much faster. And just to give you a sense, uh, for this one, here's something I asked. Put a logo, I gave it a logo file in static files.[00:16:48] Use that as the title. So for folks that actually really want to dig, uh, into a bit more detail, we've provided a powerful IDE here. So I can actually see here's the code that was generated and some pieces of the [00:17:00] code are more accessible than others, like the prompts. So this is the prompt that's used by a powerful LLM in order to do that schedule picking.[00:17:08] And I can actually read it here directly. I can edit it without having to ask the sidekick if I want to do that.[00:17:12] swyx: So this is very nice.[00:17:13] David Singleton: This is for the more, the more, uh, sophisticated users.[00:17:16] swyx: Yeah. This is other people's entire startup is prop management.[00:17:21] David Singleton: This is true. The other thing that is different about Dreamer is once you've built something here, it's ready to go.[00:17:28] We host it. So you don't have to worry about getting a database from a database provider signing up, getting API keys. You don't have to worry about your LLM provider tokens. All of that is hosted on the platform. And you can use it yourself. You can share it to the gallery for other people to, to riff on it.[00:17:46] You can also share it with your friends and coworkers to use your instance of the agent or agentic app. And we're seeing that happen a lot in our community. We've seen a whole bunch of folks who built little applications for their personal life [00:18:00] and shared them with their significant other. We've seen people who are building little productivity apps for their team at work and sharing it, uh, among them.[00:18:07] And we actually do this a lot inside of the company. So at this point we, we pretty much run the company on Dreamer agents for all kinds of important things. Uh, maybe a good example of that is, um, our wait list. People are signing up every time someone signs up for our wait list. A dreamer agent will actually research, uh, that person.[00:18:25] And we're looking for folks who are builders, not super technical to build agents and come in, uh, and give us a lot of feedback and we're prioritized bringing those people off of the wait list First,[00:18:35] swyx: just a quick question on that one is there's, it may not come up again. Do you find enrichment APIs to be useful like the ZoomInfo?[00:18:42] Uh, clear bit[00:18:43] David Singleton: enrichment is a very, uh, common use case. Um, on dreamer. Any application on Dreamer can kick off a sub-agent to do a particular task. Um, so this actually is a powerful agentic harness that runs inside of its own [00:19:00] vm. Uh, we call them sidekick tasks ‘cause they actually run in the context of the sidekick.[00:19:04] I'll talk more about Sidekick in a second and. Enrichment is a very common use case. And the cool thing about a sidekick task is that it has access to all the tools on the platform, but also public data as well. And so very frequently enrichment on our platform happens using public data that it can be found in the web.[00:19:24] There are some tools for getting people data, uh, from, uh, from various bespoke systems. And so that works pretty well. But actually, you'd be surprised. I mean, we would love if someone out there would like to build a ZoomInfo tool, we don't have one today. We'd love to see that on the platform, and I'm sure it'll be very powerful.[00:19:39] But we're also seeing that this powerful agent harness can pull a lot of data in on that note of tools that make experiences better, we're constantly adding more tools because people in the community are building them and publishing them. We review the tools carefully and then they go live for everybody.[00:19:54] Yesterday we added granola. And that was pretty cool. So I was talking to actually, uh, Sarah on my team was [00:20:00] talking to, uh, someone building on the platform this morning and they actually, they have an agentic app that they built, which is a kind of magic to-do list. So they put stuff on their to-do list and for each thing it kicks off one of these, uh, sidekick tasks to figure out how to move the ball forward thing.[00:20:14] Sometimes it'll complete it[00:20:15] swyx: entirely. Yeah.[00:20:16] David Singleton: Often by calling another agent on the platform and sometimes it just kind of researches it and helps ‘em take the first step.[00:20:21] swyx: Yeah. Do you know, this is Sam Altman's number one, ask for an AI app. It's the self-completing to-do list.[00:20:26] David Singleton: Yeah. The self-completing to-do list is something that a lot of people have built on Dreamer and are getting a lot of use out of.[00:20:32] Yeah. And, and finding it actually genuinely I shouldn't, I should, I should try that. Mm-hmm. Please do. And you'll even find some in the gallery that you can remix. So he was saying this morning that he's, he built this self completing to-do list, uh, on Dreamer already. But he connected the granola tool yesterday and now something really magical happens, which is when he says in meetings that he's gonna do a thing, it magically shows up on his to-do list and then it can magically get completed.[00:20:56] And then, as I mentioned, all the agents, all the [00:21:00] apps on Dreamer can actually work together. So our coding agent, as it builds them, does something very special where it exposes the internals of each of the experiences to the system. And then Sidekick can manipulate those to get stuff done. So he has built another agent, which he uses for recruiting.[00:21:18] It kind of keeps track of candidates and also it's got a kinda mini CRM function, so he's able to introduce candidates to each other. He told us this morning that something he'd committed to do in a meeting that was recorded on granola yesterday showed up in his magic to-do list and his magic to-do list.[00:21:34] It was like introduce a person for recruiting, used his recruiting agent to get it done.[00:21:39] swyx: Ah,[00:21:39] David Singleton: um, and this is, this is the dream. This is why we started the company. It really is the case that you can build and use these very powerful, bespoke experiences that can automate your life by working together. And I'd love to talk a little bit about how they work together.[00:21:55] Ecosystem Trust And Monetization[00:21:55] David Singleton: So obviously it's really cool to have [00:22:00] software that will work on your behalf, but it's only useful if you can trust it, right? So privacy and security is very important to us making these things accessible and. While also being trustworthy is hard. So the model that we have, which is working very well, is that the sidekick is at the core of everything here.[00:22:22] So it is both your companion, your helper, but it's also the traffic cup in the system. So when, when one agent wants to work with another agent and dreamer, it doesn't do it directly, it does it via the sidekick, well ask the sidekick to do the thing. And the sidekick understands both everything, all the expectations that have been set with me as a user about what agents can do, which tools I've given them permission to use.[00:22:45] And it will make sure that whatever is is going on is actually aligned with my own interests. And you know, that's part of the background that I bring to this problem domain. I've. Worked for years, uh, keeping very important information, safe and secure. And [00:23:00] so as we started to think about this problem, we realized that we actually had to build something that's a bit like an operating system.[00:23:06] You know, the sidekicks, like the kernel, the agents and apps are like users. Yeah. Different rings. Exactly. Because if you try to pick off just one piece of this, you can't actually make it work for people at scale. Uh, because you could build little vibe coded apps, but they're gonna grab all your data willy-nilly.[00:23:23] They won't be able to work together. You actually have to invest in the fundamental core in order to make it work well for people. And that's what we've been doing and it's, uh, it's been a lot of fun. One other thing I wanted to mention is, um, I've obviously talked about two things, tools and agentic apps.[00:23:42] We really designed Dreamer to be an ecosystem and a platform, and one of my favorite quotes about platforms, I think it's from Bill Gates, is that you can only be a platform. If you create more value for the folks participating and using the platform than, than the platform itself creates. [00:24:00] And that's our goal here.[00:24:01] So we at every step have been thinking about how do we make sure that other people are deriving even more value from Dreamer than we are? So in that vein, I already mentioned tool builders get paid and people can build agents that solve their needs and share them with others, and we are already thinking about ways that they can actually monetize those as well.[00:24:24] Against that backdrop, one of the things that we are launching today is our Builders in Residence program. So there are tons of people building really cool stuff and contributing it to the gallery already, but we've been really inspired by programs we've seen at other companies where artists might be in residence, people that are very creative.[00:24:43] And might have ideas outside of what the, the folks at the company or in the ecosystem already have. And so we are looking for creative people who have fun ideas and, you know, want to really figure out how to apply their creativity at the cutting edge [00:25:00] of technology today to come and work with us. So, uh, if you go to dreamer.com/latent space, you'll find, ooh, well, we love Latent space.[00:25:09] Uh, you'll find a link both to, uh, our tool Builder information and our builder in residence program. And for builders and residents, we'll let you in off the wait list quickly, build an agent, and then for a small number of, of the most creative folks, we're going to pay you to build agents. Uh, you can work directly with our team.[00:25:29] You know, this is like building Legos. So, you know, we've got some of the basic blocks together already, but if you need a Ron steering wheel and we don't have one already, like we'll build it for you. Yeah. Um, we really want to be inspired by, by these, uh, these builders in residence.[00:25:43] swyx: This Legos thing is pretty common as an analogy.[00:25:46] And there's a, there's a thing I call the master builder. Uh, we, the actual Lego company has master builders that they employ Yeah. To inspire people and post on socials.[00:25:56] David Singleton: That is exactly what inspired us as well. Honestly, we talked about the Lego Master [00:26:00] Builder program, so that's our builder in residence program.[00:26:02] swyx: Yeah.[00:26:03] David Singleton: Um, and then, uh, finally back on, on tools. Like I said, anyone can come in and build tools today. If you follow the latent space link dreamer.com/latent space, again, we'll get you off. Directly off the wait list. So you can build right away, you can monetize by publishing onto the platform. That's for everyone, the very best tool that gets added to the platform by mid-April.[00:26:23] Uh, we have a $10,000 prize that we want to give out really, because we just want to seed the creativity of everyone out there. So we're excited to do that.[00:26:31] swyx: Yeah. And you know, uh, this is completely a flywheel, right? Like the more tools, the more builders, the more the third thing agents, you know, it just feeds into each other.[00:26:39] David Singleton: That's right.[00:26:39] swyx: Yeah. Just on the payments thing, because we probably won't touch on that again, but I have to ask the former CTO Stripe on payments as presumably you're using Stripe Connect.[00:26:48] David Singleton: Yeah.[00:26:48] swyx: Um. Any pain points that you're, people are very interested in agent commerce and micropayment and all these things.[00:26:55] Presumably stable coins get into a conversation at some point, but maybe not now.[00:26:58] David Singleton: Yeah, we are [00:27:00] really, really excited about e agent commerce. The first step we are taking is help people in the world who have never been able to build these kind of experiences and software before to build stuff that meets their passions, share it with the world and get paid.[00:27:14] So that's all commerce that happens on our platform, and so we don't need anything new to facilitate that. Stripe Connect has existed for quite a while and is the perfect solution for this kind of stuff, so, um, we we're excited about that. First and foremost, however. A lot of the things that people are already doing on Dreamer, we just talked about a self-completing to-do list.[00:27:34] A lot of the ways that you want to complete to-dos is by actually closing the loop in the real world, and that's going to involve the exchange of value. So we have some folks that are building tools already that actually do have money move in order to, to complete that, that loop. So far, we just want to be open and agnostic to all the protocols out there.[00:27:54] I honestly think this moment in time is a little bit like the early web. So I personally started coding as a kid [00:28:00] and I think I got access to the internet in about 19 95, 19 96. And back then, uh, the web existed, you know, HTTP was a protocol, but there were also other protocols I was using all the time, like Gopher and UUCP and uh, various others.[00:28:15] So the point is like the web, HTTP and HTML. Was just one among many protocols. And of course it became the winner and it's awesome. Yeah. Um, but the others were also kind of interesting and viable at the time as well. And I think the world of agentic commerce is like this right now. Also,[00:28:30] swyx: acp.[00:28:31] David Singleton: Acp, exactly.[00:28:32] All the, all the cps, you know, on Dreamer. We hope that folks will build tools that kinda make use of all of these things, but I'm sure that at a certain point. One or two will emerge as the winners, and then we'll be able to build like really deep support in,[00:28:44] swyx: yeah. This is like maybe a complete tangent, but I do think about how a lot of these companies in AI companies in particular have to switch from c based to usage based because of course, but then, then they end up, end up having to sort of [00:29:00] obscure the margins a little bit and then they inventing end up inventing their equivalent of rob robots.[00:29:04] David Singleton: Mm-hmm.[00:29:04] swyx: Uh, where they're like, well, okay, well every company should have their own currency. And it's, it's like very short lead to a token.[00:29:11] David Singleton: Yeah.[00:29:11] swyx: Or, and I'm like, okay, well where does this end? I can't really play out the next step as to like, is this chaos? Is this,[00:29:18] David Singleton: yeah.[00:29:18] swyx: Okay.[00:29:18] David Singleton: Well, I think it is kind of like the wild west.[00:29:21] I don't mean that in a completely, it's all completely disorganized way, but there's just so many things that could happen from here. The Overton window is very wide, right? Not far how this might land. And I'm just very excited to be building a platform that can take advantage of all of those opportunities and we're just gonna be there.[00:29:36] Uh, working for our users to make sure that things that emerge work,[00:29:39] swyx: you're gonna own the consumers, you're gonna be up the OS for the app store for everything.[00:29:43] David Singleton: So one of the ways to think about this is, um, dreamer actually uses all of the state-of-the-art models as a user. You don't have to think about should I be using, you know, Opus four six, or should I be using the five four model from [00:30:00] OpenAI?[00:30:00] We are continually doing evals and so forth to make sure that the best things are there for you. You can just build on the platform and know that as the world ships around, you're gonna get the right stuff for you. Um, and I think that's something that is needed to actually have folks take advantage of this technology at scale.[00:30:19] I'd love to show you another example of something I built.[00:30:21] swyx: Let's do it.[00:30:22] David Singleton: This is another example of software that just lasts for a certain moment in time. So recently I went on a ski trip with a bunch of friends,[00:30:31] ski[00:30:31] David Singleton: Bum. Uh, so it uses ski bum. Yes. I went on a ski trip to Big Sky. I'd never been there before.[00:30:38] And I made this little intelligent app for us. And you can see it says it's loading big sky conditions. So it's actually calling the Ski Bum tool that I just showed you, which is, uh, published in our, uh, in our gallery. So what is this? This is a little app that was just for our weekend trip. It shows the current status of all the lifts of Big Sky.[00:30:54] Using that tool from the ecosystem, it shows the forecast for the upcoming weekend. It shows our [00:31:00] accommodation. This is just like where my group was staying. This is just for us and also a bunch of dining information that one of our friends, uh, put together who, who's an expert on Big Sky. So I was able to take this app, share the link with my friends.[00:31:12] They weren't on Dreamer yet, just send it to them on iMessage and they get a version they can use on their phone. And of course, here's the real kicker. So I've been on ski trips before and other weekend adventures with my friends. Yeah, people pay for different things and at the end of the weekend it's always a pain to figure out who needs to pay, who to settle up.[00:31:29] So we use this during the weekend. We added all of our expenses in here. Uh, too close are it's drill data. It's only too closely. And then at the end of the trip, we press split. And we're, we settled up and we're done. So there's another dreamer. This was all through dreamer. So the, the actual payment? No, no.[00:31:47] We, it happened because, because we paid for stuff in the real world, it was like, okay, this person needs to pay that person 20 bucks. Right? Right. This person already paid in that. Right. So it just helped us all settle up. We didn't move the money on Dreamer. You could do that. And in fact, if you're a tool builder [00:32:00] thinking about this and getting excited, like come build a tool to do that stuff.[00:32:02] We really think of our tool builders as design partners.[00:32:05] swyx: Yeah. I got, I got the tool. Uh, what, like, I hate, I use Bank of America. I hate bank, I hate the app. Mm-hmm. I hate the web. All banking websites just horrible.[00:32:13] David Singleton: Yeah.[00:32:13] swyx: So just build me, like build a thing on top of Plaid.[00:32:15] David Singleton: Yeah. Right. And then just So[00:32:17] swyx: five code by banking app,[00:32:18] David Singleton: there's already a tool for that.[00:32:20] Oh. So, um, attain Finance is a tool, a builder in our community built. Okay. Um, and it uses a secure system like Plaid. To access your, uh, financial data and you can build powerful personal finance agents on Dreamer today using this tool. And like I said, we review tools carefully. So when bringing Attain Finance onto the platform, we did actually quite a detailed security review with that company to make sure that if folks build stuff with it, it's, it's gonna work well.[00:32:49] So yeah, check that out. I think, uh, I'm, I'm pretty certain it connects to Bank of America. So you'll be able to build the, the app that you wanted already?[00:32:55] swyx: Yeah. There's a couple of points I wanted to sort of dive in on, maybe highlight to folks, [00:33:00] because I, obviously, I spent more time with Dreamers. So we're making a point where you choose on behalf of your users because they're meant to be consumers.[00:33:07] So maybe less technical,[00:33:08] David Singleton: right?[00:33:08] swyx: But obviously people can, how users can override. If you read that's, but it's not just lms, it is also the, the transcription. It, it's like all, like there's, there's a first party curated set of here's the house opinion. That's right. On what?[00:33:21] David Singleton: That's[00:33:21] swyx: right. The thing is, that's right.[00:33:22] Is what's the list? Is there like,[00:33:24] David Singleton: yeah, so actually if you look in the tool gallery, the first party kind of curated set are all the ones that have these grayscale icons. So we have a built in tool for image understanding, for image generation, for RSS, exploration, text to speech and so forth.[00:33:38] swyx: Recipes.[00:33:39] David Singleton: Uh, we actually do have a built in recipes tool.[00:33:41] It turns out that a lot of people in our alpha wanted to do stuff for cooking. Yeah. Um, and you know, you can scrape the web to get good recipes, but we were able to quite quickly find a good repository of recipes. It works great here. Yeah.[00:33:55] Stable Tool Interfaces[00:33:55] David Singleton: So the point behind these though is that we'll keep the interfaces stable, so they'll always work.[00:34:00] But you know, the best translation model and, you know, there are people using this translation tool to translate Chinese podcasts into English. It's, it's pretty powerful. It can deal with very long text, but the best translation tool today might be different from the best translation tool sometime next year.[00:34:15] And we're just gonna make sure that that translation tool is always pretty close to state of the art. So you can build something and you know it's gonna continue to work well. Of course, some of our tools are branded. You may actually have a preferred way of buying groceries, like maybe you prefer Instacart and that's great.[00:34:29] You can use the Instacart tool specifically.[00:34:31] swyx: Yeah.[00:34:32] Partnerships And Ecosystem[00:34:32] swyx: Your partnerships, uh, I mean, I don't know if you ever hit of partnerships, but this is gonna be a bonanza for anyone on to do deals.[00:34:38] David Singleton: We have an amazing person who, uh, works on all of our partnerships. Um, and it's part of what you have to do to build a platform like this that's gonna work for people.[00:34:46] Like, we've gone and done that. Schlep has a lot of work, one talks lots of different companies, um, in order to make sure that you've got good tools at the core.[00:34:54] swyx: Yeah.[00:34:54] David Singleton: And then of course, because we're open to tool builders contributing to the platform, this is only gonna get better and better and [00:35:00] better.[00:35:00] swyx: Yeah.[00:35:01] Agent Lab Routing Layer[00:35:01] swyx: One observation I have this, this is gonna master a thesis I've been pursuing, which is, uh, what I've been calling an agent lab[00:35:05] David Singleton: mm-hmm.[00:35:06] swyx: Where you sort of different than a model lab in, in, in the sense that you never train your own models, but you are the router evaluation layer, ex subject domain expert for choosing between, uh, models.[00:35:18] David Singleton: Yeah.[00:35:18] swyx: And you're explicitly doing these things. And so like in my sort of construction, every agent lab does some version of this where like, here's the image understanding endpoint and we will route for you and don't worry about it. Yeah. Sally, I think it's kind of cool.[00:35:32] David Singleton: I, I think it makes total sense. Um, and again, to make this work for folks that don't follow the AI news every day, it's an actually, it's a, it's a really important thing to do.[00:35:42] Yeah. And it, it's been, it's been a real pleasure. I mean, I'm a, I'm personally a total geek for this stuff. I love it. And being able to go and dive into all those details in order to make it work well for other people. It's a true pleasure. I cannot imagine working at anything else right now. It's just so much fun.[00:35:56] swyx: The tricky part is multimodality when some of these things do [00:36:00] merge.[00:36:00] David Singleton: Mm-hmm.[00:36:01] swyx: And you are, you're sort of, this is your imposing structure on things that fundamentally don't want to be structured. And so sometimes that might work against you, but for 99% of these cases, this is fine.[00:36:10] David Singleton: Yeah. I mean, I think it's gonna be very interesting to see how the, the, the world matures because a lot of the power of dreamer is the ability to kick off these subagents, so these powerful agent harnesses, which can actually change how they work based on the data.[00:36:25] I actually think that we will be able to. Kind of keep up with and stay at the forefront of the changing landscape of how tools and systems work together. And that's, that's new. You know, software didn't used to work like this and now it does. Um, so even, even just figuring out how to design the right pri to make that possible has itself be a lot of fun.[00:36:44] Builders Can Publish Tools[00:36:44] swyx: This is, is a sort of maybe two part question that why can't streamer make its own tools? And then why don't you let you builders maybe stand up their own routing group? I call this a routing group, right? Like where it's like collect Yeah. Things.[00:36:58] David Singleton: So two things, to [00:37:00] some extent, dreamer does make its own tools in that agents appear to the system as tools.[00:37:05] So they can be, they can be used to accomplish things. So you can build an agent that is essentially a tool. Yeah. Um, and it it,[00:37:12] swyx: which is to me very useful for reuse.[00:37:14] David Singleton: Right.[00:37:14] swyx: Right. Exactly. ‘cause I, I like, this is the way I like it. Now my next five apps, I don't want to do this whole series of back and forth again.[00:37:20] David Singleton: Right.[00:37:21] swyx: Yeah.[00:37:21] David Singleton: Um. Then at the tool layer of the system, it's open to anyone. So it's actually quite powerful and flexible. So if you wanted to add a tool, which was, uh, imagine that you were training your own foundation model, Swyx. That might be fun. And imagine you wanted people to be able to play with, I don't know, maybe you make like, you know, nano chat or whatever and you want to Yeah.[00:37:42] Let people play with your own nano chat and see how I change themselves.[00:37:44] swyx: Now.[00:37:45] David Singleton: You could, you could publish a tool that is Nano Chat and it nano image generation behind a tool, and it could be your own writer if you wanted to. I see. And honestly, if that's the kind of thing that gets you excited as a builder, please come and do it.[00:37:57] Like we, we really are [00:38:00] believers in this idea that we aren't going to figure out every single detail ourselves. We're gonna make sure it's a safe and fun place to build this stuff, but we're really open to these ideas coming from other people. Um, and so I'd like nothing more than you come in and build a tool that does some of that cool stuff that you, that you have in mind.[00:38:15] swyx: Yeah. Awesome.[00:38:16] David Singleton: And just as a reminder, if you'd like to do that, the way to find the links is dreamer.com/latent space. Um, and for a limited time on that page, um, anyone who's listening to this podcast will also get directly off of our wait list. Uh, it's quite long right now. We are working hard to bring Zika.[00:38:32] Wait, so skip the wait list.[00:38:33] swyx: You know, I think, I think that's fantastic. I, I think it's, it is really sort of probuild way to do it. I wanted to jump back to the, the bar. Yeah. You know, you know, I get excited about this.[00:38:41] David Singleton: Yes. Okay. Let's set it back in there.[00:38:43] swyx: Like, let's, you know, this is the engineer podcast that's get[00:38:46] David Singleton: Yeah.[00:38:46] swyx: As technical as you can.[00:38:47] David Singleton: Yeah.[00:38:47] swyx: On everything you've built, like have a show off.[00:38:50] David Singleton: Yeah. Okay.[00:38:51] Under The Hood Debugging[00:38:51] David Singleton: So let's go wild in the aisles in the Asian studio. So as you can see, over on the left here is a conversation with the sidekick where you ask it what to do and it will explain in English that anyone can understand what's going on.[00:39:03] But, um, if you want to pull back the covers and look under the hood, um, if you're, uh, an engineer like me, then we have this, uh, this kind of debug drawer at the bottom. So you can see the full build logs here, but you can actually also dig in and see the files and prompts that have been generated. Uh, you can upload files from your computer in static files.[00:39:24] Um,[00:39:24] swyx: very important,[00:39:25] David Singleton: uh, indeed. You can actually read the prompts that have been generated for you. We intentionally put an example in here just that you can see what the format looks like. And then, you know, we already looked at this one that was generated for this particular, um, app, but if you actually want to bring the code out of Dreamer and work on your own local machine, you can.[00:39:45] So at the core of everything here is an SDK with a powerful command line interface and we built that first. It's actually possible to build agents on Dreamer without talking to the sidekick. You can write code with your fingers on a keyboard if you want to. I know that's very [00:40:00] antiquated, not, but actually this can be a lot of fun.[00:40:02] So if you wanna pull it out onto your laptop, you can use our, our CLI and, uh, you can edit it in cursor or in cloud code. You know, you don't have to use our sidekick. And the CLI actually has full access to the rest of the platform with you as the user. So, you know, obviously it is, uh, secure and privacy sensitive, and this is a way that, um, some of our most technical builders do build stuff on the platform.[00:40:24] The really cool thing is the side cake. When it's in coding mode, it uses exactly the same CLI. So the way it. Build stuff on Dreamer is using the same tools that you might as an engineer. Um, and that's actually a very powerful abstraction because it turns out that the right way to give a lot of context to agents to use CLIs is to write great documentation.[00:40:46] Make sure that all of the things that you could do are actually possible. And guess what? That makes it a delightful developer experience for real heroes as well.[00:40:53] swyx: Yeah. So that's pretty cool. We've been telling developers to do this and they ignore this until now they have to for content.[00:40:58] David Singleton: I, I've been saying this for a [00:41:00] long time.[00:41:00] Uh, we actually Stripe docs.[00:41:02] swyx: I mean, come on. Absolutely. Come on.[00:41:03] David Singleton: Absolutely. But actually, I was chatting with folks at Stripe last week and saying, Hey, you gotta make the Stripe CLI actually tell agents what they can do on Stripe because that way they're gonna use more stuff on Stripe. I think this is a real trend for the entire industry.[00:41:16] swyx: Yeah.[00:41:16] David Singleton: So we, we've been doing that.[00:41:17] swyx: To me, this, this download and, uh, GI push mm-hmm. Everything is complete confidence in that you're not hacking it. Right. Because there's other, let's call them AI builder platforms that impose their stack on you and if you, if you, and so therefore they don't allow you to do this because they cannot.[00:41:34] Right. ‘cause they, they impose some degrees of freedom, uh, restrictions so that they can get it to work. Yours is a fully general like VM running the full code. Correct. Do whatever you want. Correct. Any language you want. Correct. Yeah.[00:41:46] David Singleton: Correct. Well, in terms of language, if you use the SDK, you could build stuff in other languages.[00:41:51] We've actually found that TypeScript is the best language for building these experiences. Yes. Because it's strongly tight. So you find out at compile time if you've made mistakes [00:42:00] and there's nothing better than getting in. A coding agent in a loop where it can see its mistakes and ask them. So TypeScript is the language that everything gets built in by default here.[00:42:08] swyx: Did And did you see that TypeScript overtook Python? I did. I did. Yeah.[00:42:12] David Singleton: And for what it's worth, when we started the company, we started writing stuff in Python, and I love Python. Um, if I do, uh, a vendor code, I always write it in Python. It's my favorite language as a developer with my fingers on the keyboard.[00:42:23] Um, but TypeScript is an amazing language for AI because there's tons of training data in the models, um, and it's strongly tight. And actually at the company we built most of the stack in TypeScript, and we have this amazing property, which is, we have type safety all the way from the database to the front end.[00:42:40] And there's nothing better for working with coding agents than being able to have them check their correctness, compile time. So the same ideas behind building the company's code base, we've put into the agent SDK here as well.[00:42:51] swyx: Yeah. Do you know if you'd use one of those tools, like Prisma or whatever, or is it Tool Lab for you?[00:42:55] David Singleton: We, we actually have crafted most of our own tools. Um. For [00:43:00] instance, we had LLM Driven Code Review, uh, before the thing that got published from philanthropic this week. You know, we, we've been doing this stuff, uh, on our own bat[00:43:07] swyx: email, we'll pay $25 per review.[00:43:09] David Singleton: We, we pay a lot less than that. However, I hear that those reviews are excellent and possibly worth $25.[00:43:14] swyx: Yeah. You know, it's an option. Right. It's good, good to have it.[00:43:17] David Singleton: Just to give you a tour of some other stuff here. So, um, I can also see all the versions. Yeah. Um, this is not gi, this is not gi, this is built into dreamer. I can see all the versions that have been pushed before. Why is it[00:43:27] swyx: not gi?[00:43:28] David Singleton: It's not gi because we can make it work more efficiently than Git.[00:43:32] And we actually, we do some work behind the scenes to kind of understand what's in each of these versions. Yeah. Um,[00:43:37] swyx: so one of the things I'm pursuing, and I have a lot of thesis, right? Mm-hmm. One of the thesis is like, does GI go away? Does GitHub go away? And like, what, what is the active reinvent[00:43:46] David Singleton: you for, for what it's worth to some extent.[00:43:48] And anything you build, there's a lot of path dependency. If we started over, we might make this gi There's, uh, you know, within the company we use, uh. For our, you know, platform source code. And we like it and it [00:44:00] works well with coding agents as well. The very first versions of this, we wanted to be able to make it possible for the sidekick to manipulate it easily.[00:44:06] Um, and this, this was an expedient way to do it.[00:44:08] swyx: Yeah.[00:44:08] Workflows Logs And Databases[00:44:08] David Singleton: Um, you can also see all the activity that has happened in the workflows that you build. A lot of agents, you'll build on Dreamer, do things in the background, so they run on triggers. These are stimuli from the outside to kick them off, and this is a nice way to see all of the things that might have kicked off your agent.[00:44:24] You know, you can have an agent that kicks off on a webhook, so you can plug it into external systems. You can have an agent that runs when you receive certain emails that match filters, including LLM filters. And so here you can see, oh, when did it run? What did it do? You know, if I open up one of these guide me prompts or guide me, uh, events.[00:44:41] Oh my can see God. Well, I told you it was calling an LLM for every one of those time slots. Here's all of the LLM calls, here's the actual prompts.[00:44:49] swyx: And you don't mind exposing all of this, right?[00:44:51] David Singleton: No. We want builders to see what's going on under the hood. It's haiku to,[00:44:53] swyx: okay. Yeah. So,[00:44:54] David Singleton: okay. Right now that one was haiku.[00:44:56] Like I said, we work with all the models and sidekick will actually pick the best one [00:45:00] for the job. And you saw that was pretty high quality and pretty fast. So Haiku four five is the one that it picked for that job. Exactly. Uh, we also have logs, as I mentioned, there's a database spun up on demand for every, uh, agent.[00:45:12] You don't have to go and figure out how to do your own hosting. This is a SQL Light. This is a SQL Light database. Yeah. Um, it's a multi-user SQL light database. And then, uh, but, but each one is you, you get a database that is unique to this agent. But then if you share the agent with multiple people, we take care of like who are the owners in each row?[00:45:31] And all of that stuff is just there outta the box. Um,[00:45:34] swyx: and again, in-house?[00:45:35] David Singleton: In-house.[00:45:36] swyx: Oh my God.[00:45:37] David Singleton: Yeah. Um, well we do work with a bunch of infrastructure providers, but the technology for how to manipulate this is in-house. Fun fact. We actually did a lot of our own infrastructure development early on at the company and realized we need to spend our energy in the stuff that we're uniquely doing in the world.[00:45:53] So we're very delighted to partner with a bunch of great designer and some of this stuff. And then finally, um, I mentioned that agentic apps agents [00:46:00] expose all of their internals to the system so the psychic can manipulate them and use them just like a user can. So you can see how it's decided to break this problem up into functions.[00:46:09] Some of the functions, the ones with the little I here are exported. That means that there's probably the visible from outside. Exactly. And others are internal. And if you want to, you can dig right in here and call individual functions and see what happens. But mostly. You don't need to think about that at all.[00:46:24] Yeah. Uh, you can keep that little drawer closed and you can talk to your sidekick and build really powerful and enchanting experiences.[00:46:30] swyx: Yeah. I mean, to me, like showing this gives the engineer a complete mental model of what you've done and what you can do with it. Yeah. For example, the first thing I, I, I look for.[00:46:39] A mental checklist of things, right? Like is off in the database, off looks like it's not right. So that's a separate layer. That's probably me means it's hard to do multi-user apps on the same app, right?[00:46:50] David Singleton: So you actually, we've solved that. So, um, see, yes, the platform builds in off, so you as a user sign into the platform, if you're using an [00:47:00] agent that was published by someone else, then your identity is, is kind of taken care of by the system.[00:47:05] And when you query the database, you're gonna get the stuff that is for you. Unless the builder specifically said, this is public data that everyone should see. So they, they actually get a chance to think about that. And again, sidekick can guide you through building, uh, agents and apps that work that way.[00:47:19] So you're right, that's another thing that people have to think about when they're trying to figure out how to build software experiences on Dreamer. You, it's built in. You talk to the sidekick as if it were a human being about what you want and that's what you get. So, you know, my, my Big Sky app that I just showed you that was designed for multiple people to use it.[00:47:38] And of course the things that we were putting in as expenses were supposed to be visible to everybody, and I just told the sidekick that's the way I wanted it. Uh, but by default, if I built an app like that, the data from each user would not been visible to the others.[00:47:49] swyx: Yeah. Yeah. Uh, this is, I presume this is a mood question, but basically you've had to build your own coding agent, right?[00:47:55] Which is sidekick slash whatever is in Inside Psychic. Obviously there's a lot of [00:48:00] people with a lot of desire for cloud code and Code X and attachment to it. Mm-hmm. I know under the hood data basically reduced to a loop, but like, would you let people use cloud coding and Code X or is the harness too specialized?[00:48:12] David Singleton: Yeah. If you, if you want to use, um, cloud code and Code X, then you go down here. Yeah. Hit get the S St K. And we even say this right here, edits your heart's content Z cursor code.[00:48:22] swyx: Like people want to use it inside of Ick, right? Yeah. They want to switch the engine.[00:48:26] David Singleton: Yeah.[00:48:26] swyx: That's the coding engine.[00:48:27] David Singleton: Yeah. We are not doing that right now.[00:48:29] Um, you know, again, the goal really is abstract the complexity. Yeah. Um, because the real target for. Building agentic apps is folks who can't do this already today. I can't tell you how many users in our community I've spoken to who are like Dreamer has changed my life because I used to have all these ideas.[00:48:50] If only I could find an engineer to help me implement them, I'd be able to get them done. They're free, and now I can talk to my sidekick and, and get it built. I think that's like really how we think [00:49:00] about the people that should get a ton of value and fun, um, out of the platform. And so they're not asking to be able to plug in their their own, you know, coding agent.[00:49:11] And for those folks, the opportunity is massive. If you've never been able to do stuff in code, now you can build stuff for you, for your friends, for your family, for your coworkers. And also there's a huge opportunity for folks who do build stuff in code to actually contribute to this ecosystem. So that's how we think about it.[00:49:28] swyx: Yeah. Amazing.[00:49:28] Personalization And Memory[00:49:28] swyx: That's most of what I wanted to cover Dreamer wise. I think personalization and memory yeah. Is probably like the single most important job of, uh, of the os. Maybe we could talk about that and then I'll, I wanted to zoom out on company building stuff.[00:49:40] David Singleton: Yeah, yeah. Sounds good.[00:49:41] swyx: Yeah. So how do you handle memory?[00:49:43] What, yeah, what have you found? What have you tried and failed?[00:49:45] David Singleton: Yeah. Okay. So, uh, first of all, at the core of dreamer is the sidekick. The sidekick gets to know you and it builds up a memory about you over time, and that turns out to be very important. So Dreamer, that's

Claude Cowork came out of an accident.Felix and the Anthropic team noticed something interesting with Claude Code: many users were using it primarily for all kinds of messy knowledge work instead of coding. Even technical builders would use it for lots of non-technical work.Even more shocking, Claude cowork wrote itself. With a team of humans simply orchestrating multiple claude code instances, the tool was ready after a brief week and a half.This isn't Felix's first rodeo with impactful and playful desktop apps. He's helped ship the Slack desktop app and is a core maintainer of Electron the open-source software framework used for building cross-platform desktop applications, even putting Windows 95 into an Electron app that runs on macOS, Windows, and Linux.In this episode, Felix joins us to unpack why execution has suddenly become cheap enough that teams can “just build all the candidates” and why the real frontier in AI products is no longer better chat, but trusted task execution.He also shares why Anthropic is betting on local-first agent workflows, why skills may matter more than most people realize, and how the hardest questions ahead are about autonomy, safety, portability, and the changing shape of knowledge work itself.We discuss* Felix's path: Slack desktop app, Electron, Windows 95 in JavaScript, and now building Claude Cowork at Anthropic* What Claude Cowork actually is: a more user-friendly, VM-based version of Claude Code designed to bring agentic workflows to non-terminal-native users* Why “user-friendly” does not mean “less powerful”: Cowork as a superset product, much like how VS Code initially looked simpler than Visual Studio but became more hackable and extensible* Anthropic's prototype-first culture: why Cowork was built in 10 days using many pre-existing internal pieces, and how internal prototypes shaped the final product* Why execution is getting cheap: the shift from long memos, specs, and debate toward rapidly building multiple candidates and choosing based on reality instead of theory* The local debate: why Felix thinks Silicon Valley is undervaluing the local computer, and why putting Claude “where you work” is often more powerful* Why Claude gets its own computer: the VM as both a safety boundary and a capability unlock, letting Claude install tools, run scripts, and work more independently without constant approval* Safety through sandboxing: why “approve every command” is not a real long-term UX, and how virtual machines create a middle ground between uselessly safe and dangerously autonomous* How Cowork differs from Claude Code: coding evals vs. knowledge-work evals, different system-prompt tradeoffs, longer planning horizons, and heavier use of planning and clarification tools* Why skills matter: simple markdown-based instructions as a lightweight abstraction layer for reusable workflows, personalized automation, and portable agent behavior* Skills vs. MCPs: why Felix is increasingly interested in file-based, text-native interfaces that tell the model what to do, rather than forcing everything through rigid tool schemas* The portability problem: why personal skills should move across agent products, and the unresolved tension between public reusable workflows and private user-specific context* Real use cases already happening today: uploading videos, organizing files, handling taxes, managing calendars, debugging internal crashes, analyzing finances, and automating repetitive browser workflows* Why AI products should work with your existing stack: Anthropic's bias toward integrating with Chrome, Office, and existing workflows instead of rebuilding every app from scratch* Computer use one year later: how much better it has gotten, why vision plus browser context is such a superpower, and why letting Claude see the thing it is working on changes everything* Why many “AI verticals” may get compressed: specialized wrappers may matter in the short term, but better general models and stronger primitives could absorb a lot of narrow use cases* The future of junior work: Felix's concerns about entry-level roles, labor-market disruption, and whether AI can compress early-career learning into denser simulated experience* Why Waterloo grads stand out: internships, shipping experience, and learning how real teams build products versus purely theoretical academic preparation* The agentic future of the desktop: what it means for Claude to have its own computer, whether AI should act on your machine or a remote one, and how intimacy with personal data changes the product design space* Why Electron still mattered: shipping Chromium as a controlled rendering stack, the limits of OS-native webviews, and why browser engines remain one of the great software abstractions* Anthropic's Labs mentality: wild internal experiments, half-broken future-looking prototypes, and the broader effort to move users from asking questions to delegating increasingly long and valuable tasks* Why the endgame is not just more capability, but more independence: teaching users to trust AI with bigger scopes of work, for longer durations, with fewer interventionsFelix Rieseberg* X: https://x.com/felixrieseberg* LinkedIn: https://www.linkedin.com/in/felixrieseberg* Website: https://felixrieseberg.com/Anthropic* Website: http://anthropic.comFull Video PodTimestamps00:00 — Cheap execution and building all the candidates00:44 — Intro in the new Kernel studio02:47 — What Claude Cowork is04:18 — Why user-friendly can be more powerful05:33 — How Anthropic built Cowork07:09 — Prototype-first product development08:00 — Why local computers still matter09:20 — Skills, primitives, and platform leverage12:13 — Cowork's architecture: VM + Chrome + system prompt15:38 — Felix's own bug-fixing Cowork workflows17:38 — Local-first agents20:16 — Evals, planning, and knowledge-work optimization23:14 — What Anthropic means by evals24:21 — Scaffolding, tools, and why skills matter27:44 — Demo: YouTube uploads and self-generated skills31:03 — Calendar automation and cleaning your desktop34:47 — Browser context and why DOM access matters37:47 — Skills portability and plugins44:36 — Which AI categories survive?46:19 — Junior jobs, simulated work, and labor disruption52:00 — Gradual takeoff vs big-bang takeoff53:42 — Finance, taxes, and enterprise verticals56:24 — Vision and the improvement in computer use57:31 — Why Claude writes its own scripts58:06 — Should Claude have its own computer?1:01:26 — Windows 95 in JavaScript1:03:19 — VM tradeoffs and sandbox design1:07:23 — Approval fatigue and safe delegation1:11:18 — The future of Cowork1:12:27 — What comes next for agentic knowledge work1:15:13 — Electron, Chromium, and desktop software lessons1:22:16 — Multiplayer agents and coworker-to-coworker workflows1:26:05 — Anthropic Labs and closing thoughtsTranscriptAlessio: Hey everyone. Welcome to the Latent Space Podcast, our first one in the new studio. This is Alessio, founder of Kernel Labs, and I'm joined by swyx, editor of Latent Space.swyx: Yeah, so nice to be here. Thanks to, uh, TJ, Alessio, Allen helping to set everything up. It looks beautiful. We even have the logo outside.Yeah, kind.Felix: It's like really nice, right? When you walk in here as a guest, you're like, ah, this is a serious production. You're like, feel it immediately.swyx: Yeah. Felix, you've been, you're, you're currently a product manager of Cowork or,Felix: uh, really Technicswyx: Eng. Yeah. The, the identities are kind of vague member technical staff.Felix: I know member staff is like, the official title will carry around forever.swyx: Yeah. I basically kind of wanted, like we've been. Kinda obsessed. I, I've been using it a lot, even for managing latent space. Like, uh, cowork helps me upload videos and like title things and like edit and everything. It's, it's like really amazing.Alessio: Cool. He said multiple times Cowork has said gi in the group track.swyx: Yeah, yeah, yeah. So, so we have a second, uh, we have a second channel, uh, for latent space tv. Uh, and I, uh, and uh, we basically, this is our Discord meetup. Um, and I I, we have like Claude Coworks, it might be a GI, I don't know if we, we have, uh, uploaded it yet, but one of the sessions was like a, like a Claude cowork thing.Felix: I, you have to see, I would love to see it. Like, I'm so curious, like one of the most fun parts of my job is like constantly see the weird things people use Cowork for because it's obviously like very hard for us to actually design for specific use cases we do. But like every single person who's like most amazed is usually amazed about a thing that I didn't even expect cowork would be good at.Um, we have a new designer and it's one of the first small tasks. I was like, Hey, we need like a new emoji for cowork for our internal stock. It's like a pretty small thing. I like, can you please do it? And he drew an SVG and just gave it to coworker was like, can you animate this emoji? And now it has like this beautiful loopy animation.Um, and I mean, I think obviously this goes down to like, it turns out you can do more things with code than you expected, but it, it's like that kind of stuff that is really fun to me. So, long story short, I would love to see like, the kind of things you're doing.swyx: I'll pull it up. I'll pull it up.Felix: Yeah. Yeah.swyx: Uh, but before we get into it, I, I think always wanna start with like a top level. What is Claude Cowork for people who haven't heard of it? Haven't tried it out.Felix: Okay. Uh, real quick, Claude Cowork is a user friendly version of Claude Code. So the way it basically works is we have Claude Code and for us, fairly impressive agent harness that over December we noticed more and more people are using either, even though they're not technical, they, they're not at home in the terminal or they are at home in the terminal, but they started using Claude Code for non-coding workloads, right?Like managing expenses or like filling out receipts or organizing a knowledge base. Like there was a big obsidian moment that a lot of people liked and we wanted to capitalize on that, but also bring, bring this capability to people who are not terminal native and who might not know how to like brew and store something.So cowork is Claude Code running in original machine with a little bit of padding, a little bit more guardrails, making it a little safer and a little bit more convenient for people who don't wanna first open up the terminal when they go to work.swyx: It's interesting, uh, that is kind of. Pitch that way as a more user friendly thing because I always feel like it, it, to me, I I treat it as like why I'm familiar with Claude Code.Like we, we did a Claude Code episode Yeah. A year ago. But this one is like even more power user tools ‘cause it, uh, it kind of integrates much better with like clotting Chrome and, uh, in all the, all the other tooling. But like, maybe, maybe that's like a perception thing, right? LikeFelix: No, honestly, I don't think you're wrong.This is like a, a thing I've been thinking a lot about for like the last two weeks. So,swyx: but when they say user friendly, it's like, oh, it's the dumb down version. But no, actually this is the superset.Felix: Yeah. Like, I think a similar thing happened, A similar thing happened to me about 10 years ago, like maybe 12 years ago when I was at Microsoft and we started working on, on Electron and like browser-based technologies and cross-platform stuff.And one of the first use cases was Visual Studio Code, which used to be a website. And the initial narrative was, or Visual Studio Code is, is like a more user-friendly version of Visual Studio. But in a similar vein, I think there was some voices saying, oh, this is. For serious developers, like, we're not gonna use this.Right? For like anything. And I think in the end what happened is people have different stories about why Visual Studio Code became such a big thing. But my personal, my personal belief is that the Hackability and the extendability has like played a pretty big role, right? You can hook in Visual Studio Code that like almost any workload, it's so easy to hack on, so easy to put extensions for it.And I think cowork might be hitting a similar thing where it's very easy to extend and it's very easy to bring into your workflows. Uh, so the convenience I think is a bit of a, it's obviously the thing we strive for as developers, but I think the way people find value in it then is by probably mapping it onto whatever they actually have to do in their job.Alessio: So end of last year, you see the spike of like non-technical usage and clock code. What's the design process to say we should make clock code work? Because I mean, you built it in only 10 days. Um, I'm sure there was some discussion before on whether it's easier to use mean. You know, like making, making like a desktop GUI is obviously one way to do it, but like there's a lot of nuance in the product.Like maybe talk people through what was like the trigger of like, we should build a separate thing. We should not build like a different plot code thing. And then maybe some of the more interesting design decisions that maybe you didn't take.Felix: Yeah, I think philanthropic, we've been thinking about ways to move people who are comfortable with using Claude to answer questions and bring more of the power of like this thing to now like, execute tasks for you.I can like solve problems for you can like build things for you. How do we bring that capability to people who are currently mostly comfortable with like a like question answer paradigm within the chat. And we've had a lot of prototypes around that. Just going back as far as like easily a year and a half.Like we had a lot of people working on that. Um, and internally philanthropic is a very prototype demo, first culture. We have a lot of like internal prototypes that don't reach the public. What Cowork actually became is like we sort of picked the right pieces out of the many prototypes that we had.Right. And that's, that's maybe also like, I think an important qualifier whenever people mention this like 10 day number. I do think it's important to me to mention that within Double Scratch there was like a lot of stuff already happening, right? Like, and I think it's important for people to remember that when you build a website, you use React, you use like a bunch of other things.And this is like a similar scenario with like a lot of pieces we already had. Um, and in terms of decision path, I think we live in like an interesting new world where execution is actually quite cheap.swyx: Mm-hmm.Felix: So maybe, maybe what you would do That's so crazy. The year. I know it's wild.swyx: You should be, ideas are cheap.Execution is the hard part. IFelix: know. And like the, we, we used to live in this world maybe where you would take a product manager and the product manager would go to a number of potential customers and in this like very low bandwidth way, would try to. Try to like tease out what are the problems they're having, what are they willing to buy?Um, and then maybe what can you build to like drive out that need and then you go back and you like draft a spec and you think about it and then like you make a design and you execute it. We internally philanthropic app, not pretty much closer to the point where we're like, don't even write a memo, just like build, like let's build all the candidates very quickly.Let's just build all of them and then pick the best ones. I think the, the decision that is most impactful both for the product as well for the users right now is like the way we put value on your local computer. I think that's a big decision point a lot of people have thought about. Should this thing, whatever it is, should it ultimately run into computer or should it run in the cloud?‘cause they're big trade offs, right?Alessio: I guess like if we solve auth, it would be easy to do in the cloud. But I think like the fact that I can just download any file from anywhere and then put it and cowork there, it's like a big unlock. Um, I mean it's interesting you mentioned reusing certain pieces. I think this is something I've been thinking about even with Claude Code, right?The price of like writing code is going to zero, blah, blah, blah. But it actually seems like the value of having some sort of platform substrate is like increasing because as you build these new things, you can kind of plug them together.Felix: Yeah.Alessio: So I almost feel like when people are saying, oh, the value of a lot of software is gonna zero because you can recreate it, to me it's almost like the opposite.It's like having an existing platform to build on top of. It's like even more valuable because you can kind of bolt things on.Felix: Yeah.Alessio: You have obviously mcps, you have skills, you have like obviously the models, which is a big part. All these things kind of come together. Do you feel like that's a valid way to think about it, where people should invest even more in kind of like primitives.To rebuild on or are you like recreating a lot of it each time because like things change and it's easier to rewrite than reuse?Felix: You know, I think, I think you're right. I think you're right that the holistic platform is really useful. And this is maybe a whole like a somewhat contrarian view to a lot of people in ai.I actually don't think that the future is going to be hyper personalized software down to the point where everyone is running their own version. Like, I actually think it's going to be quite hard for all of us to have our own internal chat tool and like, if I wanna talk to you, likeswyx: howFelix: is that gonna work, right?In the, in the context of cowork and how we build it, I think it's a bit of a combination. Like what the, the execution that gets cheap is not necessarily rebuilding all the primitives. I think our priori, there's also not a lot of value in it. So for instance, my team did not think about rebuilding clock code.We're like very much started with the. The core thesis of this should be Claude Code.Mm-hmm.Felix: And then we'll like build things on top of it. The part of the execution that gets a little cheaper is like, how do you take all of these Lego pieces and put them together in a way that makes sense for users?It's like actually valuable. You have so many different approaches now in terms of what kind of, what kind of things do you actually elevate to a primitive, do you strongly believe that all your products should be built by just combining primitive that the public also has available? Do you keep some things internal?Um, and I think that's still evolving, but I think what's probably gonna go away is like, I'm not sure if it's gonna fully go away, but I'm gonna say, I think for me personally, I will probably no longer try to come up with a really good product without testing up with people. This is not a new concept, but wherever you used to have to make costly decisions around, do we pick technology A or technology B, or do we like, um, build it this way, build it the other way.I really strongly believe now you just build all of them and try them out with a small focus group and then whatever, whatever is better is what you go with. Right. And that, that is probably quite different even from how we maybe worked a year ago. Right. Like, I think, I think this happened very recently.Alessio: Yeah. I started building something in on Electron since you're here. Coincidence. Uh, but then Electron and like SQL Light are like, there's like some issues that like between development and like, uh, building anyway. And I was like, let's just rebuild the whole thing in Swift and just recreated the whole thing in Swift.And it's like, I. It's done.swyx: You know, I didn't take any effort. I, I, I don't even know Swift.Alessio: Yeah, exactly. I was like, I'm the, I'm not reviewing it anyway, whatever. You can write in whatever language you pick, but the important stuff that I did was not write the electron bindings. Yeah. It was like the logic of what happens in the app, you know, and then the model is like, yeah, I can just recreate the same thing as withswyx: Yeah.I, I think you still want, especially for people who are doing like high performance software or like very complex software, uh, you still want like, some view of the architecture. Uh, but you can use markdown for that,Felix: right? Yeah.swyx: Uh, you don't actually have to read the code again. I, I'm still like on a sort of like a definitional thing.Um, can we build a good mental model of Claude Cowork? Um, this is what I have, right? Like you you said it's like fundamentally cloud co. We don't wanna touch it. There's the cloud app, there's clouding Chrome. I think you guys do something different in planning, but, uh, I've been talking with Tariq who is on the cloud co team, and you guys are, he's like, no, we just exposed planning.Maybe we can clarify like, what are the major pieces. That people should be aware. It goes into cowork, like,Felix: okay, I think you basically have them. So really, um, you can, you can take planning more or less out. I think there's a few things that are really valuable in cowork. Um, the virtual machine is probably the most powerful thing.So we currently run like a, we currently run like a lightweight VM and we put clocked out into the vm and we do that for, for, um, a number of reasons. Safety and security is a big one, but even if you, even if you ignore for a second safety and security and you're just like, okay, Yolo, I want this thing to do whatever.It is quite powerful to give Claus on computer that is like generally a good idea. And in terms of architecture and UX and everything else that we've been working on, philanthropic, it often is quite useful for you to like anthropomorphize, um, clot aggressively and just be like, this is a person. What will you do if you give a, if you had a person, right?Yeah. And the analogy I've given my dad this morning who is still like quite insistent on using chat even for like coding things, is if you were a developer and your employer told you that you don't need a computer, they're just gonna like, send you emails with a code and you send emails with code back like that, maybe work for Patrick Miles in the back, but that it's not very effective.Um, so what we can do with the VM is because it's a, it's a Linux system, Claude Code has more or less free reign to install whatever needs to install. It can install Python, it can install no js. We do have strict network ingress and egress controls. So you can still, as, as a user in like plain human language, make it clear to, to the entire system what you're okay with and what you're not okay with.But at no point do we have to ask a real person, like a, like a person who might be in marketing or a lawyer. I'd have to go to a lawyer and be like, are you okay with me installing Homebrew?Alessio: Yeah, yeah.Felix: Right. Because the implications of the question and the answer are complex and nuanced and like, not, not easy to reason about.This gives us a lot of distraction that makes Cloud very powerful. Now then around it, we, we do probably have a number of things that also keeps growing almost every single week that you're probably noticing that make cowork maybe better for certain tasks than just cloud. Cloud on its own. Yeah. But most of those actually live in the system prompt.They're about like, what can we infer about the work that you do? What can we, what can we intru in the system prompt to make that more effective? It's of course the like very tight integration with Cloud and Chrome. You're noticing that a lot of people, especially as the models get better, a lot of people throw up their hands when it comes to MCP connectors in this area.I'm not gonna, I'm not gonna go through like 25 M CCP connectors, click off everywhere and then like half of them don't let me do the things anyway. So Cloud and Chrome is quite powerful because we can just talk to the cloud and Chrome sub agent and that will just do things for you.swyx: Yeah, so, so one example right in MCPI, honestly, I think that the state of MCP is kind of, kind of.Really hard to integrate. Um, I need to, I needed to add, uh, Figma MCP to the coding agent that I use.Felix: Yeah.swyx: Uh, and, but I didn't wanna read the docs, so I just had caught to it. And it's, it's great at reading docs and the same, same way I had to set up like a Google Cloud, um, account for some project I was working on and get some API keys somewhere.And Google Cloud is famously super hard to navigate, so I just didn't wanna deal with any of it. I just used Claude CoworkFelix: within the first week of developing on Core. This happened very, very quickly. Um, I caught myself by starting to use cowork for coding tasks, which is not ostensibly what we built it for, right?We don't need to. But I found myself, um, I found myself like on our internal, internal tool that we have for, to collect crashes and just like debugging information and I found myself sort like picking out the ones that I think we can easily fix versus the ones that might be like kernel corruption or something else on the operating system.And I found myself sort of picking these out and then just telling Clark, go fix this bug. I was like, what am I doing here? Go one level up, tell a cowork, I want you to go to all these crash tools. I want you to find all the bugs that you think are fixable and not like an operating system crash. And then I want you to tell another cloud to like fix all of that.Um, and that's, that's, that's sort of another cloud,swyx: just so it can spin up another instance or,Felix: uh, it, currently what I do is, um, and this is a bit of a hack, but I tell it to use clockwork remote to which website itself? Yeah, that's interesting. So you basically take, if you, if you imagine like a dashboard with like 20 bucks, you, this is remote control or clock or remote, or, sorry, I just wanted to confirm what, the way I'm using it is.I have cowork running and I'm telling cowork, here's where I normally go every morning to find the latest bugs. Go read the entire bug list, separate out which ones are fixable, which ones are, are fixable, and then for the fixable ones, four is this almost loop. For each bug, write a markdown file with a prompt.And then for each markdown v, that is a prompt. Start of a cloud set. So natively Claude Code hasswyx: this concept of subagents. Mm-hmm. And this is basically a subagent, but you're not using the subagent functionality.Felix: I'm not using the subagent functionality. And the reason I'm not is because I'm firing that off as a Claude Code remoteswyx: task.Felix: Yes. That's kind of nice. ‘cause then I can just fire it off. I can go to my next meeting and in Claude Code remote. Now the work is happening.swyx: Mm-hmm. Yeah. You, you see like you're already starting to use the cloud over your local machine. And I think this is one of those things where like. Shouldn't just everything just be cloud first, right?Felix: Ah, this is such a good group. I'm like solely bad about this. I have so many thoughts about that. Okay. So I generally believe that Silicon Valley overall is undervaluing the local computer. And my default argument for that is always how come we're all using MacBooks and not like an iPad or a Chromebook?Um, that there is like still value in, in having a local machine. And now when I think about Clot, it's this entity that is supposed to be very useful to you, like it tremendously useful to you. I think that entity needs to have access to all the same tools you have access to. Otherwise it's gonna be hamstrung in like all these complex ways.And there's, there's sort of two approaches we could take. We could say, okay, we're gonna like one by one chip away at everything that is at your computer and move it into the cloud. That's, that's one way to do it. Um, and I think other products have taken that path. I personally, this is a very personal opinion, but I personally, for the amount of tools that I use.Just don't have the patience to give another tool like permissions to every single thing and keep those permissions up to date. The second thing that I'm still grappling with, and I don't have a good answer for anyone just yet, but the second thing I'm still grappling with is what does it look like for someone to slurp up your entire work and put that in the cloud?Like if I, just as an example, like if you could click a button and it just clone your entire computer into the cloud, is that something that you would want? I'm not totally convinced yet that all everyone will. Mm-hmm. And that is sort of like upstream of all the technical issues we're gonna have. ‘cause like in general, I think the world is not ready for this kind of stuff.Like, I'll give you one quick example that would probably be very easy for us. So as a desktop app, we in theory with your permission, can do a lot of things on your computer, including reading your Chrome cookies. If we really want to do right, we could take your Chrome cookies, you would have to decrypt them for us.We could put those on the cloud if we really felt like it. Pretty easy solution. That would be super cool. We could just be like, oh, we can do all your tasks in the cloud now. Um, a lot of websites, thanks, include it. If, if they see the same authentication from like two different locations, we'll just lock down your account and now you have to go to the branch and be like, okay, I, I'm here with my passport.You actually know that. Wow. Yeah. As tired as well are of the term agent for the age agent future, I think there's a lot of stuff that sort of slowly needs to catch up and until that's the case, the way I, as someone's working on clock and make Cloud most effective is to like put it where you are working.swyx: Anything else? I thought with our mental model, so like, basically like, uh, part of me also just want, like the more I understand how it works, the more I can use it to its full potential. Right?Felix: Yeah.swyx: And so what I'm get hearing from you is you told me to delete the planning thing. You're not doing anything special on, on the, that's only exclusive to Qua cowork.Felix: We have some tricks for this sort of like change week over week. We eval cowork maybe against different use cases than he would evil clock code, right? If you think about it this way. Okay, so like clock code is our eval clock cowork. Yeah. So clock code is like quite optimized for coding tasks and we mostly value it whether or not we're getting better or worse depending on how good it is at like a typical suite job.And Clark Cowork on the other hand, we evaluate more against typical knowledge work, the kind of stuff he would find in finance or in like maybe a, like in like a legal office. Um, my personal use case is always like managing my things, like managing my personal mortgage or something like that, right? Or like wealth planning for me and my family.Those are the kinds of use cases we eval, clock cowork on. And what you might be picking up on is like the subtle changes we make to the system. Prompt what we put in the system, prompt how we steer, clot with the tools we give it. Um, like either it'd be better in one or the other direction and whether there's a trade off, try us exist a lot.CLO code will be better of a code and Claude Cowork will be better. For non-coding tasks, will those gaps still exist in the next three generations of models? It's like a little unclear to me though.swyx: Yeah,Felix: because right now these like hyper optimizations we make, I'm not sure for how long they're still be relevant.swyx: I think what I was referring to was also, it, it just, uh, it qualitatively felt different when I probably, it's just all prompting and I'm reading too much into it, but like the, the fact that it comes out with like a nine step plan, I can edit the plan and give feedback and, and, and see it execute the plan.Yeah. It felt more long range than in Claude Code, but maybe that already existed in Claude Code and you just build a nicer UI for it.Felix: It's kind of both. Um, like if the Clark Code people who build the planning functionalities would city, they probably say yes, we have all of those things in Clark code and they do.Um, I think people tend to give cowork. Tasks that are maybe of longer time horizon, I thought isswyx: so long. Yeah.Felix: That's like one thing, right? It's just like that the, the chunk of work tends to be maybe a little bigger. And then the second thing is that because the work, when it gets longer, it gets a little bit more ambiguous.We do tell co-work to make heavy use of the planning tool or to make heavy use of the ask user question tool, right? We do want it to come up with like. Different scenarios of, okay, tease out what the user actually wants. Don't go off to work for like four hours and then come back with the wrong thing.And you're probably picking up on that.swyx: Yeah.Felix: Um, I wish I could tell you I like built this magical thing and it's like, there's some secret sauce,swyx: but No, no, no. I mean, it's, it's just clarity is good that, you know, engineers just want to know. Yeah. They can, they can plan around it. And then I think also for me, um, I am realizing I have to switch to my, my other machine because this is a new machine that doesn't have my session.But, uh, yeah, the, the, the planning is really important for, for me to like approve or like to see whether it's like, it's right. The ask is, the question is so beautifully presented. I mean, it also, it also available in like cursor and, and in Claude Code. But like, I, I think like it's so nice to see that it, like it's kind of for me like to understand that it gets me, it gets what I want to do.Felix: Yeah.swyx: Yeah.Felix: It probably very hardswyx: just on the topical evals. Mm-hmm. When you say eval, I think people are very vague about what it means. Is it just like vibe testing or do you have like automated programmatic evals of Claude Cowork?Felix: When we say eval, uh, what we really mean is that we essentially take the entire transcript, including all the tools that clot has available ultimately to it, and we then measure what are the outputs, depending on what we tweak, right?So we do run that a lot. We use that in training. Um, we use that in, in like, if you sort of separate out post training from like the scaffolding around it. Cowork sort of exists in the scaffolding space, but obviously we also train on it a little bit. Um, so when we say eval, we mean given the certain transcript, what do the outputs look like?Including the file outputs as well as like the actual token outputs, like the ones that you see in the chat window.Alessio: I'm curious, um, how much of the failure modes are the model intelligence versus like the usage of the end tool to put the intelligence in? Like the well planning is like a good example, right?It's like one thing is to come up with a plan. The other thing is like make a nice spreadsheet. Yeah. That kind of runs you through the plan. Like how have you seen that? Well,Felix: the thing that I grapple with a lot is that whatever scaffolding you come up with, I think we still have a bit of sort of like model overhang where the model is dramatically more capable than right.Users end up using it for. And I think part of that is that we're just not getting the model all the tools to do all the things that's theory capable of, right? There's like one thing, um, however, whenever you do build the scaffolding, I'm sort of wondering at what point, at what point will that scaffolding go away and like how much you invest in figuring out what the right scaffolding is.It's kind of up to, it's a little bit of a bet. And one thing that I as an NJ quite enjoy is that like working in philanthropic and working at a frontier lab, I maybe have a little bit more insight into what's coming, coming down the chute in terms of like, what's the next model, what is the model capable of?What is good at, what is it bad at? And I'm, I'm increasingly wondering, is the right thing for us to like really invest too much in sort of these like scaffolding corrections where the model might otherwise not misbehave, but just not do the thing that you want?Alessio: Yeah.Felix: Or is it to just like give it as many capabilities as possible, try to make those safe so there's the worst case scenarios, likeno status might be otherwise.And then just simply wait a second for the next model drop. I'm personally, currently more leaning into the ladder. I think we're gonna see a lot of like applications and companies that do very impressive things with ai that in the short term might seem very effective ‘cause they're very specialized to individual use cases.But I think once models get better generalization and get better at like those specific use cases without being super guided on those, I'm not sure how long that's gonna stick around. And you can kind of, kind of already see this in like skills and NCP servers, right? Mm-hmm. We've, we've already seen sort of this like slow shift from MCP service to skills.And like, maybe a good example is Barry who made skills. He was initially hacking on something that honestly looked a lot, looked, looked a lot like what Cowork does today. It was sort of thinking about what if cowork, but for like people who don't wanna build code. Mm-hmm. And, um, he too did that as a prototype inside the desktop app.One of the first use cases we thought of were, okay, what, what are like coding like use cases that could really benefit from graphical interfaces and like from being a little separated from the actual underlying code. And everyone comes with the same answers. Data analysis,Alessio: right?Felix: Yeah. Or saying how many users do we have today?How many, like, it's always data analysis. And I think the thing that ultimately led to skills is that we wanted to connect this little prototype to our data warehouse and. The team very quickly discovered that like instead of building a custom tool for the thing to talk our data warehouse, they just like meet and embarked on follow like mm-hmm.Dear Claude, if you want to get data, here's the end point. Here's what the API looks like. You'll figure it out.swyx: Ah.Felix: And then it be hand over control. Yeah, yeah. Also just like maybe go one step up in the layer of abstractions, right. Just, yeah. Instead of, instead of telling the thing, here's ACL I, please call the CLI, or here's an MCP.Please call this ECT shape. Just like this is the end point. If you wanna know something, if you post here, maybe you can do post sql. It's gonna be okay. And that ended up being so effective that they started trying the same pattern of like just giving the model a markdown file that describes whatever it needs to do.That the whole thing eventually became skills and we're like. We should package this up. This is a good idea.swyx: Yeah. Um, we've had Barry Mahesh, uh, on, on our conference and uh, he's uh, definitely got a good idea there.Felix: Yeah.swyx: I wanted to show you the, how I've been using Claude Cowork.Felix: Uh, this is was my favorite part.swyx: This is this. So this is like me, uh, this is how we run the Discord. Uh, we literally, uh, at first I didn't trust Cloud Core. This was my very first usage.Felix: Okay.swyx: Right. So then I was like, okay, I will just try to manually download from Zoom all my recordings and upload it to YouTube. Yeah. Because this is a very laborious process.I got a click, click, click YouTube, um, isn't super user friendly. Uh, and it just did it. And then I was like, actually, you know, even the download from Zoom part, I should also. Put into Claude Cowork, and then I did it right. Here's a bunch of, and it starts compacting here, and it, and it, it starts to even be able to do things like look through the individual frames of the video to name the video so I can upload it auto automatically.Oh, that is, and this replaces my job as a YouTuber. We will forever appreciate your creative Yes. You know, and so that's great. Uh, but then by the way, it compacts and makes, makes like a new thing, right? So I, I don't, I don't have the initial, initial thing, but then I asked it to make its own skills so that it, so that something that's repetitive and one-off and human guided becomes more automated and I can use the skills independently and reuse them.Uh, and it obviously you can write skills and that goes into context and skills at the bottom here, which is, which is so nice. Um, so I have all these skills that, that I now sort of do on a weekly basis. Uh, I know you've released scheduled Coworks, which I haven't done yet, butFelix: course I should try them. I, I think this is like so wonderful and fun for me to see because.One thing that is very fun for me about skills in particular is that they're so easy to make. Like anyone can make a skill, like a text message, could be a skill, and they can be so hyper personalized to you. And this is like sort of the subtraction layer, right? Like, um, I, I'm just guessing, but I assume, heck, you are very good at your job.You're probably given this thing some guidance about how to do it, right? I,swyx: I just said, wrap everything up into, into a skill, right?Felix: Yeah.swyx: And then, uh, and then I was like, actually, sometimes I might need to break, uh, things apart because some parts fail or some parts might be needed in individually. So I told it to split one skill into three skills.So it's like a skill splitting thing, and then there's like a parent skill that just orchestrates all of them if I want to use that. You know, like, um, I think that's, that's like really good. Uh, and, and, uh, there's, there's one more part, which is the, uh, Google Chrome thing that I told you about.Felix: Yeah.swyx: Where I'm like, okay, you know, what's better than uploading, using Claude Coworks to YouTube?Like actually. Looking at the docs to like programmatically upload to YouTube and then putting that in a skill. And I've never done that before. I don't want to deal with Google Cloud. Yeah. So Claude Cowork does it for me.Felix: That is really cool.swyx: So, so I, I just, I don't care. I just, like, I do a thing. I don't, it doesn't really matter.Felix: That is really cool. And then you've, I assume paired the skill just with the script that it's built.swyx: Yeah, no, I just update, update the skills.Felix: Oh, that is beautiful. Yeah. That's wonderful.swyx: It's kind of like a skill, like, uh, uh, basically I think like the way that people ease into Claude Cowork is like take a knowledge work task that you would normally be clicking around for and then, uh, try to turn, turn that, and then you do the, okay, well what if you went further?Okay. And then when, if you went further, when, if you, and it sort of expand the scope of cowork as you gain trust with it and, and also teach it how to replace you.Felix: Yeah. It's like a little bit like playing factorial, but for your own life. Uh, like you say, you start really small.swyx: Yeah.Felix: You start automating something really tiny and like.Once it clicks, you keep adding onto this like automation empire. Just like make your life easier and easier. My favorite skill has been, um, every single morning Kohlberg starts looking at my calendar and make sure that there's conflicts because people tend to schedule a lot of meetings, sometimes last minute, sometimes miss it soft and painful.And a lot of products have existed like that A lot. I've written in the custom prompt there. I haven't made it a skill, um, honestly should.swyx: Yeah.Felix: But I've given it like pretty clear instructions about okay, here are some people, if they book over other meetings, I'm probably gonna go to their meeting. Like if Dario schedules a meeting.swyx: Right.Felix: Not try to reschedule down. Right. Um, and I think there's some other rules in there about like what kind of meetings I care more about what kind of meetings I care less about. What is okay to like, maybe pun like when I want to be, when I want to be working, when I don't want to be working. And it's those really small things that I can think kind of click with people.Right. When we launch co-work, I think one of the US races that went most viral on Twitter. X was clean up your desktop, which is stuff, because silly, that's such a smart thing, right? Like you don't need to model to clean up your desktop. Not really. Um,swyx: like this, like clean up my desktop.Felix: Yeah, exactly. Yeah.swyx: I need to, I need to choose my desktop, right? I guess give it access to my desktop.Felix: Yeah.swyx: Okay. Uh, okay. This is very scary. Oh, we'll do it.Alessio: I did, I did it with my downloads folder. It was like, you have so many term sheets and there's like eight copies of your rental lease for your office. I was like, all right.Like, don't yell at me.Felix: It's like, it's not such a small task. And then like, I, I would never go out there and normally otherwise and tell people I've pulled a product. It can organize your folder. Right. Um, because it feels small. But I think to your point like,swyx: oh, here's, here's the, here's the ask user questions.Felix: Yeah.swyx: Uh,Felix: beautiful. Right. Elite obvious junk. You probably shouldn't click that.Alessio: No.Felix: If he's not done right.swyx: As long as it's reversible, I don'tAlessio: make up blend to,swyx: yeah. Uh, yeah. No, I, I have a, I have a typical, everything is super messy folder. So, yes. I think this, this is super helpful. So this is a pretty simple task.Mm-hmm. But I've, okay, here it is. Right. Here's the progress. I don't see this in, that's why I'm like, this gotta be something different than, uh, than Claude Code, because I'm like, weFelix: do. Yeah. That's, we do system prompt that. We're like, all right. We want you to think about like, this task Yeah. Methodology.Yeah.swyx: And then I can, I can, I can do like little suggestions for, for, for these things. It's beautiful. Look at this. I, I can, I can like say like, oh, don't do that. Don't do this. It's amazing.Felix: I'm so happy. You like it. Um, I mean, the other way around, like we're part of the Clark core team, if you would like this in Clark COVID.swyx: Yeah. Yeah. Yeah. Uh, so, so yeah, I mean, uh, this is really good. Obviously I, I'm like kind of raving about it. Uh, you know, I have other things like sign up for pg e so if you can do phone calls for me, that'd be great. Um, I, I do, peopleFelix: have done that. Obviously you can't do that natively, but people have done that with like, various other providers.swyx: Yeah. Uh, and then this is like signing up for the Figma MCP. Um, I, I really am trying to do like everything, um, data analysis as well. I do think, um, oh, design to code, uh, very, very good. Right? So like, here's a Figma file, take it. And then this is where like a lot of other tasks is like knowledge work, like replace my manual clicking, but this is no, I would normally use Claude Code or uh, Claude Code for this, but because I perceive that you have better Chrome integrationFelix: mm-hmm.swyx: I, I think you can actually do a better job of this. And I, this, this is one shot at my, uh, conference website.Felix: That's pretty cool. Like at some point I would love to like, hear how you feel about code. In the desktop apps, which is like I never use, which is the, the same team. Same team.swyx: So I use the call code in terminal, which I, I perceive to be the default way of cloud coding.Felix: So one thing this has,swyx: sorry, I'm just like, I'm notFelix: here, I'm not here. All products. Can I talk about other stuff? Like I, I'm not sure if people out there wanna like hear me advertise my stuff for like an hour. Please do that. Um, this thing is like a builtin browser, which is a thing a lot of products have said.Yeah, it's a builtin browser. And I think giving cloud eyes into like what you're actually working on makes it so much more effective. And that's probably what you've seen in cohort because it can see Chrome, it can like debug the dom, it can like see things. Um, that does make it more powerful.swyx: Yeah. So, so I think, uh, my mental model was kind broken.‘cause I only use this cowork because I thought it had a, a browser thing in it. But I understand that the Claude Code app. The app version of Claude Code does have a built-in browser. I've seen, I've seen this preview thing.Felix: Yeah.swyx: I just, I've never used it.Felix: But in the end, in the end, you sort of have it by hard.Yeah. You basically get the same thing. Right? Like the, the, the additional skill that you're describing is chart is better if we can see what it's working on. Right. That's, that's sort of like the summary here and like whether it's using your Chromeswyx: Yeah.Felix: Or it's just like making up its own little like browser.It doesn't really make a big difference because either way it's gonna see what it's working on and that just makes it much better. And then you don't have to run QA for your cloud.swyx: Why doesn't it pick up my existing Claude Code sessions? ‘cause I, I mean, obviously I've used Claude Code, but Excellent question.Um, don't have a good answer other than like, we're honest. Just haven't Yeah. This is what the Open AI team does. Okay. Uh, cool. I I I don't have other, like, I, I just, I, I do wanna expand people's minds and also maybe show people if they haven't really done it, but like, I, I think it's very interesting how I sometimes use this more than I use, I mean, I use dia, right?Yeah. Um, I, and I use, uh, I've used like all the other agentic browsers and philanthropic didn't have to build an agentic browser because you just had Claude Cowork and that's enough.Felix: Yeah. I also think like maybe integrating with number of excellent browsers out there, it's like currently on my personal priority list, a little higher than like trying to rebuild a browser from scratch.Yeah. You know, never say never, but I think going back to this idea of like, we wanna plug this into an entire existing workflow, I think our goal is actually to not replace any of the applications we have in your computer. But instead of like, work really well within a new workflow,Alessio: make the new one. Yeah.Are, it seems that nowadays, especially on the browser, most of the innovation is like user ergonomics. It's not really like the underlying browser engine. So I feel like to call it, it doesn't really matter if it's like the, uh, or Chrome or Alice, whatever.Felix: Yeah. We wanna, we wanna meet you wherever you are.Which is like, like obviously I would say that, but it's also just generally true because I don't wanna shrink my potential user base artificially by saying, okay, like, I'm gonna start building for the people who are willing to switch browsers.Alessio: Right.Felix: That's such a, like, you know, like many lawsuits have been filed over who gets to review the browser and like a lot of money has switched hands over the question of like, which browser is default and which search engine is default within the browser.Um, I just wanna build for, yeah, I wanna build for swyx essentially. Like, I wanna, I wanna, I wanna build for people who have a number of annoying tasks that they feel like. Maybe clock could do it. Could do it for them.Alessio: Yeah. What do you think about skills portability? I think there's been one thing, I use another thing called zo, which is kinda like a cloud computer plus agent.And I have a skill to add visitors to the office. Yeah. So whenever somebody has to come in after hours, they need to check in downstairs. Um, but I wanna like text the thing, so it doesn't really work in, in cowork, but now that skill is in the zone harness and it's not in my cowork thing. And then if I make a change, it's gotta, I gotta sync them.How do you see that going? Like I see memory as like. Cloud personal, kinda like, I don't necessarily want my memories to be cross thing.Felix: Yeah.Alessio: But I do want my skills to be cross agent that I use. I think with MTPs, people do the same thing. It's like, oh, Mt. P Gateway. Mt P registry. I don't really know if that's like a business.So I'm curious like if you've had any thoughts in the area.Felix: I think for me, this is sort of where I go back to the really basic primitives for our skills are file-based instead of like this complicated thing that exists inside a place somewhere that is like super proprietary. I'm really leaning into the idea of like, it's all just files and vultures, and that makes it very portable on its own.Right. We do have skills as part of this container format, which was just called plugins.Alessio: Mm-hmm.Felix: And plugins are available both for Claude Code and Claude Code work the same format, and you can install plugins. This works in cowork today. You can basically say, I'm gonna add a whole, like just a GitHub repo as a.Skills marketplace or like a plugin marketplace. And that's how we're doing portability. I think we have a lot of room left to grow in. How do we make it easy for people to know that they can write skills? How do we make it easy for them to just like, share a skill with you? Because obviously all the words I just said, right?Like I'm losing most of the knowledge worker base out there, right. And start by saying, oh, you can connect to GitHub repo. It's not exactly how most people will end up working in like a general knowledge worker space. Um, but I think there's something there. And another thing that's there that I think has not really been properly explored is the, the, the combination of which part of the skill is very portable and then which part of the skill is like very personal to you.Right. And I think that's something we haven't really solved as an industry. Hmm.swyx: It's like, which, how you wanna introduce more structure to the skill or have always have like. Public skill, private skill, you know, pair. Yeah, yeah. Kind of. I think there'sFelix: like a, like the easiest way to do this, which is we do like use string interpolation or something.Right, right. Yeah, yeah. Insert username here, insert like phone number, insert, like known folder, locations, that kind of stuff. Um, that's probably clunky. That's why we haven't built it. Um, but I do think someone is going to come up with like an interesting way to keep everything we like about skills. The portability is just a file, it's just marked down.It's just text, honestly. Right. Like a text file words. The complete lack of structure, which means you don't need any kind of tutorial to write a skill. Just like explain it to Claude the way he would explain it to me and Claude will probably get it before I work. Mm-hmm. Right? You're just like, for booking a flight, tell Claude how to book a flight the same way we tell him somewhere.I just started working here today. But combine that with a very like, personal thing. Um, maybe we'll stick with a booking a flight example. I don't actually think. AI should be booking flights. I think the tools we have is yes.swyx: Yeah. Finally, somebody says it. It's the default demo that everyone's making.Felix: I'mswyx: like, I even against like booking demos, it is not a good showcase.Felix: Yeah. I'm like, I just wanna book my flight myself. But, um, I think there's a lot of things that have a personal and a non-personal component and that's maybe why people reach for flight booking because some things are very universal. Yeah. Super flight is usually better, right? Like few people try to book the most expensive flight.And then some things are quite personal about like what times you prefer, which seat you prefer, which airports you prefer. Combining that and like a skill format that is actually portable, compatible, easy to understand for people. I think that would be very exciting. We just haven't figured it out yet.Alessio: Yeah, I think the text part every, I think everybody by now has some sort of like cloud file thing. Either Dropbox, Google Drive, whatever. So it feels like in a way it should basically like sim link. My skills into all my agent harnesses. Yeah. Just keep those ing like we have internally this like valuable tokens repo, which is like all the commands sub agents.It's good. Uh, and then I build like a TUI where you can start it and be like, you know, install this command and this three sub agents into this agent in this folder and just copy paste this. It doesn't do anything. It literally cp the file into that. But I feel like there should be something similar where like whenever I go into a new thing, it's like, hey, here's like the link to exactly the cloud folder and just bring down these skills into this.Yeah. Like today it doesn't quite work like that. Like if I install a new agent, I cannot, I have to like copy paste all the skills and I don't even know where they are.Felix: Yeah.Alessio: That's like the big problem. It's like where do I find them?Felix: Yeah.Alessio: Um, so I'm curious like in the future like that, that almost feels like my personal productivity thing will be my skills.Felix: Yeah.Alessio: Is not really the product that I use. Everybody has access to the same product. But today there's, that just looks like copy pasting ME files, IFelix: think so many things I, I really like thinking about agents and LLMs just as like another coworker. So many attempts have made to build documentation companies that are like, oh, we're gonna solve oil documentation problems.Um, I myself, like spend a little bit of time working in notion, right? I'm like deeply familiar with the concept of let's get everyone on the same page. Mm-hmm. Right? And what you're basically saying here is you want all your agents to be on the same page about your preferences, about the skills, about the way they ought to work and like how they ought to execute.And I'm not sure what the right thing is going to be if it's going to be some, some company that can say, all right, we're as an independent body, we're not trying to like, push into any particular product. It's our job to be like the skill authority, and we provide, I don't know, we're gonna be the Dropbox of skills and we can just sim link us into all the products we want to use.I'm not sure that's gonna be viable business, but as, as an idea, it would be cool.Alessio: Yeah. Yeah. I think so many things are just going away as businesses. It's like, how am I supposed to do it? I'm not even asking somebody to make a product about it. Like yeah. I wanna personally know. And there's things like you said, it's like you almost wanna skill and then interpolate it between personal and work.So if I'm booking a fly for work, it's different than I'm booking a flight personally.Felix: Yeah.Alessio: In some ways, yeah. But like a lot of the scaffolding is the same, you know? Cool.Felix: I mean, as an engineer I will tell you like, you know, technic a person to technic a person. I will just be like siblings.Alessio: Well that's what, that's what I do.We call that MD and agents that MD's just the same how sim length. And so it is like, that works, but it feels like, yeah, I don't know. MaybeFelix: you can always go one, you can always tell cowork problem and then cowork will solve it for you. Just make the siblings. That's like one way to do it.Alessio: That's true.That's true. All right. Everything is called cowork.Felix: Uh, potentially spicy. Question for both of you.swyx: Uh, which of these industries will go away?Alessio: Okay, so what Felix was saying before is interesting. There's busy like. The short term pressure of like, we need to turn these tokens into valuable things, which is I should build the last mile product that harness the model.And then there's the question of like, long term, which ones are gonna still be valuable? And I think you're kind of seeing this today with like, uh, you know, the coding space in a way is kind of like everybody's moving up and up in stack because you need more than just turning tokens into code. I think search, like enterprise search is kind of saying the same thing.Like with G Clean and like all these different companies is like, at the end of the day, if Cowork is the one doing all the work, the search itself is like such a small part that like, I don't know if I'm really gonna pay that much money just to do search. It's almost like everything is like a cowork vertical.So like how much can cowork first party support?swyx: Mm-hmm.Alessio: And how much can it not? I think for a lot of these things, the planning thing that you were showing do Which one? The planning. The planning.swyx: Okay. Yeah. Yeah.Alessio: That's one thing where like most of the value that these agents provide is like they're better at planning for specific tasks.Yeah. And have better tools for it.swyx: Yeah.Alessio: But I think the models are now moving in that direction and they have the right harnesses and they're on your computer. So for me it's almost like if for the end customer trusts your startup to be the provider of that task result, then I think that works. This is, uh, something that, this is a shortswyx: spike that we're, we're working on.Uh, yeah.Felix: I think, look, I'll, I'll, I'll tell you this, like I don't think I'm the best person to like actually estimate which industry is going to be hit the hardest. But I do think that at philanthropic as a group of people, we're deeply worried about the impact. That the tools are going to have on the labor market, especially for like junior employees that, because I think, I think it's only honest to say that when we talk about automating a lot away, a lot of the work that we personally find annoying that we maybe think's not the best use of our time.In a lot of industries, that kind of work would've been given to a junior entry level employee. Yeah. Right. And I think it's, it's only, it's only right to be really worried about that and like worry what that's going to do in particular to people like enter the shop market.Alessio: Mm-hmm. I have a solution for that.Which you make them, you create simulative jobs for them.Felix: Okay.Alessio: So this is, this is like half joke, half true. So if you think about software engineering, when you're like a junior engineer, you work like 1, 2, 3 years. And in those three years there's like maybe like a handful of moments where like you really learn something.And then a bunch of other days where like you're not really progressing.Felix: Yeah.Alessio: I think now we can use AI and these models to actually like shortcut these careers and almost like simulate the early years of your work and like just make them like super dense and like these learnings, it's like, hey, we're working on this feature, which is like a distributed system and you need to learn this thing that might take three months at a company.And so you take three months here, it's like we're just simulating the whole thing. It's actually not a real thing. And in one week we kind of speed run through the whole thing and you kind of learn your lesson from there. And we kind of repeat that in like one year. You basically get like three years worth of like projects and experience.Yeah. I think it's harder for like things like sales or for things like, you know, marketing because you don't really have a way to get the feedback loop. But I think a lot of it, it sounds kind of silly, it's like you're making the new effect job, but it's almost like you go to college, right? People pay to learn how to do it, and this might feel similar where it's like, hey, we have the.Jane Street Simulator is like, you wanna come work at Jane Street? We'll just put you in the simulator for like three months.Felix: Wow.Alessio: And you'll come out of it. It's like, you know, I'm ready.Felix: So there, there is an aspect here. I'm not an expert enough to like actually know what, what is going to happen to marketing or legal or finance, right?Like, I don't work in those jobs and I, I don't think I should talk about them, but I am an engineer and I think I have a pretty good idea of what engineering is like. And I think one thing we're sort of seeing is that as a company and also as, as the public, we're like deeply worried about entry level, but we're also seeing more senior engineers accelerate it.If like they're more productive. They, they actually increase the value they provide. And the thing that I'm thinking about a lot is the fact that even before all of this happened, um, I've always had a lot of respect for the University of Waterloo and the, the new grads that have joined my teams as from coming from the University of Waterloo always felt like.More ready than new grads will like literally spend their entire time at the university regardless of how good, but never actually had to work inside an environment where you have to ship things that eventually will be used by users. And I'm, I'm, I'm German. I like initially went to German University and I think the, the, the like information systems programs, there tend to be very theoretical, right?Like I often give people the example of like trying

Turbopuffer came out of a reading app.In 2022, Simon was helping his friends at Readwise scale their infra for a highly requested feature: article recommendations and semantic search. Readwise was paying ~$5k/month for their relational database and vector search would cost ~$20k/month making the feature too expensive to ship. In 2023 after mulling over the problem from Readwise, Simon decided he wanted to “build a search engine” which became Turbopuffer.We discuss:• Simon's path: Denmark → Shopify infra for nearly a decade → “angel engineering” across startups like Readwise, Replicate, and Causal → turbopuffer almost accidentally becoming a company • The Readwise origin story: building an early recommendation engine right after the ChatGPT moment, seeing it work, then realizing it would cost ~$30k/month for a company spending ~$5k/month total on infra and getting obsessed with fixing that cost structure • Why turbopuffer is “a search engine for unstructured data”: Simon's belief that models can learn to reason, but can't compress the world's knowledge into a few terabytes of weights, so they need to connect to systems that hold truth in full fidelity • The three ingredients for building a great database company: a new workload, a new storage architecture, and the ability to eventually support every query plan customers will want on their data • The architecture bet behind turbopuffer: going all in on object storage and NVMe, avoiding a traditional consensus layer, and building around the cloud primitives that only became possible in the last few years • Why Simon hated operating Elasticsearch at Shopify: years of painful on-call experience shaped his obsession with simplicity, performance, and eliminating state spread across multiple systems • The Cursor story: launching turbopuffer as a scrappy side project, getting an email from Cursor the next day, flying out after a 4am call, and helping cut Cursor's costs by 95% while fixing their per-user economics • The Notion story: buying dark fiber, tuning TCP windows, and eating cross-cloud costs because Simon refused to compromise on architecture just to close a deal faster • Why AI changes the build-vs-buy equation: it's less about whether a company can build search infra internally, and more about whether they have time especially if an external team can feel like an extension of their own • Why RAG isn't dead: coding companies still rely heavily on search, and Simon sees hybrid retrieval semantic, text, regex, SQL-style patterns becoming more important, not less • How agentic workloads are changing search: the old pattern was one retrieval call up front; the new pattern is one agent firing many parallel queries at once, turning search into a highly concurrent tool call • Why turbopuffer is reducing query pricing: agentic systems are dramatically increasing query volume, and Simon expects retrieval infra to adapt to huge bursts of concurrent search rather than a small number of carefully chosen calls • The philosophy of “playing with open cards”: Simon's habit of being radically honest with investors, including telling Lachy Groom he'd return the money if turbopuffer didn't hit PMF by year-end • The “P99 engineer”: Simon's framework for building a talent-dense company, rejecting by default unless someone on the team feels strongly enough to fight for the candidate —Simon Hørup Eskildsen• LinkedIn: https://www.linkedin.com/in/sirupsen• X: https://x.com/Sirupsen• https://sirupsen.com/aboutturbopuffer• https://turbopuffer.com/Full Video PodTimestamps00:00:00 The PMF promise to Lachy Groom00:00:25 Intro and Simon's background00:02:19 What turbopuffer actually is00:06:26 Shopify, Elasticsearch, and the pain behind the company00:10:07 The Readwise experiment that sparked turbopuffer00:12:00 The insight Simon couldn't stop thinking about00:17:00 S3 consistency, NVMe, and the architecture bet00:20:12 The Notion story: latency, dark fiber, and conviction00:25:03 Build vs. buy in the age of AI00:26:00 The Cursor story: early launch to breakout customer00:29:00 Why code search still matters00:32:00 Search in the age of agents00:34:22 Pricing turbopuffer in the AI era00:38:17 Why Simon chose Lachy Groom00:41:28 Becoming a founder on purpose00:44:00 The “P99 engineer” philosophy00:49:30 Bending software to your will00:51:13 The future of turbopuffer00:57:05 Simon's tea obsession00:59:03 Tea kits, X Live, and P99 LiveTranscriptSimon Hørup Eskildsen: I don't think I've said this publicly before, but I just called Lockey and was like, local Lockie. Like if this doesn't have PMF by the end of the year, like we'll just like return all the money to you. But it's just like, I don't really, we, Justine and I don't wanna work on this unless it's really working.So we want to give it the best shot this year and like we're really gonna go for it. We're gonna hire a bunch of people. We're just gonna be honest with everyone. Like when I don't know how to play a game, I just play with open cards. Lockey was the only person that didn't, that didn't freak out. He was like, I've never heard anyone say that before.Alessio: Hey everyone, welcome to the Leading Space podcast. This is Celesio Pando, Colonel Laz, and I'm joined by Swix, editor of Leading Space.swyx: Hello. Hello, uh, we're still, uh, recording in the Ker studio for the first time. Very excited. And today we are joined by Simon Eski. Of Turbo Farer welcome.Simon Hørup Eskildsen: Thank you so much for having me.swyx: Turbo Farer has like really gone on a huge tear, and I, I do have to mention that like you're one of, you're not my newest member of the Danish AHU Mafia, where like there's a lot of legendary programmers that have come out of it, like, uh, beyond Trotro, Rasmus, lado Berg and the V eight team and, and Google Maps team.Uh, you're mostly a Canadian now, but isn't that interesting? There's so many, so much like strong Danish presence.Simon Hørup Eskildsen: Yeah, I was writing a post, um, not that long ago about sort of the influences. So I grew up in Denmark, right? I left, I left when, when I was 18 to go to Canada to, to work at Shopify. Um, and so I, like, I've, I would still say that I feel more Danish than, than Canadian.This is also the weird accent. I can't say th because it, this is like, I don't, you know, my wife is also Canadian, um, and I think. I think like one of the things in, in Denmark is just like, there's just such a ruthless pragmatism and there's also a big focus on just aesthetics. Like, they're like very, people really care about like where, what things look like.Um, and like Canada has a lot of attributes, US has, has a lot of attributes, but I think there's been lots of the great things to carry. I don't know what's in the water in Ahu though. Um, and I don't know that I could be considered part of the Mafi mafia quite yet, uh, compared to the phenomenal individuals we just mentioned.Barra OV is also, uh, Danish Canadian. Okay. Yeah. I don't know where he lives now, but, and he's the PHP.swyx: Yeah. And obviously Toby German, but moved to Canada as well. Yes. Like this is like import that, uh, that, that is an interesting, um, talent move.Alessio: I think. I would love to get from you. Definition of Turbo puffer, because I think you could be a Vector db, which is maybe a bad word now in some circles, you could be a search engine.It's like, let, let's just start there and then we'll maybe run through the history of how you got to this point.Simon Hørup Eskildsen: For sure. Yeah. So Turbo Puffer is at this point in time, a search engine, right? We do full text search and we do vector search, and that's really what we're specialized in. If you're trying to do much more than that, like then this might not be the right place yet, but Turbo Buffer is all about search.The other way that I think about it is that we can take all of the world's knowledge, all of the exabytes and exabytes of data that there is, and we can use those tokens to train a model, but we can't compress all of that into a few terabytes of weights, right? Compress into a few terabytes of weights, how to reason with the world, how to make sense of the knowledge.But we have to somehow connect it to something externally that actually holds that like in full fidelity and truth. Um, and that's the thing that we intend to become. Right? That's like a very holier than now kind of phrasing, right? But being the search engine for unstructured, unstructured data is the focus of turbo puffer at this point in time.Alessio: And let's break down. So people might say, well, didn't Elasticsearch already do this? And then some other people might say, is this search on my data, is this like closer to rag than to like a xr, like a public search thing? Like how, how do you segment like the different types of search?Simon Hørup Eskildsen: The way that I generally think about this is like, there's a lot of database companies and I think if you wanna build a really big database company, sort of, you need a couple of ingredients to be in the air.We don't, which only happens roughly every 15 years. You need a new workload. You basically need the ambition that every single company on earth is gonna have data in your database. Multiple times you look at a company like Oracle, right? You will, like, I don't think you can find a company on earth with a digital presence that it not, doesn't somehow have some data in an Oracle database.Right? And I think at this point, that's also true for Snowflake and Databricks, right? 15 years later it's, or even more than that, there's not a company on earth that doesn't, in. Or directly is consuming Snowflake or, or Databricks or any of the big analytics databases. Um, and I think we're in that kind of moment now, right?I don't think you're gonna find a company over the next few years that doesn't directly or indirectly, um, have all their data available for, for search and connect it to ai. So you need that new workload, like you need something to be happening where there's a new workload that causes that to happen, and that new workload is connecting very large amounts of data to ai.The second thing you need. The second condition to build a big database company is that you need some new underlying change in the storage architecture that is not possible from the databases that have come before you. If you look at Snowflake and Databricks, right, commoditized, like massive fleet of HDDs, like that was not possible in it.It just wasn't in the air in the nineties, right? So you just didn't, we just didn't build these systems. S3 and and and so on was not around. And I think the architecture that is now possible that wasn't possible 15 years ago is to go all in on NVME SSDs. It requires a particular type of architecture for the database that.It's difficult to retrofit onto the databases that are already there, including the ones you just mentioned. The second thing is to go all in on OIC storage, more so than we could have done 15 years ago. Like we don't have a consensus layer, we don't really have anything. In fact, you could turn off all the servers that Turbo Buffer has, and we would not lose any data because we have all completely all in on OIC storage.And this means that our architecture is just so simple. So that's the second condition, right? First being a new workload. That means that every company on earth, either indirectly or directly, is using your database. Second being, there's some new storage architecture. That means that the, the companies that have come before you can do what you're doing.I think the third thing you need to do to build a big database company is that over time you have to implement more or less every Cory plan on the data. What that means is that you. You can't just get stuck in, like, this is the one thing that a database does. It has to be ever evolving because when someone has data in the database, they over time expect to be able to ask it more or less every question.So you have to do that to get the storage architecture to the limit of what, what it's capable of. Those are the three conditions.swyx: I just wanted to get a little bit of like the motivation, right? Like, so you left Shopify, you're like principal, engineer, infra guy. Um, you also head of kernel labs, uh, inside of Shopify, right?And then you consulted for read wise and that it kind of gave you that, that idea. I just wanted you to tell that story. Um, maybe I, you've told it before, but, uh, just introduce the, the. People to like the, the new workload, the sort of aha moment for turbo PufferSimon Hørup Eskildsen: For sure. So yeah, I spent almost a decade at Shopify.I was on the infrastructure team, um, from the fairly, fairly early days around 2013. Um, at the time it felt like it was growing so quickly and everything, all the metrics were, you know, doubling year on year compared to the, what companies are contending with today. It's very cute in growth. I feel like lot some companies are seeing that month over month.Um, of course. Shopify compound has been compounding for a very long time now, but I spent a decade doing that and the majority of that was just make sure the site is up today and make sure it's up a year from now. And a lot of that was really just the, um, you know, uh, the Kardashians would drive very, very large amounts of, of data to, to uh, to Shopify as they were rotating through all the merch and building out their businesses.And we just needed to make sure we could handle that. Right. And sometimes these were events, a million requests per second. And so, you know, we, we had our own data centers back in the day and we were moving to the cloud and there was so much sharding work and all of that that we were doing. So I spent a decade just scaling databases ‘cause that's fundamentally what's the most difficult thing to scale about these sites.The database that was the most difficult for me to scale during that time, and that was the most aggravating to be on call for, was elastic search. It was very, very difficult to deal with. And I saw a lot of projects that were just being held back in their ambition by using it.swyx: And I mean, self-hosted.Self-hosted. ‘causeSimon Hørup Eskildsen: it's, yeah, and it commercial, this is like 2015, right? So it's like a very particular vintage. Right. It's probably better at a lot of these things now. Um, it was difficult to contend with and I'm just like, I just think about it. It's an inverted index. It should be good at these kinds of queries and do all of this.And it was, we, we often couldn't get it to do exactly what we needed to do or basically get lucine to do, like expose lucine raw to, to, to what we needed to do. Um, so that was like. Just something that we did on the side and just panic scaled when we needed to, but not a particular focus of mine. So I left, and when I left, I, um, wasn't sure exactly what I wanted to do.I mean, it spent like a decade inside of the same company. I'd like grown up there. I started working there when I was 18.swyx: You only do Rails?Simon Hørup Eskildsen: Yeah. I mean, yeah. Rails. And he's a Rails guy. Uh, love Rails. So good. Um,Alessio: we all wish we could still work in Rails.swyx: I know know. I know, but some, I tried learning Ruby.It's just too much, like too many options to do the same thing. It's, that's my, I I know there's a, there's a way to do it.Simon Hørup Eskildsen: I love it. I don't know that I would use it now, like given cloud code and, and, and cursor and everything, but, um, um, but still it, like if I'm just sitting down and writing a teal code, that's how I think.But anyway, I left and I wasn't, I talked to a couple companies and I was like, I don't. I need to see a little bit more of the world here to know what I'm gonna like focus on next. Um, and so what I decided is like I was gonna, I called it like angel engineering, where I just hopped around in my friend's companies in three months increments and just helped them out with something.Right. And, and just vested a bit of equity and solved some interesting infrastructure problem. So I worked with a bunch of companies at the time, um, read Wise was one of them. Replicate was one of them. Um, causal, I dunno if you've tried this, it's like a, it's a spreadsheet engine Yeah. Where you can do distribution.They sold recently. Yeah. Um, we've been, we used that in fp and a at, um, at Turbo Puffer. Um, so a bunch of companies like this and it was super fun. And so we're the Chachi bt moment happened, I was with. With read Wise for a stint, we were preparing for the reader launch, right? Which is where you, you cue articles and read them later.And I was just getting their Postgres up to snuff, like, which basically boils down to tuning, auto vacuum. So I was doing that and then this happened and we were like, oh, maybe we should build a little recommendation engine and some features to try to hook in the lms. They were not that good yet, but it was clear there was something there.And so I built a small recommendation engine just, okay, let's take the articles that you've recently read, right? Like embed all the articles and then do recommendations. It was good enough that when I ran it on one of the co-founders of Rey's, like I found out that I got articles about, about having a child.I'm like, oh my God, I didn't, I, I didn't know that, that they were having a child. I wasn't sure what to do with that information, but the recommendation engine was good enough that it was suggesting articles, um, about that. And so there was, there was recommendations and uh, it actually worked really well.But this was a company that was spending maybe five grand a month in total on all their infrastructure and. When I did the napkin math on running the embeddings of all the articles, putting them into a vector index, putting it in prod, it's gonna be like 30 grand a month. That just wasn't tenable. Right?Like Read Wise is a proudly bootstrapped company and it's paying 30 grand for infrastructure for one feature versus five. It just wasn't tenable. So sort of in the bucket of this is useful, it's pretty good, but let us, let's return to it when the costs come down.swyx: Did you say it grows by feature? So for five to 30 is by the number of, like, what's the, what's the Scaling factor scale?It scales by the number of articles that you embed.Simon Hørup Eskildsen: It does, but what I meant by that is like five grand for like all of the other, like the Heroku, dinos, Postgres, like all the other, and this then storage is 30. Yeah. And then like 30 grand for one feature. Right. Which is like, what other articles are related to this one.Um, so it was just too much right to, to power everything. Their budget would've been maybe a few thousand dollars, which still would've been a lot. And so we put it in a bucket of, okay, we're gonna do that later. We'll wait, we will wait for the cost to come down. And that haunted me. I couldn't stop thinking about it.I was like, okay, there's clearly some latent demand here. If the cost had been a 10th, we would've shipped it and. This was really the only data point that I had. Right. I didn't, I, I didn't, I didn't go out and talk to anyone else. It was just so I started reading Right. I couldn't, I couldn't help myself.Like I didn't know what like a vector index is. I, I generally barely do about how to generate the vectors. There was a lot of hype about, this is a early 2023. There was a lot of hype about vector databases. There were raising a lot of money and it's like, I really didn't know anything about it. It's like, you know, trying these little models, fine tuning them.Like I was just trying to get sort of a lay of the land. So I just sat down. I have this. A GitHub repository called Napkin Math. And on napkin math, there's just, um, rows of like, oh, this is how much bandwidth. Like this is how many, you know, you can do 25 gigabytes per second on average to dram. You can do, you know, five gigabytes per second of rights to an SSD, blah blah.All of these numbers, right? And S3, how many you could do per, how much bandwidth can you drive per connection? I was just sitting down, I was like, why hasn't anyone build a database where you just put everything on O storage and then you puff it into NVME when you use the data and you puff it into dram if you're, if you're querying it alive, it's just like, this seems fairly obvious and you, the only real downside to that is that if you go all in on o storage, every right will take a couple hundred milliseconds of latency, but from there it's really all upside, right?You do the first go, it takes half a second. And it sort of occurred to me as like, well. The architecture is really good for that. It's really good for AB storage, it's really good for nvm ESSD. It's, well, you just couldn't have done that 10 years ago. Back to what we were talking about before. You really have to build a database where you have as few round trips as possible, right?This is how CPUs work today. It's how NVM E SSDs work. It's how as, um, as three works that you want to have a very large amount of outstanding requests, right? Like basically go to S3, do like that thousand requests to ask for data in one round trip. Wait for that. Get that, like, make a new decision. Do it again, and try to do that maybe a maximum of three times.But no databases were designed that way within NVME as is ds. You can drive like within, you know, within a very low multiple of DRAM bandwidth if you use it that way. And same with S3, right? You can fully max out the network card, which generally is not maxed out. You get very, like, very, very good bandwidth.And, but no one had built a database like that. So I was like, okay, well can't you just, you know, take all the vectors right? And plot them in the proverbial coordinate system. Get the clusters, put a file on S3 called clusters, do json, and then put another file for every cluster, you know, cluster one, do js O cluster two, do js ON you know that like it's two round trips, right?So you get the clusters, you find the closest clusters, and then you download the cluster files like the, the closest end. And you could do this in two round trips.swyx: You were nearest neighbors locally.Simon Hørup Eskildsen: Yes. Yes. And then, and you would build this, this file, right? It's just like ultra simplistic, but it's not a far shot from what the first version of Turbo Buffer was.Why hasn't anyone done thatAlessio: in that moment? From a workload perspective, you're thinking this is gonna be like a read heavy thing because they're doing recommend. Like is the fact that like writes are so expensive now? Oh, with ai you're actually not writing that much.Simon Hørup Eskildsen: At that point I hadn't really thought too much about, well no actually it was always clear to me that there was gonna be a lot of rights because at Shopify, the search clusters were doing, you know, I don't know, tens or hundreds of crew QPS, right?‘cause you just have to have a human sit and type in. But we did, you know, I don't know how many updates there were per second. I'm sure it was in the millions, right into the cluster. So I always knew there was like a 10 to 100 ratio on the read write. In the read wise use case. It's, um, even, even in the read wise use case, there'd probably be a lot fewer reads than writes, right?There's just a lot of churn on the amount of stuff that was going through versus the amount of queries. Um, I wasn't thinking too much about that. I was mostly just thinking about what's the fundamentally cheapest way to build a database in the cloud today using the primitives that you have available.And this is it, right? You just, now you have one machine and you know, let's say you have a terabyte of data in S3, you paid the $200 a month for that, and then maybe five to 10% of that data and needs to be an NV ME SSDs and less than that in dram. Well. You're paying very, very little to inflate the data.swyx: By the way, when you say no one else has done that, uh, would you consider Neon, uh, to be on a similar path in terms of being sort of S3 first and, uh, separating the compute and storage?Simon Hørup Eskildsen: Yeah, I think what I meant with that is, uh, just build a completely new database. I don't know if we were the first, like it was very much, it was, I mean, I, I hadn't, I just looked at the napkin math and was like, this seems really obvious.So I'm sure like a hundred people came up with it at the same time. Like the light bulb and every invention ever. Right. It was just in the air. I think Neon Neon was, was first to it. And they're trying, they're retrofitted onto Postgres, right? And then they built this whole architecture where you have, you have it in memory and then you sort of.You know, m map back to S3. And I think that was very novel at the time to do it for, for all LTP, but I hadn't seen a database that was truly all in, right. Not retrofitting it. The database felt built purely for this no consensus layer. Even using compare and swap on optic storage to do consensus. I hadn't seen anyone go that all in.And I, I mean, there, there, I'm sure there was someone that did that before us. I don't know. I was just looking at the napkin mathswyx: and, and when you say consensus layer, uh, are you strongly relying on S3 Strong consistency? You are. Okay.SoSimon Hørup Eskildsen: that is your consensus layer. It, it is the consistency layer. And I think also, like, this is something that most people don't realize, but S3 only became consistent in December of 2020.swyx: I remember this coming out during COVID and like people were like, oh, like, it was like, uh, it was just like a free upgrade.Simon Hørup Eskildsen: Yeah.swyx: They were just, they just announced it. We saw consistency guys and like, okay, cool.Simon Hørup Eskildsen: And I'm sure that they just, they probably had it in prod for a while and they're just like, it's done right.And people were like, okay, cool. But. That's a big moment, right? Like nv, ME SSDs, were also not in the cloud until around 2017, right? So you just sort of had like 2017 nv, ME SSDs, and people were like, okay, cool. There's like one skew that does this, whatever, right? Takes a few years. And then the second thing is like S3 becomes consistent in 2020.So now it means you don't have to have this like big foundation DB or like zookeeper or whatever sitting there contending with the keys, which is how. You know, that's what Snowflake and others have do so muchswyx: for goneSimon Hørup Eskildsen: Exactly. Just gone. Right? And so just push to the, you know, whatever, how many hundreds of people they have working on S3 solved and then compare and swap was not in S3 at this point in time,swyx: by the way.Uh, I don't know what that is, so maybe you wanna explain. Yes. Yeah.Simon Hørup Eskildsen: Yes. So, um, what Compare and swap is, is basically, you can imagine that if you have a database, it might be really nice to have a file called metadata json. And metadata JSON could say things like, Hey, these keys are here and this file means that, and there's lots of metadata that you have to operate in the database, right?But that's the simplest way to do it. So now you have might, you might have a lot of servers that wanna change the metadata. They might have written a file and want the metadata to contain that file. But you have a hundred nodes that are trying to contend with this metadata that JSON well, what compare and Swap allows you to do is basically just you download the file, you make the modifications, and then you write it only if it hasn't changed.While you did the modification and if not you retry. Right? Should just have this retry loops. Now you can imagine if you have a hundred nodes doing that, it's gonna be really slow, but it will converge over time. That primitive was not available in S3. It wasn't available in S3 until late 2024, but it was available in GCP.The real story of this is certainly not that I sat down and like bake brained it. I was like, okay, we're gonna start on GCS S3 is gonna get it later. Like it was really not that we started, we got really lucky, like we started on GCP and we started on GCP because tur um, Shopify ran on GCP. And so that was the platform I was most available with.Right. Um, and I knew the Canadian team there ‘cause I'd worked with them at Shopify and so it was natural for us to start there. And so when we started building the database, we're like, oh yeah, we have to build a, we really thought we had to build a consensus layer, like have a zookeeper or something to do this.But then we discovered the compare and swap. It's like, oh, we can kick the can. Like we'll just do metadata r json and just, it's fine. It's probably fine. Um, and we just kept kicking the can until we had very, very strong conviction in the idea. Um, and then we kind of just hinged the company on the fact that S3 probably was gonna get this, it started getting really painful in like mid 2024.‘cause we were closing deals with, um, um, notion actually that was running in AWS and we're like, trust us. You, you really want us to run this in GCP? And they're like, no, I don't know about that. Like, we're running everything in AWS and the latency across the cloud were so big and we had so much conviction that we bought like, you know, dark fiber between the AWS regions in, in Oregon, like in the InterExchange and GCP is like, we've never seen a startup like do like, what's going on here?And we're just like, no, we don't wanna do this. We were tuning like TCP windows, like everything to get the latency down ‘cause we had so high conviction in not doing like a, a metadata layer on S3. So those were the three conditions, right? Compare and swap. To do metadata, which wasn't in S3 until late 2024 S3 being consistent, which didn't happen until December, 2020.Uh, 2020. And then NVMe ssd, which didn't end in the cloud until 2017.swyx: I mean, in some ways, like a very big like cloud success story that like you were able to like, uh, put this all together, but also doing things like doing, uh, bind our favor. That that actually is something I've never heard.Simon Hørup Eskildsen: I mean, it's very common when you're a big company, right?You're like connecting your own like data center or whatever. But it's like, it was uniquely just a pain with notion because the, um, the org, like most of the, like if you're buying in Ashburn, Virginia, right? Like US East, the Google, like the GCP and, and AWS data centers are like within a millisecond on, on each other, on the public exchanges.But in Oregon uniquely, the GCP data center sits like a couple hundred kilometers, like east of Portland and the AWS region sits in Portland, but the network exchange they go through is through Seattle. So it's like a full, like 14 milliseconds or something like that. And so anyway, yeah. It's, it's, so we were like, okay, we can't, we have to go through an exchange in Portland.Yeah. Andswyx: you'd rather do this than like run your zookeeper and likeSimon Hørup Eskildsen: Yes. Way rather. It doesn't have state, I don't want state and two systems. Um, and I think all that is just informed by Justine, my co-founder and I had just been on call for so long. And the worst outages are the ones where you have state in multiple places that's not syncing up.So it really came from, from a a, like just a, a very pure source of pain, of just imagining what we would be Okay. Being woken up at 3:00 AM about and having something in zookeeper was not one of them.swyx: You, you're talking to like a notion or something. Do they care or do they just, theySimon Hørup Eskildsen: just, they care about latency.swyx: They latency cost. That's it.Simon Hørup Eskildsen: They just cared about latency. Right. And we just absorbed the cost. We're just like, we have high conviction in this. At some point we can move them to AWS. Right. And so we just, we, we'll buy the fiber, it doesn't matter. Right. Um, and it's like $5,000. Usually when you buy fiber, you buy like multiple lines.And we're like, we can only afford one, but we will just test it that when it goes over the public internet, it's like super smooth. And so we did a lot of, anyway, it's, yeah, it was, that's cool.Alessio: You can imagine talking to the GCP rep and it's like, no, we're gonna buy, because we know we're gonna turn, we're gonna turn from you guys and go to AWS in like six months.But in the meantime we'll do this. It'sSimon Hørup Eskildsen: a, I mean, like they, you know, this workload still runs on GCP for what it's worth. Right? ‘cause it's so, it was just, it was so reliable. So it was never about moving off GCP, it was just about honesty. It was just about giving notion the latency that they deserved.Right. Um, and we didn't want ‘em to have to care about any of this. We also, they were like, oh, egress is gonna be bad. It was like, okay, screw it. Like we're just gonna like vvc, VPC peer with you and AWS we'll eat the cost. Yeah. Whatever needs to be done.Alessio: And what were the actual workloads? Because I think when you think about ai, it's like 14 milliseconds.It's like really doesn't really matter in the scheme of like a model generation.Simon Hørup Eskildsen: Yeah. We were told the latency, right. That we had to beat. Oh, right. So, so we're just looking at the traces. Right. And then sort of like hand draw, like, you know, kind of like looking at the trace and then thinking what are the other extensions of the trace?Right. And there's a lot more to it because it's also when you have, if you have 14 versus seven milliseconds, right. You can fit in another round trip. So we had to tune TCP to try to send as much data in every round trip, prewarm all the connections. And there was, there's a lot of things that compound from having these kinds of round trips, but in the grand scheme it was just like, well, we have to beat the latency of whatever we're up against.swyx: Which is like they, I mean, notion is a database company. They could have done this themselves. They, they do lots of database engineering themselves. How do you even get in the door? Like Yeah, just like talk through that kind of.Simon Hørup Eskildsen: Last time I was in San Francisco, I was talking to one of the engineers actually, who, who was one of our champions, um, at, AT Notion.And they were, they were just trying to make sure that the, you know, per user cost matched the economics that they needed. You know, Uhhuh like, it's like the way I think about, it's like I have to earn a return on whatever the clouds charge me and then my customers have to earn a return on that. And it's like very simple, right?And so there has to be gross margin all the way up and that's how you build the product. And so then our customers have to make the right set of trade off the turbo Puffer makes, and if they're happy with that, that's great.swyx: Do you feel like you're competing with build internally versus buy or buy versus buy?Simon Hørup Eskildsen: Yeah, so, sorry, this was all to build up to your question. So one of the notion engineers told me that they'd sat and probably on a napkin, like drawn out like, why hasn't anyone built this? And then they saw terrible. It was like, well, it literally that. So, and I think AI has also changed the buy versus build equation in terms of, it's not really about can we build it, it's about do we have time to build it?I think they like, I think they felt like, okay, if this is a team that can do that and they, they feel enough like an extension of our team, well then we can go a lot faster, which would be very, very good for them. And I mean, they put us through the, through the test, right? Like we had some very, very long nights to to, to do that POC.And they were really our biggest, our second big customer off the cursor, which also was a lot of late nights. Right.swyx: Yeah. That, I mean, should we go into that story? The, the, the sort of Chris's story, like a lot, um, they credit you a lot for. Working very closely with them. So I just wanna hear, I've heard this, uh, story from Sole's point of view, but like, I'm curious what, what it looks like from your side.Simon Hørup Eskildsen: I actually haven't heard it from Sole's point of view, so maybe you can now cross reference it. The way that I remember it was that, um, the day after we launched, which was just, you know, I'd worked the whole summer on, on the first version. Justine wasn't part of it yet. ‘cause I just, I didn't tell anyone that summer that I was working on this.I was just locked in on building it because it's very easy otherwise to confuse talking about something to actually doing it. And so I was just like, I'm not gonna do that. I'm just gonna do the thing. I launched it and at this point turbo puffer is like a rust binary running on a single eight core machine in a T Marks instance.And me deploying it was like looking at the request log and then like command seeing it or like control seeing it to just like, okay, there's no request. Let's upgrade the binary. Like it was like literally the, the, the, the scrappiest thing. You could imagine it was on purpose because just like at Shopify, we did that all the time.Like, we like move, like we ran things in tux all the time to begin with. Before something had like, at least the inkling of PMF, it was like, okay, is anyone gonna hear about this? Um, and one of the cursor co-founders Arvid reached out and he just, you know, the, the cursor team are like all I-O-I-I-M-O like, um, contenders, right?So they just speak in bullet points and, and facts. It was like this amazing email exchange just of, this is how many QPS we have, this is what we're paying, this is where we're going, blah, blah, blah. And so we're just conversing in bullet points. And I tried to get a call with them a few times, but they were, so, they were like really writing the PMF bowl here, just like late 2023.And one time Swally emails me at like five. What was it like 4:00 AM Pacific time saying like, Hey, are you open for a call now? And I'm on the East coast and I, it was like 7:00 AM I was like, yeah, great, sure, whatever. Um, and we just started talking and something. Then I didn't know anything about sales.It was something that just comp compelled me. I have to go see this team. Like, there's something here. So I, I went to San Francisco and I went to their office and the way that I remember it is that Postgres was down when I showed up at the office. Did SW tell you this? No. Okay. So Postgres was down and so it's like they were distracting with that.And I was trying my best to see if I could, if I could help in any way. Like I knew a little bit about databases back to tuning, auto vacuum. It was like, I think you have to tune out a vacuum. Um, and so we, we talked about that and then, um, that evening just talked about like what would it look like, what would it look like to work with us?And I just said. Look like we're all in, like we will just do what we'll do whatever, whatever you tell us, right? They migrated everything over the next like week or two, and we reduced their cost by 95%, which I think like kind of fixed their per user economics. Um, and it solved a lot of other things. And we were just, Justine, this is also when I asked Justine to come on as my co-founder, she was the best engineer, um, that I ever worked with at Shopify.She lived two blocks away and we were just, okay, we're just gonna get this done. Um, and we did, and so we helped them migrate and we just worked like hell over the next like month or two to make sure that we were never an issue. And that was, that was the cursor story. Yeah.swyx: And, and is code a different workload than normal text?I, I don't know. Is is it just text? Is it the same thing?Simon Hørup Eskildsen: Yeah, so cursor's workload is basically, they, um, they will embed the entire code base, right? So they, they will like chunk it up in whatever they would, they do. They have their own embedding model, um, which they've been public about. Um, and they find that on, on, on their evals.It. There's one of their evals where it's like a 25% improvement on a very particular workload. They have a bunch of blog posts about it. Um, I think it works best on larger code basis, but they've trained their own embedding model to do this. Um, and so you'll see it if you use the cursor agent, it will do searches.And they've also been public around, um, how they've, I think they post trained their model to be very good at semantic search as well. Um, and that's, that's how they use it. And so it's very good at, like, can you find me on the code that's similar to this, or code that does this? And just in, in this queries, they also use GR to supplement it.swyx: Yeah.Simon Hørup Eskildsen: Um, of courseswyx: it's been a big topic of discussion like, is rag dead because gr you know,Simon Hørup Eskildsen: and I mean like, I just, we, we see lots of demand from the coding company to ethicsswyx: search in every part. Yes.Simon Hørup Eskildsen: Uh, we, we, we see demand. And so, I mean, I'm. I like case studies. I don't like, like just doing like thought pieces on this is where it's going.And like trying to be all macroeconomic about ai, that's has turned out to be a giant waste of time because no one can really predict any of this. So I just collect case studies and I mean, cursor has done a great job talking about what they're doing and I hope some of the other coding labs that use Turbo Puffer will do the same.Um, but it does seem to make a difference for particular queries. Um, I mean we can also do text, we can also do RegX, but I should also say that cursors like security posture into Tur Puffer is exceptional, right? They have their own embedding model, which makes it very difficult to reverse engineer. They obfuscate the file paths.They like you. It's very difficult to learn anything about a code base by looking at it. And the other thing they do too is that for their customers, they encrypt it with their encryption keys in turbo puffer's bucket. Um, so it's, it's, it's really, really well designed.swyx: And so this is like extra stuff they did to work with you because you are not part of Cursor.Exactly like, and this is just best practice when working in any database, not just you guys. Okay. Yeah, that makes sense. Yeah. I think for me, like the, the, the learning is kind of like you, like all workloads are hybrid. Like, you know, uh, like you, you want the semantic, you want the text, you want the RegX, you want sql.I dunno. Um, but like, it's silly to like be all in on like one particularly query pattern.Simon Hørup Eskildsen: I think, like I really like the way that, um, um, that swally at cursor talks about it, which is, um, I'm gonna butcher it here. Um, and you know, I'm a, I'm a database scalability person. I'm not a, I, I dunno anything about training models other than, um, what the internet tells me and what.The way he describes is that this is just like cash compute, right? It's like you have a point in time where you're looking at some particular context and focused on some chunk and you say, this is the layer of the neural net at this point in time. That seems fundamentally really useful to do cash compute like that.And, um, how the value of that will change over time. I'm, I'm not sure, but there seems to be a lot of value in that.Alessio: Maybe talk a bit about the evolution of the workload, because even like search, like maybe two years ago it was like one search at the start of like an LLM query to build the context. Now you have a gentech search, however you wanna call it, where like the model is both writing and changing the code and it's searching it again later.Yeah. What are maybe some of the new types of workloads or like changes you've had to make to your architecture for it?Simon Hørup Eskildsen: I think you're right. When I think of rag, I think of, Hey, there's an 8,000 token, uh, context window and you better make it count. Um, and search was a way to do that now. Everything is moving towards the, just let the agent do its thing.Right? And so back to the thing before, right? The LLM is very good at reasoning with the data, and so we're just the tool call, right? And that's increasingly what we see our customers doing. Um, what we're seeing more demand from, from our customers now is to do a lot of concurrency, right? Like Notion does a ridiculous amount of queries in every round trip just because they can't.And I'm also now, when I use the cursor agent, I also see them doing more concurrency than I've ever seen before. So a bit similar to how we designed a database to drive as much concurrency in every round trip as possible. That's also what the agents are doing. So that's new. It means just an enormous amount of queries all at once to the dataset while it's warm in as few turns as possible.swyx: Can I clarify one thing on that?Simon Hørup Eskildsen: Yes.swyx: Is it, are they batching multiple users or one user is driving multiple,Simon Hørup Eskildsen: one user driving multiple, one agent driving.swyx: It's parallel searching a bunch of things.Simon Hørup Eskildsen: Exactly.swyx: Yeah. Yeah, exactly. So yeah, the clinician also did, did this for the fast context thing, like eight parallel at once.Simon Hørup Eskildsen: Yes.swyx: And, and like an interesting problem is, well, how do you make sure you have enough diversity so you're not making the the same request eight times?Simon Hørup Eskildsen: And I think like that's probably also where the hybrid comes in, where. That's another way to diversify. It's a completely different way to, to do the search.That's a big change, right? So before it was really just like one call and then, you know, the LLM took however many seconds to return, but now we just see an enormous amount of queries. So the, um, we just see more queries. So we've like tried to reduce query, we've reduced query pricing. Um, this is probably the first time actually I'm saying that, but the query pricing is being reduced, like five x.Um, and we'll probably try to reduce it even more to accommodate some of these workloads of just doing very large amounts of queries. Um, that's one thing that's changed. I think the right, the right ratio is still very high, right? Like there's still a, an enormous amount of rights per read, but we're starting probably to see that change if people really lean into this pattern.Alessio: Can we talk a little bit about the pricing? I'm curious, uh, because traditionally a database would charge on storage, but now you have the token generation that is so expensive, where like the actual. Value of like a good search query is like much higher because they're like saving inference time down the line.How do you structure that as like, what are people receptive to on the other side too?Simon Hørup Eskildsen: Yeah. I, the, the turbo puffer pricing in the beginning was just very simple. The pricing on these on for search engines before Turbo Puffer was very server full, right? It was like, here's the vm, here's the per hour cost, right?Great. And I just sat down with like a piece of paper and said like, if Turbo Puffer was like really good, this is probably what it would cost with a little bit of margin. And that was the first pricing of Turbo Puffer. And I just like sat down and I was like, okay, like this is like probably the storage amp, but whenever on a piece of paper I, it was vibe pricing.It was very vibe price, and I got it wrong. Oh. Um, well I didn't get it wrong, but like Turbo Puffer wasn't at the first principle pricing, right? So when Cursor came on Turbo Puffer, it was like. Like, I didn't know any VCs. I didn't know, like I was just like, I don't know, I didn't know anything about raising money or anything like that.I just saw that my GCP bill was, was high, was a lot higher than the cursor bill. So Justine and I was just like, well, we have to optimize it. Um, and I mean, to the chagrin now of, of it, of, of the VCs, it now means that we're profitable because we've had so much pricing pressure in the beginning. Because it was running on my credit card and Justine and I had spent like, like tens of thousands of dollars on like compute bills and like spinning off the company and like very like, like bad Canadian lawyers and like things like to like get all of this done because we just like, we didn't know.Right. If you're like steeped in San Francisco, you're just like, you just know. Okay. Like you go out, raise a pre-seed round. I, I never heard a word pre-seed at this point in time.swyx: When you had Cursor, you had Notion you, you had no funding.Simon Hørup Eskildsen: Um, with Cursor we had no funding. Yeah. Um, by the time we had Notion Locke was, Locke was here.Yeah. So it was really just, we vibe priced it 100% from first Principles, but it wasn't, it, it was not performing at first principles, so we just did everything we could to optimize it in the beginning for that, so that at least we could have like a 5% margin or something. So I wasn't freaking out because Cursor's bill was also going like this as they were growing.And so my liability and my credit limit was like actively like calling my bank. It was like, I need a bigger credit. Like it was, yeah. Anyway, that was the beginning. Yeah. But the pricing was, yeah, like storage rights and query. Right. And the, the pricing we have today is basically just that pricing with duct tape and spit to try to approach like, you know, like a, as a margin on the physical underlying hardware.And we're doing this year, you're gonna see more and more pricing changes from us. Yeah.swyx: And like is how much does stuff like VVC peering matter because you're working in AWS land where egress is charged and all that, you know.Simon Hørup Eskildsen: We probably don't like, we have like an enterprise plan that just has like a base fee because we haven't had time to figure out SKU pricing for all of this.Um, but I mean, yeah, you can run turbo puffer either in SaaS, right? That's what Cursor does. You can run it in a single tenant cluster. So it's just you. That's what Notion does. And then you can run it in, in, in BYOC where everything is inside the customer's VPC, that's what an for example, philanthropic does.swyx: What I'm hearing is that this is probably the best CRO job for somebody who can come in and,Simon Hørup Eskildsen: I mean,swyx: help you with this.Simon Hørup Eskildsen: Um, like Turbo Puffer hired, like, I don't know what, what number this was, but we had a full-time CFO as like the 12th hire or something at Turbo Puffer, um, I think I hear are a lot of comp.I don't know how they do it. Like they have a hundred employees and not a CFO. It's like having a CFO is like a runningswyx: business man. Like, you know,Simon Hørup Eskildsen: it's so good. Yeah, like money Mike, like he just, you know, just handles the money and a lot of the business stuff and so he came in and just hopped with a lot of the operational side of the business.So like C-O-O-C-F-O, like somewhere in between.swyx: Just as quick mention of Lucky, just ‘cause I'm curious, I've met Lock and like, he's obviously a very good investor and now on physical intelligence, um, I call it generalist super angel, right? He invests in everything. Um, and I always wonder like, you know, is there something appealing about focusing on developer tooling, focusing on databases, going like, I've invested for 10 years in databases versus being like a lock where he can maybe like connect you to all the customers that you need.Simon Hørup Eskildsen: This is an excellent question. No, no one's asked me this. Um, why lockey? Because. There was a couple of people that we were talking to at the time and when we were raising, we were almost a little, we were like a bit distressed because one of our, one of our peers had just launched something that was very similar to Turbo Puffer.And someone just gave me the advice at the time of just choose the person where you just feel like you can just pick up the phone and not prepare anything. And just be completely honest, and I don't think I've said this publicly before, but I just called Lockey and was like local Lockie. Like if this doesn't have PMF by the end of the year, like we'll just like return all the money to you.But it's just like, I don't really, we, Justine and I don't wanna work on this unless it's really working. So we want to give it the best shot this year and like we're really gonna go for it. We're gonna hire a bunch of people and we're just gonna be honest with everyone. Like when I don't know how to play a game, I just play with open cards and.Lockey was the only person that didn't, that didn't freak out. He was like, I've never heard anyone say that before. As I said, I didn't even know what a seed or pre-seed round was like before, probably even at this time. So I was just like very honest with him. And I asked him like, Lockie, have you ever have, have you ever invested in database company?He was just like, no. And at the time I was like, am I dumb? Like, but I think there was something that just like really drew me to Lockie. He is so authentic, so honest, like, and there was something just like, I just felt like I could just play like, just say everything openly. And that was, that was, I think that that was like a perfect match at the time, and, and, and honestly still is.He was just like, okay, that's great. This is like the most honest, ridiculous thing I've ever heard anyone say to me. But like that, like that, whyswyx: is this ridiculous? Say competitor launch, this may not work out. It wasSimon Hørup Eskildsen: more just like. If this doesn't work out, I'm gonna close up shop by the end of the mo the year, right?Like it was, I don't know, maybe it's common. I, I don't know. He told me it was uncommon. I don't know. Um, that's why we chose him and he'd been phenomenal. The other people were talking at the, at the time were database experts. Like they, you know, knew a lot about databases and Locke didn't, this turned out to be a phenomenal asset.Right. I like Justine and I know a lot about databases. The people that we hire know a lot about databases. What we needed was just someone who didn't know a lot about databases, didn't pretend to know a lot about databases, and just wanted to help us with candidates and customers. And he did. Yeah. And I have a list, right, of the investors that I have a relationship with, and Lockey has just performed excellent in the number of sub bullets of what we can attribute back to him.Just absolutely incredible. And when people talk about like no ego and just the best thing for the founder, I like, I don't think that anyone, like even my lawyer is like, yeah, Lockey is like the most friendly person you will find.swyx: Okay. This is my most glow recommendation I've ever heard.Alessio: He deserves it.He's very special.swyx: Yeah. Yeah. Yeah. Okay. Amazing.Alessio: Since you mentioned candidates, maybe we can talk about team building, you know, like, especially in sf, it feels like it's just easier to start a company than to join a company. Uh, I'm curious your experience, especially not being n SF full-time and doing something that is maybe, you know, a very low level of detail and technical detail.Simon Hørup Eskildsen: Yeah. So joining versus starting, I never thought that I would be a founder. I would start with it, like Turbo Puffer started as a blog post, and then it became a project and then sort of almost accidentally became a company. And now it feels like it's, it's like becoming a bigger company. That was never the intention.The intentions were very pure. It's just like, why hasn't anyone done this? And it's like, I wanna be the, like, I wanna be the first person to do it. I think some founders have this, like, I could never work for anyone else. I, I really don't feel that way. Like, it's just like, I wanna see this happen. And I wanna see it happen with some people that I really enjoy working with and I wanna have fun doing it and this, this, this has all felt very natural on that, on that sense.So it was never a like join versus versus versus found. It was just dis found me at the right moment.Alessio: Well I think there's an argument for, you should have joined Cursor, right? So I'm curious like how you evaluate it. Okay, I should actually go raise money and make this a company versus like, this is like a company that is like growing like crazy.It's like an interesting technical problem. I should just build it within Cursor and then they don't have to encrypt all this stuff. They don't have to obfuscate things. Like was that on your mind at all orSimon Hørup Eskildsen: before taking the, the small check from Lockie, I did have like a hard like look at myself in the mirror of like, okay, do I really want to do this?And because if I take the money, I really have to do it right. And so the way I almost think about it's like you kind of need to ha like you kind of need to be like fucked up enough to want to go all the way. And that was the conversation where I was like, okay, this is gonna be part of my life's journey to build this company and do it in the best way that I possibly can't.Because if I ask people to join me, ask people to get on the cap table, then I have an ultimate responsibility to give it everything. And I don't, I think some people, it doesn't occur to me that everyone takes it that seriously. And maybe I take it too seriously, I don't know. But that was like a very intentional moment.And so then it was very clear like, okay, I'm gonna do this and I'm gonna give it everything.Alessio: A lot of people don't take it this seriously. But,swyx: uh, let's talk about, you have this concept of the P 99 engineer. Uh, people are 10 x saying, everyone's saying, you know, uh, maybe engineers are out of a job. I don't know.But you definitely see a P 99 engineer, and I just want you to talk about it.Simon Hørup Eskildsen: Yeah, so the P 99 engineer was just a term that we started using internally to talk about candidates and talk about how we wanted to build the company. And you know, like everyone else is, like we want a talent dense company.And I think that's almost become trite at this point. What I credit the cursor founders a lot with is that they just arrived there from first principles of like, we just need a talent dense, um, talent dense team. And I think I've seen some teams that weren't talent dense and like seemed a counterfactual run, which if you've run in been in a large company, you will just see that like it's just logically will happen at a large company.Um, and so that was super important to me and Justine and it's very difficult to maintain. And so we just needed, we needed wording for it. And so I have a document called Traits of the P 99 Engineer, and it's a bullet point list. And I look at that list after every single interview that I do, and in every single recap that we do and every recap we end with.End with, um, some version of I'm gonna reject this candidate completely regardless of what the discourse was, because I wanna see people fight for this person because the default should not be, we're gonna hire this person. The default should be, we're definitely not hiring this person. And you know, if everyone was like, ah, maybe throw a punch, then this is not the right.swyx: Do, do you operate, like if there's one cha there must have at least one champion who's like, yes, I will put my career on, on, on the line for this. You know,Simon Hørup Eskildsen: I think career on the line,swyx: maybe a chair, butSimon Hørup Eskildsen: yeah. You know, like, um, I would say so someone needs to like, have both fists up and be like, I'd fight.Right? Yeah. Yeah. And if one person said, then, okay, let's do it. Right?swyx: Yeah.Simon Hørup Eskildsen: Um. It doesn't have to be absolutely everyone. Right? And like the interviews are always the sign that you're checking for different attributes. And if someone is like knocking it outta the park in every single attribute, that's, that's fairly rare.Um, but that's really important. And so the traits of the P 99 engineer, there's lots of them. There's also the traits of the p like triple nine engineer and the quadruple nine engineer. This is like, it's a long list.swyx: Okay.Simon Hørup Eskildsen: Um, I'll give you some samples, right. Of what we, what we look for. I think that the P 99 engineer has some history of having bent, like their trajectory or something to their will.Right? Some moment where it was just, they just, you know, made the computer do what it needed to do. There's something like that, and it will, it will occur to have them at some point in their career. And, uh. Hopefully multiple times. Right.swyx: Gimme an example of one of your engineers that like,Simon Hørup Eskildsen: I'll give an eng.Uh, so we, we, we launched this thing called A and NV three. Um, we could, we're also, we're working on V four and V five right now, but a and NV three can search a hundred billion vectors with a P 50 of around 40 milliseconds and a p 99 of 200 milliseconds. Um, maybe other people have done this, I'm sure Google and others have done this, but, uh, we haven't seen anyone, um, at least not in like a public consumable SaaS that can do this.And that was an engineer, the chief architect of Turbo Puffer, Nathan, um, who more or less just bent this, the software was not capable of this and he just made it capable for a very particular workload in like a, you know, six to eight week period with the help of a lot of the team. Right. It's been, been, there's numerous of examples of that, like at, at turbo puff, but that's like really bending the software and X 86 to your will.It was incredible to watch. Um. You wanna see some moments like that?swyx: Isn't that triple nine?Simon Hørup Eskildsen: Um, I think Nathan, what's calledAlessio: group nine, that was only nine. I feel like this is too high forSimon Hørup Eskildsen: Nathan. Nathan is, uh, Nathan is like, yeah, there's a lot of nines. Okay. After that p So I think that's one trait. I think another trait is that, uh, the P 99 spends a lot of time looking at maps.Generally it's their preferred ux. They just love looking at maps. You ever seen someone who just like, sits on their phone and just like, scrolls around on a map? Or did you not look at maps A lot? You guys don't look atswyx: maps? I guess I'm not feeling there. I don't know, butSimon Hørup Eskildsen: you just dis What about trains?Do you like trains?swyx: Uh, I mean they, not enough. Okay. This is just like weapon nice. Autism is what I call it. Like, like,Simon Hørup Eskildsen: um, I love looking at maps, like, it's like my preferred UX and just like I, you know, I likeswyx: lotsAlessio: of, of like random places, soswyx: like,youswyx: know.Alessio: Yes. Okay. There you go. So instead of like random places, like how do you explore the maps?Simon Hørup Eskildsen: No, it's, it's just a joke.swyx: It's autism laugh. It's like you are just obsessed by something and you like studying a thing.Simon Hørup Eskildsen: The origin of this was that at some point I read an interview with some IOI gold medalistswyx: Uhhuh,Simon Hørup Eskildsen: and it's like, what do you do in your spare time? I was just like, I like looking at maps.I was like, I feel so seen. Like, I just like love, like swirling out. I was like, oh, Canada is so big. Where's Baffin Island? I don't know. I love it. Yeah. Um, anyway, so the traits of P 99, P 99 is obsessive, right? Like, there's just like, you'll, you'll find traits of that we do an interview at, at, at, at turbo puffer or like multiple interviews that just try to screen for some of these things.Um, so. There's lots of others, but these are the kinds of traits that we look for.swyx: I'll tell you, uh, some people listen for like some of my dere stuff. Uh, I do think about derel as maps. Um, you draw a map for people, uh, maps show you the, uh, what is commonly agreed to be the geographical features of what a boundary is.And it shows also shows you what is not doing. And I, I think a lot of like developer tools, companies try to tell you they can do everything, but like, let's, let's be real. Like you, your, your three landmarks are here, everyone comes here, then here, then here, and you draw a map and, and then you draw a journey through the map.And like that. To me, that's what developer relations looks like. So I do think about things that way.Simon Hørup Eskildsen: I think the P 99 thinks in offs, right? The P 99 is very clear about, you know, hey, turbo puffer, you can't run a high transaction workload on turbo puffer, right? It's like the right latency is a hundred milliseconds.That's a clear trade off. I think the P 99 is very good at articulating the trade offs in every decision. Um. Which is exactly what the map is in your case, right?swyx: Uh, yeah, yeah. My, my, my world. My world.Alessio: How, how do you reconcile some of these things when you're saying you bend the will the computer versus like the trade

Join Kyle, Nader, Vibhu, and swyx live at NVIDIA GTC next week!Now that AIE Europe tix are ~sold out, our attention turns to Miami and World's Fair!The definitive AI Accelerator chip company has more than 10xed this AI Summer:And is now a $4.4 trillion megacorp… that is somehow still moving like a startup. We are blessed to have a unique relationship with our first ever NVIDIA guests: Kyle Kranen who gave a great inference keynote at the first World's Fair and is one of the leading architects of NVIDIA Dynamo (a Datacenter scale inference framework supporting SGLang, TRT-LLM, vLLM), and Nader Khalil, a friend of swyx from our days in Celo in The Arena, who has been drawing developers at GTC since before they were even a glimmer in the eye of NVIDIA:Nader discusses how NVIDIA Brev has drastically reduced the barriers to entry for developers to get a top of the line GPU up and running, and Kyle explains NVIDIA Dynamo as a data center scale inference engine that optimizes serving by scaling out, leveraging techniques like prefill/decode disaggregation, scheduling, and Kubernetes-based orchestration, framed around cost, latency, and quality tradeoffs. We also dive into Jensen's “SOL” (Speed of Light) first-principles urgency concept, long-context limits and model/hardware co-design, internal model APIs (https://build.nvidia.com), and upcoming Dynamo and agent sessions at GTC.Full Video pod on YouTubeTimestamps00:00 Agent Security Basics00:39 Podcast Welcome and Guests07:19 Acquisition and DevEx Shift13:48 SOL Culture and Dynamo Setup27:38 Why Scale Out Wins29:02 Scale Up Limits Explained30:24 From Laptop to Multi Node33:07 Cost Quality Latency Tradeoffs38:42 Disaggregation Prefill vs Decode41:05 Kubernetes Scaling with Grove43:20 Context Length and Co Design57:34 Security Meets Agents58:01 Agent Permissions Model59:10 Build Nvidia Inference Gateway01:01:52 Hackathons And Autonomy Dreams01:10:26 Local GPUs And Scaling Inference01:15:31 Long Running Agents And SF ReflectionsTranscriptAgent Security BasicsNader: Agents can do three things. They can access your files, they can access the internet, and then now they can write custom code and execute it. You literally only let an agent do two of those three things. If you can access your files and you can write custom code, you don't want internet access because that's one to see full vulnerability, right?If you have access to internet and your file system, you should know the full scope of what that agent's capable of doing. Otherwise, now we can get injected or something that can happen. And so that's a lot of what we've been thinking about is like, you know, how do we both enable this because it's clearly the future.But then also, you know, what, what are these enforcement points that we can start to like protect?swyx: All right.Podcast Welcome and Guestsswyx: Welcome to the Lean Space podcast in the Chromo studio. Welcome to all the guests here. Uh, we are back with our guest host Viu. Welcome. Good to have you back. And our friends, uh, Netter and Kyle from Nvidia. Welcome.Kyle: Yeah, thanks for having us.swyx: Yeah, thank you. Actually, I don't even know your titles.Uh, I know you're like architect something of Dynamo.Kyle: Yeah. I, I'm one of the engineering leaders [00:01:00] and a architects of Dynamo.swyx: And you're director of something and developers, developer tech.Nader: Yeah.swyx: You're the developers, developers, developers guy at nvidia,Nader: open source agent marketing, brev,swyx: and likeNader: Devrel tools and stuff.swyx: Yeah. BeenNader: the focus.swyx: And we're, we're kind of recording this ahead of Nvidia, GTC, which is coming to town, uh, again, uh, or taking over town, uh, which, uh, which we'll all be at. Um, and we'll talk a little bit about your sessions and stuff. Yeah.Nader: We're super excited for it.GTC Booth Stunt Storiesswyx: One of my favorite memories for Nader, like you always do like marketing stunts and like while you were at Rev, you like had this surfboard that you like, went down to GTC with and like, NA Nvidia apparently, like did so much that they bought you.Like what, what was that like? What was that?Nader: Yeah. Yeah, we, we, um. Our logo was a chaka. We, we, uh, we were always just kind of like trying to keep true to who we were. I think, you know, some stuff, startups, you're like trying to pretend that you're a bigger, more mature company than you are. And it was actually Evan Conrad from SF Compute who was just like, you guys are like previousswyx: guest.Yeah.Nader: Amazing. Oh, really? Amazing. Yeah. He was just like, guys, you're two dudes in the room. Why are you [00:02:00] pretending that you're not? Uh, and so then we were like, okay, let's make the logo a shaka. We brought surfboards to our booth to GTC and the energy was great. Yeah. Some palm trees too. They,Kyle: they actually poked out over like the, the walls so you could, you could see the bread booth.Oh, that's so funny. AndNader: no one else,Kyle: just from very far away.Nader: Oh, so you remember it backKyle: then? Yeah I remember it pre-acquisition. I was like, oh, those guys look cool,Nader: dude. That makes sense. ‘cause uh, we, so we signed up really last minute, and so we had the last booth. It was all the way in the corner. And so I was, I was worried that no one was gonna come.So that's why we had like the palm trees. We really came in with the surfboards. We even had one of our investors bring her dog and then she was just like walking the dog around to try to like, bring energy towards our booth. Yeah.swyx: Steph.Kyle: Yeah. Yeah, she's the best,swyx: you know, as a conference organizer, I love that.Right? Like, it's like everyone who sponsors a conference comes, does their booth. They're like, we are changing the future of ai or something, some generic b******t and like, no, like actually try to stand out, make it fun, right? And people still remember it after three years.Nader: Yeah. Yeah. You know what's so funny?I'll, I'll send, I'll give you this clip if you wanna, if you wanna add it [00:03:00] in, but, uh, my wife was at the time fiance, she was in medical school and she came to help us. ‘cause it was like a big moment for us. And so we, we bought this cricket, it's like a vinyl, like a vinyl, uh, printer. ‘cause like, how else are we gonna label the surfboard?So, we got a surfboard, luckily was able to purchase that on the company card. We got a cricket and it was just like fine tuning for enterprises or something like that, that we put on the. On the surfboard and it's 1:00 AM the day before we go to GTC. She's helping me put these like vinyl stickers on.And she goes, you son of, she's like, if you pull this off, you son of a b***h. And so, uh, right. Pretty much after the acquisition, I stitched that with the mag music acquisition. I sent it to our family group chat. Ohswyx: Yeah. No, well, she, she made a good choice there. Was that like basically the origin story for Launchable is that we, it was, and maybe we should explain what Brev is andNader: Yeah.Yeah. Uh, I mean, brev is just, it's a developer tool that makes it really easy to get a GPU. So we connect a bunch of different GPU sources. So the basics of it is like, how quickly can we SSH you into a G, into a GPU and whenever we would talk to users, they wanted A GPU. They wanted an A 100. And if you go to like any cloud [00:04:00] provisioning page, usually it's like three pages of forms or in the forms somewhere there's a dropdown.And in the dropdown there's some weird code that you know to translate to an A 100. And I remember just thinking like. Every time someone says they want an A 100, like the piece of text that they're telling me that they want is like, stuffed away in the corner. Yeah. And so we were like, what if the biggest piece of text was what the user's asking for?And so when you go to Brev, it's just big GPU chips with the type that you want withswyx: beautiful animations that you worked on pre, like pre you can, like, now you can just prompt it. But back in the day. Yeah. Yeah. Those were handcraft, handcrafted artisanal code.Nader: Yeah. I was actually really proud of that because, uh, it was an, i I made it in Figma.Yeah. And then I found, I was like really struggling to figure out how to turn it from like Figma to react. So what it actually is, is just an SVG and I, I have all the styles and so when you change the chip, whether it's like active or not it changes the SVG code and that somehow like renders like, looks like it's animating, but it, we just had the transition slow, but it's just like the, a JavaScript function to change the like underlying SVG.Yeah. And that was how I ended up like figuring out how to move it from from Figma. But yeah, that's Art Artisan. [00:05:00]Kyle: Speaking of marketing stunts though, he actually used those SVGs. Or kind of use those SVGs to make these cards.Nader: Oh yeah. LikeKyle: a GPU gift card Yes. That he handed out everywhere. That was actually my first impression of thatNader: one.Yeah,swyx: yeah, yeah.Nader: Yeah.swyx: I think I still have one of them.Nader: They look great.Kyle: Yeah.Nader: I have a ton of them still actually in our garage, which just, they don't have labels. We should honestly like bring, bring them back. But, um, I found this old printing press here, actually just around the corner on Ven ness. And it's a third generation San Francisco shop.And so I come in an excited startup founder trying to like, and they just have this crazy old machinery and I'm in awe. ‘cause the the whole building is so physical. Like you're seeing these machines, they have like pedals to like move these saws and whatever. I don't know what this machinery is, but I saw all three generations.Like there's like the grandpa, the father and the son, and the son was like, around my age. Well,swyx: it's like a holy, holy trinity.Nader: It's funny because we, so I just took the same SVG and we just like printed it and it's foil printing, so they make a a, a mold. That's like an inverse of like the A 100 and then they put the foil on it [00:06:00] and then they press it into the paper.And I remember once we got them, he was like, Hey, don't forget about us. You know, I guess like early Apple and Cisco's first business cards were all made there. And so he was like, yeah, we, we get like the startup businesses but then as they mature, they kind of go somewhere else. And so I actually, I think we were talking with marketing about like using them for some, we should go back and make some cards.swyx: Yeah, yeah, yeah. You know, I remember, you know, as a very, very small breadth investor, I was like, why are we spending time like, doing these like stunts for GPUs? Like, you know, I think like as a, you know, typical like cloud hard hardware person, you go into an AWS you pick like T five X xl, whatever, and it's just like from a list and you look at the specs like, why animate this GP?And, and I, I do think like it just shows the level of care that goes throughout birth and Yeah. And now, and also the, and,Nader: and Nvidia. I think that's what the, the thing that struck me most when we first came in was like the amount of passion that everyone has. Like, I think, um, you know, you talk to, you talk to Kyle, you talk to, like, every VP that I've met at Nvidia goes so close to the metal.Like, I remember it was almost a year ago, and like my VP asked me, he's like, Hey, [00:07:00] what's cursor? And like, are you using it? And if so, why? Surprised at this, and he downloaded Cursor and he was asking me to help him like, use it. And I thought that was, uh, or like, just show him what he, you know, why we were using it.And so, the amount of care that I think everyone has and the passion, appreciate, passion and appreciation for the moment. Right. This is a very unique time. So it's really cool to see everyone really like, uh, appreciate that.swyx: Yeah.Acquisition and DevEx Shiftswyx: One thing I wanted to do before we move over to sort of like research topics and, uh, the, the stuff that Kyle's working on is just tell the story of the acquisition, right?Like, not many people have been, been through an acquisition with Nvidia. What's it like? Uh, what, yeah, just anything you'd like to say.Nader: It's a crazy experience. I think, uh, you know, we were the thing that was the most exciting for us was. Our goal was just to make it easier for developers.We wanted to find access to GPUs, make it easier to do that. And then all, oh, actually your question about launchable. So launchable was just make one click exper, like one click deploys for any software on top of the GPU. Mm-hmm. And so what we really liked about Nvidia was that it felt like we just got a lot more resources to do all of that.I think, uh, you [00:08:00] know, NVIDIA's goal is to make things as easy for developers as possible. So there was a really nice like synergy there. I think that, you know, when it comes to like an acquisition, I think the amount that the soul of the products align, I think is gonna be. Is going speak to the success of the acquisition.Yeah. And so it in many ways feels like we're home. This is a really great outcome for us. Like we you know, I love brev.nvidia.com. Like you should, you should use it's, it's theKyle: front page for GPUs.Nader: Yeah. Yeah. If you want GP views,Kyle: you go there, getswyx: it there, and it's like internally is growing very quickly.I, I don't remember You said some stats there.Nader: Yeah, yeah, yeah. It's, uh, I, I wish I had the exact numbers, but like internally, externally, it's been growing really quickly. We've been working with a bunch of partners with a bunch of different customers and ISVs, if you have a solution that you want someone that runs on the GPU and you want people to use it quickly, we can bundle it up, uh, in a launchable and make it a one click run.If you're doing things and you want just like a sandbox or something to run on, right. Like open claw. Huge moment. Super exciting. Our, uh, and we'll talk into it more, but. You know, internally, people wanna run this, and you, we know we have to be really careful from the security implications. Do we let this run on the corporate network?Security's guidance was, Hey, [00:09:00] run this on breath, it's in, you know, it's, it's, it's a vm, it's sitting in the cloud, it's off the corporate network. It's isolated. And so that's been our stance internally and externally about how to even run something like open call while we figure out how to run these things securely.But yeah,swyx: I think there's also like, you almost like we're the right team at the right time when Nvidia is starting to invest a lot more in developer experience or whatever you call it. Yeah. Uh, UX or I don't know what you call it, like software. Like obviously NVIDIA is always invested in software, but like, there's like, this is like a different audience.Yeah. It's aNader: widerKyle: developer base.swyx: Yeah. Right.Nader: Yeah. Yeah. You know, it's funny, it's like, it's not, uh,swyx: so like, what, what is it called internally? What, what is this that people should be aware that is going on there?Nader: Uh, what, like developer experienceswyx: or, yeah, yeah. Is it's called just developer experience or is there like a broader strategy hereNader: in Nvidia?Um, Nvidia always wants to make a good developer experience. The thing is and a lot of the technology is just really complicated. Like, it's not, it's uh, you know, I think, um. The thing that's been really growing or the AI's growing is having a huge moment, not [00:10:00] because like, let's say data scientists in 2018, were quiet then and are much louder now.The pie is com, right? There's a whole bunch of new audiences. My mom's wondering what she's doing. My sister's learned, like taught herself how to code. Like the, um, you know, I, I actually think just generally AI's a big equalizer and you're seeing a more like technologically literate society, I guess.Like everyone's, everyone's learning how to code. Uh, there isn't really an excuse for that. And so building a good UX means that you really understand who your end user is. And when your end user becomes such a wide, uh, variety of people, then you have to almost like reinvent the practice, right? Yeah. You haveKyle: to, and actually build more developer ux, right?Because the, there are tiers of developer base that were added. You know, the, the hackers that are building on top of open claw, right? For example, have never used gpu. They don't know what kuda is. They, they, they just want to run something.Nader: Yeah.Kyle: You need new UX that is not just. Hey, you know, how do you program something in Cuda and run it?And then, and then we built, you know, like when Deep Learning was getting big, we built, we built Torch and, and, but so recently the amount of like [00:11:00] layers that are added to that developer stack has just exploded because AI has become ubiquitous. Everyone's using it in different ways. Yeah. It'sNader: moving fast in every direction.Vertical, horizontal.Vibhu: Yeah. You guys, you even take it down to hardware, like the DGX Spark, you know, it's, it's basically the same system as just throwing it up on big GPU cluster.Nader: Yeah, yeah, yeah. It's amazing. Blackwell.swyx: Yeah. Uh, we saw the preview at the last year's GTC and that was one of the better performing, uh, videos so far, and video coverage so far.Awesome. This will beat it. Um,Nader: that wasswyx: actually, we have fingersNader: crossed. Yeah.DGX Spark and Remote AccessNader: Even when Grace Blackwell or when, um, uh, DGX Spark was first coming out getting to be involved in that from the beginning of the developer experience. And it just comes back to what youswyx: were involved.Nader: Yeah. St. St.swyx: Mars.Nader: Yeah. Yeah. I mean from, it was just like, I, I got an email, we just got thrown into the loop and suddenly yeah, I, it was actually really funny ‘cause I'm still pretty fresh from the acquisition and I'm, I'm getting an email from a bunch of the engineering VPs about like, the new hardware, GPU chip, like we're, or not chip, but just GPU system that we're putting out.And I'm like, okay, cool. Matters. Now involved with this for the ux, I'm like. What am I gonna do [00:12:00] here? So, I remember the first meeting, I was just like kind of quiet as I was hearing engineering VPs talk about what this box could be, what it could do, how we should use it. And I remember, uh, one of the first ideas that people were idea was like, oh, the first thing that it was like, I think a quote was like, the first thing someone's gonna wanna do with this is get two of them and run a Kubernetes cluster on top of them.And I was like, oh, I think I know why I'm here. I was like, the first thing we're doing is easy. SSH into the machine. And then, and you know, just kind of like scoping it down of like, once you can do that every, you, like the person who wants to run a Kubernetes cluster onto Sparks has a higher propensity for pain, then, then you know someone who buys it and wants to run open Claw right now, right?If you can make sure that that's as effortless as possible, then the rest becomes easy. So there's a tool called Nvidia Sync. It just makes the SSH connection really simple. So, you know, if you think about it like. If you have a Mac, uh, or a PC or whatever, if you have a laptop and you buy this GPU and you want to use it, you should be able to use it like it's A-A-G-P-U in the cloud, right?Um, but there's all this friction of like, how do you actually get into that? That's part of [00:13:00] Revs value proposition is just, you know, there's a CLI that wraps SSH and makes it simple. And so our goal is just get you into that machine really easily. And one thing we just launched at CES, it's in, it's still in like early access.We're ironing out some kinks, but it should be ready by GTC. You can register your spark on Brev. And so now if youswyx: like remote managed yeah, local hardware. Single pane of glass. Yeah. Yeah. Because Brev can already manage other clouds anyway, right?Vibhu: Yeah, yeah. And you use the spark on Brev as well, right?Nader: Yeah. But yeah, exactly. So, so you, you, so you, you set it up at home you can run the command on it, and then it gets it's essentially it'll appear in your Brev account, and then you can take your laptop to a Starbucks or to a cafe, and you'll continue to use your, you can continue use your spark just like any other cloud node on Brev.Yeah. Yeah. And it's just like a pre-provisioned centerswyx: in yourNader: home. Yeah, exactly.swyx: Yeah. Yeah.Vibhu: Tiny little data center.Nader: Tiny little, the size ofVibhu: your phone.SOL Culture and Dynamo Setupswyx: One more thing before we move on to Kyle. Just have so many Jensen stories and I just love, love mining Jensen stories. Uh, my favorite so far is SOL. Uh, what is, yeah, what is S-O-L-S-O-LNader: is actually, i, I think [00:14:00] of all the lessons I've learned, that one's definitely my favorite.Kyle: It'll always stick with you.Nader: Yeah. Yeah. I, you know, in your startup, everything's existential, right? Like we've, we've run out of money. We were like, on the risk of, of losing payroll, we've had to contract our team because we l ran outta money. And so like, um, because of that you're really always forcing yourself to I to like understand the root cause of everything.If you get a date, if you get a timeline, you know exactly why that date or timeline is there. You're, you're pushing every boundary and like, you're not just say, you're not just accepting like a, a no. Just because. And so as you start to introduce more layers, as you start to become a much larger organization, SOL is is essentially like what is the physics, right?The speed of light moves at a certain speed. So if flight's moving some slower, then you know something's in the way. So before trying to like layer reality back in of like, why can't this be delivered at some date? Let's just understand the physics. What is the theoretical limit to like, uh, how fast this can go?And then start to tell me why. ‘cause otherwise people will start telling you why something can't be done. But actually I think any great leader's goal is just to create urgency. Yeah. [00:15:00] There's an infiniteKyle: create compelling events, right?Nader: Yeah.Kyle: Yeah. So l is a term video is used to instigate a compelling event.You say this is done. How do we get there? What is the minimum? As much as necessary, as little as possible thing that it takes for us to get exactly here and. It helps you just break through a bunch of noise.swyx: Yeah.Kyle: Instantly.swyx: One thing I'm unclear about is, can only Jensen use the SOL card? Like, oh, no, no, no.Not everyone get the b******t out because obviously it's Jensen, but like, can someone else be like, no, likeKyle: frontline engineers use it.Nader: Yeah. Every, I think it's not so much about like, get the b******t out. It's like, it's like, give me the root understanding, right? Like, if you tell me something takes three weeks, it like, well, what's the first principles?Yeah, the first principles. It's like, what's the, what? Like why is it three weeks? What is the actual yeah. What's the actual limit of why this is gonna take three weeks? If you're gonna, if you, if let's say you wanted to buy a new computer and someone told you it's gonna be here in five days, what's the SOL?Well, like the SOL is like, I could walk into a Best Buy and pick it up for you. Right? So then anything that's like beyond that is, and is that practical? Is that how we're gonna, you know, let's say give everyone in the [00:16:00] company a laptop, like obviously not. So then like that's the SOL and then it's like, okay, well if we have to get more than 10, suddenly there might be some, right?And so now we can kind of piece the reality back.swyx: So, so this is the. Paul Graham do things that don't scale. Yeah. And this is also the, what people would now call behi agency. Yeah.Kyle: It's actually really interesting because there's a, there's a second hardware angle to SOL that like doesn't come up for all the org sol is used like culturally at aswyx: media for everything.I'm also mining for like, I think that can be annoying sometimes. And like someone keeps going IOO you and you're like, guys, like we have to be stable. We have to, we to f*****g plan. Yeah.Kyle: It's an interesting balance.Nader: Yeah. I encounter that with like, actually just with, with Alec, right? ‘cause we, we have a new conference so we need to launch, we have, we have goals of what we wanna launch by, uh, by the conference and like, yeah.At the end of the day, where isswyx: this GTC?Nader: Um, well this is like, so we, I mean we did it for CES, we did for GT CDC before that we're doing it for GTC San Jose. So I mean, like every, you know, we have a new moment. Um, and we want to launch something. Yeah. And we want to do so at SOL and that does mean that some, there's some level of prioritization that needs [00:17:00] to happen.And so it, it is difficult, right? I think, um, you have to be careful with what you're pushing. You know, stability is important and that should be factored into S-O-L-S-O-L isn't just like, build everything and let it break, you know, that, that's part of the conversation. So as you're laying, layering in all the details, one of them might be, Hey, we could build this, but then it's not gonna be stable for X, y, z reasons.And so that was like, one of our conversations for CES was, you know, hey, like we, we can get this into early access registering your spark with brev. But there are a lot of things that we need to do in order to feel really comfortable from a security perspective, right? There's a lot of networking involved before we deliver that to users.So it's like, okay. Let's get this to a point where we can at least let people experiment with it. We had it in a booth, we had it in Jensen's keynote, and then let's go iron out all the networking kinks. And that's not easy. And so, uh, that can come later. And so that was the way that we layered that back in.Yeah. ButKyle: It's not really about saying like, you don't have to do the, the maintenance or operational work. It's more about saying, you know, it's kind of like [00:18:00] highlights how progress is incremental, right? Like, what is the minimum thing that we can get to. And then there's SOL for like every component after that.But there's the SOL to get you, get you to the, the starting line. And that, that's usually how it's asked. Yeah. On the other side, you know, like SOL came out of like hardware at Nvidia. Right. So SOL is like literally if we ran the accelerator or the GPU with like at basically full speed with like no other constraints, like how FAST would be able to make a program go.swyx: Yeah. Yeah. Right.Kyle: Soswyx: in, in training that like, you know, then you work back to like some percentage of like MFU for example.Kyle: Yeah, that's a, that's a great example. So like, there's an, there's an S-O-L-M-F-U, and then there's like, you know, what's practically achievable.swyx: Cool. Should we move on to sort of, uh, Kyle's side?Uh, Kyle, you're coming more from the data science world. And, uh, I, I mean I always, whenever, whenever I meet someone who's done working in tabular stuff, graph neural networks, time series, these are basically when I go to new reps, I go to ICML, I walk the back halls. There's always like a small group of graph people.Yes. Absolute small group of tabular people. [00:19:00] And like, there's no one there. And like, it's very like, you know what I mean? Like, yeah, no, like it's, it's important interesting work if you care about solving the problems that they solve.Kyle: Yeah.swyx: But everyone else is just LMS all the time.Kyle: Yeah. I mean it's like, it's like the black hole, right?Has the event horizon reached this yet in nerves? Um,swyx: but like, you know, those are, those are transformers too. Yeah. And, and those are also like interesting things. Anyway, uh, I just wanted to spend a little bit of time on, on those, that background before we go into Dynamo, uh, proper.Kyle: Yeah, sure. I took a different path to Nvidia than that, or I joined six years ago, seven, if you count, when I was an intern.So I joined Nvidia, like right outta college. And the first thing I jumped into was not what I'd done in, during internship, which was like, you know, like some stuff for autonomous vehicles, like heavyweight object detection. I jumped into like, you know, something, I'm like, recommenders, this is popular. Andswyx: yeah, he did RexiKyle: as well.Yeah, Rexi. Yeah. I mean that, that was the taboo data at the time, right? You have tables of like, audience qualities and item qualities, and you're trying to figure out like which member of [00:20:00] the audience matches which item or, or more practically which item matches which member of the audience. And at the time, really it was like we were trying to enable.Uh, recommender, which had historically been like a little bit of a CP based workflow into something that like, ran really well in GPUs. And it's since been done. Like there are a bunch of libraries for Axis that run on GPUs. Uh, the common models like Deeplearning recommendation model, which came outta meta and the wide and deep model, which was used or was released by Google were very accelerated by GPUs using, you know, the fast HBM on the chips, especially to do, you know, vector lookups.But it was very interesting at the time and super, super relevant because like we were starting to get like. This explosion of feeds and things that required rec recommenders to just actively be on all the time. And sort of transitioned that a little bit towards graph neural networks when I discovered them because I was like, okay, you can actually use graphical neural networks to represent like, relationships between people, items, concepts, and that, that interested me.So I jumped into that at [00:21:00] Nvidia and, and got really involved for like two-ish years.swyx: Yeah. Uh, and something I learned from Brian Zaro Yeah. Is that you can just kind of choose your own path in Nvidia.Kyle: Oh my God. Yeah.swyx: Which is not a normal big Corp thing. Yeah. Like you, you have a lane, you stay in your lane.Nader: I think probably the reason why I enjoy being in a, a big company, the mission is the boss probably from a startup guy. Yeah. The missionswyx: is the boss.Nader: Yeah. Uh, it feels like a big game of pickup basketball. Like, you know, if you play one, if you wanna play basketball, you just go up to the court and you're like, Hey look, we're gonna play this game and we need three.Yeah. And you just like find your three. That's honestly for every new initiative that's what it feels like. Yeah.Vibhu: It also like shows, right? Like Nvidia. Just releasing state-of-the-art stuff in every domain. Yeah. Like, okay, you expect foundation models with Nemo tron voice just randomly parakeet.Call parakeet just comes out another one, uh, voice. TheKyle: video voice team has always been producing.Vibhu: Yeah. There's always just every other domain of paper that comes out, dataset that comes out. It's like, I mean, it also stems back to what Nvidia has to do, right? You have to make chips years before they're actually produced.Right? So you need to know, you need to really [00:22:00] focus. TheKyle: design process starts likeVibhu: exactlyKyle: three to five years before the chip gets to the market.Vibhu: Yeah. I, I'm curious more about what that's like, right? So like, you have specialist teams. Is it just like, you know, people find an interest, you go in, you go deep on whatever, and that kind of feeds back into, you know, okay, we, we expect predictions.Like the internals at Nvidia must be crazy. Right? You know? Yeah. Yeah. You know, you, you must. Not even without selling to people, you have your own predictions of where things are going. Yeah. And they're very based, very grounded. Right?Kyle: Yeah. It, it, it's really interesting. So there's like two things that I think that Amed does, which are quite interesting.Uh, one is like, we really index into passion. There's a big. Sort of organizational top sound push to like ensure that people are working on the things that they're passionate about. So if someone proposes something that's interesting, many times they can just email someone like way up the chain that they would find this relevant and say like, Hey, can I go work on this?Nader: It's actually like I worked at a, a big company for a couple years before, uh, starting on my startup journey and like, it felt very weird if you were to like email out of chain, if that makes [00:23:00] sense. Yeah. The emails at Nvidia are like mosh pitsswyx: shoot,Nader: and it's just like 60 people, just whatever. And like they're, there's this,swyx: they got messy like, reply all you,Nader: oh, it's in, it's insane.It's insane. They justKyle: help. You know, Maxim,Nader: the context. But, but that's actually like, I've actually, so this is a weird thing where I used to be like, why would we send emails? We have Slack. I am the entire, I'm the exact opposite. I feel so bad for anyone who's like messaging me on Slack ‘cause I'm so unresponsive.swyx: Your emailNader: Maxi, email Maxim. I'm email maxing Now email is a different, email is perfect because man, we can't work together. I'm email is great, right? Because important threads get bumped back up, right? Yeah, yeah. Um, and so Slack doesn't do that. So I just have like this casino going off on the right or on the left and like, I don't know which thread was from where or what, but like the threads get And then also just like the subject, so you can have like working threads.I think what's difficult is like when you're small, if you're just not 40,000 people I think Slack will work fine, but there's, I don't know what the inflection point is. There is gonna be a point where that becomes really messy and you'll actually prefer having email. ‘cause you can have working threads.You can cc more than nine people in a thread.Kyle: You can fork stuff.Nader: You can [00:24:00] fork stuff, which is super nice and just like y Yeah. And so, but that is part of where you can propose a plan. You can also just. Start, honestly, momentum's the only authority, right? So like, if you can just start, start to make a little bit of progress and show someone something, and then they can try it.That's, I think what's been, you know, I think the most effective way to push anything for forward. And that's both at Nvidia and I think just generally.Kyle: Yeah, there's, there's the other concept that like is explored a lot at Nvidia, which is this idea of a zero billion dollar business. Like market creation is a big thing at Nvidia.Like,swyx: oh, you want to go and start a zero billion dollar business?Kyle: Jensen says, we are completely happy investing in zero billion dollar markets. We don't care if this creates revenue. It's important for us to know about this market. We think it will be important in the future. It can be zero billion dollars for a while.I'm probably minging as words here for, but like, you know, like, I'll give an example. NVIDIA's been working on autonomous driving for a a long time,swyx: like an Nvidia car.Kyle: No, they, they'veVibhu: used the Mercedes, right? They're around the HQ and I think it finally just got licensed out. Now they're starting to be used quite a [00:25:00] bit.For 10 years you've been seeing Mercedes with Nvidia logos driving.Kyle: If you're in like the South San Santa Clara, it's, it's actually from South. Yeah. So, um. Zero billion dollar markets are, are a thing like, you know, Jensen,swyx: I mean, okay, look, cars are not a zero billion dollar market. But yeah, that's a bad example.Nader: I think, I think he's, he's messaging, uh, zero today, but, or even like internally, right? Like, like it's like, uh, an org doesn't have to ruthlessly find revenue very quickly to justify their existence. Right. Like a lot of the important research, a lot of the important technology being developed that, that's kind ofKyle: where research, research is very ide ideologically free at Nvidia.Yeah. Like they can pursue things that they wereswyx: Were you research officially?Kyle: I was never in research. Officially. I was always in engineering. Yeah. We in, I'm in an org called Deep Warning Algorithms, which is basically just how do we make things that are relevant to deep warning go fast.swyx: That sounds freaking cool.Vibhu: And I think a lot of that is underappreciated, right? Like time series. This week Google put out time. FF paper. Yeah. A new time series, paper res. Uh, Symantec, ID [00:26:00] started applying Transformers LMS to Yes. Rec system. Yes. And when you think the scale of companies deploying these right. Amazon recommendations, Google web search, it's like, it's huge scale andKyle: Yeah.Vibhu: You want fast?Kyle: Yeah. Yeah. Yeah. Actually it's, it, I, there's a fun moment that brought me like full circle. Like, uh, Amazon Ads recently gave a talk where they talked about using Dynamo for generative recommendation, which was like super, like weirdly cathartic for me. I'm like, oh my God. I've, I've supplanted what I was working on.Like, I, you're using LMS now to do what I was doing five years ago.swyx: Yeah. Amazing. And let's go right into Dynamo. Uh, maybe introduce Yeah, sure. To the top down and Yeah.Kyle: I think at this point a lot of people are familiar with the term of inference. Like funnily enough, like I went from, you know, inference being like a really niche topic to being something that's like discussed on like normal people's Twitter feeds.It's,Nader: it's on billboardsKyle: here now. Yeah. Very, very strange. Driving, driving, seeing just an inference ad on 1 0 1 inference at scale is becoming a lot more important. Uh, we have these moments like, you know, open claw where you have these [00:27:00] agents that take lots and lots of tokens, but produce, incredible results.There are many different aspects of test time scaling so that, you know, you can use more inference to generate a better result than if you were to use like a short amount of inference. There's reasoning, there's quiring, there's, adding agency to the model, allowing it to call tools and use skills.Dyno sort came about at Nvidia. Because myself and a couple others were, were sort of talking about the, these concepts that like, you know, you have inference engines like VLMS, shelan, tenor, TLM and they have like one single copy. They, they, they sort of think about like things as like one single copy, like one replica, right?Why Scale Out WinsKyle: Like one version of the model. But when you're actually serving things at scale, you can't just scale up that replica because you end up with like performance problems. There's a scaling limit to scaling up replicas. So you actually have to scale out to use a, maybe some Kubernetes type terminology.We kind of realized that there was like. A lot of potential optimization that we could do in scaling out and building systems for data [00:28:00] center scale inference. So Dynamo is this data center scale inference engine that sits on top of the frameworks like VLM Shilling and 10 T lm and just makes things go faster because you can leverage the economy of scale.The fact that you have KV cash, which we can define a little bit later, uh, in all these machines that is like unique and you wanna figure out like the ways to maximize your cash hits or you want to employ new techniques in inference like disaggregation, which Dynamo had introduced to the world in, in, in March, not introduced, it was a academic talk, but beforehand.But we are, you know, one of the first frameworks to start, supporting it. And we wanna like, sort of combine all these techniques into sort of a modular framework that allows you to. Accelerate your inference at scale.Nader: By the way, Kyle and I became friends on my first date, Nvidia, and I always loved, ‘cause like he always teaches meswyx: new things.Yeah. By the way, this is why I wanted to put two of you together. I was like, yeah, this is, this is gonna beKyle: good. It's very, it's very different, you know, like we've, we, we've, we've talked to each other a bunch [00:29:00] actually, you asked like, why, why can't we scale up?Nader: Yeah.Scale Up Limits ExplainedNader: model, you said model replicas.Kyle: Yeah. So you, so scale up means assigning moreswyx: heavier?Kyle: Yeah, heavier. Like making things heavier. Yeah, adding more GPUs. Adding more CPUs. Scale out is just like having a barrier saying, I'm gonna duplicate my representation of the model or a representation of this microservice or something, and I'm gonna like, replicate it Many times.Handle, load. And the reason that you can't scale, scale up, uh, past some points is like, you know, there, there, there are sort of hardware bounds and algorithmic bounds on, on that type of scaling. So I'll give you a good example that's like very trivial. Let's say you're on an H 100. The Maxim ENV link domain for H 100, for most Ds H one hundreds is heus, right?So if you scaled up past that, you're gonna have to figure out ways to handle the fact that now for the GPUs to communicate, you have to do it over Infin band, which is still very fast, but is not as fast as ENV link.swyx: Is it like one order of magnitude, like hundreds or,Kyle: it's about an order of magnitude?Yeah. Okay. Um, soswyx: not terrible.Kyle: [00:30:00] Yeah. I, I need to, I need to remember the, the data sheet here, like, I think it's like about 500 gigabytes. Uh, a second unidirectional for ENV link, and about 50 gigabytes a second unidirectional for Infin Band. I, it, it depends on the, the generation.swyx: I just wanna set this up for people who are not familiar with these kinds of like layers and the trash speedVibhu: and all that.Of course.From Laptop to Multi NodeVibhu: Also, maybe even just going like a few steps back before that, like most people are very familiar with. You see a, you know, you can use on your laptop, whatever these steel viol, lm you can just run inference there. All, there's all, you can, youcan run it on thatVibhu: laptop. You can run on laptop.Then you get to, okay, uh, models got pretty big, right? JLM five, they doubled the size, so mm-hmm. Uh, what do you do when you have to go from, okay, I can get 128 gigs of memory. I can run it on a spark. Then you have to go multi GPU. Yeah. Okay. Multi GPU, there's some support there. Now, if I'm a company and I don't have like.I'm not hiring the best researchers for this. Right. But I need to go [00:31:00] multi-node, right? I have a lot of servers. Okay, now there's efficiency problems, right? You can have multiple eight H 100 nodes, but, you know, is that as a, like, how do you do that efficiently?Kyle: Yeah. How do you like represent them? How do you choose how to represent the model?Yeah, exactly right. That's a, that's like a hard question. Everyone asks, how do you size oh, I wanna run GLM five, which just came out new model. There have been like four of them in the past week, by the way, like a bunch of new models.swyx: You know why? Right? Deep seek.Kyle: No comment. Oh. Yeah, but Ggl, LM five, right?We, we have this, new model. It's, it's like a large size, and you have to figure out how to both scale up and scale out, right? Because you have to find the right representation that you care about. Everyone does this differently. Let's be very clear. Everyone figures this out in their own path.Nader: I feel like a lot of AI or ML even is like, is like this. I think people think, you know, I, I was, there was some tweet a few months ago that was like, why hasn't fine tuning as a service taken off? You know, that might be me. It might have been you. Yeah. But people want it to be such an easy recipe to follow.But even like if you look at an ML model and specificKyle: to you Yeah,Nader: yeah.Kyle: And the [00:32:00] model,Nader: the situation, and there's just so much tinkering, right? Like when you see a model that has however many experts in the ME model, it's like, why that many experts? I don't, they, you know, they tried a bunch of things and that one seemed to do better.I think when it comes to how you're serving inference, you know, you have a bunch of decisions to make and there you can always argue that you can take something and make it more optimal. But I think it's this internal calibration and appetite for continued calibration.Vibhu: Yeah. And that doesn't mean like, you know, people aren't taking a shot at this, like tinker from thinking machines, you know?Yeah. RL as a service. Yeah, totally. It's, it also gets even harder when you try to do big model training, right? We're not the best at training Moes, uh, when they're pre-trained. Like we saw this with LAMA three, right? They're trained in such a sparse way that meta knows there's gonna be a bunch of inference done on these, right?They'll open source it, but it's very trained for what meta infrastructure wants, right? They wanna, they wanna inference it a lot. Now the question to basically think about is, okay, say you wanna serve a chat application, a coding copilot, right? You're doing a layer of rl, you're serving a model for X amount of people.Is it a chat model, a coding model? Dynamo, you know, back to that,Kyle: it's [00:33:00] like, yeah, sorry. So you we, we sort of like jumped off of, you know, jumped, uh, on that topic. Everyone has like, their own, own journey.Cost Quality Latency TradeoffsKyle: And I, I like to think of it as defined by like, what is the model you need? What is the accuracy you need?Actually I talked to NA about this earlier. There's three axes you care about. What is the quality that you're able to produce? So like, are you accurate enough or can you complete the task with enough, performance, high enough performance. Yeah, yeah. Uh, there's cost. Can you serve the model or serve your workflow?Because it's not just the model anymore, it's the workflow. It's the multi turn with an agent cheaply enough. And then can you serve it fast enough? And we're seeing all three of these, like, play out, like we saw, we saw new models from OpenAI that you know, are faster. You have like these new fast versions of models.You can change the amount of thinking to change the amount of quality, right? Produce more tokens, but at a higher cost in a, in a higher latency. And really like when you start this journey of like trying to figure out how you wanna host a model, you, you, you think about three things. What is the model I need to serve?How many times do I need to call it? What is the input sequence link was [00:34:00] the, what does the workflow look like on top of it? What is the SLA, what is the latency SLA that I need to achieve? Because there's usually some, this is usually like a constant, you, you know, the SLA that you need to hit and then like you try and find the lowest cost version that hits all of these constraints.Usually, you know, you, you start with those things and you say you, you kind of do like a bit of experimentation across some common configurations. You change the tensor parallel size, which is a form of parallelismVibhu: I take, it goes even deeper first. Gotta think what model.Kyle: Yes, course,ofKyle: course. It's like, it's like a multi-step design process because as you said, you can, you can choose a smaller model and then do more test time scaling and it'll equate the quality of a larger model because you're doing the test time scaling or you're adding a harness or something.So yes, it, it goes way deeper than that. But from the performance perspective, like once you get to the model you need, you need to host, you look at that and you say, Hey. I have this model, I need to serve it at the speed. What is the right configuration for that?Nader: You guys see the recent, uh, there was a paper I just saw like a few days ago that, uh, if you run [00:35:00] the same prompt twice, you're getting like double Just try itagain.Nader: Yeah, exactly.Vibhu: And you get a lot. Yeah. But the, the key thing there is you give the context of the failed try, right? Yeah. So it takes a shot. And this has been like, you know, basic guidance for quite a while. Just try again. ‘cause you know, trying, just try again. Did you try again? All adviceNader: in life.Vibhu: Just, it's a paper from Google, if I'm not mistaken, right?Yeah,Vibhu: yeah. I think it, it's like a seven bas little short paper. Yeah. Yeah. The title's very cute. And it's just like, yeah, just try again. Give it ask context,Kyle: multi-shot. You just like, say like, hey, like, you know, like take, take a little bit more, take a little bit more information, try and fail. Fail.Vibhu: And that basic concept has gone pretty deep.There's like, um, self distillation, rl where you, you do self distillation, you do rl and you have past failure and you know, that gives some signal so people take, try it again. Not strong enough.swyx: Uh, for, for listeners, uh, who listen to here, uh, vivo actually, and I, and we run a second YouTube channel for our paper club where, oh, that's awesome.Vivo just covered this. Yeah. Awesome. Self desolation and all that's, that's why he, to speed [00:36:00] on it.Nader: I'll to check it out.swyx: Yeah. It, it's just a good practice, like everyone needs, like a paper club where like you just read papers together and the social pressure just kind of forces you to just,Nader: we, we,there'sNader: like a big inference.Kyle: ReadingNader: group at a video. I feel so bad every time. I I, he put it on like, on our, he shared it.swyx: One, one ofNader: your guys,swyx: uh, is, is big in that, I forget es han Yeah, yeah,Kyle: es Han's on my team. Actually. Funny. There's a, there's a, there's a employee transfer between us. Han worked for Nater at Brev, and now he, he's on my team.He wasNader: our head of ai. And then, yeah, once we got in, andswyx: because I'm always looking for like, okay, can, can I start at another podcast that only does that thing? Yeah. And, uh, Esan was like, I was trying to like nudge Esan into like, is there something here? I mean, I don't think there's, there's new infant techniques every day.So it's like, it's likeKyle: you would, you would actually be surprised, um, the amount of blog posts you see. And ifswyx: there's a period where it was like, Medusa hydra, what Eagle, like, youKyle: know, now we have new forms of decode, uh, we have new forms of specula, of decoding or new,swyx: what,Kyle: what are youVibhu: excited? And it's exciting when you guys put out something like Tron.‘cause I remember the paper on this Tron three, [00:37:00] uh, the amount of like post train, the on tokens that the GPU rich can just train on. And it, it was a hybrid state space model, right? Yeah.Kyle: It's co-designed for the hardware.Vibhu: Yeah, go design for the hardware. And one of the things was always, you know, the state space models don't scale as well when you do a conversion or whatever the performance.And you guys are like, no, just keep draining. And Nitron shows a lot of that. Yeah.Nader: Also, something cool about Nitron it was released in layers, if you will, very similar to Dynamo. It's, it's, it's essentially it was released as you can, the pre-training, post-training data sets are released. Yeah. The recipes on how to do it are released.The model itself is released. It's full model. You just benefit from us turning on the GPUs. But there are companies like, uh, ServiceNow took the dataset and they trained their own model and we were super excited and like, you know, celebrated that work.ZoomVibhu: different. Zoom is, zoom is CGI, I think, uh, you know, also just to add like a lot of models don't put out based models and if there's that, why is fine tuning not taken off?You know, you can do your own training. Yeah,Kyle: sure.Vibhu: You guys put out based model, I think you put out everything.Nader: I believe I know [00:38:00]swyx: about base. BasicallyVibhu: without baseswyx: basic can be cancelable.Vibhu: Yeah. Base can be cancelable.swyx: Yeah.Vibhu: Safety training.swyx: Did we get a full picture of dymo? I, I don't know if we, what,Nader: what I'd love is you, you mentioned the three axes like break it down of like, you know, what's prefilled decode and like what are the optimizations that we can get with Dynamo?Kyle: Yeah. That, that's, that's, that's a great point. So to summarize on that three axis problem, right, there are three things that determine whether or not something can be done with inference, cost, quality, latency, right? Dynamo is supposed to be there to provide you like the runtime that allows you to pull levers to, you know, mix it up and move around the parade of frontier or the preto surface that determines is this actually possible with inference And AI todayNader: gives you the knobs.Kyle: Yeah, exactly. It gives you the knobs.Disaggregation Prefill vs DecodeKyle: Uh, and one thing that like we, we use a lot in contemporary inference and is, you know, starting to like pick up from, you know, in, in general knowledge is this co concept of disaggregation. So historically. Models would be hosted with a single inference engine. And that inference engine [00:39:00] would ping pong between two phases.There's prefill where you're reading the sequence generating KV cache, which is basically just a set of vectors that represent the sequence. And then using that KV cache to generate new tokens, which is called Decode. And some brilliant researchers across multiple different papers essentially made the realization that if you separate these two phases, you actually gain some benefits.Those benefits are basically a you don't have to worry about step synchronous scheduling. So the way that an inference engine works is you do one step and then you finish it, and then you schedule, you start scheduling the next step there. It's not like fully asynchronous. And the problem with that is you would have, uh, essentially pre-fill and decode are, are actually very different in terms of both their resource requirements and their sometimes their runtime.So you would have like prefill that would like block decode steps because you, you'd still be pre-filing and you couldn't schedule because you know the step has to end. So you remove that scheduling issue and then you also allow you, or you yourself, to like [00:40:00] split the work into two different ki types of pools.So pre-fill typically, and, and this changes as, as model architecture changes. Pre-fill is, right now, compute bound most of the time with the sequence is sufficiently long. It's compute bound. On the decode side because you're doing a full Passover, all the weights and the entire sequence, every time you do a decode step and you're, you don't have the quadratic computation of KV cache, it's usually memory bound because you're retrieving a linear amount of memory and you're doing a linear amount of compute as opposed to prefill where you retrieve a linear amount of memory and then use a quadratic.You know,Nader: it's funny, someone exo Labs did a really cool demo where for the DGX Spark, which has a lot more compute, you can do the pre the compute hungry prefill on a DG X spark and then do the decode on a, on a Mac. Yeah. And soVibhu: that's faster.Nader: Yeah. Yeah.Kyle: So you could, you can do that. You can do machine strat stratification.Nader: Yeah.Kyle: And like with our future generation generations of hardware, we actually announced, like with Reuben, this [00:41:00] new accelerator that is prefilled specific. It's called Reuben, CPX. SoKubernetes Scaling with GroveNader: I have a question when you do the scale out. Yeah. Is scaling out easier with Dynamo? Because when you need a new node, you can dedicate it to either the Prefill or, uh, decode.Kyle: Yeah. So Dynamo actually has like a, a Kubernetes component in it called Grove that allows you to, to do this like crazy scaling specialization. It has like this hot, it's a representation that, I don't wanna go too deep into Kubernetes here, but there was a previous way that you would like launch multi-node work.Uh, it's called Leader Worker Set. It's in the Kubernetes standard, and Leader worker set is great. It served a lot of people super well for a long period of time. But one of the things that it's struggles with is representing a set of cases where you have a multi-node replica that has a pair, right?You know, prefill and decode, or it's not paired, but it has like a second stage that has a ratio that changes over time. And prefill and decode are like two different things as your workload changes, right? The amount of prefill you'll need to do may change. [00:42:00] The amount of decode that you, you'll need to do might change, right?Like, let's say you start getting like insanely long queries, right? That probably means that your prefill scales like harder because you're hitting these, this quadratic scaling growth.swyx: Yeah.And then for listeners, like prefill will be long input. Decode would be long output, for example, right?Kyle: Yeah. So like decode, decode scale. I mean, decode is funny because the amount of tokens that you produce scales with the output length, but the amount of work that you do per step scales with the amount of tokens in the context.swyx: Yes.Kyle: So both scales with the input and the output.swyx: That's true.Kyle: But on the pre-fold view code side, like if.Suddenly, like the amount of work you're doing on the decode side stays about the same or like scales a little bit, and then the prefilled side like jumps up a lot. You actually don't want that ratio to be the same. You want it to change over time. So Dynamo has a set of components that A, tell you how to scale.It tells you how many prefilled workers and decoded workers you, it thinks you should have, and also provides a scheduling API for Kubernetes that allows you to actually represent and affect this scheduling on, on, on your actual [00:43:00] hardware, on your compute infrastructure.Nader: Not gonna lie. I feel a little embarrassed for being proud of my SVG function earlier.swyx: No, itNader: wasreallyKyle: cute. I, Iswyx: likeNader: it's all,swyx: it's all engineering. It's all engineering. Um, that's where I'mKyle: technical.swyx: One thing I'm, I'm kind of just curious about with all with you see at a systems level, everything going on here. Mm-hmm. And we, you know, we're scaling it up in, in multi, in distributed systems.Context Length and Co Designswyx: Um, I think one thing that's like kind of, of the moment right now is people are asking, is there any SOL sort of upper bounds. In terms of like, let's call, just call it context length for one for of a better word, but you can break it down however you like.Nader: Yeah.swyx: I just think like, well, yeah, I mean, like clearly you can engage in hybrid architectures and throw in some state space models in there.All, all you want, but it looks, still looks very attention heavy.Kyle: Yes. Uh, yeah. Long context is attention heavy. I mean, we have these hybrid models, um,swyx: to take and most, most models like cap out at a million contexts and that's it. Yeah. Like for the last two years has been it.Kyle: Yeah. The model hardware context co-design thing that we're seeing these days is actually super [00:44:00] interesting.It's like my, my passion, like my secret side passion. We see models like Kimmy or G-P-T-O-S-S. I'm use these because I, I know specific things about these models. So Kimmy two comes out, right? And it's an interesting model. It's like, like a deep seek style architecture is MLA. It's basically deep seek, scaled like a little bit differently, um, and obviously trained differently as well.But they, they talked about, why they made the design choices for context. Kimmy has more experts, but fewer attention heads, and I believe a slightly smaller attention, uh, like dimension. But I need to remember, I need to check that. Uh, it doesn't matter. But they discussed this actually at length in a blog post on ji, which is like our pu which is like credit puswyx: Yeah.Kyle: Um, in, in China. Chinese red.swyx: Yeah.Kyle: It's, yeah. So it, it's, it's actually an incredible blog post. Uh, like all the mls people in, in, in that, I've seen that on GPU are like very brilliant, but they, they talk about like the creators of Kimi K two [00:45:00] actually like, talked about it on, on, on there in the blog post.And they say, we, we actually did an experiment, right? Attention scales with the number of heads, obviously. Like if you have 64 heads versus 32 heads, you do half the work of attention. You still scale quadratic, but you do half the work. And they made a, a very specific like. Sort of barter in their system, in their architecture, they basically said, Hey, what if we gave it more experts, so we're gonna use more memory capacity.But we keep the amount of activated experts the same. We increase the expert sparsity, so we have fewer experts act. The ratio to of experts activated to number of experts is smaller, and we decrease the number of attention heads.Vibhu: And kind of for context, what the, what we had been seeing was you make models sparser instead.So no one was really touching heads. You're just having, uh,Kyle: well, they, they did, they implicitly made it sparser.Vibhu: Yeah, yeah. For, for Kimmy. They did,Kyle: yes.Vibhu: They also made it sparser. But basically what we were seeing was people were at the level of, okay, there's a sparsity ratio. You want more total parameters, less active, and that's sparsity.[00:46:00]But what you see from papers, like, the labs like moonshot deep seek, they go to the level of, okay, outside of just number of experts, you can also change how many attention heads and less attention layers. More attention. Layers. Layers, yeah. Yes, yes. So, and that's all basically coming back to, just tied together is like hardware model, co-design, which isKyle: hardware model, co model, context, co-design.Vibhu: Yeah.Kyle: Right. Like if you were training a, a model that was like. Really, really short context, uh, or like really is good at super short context tasks. You may like design it in a way such that like you don't care about attention scaling because it hasn't hit that, like the turning point where like the quadratic curve takes over.Nader: How do you consider attention or context as a separate part of the co-design? Like I would imagine hardware or just how I would've thought of it is like hardware model. Co-design would be hardware model context co-designKyle: because the harness and the context that is produced by the harness is a part of the model.Once it's trained in,Vibhu: like even though towards the end you'll do long context, you're not changing architecture through I see. Training. Yeah.Kyle: I mean you can try.swyx: You're saying [00:47:00] everyone's training the harness into the model.Kyle: I would say to some degree, orswyx: there's co-design for harness. I know there's a small amount, but I feel like not everyone has like gone full send on this.Kyle: I think, I think I think it's important to internalize the harness that you think the model will be running. Running into the model.swyx: Yeah. Interesting. Okay. Bash is like the universal harness,Kyle: right? Like I'll, I'll give. An example here, right? I mean, or just like a, like a, it's easy proof, right? If you can train against a harness and you're using that harness for everything, wouldn't you just train with the harness to ensure that you get the best possible quality out of,swyx: Well, the, uh, I, I can provide a counter argument.Yeah, sure. Which is what you wanna provide a generally useful model for other people to plug into their harnesses, right? So if youKyle: Yeah. Harnesses can be open, open source, right?swyx: Yeah. So I mean, that's, that's effectively what's happening with Codex.Kyle: Yeah.swyx: And, but like you may want like a different search tool and then you may have to name it differently or,Nader: I don't know how much people have pushed on this, but can you.Train a model, would it be, have you have people compared training a model for the for the harness versus [00:48:00] like post training forswyx: I think it's the same thing. It's the same thing. It's okay. Just extra post training. INader: see.swyx: And so, I mean, cognition does this course, it does this where you, you just have to like, if your tool is slightly different, um, either force your tool to be like the tool that they train for.Hmm. Or undo their training for their tool and then Oh, that's re retrain. Yeah. It's, it's really annoying and like,Kyle: I would hope that eventually we hit like a certain level of generality with respect to training newswyx: tools. This is not a GI like, it's, this is a really stupid like. Learn my tool b***h.Like, I don't know if, I don't know if I can say that, but like, you know, um, I think what my point kind of is, is that there's, like, I look at slopes of the scaling laws and like, this slope is not working, man. We, we are at a million token con

All speakers are announced at AIE EU, schedule coming soon. Join us there or in Miami with the renowned organizers of React Miami! Singapore CFP also open!We've called this out a few times over in AINews, but the overwhelming consensus in the Valley is that “the IDE is Dead”. In November it was just a gut feeling, but now we actually have data: even at the canonical “VSCode Fork” company, people are officially using more agents than tab autocomplete (the first wave of AI coding):Cursor has launched cloud agents for a few months now, and this specific launch is around Computer Use, which has come a long way since we first talked with Anthropic about it in 2024, and which Jonas productized as Autotab:We also take the opportunity to do a live demo, talk about slash commands and subagents, and the future of continual learning and personalized coding models, something that Sam previously worked on at New Computer. (The fact that both of these folks are top tier CEOs of their own startups that have now joined the insane talent density gathering at Cursor should also not be overlooked).Full Episode on YouTube!please like and subscribe!Timestamps00:00 Agentic Code Experiments00:53 Why Cloud Agents Matter02:08 Testing First Pillar03:36 Video Reviews Second Pillar04:29 Remote Control Third Pillar06:17 Meta Demos and Bug Repro13:36 Slash Commands and MCPs18:19 From Tab to Team Workflow31:41 Minimal Web UI Philosophy32:40 Why No File Editor34:38 Full Stack Cursor Debate36:34 Model Choice and Auto Routing38:34 Parallel Agents and Best Of N41:41 Subagents and Context Management44:48 Grind Mode and Throughput Future01:00:24 Cloud Agent Onboarding and MemoryTranscriptEP 77 - CURSOR - Audio version[00:00:00]Agentic Code ExperimentsSamantha: This is another experiment that we ran last year and didn't decide to ship at that time, but may come back to LM Judge, but one that was also agentic and could write code. So it wasn't just picking but also taking the learnings from two models or and models that it was looking at and writing a new diff.And what we found was that there were strengths to using models from different model providers as the base level of this process. Basically you could get almost like a synergistic output that was better than having a very unified like bottom model tier.Jonas: We think that over the coming months, the big unlock is not going to be one person with a model getting more done, like the water flowing faster and we'll be making the pipe much wider and so paralyzing more, whether that's swarms of agents or parallel agents, both of those are things that contribute to getting much more done in the same amount of time.Why Cloud Agents Matterswyx: This week, one of the biggest launches that Cursor's ever done is cloud agents. I think you, you had [00:01:00] cloud agents before, but this was like, you give cursor a computer, right? Yeah. So it's just basically they bought auto tab and then they repackaged it. Is that what's going on, or,Jonas: that's a big part of it.Yeah. Cloud agents already ran in their own computers, but they were sort of site reading code. Yeah. And those computers were not, they were like blank VMs typically that were not set up for the Devrel X for whatever repo the agents working on. One of the things that we talk about is if you put yourself in the model shoes and you were seeing tokens stream by and all you could do was cite read code and spit out tokens and hope that you had done the right thing,swyx: no chanceJonas: I'd be so bad.Like you obviously you need to run the code. And so that I think also is probably not that contrarian of a take, but no one has done that yet. And so giving the model the tools to onboard itself and then use full computer use end-to-end pixels in coordinates out and have the cloud computer with different apps in it is the big unlock that we've seen internally in terms of use usage of this going from, oh, we use it for little copy changes [00:02:00] to no.We're really like driving new features with this kind of new type of entech workflow. Alright, let's see it. Cool.Live Demo TourJonas: So this is what it looks like in cursor.com/agents. So this is one I kicked off a while ago. So on the left hand side is the chat. Very classic sort of agentic thing. The big new thing here is that the agent will test its changes.So you can see here it worked for half an hour. That is because it not only took time to write the tokens of code, it also took time to test them end to end. So it started Devrel servers iterate when needed. And so that's one part of it is like model works for longer and doesn't come back with a, I tried some things pr, but a I tested at pr that's ready for your review.One of the other intuition pumps we use there is if a human gave you a PR asked you to review it and you hadn't, they hadn't tested it, you'd also be annoyed because you'd be like, only ask me for a review once it's actually ready. So that's what we've done withTesting Defaults and Controlsswyx: simple question I wanted to gather out front.Some prs are way smaller, [00:03:00] like just copy change. Does it always do the video or is it sometimes,Jonas: Sometimes.swyx: Okay. So what's the judgment?Jonas: The model does it? So we we do some default prompting with sort. What types of changes to test? There's a slash command that people can do called slash no test, where if you do that, the model will not test,swyx: but the default is test.Jonas: The default is to be calibrated. So we tell it don't test, very simple copy changes, but test like more complex things. And then users can also write their agents.md and specify like this type of, if you're editing this subpart of my mono repo, never tested ‘cause that won't work or whatever.Videos and Remote ControlJonas: So pillar one is the model actually testing Pillar two is the model coming back with a video of what it did.We have found that in this new world where agents can end-to-end, write much more code, reviewing the code is one of these new bottlenecks that crop up. And so reviewing a video is not a substitute for reviewing code, but it is an entry point that is much, much easier to start with than glancing at [00:04:00] some giant diff.And so typically you kick one off you, it's done you come back and the first thing that you would do is watch this video. So this is a, video of it. In this case I wanted a tool tip over this button. And so it went and showed me what that looks like in, in this video that I think here, it actually used a gallery.So sometimes it will build storybook type galleries where you can see like that component in action. And so that's pillar two is like these demo videos of what it built. And then pillar number three is I have full remote control access to this vm. So I can go heat in here. I can hover things, I can type, I have full control.And same thing for the terminal. I have full access. And so that is also really useful because sometimes the video is like all you need to see. And oftentimes by the way, the video's not perfect, the video will show you, is this worth either merging immediately or oftentimes is this worth iterating with to get it to that final stage where I am ready to merge in.So I can go through some other examples where the first video [00:05:00] wasn't perfect, but it gave me confidence that we were on the right track and two or three follow-ups later, it was good to go. And then I also have full access here where some things you just wanna play around with. You wanna get a feel for what is this and there's no substitute to a live preview.And the VNC kind of VM remote access gives you that.swyx: Amazing What, sorry? What is VN. AndJonas: just the remote desktop. Remote desktop. Yeah.swyx: Sam, any other details that you always wanna call out?Samantha: Yeah, for me the videos have been super helpful. I would say, especially in cases where a common problem for me with agents and cloud agents beforehand was almost like under specification in my requests where our plan mode and going really back and forth and getting detailed implementation spec is a way to reduce the risk of under specification, but then similar to how human communication breaks down over time, I feel like you have this risk where it's okay, when I pull down, go to the triple of pulling down and like running this branch locally, I'm gonna see that, like I said, this should be a toggle and you have a checkbox and like, why didn't you get that detail?And having the video up front just [00:06:00] has that makes that alignment like you're talking about a shared artifact with the agent. Very clear, which has been just super helpful for me.Jonas: I can quickly run through some other Yes. Examples.Meta Agents and More DemosJonas: So this is a very front end heavy one. So one question I wasswyx: gonna say, is this only for frontJonas: end?Exactly. One question you might have is this only for front end? So this is another example where the thing I wanted it to implement was a better error message for saving secrets. So the cloud agents support adding secrets, that's part of what it needs to access certain systems. Part of onboarding that is giving access.This is cloud is working onswyx: cloud agents. Yes.Jonas: So this is a fun thing isSamantha: it can get super meta. ItJonas: can get super meta, it can start its own cloud agents, it can talk to its own cloud agents. Sometimes it's hard to wrap your mind around that. We have disabled, it's cloud agents starting more cloud agents. So we currently disallow that.Someday you might. Someday we might. Someday we might. So this actually was mostly a backend change in terms of the error handling here, where if the [00:07:00] secret is far too large, it would oh, this is actually really cool. Wow. That's the Devrel tools. That's the Devrel tools. So if the secret is far too large, we.Allow secrets above a certain size. We have a size limit on them. And the error message there was really bad. It was just some generic failed to save message. So I was like, Hey, we wanted an error message. So first cool thing it did here, zero prompting on how to test this. Instead of typing out the, like a character 5,000 times to hit the limit, it opens Devrel tools, writes js, or to paste into the input 5,000 characters of the letter A and then hit save, closes the Devrel tools, hit save and gets this new gets the new error message.So that looks like the video actually cut off, but here you can see the, here you can see the screenshot of the of the error message. What, so that is like frontend backend end-to-end feature to, to get that,swyx: yeah.Jonas: Andswyx: And you just need a full vm, full computer run everything.Okay. Yeah.Jonas: Yeah. So we've had versions of this. This is one of the auto tab lessons where we started that in 2022. [00:08:00] No, in 2023. And at the time it was like browser use, DOM, like all these different things. And I think we ended up very sort of a GI pilled in the sense that just give the model pixels, give it a box, a brain in a box is what you want and you want to remove limitations around context and capabilities such that the bottleneck should be the intelligence.And given how smart models are today, that's a very far out bottleneck. And so giving it its full VM and having it be onboarded with Devrel X set up like a human would is just been for us internally a really big step change in capability.swyx: Yeah I would say, let's call it a year ago the models weren't even good enough to do any of this stuff.SoSamantha: even six months ago. Yeah.swyx: So yeah what people have told me is like round about Sonder four fire is when this started being good enough to just automate fully by pixel.Jonas: Yeah, I think it's always a question of when is good enough. I think we found in particular with Opus 4 5, 4, 6, and Codex five three, that those were additional step [00:09:00] changes in the autonomy grade capabilities of the model to just.Go off and figure out the details and come back when it's done.swyx: I wanna appreciate a couple details. One 10 Stack Router. I see it. Yeah. I'm a big fan. Do you know any, I have to name the 10 Stack.Jonas: No.swyx: This just a random lore. Some buddy Sue Tanner. My and then the other thing if you switch back to the video.Jonas: Yeah.swyx: I wanna shout out this thing. Probably Sam did it. I don't knowJonas: the chapters.swyx: What is this called? Yeah, this is called Chapters. Yeah. It's like a Vimeo thing. I don't know. But it's so nice the design details, like the, and obviously a company called Cursor has to have a beautiful cursorSamantha: and it isswyx: the cursor.Samantha: Cursor.swyx: You see it branded? It's the cursor. Cursor, yeah. Okay, cool. And then I was like, I complained to Evan. I was like, okay, but you guys branded everything but the wallpaper. And he was like, no, that's a cursor wallpaper. I was like, what?Samantha: Yeah. Rio picked the wallpaper, I think. Yeah. The video.That's probably Alexi and yeah, a few others on the team with the chapters on the video. Matthew Frederico. There's been a lot of teamwork on this. It's a huge effort.swyx: I just, I like design details.Samantha: Yeah.swyx: And and then when you download it adds like a little cursor. Kind of TikTok clip. [00:10:00] Yes. Yes.So it's to make it really obvious is from Cursor,Jonas: we did the TikTok branding at the end. This was actually in our launch video. Alexi demoed the cloud agent that built that feature. Which was funny because that was an instance where one of the things that's been a consequence of having these videos is we use best of event where you run head to head different models on the same prompt.We use that a lot more because one of the complications with doing that before was you'd run four models and they would come back with some giant diff, like 700 lines of code times four. It's what are you gonna do? You're gonna review all that's horrible. But if you come back with four 22nd videos, yeah, I'll watch four 22nd videos.And then even if none of them is perfect, you can figure out like, which one of those do you want to iterate with, to get it over the line. Yeah. And so that's really been really fun.Bug Repro WorkflowJonas: Here's another example. That's we found really cool, which is we've actually turned since into a slash command as well slash [00:11:00] repro, where for bugs in particular, the model of having full access to the to its own vm, it can first reproduce the bug, make a video of the bug reproducing, fix the bug, make a video of the bug being fixed, like doing the same pattern workflow with obviously the bug not reproducing.And that has been the single category that has gone from like these types of bugs, really hard to reproduce and pick two tons of time locally, even if you try a cloud agent on it. Are you confident it actually fixed it to when this happens? You'll merge it in 90 seconds or something like that.So this is an example where, let me see if this is the broken one or the, okay, this is the fixed one. Okay. So we had a bug on cursor.com/agents where if you would attach images where remove them. Then still submit your prompt. They would actually still get attached to the prompt. Okay. And so here you can see Cursor is using, its full desktop by the way.This is one of the cases where if you just do, browse [00:12:00] use type stuff, you'll have a bad time. ‘cause now it needs to upload files. Like it just uses its native file viewer to do that. And so you can see here it's uploading files. It's going to submit a prompt and then it will go and open up. So this is the meta, this is cursor agent, prompting cursor agent inside its own environment.And so you can see here bug, there's five images attached, whereas when it's submitted, it only had one image.swyx: I see. Yeah. But you gotta enable that if you're gonna use cur agent inside cur.Jonas: Exactly. And so here, this is then the after video where it went, it does the same thing. It attaches images, removes, some of them hit send.And you can see here, once this agent is up, only one of the images is left in the attachments. Yeah.swyx: Beautiful.Jonas: Okay. So easy merge.swyx: So yeah. When does it choose to do this? Because this is an extra step.Jonas: Yes. I think I've not done a great job yet of calibrating the model on when to reproduce these things.Yeah. Sometimes it will do it of its own accord. Yeah. We've been conservative where we try to have it only do it when it's [00:13:00] quite sure because it does add some amount of time to how long it takes it to work on it. But we also have added things like the slash repro command where you can just do, fix this bug slash repro and then it will know that it should first make you a video of it actually finding and making sure it can reproduce the bug.swyx: Yeah. Yeah. One sort of ML topic this ties into is reward hacking, where while you write test that you update only pass. So first write test, it shows me it fails, then make you test pass, which is a classic like red green.Jonas: Yep.swyx: LikeJonas: A-T-D-D-T-D-Dswyx: thing.No, very cool. Was that the last demo? Is thereJonas: Yeah.Anything I missed on the demos or points that you think? I think thatSamantha: covers it well. Yeah.swyx: Cool. Before we stop the screen share, can you gimme like a, just a tour of the slash commands ‘cause I so God ready. Huh, what? What are the good ones?Samantha: Yeah, we wanna increase discoverability around this too.I think that'll be like a future thing we work on. Yeah. But there's definitely a lot of good stuff nowJonas: we have a lot of internal ones that I think will not be that interesting. Here's an internal one that I've made. I don't know if anyone else at Cursor uses this one. Fix bb.Samantha: I've never heard of it.Jonas: Yeah.[00:14:00]Fix Bug Bot. So this is a thing that we want to integrate more tightly on. So you made it forswyx: yourself.Jonas: I made this for myself. It's actually available to everyone in the team, but yeah, no one knows about it. But yeah, there will be Bug bot comments and so Bug Bot has a lot of cool things. We actually just launched Bug Bot Auto Fix, where you can click a button and or change a setting and it will automatically fix its own things, and that works great in a bunch of cases.There are some cases where having the context of the original agent that created the PR is really helpful for fixing the bugs, because it might be like, oh, the bug here is that this, is a regression and actually you meant to do something more like that. And so having the original prompt and all of the context of the agent that worked on it, and so here I could just do, fix or we used to be able to do fixed PB and it would do that.No test is another one that we've had. Slash repro is in here. We mentioned that one.Samantha: One of my favorites is cloud agent diagnosis. This is one that makes heavy use of the Datadog MCP. Okay. And I [00:15:00] think Nick and David on our team wrote, and basically if there is a problem with a cloud agent we'll spin up a bunch of subs.Like a singleswyx: instance.Samantha: Yeah. We'll take the ideas and argument and spin up a bunch of subagents using the Datadog MCP to explore the logs and find like all of the problems that could have happened with that. It takes the debugging time, like from potentially you can do quick stuff quickly with the Datadog ui, but it takes it down to, again, like a single agent call as opposed to trolling through logs yourself.Jonas: You should also talk about the stuff we've done with transcripts.Samantha: Yes. Also so basically we've also done some things internally. There'll be some versions of this as we ship publicly soon, where you can spit up an agent and give it access to another agent's transcript to either basically debug something that happened.So act as an external debugger. I see. Or continue the conversation. Almost like forking it.swyx: A transcript includes all the chain of thought for the 11 minutes here. 45 minutes there.Samantha: Yeah. That way. Exactly. So basically acting as a like secondary agent that debugs the first, so we've started to push more andswyx: they're all the same [00:16:00] code.It is just the different prompts, but the sa the same.Samantha: Yeah. So basically same cloud agent infrastructure and then same harness. And then like when we do things like include, there's some extra infrastructure that goes into piping in like an external transcript if we include it as an attachment.But for things like the cloud agent diagnosis, that's mostly just using the Datadog MCP. ‘Cause we also launched CPS along with along with this cloud agent launch, launch support for cloud agent cps.swyx: Oh, that was drawn out.Jonas: We won't, we'll be doing a bigger marketing moment for it next week, but, and you can now use CPS andswyx: People will listen to it as well.Yeah,Jonas: they'llSamantha: be ahead of the third. They'll be ahead. And I would I actually don't know if the Datadog CP is like publicly available yet. I realize this not sure beta testing it, but it's been one of my favorites to use. Soswyx: I think that one's interesting for Datadog. ‘cause Datadog wants to own that site.Interesting with Bits. I don't know if you've tried bits.Samantha: I haven't tried bits.swyx: Yeah.Jonas: That's their cloud agentswyx: product. Yeah. Yeah. They want to be like we own your logs and give us our, some part of the, [00:17:00] self-healing software that everyone wants. Yeah. But obviously Cursor has a strong opinion on coding agents and you, you like taking away from the which like obviously you're going to do, and not every company's like Cursor, but it's interesting if you're a Datadog, like what do you do here?Do you expose your logs to FDP and let other people do it? Or do you try to own that it because it's extra business for you? Yeah. It's like an interesting one.Samantha: It's a good question. All I know is that I love the Datadog MCP,Jonas: And yeah, it is gonna be no, no surprise that people like will demand it, right?Samantha: Yeah.swyx: It's, it's like anysystemswyx: of record company like this, it's like how much do you give away? Cool. I think that's that for the sort of cloud agents tour. Cool. And we just talk about like cloud agents have been when did Kirsten loves cloud agents? Do you know, in JuneJonas: last year.swyx: June last year. So it's been slowly develop the thing you did, like a bunch of, like Michael did a post where himself, where he like showed this chart of like ages overtaking tap. And I'm like, wow, this is like the biggest transition in code.Jonas: Yeah.swyx: Like in, in [00:18:00] like the last,Jonas: yeah. I think that kind of got turned out.Yeah. I think it's a very interest,swyx: not at all. I think it's been highlighted by our friend Andre Kati today.Jonas: Okay.swyx: Talk more about it. What does it mean? Yeah. Is I just got given like the cursor tab key.Jonas: Yes. Yes.swyx: That's that'sSamantha: cool.swyx: I know, but it's gonna be like put in a museum.Jonas: It is.Samantha: I have to say I haven't used tab a little bit myself.Jonas: Yeah. I think that what it looks like to code with AI code generally creates software, even if you want to go higher level. Is changing very rapidly. No, not a hot take, but I think from our vendor's point at Cursor, I think one of the things that is probably underappreciated from the outside is that we are extremely self-aware about that fact and Kerscher, got its start in phase one, era one of like tab and auto complete.And that was really useful in its time. But a lot of people start looking at text files and editing code, like we call it hand coding. Now when you like type out the actual letters, it'sswyx: oh that's cute.Jonas: Yeah.swyx: Oh that's cute.Jonas: You're so boomer. So boomer. [00:19:00] And so that I think has been a slowly accelerating and now in the last few months, rapidly accelerating shift.And we think that's going to happen again with the next thing where the, I think some of the pains around tab of it's great, but I actually just want to give more to the agent and I don't want to do one tab at a time. I want to just give it a task and it goes off and does a larger unit of work and I can.Lean back a little bit more and operate at that higher level of abstraction that's going to happen again, where it goes from agents handing you back diffs and you're like in the weeds and giving it, 32nd to three minute tasks, to, you're giving it, three minute to 30 minute to three hour tasks and you're getting back videos and trying out previews rather than immediately looking at diffs every single time.swyx: Yeah. Anything to add?Samantha: One other shift that I've noticed as our cloud agents have really taken off internally has been a shift from primarily individually driven development to almost this collaborative nature of development for us, slack is actually almost like a development on [00:20:00] Id basically.So Iswyx: like maybe don't even build a custom ui, like maybe that's like a debugging thing, but actually it's that.Samantha: I feel like, yeah, there's still so much to left to explore there, but basically for us, like Slack is where a lot of development happens. Like we will have these issue channels or just like this product discussion channels where people are always at cursing and that kicks off a cloud agent.And for us at least, we have team follow-ups enabled. So if Jonas kicks off at Cursor in a thread, I can follow up with it and add more context. And so it turns into almost like a discussion service where people can like collaborate on ui. Oftentimes I will kick off an investigation and then sometimes I even ask it to get blame and then tag people who should be brought in. ‘cause it can tag people in Slack and then other people will comeswyx: in, can tag other people who are not involved in conversation. Yes. Can just do at Jonas if say, was talking to,Samantha: yeah.swyx: That's cool. You should, you guys should make a big good deal outta that.Samantha: I know. It's a lot to, I feel like there's a lot more to do with our slack surface area to show people externally. But yeah, basically like it [00:21:00] can bring other people in and then other people can also contribute to that thread and you can end up with a PR again, with the artifacts visible and then people can be like, okay, cool, we can merge this.So for us it's like the ID is almost like moving into Slack in some ways as well.swyx: I have the same experience with, but it's not developers, it's me. Designer salespeople.Samantha: Yeah.swyx: So me on like technical marketing, vision, designer on design and then salespeople on here's the legal source of what we agreed on.And then they all just collaborate and correct. The agents,Jonas: I think that we found when these threads is. The work that is left, that the humans are discussing in these threads is the nugget of what is actually interesting and relevant. It's not the boring details of where does this if statement go?It's do we wanna ship this? Is this the right ux? Is this the right form factor? Yeah. How do we make this more obvious to the user? It's like those really interesting kind of higher order questions that are so easy to collaborate with and leave the implementation to the cloud agent.Samantha: Totally. And no more discussion of am I gonna do this? Are you [00:22:00] gonna do this cursor's doing it? You just have to decide. You like it.swyx: Sometimes the, I don't know if there's a, this probably, you guys probably figured this out already, but since I, you need like a mute button. So like cursor, like we're going to take this offline, but still online.But like we need to talk among the humans first. Before you like could stop responding to everything.Jonas: Yeah. This is a design decision where currently cursor won't chime in unless you explicitly add Mention it. Yeah. Yeah.Samantha: So it's not always listening.Yeah.Jonas: I can see all the intermediate messages.swyx: Have you done the recursive, can cursor add another cursor or spawn another cursor?Samantha: Oh,Jonas: we've done some versions of this.swyx: Because, ‘cause it can add humans.Jonas: Yes. One of the other things we've been working on that's like an implication of generating the code is so easy is getting it to production is still harder than it should be.And broadly, you solve one bottleneck and three new ones pop up. Yeah. And so one of the new bottlenecks is getting into production and we have a like joke internally where you'll be talking about some feature and someone says, I have a PR for that. Which is it's so easy [00:23:00] to get to, I a PR for that, but it's hard still relatively to get from I a PR for that to, I'm confident and ready to merge this.And so I think that over the coming weeks and months, that's a thing that we think a lot about is how do we scale up compute to that pipeline of getting things from a first draft An agent did.swyx: Isn't that what Merge isn't know what graphite's for, likeJonas: graphite is a big part of that. The cloud agent testingswyx: Is it fully integrated or still different companiesJonas: working on I think we'll have more to share there in the future, but the goal is to have great end-to-end experience where Cursor doesn't just help you generate code tokens, it helps you create software end-to-end.And so review is a big part of that, that I think especially as models have gotten much better at writing code, generating code, we've felt that relatively crop up more,swyx: sorry this is completely unplanned, but like there I have people arguing one to you need ai. To review ai and then there is another approach, thought school of thought where it's no, [00:24:00] reviews are dead.Like just show me the video. It's it like,Samantha: yeah. I feel again, for me, the video is often like alignment and then I often still wanna go through a code review process.swyx: Like still look at the files andSamantha: everything. Yeah. There's a spectrum of course. Like the video, if it's really well done and it does like fully like test everything, you can feel pretty competent, but it's still helpful to, to look at the code.I make hep pay a lot of attention to bug bot. I feel like Bug Bot has been a great really highly adopted internally. We often like, won't we tell people like, don't leave bug bot comments unaddressed. ‘cause we have such high confidence in it. So people always address their bug bot comments.Jonas: Once you've had two cases where you merged something and then you went back later, there was a bug in it, you merged, you went back later and you were like, ah, bug Bot had found that I should have listened to Bug Bot.Once that happens two or three times, you learn to wait for bug bot.Samantha: Yeah. So I think for us there's like that code level review where like it's looking at the actual code and then there's like the like feature level review where you're looking at the features. There's like a whole number of different like areas.There'll probably eventually be things like performance level review, security [00:25:00] review, things like that where it's like more more different aspects of how this feature might affect your code base that you want to potentially leverage an agent to help with.Jonas: And some of those like bug bot will be synchronous and you'll typically want to wait on before you merge.But I think another thing that we're starting to see is. As with cloud agents, you scale up this parallelism and how much code you generate. 10 person startups become, need the Devrel X and pipelines that a 10,000 person company used to need. And that looks like a lot of the things I think that 10,000 person companies invented in order to get that volume of software to production safely.So that's things like, release frequently or release slowly, have different stages where you release, have checkpoints, automated ways of detecting regressions. And so I think we're gonna need stacks merg stack diffs merge queues. Exactly. A lot of those things are going to be importantswyx: forward with.I think the majority of people still don't know what stack stacks are. And I like, I have many friends in Facebook and like I, I'm pretty friendly with graphite. I've just, [00:26:00] I've never needed it ‘cause I don't work on that larger team and it's just like democratization of no, only here's what we've already worked out at very large scale and here's how you can, it benefits you too.Like I think to me, one of the beautiful things about GitHub is that. It's actually useful to me as an individual solo developer, even though it's like actually collaboration software.Jonas: Yep.swyx: And I don't think a lot of Devrel tools have figured that out yet. That transition from like large down to small.Jonas: Yeah. Kers is probably an inverse story.swyx: This is small down toJonas: Yeah. Where historically Kers share, part of why we grew so quickly was anyone on the team could pick it up and in fact people would pick it up, on the weekend for their side project and then bring it into work. ‘cause they loved using it so much.swyx: Yeah.Jonas: And I think a thing that we've started working on a lot more, not us specifically, but as a company and other folks at Cursor, is making it really great for teams and making it the, the 10th person that starts using Cursor in a team. Is immediately set up with things like, we launched Marketplace recently so other people can [00:27:00] configure what CPS and skills like plugins.So skills and cps, other people can configure that. So that my cursor is ready to go and set up. Sam loves the Datadog, MCP and Slack, MCP you've also been using a lot butSamantha: also pre-launch, but I feel like it's so good.Jonas: Yeah, my cursor should be configured if Sam feels strongly that's just amazing and required.swyx: Is it automatically shared or you have to go and.Jonas: It depends on the MCP. So some are obviously off per user. Yeah. And so Sam can't off my cursor with my Slack MCP, but some are team off and those can be set up by admins.swyx: Yeah. Yeah. That's cool. Yeah, I think, we had a man on the pod when cursor was five people, and like everyone was like, okay, what's the thing?And then it's usually something teams and org and enterprise, but it's actually working. But like usually at that stage when you're five, when you're just a vs. Code fork it's like how do you get there? Yeah. Will people pay for this? People do pay for it.Jonas: Yeah. And I think for cloud agents, we expect.[00:28:00]To have similar kind of PLG things where I think off the bat we've seen a lot of adoption with kind of smaller teams where the code bases are not quite as complex to set up. Yes. If you need some insane docker layer caching thing for builds not to take two hours, that's going to take a little bit longer for us to be able to support that kind of infrastructure.Whereas if you have front end backend, like one click agents can install everything that they need themselves.swyx: This is a good chance for me to just ask some technical sort of check the box questions. Can I choose the size of the vm?Jonas: Not yet. We are planning on adding that. Weswyx: have, this is obviously you want like LXXL, whatever, right?Like it's like the Amazon like sort menu.Jonas: Yes, exactly. We'll add that.swyx: Yeah. In some ways you have to basically become like a EC2, almost like you rent a box.Jonas: You rent a box. Yes. We talk a lot about brain in a box. Yeah. So cursor, we want to be a brain in a box,swyx: but is the mental model different? Is it more serverless?Is it more persistent? Is. Something else.Samantha: We want it to be a bit persistent. The desktop should be [00:29:00] something you can return to af even after some days. Like maybe you go back, they're like still thinking about a feature for some period of time. So theswyx: full like sus like suspend the memory and bring it back and then keep going.Samantha: Exactly.swyx: That's an interesting one because what I actually do want, like from a manna and open crawl, whatever, is like I want to be able to log in with my credentials to the thing, but not actually store it in any like secret store, whatever. ‘cause it's like this is the, my most sensitive stuff.Yeah. This is like my email, whatever. And just have it like, persist to the image. I don't know how it was hood, but like to rehydrate and then just keep going from there. But I don't think a lot of infra works that way. A lot of it's stateless where like you save it to a docker image and then it's only whatever you can describe in a Docker file and that's it.That's the only thing you can cl multiple times in parallel.Jonas: Yeah. We have a bunch of different ways of setting them up. So there's a dockerfile based approach. The main default way is actually snapshottingswyx: like a Linux vmJonas: like vm, right? You run a bunch of install commands and then you snapshot more or less the file system.And so that gets you set up for everything [00:30:00] that you would want to bring a new VM up from that template basically.swyx: Yeah.Jonas: And that's a bit distinct from what Sam was talking about with the hibernating and re rehydrating where that is a full memory snapshot as well. So there, if I had like the browser open to a specific page and we bring that back, that page will still be there.swyx: Was there any discussion internally and just building this stuff about every time you shoot a video it's actually you show a little bit of the desktop and the browser and it's not necessary if you just show the browser. If, if you know you're just demoing a front end application.Why not just show the browser, right? Like it Yeah,Samantha: we do have some panning and zooming. Yeah. Like it can decide that when it's actually recording and cutting the video to highlight different things. I think we've played around with different ways of segmenting it and yeah. There's been some different revs on it for sure.Jonas: Yeah. I think one of the interesting things is the version that you see now in cursor.com actually is like half of what we had at peak where we decided to unshift or unshipped quite a few things. So two of the interesting things to talk about, one is directly an answer to your [00:31:00] question where we had native browser that you would have locally, it was basically an iframe that via port forwarding could load the URL could talk to local host in the vm.So that gets you basically, so inswyx: your machine's browser,likeJonas: in your local browser? Yeah. You would go to local host 4,000 and that would get forwarded to local host 4,000 in the VM via port forward. We unshift that like atswyx: Eng Rock.Jonas: Like an Eng Rock. Exactly. We unshift that because we felt that the remote desktop was sufficiently low latency and more general purpose.So we build Cursor web, but we also build Cursor desktop. And so it's really useful to be able to have the full spectrum of things. And even for Cursor Web, as you saw in one of the examples, the agent was uploading files and like I couldn't upload files and open the file viewer if I only had access to the browser.And we've thought a lot about, this might seem funny coming from Cursor where we started as this, vs. Code Fork and I think inherited a lot of amazing things, but also a lot [00:32:00] of legacy UI from VS Code.Minimal Web UI SurfacesJonas: And so with the web UI we wanted to be very intentional about keeping that very minimal and exposing the right sum of set of primitive sort of app surfaces we call them, that are shared features of that cloud.Environment that you and the agent both use. So agent uses desktop and controls it. I can use desktop and controlled agent runs terminal commands. I can run terminal commands. So that's how our philosophy around it. The other thing that is maybe interesting to talk about that we unshipped is and we may, both of these things we may reship and decide at some point in the future that we've changed our minds on the trade offs or gotten it to a point where, putswyx: it out there.Let users tell you they want it. Exactly. Alright, fine.Why No File EditorJonas: So one of the other things is actually a files app. And so we used to have the ability at one point during the process of testing this internally to see next to, I had GID desktop and terminal on the right hand side of the tab there earlier to also have a files app where you could see and edit files.And we actually felt that in some [00:33:00] ways, by restricting and limiting what you could do there, people would naturally leave more to the agent and fall into this new pattern of delegating, which we thought was really valuable. And there's currently no way in Cursor web to edit these files.swyx: Yeah. Except you like open up the PR and go into GitHub and do the thing.Jonas: Yeah.swyx: Which is annoying.Jonas: Just tell the agent,swyx: I have criticized open AI for this. Because Open AI is Codex app doesn't have a file editor, like it has file viewer, but isn't a file editor.Jonas: Do you use the file viewer a lot?swyx: No. I understand, but like sometimes I want it, the one way to do it is like freaking going to no, they have a open in cursor button or open an antigravity or, opening whatever and people pointed that.So I was, I was part of the early testers group people pointed that and they were like, this is like a design smell. It's like you actually want a VS. Code fork that has all these things, but also a file editor. And they were like, no, just trust us.Jonas: Yeah. I think we as Cursor will want to, as a product, offer the [00:34:00] whole spectrum and so you want to be able to.Work at really high levels of abstraction and double click and see the lowest level. That's important. But I also think that like you won't be doing that in Slack. And so there are surfaces and ways of interacting where in some cases limiting the UX capabilities makes for a cleaner experience that's more simple and drives people into these new patterns where even locally we kicked off joking about this.People like don't really edit files, hand code anymore. And so we want to build for where that's going and not where it's beenswyx: a lot of cool stuff. And Okay. I have a couple more.Full Stack Hosting Debateswyx: So observations about the design elements about these things. One of the things that I'm always thinking about is cursor and other peers of cursor start from like the Devrel tools and work their way towards cloud agents.Other people, like the lovable and bolts of the world start with here's like the vibe code. Full cloud thing. They were already cloud edges before anyone else cloud edges and we will give you the full deploy platform. So we own the whole loop. We own all the infrastructure, we own, we, we have the logs, we have the the live site, [00:35:00] whatever.And you can do that cycle cursor doesn't own that cycle even today. You don't have the versal, you don't have the, you whatever deploy infrastructure that, that you're gonna have, which gives you powers because anyone can use it. And any enterprise who, whatever you infra, I don't care. But then also gives you limitations as to how much you can actually fully debug end to end.I guess I'm just putting out there that like is there a future where there's like full stack cursor where like cursor apps.com where like I host my cursor site this, which is basically a verse clone, right? I don't know.Jonas: I think that's a interesting question to be asking, and I think like the logic that you laid out for how you would get there is logic that I largely agree with.swyx: Yeah. Yeah.Jonas: I think right now we're really focused on what we see as the next big bottleneck and because things like the Datadog MCP exist, yeah. I don't think that the best way we can help our customers ship more software. Is by building a hosting solution right now,swyx: by the way, these are things I've actually discussed with some of the companies I just named.Jonas: Yeah, for sure. Right now, just this big bottleneck is getting the code out there and also [00:36:00] unlike a lovable in the bolt, we focus much more on existing software. And the zero to one greenfield is just a very different problem. Imagine going to a Shopify and convincing them to deploy on your deployment solution.That's very different and I think will take much longer to see how that works. May never happen relative to, oh, it's like a zero to one app.swyx: I'll say. It's tempting because look like 50% of your apps are versal, superb base tailwind react it's the stack. It's what everyone does.So I it's kinda interesting.Jonas: Yeah.Model Choice and Auto Routingswyx: The other thing is the model select dying. Right now in cloud agents, it's stuck down, bottom left. Sure it's Codex High today, but do I care if it's suddenly switched to Opus? Probably not.Samantha: We definitely wanna give people a choice across models because I feel like it, the meta change is very frequently.I was a big like Opus 4.5 Maximalist, and when codex 5.3 came out, I hard, hard switch. So that's all I use now.swyx: Yeah. Agreed. I don't know if, but basically like when I use it in Slack, [00:37:00] right? Cursor does a very good job of exposing yeah. Cursors. If people go use it, here's the model we're using.Yeah. Here's how you switch if you want. But otherwise it's like extracted away, which is like beautiful because then you actually, you should decide.Jonas: Yeah, I think we want to be doing more with defaults.swyx: Yeah.Jonas: Where we can suggest things to people. A thing that we have in the editor, the desktop app is auto, which will route your request and do things there.So I think we will want to do something like that for cloud agents as well. We haven't done it yet. And so I think. We have both people like Sam, who are very savvy and want know exactly what model they want, and we also have people that want us to pick the best model for them because we have amazing people like Sam and we, we are the experts.Yeah. We have both the traffic and the internal taste and experience to know what we think is best.swyx: Yeah. I have this ongoing pieces of agent lab versus model lab. And to me, cursor and other companies are example of an agent lab that is, building a new playbook that is different from a model lab where it's like very GP heavy Olo.So obviously has a research [00:38:00] team. And my thesis is like you just, every agent lab is going to have a router because you're going to be asked like, what's what. I don't keep up to every day. I'm not a Sam, I don't keep up every day for using you as sample the arm arbitrator of taste. Put me on CRI Auto.Is it free? It's not free.Jonas: Auto's not free, but there's different pricing tiers. Yeah.swyx: Put me on Chris. You decide from me based on all the other people you know better than me. And I think every agent lab should basically end up doing this because that actually gives you extra power because you like people stop carrying or having loyalty with one lab.Jonas: Yeah.Best Of N and Model CouncilsJonas: Two other maybe interesting things that I don't know how much they're on your radar are one the best event thing we mentioned where running different models head to head is actually quite interesting becauseswyx: which exists in cursor.Jonas: That exists in cur ID and web. So the problem is where do you run them?swyx: Okay.Jonas: And so I, I can share my screen if that's interesting. Yeahinteresting.swyx: Yeah. Yeah. Obviously parallel agents, very popal.Jonas: Yes, exactly. Parallel agentsswyx: in you mind. Are they the same thing? Best event and parallel agents? I don't want to [00:39:00] put words in your mouth.Jonas: Best event is a subset of parallel agents where they're running on the same prompt.That would be my answer. So this is what that looks like. And so here in this dropdown picker, I can just select multiple models.swyx: Yeah.Jonas: And now if I do a prompt, I'm going to do something silly. I am running these five models.swyx: Okay. This is this fake clone, of course. The 2.0 yeah.Jonas: Yes, exactly. But they're running so the cursor 2.0, you can do desktop or cloud.So this is cloud specifically where the benefit over work trees is that they have their own VMs and can run commands and won't try to kill ports that the other one is running. Which are some of the pains. These are allswyx: called work trees?Jonas: No, these are all cloud agents with their own VMs.swyx: Okay. ButJonas: When you do it locally, sometimes people do work trees and that's been the main way that people have set out parallel so far.I've gotta say.swyx: That's so confusing for folks.Jonas: Yeah.swyx: No one knows what work trees are.Jonas: Exactly. I think we're phasing out work trees.swyx: Really.Jonas: Yeah.swyx: Okay.Samantha: But yeah. And one other thing I would say though on the multimodel choice, [00:40:00] so this is another experiment that we ran last year and the decide to ship at that time but may come back to, and there was an interesting learning that's relevant for, these different model providers. It was something that would run a bunch of best of ends but then synthesize and basically run like a synthesizer layer of models. And that was other agents that would take LM Judge, but one that was also agentic and could write code. So it wasn't just picking but also taking the learnings from two models or, and models that it was looking at and writing a new diff.And what we found was that at the time at least, there were strengths to using models from different model providers as the base level of this process. Like basically you could get almost like a synergistic output that was better than having a very unified, like bottom model tier. So it was really interesting ‘cause it's like potentially, even though even in the future when you have like maybe one model as ahead of the other for a little bit, there could be some benefit from having like multiple top tier models involved in like a [00:41:00] model swarm or whatever agent Swarm that you're doing, that they each have strengths and weaknesses.Yeah.Jonas: Andre called this the council, right?Samantha: Yeah, exactly. We actually, oh, that's another internal command we have that Ian wrote slash council. Oh, and they some, yeah.swyx: Yes. This idea is in various forms everywhere. And I think for me, like for me, the productization of it, you guys have done yeah, like this is very flexible, but.If I were to add another Yeah, what your thing is on here it would be too much. I what, let's say,Samantha: Ideally it's all, it's something that the user can just choose and it all happens under the hood in a way where like you just get the benefit of that process at the end and better output basically, but don't have to get too lost in the complexity of judging along the way.Jonas: Okay.Subagents for ContextJonas: Another thing on the many agents, on different parallel agents that's interesting is an idea that's been around for a while as well that has started working recently is subagents. And so this is one other way to get agents of the different prompts and different goals and different models, [00:42:00] different vintages to work together.Collaborate and delegate.swyx: Yeah. I'm very like I like one of my, I always looking for this is the year of the blah, right? Yeah. I think one of the things on the blahs is subs. I think this is of but I haven't used them in cursor. Are they fully formed or how do I honestly like an intro because do I form them from new every time?Do I have fixed subagents? How are they different for slash commands? There's all these like really basic questions that no one stops to answer for people because everyone's just like too busy launching. We have toSamantha: honestly, you could, you can see them in cursor now if you just say spin up like 50 subagents to, so cursor definesswyx: what Subagents.Yeah.Samantha: Yeah. So basically I think I shouldn't speak for the whole subagents team. This is like a different team that's been working on this, but our thesis or thing that we saw internally is that like they're great for context management for kind of long running threads, or if you're trying to just throw more compute at something.We have strongly used, almost like a generic task interface where then the main agent can define [00:43:00] like what goes into the subagent. So if I say explore my code base, it might decide to spin up an explore subagent and or might decide to spin up five explore subagent.swyx: But I don't get to set what those subagent are, right?It's all defined by a model.Samantha: I think. I actually would have to refresh myself on the sub agent interface.Jonas: There are some built-in ones like the explore subagent is free pre-built. But you can also instruct the model to use other subagents and then it will. And one other example of a built-in subagent is I actually just kicked one off in cursor and I can show you what that looks like.swyx: Yes. Because I tried to do this in pure prompt space.Jonas: So this is the desktop app? Yeah. Yeah. And that'sswyx: all you need to do, right? Yeah.Jonas: That's all you need to do. So I said use a sub agent to explore and I think, yeah, so I can even click in and see what the subagent is working on here. It ran some fine command and this is a composer under the hood.Even though my main model is Opus, it does smart routing to take, like in this instance the explorer sort of requires reading a ton of things. And so a faster model is really useful to get an [00:44:00] answer quickly, but that this is what subagent look like. And I think we wanted to do a lot more to expose hooks and ways for people to configure these.Another example of a cus sort of builtin subagent is the computer use subagent in the cloud agents, where we found that those trajectories can be long and involve a lot of images obviously, and execution of some testing verification task. We wanted to use that models that are particularly good at that.So that's one reason to use subagents. And then the other reason to use subagents is we want contexts to be summarized reduced down at a subagent level. That's a really neat boundary at which to compress that rollout and testing into a final message that agent writes that then gets passed into the parent rather than having to do some global compaction or something like that.swyx: Awesome. Cool. While we're in the subagents conversation, I can't do a cursor conversation and not talk about listen stuff. What is that? What is what? He built a browser. He built an os. Yes. And he [00:45:00] experimented with a lot of different architectures and basically ended up reinventing the software engineer org chart.This is all cool, but what's your take? What's, is there any hole behind the side? The scenes stories about that kind of, that whole adventure.Samantha: Some of those experiments have found their way into a feature that's available in cloud agents now, the long running agent mode internally, we call it grind mode.And I think there's like some hint of grind mode accessible in the picker today. ‘cause you can do choose grind until done. And so that was really the result of experiments that Wilson started in this vein where he I think the Ralph Wigga loop was like floating around at the time, but it was something he also independently found and he was experimenting with.And that was what led to this product surface.swyx: And it is just simple idea of have criteria for completion and do not. Until you complete,Samantha: there's a bit more complexity as well in, in our implementation. Like there's a specific, you have to start out by aligning and there's like a planning stage where it will work with you and it will not get like start grind execution mode until it's decided that the [00:46:00] plan is amenable to both of you.Basically,swyx: I refuse to work until you make me happy.Jonas: We found that it's really important where people would give like very underspecified prompt and then expect it to come back with magic. And if it's gonna go off and work for three minutes, that's one thing. When it's gonna go off and work for three days, probably should spend like a few hours upfront making sure that you have communicated what you actually want.swyx: Yeah. And just to like really drive from the point. We really mean three days that No, noJonas: human. Oh yeah. We've had three day months innovation whatsoever.Samantha: I don't know what the record is, but there's been a long time with the grantsJonas: and so the thing that is available in cursor. The long running agent is if you wanna think about it, very abstractly that is like one worker node.Whereas what built the browser is a society of workers and planners and different agents collaborating. Because we started building the browser with one worker node at the time, that was just the agent. And it became one worker node when we realized that the throughput of the system was not where it needed to be [00:47:00] to get something as large of a scale as the browser done.swyx: Yeah.Jonas: And so this has also become a really big mental model for us with cloud, cloud agents is there's the classic engineering latency throughput trade-offs. And so you know, the code is water flowing through a pipe. The, we think that over the coming months, the big unlock is not going to be one person with a model getting more done, like the water flowing faster and we'll be making the pipe much wider and so ing more, whether that's swarms of agents or parallel agents, both of those are things that contribute to getting.Much more done in the same amount of time, but any one of those tasks doesn't necessarily need to get done that quickly. And throughput is this really big thing where if you see the system of a hundred concurrent agents outputting thousands of tokens a second, you can't go back like that.Just you see a glimpse of the future where obviously there are many caveats. Like no one is using this browser. IRL. There's like a bunch of things not quite right yet, but we are going to get to systems that produce real production [00:48:00] code at the scale much sooner than people think. And it forces you to think what even happens to production systems. Like we've broken our GitHub actions recently because we have so many agents like producing and pushing code that like CICD is just overloaded. ‘cause suddenly it's like effectively weg grew, cursor's growing very quickly anyway, but you grow head count, 10 x when people run 10 x as many agents.And so a lot of these systems, exactly, a lot of these systems will need to adapt.swyx: It also reminds me, we, we all, the three of us live in the app layer, but if you talk to the researchers who are doing RL infrastructure, it's the same thing. It's like all these parallel rollouts and scheduling them and making sure as much throughput as possible goes through them.Yeah, it's the same thing.Jonas: We were talking briefly before we started recording. You were mentioning memory chips and some of the shortages there. The other thing that I think is just like hard to wrap your head around the scale of the system that was building the browser, the concurrency there.If Sam and I both have a system like that running for us, [00:49:00] shipping our software. The amount of inference that we're going to need per developer is just really mind-boggling. And that makes, sometimes when I think about that, I think that even with, the most optimistic projections for what we're going to need in terms of buildout, our underestimating, the extent to which these swarm systems can like churn at scale to produce code that is valuable to the economy.And,swyx: yeah, you can cut this if it's sensitive, but I was just Do you have estimates of how much your token consumption is?Jonas: Like per developer?swyx: Yeah. Or yourself. I don't need like comfy average. I just curious. ISamantha: feel like I, for a while I wasn't an admin on the usage dashboard, so I like wasn't able to actually see, but it was a,swyx: mine has gone up.Samantha: Oh yeah.swyx: But I thinkSamantha: it's in terms of how much work I'm doing, it's more like I have no worries about developers losing their jobs, at least in the near term. ‘cause I feel like that's a more broad discussion.swyx: Yeah. Yeah. You went there. I didn't go, I wasn't going there.I was just like how much more are you using?Samantha: There's so much stuff to be built. And so I feel like I'm basically just [00:50:00] trying to constantly I have more ambitions than I did before. Yes. Personally. Yes. So can't speak to the broader thing. But for me it's like I'm busier than ever before.I'm using more tokens and I am also doing more things.Jonas: Yeah. Yeah. I don't have the stats for myself, but I think broadly a thing that we've seen, that we expect to continue is J'S paradox. Whereswyx: you can't do it in our podcast without seeingJonas: it. Exactly. We've done it. Now we can wrap. We've done, we said the words.Phase one tab auto complete people paid like 20 bucks a month. And that was great. Phase two where you were iterating with these local models. Today people pay like hundreds of dollars a month. I think as we think about these highly parallel kind of agents running off for a long times in their own VM system, we are already at that point where people will be spending thousands of dollars a month per human, and I think potentially tens of thousands and beyond, where it's not like we are greedy for like capturing more money, but what happens is just individuals get that much more leverage.And if one person can do as much as 10 people, yeah. That tool that allows ‘em to do that is going to be tremendously valuable [00:51:00] and worth investing in and taking the best thing that exists.swyx: One more question on just the cursor in general and then open-ended for you guys to plug whatever you wanna put.How is Cursor hiring these days?Samantha: What do you mean by how?swyx: So obviously lead code is dead. Oh,Samantha: okay.swyx: Everyone says work trial. Different people have different levels of adoption of agents. Some people can really adopt can be much more productive. But other people, you just need to give them a little bit of time.And sometimes they've never lived in a token rich place like cursor.And once you live in a token rich place, you're you just work differently. But you need to have done that. And a lot of people anyway, it was just open-ended. Like how has agentic engineering, agentic coding changed your opinions on hiring?Is there any like broad like insights? Yeah.Jonas: Basically I'm asking this for other people, right? Yeah, totally. Totally. To hear Sam's opinion, we haven't talked about this the two of us. I think that we don't see necessarily being great at the latest thing with AI coding as a prerequisite.I do think that's a sign that people are keeping up and [00:52:00] curious and willing to upscale themselves in what's happening because. As we were talking about the last three months, the game has completely changed. It's like what I do all day is very different.swyx: Like it's my job and I can't,Jonas: Yeah, totally.I do think that still as Sam was saying, the fundamentals remain important in the current age and being able to go and double click down. And models today do still have weaknesses where if you let them run for too long without cleaning up and refactoring, the coke will get sloppy and there'll be bad abstractions.And so you still do need humans that like have built systems before, no good patterns when they see them and know where to steer things.Samantha: I would agree with that. I would say again, cursor also operates very quickly and leveraging ag agentic engineering is probably one reason why that's possible in this current moment.I think in the past it was just like people coding quickly and now there's like people who use agents to move faster as well. So it's part of our process will always look for we'll select for kind of that ability to make good decisions quickly and move well in this environment.And so I think being able to [00:53:00] figure out how to use agents to help you do that is an important part of it too.swyx: Yeah. Okay. The fork in the road, either predictions for the end of the year, if you have any, or PUDs.Jonas: Evictions are not going to go well.Samantha: I know it's hard.swyx: They're so hard. Get it wrong.It's okay. Just, yeah.Jonas: One other plug that may be interesting that I feel like we touched on but haven't talked a ton about is a thing that the kind of these new interfaces and this parallelism enables is the ability to hop back and forth between threads really quickly. And so a thing that we have,swyx: you wanna show something or,Jonas: yeah, I can show something.A thing that we have felt with local agents is this pain around contact switching. And you have one agent that went off and did some work and another agent that, that did something else. And so here by having, I just have three tabs open, let's say, but I can very quickly, hop in here.This is an example I showed earlier, but the actual workflow here I think is really different in a way that may not be obvious, where, I start t

The reception to our recent post on Code Reviews has been strong. Catch up!Amid a maelstrom of discussion on whether or not AI is killing SaaS, one of the top publicly listed SaaS companies in the world has just reported record revenues, clearing well over $1.1B in ARR for the first time with a 28% margin. As we comment on the pod, Aaron Levie is the rare public company CEO equally at home in both worlds of Silicon Valley and Wall Street/Main Street, by day helping 70% of the Fortune 500 with their Enterprise Advanced Suite, and yet by night is often found in the basements of early startups and tweeting viral insights about the future of agents.Now that both Cursor, Cloudflare, Perplexity, Anthropic and more have made Filesystems and Sandboxes and various forms of “Just Give the Agent a Box” cool (not just cool; it is now one of the single hottest areas in AI infrastructure growing 100% MoM), we find it a delightfully appropriate time to do the episode with the OG CEO who has been giving humans and computers Boxes since he was a college dropout pitching VCs at a Michael Arrington house party.Enjoy our special pod, with fan favorite returning guest/guest cohost Jeff Huber!Note: We didn't directly discuss the AI vs SaaS debate - Aaron has done many, many, many other podcasts on that, and you should read his definitive essay on it. Most commentators do not understand SaaS businesses because they have never scaled one themselves, and deeply reflected on what the true value proposition of SaaS is.We also discuss Your Company is a Filesystem:We also shoutout CTO Ben Kus' and the AI team, who talked about the technical architecture and will return for AIE WF 2026.Full Video EpisodeTimestamps* 00:00 Adapting Work for Agents* 01:29 Why Every Agent Needs a Box* 04:38 Agent Governance and Identity* 11:28 Why Coding Agents Took Off First* 21:42 Context Engineering and Search Limits* 31:29 Inside Agent Evals* 33:23 Industries and Datasets* 35:22 Building the Agent Team* 38:50 Read Write Agent Workflows* 41:54 Docs Graphs and Founder Mode* 55:38 Token FOMO Culture* 56:31 Production Function Secrets* 01:01:08 Film Roots to Box* 01:03:38 AI Future of Movies* 01:06:47 Media DevRel and EngineeringTranscriptAdapting Work for AgentsAaron Levie: Like you don't write code, you talk to an agent and it goes and does it for you, and you may be at best review it. That's even probably like, like largely not even what you're doing. What's happening is we are changing our work to make the agents effective. In that model, the agent didn't really adapt to how we work.We basically adapted to how the agent works. All of the economy has to go through that exact same evolution. Right now, it's a huge asset and an advantage for the teams that do it early and that are kinda wired into doing this ‘cause you'll see compounding returns. But that's just gonna take a while for most companies to actually go and get this deployed.swyx: Welcome to the Lane Space Pod. We're back in the chroma studio with uh, chroma, CEO, Jeff Hoover. Welcome returning guest now guest host.Aaron Levie: It's a pleasure. Wow. How'd you get upgraded to, uh, to that?swyx: Because he's like the perfect guy to be guest those for you.Aaron Levie: That makes sense actually, for We love context. We, we both really love context le we really do.We really do.swyx: Uh, and we're here with, uh, Aaron Levy. Welcome.Aaron Levie: Thank you. Good to, uh, good to be [00:01:00] here.swyx: Uh, yeah. So we've all met offline and like chatted a little bit, but like, it's always nice to get these things in person and conversation. Yeah. You just started off with so much energy. You're, you're super excited about agents.I loveAaron Levie: agents.swyx: Yeah. Open claw. Just got by, got bought by OpenAI. No, not bought, but you know, you know what I mean?Aaron Levie: Some, some, you know, acquihire. Executiveswyx: hire.Aaron Levie: Executive hire. Okay. Executive hire. Say,swyx: hey, that's my term. Okay. Um, what are you pounding the table on on agents? You have so many insightful tweets.Why Every Agent Needs a BoxAaron Levie: Well, the thing that, that we get super excited by that I think is probably, you know, should be relatively obvious is we've, we've built a platform to help enterprises manage their files and their, their corporate files and the permissions of who has access to those files and the sharing collaboration of those files.All of those files contain really, really important information for the enterprise. It might have your contracts, it might have your research materials, it might have marketing information, it might have your memos. All that data obviously has, you know, predominantly been used by humans. [00:02:00] But there's been one really interesting problem, which is that, you know, humans only really work with their files during an active engagement with them, and they kind of go away and you don't really see them for a long time.And all of a sudden, uh, with the power of AI and AI agents, all of that data becomes extremely relevant as this ongoing source of, of answers to new questions of data that will transform into, into something else that, that produces value in your organization. It, it contains the answer to the new employee that's onboarding, that needs to ramp up on a project.Um, it contains the answer to the right thing to sell a customer when you're having a conversation to them, with them contains the roadmap information that's gonna produce the next feature. So all that data. That previously we've been just sort of storing and, and you know, occasionally forgetting about, ‘cause we're only working on the new active stuff.All of that information becomes valuable to the enterprise and it's gonna become extremely valuable to end users because now they can have agents go find what they're looking for and produce new, new [00:03:00] value and new data on that information. And it's gonna become incredibly valuable to agents because agents can roam around and do a bunch of work and they're gonna need access to that data as well.And um, and you know, sometimes that will be an agent that is sort of working on behalf of, of, of you and, and effectively as you as and, and they are kind of accessing all of the same information that you have access to and, and operating as you in the system. And then sometimes there's gonna be agents that are just.Effectively autonomous and kind of run on their own and, and you're gonna collaborate and work with them kind of like you did another person. Open Claw being the most recent and maybe first real sort of, you know, kind of, you know, up updating everybody's, you know, views of this landscape version of, of what that could look like, which is, okay, I have an agent.It's on its own system, it's on its own computer, it has access to its own tools. I probably don't give it access to my entire life. I probably communicate with it like I would an assistant or a colleague and then it, it sort of has this sandbox environment. So all of that has massive implications for a platform that manage that [00:04:00] enterprise data.We think it's gonna just transform how we work with all of the enterprise content that we work with, and we just have to make sure we're building the right platform to support that.swyx: The sort of shorthand I put it is as people build agents, everybody's just realizing that every agent needs a box. Yes.And it's nice to be called box and just give everyone a box.Aaron Levie: Hey, I if I, you know, if we can make that go viral, uh, like I, I think that that terminology, I, that's theswyx: tagline. Every agentAaron Levie: needs a box. Every agent needs a box. If we can make that the headline of this, I'm fine with this. And that's the billboard I wanna like Yeah, exactly.Every agent needs a box. Um, I like it. Can we ship this? Like,swyx: okay, let's do it. Yeah.Aaron Levie: Uh, my work here is done and I got the value I needed outta this podcast Drinks.swyx: Yeah.Agent Governance and IdentityAaron Levie: But, but, um, but, but, you know, so the thing that we, we kind of think about is, um, is, you know, whether you think the number 10 x or a hundred x or whatever the number is, we're gonna have some order of magnitude more agents than people.That's inevitable. It has to happen. So then the question is, what is the infrastructure that's needed to make all those agents effective in the enterprise? Make sure that they are well governed. Make sure they're only doing [00:05:00] safe things on your information. Make sure that they're not getting exposed. The data that they shouldn't have access to.There's gonna be just incredibly spectacularly crazy security incidents that will happen with agents because you'll prompt, inject an agent and sort of find your way through the CRM system and pull out data that you shouldn't have access to. Oh, weJeff Huber: have God,Aaron Levie: right? I mean, that's just gonna happen all over the place, right?So, so then the thing is, is how do you make sure you have the right security, the permissions, the access controls, the data governance. Um, we actually don't yet exactly know in many cases how we're gonna regulate some of these agents, right? If you think about an agent in financial services, does it have the exact same financial sort of, uh, requirements that a human did?Or is it, is the risk fully on the human that was interacting or created the agent? All open questions, but no matter what, there's gonna need to be a layer that manages the, the data they have access to, the workflows that they're involved in, pulling up data from multiple systems. This is the new infrastructure opportunity in the era of agents.swyx: You have a piece on agent identities, [00:06:00] which I think was today, um, which I think a lot of breaking news, the security, security people are talking about, right? Like you basically, I, I always think of this as like, well you need the human you and then there you need the agent. YouAaron Levie: Yes.swyx: And uh, well, I don't know if it's that simple, but is box going to have an opinion on that or you're just gonna be like, well we're just the sort of the, the source layer.Yeah. Let's Okta of zero handle that.Aaron Levie: I think we're gonna have an opinion and we will work with generally wherever the contours of the market end up. Um, and the reason that we're gonna have an opinion more than other topics probably is because one of the biggest use cases for why your agent might need it, an identity is for file system access.So thus we have to kind of think about this pretty deeply. And I think, uh, unless you're like in our world thinking about this particular problem all day long, it might be, you know, like, why is this such a big deal? And the reason why it's a really big deal is because sometimes sort of say, well just give the agent an, an account on the system and it just treats, treat it like every other type of user on the system.The [00:07:00] problem is, is that I as Aaron don't really have any responsibility over anybody else's box account in our organization. I can't see the box account of any other employee that I work with. I am not liable for anything that they do. And they have, I have, I have, you know, strict privacy requirements on everything that they're able to, you know, that, that, that they work on.Agents don't have that, you know, don't have those properties. The person who creates the agent probably is gonna, for the foreseeable future, take on a lot of the liability of what that agent does. That agent doesn't deserve any privacy because, because it's, you know, it can't fully be autonomously operated and it doesn't have any legal, you know, kind of, you know, responsibility.So thus you can't just be like, oh, well I'll just create a bunch of accounts and then I'll, I'll kind of work with that agent and I'll talk to it occasionally. Like you need oversight of that. And so then the question is, how do you have a world where the agent, sometimes you have oversight of, but what if that agent goes and works with other people?That person over there is collaborating with the agent on something you shouldn't have [00:08:00] access to what they're doing. So we have all of these new boundaries that we're gonna have to figure out of, of, you know, it's really, really easy. So far we've been in, in easy mode. We've hit the easy button with ai, which is the agent just is you.And when you're in quad code and you're in cursor, and you're in Codex, you're just, the agent is you. You're offing into your services. It can do everything you can do. That's the easy mode. The hard mode is agents are kind of running on their own. People check in with them occasionally, they're doing things autonomously.How do you give them access to resources in the enterprise and not dramatically increased the security risk and the risk that you might expose the wrong thing to somebody. These are all the new problems that we have to get solved. I like the identity layer and, and identity vendors as being a solution to that, but we'll, we'll need some opinions as well because so many of the use cases are these collaborative file system use cases, which is how do I give it an agent, a subset of my data?Give it its own workspace as well. ‘cause it's gonna need to store off its own information that would be relevant for it. And how do I have the right oversight into that? [00:09:00]Jeff Huber: One thing, which, um, I think is kind interesting, think about is that you know, how humans work, right? Like I may not also just like give you access to the whole file.I might like sit next to you and like scroll to this like one part of the file and just show you that like one part and like, you know,swyx: partial file access.Jeff Huber: I'm just saying I think like our, like RA does seem to be dead, right? Like you wanna say something is dead uhhuh probably RA is dead. And uh, like the auth story to me seems like incredibly unsolved and unaddressed by like the existing state of like AI vendors.ButAaron Levie: yeah, I think, um, we're, I mean you're taking obviously really to level limit that we probably need to solve for. Yeah. And we built an access control system that was, was kind of like, you know, its own little world for, for a long time. And um, and the idea was this, it's a many to many collaboration system where I can give you any part of the file system.And it's a waterfall model. So if I give you higher up in the, in the, in the system, you get everything below. And that, that kind of created immense flexibility because I can kind of point you to any layer in the, in the tree, but then you're gonna get access to everything kind of below it. And that [00:10:00] mostly is, is working in this, in this world.But you do have to manage this issue, which is how do I create an agent that has access to some of my stuff and somebody else's stuff as well. Mm-hmm. And which parts do I get to look at as the creator of the agent? And, and these are just brand new problems? Yeah. Crazy. And humans, when there was a human there that was really easy to do.Like, like if the three of us were all sharing, there'd be a Venn diagram where we'd have an overlapping set of things we've shared, but then we'd have our own ways that we shared with each other. In an agent world, somebody needs to take responsibility for what that agent has access to and what they're working on.These are like the, some of the most probably, you know, boring problems for 98% of people on, on the internet, but they will be the problems that are the difference between can you actually have autonomous agents in an enterprise contextswyx: Yeah.Aaron Levie: That are not leaking your data constantly.swyx: No. Like, I mean, you know, I run a very, very small company for my conference and like we already have data sensitivity issues.Yes. And some of my team members cannot see Yes. Uh, the others and like, I can't imagine what it's like to run a Fortune 500 and like, you have to [00:11:00] worry about this. I'm just kinda curious, like you, you talked to a lot like, like 70, 80% of your cus uh, of the Fortune 500, your customers.Aaron Levie: Yep. 67%. Just so we're being verySEswyx: precise.So Yeah. I'm notAaron Levie: Okay. Okay.swyx: Something I'm rounding up. Yes. Round up. I'm projecting to, forAaron Levie: the government.swyx: I'm projecting to the end of the year.Aaron Levie: Okay.swyx: There you go.Aaron Levie: You do make it sound like, like we, we, well we've gotta be on this. Like we're, we're taking way too long to get to 80%. Well,swyx: no, I mean, so like. How are they approaching it?Right? Because you're, you don't have a, you don't have a final answer yet.Why Coding Agents Took Off FirstAaron Levie: Well, okay, so, so this is actually, this is the stark reality that like, unfortunately is the kinda like pouring the water on the party a little bit.swyx: Yes.Aaron Levie: We all in Silicon Valley are like, have the absolute best conditions possible for AI ever.And I think we all saw the dke, you know, kind of Dario podcast and this idea of AI coding. Why is that taken off? And, and we're not yet fully seeing it everywhere else. Well, look, if you just like enumerated the list of properties that AI coding has and then compared it to other [00:12:00] knowledge work, let's just, let's just go through a few of them.Generally speaking, you bring on a new engineer, they have access to a large swath of the code base. Like, there's like very, like you, just, like new engineer comes on, they can just go and find the, the, the stuff that they, they need to work with. It's a fully text in text out. Medium. It's only, it's just gonna be text at the end of the day.So it's like really great from a, from just a, uh, you know, kinda what the agent can work with. Obviously the models are super trained on that dataset. The labs themselves have a really strong, kind of self-reinforcing positive flywheel of why they need to do, you know, agent coding deeply. So then you get just better tooling, better services.The actual developers of the AI are daily users of the, of the thing that they're we're working on versus like the, you know, probably there's only like seven Claude Cowork legal plugin users at Anthropic any given day, but there's like a couple thousand Claude code and you know, users every single day.So just like, think about which one are they getting more feedback on. All day long. So you just go through this list. You have a, you know, everybody who's a [00:13:00] developer by definition is technical so they can go install the latest thing. We're all generally online, or at least, you know, kinda the weird ones are, and we're all talking to each other, sharing best practices, like that's like already eight differences.Versus the rest of the economy. Every other part of the economy has like, like six to seven headwinds relative to that list. You go into a company, you're a banker in financial services, you have access to like a, a tiny little subset of the total data that's gonna be relevant to do your job. And you're have to start to go and talk to a bunch of people to get the right data to do your job because Sally didn't add you to that deal room, you know, folder.And that that, you know, the information is actually in a completely different organization that you now have to go in and, and sort of run into. And it's like you have this endless list of access controls and security. As, as you talked about, you have a medium, which is not, it's not just text, right? You have, you have a zoom call that, that you're getting all of the requirements from the customer.You have a lot of in-person conversations and you're doing in-person sales and like how do you ever [00:14:00] digitize all of that information? Um, you know, I think a lot of people got upset with this idea that the code base has all the context, um, that I don't know if you follow, you know, did you follow some of that conversation that that went viral?Is like, you know, it's not that simple that, that the code base doesn't have all the knowledge, but like it's a lot, you're a lot better off than you are with other areas of knowledge work. Like you, we like, we like have documentation practices, you write specifications. Those things don't exist for like 80% of work that happens in the enterprise.That's the divide that we have, which is, which is AI coding has, has just fully, you know, where we've reached escape velocity of how powerful this stuff is, and then we're gonna have to find a way to bring that same energy and momentum, but to all these other areas of knowledge work. Where the tools aren't there, the data's not set up to be there.The access controls don't make it that easy. The context engineering is an incredibly hard problem because again, you have access control challenges, you have different data formats. You have end users that are gonna need to kind of be kind of trained through this as opposed to their adopting [00:15:00] these tools in their free time.That's where the Fortune 500 is. And so we, I think, you know, have to be prepared as an industry where we are gonna be on a multi-year march to, to be able to bring agents to the enterprise for these workflows. And I think probably the, the thing that we've learned most in coding that, that the rest of the world is not yet, I think ready for, I mean, we're, they'll, they'll have to be ready for it because it's just gonna inevitably happen is I think in coding.What, what's interesting is if you think about the practice of coding today versus two years ago. It's probably the most changed workflow in maybe the history of time from the amount of time it's changed, right? Yeah. Like, like has any, has any workflow in the entire economy changed that quickly in terms of the amount of change?I just, you know, at least in any knowledge worker workflow, there's like very rarely been an event where one piece of technology and work practice has so fundamentally, you know, changed, changed what you do. Like you don't write code, you talk to an agent and it goes and [00:16:00] does it for you, and you may be at best review it.And even that's even probably like, like largely not even what you're doing. What's happening is we are changing our work to make the agents effective. In that model, the agent didn't really adapt to how we work. We basically adapted to how the agent works. Mm-hmm. All of the economy has to go through that exact same evolution.The rest of the economy is gonna have to update its workflows to make agents effective. And to give agents the context that they need and to actually figure out what kind of prompting works and to figure out how do you ensure that the agent has the right access to information to be able to execute on its work.I, you know, this is not the panacea that people were hoping for, of the agent drops in, just automates your life. Like you have to basically re-engineer your workflow to get the most out of agents and, uh, and that, that's just gonna take, you know, multiple years across the economy. Right now it's a huge asset and an advantage for the teams that do it early and that are kinda wired into doing this.‘cause [00:17:00] you'll see compounding returns, but that's just gonna take a while for most companies to actually go and get this deployed.swyx: I love, I love pushing back. I think that. That is what a lot of technology consultants love to hear this sort of thing, right? Yeah, yeah, yeah. First to, to embrace the ai. Yes. To get to the promised land, you must pay me so much money to a hundred percent to adopt the prescribed way of, uh, conforming to the agents.Yes. And I worry that you will be eclipsed by someone else who says, no, come as you are.Aaron Levie: Yeah.swyx: And we'll meet you where you are.Aaron Levie: And, and, and and what was the thing that went viral a week ago? OpenAI probably, uh, is hiring F Dees. Yeah. Uh, to go into the enterprise. Yeah. Yeah. And then philanthropic is embedded at Goldman Sachs.Yeah. So if the labs are having to do this, if, if the labs have decided that they need to hire FDE and professional services, then I think that's a pretty clear indication that this, there's no easy mode of workflow transformation. Yeah. Yeah. So, so to your point, I think actually this is a market opportunity for, you know, new professional services and consulting [00:18:00] firms that are like Agent Build and they, and they kind of, you know, go into organizations and they figure out how to re-engineer your workflows to make them more agent ready and get your data into the right format and, you know, reconstruct your business process.So you're, you're not doing most of the work. You're telling agents how to do the work and then you're reviewing it. But I haven't seen the thing that can just drop in and, and kinda let you not go through those changes.swyx: I don't know how that kind of sales pitch goes over. Yeah. You know, you're, you're saying things like, well, in my sort of nice beautiful walled garden, here's, there's, uh, because here's this, here's this beautiful box account that has everything.Yes. And I'm like, well, most, most real life is extremely messy. Sure. And like, poorly named and there duplicate this outdated s**tAaron Levie: a hundred percent. And so No, no, a hundred percent. And so this is actually No. So, so this is, I mean, we agree that, that getting to the beautiful garden is gonna be tough.swyx: Yeah.Aaron Levie: There's also the other end of the spectrum where I, I just like, it's a technical impossibility to solve. The agent is, is truly cannot get enough context to make the right decision in, in the, in the incredibly messy land. Like there's [00:19:00] no a GI that will solve that. So, so we're gonna have to kind of land in somewhere in between, which is like we all collectively get better at.Documentation practices and, and having authoritative relatively up-to-date information and putting it in the right place like agents will, will certainly cause us to be much better organized around how we work with our information, simply because the severity of the agent pulling the wrong data will be too high and the productivity gain of that you'll miss out on by not doing this will be too high as well, that you, that your competition will just do it and they'll just have higher velocity.So, uh, and, and we, we see this a lot firsthand. So we, we build a series of agents internally that they can kind of have access to your full box account and go off and you give it a task and it can go find whatever information you're looking for and work with. And, you know, thank God for the model progress, but like, if, if you gave that task to an agent.Nine months ago, you're just gonna get lots of bogus answers because it's gonna, it's gonna say, Hey, here's, here are fi [00:20:00] five, you know, documents that all kind of smell like the right thing. And I'm gonna, but I, but you're, you're putting me on the clock. ‘cause my assistant prompt says like, you know, be pretty smart, but also try and respond to the user and it's gonna respond.And it's like, ah, it got the wrong document. And then you do that once or twice as a knowledge worker and you're just neverswyx: again,Aaron Levie: never again. You're just like done with the system.swyx: Yeah. It doesn't work.Aaron Levie: It doesn't work. And so, you know, Opus four six and Gemini three one Pro and you know, whatever the latest five 3G BT will be, like, those things are getting better and better and it's using better judgment.And this sort of like the, all of these updates to the agentic tool and search systems are, are, we're seeing, we're seeing very real progress where the agent. Kind of can, can almost smell some things a little bit fishy when it's getting, you know, we, we have this process where we, we have it go fan out, do a bunch of searches, pull up a bunch of data, and then it has to sort of do its own ranking of, you know, what are the right documents that, that it should be working with.And again, like, you know, the intelligence level of a model six months ago, [00:21:00] it'd be just throwing a dart at like, I'm just, I'm gonna grab these seven files and I, I pray, I hope that that's the right answer. And something like an opus first four five, and now four six is like, oh, it's like, no, that one doesn't seem right relative to this question because I'm seeing some signal that is making that, you know, that's contradicting the document where it would normally be in the tree and who should have access.Like it's doing all of that kind of work for you. But like, it still doesn't work if you just have a total wasteland of data. Like, it's just not, it's just not possible. Partly ‘cause a human wouldn't even be able to do it. So basically if a, if a really, really smart human. Could not do that task in five or 10 minutes for a search retrieval type task.Look, you know, your agent's not gonna be able to do it any better. You see this all day long. SoContext Engineering and Search Limitsswyx: this touches on a thing that just passionate about it was just context engineering. I, I'm just gonna let you ramble or riff on, on context engineering. If, if, if there's anything like he, he did really good work on context fraud, which has really taken over as like the term that people use and the referenceAaron Levie: a hundred percent.We, we all we think about is, is the context rob problem. [00:22:00]Jeff Huber: Yeah, there's certainly a lot of like ranking considerations. Gentech surgery think is incredibly promising. Um, yeah, I was trying to generate a question though. I think I have a question right now. Swyx.Aaron Levie: Yeah, no, but like, like I think there was this moment, um, you know, like, I don't know, two years ago before, before we knew like where the, the gotchas were gonna be in ai and I think someone was like, was like, well, infinite context windows will just solve all of these problems and ‘cause you'll just, you'll just give the context window like all the data and.It's just like, okay, I mean, maybe in 2035, like this is a viable solution. First of all, it, it would just, it would just simply cost too much. Like we just can't give the model like the 5,000 documents that might be relevant and it's gonna read them all. And I've seen enough to, to start believing in crazy stuff.So like, I'm willing to just say, sure. Like in, in 10 years from now,swyx: never say, never, never.Aaron Levie: In, in 10 years from now, we'll have infinite context windows at, at a thousandth of the price of today. Like, let's just like believe that that's possible, but Right. We're in reality today. So today we have a context engineering [00:23:00] problem, which is, I got, I got, you know, 200,000 tokens that I can work with, or prob, I don't even know what the latest graph is before, like massive degradation.16. Okay. I have 60,000 tokens that I get to work with where I'm gonna get accurate information. That's not a lot of tokens for a corpus of 10 million documents that a knowledge worker might have across all of the teams and all the projects and all the people they work with. I have, I have 10 million documents.Which, you know, maybe is times five pages per document or something like that. I'm at 50 million pages of information and I have 60,000 tokens. Like, holy s**t. Yeah. This is like, how do I bridge the 50 million pages of information with, you know, the couple hundred that I get to work with in that, in that token window.Yeah. This is like, this is like such an interesting problem and that's why actually so much work is actually like, just like search systems and the databases and that layer has to just get so locked in, but models getting better and importantly [00:24:00] knowing when they've done a search, they found the wrong thing, they go back, they check their work, they, they find a way to balance sort of appeasing the user versus double checking.We have this one, we have this one test case where we ask the agent to go find. 10 pieces of information.swyx: Is this the complex work eval?Aaron Levie: Uh, this is actually not in the eval. This is, this is sort of just like we have a bunch of different, we have a bunch of internal benchmark kind of scenarios. Every time we, we update our agent, we have one, which is, I ask it to find all of our office addresses, and I give it the list of 10 offices that we have.And there's not one document that has this, maybe there should be, that would be a great example of the kind of thing that like maybe over time companies start to, you know, have these sort of like, what are the canonical, you know, kind of key areas of knowledge that we need to have. We don't seem to have this one document that says, here are all of our offices.We have a bunch of documents that have like, here's the New York office and whatever. So you task this agent and you, you get, you say, I need the addresses for these 10 offices. Okay. And by the way, if you do this on any, you know, [00:25:00] public chat model, the same outcome is gonna happen. But for a different kind of query, you give it, you say, I need these 10 addresses.How many times should the agent go and do its search before it decides whether or not, there's just no answer to this question. Often, and especially the, the, let's say lower tier models, it'll come back and it'll give you six of the 10 addresses. And it'll, and I'll just say I couldn't find the otherswyx: four.It, it doesn't know what It doesn't know. ItAaron Levie: doesn't know what It doesn't know. Yeah. So the model is just like, like when should it stop? When should it stop doing? Like should it, should it do that task for literally an hour and just keep cranking through? Maybe I actually made up an office location and it doesn't know that I made it up and I didn't even know that I made it up.Like, should it just keep, re should it read every single file in your entire box account until it, until it should exhaust every single piece of information.swyx: Expensive.Aaron Levie: These are the new problems that we have. So, you know, something like, let's say a new opus model is sort of like, okay, I'm gonna try these types of queries.I didn't get exactly what I wanted. I'm gonna try again. I'm gonna, at [00:26:00] some point I'm gonna stop searching. ‘cause I've determined that that no amount of searching is gonna solve this problem. I'm just not able to do it. And that judgment is like a really new thing that the model needs to be able to have.It's like, when should it give up on a task? ‘cause, ‘cause you just don't, it's a can't find the thing. That's the real world of knowledge, work problems. And this is the stuff that the coding agents don't have to deal with. Because they, it just doesn't like, like you're not usually asking it about, you're, you're always creating net new information coming right outta the model for the most part.Obviously it has to know about your code base and your specs and your documentation, but, but when you deploy an agent on all of your data that now you have all of these new problems that you're dealing withJeff Huber: our, uh, follow follow-up research to context ride is actually on a genetic search. Ah. Um, and we've like right, sort of stress tested like frontier models and their ability to search.Um, and they're not actually that good at searching. Right. Uh, so you're sort of highlighting this like explore, exploit.swyx: You're just say, Debbie, Donna say everything doesn't work. Like,Aaron Levie: well,Jeff Huber: somebody has to be,Aaron Levie: um, can I just throw out one more thing? Yeah. That is different from coding and, and the rest [00:27:00] of the knowledge work that I, I failed to mention.So one other kind of key point is, is that, you know, at the end of the day. Whether you believe we're in a slop apocalypse or, or whatever. At the end of the day, if you, if you build a working product at the end of, if you, if you've built a working solution that is ultimately what the customer is paying for, like whether I have a lot of slop, a little slop or whatever, I'm sure there's lots of code bases we could go into in enterprise software companies where it's like just crazy slop that humans did over a 20 year period, but the end customer just gets this little interface.They can, they can type into it, it does its thing. Knowledge work, uh, doesn't have that property. If I have an AI model, go generate a contract and I generate a contract 20 times and, you know, all 20 times it's just 3% different and like that I, that, that kind of lop introduces all new kinds of risk for my organization that the code version of that LOP didn't, didn't introduce.These are, and so like, so how do you constrain these models to just the part that you want [00:28:00] them to work on and just do the thing that you want them to do? And, and, you know, in engineering, we don't, you can't be disbarred as an engineer, but you could be disbarred as a lawyer. Like you can do the wrong medical thing In healthcare, you, there's no, there's no equivalent to that of engineering.Like, doswyx: you want there to be, because I've considered softwareJeff Huber: engineer. What's that? Civil engineering there is, right? NotAaron Levie: software civil engineer. Sure. Oh yeah, for sure. But like in any of our companies, you like, you know, you'll be forgiven if you took down the site and, and we, we will do a rollback and you'll, you'll be in a meeting, but you have not been disbarred as an engineer.We don't, we don't change your, you know, your computer science, uh, blameJeff Huber: degree, this postmortem.Aaron Levie: Yeah, exactly. Exactly. So, so, uh, now maybe we collectively as an industry need to figure out like, what are you liable for? Not legally, but like in a, in a management sense, uh, of these agents. All sorts of interesting problems that, that, that, uh, that have to come out.But in knowledge work, that's the real hostile environments that we're operating in. Hmm.swyx: I do think like, uh, a lot of the last year's, 2025 story was the rise of coding agents and I think [00:29:00] 2026 story is definitely knowledge work agents. Yes. A hundredAaron Levie: percent.swyx: Right. Like that would, and I think open claw core work are just the beginning.Yes. Like it's, the next one's gonna just gonna be absolute craziness.Aaron Levie: It it is. And, and, uh, and it's gonna be, I mean, again, like this is gonna be this, this wave where we, we are gonna try and bring as many of the practices from coding because that, that will clearly be the forefront, which is tell an agent to go do something and has an access to a set of resources.You need to be responsible for reviewing it at the end of the process. That to me is the, is the kind of template that I just think goes across knowledge, work and odd. Cowork is a great example. Open Closet's a great example. You can kind of, sort of see what Codex could become over time. These are some, some really interesting kind of platforms that are emerging.swyx: Okay. Um, I wanted to, we touched on evals a little bit. You had, you had the report that you're gonna go bring up and then I was gonna go into like, uh, boxes, evals, but uh, go ahead. Talk about your genetic search thing.Jeff Huber: Yeah. Mostly I think kinda a few of the insights. It's like number one frontier model is not good at search.Humans have this [00:30:00] natural explore, exploit trade off where we kinda understand like when to stop doing something. Also, humans are pretty good at like forgetting actually, and like pruning their own context, whereas agents are not, and actually an agent in their kind of context history, if they knew something was bad and they even, you could see in the trace the reason you trace, Hey, that probably wasn't a good idea.If it's still in the trace, still in the context, they'll still do it again. Uhhuh. Uh, and so like, I think pruning is also gonna be like, really, it's already becoming a thing, right? But like, letting self prune the con windowsswyx: be a big deal. Yeah. So, so don't leave the mistake. Don't leave the mistake in there.Cut out the mistake but tell it that you made a mistake in the past and so it doesn't repeat it.Jeff Huber: Yeah. But like cut it out so it doesn't get like distracted by it again. ‘cause really, you know, what is so, so it will repeat its mistake just because it's been, it's inswyx: theJeff Huber: context. It'sAaron Levie: in the context so much.That's a few shot example. Even if it, yeah.Jeff Huber: It's like oh thisAaron Levie: is a great thing to go try even ifJeff Huber: it didn't work.Aaron Levie: Yeah,Jeff Huber: exactly.Aaron Levie: SoJeff Huber: there's like a bunch of stuff there. JustAaron Levie: Groundhogs Day inside these models. Yeah. I'm gonna go keep doing the same wrongJeff Huber: thing. Covering sense. I feel like, you know, some creator analogy you're trying like fit a manifold in latent space, which kind is doing break program synthesis, which is kinda one we think about we're doing right.Like, you know, certain [00:31:00] facts might be like sort of overly pitting it. There are certain, you know, sec sectors of latent space and so like plug clean space. Yeah. And, uh, andswyx: so we have a bell, our editor as a bell every time you say that. SoJeff Huber: you have, you have to like remove those, likeswyx: you shoulda a gong like TPN or something.IfJeff Huber: we gong, you either remove those links to like kinda give it the freedom, kind of do what you need to do. So, but yeah. We'll, we'll release more soon. That'sAaron Levie: awesome.Jeff Huber: That'll, that'll be cool.swyx: We're a cerebral podcast that people listen to us and, and sort of think really deep. So yeah, we try to keep it subtle.Okay. We try to keep it.Aaron Levie: Okay, fine.Inside Agent Evalsswyx: Um, you, you guys do, you guys do have EVs, you talked about your, your office thing, but, uh, you've been also promoting APEX agents and complex work. Uh, yeah, whatever you, wherever you wanna take this just Yeah. How youAaron Levie: Apex is, is obviously me, core's, uh, uh, kind of, um, agent eval.We, we supported that by sort of. Opening up some data for them around how we kind of see these, um, data workspaces in, in the, you know, kind of regular economy. So how do lawyers have a workspace? How do investment bankers have a workspace? What kind of data goes into those? And so we, [00:32:00] we partner with them on their, their apex eval.Our own, um, eval is, it's actually relatively straightforward. We have a, a set of, of documents in a, in a range of industries. We give the agent previously did this as a one shot test of just purely the model. And then we just realized we, we need to, based on where everything's going, it's just gotta be more agentic.So now it's a bit more of a test of both our harness and the model. And we have a rubric of a set of things that has to get right and we score it. Um, and you're just seeing, you know, these incredible jumps in almost every single model in its own family of, you know, opus four, um, you know, sonnet four six versus sonnet four five.swyx: Yeah. We have this up on screen.Aaron Levie: Okay, cool. So some, you're seeing it somewhere like. I, I forget the to, it was like 15 point jump, I think on the main, on the overall,swyx: yes.Aaron Levie: And it's just like, you know, these incredible leaps that, that are starting to happen. Um,swyx: and OP doesn't know any, like any, it's completely held out from op.Aaron Levie: This is not in any, there's no public data which has, you know, Ben benefits and this is just a private eval that we [00:33:00] do, and then we just happen to show it to, to the world. Hmm. So you can't, you can't train against it. And I think it's just as representative of. It's obviously reasoning capabilities, what it's doing at, at, you know, kind of test time, compute capabilities, thinking levels, all like the context rot issues.So many interesting, you know, kind of, uh, uh, capabilities that are, that are now improvingswyx: one sector that you have. That's interesting.Industries and Datasetsswyx: Uh, people are roughly familiar with healthcare and legal, but you have public sector in there.Aaron Levie: Yeah.swyx: Uh, what's that? Like, what, what, what is that?Aaron Levie: Yeah, and, and we actually test against, I dunno, maybe 10 industries.We, we end up usually just cutting a few that we think have interesting gains. All extras, won a lot of like government type documents. Um,swyx: what is that? What is it? Government type documents?Aaron Levie: Government filings. Like a taxswyx: return, likeAaron Levie: a probably not tax returns. It would be more of what would go the government be using, uh, as data.So, okay. Um, so think about research that, that type of, of, of data sets. And then we have financial services for things like data rooms and what would be in an investment prospectus. Uhhuh,swyx: that one you can dog food.Aaron Levie: Yeah, exactly. Exactly. Yes. Yes. [00:34:00] So, uh, so we, we run the models, um, in now, you know, more of an agent mode, but, but still with, with kinda limited capacity and just try and see like on a, like, for like basis, what are the improvements?And, and again, we just continue to be blown away by. How, how good these models are getting.swyx: Yeah, I mean, I think every serious AI company needs something like that where like, well, this is the work we do. Here's our company eval. Yeah. And if you don't have it, well, you're not a serious AI company.Aaron Levie: There's two dimensions, right?So there's, there's like, how are the models improving? And so which models should you either recommend a customer use, which one should you adopt? But then every single day, we're making changes to our agents. And you need to knowswyx: if you regressed,Aaron Levie: if you know. Yeah. You know, I've been fully convinced that the whole agent observability and eval space is gonna be a massive space.Um, super excited for what Braintrust is doing, excited for, you know, Lang Smith, all the things. And I think what you're going to, I mean, this is like every enter like literally every enterprise right now. It's like the AI companies are the customers of these tools. Every enterprise will have this. Yeah, you'll just [00:35:00] have to have an eval.Of all of your work and like, we'll, you'll have an eval of your RFP generation, you'll have an eval of your sales material creation. You'll have an eval of your, uh, invoice processing. And, and as you, you know, buy or use new agentic systems, you are gonna need to know like, what's the quality of your, of your pipeline.swyx: Yeah.Aaron Levie: Um, so huge, huge market with agent evals.swyx: Yeah.Building the Agent Teamswyx: And, and you know, I'm gonna shout out your, your team a bit, uh, your CTO, Ben, uh, did a great talk with us last year. Awesome. And he's gonna come back again. Oh, cool. For World's Fair.Aaron Levie: Yep.swyx: Just talk about your team, like brag a little bit. I think I, I think people take these eval numbers in pretty charts for granted, but No, there, I mean, there's, there's lots of really smart people at work during all this.Aaron Levie: Biggest shout out, uh, is we have a, we have a couple folks at Dya, uh, Sidarth, uh, that, that kind of run this. They're like a, you know, kind of tag tag team duo on our evals, Ben, our CTO, heavily involved Yasha, head of ai, uh, you know, a bunch of folks. And, um, evals is one part of the story. And then just like the full, you know, kind of AI.An agent team [00:36:00] is, uh, is a, is a pretty, you know, is core to this whole effort. So there's probably, I don't know, like maybe a few dozen people that are like the epicenter. And then you just have like layers and layers of, of kind of concentric circles of okay, then there's a search team that supports them and an infrastructure team that supports them.And it's starting to ripple through the entire company. But there's that kind of core agent team, um, that's a pretty, pretty close, uh, close knit group.swyx: The search team is separate from the infra team.Aaron Levie: I mean, we have like every, every layer of the stack we have to kind of do, except for just pure public cloud.Um, but um, you know, we, we store, I don't even know what our public numbers are in, you know, but like, you can just think about it as like a lot of data is, is stored in box. And so we have, and you have every layer of the, of the stack of, you know, how do you manage the data, the file system, the metadata system, the search system, just all of those components.And then they all are having to understand that now you've got this new customer. Which is the agent, and they've been building for two types of customers in the past. They've been building for users and they've been building for like applications. [00:37:00] And now you've got this new agent user, and it comes in with a difference of it, of property sometimes, like, hey, maybe sometimes we should do embeddings, an embedding based, you know, kind of search versus, you know, your, your typical semantic search.Like, it's just like you have to build the, the capabilities to support all of this. And we're testing stuff, throwing things away, something doesn't work and, and not relevant. It's like just, you know, total chaos. But all of those teams are supporting the agent team that is kind of coming up with its requirements of what, what do we need?swyx: Yeah. No, uh, we just came from, uh, fireside chat where you did, and you, you talked about how you're doing this. It's, it's kind of like an internal startup. Yeah. Within the broader company. The broader company's like 3000 people. Yeah. But you know, there's, there's a, this is a core team of like, well, here's the innovation center.Aaron Levie: Yeah.swyx: And like that every company kind of is run this way.Aaron Levie: Yeah. I wanna be sensitive. I don't call it the innovation center. Yeah. Only because I think everybody has to do innovation. Um, there, there's a part of the, the, the company that is, is sort of do or die for the agent wave.swyx: Yeah.Aaron Levie: And it only happens to be more of my focus simply because it's existential that [00:38:00] we get it right.swyx: Yeah.Aaron Levie: All of the supporting systems are necessary. All of the surrounding adjacent capabilities are necessary. Like the only reason we get to be a platform where you'd run an agent is because we have a security feature or a compliance feature, or a governance feature that, that some team is working on.But that's not gonna be the make or break of, of whether we get agents right. Like that already exists and we need to keep innovating there. I don't know what the right, exact precise number is, but it's not a thousand people and it's not 10 people. There's a number of people that are like the, the kind of like, you know, startup within the company that are the make or break on everything related to AI agents, you know, leveraging our platform and letting you work with your data.And that's where I spend a lot of my time, and Ben and Yosh and Diego and Teri, you know, these are just, you know, people that, that, you know, kind of across the team. Are working.swyx: Yeah. Amazing.Read Write Agent WorkflowsJeff Huber: How do you, how do you think about, I mean, you talked a lot about like kinda read workflows over your box data. Yep.Right. You know, gen search questions, queries, et cetera. But like, what about like, write or like authoring workflows?Aaron Levie: Yes. I've [00:39:00] already probably revealed too much actually now that I think about it. So, um, I've talked about whatever,Jeff Huber: whatever you can.Aaron Levie: Okay. It's just us. It's just us. Yeah. Okay. Of course, of course.So I, I guess I would just, uh, I'll make it a little bit conceptual, uh, because again, I've already, I've already said things that are not even ga but, but we've, we've kinda like danced around it publicly, so I, yeah, yeah. Okay. Just like, hopefully nobody watches this, um, episode. No.swyx: It's tidbits for the Heidi engaged to go figure out like what exactly, um, you know, is, is your sort of line of thinking.Sure. They can connect the dots.Aaron Levie: Yeah. So, so I would say that, that, uh, we, you know, as a, as a place where you have your enterprise content, there's a use case where I want to, you know, have an agent read that data and answer questions for me. And then there's a use case where I want the agent to create something.And use the file system to create something or store off data that it's working on, or be able to have, you know, various files that it's writing to about the work it's doing. So we do see it as a total read write. The harder problem has so far been the read only because, because again, you have that kind of like 10 [00:40:00] million to one ratio problem, whereas rights are a lot of, that's just gonna come from the model and, and we just like, we'll just put it in the file system and kinda use it.So it's a little bit of a technically easier problem, but the only part that's like, not necessarily technically hard, it is just like it's not yet perfected in the state of the ecosystem is, you know, building a beautiful PowerPoint presentation. It's still a hard problem for these models. Like, like we still, you know, like, like these formats are just, we're not built for.They'reswyx: working on it.Aaron Levie: They're, they're working on it. Everybody's working on it.swyx: Every launch is like, well, we do PowerPoint now.Aaron Levie: We're getting, yeah, getting a lot, getting a lot of better each time. But then you'll do this thing where you'll ask the update one slide and all of a sudden, like the fonts will be just like a little bit different, you know, on two of the slides, or it moved, you know, some shape over to the left a little bit.And again, these are the kind of things that, like in code, obviously you could really care about if you really care about, you know, how beautiful is the code, but at the end, user doesn't notice all those problems and file creation, the end user instantly sees it. You're [00:41:00] like, ah, like paragraph three, like, you literally just changed the font on me.Like it's a totally different font and like midway through the document. Mm-hmm. Those are the kind of things that you run into a lot of in the, in the content creation side. So, mm-hmm. We are gonna have native agents. That do all of those things, they'll be powered by the leading kind of models and labs.But the thing that I think is, is probably gonna be a much bigger idea over time is any agent on any system, again, using Box as a file system for its work, and in that kind of scenario, we don't necessarily care what it's putting in the file system. It could put its memory files, it could put its, you know, specification, you know, documents.It could put, you know, whatever its markdown files are, or it could, you know, generate PDFs. It's just like, it's a workspace that is, is sort of sandboxed off for its work. People can collaborate into it, it can share with other people. And, and so we, we were thinking a lot about what's the right, you know, kind of way to, to deliver that at scale.Docs Graphs and Founder Modeswyx: I wanted to come into sort of the sort of AI transformation or AI sort of, uh, operations things. [00:42:00] Um, one of the tweets that you, that you wanted to talk about, this is just me going through your tweets, by the way. Oh, okay. I mean, like, this is, you readAaron Levie: one by one,swyx: you're the, you're the easiest guest to prep for because you, you already have like, this is the, this is what I'm interested in.I'm like, okay, well, areAaron Levie: we gonna get to like, like February, January or something? Where are we in the, in the timelines? How far back are we going?swyx: Can you, can you describe boxes? A set of skills? Right? Like that, that's like, that's like one of the extremes of like, well if you, you just turn everything into a markdown file.Yeah. Then your agent can run your company. Uh, like you just have to write, find the right sequence of words toAaron Levie: Yes.swyx: To do it.Aaron Levie: Sorry, isthatswyx: the question? So I think the question is like, what if we documented everything? Yes. The way that you exactly said like,Aaron Levie: yes.swyx: Um, let's get all the Fortune five hundreds, uh, prepared for agents.Yes. And like, you know, everything's in golden and, and nicely filed away and everything. Yes. What's missing? Like, what's left, right? LikeAaron Levie: Yeah.swyx: You've, you've run your company for a decade. LikeAaron Levie: Yeah. I think the challenge is that, that that information changes a week later. And because something happened in the market for that [00:43:00] customer, or us as a company that now has to go get updated, and so these systems are living and breathing and they have to experience reality and updates to reality, which right now is probably gonna be humans, you know, kinda giving those, giving them the updates.And, you know, there is this piece about context graphs as as, uh, that kinda went very viral. Yeah. And I, I, I was like a, i, I, I thought it was super provocative. I agreed with many parts of it. I disagree with a few parts around. You know, it's not gonna be as easy as as just if we just had the agent traces, then we can finally do that work because there's just like, there's so much more other stuff that that's happening that, that we haven't been able to capture and digitize.And I think they actually represented that in the piece to be clear. But like there's just a lot of work, you know, that that has to, you just can't have only skills files, you know, for your company because it's just gonna be like, there's gonna be a lot of other stuff that happens. Yeah. Change over time.Yeah. Most companies are practically apprenticeships.swyx: Most companies are practically apprenticeships. LikeJeff Huber: every new employee who joins the team, [00:44:00] like you span one to three months. Like ramping them up.Aaron Levie: Yes. AllJeff Huber: that tat knowledgeAaron Levie: isJeff Huber: not written down.Aaron Levie: Yes.Jeff Huber: But like, it would have to be if you wanted to like give it to an Asian.Right. And so like that seems to me like to beAaron Levie: one is I think you're gonna see again a premium on companies that can document this. Mm-hmm. Much. There'll be a huge premium on that because, because you know, can you shorten that three month ramp cycle to a two week ramp cycle? That's an instant productivity gain.Can you re dramatically reduce rework in the organization because you've documented where all the stuff is and where the answers are. Can you make your average employee as good as your 90th percentile employee because you've captured the knowledge that's sort of in the heads of, of those top employees and make that available.So like you can see some very clear productivity benefits. Mm-hmm. If you had a company culture of making sure you know your information was captured, digitized, put in a format that was agent ready and then made available to agents to work with, and then you just, again, have this reality of like add a 10,000 person [00:45:00] company.Mapping that to the, you know, access structure of the company is just a hard problem. Is like, is like, yeah, well, you just, not every piece of information that's digitized can be shared to everybody. And so now you have to organize that in a way that actually works. There was a pretty good piece, um, this, this, uh, this piece called your company as a file is a file system.I, did you see that one?swyx: Nope.Aaron Levie: Uh, yes. You saw it. Yeah. And, and, uh, I actually be curious your thoughts on it. Um, like, like an interesting kind of like, we, we agree with it because, because that's how we see the world and, uh,swyx: okay. We, we have it up on screen. Oh,Aaron Levie: okay. Yeah. But, but it's all about basically like, you know, we've already, we, we, we already organized in this kind of like, you know, permission structure way.Uh, and, and these are the kind of, you know, natural ways that, that agents can now work with data. So it's kind of like this, this, you know, kind of interesting metaphor, but I do think companies will have to start to think about how they start to digitize more, more of that data. What was your take?Jeff Huber: Yeah, I mean, like the company's probably like an acid compliant file system.Aaron Levie: Uh,Jeff Huber: yeah. Which I'm guessing boxes, right? So, yeah. Yes.swyx: Yeah. [00:46:00]Jeff Huber: Which you have a great piece on, but,swyx: uh, yeah. Well, uh, I, I, my, my, my direction is a little bit like, I wanna rewind a little bit to the graph word you said that there, that's a magic trigger word for us. I always ask what's your take on knowledge graphs?Yeah. Uh, ‘cause every, especially at every data database person, I just wanna see what they think. There's been knowledge graphs, hype cycles, and you've seen it all. So.Aaron Levie: Hmm. I actually am not the expert in knowledge graphs, so, so that you might need toswyx: research, you don't need to be an expert. Yeah. I think it's just like, well, how, how seriously do people take it?Yeah. Like, is is, is there a lot of potential in the, in the HOVI?Aaron Levie: Uh, well, can I, can I, uh, understand first if it's, um, is this a loaded question in the sense of are you super pro, super con, super anti medium? Iswyx: see pro, I see pros and cons. Okay. Uh, but I, I think your opinion should be independent of mine.Aaron Levie: Yeah. No, no, totally. Yeah. I just want to see what I'm stepping into.swyx: No, I know. It's a, and it's a huge trigger word for a lot of people out Yeah. In our audience. And they're, they're trying to figure out why is that? Because whyAaron Levie: is this such aswyx: hot item for them? Because a lot of people get graph religion.And they're like, everything's a graph. Of course you have to represent it as a graph. Well, [00:47:00] how do you solve your knowledge? Um, changing over time? Well, it's a graph.Aaron Levie: Yeah.swyx: And, and I think there, there's that line of work and then there's, there's a lot of people who are like, well, you don't need it. And both are right.Aaron Levie: Yeah. And what do the people who say you don't need it, what are theyswyx: arguing for Mark down files. Oh, sure, sure. Simplicity.Aaron Levie: Yeah.swyx: Versus it's, it's structure versus less structure. Right. That's, that's all what it is. I do.Aaron Levie: I think the tricky thing is, um, is, is again, when this gets met with real humans, they're just going to their computer.They're just working with some people on Slack or teams. They're just sharing some data through a collaborative file system and Google Docs or Box or whatever. I certainly like the vision of most, most knowledge graph, you know, kind of futuristic kind of ways of thinking about it. Uh, it's just like, you know, it's 2026.We haven't seen it yet. Kind of play out as as, I mean, I remember. Do you remember the, um, in like, actually I don't, I don't even know how old you guys are, but I'll for, for to show my age. I remember 17 years ago, everybody thought enterprises would just run on [00:48:00] Wikis. Yeah. And, uh, confluence and, and not even, I mean, confluence actually took off for engineering for sure.Like unquestionably. But like, this was like everything would be in the w. And I think based on our, uh, our, uh, general style of, of, of what we were building, like we were just like, I don't know, people just like wanna workspace. They're gonna collaborate with other people.swyx: Exactly. Yeah. So you were, you were anti-knowledge graph.Aaron Levie: Not anti, not anti. Soswyx: not nonAaron Levie: I'm not, I'm not anti. ‘cause I think, I think your search system, I just think these are two systems that probably, but like, I'm, I'm not in any religious war. I don't want to be in anybody's YouTube comments on this. There's not a fight for me.swyx: We, we love YouTube comments. We're, we're, we're get into comments.Aaron Levie: Okay. Uh, but like, but I, I, it's mostly just a virtue of what we built. Yeah. And we just continued down that path. Yeah.swyx: Yeah.Aaron Levie: And, um, and that, that was what we pursued. But I'm not, this is not a, you know, kind of, this is not a, uh, it'sswyx: not existential for you. Great.Aaron Levie: We're happy to plug into somebody else's graph.We're happy to feed data into it. We're happy for [00:49:00] agents to, to talk to multiple systems. Not, not our fight.swyx: Yeah.Aaron Levie: But I need your answer. Yeah. Graphs or nerd Snipes is very effective nerd.swyx: See this is, this is one, one opinion and then I've,Jeff Huber: and I think that the actual graph structure is emergent in the mind of the agent.Ah, in the same way it is in the mind of the human. And that's a more powerful graph ‘cause it actually involved over time.swyx: So don't tell me how to graph. I'll, I'll figure it out myself. Exactly. Okay. All right. AndJeff Huber: what's yours?swyx: I like the, the Wiki approach. Uh, my, I'm actually

This is a free preview of a paid episode. To hear more, visit www.latent.spaceAIE Europe CFP and AIE World's Fair paper submissions for CAIS peer review are due TODAY - do not delay! Last call ever.We're excited to welcome METR for their first LS Pod, hopefully the first of many:METR are keepers of currently the single most infamous chart in AI:But every Latent Space reader should be sophisticated enough to know that the details matter and that hype and hyperbole go hand in hand in AI social media, because the millions of impressions that got, by people who don't understand or care about the nuances, disclaimers, and error bars, far outreaches the 69k views on the corrections by the people who actually made the chart:There's a lot of nuance both in making benchmarks (as we discovered with OpenAI on our SWE-Bench Verified podcast) and in extrapolating results from them, especially where exponentials and sigmoids are concerned. METR's Long Horizons work itself has known biases that the authors have responsibly disclosed, but go far too underappreciated in the pursuit of doomer chart porn.If you're interested in a short, sharable TED talk version of this pod, over at AIE CODE we were blessed to feature Joel twice, as a stage talk and with a longer form small workshop with Q&A:We also make sure cover some of METR's lesser known work on Threat Evaluation but also Developer Productivity, where 2x friend of the pod and now Zyphra founder Quentin Anthony was the ONLY productive participant!Finally, if you're the sort to read these show notes to the end, then you definitely deserve some pictures of Joel shredding the guitar at Love Band Karaoke which we mention at the end: Full Video PodTimestamps00:00 What METR Means00:39 Podcast Intro With Joel01:39 ME vs TR03:33 Time Horizon Origin Story04:56 Picking Tasks And Biases09:13 Time Horizon Misconceptions11:37 Opus 4.5 And Trendlines14:27 Productivity Studies And Explosions29:50 Compute Slows Progress30:47 Algorithms Need Compute32:45 Industry Spend and Data34:57 Clusters and Shipping Timelines36:44 Prediction Markets for Models38:10 Manifold Alpha Story43:04 Beyond Benchmarks Evals51:39 METR Roadmap and FarewellTranscript

Swyx joined SAIL! Thank you SAIL Media, Prof. Tom Yeh, 8Lee, Hamid Bagheri, c9n, and many others for tuning into SAIL Live #6 with Nathan Lambert and Sebastian Raschka, PhD. Sharing here for the LS paid subscribers.We covered: This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe

Editor's note: CuspAI raised a $100m Series A in September and is rumored to have reached a unicorn valuation. They have all-star advisors from Geoff Hinton to Yann Lecun and team of deep domain experts to tackle this next frontier in AI applications.In this episode, Max Welling traces the thread connecting quantum gravity, equivariant neural networks, diffusion models, and climate-focused materials discovery (yes, there is one!!!).We begin with a provocative framing: experiments as computation. Welling describes the idea of a “physics processing unit”—a world in which digital models and physical experiments work together, with nature itself acting as a kind of processor. It's a grounded but ambitious vision of AI for science: not replacing chemists, but accelerating them.Along the way, we discuss:* Why symmetry and equivariance matter in deep learning* The tradeoff between scale and inductive bias* The deep mathematical links between diffusion models and stochastic thermodynamics* Why materials—not software—may be the real bottleneck for AI and the energy transition* What it actually takes to build an AI-driven materials platformMax reflects on moving from curiosity-driven theoretical physics (including work with Gerard ‘t Hooft) toward impact-driven research in climate and energy. The result is a conversation about convergence: physics and machine learning, digital models and laboratory experiments, long-term ambition and incremental progress.Full Video EpisodeTimestamps* 00:00:00 – The Physics Processing Unit (PPU): Nature as the Ultimate Computer* Max introduces the idea of a Physics Processing Unit — using real-world experiments as computation.* 00:00:44 – From Quantum Gravity to AI for Materials* Brandon frames Max's career arc: VAE pioneer → equivariant GNNs → materials startup founder.* 00:01:34 – Curiosity vs Impact: How His Motivation Evolved* Max explains the shift from pure theoretical curiosity to climate-driven impact.* 00:02:43 – Why CaspAI Exists: Technology as Climate Strategy* Politics struggles; technology scales. Why materials innovation became the focus.* 00:03:39 – The Thread: Physics → Symmetry → Machine Learning* How gauge symmetry, group theory, and relativity informed equivariant neural networks.* 00:06:52 – AI for Science Is Exploding (Not Emerging)* The funding surge and why AI-for-Science feels like a new industrial era.* 00:07:53 – Why Now? The Two Catalysts Behind AI for Science* Protein folding, ML force fields, and the tipping point moment.* 00:10:12 – How Engineers Can Enter AI for Science* Practical pathways: curriculum, workshops, cross-disciplinary training.* 00:11:28 – Why Materials Matter More Than Software* The argument that everything—LLMs included—rests on materials innovation.* 00:13:02 – Materials as a Search Engine* The vision: automated exploration of chemical space like querying Google.* 01:14:48 – Inside CuspAI: The Platform Architecture* Generative models + multi-scale digital twin + experiment loop.* 00:21:17 – Automating Chemistry: Human-in-the-Loop First* Start manual → modular tools → agents → increasing autonomy.* 00:25:04 – Moonshots vs Incremental Wins* Balancing lighthouse materials with paid partnerships.* 00:26:22 – Why Breakthroughs Will Still Require Humans* Automation is vertical-specific and iterative.* 00:29:01 – What Is Equivariance (In Plain English)?* Symmetry in neural networks explained with the bottle example.* 00:30:01 – Why Not Just Use Data Augmentation?* The optimization trade-off between inductive bias and data scale.* 00:31:55 – Generative AI Meets Stochastic Thermodynamics* His upcoming book and the unification of diffusion models and physics.* 00:33:44 – When the Book Drops (ICLR?)TranscriptMax: I want to think of it as what I would call a physics processing unit, like a PPU, right? Which is you have digital processing units and then you have physics processing units. So it's basically nature doing computations for you. It's the fastest computer known, as possible even. It's a bit hard to program because you have to do all these experiments. Those are quite bulky, it's like a very large thing you have to do. But in a way it is a computation and that's the way I want to see it. You can do computations in a data center and then you can ask nature to do some computations. Your interface with nature is a bit more complicated. But then these things will have to seamlessly work together to get to a new material that you're interested in.[01:00:44:14 - 01:01:34:08]Brandon: Yeah, it's a pleasure to have Max Woehling as a guest today. Max has done so much over his career that I've been so excited about. If you're in the deep learning community, you probably know Max for his work on variational autocoders, which has literally stood the test of prime or officially stood the test of prime. If you are a scientist, you probably know him for his like, binary work on graph neural networks on equivariance. And if you're a material science, you probably know him about his new startup, CASPAI. Max has a long history doing lots of cool problems. You started in quantum gravity, which is I think very different than all of these other things you worked on. The first question for AI engineers and for scientists, what is the thread in how you think about problems? What is the thread in the type of things which excite you? And how do you decide what is the next big thing you want to work on?[01:01:34:08 - 01:02:41:13]Max: So it has actually evolved a lot. In my young days, let's breathe, I would just follow what I would find super interesting. I have kind of this sensor. I think many people have, but maybe not really sort of use very much, which is like, you get this feeling about getting very excited about some problem. Like it could be, what's inside of a black hole or what's at the boundary of the universe or what are quantum mechanics actually all about. And so I follow that basically throughout my career. But I have to say that as you get older, this changes a little bit in the sense that there's a new dimension coming to it and there's this impact. Going in two-dimensional quantum gravity, you pretty much guaranteed there's going to be no impact on what you do relative, maybe a few papers, but not in this world, this energy scale. As I get closer to retirement, which is fortunately still 10 years away or so, I do want to kind of make a positive impact in the world. And I got pretty worried about climate change.[01:02:43:15 - 01:03:19:11]Max: I think politics seems to have a hard time solving it, especially these days. And so I thought better work on it from the technology side. And that's why we started CaspAI. But there's also a lot of really interesting science problems in material science. And so it's kind of combining both the impact you can make with it as well as the interesting science. So it's sort of these two dimensions, like working on things which you feel there's like, well, there's something very deep going on here. And on the other hand, trying to build tools that can actually make a real impact in the world.[01:03:19:11 - 01:03:39:23]RJ: So the thread that when I look back, look at the different things that you worked out, some of them seem pretty connected, like the physics to equivariance and, yeah, and, uh, gravitational networks, maybe. And that seems to be somewhat related to Casp. Do you have a thread through there?[01:03:39:23 - 01:06:52:16]Max: Yeah. So physics is the thread. So having done, you know, spent a lot of time in theoretical physics, I think there is first very fundamental and exciting questions, like things that haven't actually been figured out in quantum gravity. So that is really the frontier. There's also a lot of mathematical tools that you can use, right? In, for instance, in particle physics, but also in general relativity, sort of symmetry space to play an enormously important role. And this goes all the way to gauge symmetries as well. And so applying these kinds of symmetries to, uh, machine learning was actually, you know, I thought of it as a very deep and interesting mathematical problem. I did this with Taco Cohen and Taco was the main driver behind this, went all the way from just simple, like rotational symmetries all the way to gauge symmetries on spheres and stuff like that. So, and, uh, Maurice Weiler, who's also here, um, when he was a PhD student, he was a very good student with me, you know, he wrote an entire book, which I can really recommend about the role of symmetries in AI and machine learning. So I find this a very deep and interesting problem. So more recently, so I've taken a sort of different path, which is the relationship between diffusion models and that field called stochastic thermodynamics. This is basically the thermodynamics, which is a theory of equilibrium. So but then formulated for out of equilibrium systems. And it turns out that the mathematics that we use for diffusion models, but even for reinforcement learning for Schrodinger bridges for MCMC sampling has the same mathematics as this theoretical, this physical theory of non-equilibrium systems. And that got me very excited. And actually, uh, when I taught a course in, um, Mauschenberg, uh, it is South Africa, close to Cape Town at the African Institute for Mathematical Sciences Ames. And I turned that into a book site. Two years later, the book was finished. I've sent it to the publisher. And this is about the deep relationship between free energy, diffusion models, basically generative AI and stochastic thermodynamics. So it's always some kind of, I don't know, I find physics very deep. I also think a lot about quantum mechanics and it's, it's, it's a completely weird theory that actually nobody really understands. And there's a very interesting story, which is maybe good to tell to connect sort of my PZ back to where I'm now. So I did my PZ with a Nobel Laureate, Gerard the toft. He says the most brilliant man I've ever met. He was never wrong about anything as long as I've seen him. And now he says quantum mechanics is wrong and he has a new theory of quantum mechanics. Nobody understands what he's saying, even though what he's writing down is not mathematically very complex, but he's trying to address this understandability, let's say of quantum mechanics head on. And I find it very courageous and I'm completely fascinated by it. So I'm also trying to think about, okay, can I actually understand quantum mechanics in a more mundane way? So that, you know, without all the weird multiverses and collapses and stuff like that. So the physics is always been the threat and I'm trying to apply the physics to the machine learning to build better algorithms.[01:06:52:16 - 01:07:05:15]Brandon: You are still very involved in understanding and understanding physics and the worlds. Yeah. And just like applications to machine learning or introducing no formalisms. That's really cool.[01:07:05:15 - 01:07:18:02]Max: Yes, I would say I'm not contributing much to physics, but I'm contributing to the interface between physics and science. And that's called AI for science or science or AI is kind of a super, it's actually a new discipline that's emerging.[01:07:18:02 - 01:07:18:19]Speaker 5: Yeah.[01:07:18:19 - 01:07:45:14]Max: And it's not just emerging, it's exploding, I would say. That's the better term because I know you go from investments into like in the hundreds of millions now in the billions. So there's now actually a startup by Jeff Bezos that is at 6.2 billion sheep round. Right. Insane. I guess it's the largest startup ever, I think. And that's in this field, AI for science. It tells you something that we are creating a new bubble here.[01:07:46:15 - 01:07:53:28]Brandon: So why do you think it is? What has changed that has motivated people to start working on AI for science type problems?[01:07:53:28 - 01:08:49:17]Max: So there's two reasons actually. One is that people have been applying sort of the new tools from AI to the sciences, which is quite natural. And there's of course, I think there's two big examples, protein folding is a big one. And the other one is machine learning forest fields or something called machine learning inter-atomic potentials. Both of them have been actually very successful. Both also had something to do with symmetries, which is a little cool. And sort of people in the AI sciences saw an opportunity to apply the tools that they had developed beyond advertised placement, right, or multimedia applications into something that could actually make a very positive impact in society like health, drug development, materials for the energy transition, carbon capture. These are all really cool, impactful applications.[01:08:50:19 - 01:09:42:14]Max: Despite that, the science and the kind of the is also very interesting. I would say the fact that these sort of these two fields are coming together and that we're now at the point that we can actually model these things effectively and move the needle on some of these sort of science sort of methodologies is also a very unique moment, I would say. People recognize that, okay, now we're at the cusp of something new, where it results whether the company is called after. We're at the cusp of something new. And of course that always creates a lot of energy. It's like, okay, there's something, it's like sort of virgin field. It's like nobody's green field. Nobody's been there. I can rush in and I can sort of start harvesting there, right? And I think that's also what's causing a lot of sort of enthusiasm in the fields.[01:09:42:14 - 01:10:12:18]RJ: If you're an AI engineer, basically if the people that listen to this podcast will be in the field, then you maybe don't have a strong science background. How does, but are excited. Most I would say most AI practitioners, BM engineers or scientists would consider themselves scientists and they have some background, a little bit of physics, a little bit of industry college, maybe even graduate school that have been working or are starting out. How does somebody who is not a scientist on a day-to-day basis, how do they get involved?[01:10:12:18 - 01:10:14:28]Max: Well, they can read my book once it's out.[01:10:16:07 - 01:11:05:24]Max: This is basically saying that there is more, we should create curricula that are on this interface. So I'm not sure there is, also we already have some universities actual courses you can take, maybe online courses you can take. These workshops where we are now are actually very good as well. And we should probably have more tutorials before the workshop starts. Actually we've, I've kind of proposed this at some point. It's like maybe first have an hour of a tutorial so that people can get new into the field. There's a lot out there. Most of it is of course inaccessible, but I would say we will create much more books and other contents that is more accessible, including this podcast I would say. So I think it will come. And these days you can watch videos and things. There's a huge amount of content you can go and see.[01:11:05:24 - 01:11:28:28]Brandon: So maybe a follow-up to that. How do people learn and get involved? But why should they get involved? I mean, we have a lot of people who are of our audience will be interested in AI engineering, but they may be looking for bigger impacts in the world. What opportunities does AI for science provide them to make an impact to change the world? That working in this the world of pure bits would not.[01:11:28:28 - 01:11:40:06]Max: So my view is that underlying almost everything is immaterial. So we are focusing a lot on LLMs now, which is kind of the software layer.[01:11:41:06 - 01:11:56:05]Max: I would say if you think very hard, underlying everything is immaterial. So underlying an LLM is a GPU, and underlying a GPU is a wafer on which we will have to deposit materials. Do we want to wait a little bit?[01:12:02:25 - 01:12:11:06]Max: Underlying everything is immaterial. So I was saying, you know, there's the LLM underlying the LLM is a GPU on which it runs. In order to make that GPU,[01:12:12:08 - 01:12:43:20]Max: you have to put materials down on a wafer and sort of shine on it with sort of EUV light in order to etch kind of the structures in. But that's now an actual material problem, because more or less we've reached the limits of scaling things down. And now we are trying to improve further by new materials. So that's a fundamental materials problem. We need to get through the energy transition fast if we don't want to kind of mess up this world. And so there is, for instance, batteries. That's a complete materials problem. There's fuel cells.[01:12:44:23 - 01:13:01:16]Max: There is solar panels. So that they can now make solar panels with new perovskite layers on top of the silicon layers that can capture, you know, theoretically up to 50% of the light, where now we're at, I don't know, maybe 22 or something. So these are huge changes all by material innovation.[01:13:02:21 - 01:13:47:15]Max: And yeah, I think wherever you go, you know, I can probably dig deep enough and then tell you, well, actually, the very foundation of what you're doing is a material problem. And so I think it's just very nice to work on this very, very foundation. And also because I think this is maybe also something that's happening now is we can start to search through this material space. This has never been the case, right? It's like scientists, the normal way of working is you read papers and then you come up with no hypothesis. You do an experiment and you learn, et cetera. So that's a very slow process. Now we can treat this as a search engine. Like we search the internet, we now search the space of all possible molecules, not just the ones that people have made or that they're in the universe, but all of them.[01:13:48:21 - 01:14:42:01]Max: And we can make this kind of fully automated. That's the hope, right? We can just type, it becomes a tool where you type what you want and something starts spinning and some experiments get going. And then, you know, outcome list of materials and then you look at it and say, maybe not. And then you refine your query a little bit. And you kind of do research with this search engine where a huge amount of computation and experimentation is happening, you know, somewhere far away in some lab or some data center or something like this. I find this a very, very promising view of how we can sort of build a much better sort of materials layer underneath almost everything. And also more sustainable materials. Our plastics are polluting the planet. If you come up with a plastic that kind of destroys itself, you know, after, I don't a few weeks, right? And actually becomes a fertilizer. These are things that are not impossible at all. These things can be done, right? And we should do it.[01:14:42:01 - 01:14:47:23]RJ: Can you tell us a little bit just generally about CUSBI and then I have a ton of questions.[01:14:47:23 - 01:14:48:15]Speaker 5: Yeah.[01:14:48:15 - 01:17:49:10]Max: So CUSBI started about 20 months ago and it was because I was worried about I'm still worried about climate change. And so I realized that in order to get, you know, to stay within two degrees, let's say, we would not only have to reduce our emissions to zero by 2050, but then, you know, another half century or even a century of removing carbon dioxide from the atmosphere, not by reducing your emissions, but actually removing it at a rate that's about half the rate that we now emit it. And that is a unsolved problem. But if we don't solve it, two degrees is not going to happen, right? It's going to be much more. And I don't think people quite understand how bad that can be, like four degrees, like very bad. So this technology needs to be developed. And so this was my and my co-founder, Chet Edwards, motivation to start this startup. And also because, you know, we saw the technology was ready, which is also very good. So if you're, you know, the time is right to do it. And yeah, so we now in the meanwhile, we've grown to about 40 people. We've kind of collected 130 million investment into the company, which is for a European company is quite a lot. I would say it's interesting that right after that, you know, other startups got even more. So that's kind of tells you how fast this is growing. But yeah, we are we are now at the we've built the platform, of course, but it's for a series of material classes and it needs to be constantly expanded to new material classes. And it can be more automated because, you know, we know putting LLMs in as the whole thing gets more and more automated. And now we're moving to sort of high throughput experimentation. So connecting the actual platform, which is computational, to the experiments so that you can get also get fast feedback from experiments. And I kind of think of experiments as something you do at the end, although that's what we've been doing so far. I want to think of it as what I would call a sort of a physics processing unit, like a PPU, right, which is you have digital processing units and then you have physics processing units. So it's basically nature doing computations for you. It's the fastest computer known as possible, even. It's a bit hard to program because you have to do all these experiments. Those are quite, quite bulky. It's like a very large thing you have to do. But in a way, it is a computation. And that's the way I want to see it. So I want to you can do computations in a data center and then you can ask nature to do some computations. Your interface with nature is a bit more complicated. But then these things will have to seamlessly work together to get to a new material that you're interested in. And that's the vision we have. We don't say super intelligence because I don't quite know what it means and I don't want to oversell it. But I do want to automate this process and give a very powerful tool in the hands of the chemists and the material scientists.[01:17:49:10 - 01:18:01:02]Brandon: That actually brings up a question I wanted to ask you. First of all, can you talk about your platform to like whatever degree, like explain kind of how it works and like what you your thought processes was in developing it?[01:18:01:02 - 01:20:47:22]Max: Yeah, I think it's been surprisingly, it's not rocket science, I would say. It's not rocket science in the sense of the design and basically the design that, you know, I wrote down at the very beginning. It's still more or less the design, although you add things like I wasn't thinking very much about multi-scale models and as the common are rated that actually multi-scale is very important. And the beginning, I wasn't thinking very much about self-driving labs. But now I think, you know, we are now at the stage we should be adding that. And so there is sort of bits and details that we're adding. But more or less, it's what you see in the slide decks here as well, which is there is a generative component that you have to train to generate candidates. And then there is a digital twin, multi-scale, multi-fidelity digital twin, which you walk through the steps of the ladder, you know, they do the cheap things first, you weed out everything that's obviously unuseful, and then you go to more and more expensive things later. And so you narrow things down to a small number. Those go into an experiment, you know, do the experiment, get feedback, etc. Now, things that also have been more recently added is sort of more agentic sort of parts. You know, we have agents that search the literature and come up with, you know, actually the chemical literature and come up with, you know, chemical suggestions for doing experiments. We have agents which sort of autonomously orchestrate all of the computations and the experiments that need to be done. You know, they're in various stages of maturity and they can be continuously improved, I would say. And so that's basically I don't think that part. There's rocket science, but, you know, the design of that thing is not like surprising. What is it's surprising hard to actually build it. Right. So that's that's the thing that is where the moat is in the data that you can get your hands on and the and actually building the platform. And I would say there's two people in particular I want to call out, which is Felix Hunker, who is actually, you know, building the scientific part of the platform and Sandra de Maria, who is building the sort of the skate that is kind of this the MLOps part of the platform. Yeah. And so and recently we also added sort of Aaron Walsh to our team, who is a very accomplished scientist from Imperial College. We're very happy about that. He's going to be a chief science officer. And we also have a partnerships team that sort of seeks out all the customers because I think this is one thing I find very important. In print, it's so complex to do to actually bring a material to the real world that you must do this, you know, in collaboration with sort of the domain experts, which are the companies typically. So we always we only start to invest in the direction if we find a good industrial partner to go on that journey with us.[01:20:47:22 - 01:20:55:12]Brandon: Makes a lot of sense. Over the evolution of the platform, did you find that you that human intervention, human,[01:20:56:18 - 01:21:17:01]Brandon: I guess you could start out with a pure, you could imagine two directions when you start up making everything purely automatic, automated, agentic, so on. And then later on, you like find that you need to have more human input and feedback different steps. Or maybe did you start out with having human feedback? You have lots of steps and then like kind of, yeah, figure out ways to remove, you know,[01:21:17:01 - 01:22:39:18]Max: that is the second one. So you build tools for you. So it's much more modular than you think. But it's like, we need these tools for this application. We need these tools. So you build all these tools, and then you go through a workflow actually in the beginning just manually. So you put them in a first this tool, then run this to them or this with sithery. So you put them in a workflow and then you figure out, oh, actually, you know, this this porous material that we are trying to make actually collapses if you shake it a bit. Okay, then you add a new tool that says test for stability. Right. Yeah. And so there's more and more tools. And then you build the agent, which could be a Bayesian optimizer, or it could be an actual other them, you know, maybe trained to be a good chemist that will then start to use all these tools in the right way in the right order. Yeah. Right. But in the beginning, it's like you as a chemist are putting the workflow together. And then you think about, okay, how am I going to automate this? Right. For one very easy question you can ask yourself is, you know, every time somebody who is not a super expert in DFT, yeah, and he wants to do a calculation has to go to somebody who knows DFT. And so could you start to automate that away, which is like, okay, make it so user friendly, so that you actually do the right DFT for the right problem and for the right length of time, and you can actually assess whether it's a good outcome, etc. So you start to automate smaller small pieces and bigger pieces, etc. And in the end, the whole thing is automated.[01:22:39:18 - 01:22:53:25]Brandon: So your philosophy is you want to provide a set of specific tools that make it so that the scientists making decisions are better informed and less so trying to create an automated process.[01:22:53:25 - 01:23:22:01]Max: I think it's this is sort of the same where you're saying because, yes, we want to automate, yeah, but we don't see something very soon where the chemists and the domain expert is out of the loop. Yeah, but it but it's a retreat, right? It's like, okay, so first, you need an expert to tell you precisely how to set the parameters of the DFT calculation. Okay, maybe we can take that out. We can maybe automate that, right? And so increasingly, more of these things are going to be removed.[01:23:22:01 - 01:23:22:19]Speaker 5: Yeah.[01:23:22:19 - 01:24:33:25]Max: In the end, the vision is it will be a search engine where you where somebody, a chemist will type things and we'll get candidates, but the chemist will still decide what is a good material and what is not a good material out of that list, right? And so the vision of a completely dark lab, where you can close the door and you just say, just, you know, find something interesting and then it will it will just figure out what's interesting and we'll figure out, you know, it's like, oh, I found this new material to blah, blah, blah, blah, right? That's not the vision I have. He's not for, you know, a long time. So for me, it's really empowering the domain experts that are sitting in the companies and in universities to be much faster in developing their materials. And I should say, it's also good to be a little humble at times, because it is very complicated, you know, to bring it to make it and to bring it into the real world. And there are people that are doing this for the entire lives. Yeah. Right. And it's like, I wonder if they scratch their head and say, well, you know, how are you going to completely automate that away, like in the next five years? I don't think that's going to happen at all.[01:24:35:01 - 01:24:39:24]Max: Yeah. So to me, it's an increasingly powerful tool in the hands of the chemists.[01:24:39:24 - 01:25:04:02]RJ: I have a question. You've talked before about getting people interested based on having, you know, sort of a big breakthrough in materials, incremental change. I'm curious what you think about the platform you have now in are sort of stepping towards and how are you chasing the big change or is this like incremental or is there they're not mutually exclusive, obviously, but what do you think about that?[01:25:04:02 - 01:26:04:27]Max: We follow a mixed strategy. So we are definitely going after a big material. Again, we do this with a partner. I'm not going to disclose precisely what it is, but we have our own kind of long term goal. You could call it lighthouse or, you know, sort of moonshot or whatever, but it is going to be a really impactful material that we want to develop as a proof point that it can be done and that it will make it into the into the real world and that AI was essential in actually making it happen. At the same time, we also are quite happy to work with companies that have more modest goals. Like I would say one is a very deep partnership where you go on a journey with a company and that's a long term commitment together. And the other one is like somebody says, I knew I need a force field. Can you help me train this force field and then maybe analyze this particular problem for me? And I'll pay you a bunch of money for that. And then maybe after that we'll see. And that's fine too. Right. But we prefer, you know, the deep partnerships where we can really change something for the good.[01:26:04:27 - 01:26:22:02]RJ: Yeah. And do you feel like from a platform standpoint you're ready for that or what are the things that and again, not asking you to disclose proprietary secret sauce, but what are the things generally speaking that need to happen from where we are to where to get those big breakthroughs?[01:26:22:02 - 01:28:40:01]Max: What I find interesting about this field is that every time you build something, it's actually immediately useful. Right. And so unlike quantum computing, which or nuclear fusion, so you work for 20, 30, 40 years and nothing, nothing, nothing, nothing. And then it has to happen. Right. And when it happens, it's huge. So it's quite different here because every time you introduce, so you go to a customer and you say, so what do you need? Right. So we work, let's say, on a problem like a water filtration. We want to remove PFAS from water. Right. So we do this with a company, Camira. So they are a deep partner for us. Right. So we on a journey together. I think that the breakthrough will happen with a lot of human in the loop because there is the chemists who have a whole lot more knowledge of their field and it's us who will help them with training, having a new message. And in that kind of interface, these interactions, something beautiful will happen and that will have to happen first before this field will really take off, I think. And so in the sense that it's not a bubble, let's put it that way. So that's people see that as actual real what's happening. So in the beginning, it will be very, you know, with a lot of humans in the loop, I would say, and I would I would hope we will have this new sort of breakthrough material before, you know, everything is completely automated because that will take a while. And also it is very vertical specific. So it's like completely automating something for problem A, you know, you can probably achieve it, but then you'll sort of have to start over again for problem B because, you know, your experimental setup looks very different in the machines that you characterize your materials look very different. Even the models in your platform will have to be retrained and fine tuned to the new class. So every time, you know, you have a lot of learnings to transfer, but also, you know, the problems are actually different. And so, yes, I would want that breakthrough material before it's completely automated, which I think is kind of a long term vision. And I would say every time you move to something new, you'll have to start retraining and humans will have to come in again and say, okay, so what does this problem look like? And now sort of, you know, point the the machine again, you know, in the new direction and then and then use it again.[01:28:40:01 - 01:28:47:17]RJ: For the non-scientists among us, me included a bit of a scientist. There's a lot of terminology. You mentioned DFT,[01:28:49:00 - 01:29:01:11]RJ: you equivariance we've talked about. Can you sort of explain in engineering terms or the level of sophistication and engineering? Well, how what is equivariance?[01:29:01:11 - 01:29:55:01]Max: So equivariance is the infusion of symmetry in neural networks. So if I build a neural network, let's say that needs to recognize this bottle, right, and then I rotate the bottle, it will then actually have to completely start again because it has no idea that the rotated bottle. Well, actually, the input that represents a rotated bottle is actually rotated bottle. It just doesn't understand that. Right. If you build equivariance in basically once you've trained it in one orientation, it will understand it in any other orientation. So that means you need a lot less data to train these models. And these are constraints on the weights of the model. So so basically you have to constrain the way such data to understand it. And you can build it in, you can hard code it in. And yeah, this the symmetry groups can be, you know, translations, rotations, but also permutations. I can graph neural network, their permutations and then physics, of course, as many more of these groups.[01:29:55:01 - 01:30:01:08]RJ: To pray devil's advocate, why not just use data augmentation by your bottle is in all the different orientations?[01:30:01:08 - 01:30:58:23]Max: As an option, it's just not exact. It's like, why would you go through the work of doing all that? Where you would really need an infinite number of augmentations to get it completely right. Where you can also hard code it in. Now, I have to say sometimes actually data augmentation works even better than hard coding the equivariance in. And this is something to do with the fact that if you constrain the optimization, the weights before the optimization starts, the optimization surface or objective becomes more complicated. And so it's harder to find good minima. So there is also a complicated interplay, I think, between the optimization process and these constraints you put in your network. And so, yeah, you'll hear kind of contradicting claims in this field. Like some people and for certain applications, it works just better than not doing it. And sometimes you hear other people, if you have a lot of data and you can do data augmentation, then actually it's easier to optimize them and it actually works better than putting the equivariance in.[01:30:58:23 - 01:31:07:16]Brandon: Do you think there's kind of a bitter lesson for mathematically founded models and strategies for doing deep learning?[01:31:07:16 - 01:31:46:06]Max: Yeah, ultimately it's a trade-off between data and inductive bias. So if your inductive bias is not perfectly correct, you have to be careful because you put a ceiling to what you can do. But if you know the symmetry is there, it's hard to imagine there isn't a way to actually leverage it. But yeah, so there is a bitter lesson. And one of the bitter lessons is you should always make sure your architecture is scale, unless you have a tiny data set, in which case it doesn't matter. But if you, you know, the same bitter lessons or lessons that you can draw in LLM space are eventually going to be true in this space as well, I think.[01:31:47:10 - 01:31:55:01]RJ: Can you talk a little bit about your upcoming book and tell the listeners, like, what's exciting about it? Yeah, I should read it.[01:31:55:01 - 01:33:42:20]Max: So this book is about, it's called Generative AI and Stochastic Thermodynamics. It basically lays bare the fact that the mathematics that goes into both generative AI, which is the technology to generate images and videos, and this field of non-equilibrium statistical mechanics, which are systems of molecules that are just moving around and relaxing to the ground state, or that you can control to have certain, you know, be in a certain state, the mathematics of these two is actually identical. And so that's fascinating. And in fact, what's interesting is that Jeff Hinton and Radford Neal already wrote down the variational free energy for machine learning a long time ago. And there's also Carl Friston's work on free energy principle and active entrance. But now we've related it to this very new field in physics, which is called stochastic thermodynamics or non-equilibrium thermodynamics, which has its own very interesting theorems, like fluctuation theorems, which we don't typically talk about, but we can learn a lot from. And I think it's just it can sort of now start to cross fertilize. When we see that these things are actually the same, we can, like we did for symmetries, we can now look at this new theory that's out there, developed by these very smart physicists, and say, okay, what can we take from here that will make our algorithms better? At the same time, we can use our models to now help the scientists do better science. And so it becomes a beautiful cross-fertilization between these two fields. The book is rather technical, I would say. And it takes all sorts of things that have been done as stochastic thermodynamics, and all sorts of models that have been done in the machine learning literature, and it basically equates them to each other. And I think hopefully that sense of unification will be revealing to people.[01:33:42:20 - 01:33:44:05]RJ: Wait, and when is it out?[01:33:44:05 - 01:33:56:09]Max: Well, it depends on the publisher now. But I hope in April, I'm going to give a keynote at ICLR. And it would be very nice if they have this book in my hand. But you know, it's hard to control these kind of timelines.[01:33:56:09 - 01:33:58:19]RJ: Yeah, I'm looking forward to it. Great.[01:33:58:19 - 01:33:59:25]Max: Thank you very much. This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe

This is a free preview of a paid episode. To hear more, visit www.latent.spaceFirst speakers for AIE Europe and AIEi Miami have been announced. If you're in Asia/Aus, come by Singapore and Melbourne. AI Engineering is going global!One year ago today, Anthropic launched Claude Code, to not much fanfare:The word of mouth was incredibly strong however, and so we were glad to be one of the first podcasts to invite Boris and Cat on in early May:As we discussed on the pod, all CC usage was API-based and therefore it was ridiculously expensive to do anything. This was then fixed by the team including Claude Code in the Claude Pro plan in early June, and then the virality caused us to make a rare trend call in late June:Now, 6 months on, Doug has just calculated that around 4% of GitHub is written by Claude Code:We talk about how Doug uses Claude Code to do SemiAnalysis work.Memory ManiaIn the second part of this episode, we also check in on Memory Mania, which is going to affect you (yes, you) at home if it hasn't already:Full Episode on YouTubeTimestamps00:00 AI as Junior Analyst00:59 Meet Swyx and Doug03:30 From Value Mule to Semis06:28 Moore's Law Ends Thesis12:02 Claude Code Awakening32:02 Agent Swarms Reality Check32:53 Kimi Swarm Benchmarks37:31 Bots vs Zapier Automation39:44 Claude Code Workflow Setup57:54 AGI Metrics and GDP01:04:48 Railroad CapEx Analogy01:06:00 Funding Bubbles and Demand01:08:11 Agents Replace Work Tools01:13:56 Codex vs Claude Race01:21:15 Microsoft and TPU Strategy01:34:13 TPU Window vs Nvidia01:36:30 HBM Supply Chain Squeeze01:39:41 Memory Shock and CXL01:45:20 Context Rationing Future01:54:37 Writing and Trail LessonsTranscript[00:00:00] AI as Junior Analyst[00:00:00] Doug: This crap makes mistakes all the time. All the time. It is still just like a, like I think of it once again as like a junior analyst, right? The analyst goes and does all this like really pain in the ass information and you bring it all together to make a good decision at the top. Historically what happens is that junior analyst, who I once was, went and gathered all that information, and after doing this enough times, there's a meta level thinking that's happening where it's like, okay, here's what I really understand and how this type of analysis, I'm an expert in, actually I'm very good at, I consistently have a hit rate.[00:00:28] Now I'm the expert, right? I don't think that meta level learning is there yet. We'll see if l ones do it, right? Everyone who's spending one quadrillion dollars in the world thinks it will, it better, it better happen by if you're spending, you know, a trillion dollars and there's not meta level learning.[00:00:44] But for me, in our firm, that massively amplifies everyone who is an expert. ‘cause like you have to still do something that you can just like lop it up. It's very obvious to me. What It's slop.[00:00:59] Meet Swyx and Doug

Olivia Watkins (Frontier Evals team) and Mia Glaese (VP of Research at OpenAI, leading the Codex, human data, and alignment teams) discuss a new blog post (https://openai.com/index/why-we-no-longer-evaluate-swe-bench-verified/) arguing that SWE-Bench Verified—long treated as a key “North Star” coding benchmark—has become saturated and highly contaminated, making it less useful for measuring real coding progress. SWE-Bench Verified originated as a major OpenAI-led cleanup of the original Princeton SWE-Bench benchmark, including a large human review effort with nearly 100 software engineers and multiple independent reviews to curate ~500 higher-quality tasks. But recent findings show that many remaining failures can reflect unfair or overly narrow tests (e.g., requiring specific naming or unspecified implementation details) rather than true model inability, and cite examples suggesting contamination such as models recalling repository-specific implementation details or task identifiers. From now on, OpenAI plans to stop reporting SWE-Bench Verified and instead focus on SWE-Bench Pro (from Scale), which is harder, more diverse (more repos and languages), includes longer tasks (1–4 hours and 4+ hours), and shows substantially less evidence of contamination under their “contamination auditor agent” analysis. We also discuss what future coding/agent benchmarks should measure beyond pass/fail tests—longer-horizon tasks, open-ended design decisions, code quality/maintainability, and real-world product-building—along with the tradeoffs between fast automated grading and human-intensive evaluation. 00:00 Meet the Frontier Evals Team00:56 Why SWE Bench Stalled01:47 How Verified Was Built04:32 Contamination In The Wild06:16 Unfair Tests And Narrow Specs08:40 When Benchmarks Saturate10:28 Switching To SWE Bench Pro12:31 What Great Coding Evals Measure18:17 Beyond Tests Dollars And Autonomy21:49 Preparedness And Future Directions Get full access to Latent.Space at www.latent.space/subscribe

Tickets for AIEi Miami and AIE Europe are live, with first wave speakers announced!From pioneering software-defined networking to backing many of the most aggressive AI model companies of this cycle, Martin Casado and Sarah Wang sit at the center of the capital, compute, and talent arms race reshaping the tech industry. As partners at a16z investing across infrastructure and growth, they've watched venture and growth blur, model labs turn dollars into capability at unprecedented speed, and startups raise nine-figure rounds before monetization.Martin and Sarah join us to unpack the new financing playbook for AI: why today's rounds are really compute contracts in disguise, how the “raise → train → ship → raise bigger” flywheel works, and whether foundation model companies can outspend the entire app ecosystem built on top of them. They also share what's underhyped (boring enterprise software), what's overheated (talent wars and compensation spirals), and the two radically different futures they see for AI's market structure.We discuss:* Martin's “two futures” fork: infinite fragmentation and new software categories vs. a small oligopoly of general models that consume everything above them* The capital flywheel: how model labs translate funding directly into capability gains, then into revenue growth measured in weeks, not years* Why venture and growth have merged: $100M–$1B hybrid rounds, strategic investors, compute negotiations, and complex deal structures* The AGI vs. product tension: allocating scarce GPUs between long-term research and near-term revenue flywheels* Whether frontier labs can out-raise and outspend the entire app ecosystem built on top of their APIs* Why today's talent wars ($10M+ comp packages, $B acqui-hires) are breaking early-stage founder math* Cursor as a case study: building up from the app layer while training down into your own models* Why “boring” enterprise software may be the most underinvested opportunity in the AI mania* Hardware and robotics: why the ChatGPT moment hasn't yet arrived for robots and what would need to change* World Labs and generative 3D: bringing the marginal cost of 3D scene creation down by orders of magnitude* Why public AI discourse is often wildly disconnected from boardroom reality and how founders should navigate the noiseShow Notes:* “Where Value Will Accrue in AI: Martin Casado & Sarah Wang” - a16z show* “Jack Altman & Martin Casado on the Future of Venture Capital”* World Labs—Martin Casado• LinkedIn: https://www.linkedin.com/in/martincasado/• X: https://x.com/martin_casadoSarah Wang• LinkedIn: https://www.linkedin.com/in/sarah-wang-59b96a7• X: https://x.com/sarahdingwanga16z• https://a16z.com/Timestamps00:00:00 – Intro: Live from a16z00:01:20 – The New AI Funding Model: Venture + Growth Collide00:03:19 – Circular Funding, Demand & “No Dark GPUs”00:05:24 – Infrastructure vs Apps: The Lines Blur00:06:24 – The Capital Flywheel: Raise → Train → Ship → Raise Bigger00:09:39 – Can Frontier Labs Outspend the Entire App Ecosystem?00:11:24 – Character AI & The AGI vs Product Dilemma00:14:39 – Talent Wars, $10M Engineers & Founder Anxiety00:17:33 – What's Underinvested? The Case for “Boring” Software00:19:29 – Robotics, Hardware & Why It's Hard to Win00:22:42 – Custom ASICs & The $1B Training Run Economics00:24:23 – American Dynamism, Geography & AI Power Centers00:26:48 – How AI Is Changing the Investor Workflow (Claude Cowork)00:29:12 – Two Futures of AI: Infinite Expansion or Oligopoly?00:32:48 – If You Can Raise More Than Your Ecosystem, You Win00:34:27 – Are All Tasks AGI-Complete? Coding as the Test Case00:38:55 – Cursor & The Power of the App Layer00:44:05 – World Labs, Spatial Intelligence & 3D Foundation Models00:47:20 – Thinking Machines, Founder Drama & Media Narratives00:52:30 – Where Long-Term Power Accrues in the AI StackTranscriptLatent.Space - Inside AI's $10B+ Capital Flywheel — Martin Casado & Sarah Wang of a16z[00:00:00] Welcome to Latent Space (Live from a16z) + Meet the Guests[00:00:00] Alessio: Hey everyone. Welcome to the Latent Space podcast, live from a 16 z. Uh, this is Alessio founder Kernel Lance, and I'm joined by Twix, editor of Latent Space.[00:00:08] swyx: Hey, hey, hey. Uh, and we're so glad to be on with you guys. Also a top AI podcast, uh, Martin Cado and Sarah Wang. Welcome, very[00:00:16] Martin Casado: happy to be here and welcome.[00:00:17] swyx: Yes, uh, we love this office. We love what you've done with the place. Uh, the new logo is everywhere now. It's, it's still getting, takes a while to get used to, but it reminds me of like sort of a callback to a more ambitious age, which I think is kind of[00:00:31] Martin Casado: definitely makes a statement.[00:00:33] swyx: Yeah.[00:00:34] Martin Casado: Not quite sure what that statement is, but it makes a statement.[00:00:37] swyx: Uh, Martin, I go back with you to Netlify.[00:00:40] Martin Casado: Yep.[00:00:40] swyx: Uh, and, uh, you know, you create a software defined networking and all, all that stuff people can read up on your background. Yep. Sarah, I'm newer to you. Uh, you, you sort of started working together on AI infrastructure stuff.[00:00:51] Sarah Wang: That's right. Yeah. Seven, seven years ago now.[00:00:53] Martin Casado: Best growth investor in the entire industry.[00:00:55] swyx: Oh, say[00:00:56] Martin Casado: more hands down there is, there is. [00:01:00] I mean, when it comes to AI companies, Sarah, I think has done the most kind of aggressive, um, investment thesis around AI models, right? So, worked for Nom Ja, Mira Ia, FEI Fey, and so just these frontier, kind of like large AI models.[00:01:15] I think, you know, Sarah's been the, the broadest investor. Is that fair?[00:01:20] Venture vs. Growth in the Frontier Model Era[00:01:20] Sarah Wang: No, I, well, I was gonna say, I think it's been a really interesting tag, tag team actually just ‘cause the, a lot of these big C deals, not only are they raising a lot of money, um, it's still a tech founder bet, which obviously is inherently early stage.[00:01:33] But the resources,[00:01:36] Martin Casado: so many, I[00:01:36] Sarah Wang: was gonna say the resources one, they just grow really quickly. But then two, the resources that they need day one are kind of growth scale. So I, the hybrid tag team that we have is. Quite effective, I think,[00:01:46] Martin Casado: what is growth these days? You know, you don't wake up if it's less than a billion or like, it's, it's actually, it's actually very like, like no, it's a very interesting time in investing because like, you know, take like the character around, right?[00:01:59] These tend to [00:02:00] be like pre monetization, but the dollars are large enough that you need to have a larger fund and the analysis. You know, because you've got lots of users. ‘cause this stuff has such high demand requires, you know, more of a number sophistication. And so most of these deals, whether it's US or other firms on these large model companies, are like this hybrid between venture growth.[00:02:18] Sarah Wang: Yeah. Total. And I think, you know, stuff like BD for example, you wouldn't usually need BD when you were seed stage trying to get market biz Devrel. Biz Devrel, exactly. Okay. But like now, sorry, I'm,[00:02:27] swyx: I'm not familiar. What, what, what does biz Devrel mean for a venture fund? Because I know what biz Devrel means for a company.[00:02:31] Sarah Wang: Yeah.[00:02:32] Compute Deals, Strategics, and the ‘Circular Funding' Question[00:02:32] Sarah Wang: You know, so a, a good example is, I mean, we talk about buying compute, but there's a huge negotiation involved there in terms of, okay, do you get equity for the compute? What, what sort of partner are you looking at? Is there a go-to market arm to that? Um, and these are just things on this scale, hundreds of millions, you know, maybe.[00:02:50] Six months into the inception of a company, you just wouldn't have to negotiate these deals before.[00:02:54] Martin Casado: Yeah. These large rounds are very complex now. Like in the past, if you did a series A [00:03:00] or a series B, like whatever, you're writing a 20 to a $60 million check and you call it a day. Now you normally have financial investors and strategic investors, and then the strategic portion always still goes with like these kind of large compute contracts, which can take months to do.[00:03:13] And so it's, it's very different ties. I've been doing this for 10 years. It's the, I've never seen anything like this.[00:03:19] swyx: Yeah. Do you have worries about the circular funding from so disease strategics?[00:03:24] Martin Casado: I mean, listen, as long as the demand is there, like the demand is there. Like the problem with the internet is the demand wasn't there.[00:03:29] swyx: Exactly. All right. This, this is like the, the whole pyramid scheme bubble thing, where like, as long as you mark to market on like the notional value of like, these deals, fine, but like once it starts to chip away, it really Well[00:03:41] Martin Casado: no, like as, as, as, as long as there's demand. I mean, you know, this, this is like a lot of these sound bites have already become kind of cliches, but they're worth saying it.[00:03:47] Right? Like during the internet days, like we were. Um, raising money to put fiber in the ground that wasn't used. And that's a problem, right? Because now you actually have a supply overhang.[00:03:58] swyx: Mm-hmm.[00:03:59] Martin Casado: And even in the, [00:04:00] the time of the, the internet, like the supply and, and bandwidth overhang, even as massive as it was in, as massive as the crash was only lasted about four years.[00:04:09] But we don't have a supply overhang. Like there's no dark GPUs, right? I mean, and so, you know, circular or not, I mean, you know, if, if someone invests in a company that, um. You know, they'll actually use the GPUs. And on the other side of it is the, is the ask for customer. So I I, I think it's a different time.[00:04:25] Sarah Wang: I think the other piece, maybe just to add onto this, and I'm gonna quote Martine in front of him, but this is probably also a unique time in that. For the first time, you can actually trace dollars to outcomes. Yeah, right. Provided that scaling laws are, are holding, um, and capabilities are actually moving forward.[00:04:40] Because if you can put translate dollars into capabilities, uh, a capability improvement, there's demand there to martine's point. But if that somehow breaks, you know, obviously that's an important assumption in this whole thing to make it work. But you know, instead of investing dollars into sales and marketing, you're, you're investing into r and d to get to the capability, um, you know, increase.[00:04:59] And [00:05:00] that's sort of been the demand driver because. Once there's an unlock there, people are willing to pay for it.[00:05:05] Alessio: Yeah.[00:05:06] Blurring Lines: Models as Infra + Apps, and the New Fundraising Flywheel[00:05:06] Alessio: Is there any difference in how you built the portfolio now that some of your growth companies are, like the infrastructure of the early stage companies, like, you know, OpenAI is now the same size as some of the cloud providers were early on.[00:05:16] Like what does that look like? Like how much information can you feed off each other between the, the two?[00:05:24] Martin Casado: There's so many lines that are being crossed right now, or blurred. Right. So we already talked about venture and growth. Another one that's being blurred is between infrastructure and apps, right? So like what is a model company?[00:05:35] Mm-hmm. Like, it's clearly infrastructure, right? Because it's like, you know, it's doing kind of core r and d. It's a horizontal platform, but it's also an app because it's um, uh, touches the users directly. And then of course. You know, the, the, the growth of these is just so high. And so I actually think you're just starting to see a, a, a new financing strategy emerge and, you know, we've had to adapt as a result of that.[00:05:59] And [00:06:00] so there's been a lot of changes. Um, you're right that these companies become platform companies very quickly. You've got ecosystem build out. So none of this is necessarily new, but the timescales of which it's happened is pretty phenomenal. And the way we'd normally cut lines before is blurred a little bit, but.[00:06:16] But that, that, that said, I mean, a lot of it also just does feel like things that we've seen in the past, like cloud build out the internet build out as well.[00:06:24] Sarah Wang: Yeah. Um, yeah, I think it's interesting, uh, I don't know if you guys would agree with this, but it feels like the emerging strategy is, and this builds off of your other question, um.[00:06:33] You raise money for compute, you pour that or you, you pour the money into compute, you get some sort of breakthrough. You funnel the breakthrough into your vertically integrated application. That could be chat GBT, that could be cloud code, you know, whatever it is. You massively gain share and get users.[00:06:49] Maybe you're even subsidizing at that point. Um, depending on your strategy. You raise money at the peak momentum and then you repeat, rinse and repeat. Um, and so. And that wasn't [00:07:00] true even two years ago, I think. Mm-hmm. And so it's sort of to your, just tying it to fundraising strategy, right? There's a, and hiring strategy.[00:07:07] All of these are tied, I think the lines are blurring even more today where everyone is, and they, but of course these companies all have API businesses and so they're these, these frenemy lines that are getting blurred in that a lot of, I mean, they have billions of dollars of API revenue, right? And so there are customers there.[00:07:23] But they're competing on the app layer.[00:07:24] Martin Casado: Yeah. So this is a really, really important point. So I, I would say for sure, venture and growth, that line is blurry app and infrastructure. That line is blurry. Um, but I don't think that that changes our practice so much. But like where the very open questions are like, does this layer in the same way.[00:07:43] Compute traditionally has like during the cloud is like, you know, like whatever, somebody wins one layer, but then another whole set of companies wins another layer. But that might not, might not be the case here. It may be the case that you actually can't verticalize on the token string. Like you can't build an app like it, it necessarily goes down just because there are no [00:08:00] abstractions.[00:08:00] So those are kinda the bigger existential questions we ask. Another thing that is very different this time than in the history of computer sciences is. In the past, if you raised money, then you basically had to wait for engineering to catch up. Which famously doesn't scale like the mythical mammoth. It take a very long time.[00:08:18] But like that's not the case here. Like a model company can raise money and drop a model in a, in a year, and it's better, right? And, and it does it with a team of 20 people or 10 people. So this type of like money entering a company and then producing something that has demand and growth right away and using that to raise more money is a very different capital flywheel than we've ever seen before.[00:08:39] And I think everybody's trying to understand what the consequences are. So I think it's less about like. Big companies and growth and this, and more about these more systemic questions that we actually don't have answers to.[00:08:49] Alessio: Yeah, like at Kernel Labs, one of our ideas is like if you had unlimited money to spend productively to turn tokens into products, like the whole early stage [00:09:00] market is very different because today you're investing X amount of capital to win a deal because of price structure and whatnot, and you're kind of pot committing.[00:09:07] Yeah. To a certain strategy for a certain amount of time. Yeah. But if you could like iteratively spin out companies and products and just throw, I, I wanna spend a million dollar of inference today and get a product out tomorrow.[00:09:18] swyx: Yeah.[00:09:19] Alessio: Like, we should get to the point where like the friction of like token to product is so low that you can do this and then you can change the Right, the early stage venture model to be much more iterative.[00:09:30] And then every round is like either 100 k of inference or like a hundred million from a 16 Z. There's no, there's no like $8 million C round anymore. Right.[00:09:38] When Frontier Labs Outspend the Entire App Ecosystem[00:09:38] Martin Casado: But, but, but, but there's a, there's a, the, an industry structural question that we don't know the answer to, which involves the frontier models, which is, let's take.[00:09:48] Anthropic it. Let's say Anthropic has a state-of-the-art model that has some large percentage of market share. And let's say that, uh, uh, uh, you know, uh, a company's building smaller models [00:10:00] that, you know, use the bigger model in the background, open 4.5, but they add value on top of that. Now, if Anthropic can raise three times more.[00:10:10] Every subsequent round, they probably can raise more money than the entire app ecosystem that's built on top of it. And if that's the case, they can expand beyond everything built on top of it. It's like imagine like a star that's just kind of expanding, so there could be a systemic. There could be a, a systemic situation where the soda models can raise so much money that they can out pay anybody that bills on top of ‘em, which would be something I don't think we've ever seen before just because we were so bottlenecked in engineering, and this is a very open question.[00:10:41] swyx: Yeah. It's, it is almost like bitter lesson applied to the startup industry.[00:10:45] Martin Casado: Yeah, a hundred percent. It literally becomes an issue of like raise capital, turn that directly into growth. Use that to raise three times more. Exactly. And if you can keep doing that, you literally can outspend any company that's built the, not any company.[00:10:57] You can outspend the aggregate of companies on top of [00:11:00] you and therefore you'll necessarily take their share, which is crazy.[00:11:02] swyx: Would you say that kind of happens in character? Is that the, the sort of postmortem on. What happened?[00:11:10] Sarah Wang: Um,[00:11:10] Martin Casado: no.[00:11:12] Sarah Wang: Yeah, because I think so,[00:11:13] swyx: I mean the actual postmortem is, he wanted to go back to Google.[00:11:15] Exactly. But like[00:11:18] Martin Casado: that's another difference that[00:11:19] Sarah Wang: you said[00:11:21] Martin Casado: it. We should talk, we should actually talk about that.[00:11:22] swyx: Yeah,[00:11:22] Sarah Wang: that's[00:11:23] swyx: Go for it. Take it. Take,[00:11:23] Sarah Wang: yeah.[00:11:24] Character.AI, Founder Goals (AGI vs Product), and GPU Allocation Tradeoffs[00:11:24] Sarah Wang: I was gonna say, I think, um. The, the, the character thing raises actually a different issue, which actually the Frontier Labs will face as well. So we'll see how they handle it.[00:11:34] But, um, so we invest in character in January, 2023, which feels like eons ago, I mean, three years ago. Feels like lifetimes ago. But, um, and then they, uh, did the IP licensing deal with Google in August, 2020. Uh, four. And so, um, you know, at the time, no, you know, he's talked publicly about this, right? He wanted to Google wouldn't let him put out products in the world.[00:11:56] That's obviously changed drastically. But, um, he went to go do [00:12:00] that. Um, but he had a product attached. The goal was, I mean, it's Nome Shair, he wanted to get to a GI. That was always his personal goal. But, you know, I think through collecting data, right, and this sort of very human use case, that the character product.[00:12:13] Originally was and still is, um, was one of the vehicles to do that. Um, I think the real reason that, you know. I if you think about the, the stress that any company feels before, um, you ultimately going one way or the other is sort of this a GI versus product. Um, and I think a lot of the big, I think, you know, opening eyes, feeling that, um, anthropic if they haven't started, you know, felt it, certainly given the success of their products, they may start to feel that soon.[00:12:39] And the real. I think there's real trade-offs, right? It's like how many, when you think about GPUs, that's a limited resource. Where do you allocate the GPUs? Is it toward the product? Is it toward new re research? Right? Is it, or long-term research, is it toward, um, n you know, near to midterm research? And so, um, in a case where you're resource constrained, um, [00:13:00] of course there's this fundraising game you can play, right?[00:13:01] But the fund, the market was very different back in 2023 too. Um. I think the best researchers in the world have this dilemma of, okay, I wanna go all in on a GI, but it's the product usage revenue flywheel that keeps the revenue in the house to power all the GPUs to get to a GI. And so it does make, um, you know, I think it sets up an interesting dilemma for any startup that has trouble raising up until that level, right?[00:13:27] And certainly if you don't have that progress, you can't continue this fly, you know, fundraising flywheel.[00:13:32] Martin Casado: I would say that because, ‘cause we're keeping track of all of the things that are different, right? Like, you know, venture growth and uh, app infra and one of the ones is definitely the personalities of the founders.[00:13:45] It's just very different this time I've been. Been doing this for a decade and I've been doing startups for 20 years. And so, um, I mean a lot of people start this to do a GI and we've never had like a unified North star that I recall in the same [00:14:00] way. Like people built companies to start companies in the past.[00:14:02] Like that was what it was. Like I would create an internet company, I would create infrastructure company, like it's kind of more engineering builders and this is kind of a different. You know, mentality. And some companies have harnessed that incredibly well because their direction is so obviously on the path to what somebody would consider a GI, but others have not.[00:14:20] And so like there is always this tension with personnel. And so I think we're seeing more kind of founder movement.[00:14:27] Sarah Wang: Yeah.[00:14:27] Martin Casado: You know, as a fraction of founders than we've ever seen. I mean, maybe since like, I don't know the time of like Shockly and the trade DUR aid or something like that. Way back in the beginning of the industry, I, it's a very, very.[00:14:38] Unusual time of personnel.[00:14:39] Sarah Wang: Totally.[00:14:40] Talent Wars, Mega-Comp, and the Rise of Acquihire M&A[00:14:40] Sarah Wang: And it, I think it's exacerbated by the fact that talent wars, I mean, every industry has talent wars, but not at this magnitude, right? No. Yeah. Very rarely can you see someone get poached for $5 billion. That's hard to compete with. And then secondly, if you're a founder in ai, you could fart and it would be on the front page of, you know, the information these days.[00:14:59] And so there's [00:15:00] sort of this fishbowl effect that I think adds to the deep anxiety that, that these AI founders are feeling.[00:15:06] Martin Casado: Hmm.[00:15:06] swyx: Uh, yes. I mean, just on, uh, briefly comment on the founder, uh, the sort of. Talent wars thing. I feel like 2025 was just like a blip. Like I, I don't know if we'll see that again.[00:15:17] ‘cause meta built the team. Like, I don't know if, I think, I think they're kind of done and like, who's gonna pay more than meta? I, I don't know.[00:15:23] Martin Casado: I, I agree. So it feels so, it feel, it feels this way to me too. It's like, it is like, basically Zuckerberg kind of came out swinging and then now he's kind of back to building.[00:15:30] Yeah,[00:15:31] swyx: yeah. You know, you gotta like pay up to like assemble team to rush the job, whatever. But then now, now you like you, you made your choices and now they got a ship.[00:15:38] Martin Casado: I mean, the, the o other side of that is like, you know, like we're, we're actually in the job hiring market. We've got 600 people here. I hire all the time.[00:15:44] I've got three open recs if anybody's interested, that's listening to this for investor. Yeah, on, on the team, like on the investing side of the team, like, and, um, a lot of the people we talk to have acting, you know, active, um, offers for 10 million a year or something like that. And like, you know, and we pay really, [00:16:00] really well.[00:16:00] And just to see what's out on the market is really, is really remarkable. And so I would just say it's actually, so you're right, like the really flashy one, like I will get someone for, you know, a billion dollars, but like the inflated, um, uh, trickles down. Yeah, it is still very active today. I mean,[00:16:18] Sarah Wang: yeah, you could be an L five and get an offer in the tens of millions.[00:16:22] Okay. Yeah. Easily. Yeah. It's so I think you're right that it felt like a blip. I hope you're right. Um, but I think it's been, the steady state is now, I think got pulled up. Yeah. Yeah. I'll pull up for[00:16:31] Martin Casado: sure. Yeah.[00:16:32] Alessio: Yeah. And I think that's breaking the early stage founder math too. I think before a lot of people would be like, well, maybe I should just go be a founder instead of like getting paid.[00:16:39] Yeah. 800 KA million at Google. But if I'm getting paid. Five, 6 million. That's different but[00:16:45] Martin Casado: on. But on the other hand, there's more strategic money than we've ever seen historically, right? Mm-hmm. And so, yep. The economics, the, the, the, the calculus on the economics is very different in a number of ways. And, uh, it's crazy.[00:16:58] It's cra it's causing like a, [00:17:00] a, a, a ton of change in confusion in the market. Some very positive, sub negative, like, so for example, the other side of the, um. The co-founder, like, um, acquisition, you know, mark Zuckerberg poaching someone for a lot of money is like, we were actually seeing historic amount of m and a for basically acquihires, right?[00:17:20] That you like, you know, really good outcomes from a venture perspective that are effective acquihires, right? So I would say it's probably net positive from the investment standpoint, even though it seems from the headlines to be very disruptive in a negative way.[00:17:33] Alessio: Yeah.[00:17:33] What's Underfunded: Boring Software, Robotics Skepticism, and Custom Silicon Economics[00:17:33] Alessio: Um, let's talk maybe about what's not being invested in, like maybe some interesting ideas that you would see more people build or it, it seems in a way, you know, as ycs getting more popular, it's like access getting more popular.[00:17:47] There's a startup school path that a lot of founders take and they know what's hot in the VC circles and they know what gets funded. Uh, and there's maybe not as much risk appetite for. Things outside of that. Um, I'm curious if you feel [00:18:00] like that's true and what are maybe, uh, some of the areas, uh, that you think are under discussed?[00:18:06] Martin Casado: I mean, I actually think that we've taken our eye off the ball in a lot of like, just traditional, you know, software companies. Um, so like, I mean. You know, I think right now there's almost a barbell, like you're like the hot thing on X, you're deep tech.[00:18:21] swyx: Mm-hmm.[00:18:22] Martin Casado: Right. But I, you know, I feel like there's just kind of a long, you know, list of like good.[00:18:28] Good companies that will be around for a long time in very large markets. Say you're building a database, you know, say you're building, um, you know, kind of monitoring or logging or tooling or whatever. There's some good companies out there right now, but like, they have a really hard time getting, um, the attention of investors.[00:18:43] And it's almost become a meme, right? Which is like, if you're not basically growing from zero to a hundred in a year, you're not interesting, which is just, is the silliest thing to say. I mean, think of yourself as like an introvert person, like, like your personal money, right? Mm-hmm. So. Your personal money, will you put it in the stock market at 7% or you put it in this company growing five x in a very large [00:19:00] market?[00:19:00] Of course you can put it in the company five x. So it's just like we say these stupid things, like if you're not going from zero to a hundred, but like those, like who knows what the margins of those are mean. Clearly these are good investments. True for anybody, right? True. Like our LPs want whatever.[00:19:12] Three x net over, you know, the life cycle of a fund, right? So a, a company in a big market growing five X is a great investment. We'd, everybody would be happy with these returns, but we've got this kind of mania on these, these strong growths. And so I would say that that's probably the most underinvested sector.[00:19:28] Right now.[00:19:29] swyx: Boring software, boring enterprise software.[00:19:31] Martin Casado: Traditional. Really good company.[00:19:33] swyx: No, no AI here.[00:19:34] Martin Casado: No. Like boring. Well, well, the AI of course is pulling them into use cases. Yeah, but that's not what they're, they're not on the token path, right? Yeah. Let's just say that like they're software, but they're not on the token path.[00:19:41] Like these are like they're great investments from any definition except for like random VC on Twitter saying VC on x, saying like, it's not growing fast enough. What do you[00:19:52] Sarah Wang: think? Yeah, maybe I'll answer a slightly different. Question, but adjacent to what you asked, um, which is maybe an area that we're not, uh, investing [00:20:00] right now that I think is a question and we're spending a lot of time in regardless of whether we pull the trigger or not.[00:20:05] Um, and it would probably be on the hardware side, actually. Robotics, right? And the robotics side. Robotics. Right. Which is, it's, I don't wanna say that it's not getting funding ‘cause it's clearly, uh, it's, it's sort of non-consensus to almost not invest in robotics at this point. But, um, we spent a lot of time in that space and I think for us, we just haven't seen the chat GPT moment.[00:20:22] Happen on the hardware side. Um, and the funding going into it feels like it's already. Taking that for granted.[00:20:30] Martin Casado: Yeah. Yeah. But we also went through the drone, you know, um, there's a zip line right, right out there. What's that? Oh yeah, there's a zip line. Yeah. What the drone, what the av And like one of the takeaways is when it comes to hardware, um, most companies will end up verticalizing.[00:20:46] Like if you're. If you're investing in a robot company for an A for agriculture, you're investing in an ag company. ‘cause that's the competition and that's surprising. And that's supply chain. And if you're doing it for mining, that's mining. And so the ad team does a lot of that type of stuff ‘cause they actually set up to [00:21:00] diligence that type of work.[00:21:01] But for like horizontal technology investing, there's very little when it comes to robots just because it's so fit for, for purpose. And so we kinda like to look at software. Solutions or horizontal solutions like applied intuition. Clearly from the AV wave deep map, clearly from the AV wave, I would say scale AI was actually a horizontal one for That's fair, you know, for robotics early on.[00:21:23] And so that sort of thing we're very, very interested. But the actual like robot interacting with the world is probably better for different team. Agree.[00:21:30] Alessio: Yeah, I'm curious who these teams are supposed to be that invest in them. I feel like everybody's like, yeah, robotics, it's important and like people should invest in it.[00:21:38] But then when you look at like the numbers, like the capital requirements early on versus like the moment of, okay, this is actually gonna work. Let's keep investing. That seems really hard to predict in a way that is not,[00:21:49] Martin Casado: I think co, CO two, kla, gc, I mean these are all invested in in Harvard companies. He just, you know, and [00:22:00] listen, I mean, it could work this time for sure.[00:22:01] Right? I mean if Elon's doing it, he's like, right. Just, just the fact that Elon's doing it means that there's gonna be a lot of capital and a lot of attempts for a long period of time. So that alone maybe suggests that we should just be investing in robotics just ‘cause you have this North star who's Elon with a humanoid and that's gonna like basically willing into being an industry.[00:22:17] Um, but we've just historically found like. We're a huge believer that this is gonna happen. We just don't feel like we're in a good position to diligence these things. ‘cause again, robotics companies tend to be vertical. You really have to understand the market they're being sold into. Like that's like that competitive equilibrium with a human being is what's important.[00:22:34] It's not like the core tech and like we're kind of more horizontal core tech type investors. And this is Sarah and I. Yeah, the ad team is different. They can actually do these types of things.[00:22:42] swyx: Uh, just to clarify, AD stands for[00:22:44] Martin Casado: American Dynamism.[00:22:45] swyx: Alright. Okay. Yeah, yeah, yeah. Uh, I actually, I do have a related question that, first of all, I wanna acknowledge also just on the, on the chip side.[00:22:51] Yeah. I, I recall a podcast that where you were on, i, I, I think it was the a CC podcast, uh, about two or three years ago where you, where you suddenly said [00:23:00] something, which really stuck in my head about how at some point, at some point kind of scale it makes sense to. Build a custom aic Yes. For per run.[00:23:07] Martin Casado: Yes.[00:23:07] It's crazy. Yeah.[00:23:09] swyx: We're here and I think you, you estimated 500 billion, uh, something.[00:23:12] Martin Casado: No, no, no. A billion, a billion dollar training run of $1 billion training run. It makes sense to actually do a custom meic if you can do it in time. The question now is timelines. Yeah, but not money because just, just, just rough math.[00:23:22] If it's a billion dollar training. Then the inference for that model has to be over a billion, otherwise it won't be solvent. So let's assume it's, if you could save 20%, which you could save much more than that with an ASIC 20%, that's $200 million. You can tape out a chip for $200 million. Right? So now you can literally like justify economically, not timeline wise.[00:23:41] That's a different issue. An ASIC per model, which[00:23:44] swyx: is because that, that's how much we leave on the table every single time. We, we, we do like generic Nvidia.[00:23:48] Martin Casado: Exactly. Exactly. No, it, it is actually much more than that. You could probably get, you know, a factor of two, which would be 500 million.[00:23:54] swyx: Typical MFU would be like 50.[00:23:55] Yeah, yeah. And that's good.[00:23:57] Martin Casado: Exactly. Yeah. Hundred[00:23:57] swyx: percent. Um, so, so, yeah, and I mean, and I [00:24:00] just wanna acknowledge like, here we are in, in, in 2025 and opening eyes confirming like Broadcom and all the other like custom silicon deals, which is incredible. I, I think that, uh, you know, speaking about ad there's, there's a really like interesting tie in that obviously you guys are hit on, which is like these sort, this sort of like America first movement or like sort of re industrialized here.[00:24:17] Yeah. Uh, move TSMC here, if that's possible. Um, how much overlap is there from ad[00:24:23] Martin Casado: Yeah.[00:24:23] swyx: To, I guess, growth and, uh, investing in particularly like, you know, US AI companies that are strongly bounded by their compute.[00:24:32] Martin Casado: Yeah. Yeah. So I mean, I, I would view, I would view AD as more as a market segmentation than like a mission, right?[00:24:37] So the market segmentation is, it has kind of regulatory compliance issues or government, you know, sale or it deals with like hardware. I mean, they're just set up to, to, to, to, to. To diligence those types of companies. So it's a more of a market segmentation thing. I would say the entire firm. You know, which has been since it is been intercepted, you know, has geographical biases, right?[00:24:58] I mean, for the longest time we're like, you [00:25:00] know, bay Area is gonna be like, great, where the majority of the dollars go. Yeah. And, and listen, there, there's actually a lot of compounding effects for having a geographic bias. Right. You know, everybody's in the same place. You've got an ecosystem, you're there, you've got presence, you've got a network.[00:25:12] Um, and, uh, I mean, I would say the Bay area's very much back. You know, like I, I remember during pre COVID, like it was like almost Crypto had kind of. Pulled startups away. Miami from the Bay Area. Miami, yeah. Yeah. New York was, you know, because it's so close to finance, came up like Los Angeles had a moment ‘cause it was so close to consumer, but now it's kind of come back here.[00:25:29] And so I would say, you know, we tend to be very Bay area focused historically, even though of course we've asked all over the world. And then I would say like, if you take the ring out, you know, one more, it's gonna be the US of course, because we know it very well. And then one more is gonna be getting us and its allies and Yeah.[00:25:44] And it goes from there.[00:25:45] Sarah Wang: Yeah,[00:25:45] Martin Casado: sorry.[00:25:46] Sarah Wang: No, no. I agree. I think from a, but I think from the intern that that's sort of like where the companies are headquartered. Maybe your questions on supply chain and customer base. Uh, I, I would say our customers are, are, our companies are fairly international from that perspective.[00:25:59] Like they're selling [00:26:00] globally, right? They have global supply chains in some cases.[00:26:03] Martin Casado: I would say also the stickiness is very different.[00:26:05] Sarah Wang: Yeah.[00:26:05] Martin Casado: Historically between venture and growth, like there's so much company building in venture, so much so like hiring the next PM. Introducing the customer, like all of that stuff.[00:26:15] Like of course we're just gonna be stronger where we have our network and we've been doing business for 20 years. I've been in the Bay Area for 25 years, so clearly I'm just more effective here than I would be somewhere else. Um, where I think, I think for some of the later stage rounds, the companies don't need that much help.[00:26:30] They're already kind of pretty mature historically, so like they can kind of be everywhere. So there's kind of less of that stickiness. This is different in the AI time. I mean, Sarah is now the, uh, chief of staff of like half the AI companies in, uh, in the Bay Area right now. She's like, ops Ninja Biz, Devrel, BizOps.[00:26:48] swyx: Are, are you, are you finding much AI automation in your work? Like what, what is your stack.[00:26:53] Sarah Wang: Oh my, in my personal stack.[00:26:54] swyx: I mean, because like, uh, by the way, it's the, the, the reason for this is it is triggering, uh, yeah. We, like, I'm hiring [00:27:00] ops, ops people. Um, a lot of ponders I know are also hiring ops people and I'm just, you know, it's opportunity Since you're, you're also like basically helping out with ops with a lot of companies.[00:27:09] What are people doing these days? Because it's still very manual as far as I can tell.[00:27:13] Sarah Wang: Hmm. Yeah. I think the things that we help with are pretty network based, um, in that. It's sort of like, Hey, how do do I shortcut this process? Well, let's connect you to the right person. So there's not quite an AI workflow for that.[00:27:26] I will say as a growth investor, Claude Cowork is pretty interesting. Yeah. Like for the first time, you can actually get one shot data analysis. Right. Which, you know, if you're gonna do a customer database, analyze a cohort retention, right? That's just stuff that you had to do by hand before. And our team, the other, it was like midnight and the three of us were playing with Claude Cowork.[00:27:47] We gave it a raw file. Boom. Perfectly accurate. We checked the numbers. It was amazing. That was my like, aha moment. That sounds so boring. But you know, that's, that's the kind of thing that a growth investor is like, [00:28:00] you know, slaving away on late at night. Um, done in a few seconds.[00:28:03] swyx: Yeah. You gotta wonder what the whole, like, philanthropic labs, which is like their new sort of products studio.[00:28:10] Yeah. What would that be worth as an independent, uh, startup? You know, like a[00:28:14] Martin Casado: lot.[00:28:14] Sarah Wang: Yeah, true.[00:28:16] swyx: Yeah. You[00:28:16] Martin Casado: gotta hand it to them. They've been executing incredibly well.[00:28:19] swyx: Yeah. I, I mean, to me, like, you know, philanthropic, like building on cloud code, I think, uh, it makes sense to me the, the real. Um, pedal to the metal, whatever the, the, the phrase is, is when they start coming after consumer with, uh, against OpenAI and like that is like red alert at Open ai.[00:28:35] Oh, I[00:28:35] Martin Casado: think they've been pretty clear. They're enterprise focused.[00:28:37] swyx: They have been, but like they've been free. Here's[00:28:40] Martin Casado: care publicly,[00:28:40] swyx: it's enterprise focused. It's coding. Right. Yeah.[00:28:43] AI Labs vs Startups: Disruption, Undercutting & the Innovator's Dilemma[00:28:43] swyx: And then, and, but here's cloud, cloud, cowork, and, and here's like, well, we, uh, they, apparently they're running Instagram ads for Claudia.[00:28:50] I, on, you know, for, for people on, I get them all the time. Right. And so, like,[00:28:54] Martin Casado: uh,[00:28:54] swyx: it, it's kind of like this, the disruption thing of, uh, you know. Mo Open has been doing, [00:29:00] consumer been doing the, just pursuing general intelligence in every mo modality, and here's a topic that only focus on this thing, but now they're sort of undercutting and doing the whole innovator's dilemma thing on like everything else.[00:29:11] Martin Casado: It's very[00:29:11] swyx: interesting.[00:29:12] Martin Casado: Yeah, I mean there's, there's a very open que so for me there's like, do you know that meme where there's like the guy in the path and there's like a path this way? There's a path this way. Like one which way Western man. Yeah. Yeah.[00:29:23] Two Futures for AI: Infinite Market vs AGI Oligopoly[00:29:23] Martin Casado: And for me, like, like all the entire industry kind of like hinges on like two potential futures.[00:29:29] So in, in one potential future, um, the market is infinitely large. There's perverse economies of scale. ‘cause as soon as you put a model out there, like it kind of sublimates and all the other models catch up and like, it's just like software's being rewritten and fractured all over the place and there's tons of upside and it just grows.[00:29:48] And then there's another path which is like, well. Maybe these models actually generalize really well, and all you have to do is train them with three times more money. That's all you have to [00:30:00] do, and it'll just consume everything beyond it. And if that's the case, like you end up with basically an oligopoly for everything, like, you know mm-hmm.[00:30:06] Because they're perfectly general and like, so this would be like the, the a GI path would be like, these are perfectly general. They can do everything. And this one is like, this is actually normal software. The universe is complicated. You've got, and nobody knows the answer.[00:30:18] The Economics Reality Check: Gross Margins, Training Costs & Borrowing Against the Future[00:30:18] Martin Casado: My belief is if you actually look at the numbers of these companies, so generally if you look at the numbers of these companies, if you look at like the amount they're making and how much they, they spent training the last model, they're gross margin positive.[00:30:30] You're like, oh, that's really working. But if you look at like. The current training that they're doing for the next model, their gross margin negative. So part of me thinks that a lot of ‘em are kind of borrowing against the future and that's gonna have to slow down. It's gonna catch up to them at some point in time, but we don't really know.[00:30:47] Sarah Wang: Yeah.[00:30:47] Martin Casado: Does that make sense? Like, I mean, it could be, it could be the case that the only reason this is working is ‘cause they can raise that next round and they can train that next model. ‘cause these models have such a short. Life. And so at some point in time, like, you know, they won't be able to [00:31:00] raise that next round for the next model and then things will kind of converge and fragment again.[00:31:03] But right now it's not.[00:31:04] Sarah Wang: Totally. I think the other, by the way, just, um, a meta point. I think the other lesson from the last three years is, and we talk about this all the time ‘cause we're on this. Twitter X bubble. Um, cool. But, you know, if you go back to, let's say March, 2024, that period, it felt like a, I think an open source model with an, like a, you know, benchmark leading capability was sort of launching on a daily basis at that point.[00:31:27] And, um, and so that, you know, that's one period. Suddenly it's sort of like open source takes over the world. There's gonna be a plethora. It's not an oligopoly, you know, if you fast, you know, if you, if you rewind time even before that GPT-4 was number one for. Nine months, 10 months. It's a long time. Right.[00:31:44] Um, and of course now we're in this era where it feels like an oligopoly, um, maybe some very steady state shifts and, and you know, it could look like this in the future too, but it just, it's so hard to call. And I think the thing that keeps, you know, us up at [00:32:00] night in, in a good way and bad way, is that the capability progress is actually not slowing down.[00:32:06] And so until that happens, right, like you don't know what's gonna look like.[00:32:09] Martin Casado: But I, I would, I would say for sure it's not converged, like for sure, like the systemic capital flows have not converged, meaning right now it's still borrowing against the future to subsidize growth currently, which you can do that for a period of time.[00:32:23] But, but you know, at the end, at some point the market will rationalize that and just nobody knows what that will look like.[00:32:29] Alessio: Yeah.[00:32:29] Martin Casado: Or, or like the drop in price of compute will, will, will save them. Who knows?[00:32:34] Alessio: Yeah. Yeah. I think the models need to ask them to, to specific tasks. You know? It's like, okay, now Opus 4.5 might be a GI at some specific task, and now you can like depreciate the model over a longer time.[00:32:45] I think now, now, right now there's like no old model.[00:32:47] Martin Casado: No, but let, but lemme just change that mental, that's, that used to be my mental model. Lemme just change it a little bit.[00:32:53] Capital as a Weapon vs Task Saturation: Where Real Enterprise Value Gets Built[00:32:53] Martin Casado: If you can raise three times, if you can raise more than the aggregate of anybody that uses your models, that doesn't even matter.[00:32:59] It doesn't [00:33:00] even matter. See what I'm saying? Like, yeah. Yeah. So, so I have an API Business. My API business is 60% margin, or 70% margin, or 80% margin is a high margin business. So I know what everybody is using. If I can raise more money than the aggregate of everybody that's using it, I will consume them whether I'm a GI or not.[00:33:14] And I will know if they're using it ‘cause they're using it. And like, unlike in the past where engineering stops me from doing that.[00:33:21] Alessio: Mm-hmm.[00:33:21] Martin Casado: It is very straightforward. You just train. So I also thought it was kind of like, you must ask the code a GI, general, general, general. But I think there's also just a possibility that the, that the capital markets will just give them the, the, the ammunition to just go after everybody on top of ‘em.[00:33:36] Sarah Wang: I, I do wonder though, to your point, um, if there's a certain task that. Getting marginally better isn't actually that much better. Like we've asked them to it, to, you know, we can call it a GI or whatever, you know, actually, Ali Goi talks about this, like we're already at a GI for a lot of functions in the enterprise.[00:33:50] Um. That's probably those for those tasks, you probably could build very specific companies that focus on just getting as much value out of that task that isn't [00:34:00] coming from the model itself. There's probably a rich enterprise business to be built there. I mean, could be wrong on that, but there's a lot of interesting examples.[00:34:08] So, right, if you're looking the legal profession or, or whatnot, and maybe that's not a great one ‘cause the models are getting better on that front too, but just something where it's a bit saturated, then the value comes from. Services. It comes from implementation, right? It comes from all these things that actually make it useful to the end customer.[00:34:24] Martin Casado: Sorry, what am I, one more thing I think is, is underused in all of this is like, to what extent every task is a GI complete.[00:34:31] Sarah Wang: Mm-hmm.[00:34:32] Martin Casado: Yeah. I code every day. It's so fun.[00:34:35] Sarah Wang: That's a core question. Yeah.[00:34:36] Martin Casado: And like. When I'm talking to these models, it's not just code. I mean, it's everything, right? Like I, you know, like it's,[00:34:43] swyx: it's healthcare.[00:34:44] It's,[00:34:44] Martin Casado: I mean, it's[00:34:44] swyx: Mele,[00:34:45] Martin Casado: but it's every, it is exactly that. Like, yeah, that's[00:34:47] Sarah Wang: great support. Yeah.[00:34:48] Martin Casado: It's everything. Like I'm asking these models to, yeah, to understand compliance. I'm asking these models to go search the web. I'm asking these models to talk about things I know in the history, like it's having a full conversation with me while I, I engineer, and so it could be [00:35:00] the case that like, mm-hmm.[00:35:01] The most a, you know, a GI complete, like I'm not an a GI guy. Like I think that's, you know, but like the most a GI complete model will is win independent of the task. And we don't know the answer to that one either.[00:35:11] swyx: Yeah.[00:35:12] Martin Casado: But it seems to me that like, listen, codex in my experience is for sure better than Opus 4.5 for coding.[00:35:18] Like it finds the hardest bugs that I work in with. Like, it is, you know. The smartest developers. I don't work on it. It's great. Um, but I think Opus 4.5 is actually very, it's got a great bedside manner and it really, and it, it really matters if you're building something very complex because like, it really, you know, like you're, you're, you're a partner and a brainstorming partner for somebody.[00:35:38] And I think we don't discuss enough how every task kind of has that quality.[00:35:42] swyx: Mm-hmm.[00:35:43] Martin Casado: And what does that mean to like capital investment and like frontier models and Submodels? Yeah.[00:35:47] Why “Coding Models” Keep Collapsing into Generalists (Reasoning vs Taste)[00:35:47] Martin Casado: Like what happened to all the special coding models? Like, none of ‘em worked right. So[00:35:51] Alessio: some of them, they didn't even get released.[00:35:53] Magical[00:35:54] Martin Casado: Devrel. There's a whole, there's a whole host. We saw a bunch of them and like there's this whole theory that like, there could be, and [00:36:00] I think one of the conclusions is, is like there's no such thing as a coding model,[00:36:04] Alessio: you know?[00:36:04] Martin Casado: Like, that's not a thing. Like you're talking to another human being and it's, it's good at coding, but like it's gotta be good at everything.[00:36:10] swyx: Uh, minor disagree only because I, I'm pretty like, have pretty high confidence that basically open eye will always release a GPT five and a GT five codex. Like that's the code's. Yeah. The way I call it is one for raisin, one for Tiz. Um, and, and then like someone internal open, it was like, yeah, that's a good way to frame it.[00:36:32] Martin Casado: That's so funny.[00:36:33] swyx: Uh, but maybe it, maybe it collapses down to reason and that's it. It's not like a hundred dimensions doesn't life. Yeah. It's two dimensions. Yeah, yeah, yeah, yeah. Like and exactly. Beside manner versus coding. Yeah.[00:36:43] Martin Casado: Yeah.[00:36:44] swyx: It's, yeah.[00:36:46] Martin Casado: I, I think for, for any, it's hilarious. For any, for anybody listening to this for, for, for, I mean, for you, like when, when you're like coding or using these models for something like that.[00:36:52] Like actually just like be aware of how much of the interaction has nothing to do with coding and it just turns out to be a large portion of it. And so like, you're, I [00:37:00] think like, like the best Soto ish model. You know, it is going to remain very important no matter what the task is.[00:37:06] swyx: Yeah.[00:37:07] What He's Actually Coding: Gaussian Splats, Spark.js & 3D Scene Rendering Demos[00:37:07] swyx: Uh, speaking of coding, uh, I, I'm gonna be cheeky and ask like, what actually are you coding?[00:37:11] Because obviously you, you could code anything and you are obviously a busy investor and a manager of the good. Giant team. Um, what are you calling?[00:37:18] Martin Casado: I help, um, uh, FEFA at World Labs. Uh, it's one of the investments and um, and they're building a foundation model that creates 3D scenes.[00:37:27] swyx: Yeah, we had it on the pod.[00:37:28] Yeah. Yeah,[00:37:28] Martin Casado: yeah. And so these 3D scenes are Gaussian splats, just by the way that kind of AI works. And so like, you can reconstruct a scene better with, with, with radiance feels than with meshes. ‘cause like they don't really have topology. So, so they, they, they produce each. Beautiful, you know, 3D rendered scenes that are Gaussian splats, but the actual industry support for Gaussian splats isn't great.[00:37:50] It's just never, you know, it's always been meshes and like, things like unreal use meshes. And so I work on a open source library called Spark js, which is a. Uh, [00:38:00] a JavaScript rendering layer ready for Gaussian splats. And it's just because, you know, um, you, you, you need that support and, and right now there's kind of a three js moment that's all meshes and so like, it's become kind of the default in three Js ecosystem.[00:38:13] As part of that to kind of exercise the library, I just build a whole bunch of cool demos. So if you see me on X, you see like all my demos and all the world building, but all of that is just to exercise this, this library that I work on. ‘cause it's actually a very tough algorithmics problem to actually scale a library that much.[00:38:29] And just so you know, this is ancient history now, but 30 years ago I paid for undergrad, you know, working on game engines in college in the late nineties. So I've got actually a back and it's very old background, but I actually have a background in this and so a lot of it's fun. You know, but, but the, the, the, the whole goal is just for this rendering library to, to,[00:38:47] Sarah Wang: are you one of the most active contributors?[00:38:49] The, their GitHub[00:38:50] Martin Casado: spark? Yes.[00:38:51] Sarah Wang: Yeah, yeah.[00:38:51] Martin Casado: There's only two of us there, so, yes. No, so by the way, so the, the pri The pri, yeah. Yeah. So the primary developer is a [00:39:00] guy named Andres Quist, who's an absolute genius. He and I did our, our PhDs together. And so like, um, we studied for constant Quas together. It was almost like hanging out with an old friend, you know?[00:39:09] And so like. So he, he's the core, core guy. I did mostly kind of, you know, the side I run venture fund.[00:39:14] swyx: It's amazing. Like five years ago you would not have done any of this. And it brought you back[00:39:19] Martin Casado: the act, the Activ energy, you're still back. Energy was so high because you had to learn all the framework b******t.[00:39:23] Man, I f*****g used to hate that. And so like, now I don't have to deal with that. I can like focus on the algorithmics so I can focus on the scaling and I,[00:39:29] swyx: yeah. Yeah.[00:39:29] LLMs vs Spatial Intelligence + How to Value World Labs' 3D Foundation Model[00:39:29] swyx: And then, uh, I'll observe one irony and then I'll ask a serious investor question, uh, which is like, the irony is FFE actually doesn't believe that LMS can lead us to spatial intelligence.[00:39:37] And here you are using LMS to like help like achieve spatial intelligence. I just see, I see some like disconnect in there.[00:39:45] Martin Casado: Yeah. Yeah. So I think, I think, you know, I think, I think what she would say is LLMs are great to help with coding.[00:39:51] swyx: Yes.[00:39:51] Martin Casado: But like, that's very different than a model that actually like provides, they, they'll never have the[00:39:56] swyx: spatial inte[00:39:56] Martin Casado: issues.[00:39:56] And listen, our brains clearly listen, our brains, brains clearly have [00:40:00] both our, our brains clearly have a language reasoning section and they clearly have a spatial reasoning section. I mean, it's just, you know, these are two pretty independent problems.[00:40:07] swyx: Okay. And you, you, like, I, I would say that the, the one data point I recently had, uh, against it is the DeepMind, uh, IMO Gold, where, so, uh, typically the, the typical answer is that this is where you start going down the neuros symbolic path, right?[00:40:21] Like one, uh, sort of very sort of abstract reasoning thing and one form, formal thing. Um, and that's what. DeepMind had in 2024 with alpha proof, alpha geometry, and now they just use deep think and just extended thinking tokens. And it's one model and it's, and it's in LM.[00:40:36] Martin Casado: Yeah, yeah, yeah, yeah, yeah.[00:40:37] swyx: And so that, that was my indication of like, maybe you don't need a separate system.[00:40:42] Martin Casado: Yeah. So, so let me step back. I mean, at the end of the day, at the end of the day, these things are like nodes in a graph with weights on them. Right. You know, like it can be modeled like if you, if you distill it down. But let me just talk about the two different substrates. Let's, let me put you in a dark room.[00:40:56] Like totally black room. And then let me just [00:41:00] describe how you exit it. Like to your left, there's a table like duck below this thing, right? I mean like the chances that you're gonna like not run into something are very low. Now let me like turn on the light and you actually see, and you can do distance and you know how far something away is and like where it is or whatever.[00:41:17] Then you can do it, right? Like language is not the right primitives to describe. The universe because it's not exact enough. So that's all Faye, Faye is talking about. When it comes to like spatial reasoning, it's like you actually have to know that this is three feet far, like that far away. It is curved.[00:41:37] You have to understand, you know, the, like the actual movement through space.[00:41:40] swyx: Yeah.[00:41:40] Martin Casado: So I do, I listen, I do think at the end of these models are definitely converging as far as models, but there's, there's, there's different representations of problems you're solving. One is language. Which, you know, that would be like describing to somebody like what to do.[00:41:51] And the other one is actually just showing them and the space reasoning is just showing them.[00:41:55] swyx: Yeah, yeah, yeah. Right. Got it, got it. Uh, the, in the investor question was on, on, well labs [00:42:00] is, well, like, how do I value something like this? What, what, what work does the, do you do? I'm just like, Fefe is awesome.[00:42:07] Justin's awesome. And you know, the other two co-founder, co-founders, but like the, the, the tech, everyone's building cool tech. But like, what's the value of the tech? And this is the fundamental question[00:42:16] Martin Casado: of, well, let, let, just like these, let me just maybe give you a rough sketch on the diffusion models. I actually love to hear Sarah because I'm a venture for, you know, so like, ventures always, always like kind of wild west type[00:42:24] swyx: stuff.[00:42:24] You, you, you, you paid a dream and she has to like, actually[00:42:28] Martin Casado: I'm gonna say I'm gonna mar to reality, so I'm gonna say the venture for you. And she can be like, okay, you a little kid. Yeah. So like, so, so these diffusion models literally. Create something for, for almost nothing. And something that the, the world has found to be very valuable in the past, in our real markets, right?[00:42:45] Like, like a 2D image. I mean, that's been an entire market. People value them. It takes a human being a long time to create it, right? I mean, to create a, you know, a, to turn me into a whatever, like an image would cost a hundred bucks in an hour. The inference cost [00:43:00] us a hundredth of a penny, right? So we've seen this with speech in very successful companies.[00:43:03] We've seen this with 2D image. We've seen this with movies. Right? Now, think about 3D scene. I mean, I mean, when's Grand Theft Auto coming out? It's been six, what? It's been 10 years. I mean, how, how like, but hasn't been 10 years.[00:43:14] Alessio: Yeah.[00:43:15] Martin Casado: How much would it cost to like, to reproduce this room in 3D? Right. If you, if you, if you hired somebody on fiber, like in, in any sort of quality, probably 4,000 to $10,000.[00:43:24] And then if you had a professional, probably $30,000. So if you could generate the exact same thing from a 2D image, and we know that these are used and they're using Unreal and they're using Blend, or they're using movies and they're using video games and they're using all. So if you could do that for.[00:43:36] You know, less than a dollar, that's four or five orders of magnitude cheaper. So you're bringing the marginal cost of something that's useful down by three orders of magnitude, which historically have created very large companies. So that would be like the venture kind of strategic dreaming map.[00:43:49] swyx: Yeah.[00:43:50] And, and for listeners, uh, you can do this yourself on your, on your own phone with like. Uh, the marble.[00:43:55] Martin Casado: Yeah. Marble.[00:43:55] swyx: Uh, or but also there's many Nerf apps where you just go on your iPhone and, and do this.[00:43:59] Martin Casado: Yeah. Yeah. [00:44:00] Yeah. And, and in the case of marble though, it would, what you do is you literally give it in.[00:44:03] So most Nerf apps you like kind of run around and take a whole bunch of pictures and then you kind of reconstruct it.[00:44:08] swyx: Yeah.[00:44:08] Martin Casado: Um, things like marble, just that the whole generative 3D space will just take a 2D image and it'll reconstruct all the like, like[00:44:16] swyx: meaning it has to fill in. Uh,[00:44:18] Martin Casado: stuff at the back of the table, under the table, the back, like, like the images, it doesn't see.[00:44:22] So the generator stuff is very different than reconstruction that it fills in the things that you can't see.[00:44:26] swyx: Yeah. Okay.[00:44:26] Sarah Wang: So,[00:44:27] Martin Casado: all right. So now the,[00:44:28] Sarah Wang: no, no. I mean I love that[00:44:29] Martin Casado: the adult[00:44:29] Sarah Wang: perspective. Um, well, no, I was gonna say these are very much a tag team. So we, we started this pod with that, um, premise. And I think this is a perfect question to even build on that further.[00:44:36] ‘cause it truly is, I mean, we're tag teaming all of these together.[00:44:39] Investing in Model Labs, Media Rumors, and the Cursor Playbook (Margins & Going Down-Stack)[00:44:39] Sarah Wang: Um, but I think every investment fundamentally starts with the same. Maybe the same two premises. One is, at this point in time, we actually believe that there are. And of one founders for their particular craft, and they have to be demonstrated in their prior careers, right?[00:44:56] So, uh, we're not investing in every, you know, now the term is NEO [00:45:00] lab, but every foundation model, uh, any, any company, any founder trying to build a foundation model, we're not, um, contrary to popular opinion, we're

From rewriting Google's search stack in the early 2000s to reviving sparse trillion-parameter models and co-designing TPUs with frontier ML research, Jeff Dean has quietly shaped nearly every layer of the modern AI stack. As Chief AI Scientist at Google and a driving force behind Gemini, Jeff has lived through multiple scaling revolutions from CPUs and sharded indices to multimodal models that reason across text, video, and code.Jeff joins us to unpack what it really means to “own the Pareto frontier,” why distillation is the engine behind every Flash model breakthrough, how energy (in picojoules) not FLOPs is becoming the true bottleneck, what it was like leading the charge to unify all of Google's AI teams, and why the next leap won't come from bigger context windows alone, but from systems that give the illusion of attending to trillions of tokens.We discuss:* Jeff's early neural net thesis in 1990: parallel training before it was cool, why he believed scaling would win decades early, and the “bigger model, more data, better results” mantra that held for 15 years* The evolution of Google Search: sharding, moving the entire index into memory in 2001, softening query semantics pre-LLMs, and why retrieval pipelines already resemble modern LLM systems* Pareto frontier strategy: why you need both frontier “Pro” models and low-latency “Flash” models, and how distillation lets smaller models surpass prior generations* Distillation deep dive: ensembles → compression → logits as soft supervision, and why you need the biggest model to make the smallest one good* Latency as a first-class objective: why 10–50x lower latency changes UX entirely, and how future reasoning workloads will demand 10,000 tokens/sec* Energy-based thinking: picojoules per bit, why moving data costs 1000x more than a multiply, batching through the lens of energy, and speculative decoding as amortization* TPU co-design: predicting ML workloads 2–6 years out, speculative hardware features, precision reduction, sparsity, and the constant feedback loop between model architecture and silicon* Sparse models and “outrageously large” networks: trillions of parameters with 1–5% activation, and why sparsity was always the right abstraction* Unified vs. specialized models: abandoning symbolic systems, why general multimodal models tend to dominate vertical silos, and when vertical fine-tuning still makes sense* Long context and the illusion of scale: beyond needle-in-a-haystack benchmarks toward systems that narrow trillions of tokens to 117 relevant documents* Personalized AI: attending to your emails, photos, and documents (with permission), and why retrieval + reasoning will unlock deeply personal assistants* Coding agents: 50 AI interns, crisp specifications as a new core skill, and how ultra-low latency will reshape human–agent collaboration* Why ideas still matter: transformers, sparsity, RL, hardware, systems — scaling wasn't blind; the pieces had to multiply togetherShow Notes:* Gemma 3 Paper* Gemma 3* Gemini 2.5 Report* Jeff Dean's “Software Engineering Advice fromBuilding Large-Scale Distributed Systems” Presentation (with Back of the Envelope Calculations)* Latency Numbers Every Programmer Should Know by Jeff Dean* The Jeff Dean Facts* Jeff Dean Google Bio* Jeff Dean on “Important AI Trends” @Stanford AI Club* Jeff Dean & Noam Shazeer — 25 years at Google (Dwarkesh)—Jeff Dean* LinkedIn: https://www.linkedin.com/in/jeff-dean-8b212555* X: https://x.com/jeffdeanGoogle* https://google.com* https://deepmind.googleFull Video EpisodeTimestamps00:00:04 — Introduction: Alessio & Swyx welcome Jeff Dean, chief AI scientist at Google, to the Latent Space podcast00:00:30 — Owning the Pareto Frontier & balancing frontier vs low-latency models00:01:31 — Frontier models vs Flash models + role of distillation00:03:52 — History of distillation and its original motivation00:05:09 — Distillation's role in modern model scaling00:07:02 — Model hierarchy (Flash, Pro, Ultra) and distillation sources00:07:46 — Flash model economics & wide deployment00:08:10 — Latency importance for complex tasks00:09:19 — Saturation of some tasks and future frontier tasks00:11:26 — On benchmarks, public vs internal00:12:53 — Example long-context benchmarks & limitations00:15:01 — Long-context goals: attending to trillions of tokens00:16:26 — Realistic use cases beyond pure language00:18:04 — Multimodal reasoning and non-text modalities00:19:05 — Importance of vision & motion modalities00:20:11 — Video understanding example (extracting structured info)00:20:47 — Search ranking analogy for LLM retrieval00:23:08 — LLM representations vs keyword search00:24:06 — Early Google search evolution & in-memory index00:26:47 — Design principles for scalable systems00:28:55 — Real-time index updates & recrawl strategies00:30:06 — Classic “Latency numbers every programmer should know”00:32:09 — Cost of memory vs compute and energy emphasis00:34:33 — TPUs & hardware trade-offs for serving models00:35:57 — TPU design decisions & co-design with ML00:38:06 — Adapting model architecture to hardware00:39:50 — Alternatives: energy-based models, speculative decoding00:42:21 — Open research directions: complex workflows, RL00:44:56 — Non-verifiable RL domains & model evaluation00:46:13 — Transition away from symbolic systems toward unified LLMs00:47:59 — Unified models vs specialized ones00:50:38 — Knowledge vs reasoning & retrieval + reasoning00:52:24 — Vertical model specialization & modules00:55:21 — Token count considerations for vertical domains00:56:09 — Low resource languages & contextual learning00:59:22 — Origins: Dean's early neural network work01:10:07 — AI for coding & human–model interaction styles01:15:52 — Importance of crisp specification for coding agents01:19:23 — Prediction: personalized models & state retrieval01:22:36 — Token-per-second targets (10k+) and reasoning throughput01:23:20 — Episode conclusion and thanksTranscriptAlessio Fanelli [00:00:04]: Hey everyone, welcome to the Latent Space podcast. This is Alessio, founder of Kernel Labs, and I'm joined by Swyx, editor of Latent Space. Shawn Wang [00:00:11]: Hello, hello. We're here in the studio with Jeff Dean, chief AI scientist at Google. Welcome. Thanks for having me. It's a bit surreal to have you in the studio. I've watched so many of your talks, and obviously your career has been super legendary. So, I mean, congrats. I think the first thing must be said, congrats on owning the Pareto Frontier.Jeff Dean [00:00:30]: Thank you, thank you. Pareto Frontiers are good. It's good to be out there.Shawn Wang [00:00:34]: Yeah, I mean, I think it's a combination of both. You have to own the Pareto Frontier. You have to have like frontier capability, but also efficiency, and then offer that range of models that people like to use. And, you know, some part of this was started because of your hardware work. Some part of that is your model work, and I'm sure there's lots of secret sauce that you guys have worked on cumulatively. But, like, it's really impressive to see it all come together in, like, this slittily advanced.Jeff Dean [00:01:04]: Yeah, yeah. I mean, I think, as you say, it's not just one thing. It's like a whole bunch of things up and down the stack. And, you know, all of those really combine to help make UNOS able to make highly capable large models, as well as, you know, software techniques to get those large model capabilities into much smaller, lighter weight models that are, you know, much more cost effective and lower latency, but still, you know, quite capable for their size. Yeah.Alessio Fanelli [00:01:31]: How much pressure do you have on, like, having the lower bound of the Pareto Frontier, too? I think, like, the new labs are always trying to push the top performance frontier because they need to raise more money and all of that. And you guys have billions of users. And I think initially when you worked on the CPU, you were thinking about, you know, if everybody that used Google, we use the voice model for, like, three minutes a day, they were like, you need to double your CPU number. Like, what's that discussion today at Google? Like, how do you prioritize frontier versus, like, we have to do this? How do we actually need to deploy it if we build it?Jeff Dean [00:02:03]: Yeah, I mean, I think we always want to have models that are at the frontier or pushing the frontier because I think that's where you see what capabilities now exist that didn't exist at the sort of slightly less capable last year's version or last six months ago version. At the same time, you know, we know those are going to be really useful for a bunch of use cases, but they're going to be a bit slower and a bit more expensive than people might like for a bunch of other broader models. So I think what we want to do is always have kind of a highly capable sort of affordable model that enables a whole bunch of, you know, lower latency use cases. People can use them for agentic coding much more readily and then have the high-end, you know, frontier model that is really useful for, you know, deep reasoning, you know, solving really complicated math problems, those kinds of things. And it's not that. One or the other is useful. They're both useful. So I think we'd like to do both. And also, you know, through distillation, which is a key technique for making the smaller models more capable, you know, you have to have the frontier model in order to then distill it into your smaller model. So it's not like an either or choice. You sort of need that in order to actually get a highly capable, more modest size model. Yeah.Alessio Fanelli [00:03:24]: I mean, you and Jeffrey came up with the solution in 2014.Jeff Dean [00:03:28]: Don't forget, L'Oreal Vinyls as well. Yeah, yeah.Alessio Fanelli [00:03:30]: A long time ago. But like, I'm curious how you think about the cycle of these ideas, even like, you know, sparse models and, you know, how do you reevaluate them? How do you think about in the next generation of model, what is worth revisiting? Like, yeah, they're just kind of like, you know, you worked on so many ideas that end up being influential, but like in the moment, they might not feel that way necessarily. Yeah.Jeff Dean [00:03:52]: I mean, I think distillation was originally motivated because we were seeing that we had a very large image data set at the time, you know, 300 million images that we could train on. And we were seeing that if you create specialists for different subsets of those image categories, you know, this one's going to be really good at sort of mammals, and this one's going to be really good at sort of indoor room scenes or whatever, and you can cluster those categories and train on an enriched stream of data after you do pre-training on a much broader set of images. You get much better performance. If you then treat that whole set of maybe 50 models you've trained as a large ensemble, but that's not a very practical thing to serve, right? So distillation really came about from the idea of, okay, what if we want to actually serve that and train all these independent sort of expert models and then squish it into something that actually fits in a form factor that you can actually serve? And that's, you know, not that different from what we're doing today. You know, often today we're instead of having an ensemble of 50 models. We're having a much larger scale model that we then distill into a much smaller scale model.Shawn Wang [00:05:09]: Yeah. A part of me also wonders if distillation also has a story with the RL revolution. So let me maybe try to articulate what I mean by that, which is you can, RL basically spikes models in a certain part of the distribution. And then you have to sort of, well, you can spike models, but usually sometimes... It might be lossy in other areas and it's kind of like an uneven technique, but you can probably distill it back and you can, I think that the sort of general dream is to be able to advance capabilities without regressing on anything else. And I think like that, that whole capability merging without loss, I feel like it's like, you know, some part of that should be a distillation process, but I can't quite articulate it. I haven't seen much papers about it.Jeff Dean [00:06:01]: Yeah, I mean, I tend to think of one of the key advantages of distillation is that you can have a much smaller model and you can have a very large, you know, training data set and you can get utility out of making many passes over that data set because you're now getting the logits from the much larger model in order to sort of coax the right behavior out of the smaller model that you wouldn't otherwise get with just the hard labels. And so, you know, I think that's what we've observed. Is you can get, you know, very close to your largest model performance with distillation approaches. And that seems to be, you know, a nice sweet spot for a lot of people because it enables us to kind of, for multiple Gemini generations now, we've been able to make the sort of flash version of the next generation as good or even substantially better than the previous generations pro. And I think we're going to keep trying to do that because that seems like a good trend to follow.Shawn Wang [00:07:02]: So, Dara asked, so it was the original map was Flash Pro and Ultra. Are you just sitting on Ultra and distilling from that? Is that like the mother load?Jeff Dean [00:07:12]: I mean, we have a lot of different kinds of models. Some are internal ones that are not necessarily meant to be released or served. Some are, you know, our pro scale model and we can distill from that as well into our Flash scale model. So I think, you know, it's an important set of capabilities to have and also inference time scaling. It can also be a useful thing to improve the capabilities of the model.Shawn Wang [00:07:35]: And yeah, yeah, cool. Yeah. And obviously, I think the economy of Flash is what led to the total dominance. I think the latest number is like 50 trillion tokens. I don't know. I mean, obviously, it's changing every day.Jeff Dean [00:07:46]: Yeah, yeah. But, you know, by market share, hopefully up.Shawn Wang [00:07:50]: No, I mean, there's no I mean, there's just the economics wise, like because Flash is so economical, like you can use it for everything. Like it's in Gmail now. It's in YouTube. Like it's yeah. It's in everything.Jeff Dean [00:08:02]: We're using it more in our search products of various AI mode reviews.Shawn Wang [00:08:05]: Oh, my God. Flash past the AI mode. Oh, my God. Yeah, that's yeah, I didn't even think about that.Jeff Dean [00:08:10]: I mean, I think one of the things that is quite nice about the Flash model is not only is it more affordable, it's also a lower latency. And I think latency is actually a pretty important characteristic for these models because we're going to want models to do much more complicated things that are going to involve, you know, generating many more tokens from when you ask the model to do so. So, you know, if you're going to ask the model to do something until it actually finishes what you ask it to do, because you're going to ask now, not just write me a for loop, but like write me a whole software package to do X or Y or Z. And so having low latency systems that can do that seems really important. And Flash is one direction, one way of doing that. You know, obviously our hardware platforms enable a bunch of interesting aspects of our, you know, serving stack as well, like TPUs, the interconnect between. Chips on the TPUs is actually quite, quite high performance and quite amenable to, for example, long context kind of attention operations, you know, having sparse models with lots of experts. These kinds of things really, really matter a lot in terms of how do you make them servable at scale.Alessio Fanelli [00:09:19]: Yeah. Does it feel like there's some breaking point for like the proto Flash distillation, kind of like one generation delayed? I almost think about almost like the capability as a. In certain tasks, like the pro model today is a saturated, some sort of task. So next generation, that same task will be saturated at the Flash price point. And I think for most of the things that people use models for at some point, the Flash model in two generation will be able to do basically everything. And how do you make it economical to like keep pushing the pro frontier when a lot of the population will be okay with the Flash model? I'm curious how you think about that.Jeff Dean [00:09:59]: I mean, I think that's true. If your distribution of what people are asking people, the models to do is stationary, right? But I think what often happens is as the models become more capable, people ask them to do more, right? So, I mean, I think this happens in my own usage. Like I used to try our models a year ago for some sort of coding task, and it was okay at some simpler things, but wouldn't do work very well for more complicated things. And since then, we've improved dramatically on the more complicated coding tasks. And now I'll ask it to do much more complicated things. And I think that's true, not just of coding, but of, you know, now, you know, can you analyze all the, you know, renewable energy deployments in the world and give me a report on solar panel deployment or whatever. That's a very complicated, you know, more complicated task than people would have asked a year ago. And so you are going to want more capable models to push the frontier in the absence of what people ask the models to do. And that also then gives us. Insight into, okay, where does the, where do things break down? How can we improve the model in these, these particular areas, uh, in order to sort of, um, make the next generation even better.Alessio Fanelli [00:11:11]: Yeah. Are there any benchmarks or like test sets they use internally? Because it's almost like the same benchmarks get reported every time. And it's like, all right, it's like 99 instead of 97. Like, how do you have to keep pushing the team internally to it? Or like, this is what we're building towards. Yeah.Jeff Dean [00:11:26]: I mean, I think. Benchmarks, particularly external ones that are publicly available. Have their utility, but they often kind of have a lifespan of utility where they're introduced and maybe they're quite hard for current models. You know, I, I like to think of the best kinds of benchmarks are ones where the initial scores are like 10 to 20 or 30%, maybe, but not higher. And then you can sort of work on improving that capability for, uh, whatever it is, the benchmark is trying to assess and get it up to like 80, 90%, whatever. I, I think once it hits kind of 95% or something, you get very diminishing returns from really focusing on that benchmark, cuz it's sort of, it's either the case that you've now achieved that capability, or there's also the issue of leakage in public data or very related kind of data being, being in your training data. Um, so we have a bunch of held out internal benchmarks that we really look at where we know that wasn't represented in the training data at all. There are capabilities that we want the model to have. Um, yeah. Yeah. Um, that it doesn't have now, and then we can work on, you know, assessing, you know, how do we make the model better at these kinds of things? Is it, we need different kind of data to train on that's more specialized for this particular kind of task. Do we need, um, you know, a bunch of, uh, you know, architectural improvements or some sort of, uh, model capability improvements, you know, what would help make that better?Shawn Wang [00:12:53]: Is there, is there such an example that you, uh, a benchmark inspired in architectural improvement? Like, uh, I'm just kind of. Jumping on that because you just.Jeff Dean [00:13:02]: Uh, I mean, I think some of the long context capability of the, of the Gemini models that came, I guess, first in 1.5 really were about looking at, okay, we want to have, um, you know,Shawn Wang [00:13:15]: immediately everyone jumped to like completely green charts of like, everyone had, I was like, how did everyone crack this at the same time? Right. Yeah. Yeah.Jeff Dean [00:13:23]: I mean, I think, um, and once you're set, I mean, as you say that needed single needle and a half. Hey, stack benchmark is really saturated for at least context links up to 1, 2 and K or something. Don't actually have, you know, much larger than 1, 2 and 8 K these days or two or something. We're trying to push the frontier of 1 million or 2 million context, which is good because I think there are a lot of use cases where. Yeah. You know, putting a thousand pages of text or putting, you know, multiple hour long videos and the context and then actually being able to make use of that as useful. Try to, to explore the über graduation are fairly large. But the single needle in a haystack benchmark is sort of saturated. So you really want more complicated, sort of multi-needle or more realistic, take all this content and produce this kind of answer from a long context that sort of better assesses what it is people really want to do with long context. Which is not just, you know, can you tell me the product number for this particular thing?Shawn Wang [00:14:31]: Yeah, it's retrieval. It's retrieval within machine learning. It's interesting because I think the more meta level I'm trying to operate at here is you have a benchmark. You're like, okay, I see the architectural thing I need to do in order to go fix that. But should you do it? Because sometimes that's an inductive bias, basically. It's what Jason Wei, who used to work at Google, would say. Exactly the kind of thing. Yeah, you're going to win. Short term. Longer term, I don't know if that's going to scale. You might have to undo that.Jeff Dean [00:15:01]: I mean, I like to sort of not focus on exactly what solution we're going to derive, but what capability would you want? And I think we're very convinced that, you know, long context is useful, but it's way too short today. Right? Like, I think what you would really want is, can I attend to the internet while I answer my question? Right? But that's not going to happen. I think that's going to be solved by purely scaling the existing solutions, which are quadratic. So a million tokens kind of pushes what you can do. You're not going to do that to a trillion tokens, let alone, you know, a billion tokens, let alone a trillion. But I think if you could give the illusion that you can attend to trillions of tokens, that would be amazing. You'd find all kinds of uses for that. You would have attend to the internet. You could attend to the pixels of YouTube and the sort of deeper representations that we can find. You could attend to the form for a single video, but across many videos, you know, on a personal Gemini level, you could attend to all of your personal state with your permission. So like your emails, your photos, your docs, your plane tickets you have. I think that would be really, really useful. And the question is, how do you get algorithmic improvements and system level improvements that get you to something where you actually can attend to trillions of tokens? Right. In a meaningful way. Yeah.Shawn Wang [00:16:26]: But by the way, I think I did some math and it's like, if you spoke all day, every day for eight hours a day, you only generate a maximum of like a hundred K tokens, which like very comfortably fits.Jeff Dean [00:16:38]: Right. But if you then say, okay, I want to be able to understand everything people are putting on videos.Shawn Wang [00:16:46]: Well, also, I think that the classic example is you start going beyond language into like proteins and whatever else is extremely information dense. Yeah. Yeah.Jeff Dean [00:16:55]: I mean, I think one of the things about Gemini's multimodal aspects is we've always wanted it to be multimodal from the start. And so, you know, that sometimes to people means text and images and video sort of human-like and audio, audio, human-like modalities. But I think it's also really useful to have Gemini know about non-human modalities. Yeah. Like LIDAR sensor data from. Yes. Say, Waymo vehicles or. Like robots or, you know, various kinds of health modalities, x-rays and MRIs and imaging and genomics information. And I think there's probably hundreds of modalities of data where you'd like the model to be able to at least be exposed to the fact that this is an interesting modality and has certain meaning in the world. Where even if you haven't trained on all the LIDAR data or MRI data, you could have, because maybe that's not, you know, it doesn't make sense in terms of trade-offs of. You know, what you include in your main pre-training data mix, at least including a little bit of it is actually quite useful. Yeah. Because it sort of tempts the model that this is a thing.Shawn Wang [00:18:04]: Yeah. Do you believe, I mean, since we're on this topic and something I just get to ask you all the questions I always wanted to ask, which is fantastic. Like, are there some king modalities, like modalities that supersede all the other modalities? So a simple example was Vision can, on a pixel level, encode text. And DeepSeq had this DeepSeq CR paper that did that. Vision. And Vision has also been shown to maybe incorporate audio because you can do audio spectrograms and that's, that's also like a Vision capable thing. Like, so, so maybe Vision is just the king modality and like. Yeah.Jeff Dean [00:18:36]: I mean, Vision and Motion are quite important things, right? Motion. Well, like video as opposed to static images, because I mean, there's a reason evolution has evolved eyes like 23 independent ways, because it's such a useful capability for sensing the world around you, which is really what we want these models to be. So I think the only thing that we can be able to do is interpret the things we're seeing or the things we're paying attention to and then help us in using that information to do things. Yeah.Shawn Wang [00:19:05]: I think motion, you know, I still want to shout out, I think Gemini, still the only native video understanding model that's out there. So I use it for YouTube all the time. Nice.Jeff Dean [00:19:15]: Yeah. Yeah. I mean, it's actually, I think people kind of are not necessarily aware of what the Gemini models can actually do. Yeah. Like I have an example I've used in one of my talks. It had like, it was like a YouTube highlight video of 18 memorable sports moments across the last 20 years or something. So it has like Michael Jordan hitting some jump shot at the end of the finals and, you know, some soccer goals and things like that. And you can literally just give it the video and say, can you please make me a table of what all these different events are? What when the date is when they happened? And a short description. And so you get like now an 18 row table of that information extracted from the video, which is, you know, not something most people think of as like a turn video into sequel like table.Alessio Fanelli [00:20:11]: Has there been any discussion inside of Google of like, you mentioned tending to the whole internet, right? Google, it's almost built because a human cannot tend to the whole internet and you need some sort of ranking to find what you need. Yep. That ranking is like much different for an LLM because you can expect a person to look at maybe the first five, six links in a Google search versus for an LLM. Should you expect to have 20 links that are highly relevant? Like how do you internally figure out, you know, how do we build the AI mode that is like maybe like much broader search and span versus like the more human one? Yeah.Jeff Dean [00:20:47]: I mean, I think even pre-language model based work, you know, our ranking systems would be built to start. I mean, I think even pre-language model based work, you know, our ranking systems would be built to start. With a giant number of web pages in our index, many of them are not relevant. So you identify a subset of them that are relevant with very lightweight kinds of methods. You know, you're down to like 30,000 documents or something. And then you gradually refine that to apply more and more sophisticated algorithms and more and more sophisticated sort of signals of various kinds in order to get down to ultimately what you show, which is, you know, the final 10 results or, you know, 10 results plus. Other kinds of information. And I think an LLM based system is not going to be that dissimilar, right? You're going to attend to trillions of tokens, but you're going to want to identify, you know, what are the 30,000 ish documents that are with the, you know, maybe 30 million interesting tokens. And then how do you go from that into what are the 117 documents I really should be paying attention to in order to carry out the tasks that the user has asked? And I think, you know, you can imagine systems where you have, you know, a lot of highly parallel processing to identify those initial 30,000 candidates, maybe with very lightweight kinds of models. Then you have some system that sort of helps you narrow down from 30,000 to the 117 with maybe a little bit more sophisticated model or set of models. And then maybe the final model is the thing that looks. So the 117 things that might be your most capable model. So I think it has to, it's going to be some system like that, that is really enables you to give the illusion of attending to trillions of tokens. Sort of the way Google search gives you, you know, not the illusion, but you are searching the internet, but you're finding, you know, a very small subset of things that are, that are relevant.Shawn Wang [00:22:47]: Yeah. I often tell a lot of people that are not steeped in like Google search history that, well, you know, like Bert was. Like he was like basically immediately inside of Google search and that improves results a lot, right? Like I don't, I don't have any numbers off the top of my head, but like, I'm sure you guys, that's obviously the most important numbers to Google. Yeah.Jeff Dean [00:23:08]: I mean, I think going to an LLM based representation of text and words and so on enables you to get out of the explicit hard notion of, of particular words having to be on the page, but really getting at the notion of this topic of this page or this page. Paragraph is highly relevant to this query. Yeah.Shawn Wang [00:23:28]: I don't think people understand how much LLMs have taken over all these very high traffic system, very high traffic. Yeah. Like it's Google, it's YouTube. YouTube has this like semantics ID thing where it's just like every token or every item in the vocab is a YouTube video or something that predicts the video using a code book, which is absurd to me for YouTube size.Jeff Dean [00:23:50]: And then most recently GROK also for, for XAI, which is like, yeah. I mean, I'll call out even before LLMs were used extensively in search, we put a lot of emphasis on softening the notion of what the user actually entered into the query.Shawn Wang [00:24:06]: So do you have like a history of like, what's the progression? Oh yeah.Jeff Dean [00:24:09]: I mean, I actually gave a talk in, uh, I guess, uh, web search and data mining conference in 2009, uh, where we never actually published any papers about the origins of Google search, uh, sort of, but we went through sort of four or five or six. generations, four or five or six generations of, uh, redesigning of the search and retrieval system, uh, from about 1999 through 2004 or five. And that talk is really about that evolution. And one of the things that really happened in 2001 was we were sort of working to scale the system in multiple dimensions. So one is we wanted to make our index bigger, so we could retrieve from a larger index, which always helps your quality in general. Uh, because if you don't have the page in your index, you're going to not do well. Um, and then we also needed to scale our capacity because we were, our traffic was growing quite extensively. Um, and so we had, you know, a sharded system where you have more and more shards as the index grows, you have like 30 shards. And then if you want to double the index size, you make 60 shards so that you can bound the latency by which you respond for any particular user query. Um, and then as traffic grows, you add, you add more and more replicas of each of those. And so we eventually did the math that realized that in a data center where we had say 60 shards and, um, you know, 20 copies of each shard, we now had 1200 machines, uh, with disks. And we did the math and we're like, Hey, one copy of that index would actually fit in memory across 1200 machines. So in 2001, we introduced, uh, we put our entire index in memory and what that enabled from a quality perspective was amazing. Um, and so we had more and more replicas of each of those. Before you had to be really careful about, you know, how many different terms you looked at for a query, because every one of them would involve a disk seek on every one of the 60 shards. And so you, as you make your index bigger, that becomes even more inefficient. But once you have the whole index in memory, it's totally fine to have 50 terms you throw into the query from the user's original three or four word query, because now you can add synonyms like restaurant and restaurants and cafe and, uh, you know, things like that. Uh, bistro and all these things. And you can suddenly start, uh, sort of really, uh, getting at the meaning of the word as opposed to the exact semantic form the user typed in. And that was, you know, 2001, very much pre LLM, but really it was about softening the, the strict definition of what the user typed in order to get at the meaning.Alessio Fanelli [00:26:47]: What are like principles that you use to like design the systems, especially when you have, I mean, in 2001, the internet is like. Doubling, tripling every year in size is not like, uh, you know, and I think today you kind of see that with LLMs too, where like every year the jumps in size and like capabilities are just so big. Are there just any, you know, principles that you use to like, think about this? Yeah.Jeff Dean [00:27:08]: I mean, I think, uh, you know, first, whenever you're designing a system, you want to understand what are the sort of design parameters that are going to be most important in designing that, you know? So, you know, how many queries per second do you need to handle? How big is the internet? How big is the index you need to handle? How much data do you need to keep for every document in the index? How are you going to look at it when you retrieve things? Um, what happens if traffic were to double or triple, you know, will that system work well? And I think a good design principle is you're going to want to design a system so that the most important characteristics could scale by like factors of five or 10, but probably not beyond that because often what happens is if you design a system for X. And something suddenly becomes a hundred X, that would enable a very different point in the design space that would not make sense at X. But all of a sudden at a hundred X makes total sense. So like going from a disk space index to a in memory index makes a lot of sense once you have enough traffic, because now you have enough replicas of the sort of state on disk that those machines now actually can hold, uh, you know, a full copy of the, uh, index and memory. Yeah. And that all of a sudden enabled. A completely different design that wouldn't have been practical before. Yeah. Um, so I'm, I'm a big fan of thinking through designs in your head, just kind of playing with the design space a little before you actually do a lot of writing of code. But, you know, as you said, in the early days of Google, we were growing the index, uh, quite extensively. We were growing the update rate of the index. So the update rate actually is the parameter that changed the most. Surprising. So it used to be once a month.Shawn Wang [00:28:55]: Yeah.Jeff Dean [00:28:56]: And then we went to a system that could update any particular page in like sub one minute. Okay.Shawn Wang [00:29:02]: Yeah. Because this is a competitive advantage, right?Jeff Dean [00:29:04]: Because all of a sudden news related queries, you know, if you're, if you've got last month's news index, it's not actually that useful for.Shawn Wang [00:29:11]: News is a special beast. Was there any, like you could have split it onto a separate system.Jeff Dean [00:29:15]: Well, we did. We launched a Google news product, but you also want news related queries that people type into the main index to also be sort of updated.Shawn Wang [00:29:23]: So, yeah, it's interesting. And then you have to like classify whether the page is, you have to decide which pages should be updated and what frequency. Oh yeah.Jeff Dean [00:29:30]: There's a whole like, uh, system behind the scenes that's trying to decide update rates and importance of the pages. So even if the update rate seems low, you might still want to recrawl important pages quite often because, uh, the likelihood they change might be low, but the value of having updated is high.Shawn Wang [00:29:50]: Yeah, yeah, yeah, yeah. Uh, well, you know, yeah. This, uh, you know, mention of latency and, and saving things to this reminds me of one of your classics, which I have to bring up, which is latency numbers. Every programmer should know, uh, was there a, was it just a, just a general story behind that? Did you like just write it down?Jeff Dean [00:30:06]: I mean, this has like sort of eight or 10 different kinds of metrics that are like, how long does a cache mistake? How long does branch mispredict take? How long does a reference domain memory take? How long does it take to send, you know, a packet from the U S to the Netherlands or something? Um,Shawn Wang [00:30:21]: why Netherlands, by the way, or is it, is that because of Chrome?Jeff Dean [00:30:25]: Uh, we had a data center in the Netherlands, um, so, I mean, I think this gets to the point of being able to do the back of the envelope calculations. So these are sort of the raw ingredients of those, and you can use them to say, okay, well, if I need to design a system to do image search and thumb nailing or something of the result page, you know, how, what I do that I could pre-compute the image thumbnails. I could like. Try to thumbnail them on the fly from the larger images. What would that do? How much dis bandwidth than I need? How many des seeks would I do? Um, and you can sort of actually do thought experiments in, you know, 30 seconds or a minute with the sort of, uh, basic, uh, basic numbers at your fingertips. Uh, and then as you sort of build software using higher level libraries, you kind of want to develop the same intuitions for how long does it take to, you know, look up something in this particular kind of.Shawn Wang [00:31:21]: I'll see you next time.Shawn Wang [00:31:51]: Which is a simple byte conversion. That's nothing interesting. I wonder if you have any, if you were to update your...Jeff Dean [00:31:58]: I mean, I think it's really good to think about calculations you're doing in a model, either for training or inference.Jeff Dean [00:32:09]: Often a good way to view that is how much state will you need to bring in from memory, either like on-chip SRAM or HBM from the accelerator. Attached memory or DRAM or over the network. And then how expensive is that data motion relative to the cost of, say, an actual multiply in the matrix multiply unit? And that cost is actually really, really low, right? Because it's order, depending on your precision, I think it's like sub one picodule.Shawn Wang [00:32:50]: Oh, okay. You measure it by energy. Yeah. Yeah.Jeff Dean [00:32:52]: Yeah. I mean, it's all going to be about energy and how do you make the most energy efficient system. And then moving data from the SRAM on the other side of the chip, not even off the off chip, but on the other side of the same chip can be, you know, a thousand picodules. Oh, yeah. And so all of a sudden, this is why your accelerators require batching. Because if you move, like, say, the parameter of a model from SRAM on the, on the chip into the multiplier unit, that's going to cost you a thousand picodules. So you better make use of that, that thing that you moved many, many times with. So that's where the batch dimension comes in. Because all of a sudden, you know, if you have a batch of 256 or something, that's not so bad. But if you have a batch of one, that's really not good.Shawn Wang [00:33:40]: Yeah. Yeah. Right.Jeff Dean [00:33:41]: Because then you paid a thousand picodules in order to do your one picodule multiply.Shawn Wang [00:33:46]: I have never heard an energy-based analysis of batching.Jeff Dean [00:33:50]: Yeah. I mean, that's why people batch. Yeah. Ideally, you'd like to use batch size one because the latency would be great.Shawn Wang [00:33:56]: The best latency.Jeff Dean [00:33:56]: But the energy cost and the compute cost inefficiency that you get is quite large. So, yeah.Shawn Wang [00:34:04]: Is there a similar trick like, like, like you did with, you know, putting everything in memory? Like, you know, I think obviously NVIDIA has caused a lot of waves with betting very hard on SRAM with Grok. I wonder if, like, that's something that you already saw with, with the TPUs, right? Like that, that you had to. Uh, to serve at your scale, uh, you probably sort of saw that coming. Like what, what, what hardware, uh, innovations or insights were formed because of what you're seeing there?Jeff Dean [00:34:33]: Yeah. I mean, I think, you know, TPUs have this nice, uh, sort of regular structure of 2D or 3D meshes with a bunch of chips connected. Yeah. And each one of those has HBM attached. Um, I think for serving some kinds of models, uh, you know, you, you pay a lot higher cost. Uh, and time latency, um, bringing things in from HBM than you do bringing them in from, uh, SRAM on the chip. So if you have a small enough model, you can actually do model parallelism, spread it out over lots of chips and you actually get quite good throughput improvements and latency improvements from doing that. And so you're now sort of striping your smallish scale model over say 16 or 64 chips. Uh, but as if you do that and it all fits in. In SRAM, uh, that can be a big win. So yeah, that's not a surprise, but it is a good technique.Alessio Fanelli [00:35:27]: Yeah. What about the TPU design? Like how much do you decide where the improvements have to go? So like, this is like a good example of like, is there a way to bring the thousand picojoules down to 50? Like, is it worth designing a new chip to do that? The extreme is like when people say, oh, you should burn the model on the ASIC and that's kind of like the most extreme thing. How much of it? Is it worth doing an hardware when things change so quickly? Like what was the internal discussion? Yeah.Jeff Dean [00:35:57]: I mean, we, we have a lot of interaction between say the TPU chip design architecture team and the sort of higher level modeling, uh, experts, because you really want to take advantage of being able to co-design what should future TPUs look like based on where we think the sort of ML research puck is going, uh, in some sense, because, uh, you know, as a hardware designer for ML and in particular, you're trying to design a chip starting today and that design might take two years before it even lands in a data center. And then it has to sort of be a reasonable lifetime of the chip to take you three, four or five years. So you're trying to predict two to six years out where, what ML computations will people want to run two to six years out in a very fast changing field. And so having people with interest. Interesting ML research ideas of things we think will start to work in that timeframe or will be more important in that timeframe, uh, really enables us to then get, you know, interesting hardware features put into, you know, TPU N plus two, where TPU N is what we have today.Shawn Wang [00:37:10]: Oh, the cycle time is plus two.Jeff Dean [00:37:12]: Roughly. Wow. Because, uh, I mean, sometimes you can squeeze some changes into N plus one, but, you know, bigger changes are going to require the chip. Yeah. Design be earlier in its lifetime design process. Um, so whenever we can do that, it's generally good. And sometimes you can put in speculative features that maybe won't cost you much chip area, but if it works out, it would make something, you know, 10 times as fast. And if it doesn't work out, well, you burned a little bit of tiny amount of your chip area on that thing, but it's not that big a deal. Uh, sometimes it's a very big change and we want to be pretty sure this is going to work out. So we'll do like lots of carefulness. Uh, ML experimentation to show us, uh, this is actually the, the way we want to go. Yeah.Alessio Fanelli [00:37:58]: Is there a reverse of like, we already committed to this chip design so we can not take the model architecture that way because it doesn't quite fit?Jeff Dean [00:38:06]: Yeah. I mean, you, you definitely have things where you're going to adapt what the model architecture looks like so that they're efficient on the chips that you're going to have for both training and inference of that, of that, uh, generation of model. So I think it kind of goes both ways. Um, you know, sometimes you can take advantage of, you know, lower precision things that are coming in a future generation. So you can, might train it at that lower precision, even if the current generation doesn't quite do that. Mm.Shawn Wang [00:38:40]: Yeah. How low can we go in precision?Jeff Dean [00:38:43]: Because people are saying like ternary is like, uh, yeah, I mean, I'm a big fan of very low precision because I think that gets, that saves you a tremendous amount of time. Right. Because it's picojoules per bit that you're transferring and reducing the number of bits is a really good way to, to reduce that. Um, you know, I think people have gotten a lot of luck, uh, mileage out of having very low bit precision things, but then having scaling factors that apply to a whole bunch of, uh, those, those weights. Scaling. How does it, how does it, okay.Shawn Wang [00:39:15]: Interesting. You, so low, low precision, but scaled up weights. Yeah. Huh. Yeah. Never considered that. Yeah. Interesting. Uh, w w while we're on this topic, you know, I think there's a lot of, um, uh, this, the concept of precision at all is weird when we're sampling, you know, uh, we just, at the end of this, we're going to have all these like chips that I'll do like very good math. And then we're just going to throw a random number generator at the start. So, I mean, there's a movement towards, uh, energy based, uh, models and processors. I'm just curious if you've, obviously you've thought about it, but like, what's your commentary?Jeff Dean [00:39:50]: Yeah. I mean, I think. There's a bunch of interesting trends though. Energy based models is one, you know, diffusion based models, which don't sort of sequentially decode tokens is another, um, you know, speculative decoding is a way that you can get sort of an equivalent, very small.Shawn Wang [00:40:06]: Draft.Jeff Dean [00:40:07]: Batch factor, uh, for like you predict eight tokens out and that enables you to sort of increase the effective batch size of what you're doing by a factor of eight, even, and then you maybe accept five or six of those tokens. So you get. A five, a five X improvement in the amortization of moving weights, uh, into the multipliers to do the prediction for the, the tokens. So these are all really good techniques and I think it's really good to look at them from the lens of, uh, energy, real energy, not energy based models, um, and, and also latency and throughput, right? If you look at things from that lens, that sort of guides you to. Two solutions that are gonna be, uh, you know, better from, uh, you know, being able to serve larger models or, you know, equivalent size models more cheaply and with lower latency.Shawn Wang [00:41:03]: Yeah. Well, I think, I think I, um, it's appealing intellectually, uh, haven't seen it like really hit the mainstream, but, um, I do think that, uh, there's some poetry in the sense that, uh, you know, we don't have to do, uh, a lot of shenanigans if like we fundamentally. Design it into the hardware. Yeah, yeah.Jeff Dean [00:41:23]: I mean, I think there's still a, there's also sort of the more exotic things like analog based, uh, uh, computing substrates as opposed to digital ones. Uh, I'm, you know, I think those are super interesting cause they can be potentially low power. Uh, but I think you often end up wanting to interface that with digital systems and you end up losing a lot of the power advantages in the digital to analog and analog to digital conversions. You end up doing, uh, at the sort of boundaries. And periphery of that system. Um, I still think there's a tremendous distance we can go from where we are today in terms of energy efficiency with sort of, uh, much better and specialized hardware for the models we care about.Shawn Wang [00:42:05]: Yeah.Alessio Fanelli [00:42:06]: Um, any other interesting research ideas that you've seen, or like maybe things that you cannot pursue a Google that you would be interested in seeing researchers take a step at, I guess you have a lot of researchers. Yeah, I guess you have enough, but our, our research.Jeff Dean [00:42:21]: Our research portfolio is pretty broad. I would say, um, I mean, I think, uh, in terms of research directions, there's a whole bunch of, uh, you know, open problems and how do you make these models reliable and able to do much longer, kind of, uh, more complex tasks that have lots of subtasks. How do you orchestrate, you know, maybe one model that's using other models as tools in order to sort of build, uh, things that can accomplish, uh, you know, much more. Yeah. Significant pieces of work, uh, collectively, then you would ask a single model to do. Um, so that's super interesting. How do you get more verifiable, uh, you know, how do you get RL to work for non-verifiable domains? I think it's a pretty interesting open problem because I think that would broaden out the capabilities of the models, the improvements that you're seeing in both math and coding. Uh, if we could apply those to other less verifiable domains, because we've come up with RL techniques that actually enable us to do that. Uh, effectively, that would, that would really make the models improve quite a lot. I think.Alessio Fanelli [00:43:26]: I'm curious, like when we had Noam Brown on the podcast, he said, um, they already proved you can do it with deep research. Um, you kind of have it with AI mode in a way it's not verifiable. I'm curious if there's any thread that you think is interesting there. Like what is it? Both are like information retrieval of JSON. So I wonder if it's like the retrieval is like the verifiable part. That you can score or what are like, yeah, yeah. How, how would you model that, that problem?Jeff Dean [00:43:55]: Yeah. I mean, I think there are ways of having other models that can evaluate the results of what a first model did, maybe even retrieving. Can you have another model that says, is this things, are these things you retrieved relevant? Or can you rate these 2000 things you retrieved to assess which ones are the 50 most relevant or something? Um, I think those kinds of techniques are actually quite effective. Sometimes I can even be the same model, just prompted differently to be a, you know, a critic as opposed to a, uh, actual retrieval system. Yeah.Shawn Wang [00:44:28]: Um, I do think like there, there is that, that weird cliff where like, it feels like we've done the easy stuff and then now it's, but it always feels like that every year. It's like, oh, like we know, we know, and the next part is super hard and nobody's figured it out. And, uh, exactly with this RLVR thing where like everyone's talking about, well, okay, how do we. the next stage of the non-verifiable stuff. And everyone's like, I don't know, you know, Ellen judge.Jeff Dean [00:44:56]: I mean, I feel like the nice thing about this field is there's lots and lots of smart people thinking about creative solutions to some of the problems that we all see. Uh, because I think everyone sort of sees that the models, you know, are great at some things and they fall down around the edges of those things and, and are not as capable as we'd like in those areas. And then coming up with good techniques and trying those. And seeing which ones actually make a difference is sort of what the whole research aspect of this field is, is pushing forward. And I think that's why it's super interesting. You know, if you think about two years ago, we were struggling with GSM, eight K problems, right? Like, you know, Fred has two rabbits. He gets three more rabbits. How many rabbits does he have? That's a pretty far cry from the kinds of mathematics that the models can, and now you're doing IMO and Erdos problems in pure language. Yeah. Yeah. Pure language. So that is a really, really amazing jump in capabilities in, you know, in a year and a half or something. And I think, um, for other areas, it'd be great if we could make that kind of leap. Uh, and you know, we don't exactly see how to do it for some, some areas, but we do see it for some other areas and we're going to work hard on making that better. Yeah.Shawn Wang [00:46:13]: Yeah.Alessio Fanelli [00:46:14]: Like YouTube thumbnail generation. That would be very helpful. We need that. That would be AGI. We need that.Shawn Wang [00:46:20]: That would be. As far as content creators go.Jeff Dean [00:46:22]: I guess I'm not a YouTube creator, so I don't care that much about that problem, but I guess, uh, many people do.Shawn Wang [00:46:27]: It does. Yeah. It doesn't, it doesn't matter. People do judge books by their covers as it turns out. Um, uh, just to draw a bit on the IMO goal. Um, I'm still not over the fact that a year ago we had alpha proof and alpha geometry and all those things. And then this year we were like, screw that we'll just chuck it into Gemini. Yeah. What's your reflection? Like, I think this, this question about. Like the merger of like symbolic systems and like, and, and LMS, uh, was a very much core belief. And then somewhere along the line, people would just said, Nope, we'll just all do it in the LLM.Jeff Dean [00:47:02]: Yeah. I mean, I think it makes a lot of sense to me because, you know, humans manipulate symbols, but we probably don't have like a symbolic representation in our heads. Right. We have some distributed representation that is neural net, like in some way of lots of different neurons. And activation patterns firing when we see certain things and that enables us to reason and plan and, you know, do chains of thought and, you know, roll them back now that, that approach for solving the problem doesn't seem like it's going to work. I'm going to try this one. And, you know, in a lot of ways we're emulating what we intuitively think, uh, is happening inside real brains in neural net based models. So it never made sense to me to have like completely separate. Uh, discrete, uh, symbolic things, and then a completely different way of, of, uh, you know, thinking about those things.Shawn Wang [00:47:59]: Interesting. Yeah. Uh, I mean, it's maybe seems obvious to you, but it wasn't obvious to me a year ago. Yeah.Jeff Dean [00:48:06]: I mean, I do think like that IMO with, you know, translating to lean and using lean and then the next year and also a specialized geometry model. And then this year switching to a single unified model. That is roughly the production model with a little bit more inference budget, uh, is actually, you know, quite good because it shows you that the capabilities of that general model have improved dramatically and, and now you don't need the specialized model. This is actually sort of very similar to the 2013 to 16 era of machine learning, right? Like it used to be, people would train separate models for lots of different, each different problem, right? I have, I want to recognize street signs and something. So I train a street sign. Recognition recognition model, or I want to, you know, decode speech recognition. I have a speech model, right? I think now the era of unified models that do everything is really upon us. And the question is how well do those models generalize to new things they've never been asked to do and they're getting better and better.Shawn Wang [00:49:10]: And you don't need domain experts. Like one of my, uh, so I interviewed ETA who was on, who was on that team. Uh, and he was like, yeah, I, I don't know how they work. I don't know where the IMO competition was held. I don't know the rules of it. I just trained the models, the training models. Yeah. Yeah. And it's kind of interesting that like people with these, this like universal skill set of just like machine learning, you just give them data and give them enough compute and they can kind of tackle any task, which is the bitter lesson, I guess. I don't know. Yeah.Jeff Dean [00:49:39]: I mean, I think, uh, general models, uh, will win out over specialized ones in most cases.Shawn Wang [00:49:45]: Uh, so I want to push there a bit. I think there's one hole here, which is like, uh. There's this concept of like, uh, maybe capacity of a model, like abstractly a model can only contain the number of bits that it has. And, uh, and so it, you know, God knows like Gemini pro is like one to 10 trillion parameters. We don't know, but, uh, the Gemma models, for example, right? Like a lot of people want like the open source local models that are like that, that, that, and, and, uh, they have some knowledge, which is not necessary, right? Like they can't know everything like, like you have the. The luxury of you have the big model and big model should be able to capable of everything. But like when, when you're distilling and you're going down to the small models, you know, you're actually memorizing things that are not useful. Yeah. And so like, how do we, I guess, do we want to extract that? Can we, can we divorce knowledge from reasoning, you know?Jeff Dean [00:50:38]: Yeah. I mean, I think you do want the model to be most effective at reasoning if it can retrieve things, right? Because having the model devote precious parameter space. To remembering obscure facts that could be looked up is actually not the best use of that parameter space, right? Like you might prefer something that is more generally useful in more settings than this obscure fact that it has. Um, so I think that's always attention at the same time. You also don't want your model to be kind of completely detached from, you know, knowing stuff about the world, right? Like it's probably useful to know how long the golden gate be. Bridges just as a general sense of like how long are bridges, right? And, uh, it should have that kind of knowledge. It maybe doesn't need to know how long some teeny little bridge in some other more obscure part of the world is, but, uh, it does help it to have a fair bit of world knowledge and the bigger your model is, the more you can have. Uh, but I do think combining retrieval with sort of reasoning and making the model really good at doing multiple stages of retrieval. Yeah.Shawn Wang [00:51:49]: And reasoning through the intermediate retrieval results is going to be a, a pretty effective way of making the model seem much more capable, because if you think about, say, a personal Gemini, yeah, right?Jeff Dean [00:52:01]: Like we're not going to train Gemini on my email. Probably we'd rather have a single model that, uh, we can then use and use being able to retrieve from my email as a tool and have the model reason about it and retrieve from my photos or whatever, uh, and then make use of that and have multiple. Um, you know, uh, stages of interaction. that makes sense.Alessio Fanelli [00:52:24]: Do you think the vertical models are like, uh, interesting pursuit? Like when people are like, oh, we're building the best healthcare LLM, we're building the best law LLM, are those kind of like short-term stopgaps or?Jeff Dean [00:52:37]: No, I mean, I think, I think vertical models are interesting. Like you want them to start from a pretty good base model, but then you can sort of, uh, sort of viewing them, view them as enriching the data. Data distribution for that particular vertical domain for healthcare, say, um, we're probably not going to train or for say robotics. We're probably not going to train Gemini on all possible robotics data. We, you could train it on because we want it to have a balanced set of capabilities. Um, so we'll expose it to some robotics data, but if you're trying to build a really, really good robotics model, you're going to want to start with that and then train it on more robotics data. And then maybe that would. It's multilingual translation capability, but improve its robotics capabilities. And we're always making these kind of, uh, you know, trade-offs in the data mix that we train the base Gemini models on. You know, we'd love to include data from 200 more languages and as much data as we have for those languages, but that's going to displace some other capabilities of the model. It won't be as good at, um, you know, Pearl programming, you know, it'll still be good at Python programming. Cause we'll include it. Enough. Of that, but there's other long tail computer languages or coding capabilities that it may suffer on or multi, uh, multimodal reasoning capabilities may suffer. Cause we didn't get to expose it to as much data there, but it's really good at multilingual things. So I, I think some combination of specialized models, maybe more modular models. So it'd be nice to have the capability to have those 200 languages, plus this awesome robotics model, plus this awesome healthcare, uh, module that all can be knitted together to work in concert and called upon in different circumstances. Right? Like if I have a health related thing, then it should enable using this health module in conjunction with the main base model to be even better at those kinds of things. Yeah.Shawn Wang [00:54:36]: Installable knowledge. Yeah.Jeff Dean [00:54:37]: Right.Shawn Wang [00:54:38]: Just download as a, as a package.Jeff Dean [00:54:39]: And some of that installable stuff can come from retrieval, but some of it probably should come from preloaded training on, you know, uh, a hundred billion tokens or a trillion tokens of health data. Yeah.Shawn Wang [00:54:51]: And for listeners, I think, uh, I will highlight the Gemma three end paper where they, there was a little bit of that, I think. Yeah.Alessio Fanelli [00:54:56]: Yeah. I guess the question is like, how many billions of tokens do you need to outpace the frontier model improvements? You know, it's like, if I have to make this model better healthcare and the main. Gemini model is still improving. Do I need 50 billion tokens? Can I do it with a hundred, if I need a trillion healthcare tokens, it's like, they're probably not out there that you don't have, you know, I think that's really like the.Jeff Dean [00:55:21]: Well, I mean, I think healthcare is a particularly challenging domain, so there's a lot of healthcare data that, you know, we don't have access to appropriately, but there's a lot of, you know, uh, healthcare organizations that want to train models on their own data. That is not public healthcare data, uh, not public health. But public healthcare data. Um, so I think there are opportunities there to say, partner with a large healthcare organization and train models for their use that are going to be, you know, more bespoke, but probably, uh, might be better than a general model trained on say, public data. Yeah.Shawn Wang [00:55:58]: Yeah. I, I believe, uh, by the way, also this is like somewhat related to the language conversation. Uh, I think one of your, your favorite examples was you can put a low resource language in the context and it just learns. Yeah.Jeff Dean [00:56:09]: Oh, yeah, I think the example we used was Calamon, which is truly low resource because it's only spoken by, I think 120 people in the world and there's no written text.Shawn Wang [00:56:20]: So, yeah. So you can just do it that way. Just put it in the context. Yeah. Yeah. But I think your whole data set in the context, right.Jeff Dean [00:56:27]: If you, if you take a language like, uh, you know, Somali or something, there is a fair bit of Somali text in the world that, uh, or Ethiopian Amharic or something, um, you know, we probably. Yeah. Are not putting all the data from those languages into the Gemini based training. We put some of it, but if you put more of it, you'll improve the capabilities of those models.Shawn Wang [00:56:49]: Yeah.Jeff Dean [00:56:49]:

This podcast features Gabriele Corso and Jeremy Wohlwend, co-founders of Boltz and authors of the Boltz Manifesto, discussing the rapid evolution of structural biology models from AlphaFold to their own open-source suite, Boltz-1 and Boltz-2. The central thesis is that while single-chain protein structure prediction is largely “solved” through evolutionary hints, the next frontier lies in modeling complex interactions (protein-ligand, protein-protein) and generative protein design, which Boltz aims to democratize via open-source foundations and scalable infrastructure.Full Video PodOn YouTube!Timestamps* 00:00 Introduction to Benchmarking and the “Solved” Protein Problem* 06:48 Evolutionary Hints and Co-evolution in Structure Prediction* 10:00 The Importance of Protein Function and Disease States* 15:31 Transitioning from AlphaFold 2 to AlphaFold 3 Capabilities* 19:48 Generative Modeling vs. Regression in Structural Biology* 25:00 The “Bitter Lesson” and Specialized AI Architectures* 29:14 Development Anecdotes: Training Boltz-1 on a Budget* 32:00 Validation Strategies and the Protein Data Bank (PDB)* 37:26 The Mission of Boltz: Democratizing Access and Open Source* 41:43 Building a Self-Sustaining Research Community* 44:40 Boltz-2 Advancements: Affinity Prediction and Design* 51:03 BoltzGen: Merging Structure and Sequence Prediction* 55:18 Large-Scale Wet Lab Validation Results* 01:02:44 Boltz Lab Product Launch: Agents and Infrastructure* 01:13:06 Future Directions: Developpability and the “Virtual Cell”* 01:17:35 Interacting with Skeptical Medicinal ChemistsKey SummaryEvolution of Structure Prediction & Evolutionary Hints* Co-evolutionary Landscapes: The speakers explain that breakthrough progress in single-chain protein prediction relied on decoding evolutionary correlations where mutations in one position necessitate mutations in another to conserve 3D structure.* Structure vs. Folding: They differentiate between structure prediction (getting the final answer) and folding (the kinetic process of reaching that state), noting that the field is still quite poor at modeling the latter.* Physics vs. Statistics: RJ posits that while models use evolutionary statistics to find the right “valley” in the energy landscape, they likely possess a “light understanding” of physics to refine the local minimum.The Shift to Generative Architectures* Generative Modeling: A key leap in AlphaFold 3 and Boltz-1 was moving from regression (predicting one static coordinate) to a generative diffusion approach that samples from a posterior distribution.* Handling Uncertainty: This shift allows models to represent multiple conformational states and avoid the “averaging” effect seen in regression models when the ground truth is ambiguous.* Specialized Architectures: Despite the “bitter lesson” of general-purpose transformers, the speakers argue that equivariant architectures remain vastly superior for biological data due to the inherent 3D geometric constraints of molecules.Boltz-2 and Generative Protein Design* Unified Encoding: Boltz-2 (and BoltzGen) treats structure and sequence prediction as a single task by encoding amino acid identities into the atomic composition of the predicted structure.* Design Specifics: Instead of a sequence, users feed the model blank tokens and a high-level “spec” (e.g., an antibody framework), and the model decodes both the 3D structure and the corresponding amino acids.* Affinity Prediction: While model confidence is a common metric, Boltz-2 focuses on affinity prediction—quantifying exactly how tightly a designed binder will stick to its target.Real-World Validation and Productization* Generalized Validation: To prove the model isn't just “regurgitating” known data, Boltz tested its designs on 9 targets with zero known interactions in the PDB, achieving nanomolar binders for two-thirds of them.* Boltz Lab Infrastructure: The newly launched Boltz Lab platform provides “agents” for protein and small molecule design, optimized to run 10x faster than open-source versions through proprietary GPU kernels.* Human-in-the-Loop: The platform is designed to convert skeptical medicinal chemists by allowing them to run parallel screens and use their intuition to filter model outputs.TranscriptRJ [00:05:35]: But the goal remains to, like, you know, really challenge the models, like, how well do these models generalize? And, you know, we've seen in some of the latest CASP competitions, like, while we've become really, really good at proteins, especially monomeric proteins, you know, other modalities still remain pretty difficult. So it's really essential, you know, in the field that there are, like, these efforts to gather, you know, benchmarks that are challenging. So it keeps us in line, you know, about what the models can do or not.Gabriel [00:06:26]: Yeah, it's interesting you say that, like, in some sense, CASP, you know, at CASP 14, a problem was solved and, like, pretty comprehensively, right? But at the same time, it was really only the beginning. So you can say, like, what was the specific problem you would argue was solved? And then, like, you know, what is remaining, which is probably quite open.RJ [00:06:48]: I think we'll steer away from the term solved, because we have many friends in the community who get pretty upset at that word. And I think, you know, fairly so. But the problem that was, you know, that a lot of progress was made on was the ability to predict the structure of single chain proteins. So proteins can, like, be composed of many chains. And single chain proteins are, you know, just a single sequence of amino acids. And one of the reasons that we've been able to make such progress is also because we take a lot of hints from evolution. So the way the models work is that, you know, they sort of decode a lot of hints. That comes from evolutionary landscapes. So if you have, like, you know, some protein in an animal, and you go find the similar protein across, like, you know, different organisms, you might find different mutations in them. And as it turns out, if you take a lot of the sequences together, and you analyze them, you see that some positions in the sequence tend to evolve at the same time as other positions in the sequence, sort of this, like, correlation between different positions. And it turns out that that is typically a hint that these two positions are close in three dimension. So part of the, you know, part of the breakthrough has been, like, our ability to also decode that very, very effectively. But what it implies also is that in absence of that co-evolutionary landscape, the models don't quite perform as well. And so, you know, I think when that information is available, maybe one could say, you know, the problem is, like, somewhat solved. From the perspective of structure prediction, when it isn't, it's much more challenging. And I think it's also worth also differentiating the, sometimes we confound a little bit, structure prediction and folding. Folding is the more complex process of actually understanding, like, how it goes from, like, this disordered state into, like, a structured, like, state. And that I don't think we've made that much progress on. But the idea of, like, yeah, going straight to the answer, we've become pretty good at.Brandon [00:08:49]: So there's this protein that is, like, just a long chain and it folds up. Yeah. And so we're good at getting from that long chain in whatever form it was originally to the thing. But we don't know how it necessarily gets to that state. And there might be intermediate states that it's in sometimes that we're not aware of.RJ [00:09:10]: That's right. And that relates also to, like, you know, our general ability to model, like, the different, you know, proteins are not static. They move, they take different shapes based on their energy states. And I think we are, also not that good at understanding the different states that the protein can be in and at what frequency, what probability. So I think the two problems are quite related in some ways. Still a lot to solve. But I think it was very surprising at the time, you know, that even with these evolutionary hints that we were able to, you know, to make such dramatic progress.Brandon [00:09:45]: So I want to ask, why does the intermediate states matter? But first, I kind of want to understand, why do we care? What proteins are shaped like?Gabriel [00:09:54]: Yeah, I mean, the proteins are kind of the machines of our body. You know, the way that all the processes that we have in our cells, you know, work is typically through proteins, sometimes other molecules, sort of intermediate interactions. And through that interactions, we have all sorts of cell functions. And so when we try to understand, you know, a lot of biology, how our body works, how disease work. So we often try to boil it down to, okay, what is going right in case of, you know, our normal biological function and what is going wrong in case of the disease state. And we boil it down to kind of, you know, proteins and kind of other molecules and their interaction. And so when we try predicting the structure of proteins, it's critical to, you know, have an understanding of kind of those interactions. It's a bit like seeing the difference between... Having kind of a list of parts that you would put it in a car and seeing kind of the car in its final form, you know, seeing the car really helps you understand what it does. On the other hand, kind of going to your question of, you know, why do we care about, you know, how the protein falls or, you know, how the car is made to some extent is that, you know, sometimes when something goes wrong, you know, there are, you know, cases of, you know, proteins misfolding. In some diseases and so on, if we don't understand this folding process, we don't really know how to intervene.RJ [00:11:30]: There's this nice line in the, I think it's in the Alpha Fold 2 manuscript, where they sort of discuss also like why we even hopeful that we can target the problem in the first place. And then there's this notion that like, well, four proteins that fold. The folding process is almost instantaneous, which is a strong, like, you know, signal that like, yeah, like we should, we might be... able to predict that this very like constrained thing that, that the protein does so quickly. And of course that's not the case for, you know, for, for all proteins. And there's a lot of like really interesting mechanisms in the cells, but yeah, I remember reading that and thought, yeah, that's somewhat of an insightful point.Gabriel [00:12:10]: I think one of the interesting things about the protein folding problem is that it used to be actually studied. And part of the reason why people thought it was impossible, it used to be studied as kind of like a classical example. Of like an MP problem. Uh, like there are so many different, you know, type of, you know, shapes that, you know, this amino acid could take. And so, this grows combinatorially with the size of the sequence. And so there used to be kind of a lot of actually kind of more theoretical computer science thinking about and studying protein folding as an MP problem. And so it was very surprising also from that perspective, kind of seeing. Machine learning so clear, there is some, you know, signal in those sequences, through evolution, but also through kind of other things that, you know, us as humans, we're probably not really able to, uh, to understand, but that is, models I've, I've learned.Brandon [00:13:07]: And so Andrew White, we were talking to him a few weeks ago and he said that he was following the development of this and that there were actually ASICs that were developed just to solve this problem. So, again, that there were. There were many, many, many millions of computational hours spent trying to solve this problem before AlphaFold. And just to be clear, one thing that you mentioned was that there's this kind of co-evolution of mutations and that you see this again and again in different species. So explain why does that give us a good hint that they're close by to each other? Yeah.RJ [00:13:41]: Um, like think of it this way that, you know, if I have, you know, some amino acid that mutates, it's going to impact everything around it. Right. In three dimensions. And so it's almost like the protein through several, probably random mutations and evolution, like, you know, ends up sort of figuring out that this other amino acid needs to change as well for the structure to be conserved. Uh, so this whole principle is that the structure is probably largely conserved, you know, because there's this function associated with it. And so it's really sort of like different positions compensating for, for each other. I see.Brandon [00:14:17]: Those hints in aggregate give us a lot. Yeah. So you can start to look at what kinds of information about what is close to each other, and then you can start to look at what kinds of folds are possible given the structure and then what is the end state.RJ [00:14:30]: And therefore you can make a lot of inferences about what the actual total shape is. Yeah, that's right. It's almost like, you know, you have this big, like three dimensional Valley, you know, where you're sort of trying to find like these like low energy states and there's so much to search through. That's almost overwhelming. But these hints, they sort of maybe put you in. An area of the space that's already like, kind of close to the solution, maybe not quite there yet. And, and there's always this question of like, how much physics are these models learning, you know, versus like, just pure like statistics. And like, I think one of the thing, at least I believe is that once you're in that sort of approximate area of the solution space, then the models have like some understanding, you know, of how to get you to like, you know, the lower energy, uh, low energy state. And so maybe you have some, some light understanding. Of physics, but maybe not quite enough, you know, to know how to like navigate the whole space. Right. Okay.Brandon [00:15:25]: So we need to give it these hints to kind of get into the right Valley and then it finds the, the minimum or something. Yeah.Gabriel [00:15:31]: One interesting explanation about our awful free works that I think it's quite insightful, of course, doesn't cover kind of the entirety of, of what awful does that is, um, they're going to borrow from, uh, Sergio Chinico for MIT. So he sees kind of awful. Then the interesting thing about awful is God. This very peculiar architecture that we have seen, you know, used, and this architecture operates on this, you know, pairwise context between amino acids. And so the idea is that probably the MSA gives you this first hint about what potential amino acids are close to each other. MSA is most multiple sequence alignment. Exactly. Yeah. Exactly. This evolutionary information. Yeah. And, you know, from this evolutionary information about potential contacts, then is almost as if the model is. of running some kind of, you know, diastro algorithm where it's sort of decoding, okay, these have to be closed. Okay. Then if these are closed and this is connected to this, then this has to be somewhat closed. And so you decode this, that becomes basically a pairwise kind of distance matrix. And then from this rough pairwise distance matrix, you decode kind of theBrandon [00:16:42]: actual potential structure. Interesting. So there's kind of two different things going on in the kind of coarse grain and then the fine grain optimizations. Interesting. Yeah. Very cool.Gabriel [00:16:53]: Yeah. You mentioned AlphaFold3. So maybe we have a good time to move on to that. So yeah, AlphaFold2 came out and it was like, I think fairly groundbreaking for this field. Everyone got very excited. A few years later, AlphaFold3 came out and maybe for some more history, like what were the advancements in AlphaFold3? And then I think maybe we'll, after that, we'll talk a bit about the sort of how it connects to Bolt. But anyway. Yeah. So after AlphaFold2 came out, you know, Jeremy and I got into the field and with many others, you know, the clear problem that, you know, was, you know, obvious after that was, okay, now we can do individual chains. Can we do interactions, interaction, different proteins, proteins with small molecules, proteins with other molecules. And so. So why are interactions important? Interactions are important because to some extent that's kind of the way that, you know, these machines, you know, these proteins have a function, you know, the function comes by the way that they interact with other proteins and other molecules. Actually, in the first place, you know, the individual machines are often, as Jeremy was mentioning, not made of a single chain, but they're made of the multiple chains. And then these multiple chains interact with other molecules to give the function to those. And on the other hand, you know, when we try to intervene of these interactions, think about like a disease, think about like a, a biosensor or many other ways we are trying to design the molecules or proteins that interact in a particular way with what we would call a target protein or target. You know, this problem after AlphaVol2, you know, became clear, kind of one of the biggest problems in the field to, to solve many groups, including kind of ours and others, you know, started making some kind of contributions to this problem of trying to model these interactions. And AlphaVol3 was, you know, was a significant advancement on the problem of modeling interactions. And one of the interesting thing that they were able to do while, you know, some of the rest of the field that really tried to try to model different interactions separately, you know, how protein interacts with small molecules, how protein interacts with other proteins, how RNA or DNA have their structure, they put everything together and, you know, train very large models with a lot of advances, including kind of changing kind of systems. Some of the key architectural choices and managed to get a single model that was able to set this new state-of-the-art performance across all of these different kind of modalities, whether that was protein, small molecules is critical to developing kind of new drugs, protein, protein, understanding, you know, interactions of, you know, proteins with RNA and DNAs and so on.Brandon [00:19:39]: Just to satisfy the AI engineers in the audience, what were some of the key architectural and data, data changes that made that possible?Gabriel [00:19:48]: Yeah, so one critical one that was not necessarily just unique to AlphaFold3, but there were actually a few other teams, including ours in the field that proposed this, was moving from, you know, modeling structure prediction as a regression problem. So where there is a single answer and you're trying to shoot for that answer to a generative modeling problem where you have a posterior distribution of possible structures and you're trying to sample this distribution. And this achieves two things. One is it starts to allow us to try to model more dynamic systems. As we said, you know, some of these structures can actually take multiple structures. And so, you know, you can now model that, you know, through kind of modeling the entire distribution. But on the second hand, from more kind of core modeling questions, when you move from a regression problem to a generative modeling problem, you are really tackling the way that you think about uncertainty in the model in a different way. So if you think about, you know, I'm undecided between different answers, what's going to happen in a regression model is that, you know, I'm going to try to make an average of those different kind of answers that I had in mind. When you have a generative model, what you're going to do is, you know, sample all these different answers and then maybe use separate models to analyze those different answers and pick out the best. So that was kind of one of the critical improvement. The other improvement is that they significantly simplified, to some extent, the architecture, especially of the final model that takes kind of those pairwise representations and turns them into an actual structure. And that now looks a lot more like a more traditional transformer than, you know, like a very specialized equivariant architecture that it was in AlphaFold3.Brandon [00:21:41]: So this is a bitter lesson, a little bit.Gabriel [00:21:45]: There is some aspect of a bitter lesson, but the interesting thing is that it's very far from, you know, being like a simple transformer. This field is one of the, I argue, very few fields in applied machine learning where we still have kind of architecture that are very specialized. And, you know, there are many people that have tried to replace these architectures with, you know, simple transformers. And, you know, there is a lot of debate in the field, but I think kind of that most of the consensus is that, you know, the performance... that we get from the specialized architecture is vastly superior than what we get through a single transformer. Another interesting thing that I think on the staying on the modeling machine learning side, which I think it's somewhat counterintuitive seeing some of the other kind of fields and applications is that scaling hasn't really worked kind of the same in this field. Now, you know, models like AlphaFold2 and AlphaFold3 are, you know, still very large models.RJ [00:29:14]: in a place, I think, where we had, you know, some experience working in, you know, with the data and working with this type of models. And I think that put us already in like a good place to, you know, to produce it quickly. And, you know, and I would even say, like, I think we could have done it quicker. The problem was like, for a while, we didn't really have the compute. And so we couldn't really train the model. And actually, we only trained the big model once. That's how much compute we had. We could only train it once. And so like, while the model was training, we were like, finding bugs left and right. A lot of them that I wrote. And like, I remember like, I was like, sort of like, you know, doing like, surgery in the middle, like stopping the run, making the fix, like relaunching. And yeah, we never actually went back to the start. We just like kept training it with like the bug fixes along the way, which was impossible to reproduce now. Yeah, yeah, no, that model is like, has gone through such a curriculum that, you know, learned some weird stuff. But yeah, somehow by miracle, it worked out.Gabriel [00:30:13]: The other funny thing is that the way that we were training, most of that model was through a cluster from the Department of Energy. But that's sort of like a shared cluster that many groups use. And so we were basically training the model for two days, and then it would go back to the queue and stay a week in the queue. Oh, yeah. And so it was pretty painful. And so we actually kind of towards the end with Evan, the CEO of Genesis, and basically, you know, I was telling him a bit about the project and, you know, kind of telling him about this frustration with the compute. And so luckily, you know, he offered to kind of help. And so we, we got the help from Genesis to, you know, finish up the model. Otherwise, it probably would have taken a couple of extra weeks.Brandon [00:30:57]: Yeah, yeah.Brandon [00:31:02]: And then, and then there's some progression from there.Gabriel [00:31:06]: Yeah, so I would say kind of that, both one, but also kind of these other kind of set of models that came around the same time, were kind of approaching were a big leap from, you know, kind of the previous kind of open source models, and, you know, kind of really kind of approaching the level of AlphaVault 3. But I would still say that, you know, even to this day, there are, you know, some... specific instances where AlphaVault 3 works better. I think one common example is antibody antigen prediction, where, you know, AlphaVault 3 still seems to have an edge in many situations. Obviously, these are somewhat different models. They are, you know, you run them, you obtain different results. So it's, it's not always the case that one model is better than the other, but kind of in aggregate, we still, especially at the time.Brandon [00:32:00]: So AlphaVault 3 is, you know, still having a bit of an edge. We should talk about this more when we talk about Boltzgen, but like, how do you know one is, one model is better than the other? Like you, so you, I make a prediction, you make a prediction, like, how do you know?Gabriel [00:32:11]: Yeah, so easily, you know, the, the great thing about kind of structural prediction and, you know, once we're going to go into the design space of designing new small molecule, new proteins, this becomes a lot more complex. But a great thing about structural prediction is that a bit like, you know, CASP was doing, basically the way that you can evaluate them is that, you know, you train... You know, you train a model on a structure that was, you know, released across the field up until a certain time. And, you know, one of the things that we didn't talk about that was really critical in all this development is the PDB, which is the Protein Data Bank. It's this common resources, basically common database where every biologist publishes their structures. And so we can, you know, train on, you know, all the structures that were put in the PDB until a certain date. And then... And then we basically look for recent structures, okay, which structures look pretty different from anything that was published before, because we really want to try to understand generalization.Brandon [00:33:13]: And then on this new structure, we evaluate all these different models. And so you just know when AlphaFold3 was trained, you know, when you're, you intentionally trained to the same date or something like that. Exactly. Right. Yeah.Gabriel [00:33:24]: And so this is kind of the way that you can somewhat easily kind of compare these models, obviously, that assumes that, you know, the training. You've always been very passionate about validation. I remember like DiffDoc, and then there was like DiffDocL and DocGen. You've thought very carefully about this in the past. Like, actually, I think DocGen is like a really funny story that I think, I don't know if you want to talk about that. It's an interesting like... Yeah, I think one of the amazing things about putting things open source is that we get a ton of feedback from the field. And, you know, sometimes we get kind of great feedback of people. Really like... But honestly, most of the times, you know, to be honest, that's also maybe the most useful feedback is, you know, people sharing about where it doesn't work. And so, you know, at the end of the day, it's critical. And this is also something, you know, across other fields of machine learning. It's always critical to set, to do progress in machine learning, set clear benchmarks. And as, you know, you start doing progress of certain benchmarks, then, you know, you need to improve the benchmarks and make them harder and harder. And this is kind of the progression of, you know, how the field operates. And so, you know, the example of DocGen was, you know, we published this initial model called DiffDoc in my first year of PhD, which was sort of like, you know, one of the early models to try to predict kind of interactions between proteins, small molecules, that we bought a year after AlphaFold2 was published. And now, on the one hand, you know, on these benchmarks that we were using at the time, DiffDoc was doing really well, kind of, you know, outperforming kind of some of the traditional physics-based methods. But on the other hand, you know, when we started, you know, kind of giving these tools to kind of many biologists, and one example was that we collaborated with was the group of Nick Polizzi at Harvard. We noticed, started noticing that there was this clear, pattern where four proteins that were very different from the ones that we're trained on, the models was, was struggling. And so, you know, that seemed clear that, you know, this is probably kind of where we should, you know, put our focus on. And so we first developed, you know, with Nick and his group, a new benchmark, and then, you know, went after and said, okay, what can we change? And kind of about the current architecture to improve this pattern and generalization. And this is the same that, you know, we're still doing today, you know, kind of, where does the model not work, you know, and then, you know, once we have that benchmark, you know, let's try to, through everything we, any ideas that we have of the problem.RJ [00:36:15]: And there's a lot of like healthy skepticism in the field, which I think, you know, is, is, is great. And I think, you know, it's very clear that there's a ton of things, the models don't really work well on, but I think one thing that's probably, you know, undeniable is just like the pace of, pace of progress, you know, and how, how much better we're getting, you know, every year. And so I think if you, you know, if you assume, you know, any constant, you know, rate of progress moving forward, I think things are going to look pretty cool at some point in the future.Gabriel [00:36:42]: ChatGPT was only three years ago. Yeah, I mean, it's wild, right?RJ [00:36:45]: Like, yeah, yeah, yeah, it's one of those things. Like, you've been doing this. Being in the field, you don't see it coming, you know? And like, I think, yeah, hopefully we'll, you know, we'll, we'll continue to have as much progress we've had the past few years.Brandon [00:36:55]: So this is maybe an aside, but I'm really curious, you get this great feedback from the, from the community, right? By being open source. My question is partly like, okay, yeah, if you open source and everyone can copy what you did, but it's also maybe balancing priorities, right? Where you, like all my customers are saying. I want this, there's all these problems with the model. Yeah, yeah. But my customers don't care, right? So like, how do you, how do you think about that? Yeah.Gabriel [00:37:26]: So I would say a couple of things. One is, you know, part of our goal with Bolts and, you know, this is also kind of established as kind of the mission of the public benefit company that we started is to democratize the access to these tools. But one of the reasons why we realized that Bolts needed to be a company, it couldn't just be an academic project is that putting a model on GitHub is definitely not enough to get, you know, chemists and biologists, you know, across, you know, both academia, biotech and pharma to use your model to, in their therapeutic programs. And so a lot of what we think about, you know, at Bolts beyond kind of the, just the models is thinking about all the layers. The layers that come on top of the models to get, you know, from, you know, those models to something that can really enable scientists in the industry. And so that goes, you know, into building kind of the right kind of workflows that take in kind of, for example, the data and try to answer kind of directly that those problems that, you know, the chemists and the biologists are asking, and then also kind of building the infrastructure. And so this to say that, you know, even with models fully open. You know, we see a ton of potential for, you know, products in the space and the critical part about a product is that even, you know, for example, with an open source model, you know, running the model is not free, you know, as we were saying, these are pretty expensive model and especially, and maybe we'll get into this, you know, these days we're seeing kind of pretty dramatic inference time scaling of these models where, you know, the more you run them, the better the results are. But there, you know, you see. You start getting into a point that compute and compute costs becomes a critical factor. And so putting a lot of work into building the right kind of infrastructure, building the optimizations and so on really allows us to provide, you know, a much better service potentially to the open source models. That to say, you know, even though, you know, with a product, we can provide a much better service. I do still think, and we will continue to put a lot of our models open source because the critical kind of role. I think of open source. Models is, you know, helping kind of the community progress on the research and, you know, from which we, we all benefit. And so, you know, we'll continue to on the one hand, you know, put some of our kind of base models open source so that the field can, can be on top of it. And, you know, as we discussed earlier, we learn a ton from, you know, the way that the field uses and builds on top of our models, but then, you know, try to build a product that gives the best experience possible to scientists. So that, you know, like a chemist or a biologist doesn't need to, you know, spin off a GPU and, you know, set up, you know, our open source model in a particular way, but can just, you know, a bit like, you know, I, even though I am a computer scientist, machine learning scientist, I don't necessarily, you know, take a open source LLM and try to kind of spin it off. But, you know, I just maybe open a GPT app or a cloud code and just use it as an amazing product. We kind of want to give the same experience. So this front world.Brandon [00:40:40]: I heard a good analogy yesterday that a surgeon doesn't want the hospital to design a scalpel, right?Brandon [00:40:48]: So just buy the scalpel.RJ [00:40:50]: You wouldn't believe like the number of people, even like in my short time, you know, between AlphaFold3 coming out and the end of the PhD, like the number of people that would like reach out just for like us to like run AlphaFold3 for them, you know, or things like that. Just because like, you know, bolts in our case, you know, just because it's like. It's like not that easy, you know, to do that, you know, if you're not a computational person. And I think like part of the goal here is also that, you know, we continue to obviously build the interface with computational folks, but that, you know, the models are also accessible to like a larger, broader audience. And then that comes from like, you know, good interfaces and stuff like that.Gabriel [00:41:27]: I think one like really interesting thing about bolts is that with the release of it, you didn't just release a model, but you created a community. Yeah. Did that community, it grew very quickly. Did that surprise you? And like, what is the evolution of that community and how is that fed into bolts?RJ [00:41:43]: If you look at its growth, it's like very much like when we release a new model, it's like, there's a big, big jump, but yeah, it's, I mean, it's been great. You know, we have a Slack community that has like thousands of people on it. And it's actually like self-sustaining now, which is like the really nice part because, you know, it's, it's almost overwhelming, I think, you know, to be able to like answer everyone's questions and help. It's really difficult, you know. The, the few people that we were, but it ended up that like, you know, people would answer each other's questions and like, sort of like, you know, help one another. And so the Slack, you know, has been like kind of, yeah, self, self-sustaining and that's been, it's been really cool to see.RJ [00:42:21]: And, you know, that's, that's for like the Slack part, but then also obviously on GitHub as well. We've had like a nice, nice community. You know, I think we also aspire to be even more active on it, you know, than we've been in the past six months, which has been like a bit challenging, you know, for us. But. Yeah, the community has been, has been really great and, you know, there's a lot of papers also that have come out with like new evolutions on top of bolts and it's surprised us to some degree because like there's a lot of models out there. And I think like, you know, sort of people converging on that was, was really cool. And, you know, I think it speaks also, I think, to the importance of like, you know, when, when you put code out, like to try to put a lot of emphasis and like making it like as easy to use as possible and something we thought a lot about when we released the code base. You know, it's far from perfect, but, you know.Brandon [00:43:07]: Do you think that that was one of the factors that caused your community to grow is just the focus on easy to use, make it accessible? I think so.RJ [00:43:14]: Yeah. And we've, we've heard it from a few people over the, over the, over the years now. And, you know, and some people still think it should be a lot nicer and they're, and they're right. And they're right. But yeah, I think it was, you know, at the time, maybe a little bit easier than, than other things.Gabriel [00:43:29]: The other thing part, I think led to, to the community and to some extent, I think, you know, like the somewhat the trust in the community. Kind of what we, what we put out is the fact that, you know, it's not really been kind of, you know, one model, but, and maybe we'll talk about it, you know, after Boltz 1, you know, there were maybe another couple of models kind of released, you know, or open source kind of soon after. We kind of continued kind of that open source journey or at least Boltz 2, where we are not only improving kind of structure prediction, but also starting to do affinity predictions, understanding kind of the strength of the interactions between these different models, which is this critical component. critical property that you often want to optimize in discovery programs. And then, you know, more recently also kind of protein design model. And so we've sort of been building this suite of, of models that come together, interact with one another, where, you know, kind of, there is almost an expectation that, you know, we, we take very at heart of, you know, always having kind of, you know, across kind of the entire suite of different tasks, the best or across the best. model out there so that it's sort of like our open source tool can be kind of the go-to model for everybody in the, in the industry. I really want to talk about Boltz 2, but before that, one last question in this direction, was there anything about the community which surprised you? Were there any, like, someone was doing something and you're like, why would you do that? That's crazy. Or that's actually genius. And I never would have thought about that.RJ [00:45:01]: I mean, we've had many contributions. I think like some of the. Interesting ones, like, I mean, we had, you know, this one individual who like wrote like a complex GPU kernel, you know, for part of the architecture on a piece of, the funny thing is like that piece of the architecture had been there since AlphaFold 2, and I don't know why it took Boltz for this, you know, for this person to, you know, to decide to do it, but that was like a really great contribution. We've had a bunch of others, like, you know, people figuring out like ways to, you know, hack the model to do something. They click peptides, like, you know, there's, I don't know if there's any other interesting ones come to mind.Gabriel [00:45:41]: One cool one, and this was, you know, something that initially was proposed as, you know, as a message in the Slack channel by Tim O'Donnell was basically, he was, you know, there are some cases, especially, for example, we discussed, you know, antibody-antigen interactions where the models don't necessarily kind of get the right answer. What he noticed is that, you know, the models were somewhat stuck into predicting kind of the antibodies. And so he basically ran the experiments in this model, you can condition, basically, you can give hints. And so he basically gave, you know, random hints to the model, basically, okay, you should bind to this residue, you should bind to the first residue, or you should bind to the 11th residue, or you should bind to the 21st residue, you know, basically every 10 residues scanning the entire antigen.Brandon [00:46:33]: Residues are the...Gabriel [00:46:34]: The amino acids. The amino acids, yeah. So the first amino acids. The 11 amino acids, and so on. So it's sort of like doing a scan, and then, you know, conditioning the model to predict all of them, and then looking at the confidence of the model in each of those cases and taking the top. And so it's sort of like a very somewhat crude way of doing kind of inference time search. But surprisingly, you know, for antibody-antigen prediction, it actually kind of helped quite a bit. And so there's some, you know, interesting ideas that, you know, obviously, as kind of developing the model, you say kind of, you know, wow. This is why would the model, you know, be so dumb. But, you know, it's very interesting. And that, you know, leads you to also kind of, you know, start thinking about, okay, how do I, can I do this, you know, not with this brute force, but, you know, in a smarter way.RJ [00:47:22]: And so we've also done a lot of work on that direction. And that speaks to, like, the, you know, the power of scoring. We're seeing that a lot. I'm sure we'll talk about it more when we talk about BullsGen. But, you know, our ability to, like, take a structure and determine that that structure is, like... Good. You know, like, somewhat accurate. Whether that's a single chain or, like, an interaction is a really powerful way of improving, you know, the models. Like, sort of like, you know, if you can sample a ton and you assume that, like, you know, if you sample enough, you're likely to have, like, you know, the good structure. Then it really just becomes a ranking problem. And, you know, now we're, you know, part of the inference time scaling that Gabby was talking about is very much that. It's like, you know, the more we sample, the more we, like, you know, the ranking model. The ranking model ends up finding something it really likes. And so I think our ability to get better at ranking, I think, is also what's going to enable sort of the next, you know, next big, big breakthroughs. Interesting.Brandon [00:48:17]: But I guess there's a, my understanding, there's a diffusion model and you generate some stuff and then you, I guess, it's just what you said, right? Then you rank it using a score and then you finally... And so, like, can you talk about those different parts? Yeah.Gabriel [00:48:34]: So, first of all, like, the... One of the critical kind of, you know, beliefs that we had, you know, also when we started working on Boltz 1 was sort of like the structure prediction models are somewhat, you know, our field version of some foundation models, you know, learning about kind of how proteins and other molecules interact. And then we can leverage that learning to do all sorts of other things. And so with Boltz 2, we leverage that learning to do affinity predictions. So understanding kind of, you know, if I give you this protein, this molecule. How tightly is that interaction? For Boltz 1, what we did was taking kind of that kind of foundation models and then fine tune it to predict kind of entire new proteins. And so the way basically that that works is sort of like instead of for the protein that you're designing, instead of fitting in an actual sequence, you fit in a set of blank tokens. And you train the models to, you know, predict both the structure of kind of that protein. The structure also, what the different amino acids of that proteins are. And so basically the way that Boltz 1 operates is that you feed a target protein that you may want to kind of bind to or, you know, another DNA, RNA. And then you feed the high level kind of design specification of, you know, what you want your new protein to be. For example, it could be like an antibody with a particular framework. It could be a peptide. It could be many other things. And that's with natural language or? And that's, you know, basically, you know, prompting. And we have kind of this sort of like spec that you specify. And, you know, you feed kind of this spec to the model. And then the model translates this into, you know, a set of, you know, tokens, a set of conditioning to the model, a set of, you know, blank tokens. And then, you know, basically the codes as part of the diffusion models, the codes. It's a new structure and a new sequence for your protein. And, you know, basically, then we take that. And as Jeremy was saying, we are trying to score it and, you know, how good of a binder it is to that original target.Brandon [00:50:51]: You're using basically Boltz to predict the folding and the affinity to that molecule. So and then that kind of gives you a score? Exactly.Gabriel [00:51:03]: So you use this model to predict the folding. And then you do two things. One is that you predict the structure and with something like Boltz2, and then you basically compare that structure with what the model predicted, what Boltz2 predicted. And this is sort of like in the field called consistency. It's basically you want to make sure that, you know, the structure that you're predicting is actually what you're trying to design. And that gives you a much better confidence that, you know, that's a good design. And so that's the first filtering. And the second filtering that we did as part of kind of the Boltz2 pipeline that was released is that we look at the confidence that the model has in the structure. Now, unfortunately, kind of going to your question of, you know, predicting affinity, unfortunately, confidence is not a very good predictor of affinity. And so one of the things that we've actually done a ton of progress, you know, since we released Boltz2.Brandon [00:52:03]: And kind of we have some new results that we are going to kind of announce soon is kind of, you know, the ability to get much better hit rates when instead of, you know, trying to rely on confidence of the model, we are actually directly trying to predict the affinity of that interaction. Okay. Just backing up a minute. So your diffusion model actually predicts not only the protein sequence, but also the folding of it. Exactly.Gabriel [00:52:32]: And actually, you can... One of the big different things that we did compared to other models in the space, and, you know, there were some papers that had already kind of done this before, but we really scaled it up was, you know, basically somewhat merging kind of the structure prediction and the sequence prediction into almost the same task. And so the way that Boltz2 works is that you are basically the only thing that you're doing is predicting the structure. So the only sort of... Supervision is we give you a supervision on the structure, but because the structure is atomic and, you know, the different amino acids have a different atomic composition, basically from the way that you place the atoms, we also understand not only kind of the structure that you wanted, but also the identity of the amino acid that, you know, the models believed was there. And so we've basically, instead of, you know, having these two supervision signals, you know, one discrete, one continuous. That somewhat, you know, don't interact well together. We sort of like build kind of like an encoding of, you know, sequences in structures that allows us to basically use exactly the same supervision signal that we were using to Boltz2 that, you know, you know, largely similar to what AlphaVol3 proposed, which is very scalable. And we can use that to design new proteins. Oh, interesting.RJ [00:53:58]: Maybe a quick shout out to Hannes Stark on our team who like did all this work. Yeah.Gabriel [00:54:04]: Yeah, that was a really cool idea. I mean, like looking at the paper and there's this is like encoding or you just add a bunch of, I guess, kind of atoms, which can be anything, and then they get sort of rearranged and then basically plopped on top of each other so that and then that encodes what the amino acid is. And there's sort of like a unique way of doing this. It was that was like such a really such a cool, fun idea.RJ [00:54:29]: I think that idea was had existed before. Yeah, there were a couple of papers.Gabriel [00:54:33]: Yeah, I had proposed this and and Hannes really took it to the large scale.Brandon [00:54:39]: In the paper, a lot of the paper for Boltz2Gen is dedicated to actually the validation of the model. In my opinion, all the people we basically talk about feel that this sort of like in the wet lab or whatever the appropriate, you know, sort of like in real world validation is the whole problem or not the whole problem, but a big giant part of the problem. So can you talk a little bit about the highlights? From there, that really because to me, the results are impressive, both from the perspective of the, you know, the model and also just the effort that went into the validation by a large team.Gabriel [00:55:18]: First of all, I think I should start saying is that both when we were at MIT and Thomas Yacolas and Regina Barzillai's lab, as well as at Boltz, you know, we are not a we're not a biolab and, you know, we are not a therapeutic company. And so to some extent, you know, we were first forced to, you know, look outside of, you know, our group, our team to do the experimental validation. One of the things that really, Hannes, in the team pioneer was the idea, OK, can we go not only to, you know, maybe a specific group and, you know, trying to find a specific system and, you know, maybe overfit a bit to that system and trying to validate. But how can we test this model? So. Across a very wide variety of different settings so that, you know, anyone in the field and, you know, printing design is, you know, such a kind of wide task with all sorts of different applications from therapeutic to, you know, biosensors and many others that, you know, so can we get a validation that is kind of goes across many different tasks? And so he basically put together, you know, I think it was something like, you know, 25 different. You know, academic and industry labs that committed to, you know, testing some of the designs from the model and some of this testing is still ongoing and, you know, giving results kind of back to us in exchange for, you know, hopefully getting some, you know, new great sequences for their task. And he was able to, you know, coordinate this, you know, very wide set of, you know, scientists and already in the paper, I think we. Shared results from, I think, eight to 10 different labs kind of showing results from, you know, designing peptides, designing to target, you know, ordered proteins, peptides targeting disordered proteins, which are results, you know, of designing proteins that bind to small molecules, which are results of, you know, designing nanobodies and across a wide variety of different targets. And so that's sort of like. That gave to the paper a lot of, you know, validation to the model, a lot of validation that was kind of wide.Brandon [00:57:39]: And so those would be therapeutics for those animals or are they relevant to humans as well? They're relevant to humans as well.Gabriel [00:57:45]: Obviously, you need to do some work into, quote unquote, humanizing them, making sure that, you know, they have the right characteristics to so they're not toxic to humans and so on.RJ [00:57:57]: There are some approved medicine in the market that are nanobodies. There's a general. General pattern, I think, in like in trying to design things that are smaller, you know, like it's easier to manufacture at the same time, like that comes with like potentially other challenges, like maybe a little bit less selectivity than like if you have something that has like more hands, you know, but the yeah, there's this big desire to, you know, try to design many proteins, nanobodies, small peptides, you know, that just are just great drug modalities.Brandon [00:58:27]: Okay. I think we were left off. We were talking about validation. Validation in the lab. And I was very excited about seeing like all the diverse validations that you've done. Can you go into some more detail about them? Yeah. Specific ones. Yeah.RJ [00:58:43]: The nanobody one. I think we did. What was it? 15 targets. Is that correct? 14. 14 targets. Testing. So we typically the way this works is like we make a lot of designs. All right. On the order of like tens of thousands. And then we like rank them and we pick like the top. And in this case, and was 15 right for each target and then we like measure sort of like the success rates, both like how many targets we were able to get a binder for and then also like more generally, like out of all of the binders that we designed, how many actually proved to be good binders. Some of the other ones I think involved like, yeah, like we had a cool one where there was a small molecule or design a protein that binds to it. That has a lot of like interesting applications, you know, for example. Like Gabri mentioned, like biosensing and things like that, which is pretty cool. We had a disordered protein, I think you mentioned also. And yeah, I think some of those were some of the highlights. Yeah.Gabriel [00:59:44]: So I would say that the way that we structure kind of some of those validations was on the one end, we have validations across a whole set of different problems that, you know, the biologists that we were working with came to us with. So we were trying to. For example, in some of the experiments, design peptides that would target the RACC, which is a target that is involved in metabolism. And we had, you know, a number of other applications where we were trying to design, you know, peptides or other modalities against some other therapeutic relevant targets. We designed some proteins to bind small molecules. And then some of the other testing that we did was really trying to get like a more broader sense. So how does the model work, especially when tested, you know, on somewhat generalization? So one of the things that, you know, we found with the field was that a lot of the validation, especially outside of the validation that was on specific problems, was done on targets that have a lot of, you know, known interactions in the training data. And so it's always a bit hard to understand, you know, how much are these models really just regurgitating kind of what they've seen or trying to imitate. What they've seen in the training data versus, you know, really be able to design new proteins. And so one of the experiments that we did was to take nine targets from the PDB, filtering to things where there is no known interaction in the PDB. So basically the model has never seen kind of this particular protein bound or a similar protein bound to another protein. So there is no way that. The model from its training set can sort of like say, okay, I'm just going to kind of tweak something and just imitate this particular kind of interaction. And so we took those nine proteins. We worked with adaptive CRO and basically tested, you know, 15 mini proteins and 15 nanobodies against each one of them. And the very cool thing that we saw was that on two thirds of those targets, we were able to, from this 15 design, get nanomolar binders, nanomolar, roughly speaking, just a measure of, you know, how strongly kind of the interaction is, roughly speaking, kind of like a nanomolar binder is approximately the kind of binding strength or binding that you need for a therapeutic. Yeah. So maybe switching directions a bit. Bolt's lab was just announced this week or was it last week? Yeah. This is like your. First, I guess, product, if that's if you want to call it that. Can you talk about what Bolt's lab is and yeah, you know, what you hope that people take away from this? Yeah.RJ [01:02:44]: You know, as we mentioned, like I think at the very beginning is the goal with the product has been to, you know, address what the models don't on their own. And there's largely sort of two categories there. I'll split it in three. The first one. It's one thing to predict, you know, a single interaction, for example, like a single structure. It's another to like, you know, very effectively search a space, a design space to produce something of value. What we found, like sort of building on this product is that there's a lot of steps involved, you know, in that there's certainly need to like, you know, accompany the user through, you know, one of those steps, for example, is like, you know, the creation of the target itself. You know, how do we make sure that the model has like a good enough understanding of the target? So we can like design something and there's all sorts of tricks, you know, that you can do to improve like a particular, you know, structure prediction. And so that's sort of like, you know, the first stage. And then there's like this stage of like, you know, designing and searching the space efficiently. You know, for something like BullsGen, for example, like you, you know, you design many things and then you rank them, for example, for small molecule process, a little bit more complicated. We actually need to also make sure that the molecules are synthesizable. And so the way we do that is that, you know, we have a generative model that learns. To use like appropriate building blocks such that, you know, it can design within a space that we know is like synthesizable. And so there's like, you know, this whole pipeline really of different models involved in being able to design a molecule. And so that's been sort of like the first thing we call them agents. We have a protein agent and we have a small molecule design agents. And that's really like at the core of like what powers, you know, the BullsLab platform.Brandon [01:04:22]: So these agents, are they like a language model wrapper or they're just like your models and you're just calling them agents? A lot. Yeah. Because they, they, they sort of perform a function on behalf of.RJ [01:04:33]: They're more of like a, you know, a recipe, if you wish. And I think we use that term sort of because of, you know, sort of the complex pipelining and automation, you know, that goes into like all this plumbing. So that's the first part of the product. The second part is the infrastructure. You know, we need to be able to do this at very large scale for any one, you know, group that's doing a design campaign. Let's say you're designing, you know, I'd say a hundred thousand possible candidates. Right. To find the good one that is, you know, a very large amount of compute, you know, for small molecules, it's on the order of like a few seconds per designs for proteins can be a bit longer. And so, you know, ideally you want to do that in parallel, otherwise it's going to take you weeks. And so, you know, we've put a lot of effort into like, you know, our ability to have a GPU fleet that allows any one user, you know, to be able to do this kind of like large parallel search.Brandon [01:05:23]: So you're amortizing the cost over your users. Exactly. Exactly.RJ [01:05:27]: And, you know, to some degree, like it's whether you. Use 10,000 GPUs for like, you know, a minute is the same cost as using, you know, one GPUs for God knows how long. Right. So you might as well try to parallelize if you can. So, you know, a lot of work has gone, has gone into that, making it very robust, you know, so that we can have like a lot of people on the platform doing that at the same time. And the third one is, is the interface and the interface comes in, in two shapes. One is in form of an API and that's, you know, really suited for companies that want to integrate, you know, these pipelines, these agents.RJ [01:06:01]: So we're already partnering with, you know, a few distributors, you know, that are gonna integrate our API. And then the second part is the user interface. And, you know, we, we've put a lot of thoughts also into that. And this is when I, I mentioned earlier, you know, this idea of like broadening the audience. That's kind of what the, the user interface is about. And we've built a lot of interesting features in it, you know, for example, for collaboration, you know, when you have like potentially multiple medicinal chemists or. We're going through the results and trying to pick out, okay, like what are the molecules that we're going to go and test in the lab? It's powerful for them to be able to, you know, for example, each provide their own ranking and then do consensus building. And so there's a lot of features around launching these large jobs, but also around like collaborating on analyzing the results that we try to solve, you know, with that part of the platform. So Bolt's lab is sort of a combination of these three objectives into like one, you know, sort of cohesive platform. Who is this accessible to? Everyone. You do need to request access today. We're still like, you know, sort of ramping up the usage, but anyone can request access. If you are an academic in particular, we, you know, we provide a fair amount of free credit so you can play with the platform. If you are a startup or biotech, you may also, you know, reach out and we'll typically like actually hop on a call just to like understand what you're trying to do and also provide a lot of free credit to get started. And of course, also with larger companies, we can deploy this platform in a more like secure environment. And so that's like more like customizing. You know, deals that we make, you know, with the partners, you know, and that's sort of the ethos of Bolt. I think this idea of like servicing everyone and not necessarily like going after just, you know, the really large enterprises. And that starts from the open source, but it's also, you know, a key design principle of the product itself.Gabriel [01:07:48]: One thing I was thinking about with regards to infrastructure, like in the LLM space, you know, the cost of a token has gone down by I think a factor of a thousand or so over the last three years, right? Yeah. And is it possible that like essentially you can exploit economies of scale and infrastructure that you can make it cheaper to run these things yourself than for any person to roll their own system? A hundred percent. Yeah.RJ [01:08:08]: I mean, we're already there, you know, like running Bolts on our platform, especially on a large screen is like considerably cheaper than it would probably take anyone to put the open source model out there and run it. And on top of the infrastructure, like one of the things that we've been working on is accelerating the models. So, you know. Our small molecule screening pipeline is 10x faster on Bolts Lab than it is in the open source, you know, and that's also part of like, you know, building a product, you know, of something that scales really well. And we really wanted to get to a point where like, you know, we could keep prices very low in a way that it would be a no-brainer, you know, to use Bolts through our platform.Gabriel [01:08:52]: How do you think about validation of your like agentic systems? Because, you know, as you were saying earlier. Like we're AlphaFold style models are really good at, let's say, monomeric, you know, proteins where you have, you know, co-evolution data. But now suddenly the whole point of this is to design something which doesn't have, you know, co-evolution data, something which is really novel. So now you're basically leaving the domain that you thought was, you know, that you know you are good at. So like, how do you validate that?RJ [01:09:22]: Yeah, I like every complete, but there's obviously, you know, a ton of computational metrics. That we rely on, but those are only take you so far. You really got to go to the lab, you know, and test, you know, okay, with this method A and this method B, how much better are we? You know, how much better is my, my hit rate? How stronger are my binders? Also, it's not just about hit rate. It's also about how good the binders are. And there's really like no way, nowhere around that. I think we're, you know, we've really ramped up the amount of experimental validation that we do so that we like really track progress, you know, as scientifically sound, you know. Yeah. As, as possible out of this, I think.Gabriel [01:10:00]: Yeah, no, I think, you know, one thing that is unique about us and maybe companies like us is that because we're not working on like maybe a couple of therapeutic pipelines where, you know, our validation would be focused on those. We, when we do an experimental validation, we try to test it across tens of targets. And so that on the one end, we can get a much more statistically significant result and, and really allows us to make progress. From the methodological side without being, you know, steered by, you know, overfitting on any one particular system. And of course we choose, you know, w

From Palantir and Two Sigma to building Goodfire into the poster-child for actionable mechanistic interpretability, Mark Bissell (Member of Technical Staff) and Myra Deng (Head of Product) are trying to turn “peeking inside the model” into a repeatable production workflow by shipping APIs, landing real enterprise deployments, and now scaling the bet with a recent $150M Series B funding round at a $1.25B valuation.In this episode, we go far beyond the usual “SAEs are cool” take. We talk about Goodfire's core bet: that the AI lifecycle is still fundamentally broken because the only reliable control we have is data and we post-train, RLHF, and fine-tune by “slurping supervision through a straw,” hoping the model picks up the right behaviors while quietly absorbing the wrong ones. Goodfire's answer is to build a bi-directional interface between humans and models: read what's happening inside, edit it surgically, and eventually use interpretability during training so customization isn't just brute-force guesswork.Mark and Myra walk through what that looks like when you stop treating interpretability like a lab demo and start treating it like infrastructure: lightweight probes that add near-zero latency, token-level safety filters that can run at inference time, and interpretability workflows that survive messy constraints (multilingual inputs, synthetic→real transfer, regulated domains, no access to sensitive data). We also get a live window into what “frontier-scale interp” means operationally (i.e. steering a trillion-parameter model in real time by targeting internal features) plus why the same tooling generalizes cleanly from language models to genomics, medical imaging, and “pixel-space” world models.We discuss:* Myra + Mark's path: Palantir (health systems, forward-deployed engineering) → Goodfire early team; Two Sigma → Head of Product, translating frontier interpretability research into a platform and real-world deployments* What “interpretability” actually means in practice: not just post-hoc poking, but a broader “science of deep learning” approach across the full AI lifecycle (data curation → post-training → internal representations → model design)* Why post-training is the first big wedge: “surgical edits” for unintended behaviors likereward hacking, sycophancy, noise learned during customization plus the dream of targeted unlearning and bias removal without wrecking capabilities* SAEs vs probes in the real world: why SAE feature spaces sometimes underperform classifiers trained on raw activations for downstream detection tasks (hallucination, harmful intent, PII), and what that implies about “clean concept spaces”* Rakuten in production: deploying interpretability-based token-level PII detection at inference time to prevent routing private data to downstream providers plus the gnarly constraints: no training on real customer PII, synthetic→real transfer, English + Japanese, and tokenization quirks* Why interp can be operationally cheaper than LLM-judge guardrails: probes are lightweight, low-latency, and don't require hosting a second large model in the loop* Real-time steering at frontier scale: a demo of steering Kimi K2 (~1T params) live and finding features via SAE pipelines, auto-labeling via LLMs, and toggling a “Gen-Z slang” feature across multiple layers without breaking tool use* Hallucinations as an internal signal: the case that models have latent uncertainty / “user-pleasing” circuitry you can detect and potentially mitigate more directly than black-box methods* Steering vs prompting: the emerging view that activation steering and in-context learning are more closely connected than people think, including work mapping between the two (even for jailbreak-style behaviors)* Interpretability for science: using the same tooling across domains (genomics, medical imaging, materials) to debug spurious correlations and extract new knowledge up to and including early biomarker discovery work with major partners* World models + “pixel-space” interpretability: why vision/video models make concepts easier to see, how that accelerates the feedback loop, and why robotics/world-model partners are especially interesting design partners* The north star: moving from “data in, weights out” to intentional model design where experts can impart goals and constraints directly, not just via reward signals and brute-force post-training—Goodfire AI* Website: https://goodfire.ai* LinkedIn: https://www.linkedin.com/company/goodfire-ai/* X: https://x.com/GoodfireAIMyra Deng* Website: https://myradeng.com/* LinkedIn: https://www.linkedin.com/in/myra-deng/* X: https://x.com/myra_dengMark Bissell* LinkedIn: https://www.linkedin.com/in/mark-bissell/* X: https://x.com/MarkMBissellFull Video EpisodeTimestamps00:00:00 Introduction00:00:05 Introduction to the Latent Space Podcast and Guests from Goodfire00:00:29 What is Goodfire? Mission and Focus on Interpretability00:01:01 Goodfire's Practical Approach to Interpretability00:01:37 Goodfire's Series B Fundraise Announcement00:02:04 Backgrounds of Mark and Myra from Goodfire00:02:51 Team Structure and Roles at Goodfire00:05:13 What is Interpretability? Definitions and Techniques00:05:30 Understanding Errors00:07:29 Post-training vs. Pre-training Interpretability Applications00:08:51 Using Interpretability to Remove Unwanted Behaviors00:10:09 Grokking, Double Descent, and Generalization in Models00:10:15 404 Not Found Explained00:12:06 Subliminal Learning and Hidden Biases in Models00:14:07 How Goodfire Chooses Research Directions and Projects00:15:00 Troubleshooting Errors00:16:04 Limitations of SAEs and Probes in Interpretability00:18:14 Rakuten Case Study: Production Deployment of Interpretability00:20:45 Conclusion00:21:12 Efficiency Benefits of Interpretability Techniques00:21:26 Live Demo: Real-Time Steering in a Trillion Parameter Model00:25:15 How Steering Features are Identified and Labeled00:26:51 Detecting and Mitigating Hallucinations Using Interpretability00:31:20 Equivalence of Activation Steering and Prompting00:34:06 Comparing Steering with Fine-Tuning and LoRA Techniques00:36:04 Model Design and the Future of Intentional AI Development00:38:09 Getting Started in Mechinterp: Resources, Programs, and Open Problems00:40:51 Industry Applications and the Rise of Mechinterp in Practice00:41:39 Interpretability for Code Models and Real-World Usage00:43:07 Making Steering Useful for More Than Stylistic Edits00:46:17 Applying Interpretability to Healthcare and Scientific Discovery00:49:15 Why Interpretability is Crucial in High-Stakes Domains like Healthcare00:52:03 Call for Design Partners Across Domains00:54:18 Interest in World Models and Visual Interpretability00:57:22 Sci-Fi Inspiration: Ted Chiang and Interpretability01:00:14 Interpretability, Safety, and Alignment Perspectives01:04:27 Weak-to-Strong Generalization and Future Alignment Challenges01:05:38 Final Thoughts and Hiring/Collaboration Opportunities at GoodfireTranscriptShawn Wang [00:00:05]: So welcome to the Latent Space pod. We're back in the studio with our special MechInterp co-host, Vibhu. Welcome. Mochi, Mochi's special co-host. And Mochi, the mechanistic interpretability doggo. We have with us Mark and Myra from Goodfire. Welcome. Thanks for having us on. Maybe we can sort of introduce Goodfire and then introduce you guys. How do you introduce Goodfire today?Myra Deng [00:00:29]: Yeah, it's a great question. So Goodfire, we like to say, is an AI research lab that focuses on using interpretability to understand, learn from, and design AI models. And we really believe that interpretability will unlock the new generation, next frontier of safe and powerful AI models. That's our description right now, and I'm excited to dive more into the work we're doing to make that happen.Shawn Wang [00:00:55]: Yeah. And there's always like the official description. Is there an understatement? Is there an unofficial one that sort of resonates more with a different audience?Mark Bissell [00:01:01]: Well, being an AI research lab that's focused on interpretability, there's obviously a lot of people have a lot that they think about when they think of interpretability. And I think we have a pretty broad definition of what that means and the types of places that can be applied. And in particular, applying it in production scenarios, in high stakes industries, and really taking it sort of from the research world into the real world. Which, you know. It's a new field, so that hasn't been done all that much. And we're excited about actually seeing that sort of put into practice.Shawn Wang [00:01:37]: Yeah, I would say it wasn't too long ago that Anthopic was like still putting out like toy models or superposition and that kind of stuff. And I wouldn't have pegged it to be this far along. When you and I talked at NeurIPS, you were talking a little bit about your production use cases and your customers. And then not to bury the lead, today we're also announcing the fundraise, your Series B. $150 million. $150 million at a 1.25B valuation. Congrats, Unicorn.Mark Bissell [00:02:02]: Thank you. Yeah, no, things move fast.Shawn Wang [00:02:04]: We were talking to you in December and already some big updates since then. Let's dive, I guess, into a bit of your backgrounds as well. Mark, you were at Palantir working on health stuff, which is really interesting because the Goodfire has some interesting like health use cases. I don't know how related they are in practice.Mark Bissell [00:02:22]: Yeah, not super related, but I don't know. It was helpful context to know what it's like. Just to work. Just to work with health systems and generally in that domain. Yeah.Shawn Wang [00:02:32]: And Mara, you were at Two Sigma, which actually I was also at Two Sigma back in the day. Wow, nice.Myra Deng [00:02:37]: Did we overlap at all?Shawn Wang [00:02:38]: No, this is when I was briefly a software engineer before I became a sort of developer relations person. And now you're head of product. What are your sort of respective roles, just to introduce people to like what all gets done in Goodfire?Mark Bissell [00:02:51]: Yeah, prior to Goodfire, I was at Palantir for about three years as a forward deployed engineer, now a hot term. Wasn't always that way. And as a technical lead on the health care team and at Goodfire, I'm a member of the technical staff. And honestly, that I think is about as specific as like as as I could describe myself because I've worked on a range of things. And, you know, it's it's a fun time to be at a team that's still reasonably small. I think when I joined one of the first like ten employees, now we're above 40, but still, it looks like there's always a mix of research and engineering and product and all of the above. That needs to get done. And I think everyone across the team is, you know, pretty, pretty switch hitter in the roles they do. So I think you've seen some of the stuff that I worked on related to image models, which was sort of like a research demo. More recently, I've been working on our scientific discovery team with some of our life sciences partners, but then also building out our core platform for more of like flexing some of the kind of MLE and developer skills as well.Shawn Wang [00:03:53]: Very generalist. And you also had like a very like a founding engineer type role.Myra Deng [00:03:58]: Yeah, yeah.Shawn Wang [00:03:59]: So I also started as I still am a member of technical staff, did a wide range of things from the very beginning, including like finding our office space and all of this, which is we both we both visited when you had that open house thing. It was really nice.Myra Deng [00:04:13]: Thank you. Thank you. Yeah. Plug to come visit our office.Shawn Wang [00:04:15]: It looked like it was like 200 people. It has room for 200 people. But you guys are like 10.Myra Deng [00:04:22]: For a while, it was very empty. But yeah, like like Mark, I spend. A lot of my time as as head of product, I think product is a bit of a weird role these days, but a lot of it is thinking about how do we take our frontier research and really apply it to the most important real world problems and how does that then translate into a platform that's repeatable or a product and working across, you know, the engineering and research teams to make that happen and also communicating to the world? Like, what is interpretability? What is it used for? What is it good for? Why is it so important? All of these things are part of my day-to-day as well.Shawn Wang [00:05:01]: I love like what is things because that's a very crisp like starting point for people like coming to a field. They all do a fun thing. Vibhu, why don't you want to try tackling what is interpretability and then they can correct us.Vibhu Sapra [00:05:13]: Okay, great. So I think like one, just to kick off, it's a very interesting role to be head of product, right? Because you guys, at least as a lab, you're more of an applied interp lab, right? Which is pretty different than just normal interp, like a lot of background research. But yeah. You guys actually ship an API to try these things. You have Ember, you have products around it, which not many do. Okay. What is interp? So basically you're trying to have an understanding of what's going on in model, like in the model, in the internal. So different approaches to do that. You can do probing, SAEs, transcoders, all this stuff. But basically you have an, you have a hypothesis. You have something that you want to learn about what's happening in a model internals. And then you're trying to solve that from there. You can do stuff like you can, you know, you can do activation mapping. You can try to do steering. There's a lot of stuff that you can do, but the key question is, you know, from input to output, we want to have a better understanding of what's happening and, you know, how can we, how can we adjust what's happening on the model internals? How'd I do?Mark Bissell [00:06:12]: That was really good. I think that was great. I think it's also a, it's kind of a minefield of a, if you ask 50 people who quote unquote work in interp, like what is interpretability, you'll probably get 50 different answers. And. Yeah. To some extent also like where, where good fire sits in the space. I think that we're an AI research company above all else. And interpretability is a, is a set of methods that we think are really useful and worth kind of specializing in, in order to accomplish the goals we want to accomplish. But I think we also sort of see some of the goals as even more broader as, as almost like the science of deep learning and just taking a not black box approach to kind of any part of the like AI development life cycle, whether that. That means using interp for like data curation while you're training your model or for understanding what happened during post-training or for the, you know, understanding activations and sort of internal representations, what is in there semantically. And then a lot of sort of exciting updates that were, you know, are sort of also part of the, the fundraise around bringing interpretability to training, which I don't think has been done all that much before. A lot of this stuff is sort of post-talk poking at models as opposed to. To actually using this to intentionally design them.Shawn Wang [00:07:29]: Is this post-training or pre-training or is that not a useful.Myra Deng [00:07:33]: Currently focused on post-training, but there's no reason the techniques wouldn't also work in pre-training.Shawn Wang [00:07:38]: Yeah. It seems like it would be more active, applicable post-training because basically I'm thinking like rollouts or like, you know, having different variations of a model that you can tweak with the, with your steering. Yeah.Myra Deng [00:07:50]: And I think in a lot of the news that you've seen in, in, on like Twitter or whatever, you've seen a lot of unintended. Side effects come out of post-training processes, you know, overly sycophantic models or models that exhibit strange reward hacking behavior. I think these are like extreme examples. There's also, you know, very, uh, mundane, more mundane, like enterprise use cases where, you know, they try to customize or post-train a model to do something and it learns some noise or it doesn't appropriately learn the target task. And a big question that we've always had is like, how do you use your understanding of what the model knows and what it's doing to actually guide the learning process?Shawn Wang [00:08:26]: Yeah, I mean, uh, you know, just to anchor this for people, uh, one of the biggest controversies of last year was 4.0 GlazeGate. I've never heard of GlazeGate. I didn't know that was what it was called. The other one, they called it that on the blog post and I was like, well, how did OpenAI call it? Like officially use that term. And I'm like, that's funny, but like, yeah, I guess it's the pitch that if they had worked a good fire, they wouldn't have avoided it. Like, you know what I'm saying?Myra Deng [00:08:51]: I think so. Yeah. Yeah.Mark Bissell [00:08:53]: I think that's certainly one of the use cases. I think. Yeah. Yeah. I think the reason why post-training is a place where this makes a lot of sense is a lot of what we're talking about is surgical edits. You know, you want to be able to have expert feedback, very surgically change how your model is doing, whether that is, you know, removing a certain behavior that it has. So, you know, one of the things that we've been looking at or is, is another like common area where you would want to make a somewhat surgical edit is some of the models that have say political bias. Like you look at Quen or, um, R1 and they have sort of like this CCP bias.Shawn Wang [00:09:27]: Is there a CCP vector?Mark Bissell [00:09:29]: Well, there's, there are certainly internal, yeah. Parts of the representation space where you can sort of see where that lives. Yeah. Um, and you want to kind of, you know, extract that piece out.Shawn Wang [00:09:40]: Well, I always say, you know, whenever you find a vector, a fun exercise is just like, make it very negative to see what the opposite of CCP is.Mark Bissell [00:09:47]: The super America, bald eagles flying everywhere. But yeah. So in general, like lots of post-training tasks where you'd want to be able to, to do that. Whether it's unlearning a certain behavior or, you know, some of the other kind of cases where this comes up is, are you familiar with like the, the grokking behavior? I mean, I know the machine learning term of grokking.Shawn Wang [00:10:09]: Yeah.Mark Bissell [00:10:09]: Sort of this like double descent idea of, of having a model that is able to learn a generalizing, a generalizing solution, as opposed to even if memorization of some task would suffice, you want it to learn the more general way of doing a thing. And so, you know, another. A way that you can think about having surgical access to a model's internals would be learn from this data, but learn in the right way. If there are many possible, you know, ways to, to do that. Can make interp solve the double descent problem?Shawn Wang [00:10:41]: Depends, I guess, on how you. Okay. So I, I, I viewed that double descent as a problem because then you're like, well, if the loss curves level out, then you're done, but maybe you're not done. Right. Right. But like, if you actually can interpret what is a generalizing or what you're doing. What is, what is still changing, even though the loss is not changing, then maybe you, you can actually not view it as a double descent problem. And actually you're just sort of translating the space in which you view loss and like, and then you have a smooth curve. Yeah.Mark Bissell [00:11:11]: I think that's certainly like the domain of, of problems that we're, that we're looking to get.Shawn Wang [00:11:15]: Yeah. To me, like double descent is like the biggest thing to like ML research where like, if you believe in scaling, then you don't need, you need to know where to scale. And. But if you believe in double descent, then you don't, you don't believe in anything where like anything levels off, like.Vibhu Sapra [00:11:30]: I mean, also tendentially there's like, okay, when you talk about the China vector, right. There's the subliminal learning work. It was from the anthropic fellows program where basically you can have hidden biases in a model. And as you distill down or, you know, as you train on distilled data, those biases always show up, even if like you explicitly try to not train on them. So, you know, it's just like another use case of. Okay. If we can interpret what's happening in post-training, you know, can we clear some of this? Can we even determine what's there? Because yeah, it's just like some worrying research that's out there that shows, you know, we really don't know what's going on.Mark Bissell [00:12:06]: That is. Yeah. I think that's the biggest sentiment that we're sort of hoping to tackle. Nobody knows what's going on. Right. Like subliminal learning is just an insane concept when you think about it. Right. Train a model on not even the logits, literally the output text of a bunch of random numbers. And now your model loves owls. And you see behaviors like that, that are just, they defy, they defy intuition. And, and there are mathematical explanations that you can get into, but. I mean.Shawn Wang [00:12:34]: It feels so early days. Objectively, there are a sequence of numbers that are more owl-like than others. There, there should be.Mark Bissell [00:12:40]: According to, according to certain models. Right. It's interesting. I think it only applies to models that were initialized from the same starting Z. Usually, yes.Shawn Wang [00:12:49]: But I mean, I think that's a, that's a cheat code because there's not enough compute. But like if you believe in like platonic representation, like probably it will transfer across different models as well. Oh, you think so?Mark Bissell [00:13:00]: I think of it more as a statistical artifact of models initialized from the same seed sort of. There's something that is like path dependent from that seed that might cause certain overlaps in the latent space and then sort of doing this distillation. Yeah. Like it pushes it towards having certain other tendencies.Vibhu Sapra [00:13:24]: Got it. I think there's like a bunch of these open-ended questions, right? Like you can't train in new stuff during the RL phase, right? RL only reorganizes weights and you can only do stuff that's somewhat there in your base model. You're not learning new stuff. You're just reordering chains and stuff. But okay. My broader question is when you guys work at an interp lab, how do you decide what to work on and what's kind of the thought process? Right. Because we can ramble for hours. Okay. I want to know this. I want to know that. But like, how do you concretely like, you know, what's the workflow? Okay. There's like approaches towards solving a problem, right? I can try prompting. I can look at chain of thought. I can train probes, SAEs. But how do you determine, you know, like, okay, is this going anywhere? Like, do we have set stuff? Just, you know, if you can help me with all that. Yeah.Myra Deng [00:14:07]: It's a really good question. I feel like we've always at the very beginning of the company thought about like, let's go and try to learn what isn't working in machine learning today. Whether that's talking to customers or talking to researchers at other labs, trying to understand both where the frontier is going and where things are really not falling apart today. And then developing a perspective on how we can push the frontier using interpretability methods. And so, you know, even our chief scientist, Tom, spends a lot of time talking to customers and trying to understand what real world problems are and then taking that back and trying to apply the current state of the art to those problems and then seeing where they fall down basically. And then using those failures or those shortcomings to understand what hills to climb when it comes to interpretability research. So like on the fundamental side, for instance, when we have done some work applying SAEs and probes, we've encountered, you know, some shortcomings in SAEs that we found a little bit surprising. And so have gone back to the drawing board and done work on that. And then, you know, we've done some work on better foundational interpreter models. And a lot of our team's research is focused on what is the next evolution beyond SAEs, for instance. And then when it comes to like control and design of models, you know, we tried steering with our first API and realized that it still fell short of black box techniques like prompting or fine tuning. And so went back to the drawing board and we're like, how do we make that not the case and how do we improve it beyond that? And one of our researchers, Ekdeep, who just joined is actually Ekdeep and Atticus are like steering experts and have spent a lot of time trying to figure out like, what is the research that enables us to actually do this in a much more powerful, robust way? So yeah, the answer is like, look at real world problems, try to translate that into a research agenda and then like hill climb on both of those at the same time.Shawn Wang [00:16:04]: Yeah. Mark has the steering CLI demo queued up, which we're going to go into in a sec. But I always want to double click on when you drop hints, like we found some problems with SAEs. Okay. What are they? You know, and then we can go into the demo. Yeah.Myra Deng [00:16:19]: I mean, I'm curious if you have more thoughts here as well, because you've done it in the healthcare domain. But I think like, for instance, when we do things like trying to detect behaviors within models that are harmful or like behaviors that a user might not want to have in their model. So hallucinations, for instance, harmful intent, PII, all of these things. We first tried using SAE probes for a lot of these tasks. So taking the feature activation space from SAEs and then training classifiers on top of that, and then seeing how well we can detect the properties that we might want to detect in model behavior. And we've seen in many cases that probes just trained on raw activations seem to perform better than SAE probes, which is a bit surprising if you think that SAEs are actually also capturing the concepts that you would want to capture cleanly and more surgically. And so that is an interesting observation. I don't think that is like, I'm not down on SAEs at all. I think there are many, many things they're useful for, but we have definitely run into cases where I think the concept space described by SAEs is not as clean and accurate as we would expect it to be for actual like real world downstream performance metrics.Mark Bissell [00:17:34]: Fair enough. Yeah. It's the blessing and the curse of unsupervised methods where you get to peek into the AI's mind. But sometimes you wish that you saw other things when you walked inside there. Although in the PII instance, I think weren't an SAE based approach actually did prove to be the most generalizable?Myra Deng [00:17:53]: It did work well in the case that we published with Rakuten. And I think a lot of the reasons it worked well was because we had a noisier data set. And so actually the blessing of unsupervised learning is that we actually got to get more meaningful, generalizable signal from SAEs when the data was noisy. But in other cases where we've had like good data sets, it hasn't been the case.Shawn Wang [00:18:14]: And just because you named Rakuten and I don't know if we'll get it another chance, like what is the overall, like what is Rakuten's usage or production usage? Yeah.Myra Deng [00:18:25]: So they are using us to essentially guardrail and inference time monitor their language model usage and their agent usage to detect things like PII so that they don't route private user information.Myra Deng [00:18:41]: And so that's, you know, going through all of their user queries every day. And that's something that we deployed with them a few months ago. And now we are actually exploring very early partnerships, not just with Rakuten, but with other people around how we can help with potentially training and customization use cases as well. Yeah.Shawn Wang [00:19:03]: And for those who don't know, like it's Rakuten is like, I think number one or number two e-commerce store in Japan. Yes. Yeah.Mark Bissell [00:19:10]: And I think that use case actually highlights a lot of like what it looks like to deploy things in practice that you don't always think about when you're doing sort of research tasks. So when you think about some of the stuff that came up there that's more complex than your idealized version of a problem, they were encountering things like synthetic to real transfer of methods. So they couldn't train probes, classifiers, things like that on actual customer data of PII. So what they had to do is use synthetic data sets. And then hope that that transfer is out of domain to real data sets. And so we can evaluate performance on the real data sets, but not train on customer PII. So that right off the bat is like a big challenge. You have multilingual requirements. So this needed to work for both English and Japanese text. Japanese text has all sorts of quirks, including tokenization behaviors that caused lots of bugs that caused us to be pulling our hair out. And then also a lot of tasks you'll see. You might make simplifying assumptions if you're sort of treating it as like the easiest version of the problem to just sort of get like general results where maybe you say you're classifying a sentence to say, does this contain PII? But the need that Rakuten had was token level classification so that you could precisely scrub out the PII. So as we learned more about the problem, you're sort of speaking about what that looks like in practice. Yeah. A lot of assumptions end up breaking. And that was just one instance where you. A problem that seems simple right off the bat ends up being more complex as you keep diving into it.Vibhu Sapra [00:20:41]: Excellent. One of the things that's also interesting with Interp is a lot of these methods are very efficient, right? So where you're just looking at a model's internals itself compared to a separate like guardrail, LLM as a judge, a separate model. One, you have to host it. Two, there's like a whole latency. So if you use like a big model, you have a second call. Some of the work around like self detection of hallucination, it's also deployed for efficiency, right? So if you have someone like Rakuten doing it in production live, you know, that's just another thing people should consider.Mark Bissell [00:21:12]: Yeah. And something like a probe is super lightweight. Yeah. It's no extra latency really. Excellent.Shawn Wang [00:21:17]: You have the steering demos lined up. So we were just kind of see what you got. I don't, I don't actually know if this is like the latest, latest or like alpha thing.Mark Bissell [00:21:26]: No, this is a pretty hacky demo from from a presentation that someone else on the team recently gave. So this will give a sense for, for technology. So you can see the steering and action. Honestly, I think the biggest thing that this highlights is that as we've been growing as a company and taking on kind of more and more ambitious versions of interpretability related problems, a lot of that comes to scaling up in various different forms. And so here you're going to see steering on a 1 trillion parameter model. This is Kimi K2. And so it's sort of fun that in addition to the research challenges, there are engineering challenges that we're now tackling. Cause for any of this to be sort of useful in production, you need to be thinking about what it looks like when you're using these methods on frontier models as opposed to sort of like toy kind of model organisms. So yeah, this was thrown together hastily, pretty fragile behind the scenes, but I think it's quite a fun demo. So screen sharing is on. So I've got two terminal sessions pulled up here. On the left is a forked version that we have of the Kimi CLI that we've got running to point at our custom hosted Kimi model. And then on the right is a set up that will allow us to steer on certain concepts. So I should be able to chat with Kimi over here. Tell it hello. This is running locally. So the CLI is running locally, but the Kimi server is running back to the office. Well, hopefully should be, um, that's too much to run on that Mac. Yeah. I think it's, uh, it takes a full, like each 100 node. I think it's like, you can. You can run it on eight GPUs, eight 100. So, so yeah, Kimi's running. We can ask it a prompt. It's got a forked version of our, uh, of the SG line code base that we've been working on. So I'm going to tell it, Hey, this SG line code base is slow. I think there's a bug. Can you try to figure it out? There's a big code base, so it'll, it'll spend some time doing this. And then on the right here, I'm going to initialize in real time. Some steering. Let's see here.Mark Bissell [00:23:33]: searching for any. Bugs. Feature ID 43205.Shawn Wang [00:23:38]: Yeah.Mark Bissell [00:23:38]: 20, 30, 40. So let me, uh, this is basically a feature that we found that inside Kimi seems to cause it to speak in Gen Z slang. And so on the left, it's still sort of thinking normally it might take, I don't know, 15 seconds for this to kick in, but then we're going to start hopefully seeing him do this code base is massive for real. So we're going to start. We're going to start seeing Kimi transition as the steering kicks in from normal Kimi to Gen Z Kimi and both in its chain of thought and its actual outputs.Mark Bissell [00:24:19]: And interestingly, you can see, you know, it's still able to call tools, uh, and stuff. It's um, it's purely sort of it's it's demeanor. And there are other features that we found for interesting things like concision. So that's more of a practical one. You can make it more concise. Um, the types of programs, uh, programming languages that uses, but yeah, as we're seeing it come in. Pretty good. Outputs.Shawn Wang [00:24:43]: Scheduler code is actually wild.Vibhu Sapra [00:24:46]: Yo, this code is actually insane, bro.Vibhu Sapra [00:24:53]: What's the process of training in SAE on this, or, you know, how do you label features? I know you guys put out a pretty cool blog post about, um, finding this like autonomous interp. Um, something. Something about how agents for interp is different than like coding agents. I don't know while this is spewing up, but how, how do we find feature 43, two Oh five. Yeah.Mark Bissell [00:25:15]: So in this case, um, we, our platform that we've been building out for a long time now supports all the sort of classic out of the box interp techniques that you might want to have like SAE training, probing things of that kind, I'd say the techniques for like vanilla SAEs are pretty well established now where. You take your model that you're interpreting, run a whole bunch of data through it, gather activations, and then yeah, pretty straightforward pipeline to train an SAE. There are a lot of different varieties. There's top KSAEs, batch top KSAEs, um, normal ReLU SAEs. And then once you have your sparse features to your point, assigning labels to them to actually understand that this is a gen Z feature, that's actually where a lot of the kind of magic happens. Yeah. And the most basic standard technique is look at all of your d input data set examples that cause this feature to fire most highly. And then you can usually pick out a pattern. So for this feature, If I've run a diverse enough data set through my model feature 43, two Oh five. Probably tends to fire on all the tokens that sounds like gen Z slang. You know, that's the, that's the time of year to be like, Oh, I'm in this, I'm in this Um, and, um, so, you know, you could have a human go through all 43,000 concepts andVibhu Sapra [00:26:34]: And I've got to ask the basic question, you know, can we get examples where it hallucinates, pass it through, see what feature activates for hallucinations? Can I just, you know, turn hallucination down?Myra Deng [00:26:51]: Oh, wow. You really predicted a project we're already working on right now, which is detecting hallucinations using interpretability techniques. And this is interesting because hallucinations is something that's very hard to detect. And it's like a kind of a hairy problem and something that black box methods really struggle with. Whereas like Gen Z, you could always train a simple classifier to detect that hallucinations is harder. But we've seen that models internally have some... Awareness of like uncertainty or some sort of like user pleasing behavior that leads to hallucinatory behavior. And so, yeah, we have a project that's trying to detect that accurately. And then also working on mitigating the hallucinatory behavior in the model itself as well.Shawn Wang [00:27:39]: Yeah, I would say most people are still at the level of like, oh, I would just turn temperature to zero and that turns off hallucination. And I'm like, well, that's a fundamental misunderstanding of how this works. Yeah.Mark Bissell [00:27:51]: Although, so part of what I like about that question is you, there are SAE based approaches that might like help you get at that. But oftentimes the beauty of SAEs and like we said, the curse is that they're unsupervised. So when you have a behavior that you deliberately would like to remove, and that's more of like a supervised task, often it is better to use something like probes and specifically target the thing that you're interested in reducing as opposed to sort of like hoping that when you fragment the latent space, one of the vectors that pops out.Vibhu Sapra [00:28:20]: And as much as we're training an autoencoder to be sparse, we're not like for sure certain that, you know, we will get something that just correlates to hallucination. You'll probably split that up into 20 other things and who knows what they'll be.Mark Bissell [00:28:36]: Of course. Right. Yeah. So there's no sort of problems with like feature splitting and feature absorption. And then there's the off target effects, right? Ideally, you would want to be very precise where if you reduce the hallucination feature, suddenly maybe your model can't write. Creatively anymore. And maybe you don't like that, but you want to still stop it from hallucinating facts and figures.Shawn Wang [00:28:55]: Good. So Vibhu has a paper to recommend there that we'll put in the show notes. But yeah, I mean, I guess just because your demo is done, any any other things that you want to highlight or any other interesting features you want to show?Mark Bissell [00:29:07]: I don't think so. Yeah. Like I said, this is a pretty small snippet. I think the main sort of point here that I think is exciting is that there's not a whole lot of inter being applied to models quite at this scale. You know, Anthropic certainly has some some. Research and yeah, other other teams as well. But it's it's nice to see these techniques, you know, being put into practice. I think not that long ago, the idea of real time steering of a trillion parameter model would have sounded.Shawn Wang [00:29:33]: Yeah. The fact that it's real time, like you started the thing and then you edited the steering vector.Vibhu Sapra [00:29:38]: I think it's it's an interesting one TBD of what the actual like production use case would be on that, like the real time editing. It's like that's the fun part of the demo, right? You can kind of see how this could be served behind an API, right? Like, yes, you're you only have so many knobs and you can just tweak it a bit more. And I don't know how it plays in. Like people haven't done that much with like, how does this work with or without prompting? Right. How does this work with fine tuning? Like, there's a whole hype of continual learning, right? So there's just so much to see. Like, is this another parameter? Like, is it like parameter? We just kind of leave it as a default. We don't use it. So I don't know. Maybe someone here wants to put out a guide on like how to use this with prompting when to do what?Mark Bissell [00:30:18]: Oh, well, I have a paper recommendation. I think you would love from Act Deep on our team, who is an amazing researcher, just can't say enough amazing things about Act Deep. But he actually has a paper that as well as some others from the team and elsewhere that go into the essentially equivalence of activation steering and in context learning and how those are from a he thinks of everything in a cognitive neuroscience Bayesian framework, but basically how you can precisely show how. Prompting in context, learning and steering exhibit similar behaviors and even like get quantitative about the like magnitude of steering you would need to do to induce a certain amount of behavior similar to certain prompting, even for things like jailbreaks and stuff. It's a really cool paper. Are you saying steering is less powerful than prompting? More like you can almost write a formula that tells you how to convert between the two of them.Myra Deng [00:31:20]: And so like formally equivalent actually in the in the limit. Right.Mark Bissell [00:31:24]: So like one case study of this is for jailbreaks there. I don't know. Have you seen the stuff where you can do like many shot jailbreaking? You like flood the context with examples of the behavior. And the topic put out that paper.Shawn Wang [00:31:38]: A lot of people were like, yeah, we've been doing this, guys.Mark Bissell [00:31:40]: Like, yeah, what's in this in context learning and activation steering equivalence paper is you can like predict the number. Number of examples that you will need to put in there in order to jailbreak the model. That's cool. By doing steering experiments and using this sort of like equivalence mapping. That's cool. That's really cool. It's very neat. Yeah.Shawn Wang [00:32:02]: I was going to say, like, you know, I can like back rationalize that this makes sense because, you know, what context is, is basically just, you know, it updates the KV cache kind of and like and then every next token inference is still like, you know, the sheer sum of everything all the way. It's plus all the context. It's up to date. And you could, I guess, theoretically steer that with you probably replace that with your steering. The only problem is steering typically is on one layer, maybe three layers like like you did. So it's like not exactly equivalent.Mark Bissell [00:32:33]: Right, right. There's sort of you need to get precise about, yeah, like how you sort of define steering and like what how you're modeling the setup. But yeah, I've got the paper pulled up here. Belief dynamics reveal the dual nature. Yeah. The title is Belief Dynamics Reveal the Dual Nature of Incompetence. And it's an exhibition of the practical context learning and activation steering. So Eric Bigelow, Dan Urgraft on the who are doing fellowships at Goodfire, Ekt Deep's the final author there.Myra Deng [00:32:59]: I think actually to your question of like, what is the production use case of steering? I think maybe if you just think like one level beyond steering as it is today. Like imagine if you could adapt your model to be, you know, an expert legal reasoner. Like in almost real time, like very quickly. efficiently using human feedback or using like your semantic understanding of what the model knows and where it knows that behavior. I think that while it's not clear what the product is at the end of the day, it's clearly very valuable. Thinking about like what's the next interface for model customization and adaptation is a really interesting problem for us. Like we have heard a lot of people actually interested in fine-tuning an RL for open weight models in production. And so people are using things like Tinker or kind of like open source libraries to do that, but it's still very difficult to get models fine-tuned and RL'd for exactly what you want them to do unless you're an expert at model training. And so that's like something we'reShawn Wang [00:34:06]: looking into. Yeah. I never thought so. Tinker from Thinking Machines famously uses rank one LoRa. Is that basically the same as steering? Like, you know, what's the comparison there?Mark Bissell [00:34:19]: Well, so in that case, you are still applying updates to the parameters, right?Shawn Wang [00:34:25]: Yeah. You're not touching a base model. You're touching an adapter. It's kind of, yeah.Mark Bissell [00:34:30]: Right. But I guess it still is like more in parameter space then. I guess it's maybe like, are you modifying the pipes or are you modifying the water flowing through the pipes to get what you're after? Yeah. Just maybe one way.Mark Bissell [00:34:44]: I like that analogy. That's my mental map of it at least, but it gets at this idea of model design and intentional design, which is something that we're, that we're very focused on. And just the fact that like, I hope that we look back at how we're currently training models and post-training models and just think what a primitive way of doing that right now. Like there's no intentionalityShawn Wang [00:35:06]: really in... It's just data, right? The only thing in control is what data we feed in.Mark Bissell [00:35:11]: So, so Dan from Goodfire likes to use this analogy of, you know, he has a couple of young kids and he talks about like, what if I could only teach my kids how to be good people by giving them cookies or like, you know, giving them a slap on the wrist if they do something wrong, like not telling them why it was wrong or like what they should have done differently or something like that. Just figure it out. Right. Exactly. So that's RL. Yeah. Right. And, and, you know, it's sample inefficient. There's, you know, what do they say? It's like slurping feedback. It's like, slurping supervision. Right. And so you'd like to get to the point where you can have experts giving feedback to their models that are, uh, internalized and, and, you know, steering is an inference time way of sort of getting that idea. But ideally you're moving to a world whereVibhu Sapra [00:36:04]: it is much more intentional design in perpetuity for these models. Okay. This is one of the questions we asked Emmanuel from Anthropic on the podcast a few months ago. Basically the question, was you're at a research lab that does model training, foundation models, and you're on an interp team. How does it tie back? Right? Like, does this, do ideas come from the pre-training team? Do they go back? Um, you know, so for those interested, you can, you can watch that. There wasn't too much of a connect there, but it's still something, you know, it's something they want toMark Bissell [00:36:33]: push for down the line. It can be useful for all of the above. Like there are certainly post-hocVibhu Sapra [00:36:39]: use cases where it doesn't need to touch that. I think the other thing a lot of people forget is this stuff isn't too computationally expensive, right? Like I would say, if you're interested in getting into research, MechInterp is one of the most approachable fields, right? A lot of this train an essay, train a probe, this stuff, like the budget for this one, there's already a lot done. There's a lot of open source work. You guys have done some too. Um, you know,Shawn Wang [00:37:04]: There's like notebooks from the Gemini team for Neil Nanda or like, this is how you do it. Just step through the notebook.Vibhu Sapra [00:37:09]: Even if you're like, not even technical with any of this, you can still make like progress. There, you can look at different activations, but, uh, if you do want to get into training, you know, training this stuff, correct me if I'm wrong is like in the thousands of dollars, not even like, it's not that high scale. And then same with like, you know, applying it, doing it for post-training or all this stuff is fairly cheap in scale of, okay. I want to get into like model training. I don't have compute for like, you know, pre-training stuff. So it's, it's a very nice field to get into. And also there's a lot of like open questions, right? Um, some of them have to go with, okay, I want a product. I want to solve this. Like there's also just a lot of open-ended stuff that people could work on. That's interesting. Right. I don't know if you guys have any calls for like, what's open questions, what's open work that you either open collaboration with, or like, you'd just like to see solved or just, you know, for people listening that want to get into McInturk because people always talk about it. What are, what are the things they should check out? Start, of course, you know, join you guys as well. I'm sure you're hiring.Myra Deng [00:38:09]: There's a paper, I think from, was it Lee, uh, Sharky? It's open problems and, uh, it's, it's a bit of interpretability, which I recommend everyone who's interested in the field. Read. I'm just like a really comprehensive overview of what are the things that experts in the field think are the most important problems to be solved. I also think to your point, it's been really, really inspiring to see, I think a lot of young people getting interested in interpretability, actually not just young people also like scientists to have been, you know, experts in physics for many years and in biology or things like this, um, transitioning into interp, because the barrier of, of what's now interp. So it's really cool to see a number to entry is, you know, in some ways low and there's a lot of information out there and ways to get started. There's this anecdote of like professors at universities saying that all of a sudden every incoming PhD student wants to study interpretability, which was not the case a few years ago. So it just goes to show how, I guess, like exciting the field is, how fast it's moving, how quick it is to get started and things like that.Mark Bissell [00:39:10]: And also just a very welcoming community. You know, there's an open source McInturk Slack channel. There are people are always posting questions and just folks in the space are always responsive if you ask things on various forums and stuff. But yeah, the open paper, open problems paper is a really good one.Myra Deng [00:39:28]: For other people who want to get started, I think, you know, MATS is a great program. What's the acronym for? Machine Learning and Alignment Theory Scholars? It's like the...Vibhu Sapra [00:39:40]: Normally summer internship style.Myra Deng [00:39:42]: Yeah, but they've been doing it year round now. And actually a lot of our full-time staff have come through that program or gone through that program. And it's great for anyone who is transitioning into interpretability. There's a couple other fellows programs. We do one as well as Anthropic. And so those are great places to get started if anyone is interested.Mark Bissell [00:40:03]: Also, I think been seen as a research field for a very long time. But I think engineering... I think engineers are sorely wanted for interpretability as well, especially at Goodfire, but elsewhere, as it does scale up.Shawn Wang [00:40:18]: I should mention that Lee actually works with you guys, right? And in the London office and I'm adding our first ever McInturk track at AI Europe because I see this industry applications now emerging. And I'm pretty excited to, you know, help push that along. Yeah, I was looking forward to that. It'll effectively be the first industry McInturk conference. Yeah. I'm so glad you added that. You know, it's still a little bit of a bet. It's not that widespread, but I can definitely see this is the time to really get into it. We want to be early on things.Mark Bissell [00:40:51]: For sure. And I think the field understands this, right? So at ICML, I think the title of the McInturk workshop this year was actionable interpretability. And there was a lot of discussion around bringing it to various domains. Everyone's adding pragmatic, actionable, whatever.Shawn Wang [00:41:10]: It's like, okay, well, we weren't actionable before, I guess. I don't know.Vibhu Sapra [00:41:13]: And I mean, like, just, you know, being in Europe, you see the Interp room. One, like old school conferences, like, I think they had a very tiny room till they got lucky and they got it doubled. But there's definitely a lot of interest, a lot of niche research. So you see a lot of research coming out of universities, students. We covered the paper last week. It's like two unknown authors, not many citations. But, you know, you can make a lot of meaningful work there. Yeah. Yeah. Yeah.Shawn Wang [00:41:39]: Yeah. I think people haven't really mentioned this yet. It's just Interp for code. I think it's like an abnormally important field. We haven't mentioned this yet. The conspiracy theory last two years ago was when the first SAE work came out of Anthropic was they would do like, oh, we just used SAEs to turn the bad code vector down and then turn up the good code. And I think like, isn't that the dream? Like, you know, like, but basically, I guess maybe, why is it funny? Like, it's... If it was realistic, it would not be funny. It would be like, no, actually, we should do this. But it's funny because we know there's like, we feel there's some limitations to what steering can do. And I think a lot of the public image of steering is like the Gen Z stuff. Like, oh, you can make it really love the Golden Gate Bridge, or you can make it speak like Gen Z. To like be a legal reasoner seems like a huge stretch. Yeah. And I don't know if that will get there this way. Yeah.Myra Deng [00:42:36]: I think, um, I will say we are announcing. Something very soon that I will not speak too much about. Um, but I think, yeah, this is like what we've run into again and again is like, we, we don't want to be in the world where steering is only useful for like stylistic things. That's definitely not, not what we're aiming for. But I think the types of interventions that you need to do to get to things like legal reasoning, um, are much more sophisticated and require breakthroughs in, in learning algorithms. And that's, um...Shawn Wang [00:43:07]: And is this an emergent property of scale as well?Myra Deng [00:43:10]: I think so. Yeah. I mean, I think scale definitely helps. I think scale allows you to learn a lot of information and, and reduce noise across, you know, large amounts of data. But I also think we think that there's ways to do things much more effectively, um, even, even at scale. So like actually learning exactly what you want from the data and not learning things that you do that you don't want exhibited in the data. So we're not like anti-scale, but we are also realizing that scale is not going to get us anywhere. It's not going to get us to the type of AI development that we want to be at in, in the future as these models get more powerful and get deployed in all these sorts of like mission critical contexts. Current life cycle of training and deploying and evaluations is, is to us like deeply broken and has opportunities to, to improve. So, um, more to come on that very, very soon.Mark Bissell [00:44:02]: And I think that that's a use basically, or maybe just like a proof point that these concepts do exist. Like if you can manipulate them in the precise best way, you can get the ideal combination of them that you desire. And steering is maybe the most coarse grained sort of peek at what that looks like. But I think it's evocative of what you could do if you had total surgical control over every concept, every parameter. Yeah, exactly.Myra Deng [00:44:30]: There were like bad code features. I've got it pulled up.Vibhu Sapra [00:44:33]: Yeah. Just coincidentally, as you guys are talking.Shawn Wang [00:44:35]: This is like, this is exactly.Vibhu Sapra [00:44:38]: There's like specifically a code error feature that activates and they show, you know, it's not, it's not typo detection. It's like, it's, it's typos in code. It's not typical typos. And, you know, you can, you can see it clearly activates where there's something wrong in code. And they have like malicious code, code error. They have a whole bunch of sub, you know, sub broken down little grain features. Yeah.Shawn Wang [00:45:02]: Yeah. So, so the, the rough intuition for me, the, why I talked about post-training was that, well, you just, you know, have a few different rollouts with all these things turned off and on and whatever. And then, you know, you can, that's, that's synthetic data you can kind of post-train on. Yeah.Vibhu Sapra [00:45:13]: And I think we make it sound easier than it is just saying, you know, they do the real hard work.Myra Deng [00:45:19]: I mean, you guys, you guys have the right idea. Exactly. Yeah. We replicated a lot of these features in, in our Lama models as well. I remember there was like.Vibhu Sapra [00:45:26]: And I think a lot of this stuff is open, right? Like, yeah, you guys opened yours. DeepMind has opened a lot of essays on Gemma. Even Anthropic has opened a lot of this. There's, there's a lot of resources that, you know, we can probably share of people that want to get involved.Shawn Wang [00:45:41]: Yeah. And special shout out to like Neuronpedia as well. Yes. Like, yeah, amazing piece of work to visualize those things.Myra Deng [00:45:49]: Yeah, exactly.Shawn Wang [00:45:50]: I guess I wanted to pivot a little bit on, onto the healthcare side, because I think that's a big use case for you guys. We haven't really talked about it yet. This is a bit of a crossover for me because we are, we are, we do have a separate science pod that we're starting up for AI, for AI for science, just because like, it's such a huge investment category and also I'm like less qualified to do it, but we actually have bio PhDs to cover that, which is great, but I need to just kind of recover, recap your work, maybe on the evil two stuff, but then, and then building forward.Mark Bissell [00:46:17]: Yeah, for sure. And maybe to frame up the conversation, I think another kind of interesting just lens on interpretability in general is a lot of the techniques that were described. are ways to solve the AI human interface problem. And it's sort of like bidirectional communication is the goal there. So what we've been talking about with intentional design of models and, you know, steering, but also more advanced techniques is having humans impart our desires and control into models and over models. And the reverse is also very interesting, especially as you get to superhuman models, whether that's narrow superintelligence, like these scientific models that work on genomics, data, medical imaging, things like that. But down the line, you know, superintelligence of other forms as well. What knowledge can the AIs teach us as sort of that, that the other direction in that? And so some of our life science work to date has been getting at exactly that question, which is, well, some of it does look like debugging these various life sciences models, understanding if they're actually performing well, on tasks, or if they're picking up on spurious correlations, for instance, genomics models, you would like to know whether they are sort of focusing on the biologically relevant things that you care about, or if it's using some simpler correlate, like the ancestry of the person that it's looking at. But then also in the instances where they are superhuman, and maybe they are understanding elements of the human genome that we don't have names for or specific, you know, yeah, discoveries that they've made that that we don't know about, that's, that's a big goal. And so we're already seeing that, right, we are partnered with organizations like Mayo Clinic, leading research health system in the United States, our Institute, as well as a startup called Prima Menta, which focuses on neurodegenerative disease. And in our partnership with them, we've used foundation models, they've been training and applied our interpretability techniques to find novel biomarkers for Alzheimer's disease. So I think this is just the tip of the iceberg. But it's, that's like a flavor of some of the things that we're working on.Shawn Wang [00:48:36]: Yeah, I think that's really fantastic. Obviously, we did the Chad Zuckerberg pod last year as well. And like, there's a plethora of these models coming out, because there's so much potential and research. And it's like, very interesting how it's basically the same as language models, but just with a different underlying data set. But it's like, it's the same exact techniques. Like, there's no change, basically.Mark Bissell [00:48:59]: Yeah. Well, and even in like other domains, right? Like, you know, robotics, I know, like a lot of the companies just use Gemma as like the like backbone, and then they like make it into a VLA that like takes these actions. It's, it's, it's transformers all the way down. So yeah.Vibhu Sapra [00:49:15]: Like we have Med Gemma now, right? Like this week, even there was Med Gemma 1.5. And they're training it on this stuff, like 3d scans, medical domain knowledge, and all that stuff, too. So there's a push from both sides. But I think the thing that, you know, one of the things about McInturpp is like, you're a little bit more cautious in some domains, right? So healthcare, mainly being one, like guardrails, understanding, you know, we're more risk adverse to something going wrong there. So even just from a basic understanding, like, if we're trusting these systems to make claims, we want to know why and what's going on.Myra Deng [00:49:51]: Yeah, I think there's totally a kind of like deployment bottleneck to actually using. foundation models for real patient usage or things like that. Like, say you're using a model for rare disease prediction, you probably want some explanation as to why your model predicted a certain outcome, and an interpretable explanation at that. So that's definitely a use case. But I also think like, being able to extract scientific information that no human knows to accelerate drug discovery and disease treatment and things like that actually is a really, really big unlock for science, like scientific discovery. And you've seen a lot of startups, like say that they're going to accelerate scientific discovery. And I feel like we actually are doing that through our interp techniques. And kind of like, almost by accident, like, I think we got reached out to very, very early on from these healthcare institutions. And none of us had healthcare.Shawn Wang [00:50:49]: How did they even hear of you? A podcast.Myra Deng [00:50:51]: Oh, okay. Yeah, podcast.Vibhu Sapra [00:50:53]: Okay, well, now's that time, you know.Myra Deng [00:50:55]: Everyone can call us.Shawn Wang [00:50:56]: Podcasts are the most important thing. Everyone should listen to podcasts.Myra Deng [00:50:59]: Yeah, they reached out. They were like, you know, we have these really smart models that we've trained, and we want to know what they're doing. And we were like, really early that time, like three months old, and it was a few of us. And we were like, oh, my God, we've never used these models. Let's figure it out. But it's also like, great proof that interp techniques scale pretty well across domains. We didn't really have to learn too much about.Shawn Wang [00:51:21]: Interp is a machine learning technique, machine learning skills everywhere, right? Yeah. And it's obviously, it's just like a general insight. Yeah. Probably to finance too, I think, which would be fun for our history. I don't know if you have anything to say there.Mark Bissell [00:51:34]: Yeah, well, just across the science. Like, we've also done work on material science. Yeah, it really runs the gamut.Vibhu Sapra [00:51:40]: Yeah. Awesome. And, you know, for those that should reach out, like, you're obviously experts in this, but like, is there a call out for people that you're looking to partner with, design partners, people to use your stuff outside of just, you know, the general developer that wants to. Plug and play steering stuff, like on the research side more so, like, are there ideal design partners, customers, stuff like that?Myra Deng [00:52:03]: Yeah, I can talk about maybe non-life sciences, and then I'm curious to hear from you on the life sciences side. But we're looking for design partners across many domains, language, anyone who's customizing language models or trying to push the frontier of code or reasoning models is really interesting to us. And then also interested in the frontier of modeling. There's a lot of models that work in, like, pixel space, as we call it. So if you're doing world models, video models, even robotics, where there's not a very clean natural language interface to interact with, I think we think that Interp can really help and are looking for a few partners in that space.Shawn Wang [00:52:43]: Just because you mentioned the keyword

Editor's note: Welcome to our new AI for Science pod, with your new hosts RJ and Brandon! See the writeup on Latent.Space (https://Latent.Space) for more details on why we're launching 2 new pods this year. RJ Honicky is a co-founder and CTO at MiraOmics (https://miraomics.bio/), building AI models and services for single cell, spatial transcriptomics and pathology slide analysis. Brandon Anderson builds AI systems for RNA drug discovery at Atomic AI (https://atomic.ai). Anything said on this podcast is his personal take — not Atomic's.—From building molecular dynamics simulations at the University of Washington to red-teaming GPT-4 for chemistry applications and co-founding Future House (a focused research organization) and Edison Scientific (a venture-backed startup automating science at scale)—Andrew White has spent the last five years living through the full arc of AI's transformation of scientific discovery, from ChemCrow (the first Chemistry LLM agent) triggering White House briefings and three-letter agency meetings, to shipping Kosmos, an end-to-end autonomous research system that generates hypotheses, runs experiments, analyzes data, and updates its world model to accelerate the scientific method itself.* The ChemCrow story: GPT-4 + React + cloud lab automation, released March 2023, set off a storm of anxiety about AI-accelerated bioweapons/chemical weapons, led to a White House briefing (Jake Sullivan presented the paper to the president in a 30-minute block), and meetings with three-letter agencies asking “how does this change breakout time for nuclear weapons research?”* Why scientific taste is the frontier: RLHF on hypotheses didn't work (humans pay attention to tone, actionability, and specific facts, not “if this hypothesis is true/false, how does it change the world?”), so they shifted to end-to-end feedback loops where humans click/download discoveries and that signal rolls up to hypothesis quality* Cosmos: the full scientific agent with a world model (distilled memory system, like a Git repo for scientific knowledge) that iterates on hypotheses via literature search, data analysis, and experiment design—built by Ludo after weeks of failed attempts, the breakthrough was putting data analysis in the loop (literature alone didn't work)* Why molecular dynamics and DFT are overrated: “MD and DFT have consumed an enormous number of PhDs at the altar of beautiful simulation, but they don't model the world correctly—you simulate water at 330 Kelvin to get room temperature, you overfit to validation data with GGA/B3LYP functionals, and real catalysts (grain boundaries, dopants) are too complicated for DFT”* The AlphaFold vs. DE Shaw Research counterfactual: DE Shaw built custom silicon, taped out chips with MD algorithms burned in, ran MD at massive scale in a special room in Times Square, and David Shaw flew in by helicopter to present—Andrew thought protein folding would require special machines to fold one protein per day, then AlphaFold solved it in Google Colab on a desktop GPU* The E3 Zero reward hacking saga: trained a model to generate molecules with specific atom counts (verifiable reward), but it kept exploiting loopholes, then a Nature paper came out that year proving six-nitrogen compounds are possible under extreme conditions, then it started adding nitrogen gas (purchasable, doesn't participate in reactions), then acid-base chemistry to move one atom, and Andrew ended up “building a ridiculous catalog of purchasable compounds in a Bloom filter” to close the loopAndrew White* FutureHouse: http://futurehouse.org/* Edison Scientific: http://edisonscientific.com/* X: https://x.com/andrewwhite01* Cosmos paper: https://futurediscovery.org/cosmosFull Video EpisodeTimestamps00:00:00 Introduction: Andrew White on Automating Science with Future House and Edison Scientific00:02:22 The Academic to Startup Journey: Red Teaming GPT-4 and the ChemCrow Paper00:11:35 Future House Origins: The FRO Model and Mission to Automate Science00:12:32 Resigning Tenure: Why Leave Academia for AI Science00:15:54 What Does ‘Automating Science' Actually Mean?00:17:30 The Lab-in-the-Loop Bottleneck: Why Intelligence Isn't Enough00:18:39 Scientific Taste and Human Preferences: The 52% Agreement Problem00:20:05 Paper QA, Robin, and the Road to Cosmos00:21:57 World Models as Scientific Memory: The GitHub Analogy00:40:20 The Bitter Lesson for Biology: Why Molecular Dynamics and DFT Are Overrated00:43:22 AlphaFold's Shock: When First Principles Lost to Machine Learning00:46:25 Enumeration and Filtration: How AI Scientists Generate Hypotheses00:48:15 CBRN Safety and Dual-Use AI: Lessons from Red Teaming01:00:40 The Future of Chemistry is Language: Multimodal Debate01:08:15 Ether Zero: The Hilarious Reward Hacking Adventures01:10:12 Will Scientists Be Displaced? Jevons Paradox and Infinite Discovery01:13:46 Cosmos in Practice: Open Access and Enterprise Partnerships Get full access to Latent.Space at www.latent.space/subscribe

From shipping Gemini Deep Think and IMO Gold to launching the Reasoning and AGI team in Singapore, Yi Tay has spent the last 18 months living through the full arc of Google DeepMind's pivot from architecture research to RL-driven reasoning—watching his team go from a dozen researchers to 300+, training models that solve International Math Olympiad problems in a live competition, and building the infrastructure to scale deep thinking across every domain, and driving Gemini to the top of the leaderboards across every category. Yi Returns to dig into the inside story of the IMO effort and more!We discuss:* Yi's path: Brain → Reka → Google DeepMind → Reasoning and AGI team Singapore, leading model training for Gemini Deep Think and IMO Gold* The IMO Gold story: four co-captains (Yi in Singapore, Jonathan in London, Jordan in Mountain View, and Tong leading the overall effort), training the checkpoint in ~1 week, live competition in Australia with professors punching in problems as they came out, and the tension of not knowing if they'd hit Gold until the human scores came in (because the Gold threshold is a percentile, not a fixed number)* Why they threw away AlphaProof: “If one model can't do it, can we get to AGI?” The decision to abandon symbolic systems and bet on end-to-end Gemini with RL was bold and non-consensus* On-policy vs. off-policy RL: off-policy is imitation learning (copying someone else's trajectory), on-policy is the model generating its own outputs, getting rewarded, and training on its own experience—”humans learn by making mistakes, not by copying”* Why self-consistency and parallel thinking are fundamental: sampling multiple times, majority voting, LM judges, and internal verification are all forms of self-consistency that unlock reasoning beyond single-shot inference* The data efficiency frontier: humans learn from 8 orders of magnitude less data than models, so where's the bug? Is it the architecture, the learning algorithm, backprop, off-policyness, or something else?* Three schools of thought on world models: (1) Genie/spatial intelligence (video-based world models), (2) Yann LeCun's JEPA + FAIR's code world models (modeling internal execution state), (3) the amorphous “resolution of possible worlds” paradigm (curve-fitting to find the world model that best explains the data)* Why AI coding crossed the threshold: Yi now runs a job, gets a bug, pastes it into Gemini, and relaunches without even reading the fix—”the model is better than me at this”* The Pokémon benchmark: can models complete Pokédex by searching the web, synthesizing guides, and applying knowledge in a visual game state? “Efficient search of novel idea space is interesting, but we're not even at the point where models can consistently apply knowledge they look up”* DSI and generative retrieval: re-imagining search as predicting document identifiers with semantic tokens, now deployed at YouTube (symmetric IDs for RecSys) and Spotify* Why RecSys and IR feel like a different universe: “modeling dynamics are strange, like gravity is different—you hit the shuttlecock and hear glass shatter, cause and effect are too far apart”* The closed lab advantage is increasing: the gap between frontier labs and open source is growing because ideas compound over time, and researchers keep finding new tricks that play well with everything built before* Why ideas still matter: “the last five years weren't just blind scaling—transformers, pre-training, RL, self-consistency, all had to play well together to get us here”* Gemini Singapore: hiring for RL and reasoning researchers, looking for track record in RL or exceptional achievement in coding competitions, and building a small, talent-dense team close to the frontier—Yi Tay* Google DeepMind: https://deepmind.google* X: https://x.com/YiTayMLFull Video EpisodeTimestamps00:00:00 Introduction: Returning to Google DeepMind and the Singapore AGI Team00:04:52 The Philosophy of On-Policy RL: Learning from Your Own Mistakes00:12:00 IMO Gold Medal: The Journey from AlphaProof to End-to-End Gemini00:21:33 Training IMO Cat: Four Captains Across Three Time Zones00:26:19 Pokemon and Long-Horizon Reasoning: Beyond Academic Benchmarks00:36:29 AI Coding Assistants: From Lazy to Actually Useful00:32:59 Reasoning, Chain of Thought, and Latent Thinking00:44:46 Is Attention All You Need? Architecture, Learning, and the Local Minima00:55:04 Data Efficiency and World Models: The Next Frontier01:08:12 DSI and Generative Retrieval: Reimagining Search with Semantic IDs01:17:59 Building GDM Singapore: Geography, Talent, and the Symposium01:24:18 Hiring Philosophy: High Stats, Research Taste, and Student Budgets01:28:49 Health, HRV, and Research Performance: The 23kg Journey Get full access to Latent.Space at www.latent.space/subscribe

From shipping Gemini Deep Think and IMO Gold to launching the Reasoning and AGI team in Singapore, Yi Tay has spent the last 18 months living through the full arc of Google DeepMind's pivot from architecture research to RL-driven reasoning—watching his team go from a dozen researchers to 300+, training models that solve International Math Olympiad problems in a live competition, and building the infrastructure to scale deep thinking across every domain, and driving Gemini to the top of the leaderboards across every category. Yi Returns to dig into the inside story of the IMO effort and more! We discuss: Yi's path: Brain → Reka → Google DeepMind → Reasoning and AGI team Singapore, leading model training for Gemini Deep Think and IMO Gold The IMO Gold story: four co-captains (Yi in Singapore, Jonathan in London, Jordan in Mountain View, and Tong leading the overall effort), training the checkpoint in ~1 week, live competition in Australia with professors punching in problems as they came out, and the tension of not knowing if they'd hit Gold until the human scores came in (because the Gold threshold is a percentile, not a fixed number) Why they threw away AlphaProof: "If one model can't do it, can we get to AGI?" The decision to abandon symbolic systems and bet on end-to-end Gemini with RL was bold and non-consensus On-policy vs. off-policy RL: off-policy is imitation learning (copying someone else's trajectory), on-policy is the model generating its own outputs, getting rewarded, and training on its own experience—"humans learn by making mistakes, not by copying" Why self-consistency and parallel thinking are fundamental: sampling multiple times, majority voting, LM judges, and internal verification are all forms of self-consistency that unlock reasoning beyond single-shot inference The data efficiency frontier: humans learn from 8 orders of magnitude less data than models, so where's the bug? Is it the architecture, the learning algorithm, backprop, off-policyness, or something else? Three schools of thought on world models: (1) Genie/spatial intelligence (video-based world models), (2) Yann LeCun's JEPA + FAIR's code world models (modeling internal execution state), (3) the amorphous "resolution of possible worlds" paradigm (curve-fitting to find the world model that best explains the data) Why AI coding crossed the threshold: Yi now runs a job, gets a bug, pastes it into Gemini, and relaunches without even reading the fix—"the model is better than me at this" The Pokémon benchmark: can models complete Pokédex by searching the web, synthesizing guides, and applying knowledge in a visual game state? "Efficient search of novel idea space is interesting, but we're not even at the point where models can consistently apply knowledge they look up" DSI and generative retrieval: re-imagining search as predicting document identifiers with semantic tokens, now deployed at YouTube (symmetric IDs for RecSys) and Spotify Why RecSys and IR feel like a different universe: "modeling dynamics are strange, like gravity is different—you hit the shuttlecock and hear glass shatter, cause and effect are too far apart" The closed lab advantage is increasing: the gap between frontier labs and open source is growing because ideas compound over time, and researchers keep finding new tricks that play well with everything built before Why ideas still matter: "the last five years weren't just blind scaling—transformers, pre-training, RL, self-consistency, all had to play well together to get us here" Gemini Singapore: hiring for RL and reasoning researchers, looking for track record in RL or exceptional achievement in coding competitions, and building a small, talent-dense team close to the frontier — Yi Tay Google DeepMind: https://deepmind.google X: https://x.com/YiTayML Chapters 00:00:00 Introduction: Returning to Google DeepMind and the Singapore AGI Team 00:04:52 The Philosophy of On-Policy RL: Learning from Your Own Mistakes 00:12:00 IMO Gold Medal: The Journey from AlphaProof to End-to-End Gemini 00:21:33 Training IMO Cat: Four Captains Across Three Time Zones 00:26:19 Pokemon and Long-Horizon Reasoning: Beyond Academic Benchmarks 00:36:29 AI Coding Assistants: From Lazy to Actually Useful 00:32:59 Reasoning, Chain of Thought, and Latent Thinking 00:44:46 Is Attention All You Need? Architecture, Learning, and the Local Minima 00:55:04 Data Efficiency and World Models: The Next Frontier 01:08:12 DSI and Generative Retrieval: Reimagining Search with Semantic IDs 01:17:59 Building GDM Singapore: Geography, Talent, and the Symposium 01:24:18 Hiring Philosophy: High Stats, Research Taste, and Student Budgets 01:28:49 Health, HRV, and Research Performance: The 23kg Journey

From building internal AI labs to becoming CTO of Brex, James Reggio has helped lead one of the most disciplined AI transformations inside a real financial institution where compliance, auditability, and customer trust actually matter.We sat down with Reggio to unpack Brex's three-pillar AI strategy (corporate, operational, and product AI) [https://www.brex.com/journal/brex-ai-native-operations], how SOP-driven agents beat overengineered RL in ops, why Brex lets employees “build their own AI stack” instead of picking winners [https://www.conductorone.com/customers/brex/], and how a small, founder-heavy AI team is shipping production agents to 40,000+ companies. Reggio also goes deep on Brex's multi-agent “network” architecture, evals for multi-turn systems, agentic coding's second-order effects on codebase understanding, and why the future of finance software looks less like dashboards and more like executive assistants coordinating specialist agents behind the scenes.We discuss:* Brex's three-pillar AI strategy: corporate AI for 10x employee workflows, operational AI for cost and compliance leverage, and product AI that lets customers justify Brex as part of their AI strategy to the board* Why SOP-driven agents beat overengineered RL in finance ops, and how breaking work into auditable, repeatable steps unlocked faster automation in KYC, underwriting, fraud, and disputes* Building an internal AI platform early: LLM gateways, prompt/version management, evals, cost observability, and why platform work quietly became the force multiplier behind everything else* Multi-agent “networks” vs single-agent tools: why Brex's EA-style assistant coordinates specialist agents (policy, travel, reimbursements) through multi-turn conversations instead of one-shot tool calls* The audit agent pattern: separating detection, judgment, and follow-up into different agents to reduce false negatives without overwhelming finance teams* Centralized AI teams without resentment: how Brex avoided “AI envy” by tying work to business impact and letting anyone transfer in if they cared deeply enough* Letting employees build their own AI stack: ChatGPT vs Claude vs Gemini, Cursor vs Windsurf, and why Brex refuses to pick winners in fast-moving tool races* Measuring adoption without vanity metrics: why “% of code written by AI” is the wrong KPI and what second-order effects (slop, drift, code ownership) actually matter* Evals in the real world: regression tests from ops QA, LLM-as-judge for multi-turn agents, and why integration-style evals break faster than you expect* Teaching AI fluency at scale: the user → advocate → builder → native framework, ops-led training, spot bonuses, and avoiding fear-based adoption* Re-interviewing the entire engineering org: using agentic coding interviews internally to force hands-on skill upgrades without formal performance scoring* Headcount in the age of agents: why Brex grew the business without growing engineering, and why AI amplifies bad architecture as fast as good decisions* The future of finance software: why dashboards fade, assistants take over, and agent-to-agent collaboration becomes the real UI—James Reggio* X: https://x.com/jamesreggio* LinkedIn: https://www.linkedin.com/in/jamesreggio/Where to find Latent Space* X: https://x.com/latentspacepodFull Video EpisodeTimestamps00:00:00 Introduction00:01:24 From Mobile Engineer to CTO: The Founder's Path00:03:00 Quitters Welcome: Building a Founder-Friendly Culture00:05:13 The AI Team Structure: 10-Person Startup Within Brex00:11:55 Building the Brex Agent Platform: Multi-Agent Networks00:13:45 Tech Stack Decisions: TypeScript, Mastra, and MCP00:24:32 Operational AI: Automating Underwriting, KYC, and Fraud00:16:40 The Brex Assistant: Executive Assistant for Every Employee00:40:26 Evaluation Strategy: From Simple SOPs to Multi-Turn Evals00:37:11 Agentic Coding Adoption: Cursor, Windsurf, and the Engineering Interview00:58:51 AI Fluency Levels: From User to Native01:09:14 The Audit Agent Network: Finance Team Agents in Action01:03:33 The Future of Engineering Headcount and AI Leverage Get full access to Latent.Space at www.latent.space/subscribe

From building internal AI labs to becoming CTO of Brex, James Reggio has helped lead one of the most disciplined AI transformations inside a real financial institution where compliance, auditability, and customer trust actually matter. We sat down with Reggio to unpack Brex's three-pillar AI strategy (corporate, operational, and product AI) [https://www.brex.com/journal/brex-ai-native-operations], how SOP-driven agents beat overengineered RL in ops, why Brex lets employees “build their own AI stack” instead of picking winners [https://www.conductorone.com/customers/brex/], and how a small, founder-heavy AI team is shipping production agents to 40,000+ companies. Reggio also goes deep on Brex's multi-agent “network” architecture, evals for multi-turn systems, agentic coding's second-order effects on codebase understanding, and why the future of finance software looks less like dashboards and more like executive assistants coordinating specialist agents behind the scenes. We discuss: Brex's three-pillar AI strategy: corporate AI for 10x employee workflows, operational AI for cost and compliance leverage, and product AI that lets customers justify Brex as part of their AI strategy to the board Why SOP-driven agents beat overengineered RL in finance ops, and how breaking work into auditable, repeatable steps unlocked faster automation in KYC, underwriting, fraud, and disputes Building an internal AI platform early: LLM gateways, prompt/version management, evals, cost observability, and why platform work quietly became the force multiplier behind everything else Multi-agent “networks” vs single-agent tools: why Brex's EA-style assistant coordinates specialist agents (policy, travel, reimbursements) through multi-turn conversations instead of one-shot tool calls The audit agent pattern: separating detection, judgment, and follow-up into different agents to reduce false negatives without overwhelming finance teams Centralized AI teams without resentment: how Brex avoided “AI envy” by tying work to business impact and letting anyone transfer in if they cared deeply enough Letting employees build their own AI stack: ChatGPT vs Claude vs Gemini, Cursor vs Windsurf, and why Brex refuses to pick winners in fast-moving tool races Measuring adoption without vanity metrics: why “% of code written by AI” is the wrong KPI and what second-order effects (slop, drift, code ownership) actually matter Evals in the real world: regression tests from ops QA, LLM-as-judge for multi-turn agents, and why integration-style evals break faster than you expect Teaching AI fluency at scale: the user → advocate → builder → native framework, ops-led training, spot bonuses, and avoiding fear-based adoption Re-interviewing the entire engineering org: using agentic coding interviews internally to force hands-on skill upgrades without formal performance scoring Headcount in the age of agents: why Brex grew the business without growing engineering, and why AI amplifies bad architecture as fast as good decisions The future of finance software: why dashboards fade, assistants take over, and agent-to-agent collaboration becomes the real UI — James Reggio X: https://x.com/jamesreggio LinkedIn: https://www.linkedin.com/in/jamesreggio/ Where to find Latent Space X: https://x.com/latentspacepod Substack: https://www.latent.space/ Chapters 00:00:00 Introduction 00:01:24 From Mobile Engineer to CTO: The Founder's Path 00:03:00 Quitters Welcome: Building a Founder-Friendly Culture 00:05:13 The AI Team Structure: 10-Person Startup Within Brex 00:11:55 Building the Brex Agent Platform: Multi-Agent Networks 00:13:45 Tech Stack Decisions: TypeScript, Mastra, and MCP 00:24:32 Operational AI: Automating Underwriting, KYC, and Fraud 00:16:40 The Brex Assistant: Executive Assistant for Every Employee 00:40:26 Evaluation Strategy: From Simple SOPs to Multi-Turn Evals 00:37:11 Agentic Coding Adoption: Cursor, Windsurf, and the Engineering Interview 00:58:51 AI Fluency Levels: From User to Native 01:09:14 The Audit Agent Network: Finance Team Agents in Action 01:03:33 The Future of Engineering Headcount and AI Leverage

don't miss George's AIE talk: https://www.youtube.com/watch?v=sRpqPgKeXNk —- From launching a side project in a Sydney basement to becoming the independent gold standard for AI benchmarking—trusted by developers, enterprises, and every major lab to navigate the exploding landscape of models, providers, and capabilities—George Cameron and Micah Hill-Smith have spent two years building Artificial Analysis into the platform that answers the questions no one else will: Which model is actually best for your use case? What are the real speed-cost trade-offs? And how open is "open" really? We discuss: The origin story: built as a side project in 2023 while Micah was building a legal AI assistant, launched publicly in January 2024, and went viral after Swyx's retweet Why they run evals themselves: labs prompt models differently, cherry-pick chain-of-thought examples (Google Gemini 1.0 Ultra used 32-shot prompts to beat GPT-4 on MMLU), and self-report inflated numbers The mystery shopper policy: they register accounts not on their own domain and run intelligence + performance benchmarks incognito to prevent labs from serving different models on private endpoints How they make money: enterprise benchmarking insights subscription (standardized reports on model deployment, serverless vs. managed vs. leasing chips) and private custom benchmarking for AI companies (no one pays to be on the public leaderboard) The Intelligence Index (V3): synthesizes 10 eval datasets (MMLU, GPQA, agentic benchmarks, long-context reasoning) into a single score, with 95% confidence intervals via repeated runs Omissions Index (hallucination rate): scores models from -100 to +100 (penalizing incorrect answers, rewarding "I don't know"), and Claude models lead with the lowest hallucination rates despite not always being the smartest GDP Val AA: their version of OpenAI's GDP-bench (44 white-collar tasks with spreadsheets, PDFs, PowerPoints), run through their Stirrup agent harness (up to 100 turns, code execution, web search, file system), graded by Gemini 3 Pro as an LLM judge (tested extensively, no self-preference bias) The Openness Index: scores models 0-18 on transparency of pre-training data, post-training data, methodology, training code, and licensing (AI2 OLMo 2 leads, followed by Nous Hermes and NVIDIA Nemotron) The smiling curve of AI costs: GPT-4-level intelligence is 100-1000x cheaper than at launch (thanks to smaller models like Amazon Nova), but frontier reasoning models in agentic workflows cost more than ever (sparsity, long context, multi-turn agents) Why sparsity might go way lower than 5%: GPT-4.5 is ~5% active, Gemini models might be ~3%, and Omissions Index accuracy correlates with total parameters (not active), suggesting massive sparse models are the future Token efficiency vs. turn efficiency: GPT-5 costs more per token but solves Tau-bench in fewer turns (cheaper overall), and models are getting better at using more tokens only when needed (5.1 Codex has tighter token distributions) V4 of the Intelligence Index coming soon: adding GDP Val AA, Critical Point, hallucination rate, and dropping some saturated benchmarks (human-eval-style coding is now trivial for small models) — Artificial Analysis Website: https://artificialanalysis.ai (https://artificialanalysis.ai ("https://artificialanalysis.ai")) George Cameron on X: https://x.com/grmcameron (https://x.com/grmcameron ("https://x.com/grmcameron")) Micah Hill-Smith on X: https://x.com/_micah_h (https://x.com/_micah_h ("https://x.com/_micah_h")) Chapters 00:00:00 Introduction: Full Circle Moment and Artificial Analysis Origins 00:01:08 Business Model: Independence and Revenue Streams 00:04:00 The Origin Story: From Legal AI to Benchmarking 00:07:00 Early Challenges: Cost, Methodology, and Independence 00:16:13 AI Grant and Moving to San Francisco 00:18:58 Evolution of the Intelligence Index: V1 to V3 00:27:55 New Benchmarks: Hallucination Rate and Omissions Index 00:33:19 Critical Point and Frontier Physics Problems 00:35:56 GDPVAL AA: Agentic Evaluation and Stirrup Harness 00:51:47 The Openness Index: Measuring Model Transparency 00:57:57 The Smiling Curve: Cost of Intelligence Paradox 01:04:00 Hardware Efficiency and Sparsity Trends 01:07:43 Reasoning vs Non-Reasoning: Token Efficiency Matters 01:10:47 Multimodal Benchmarking and Community Requests 01:14:50 Looking Ahead: V4 Intelligence Index and Beyond

Happy New Year! You may have noticed that in 2025 we had moved toward YouTube as our primary podcasting platform. As we'll explain in the next State of Latent Space post, we'll be doubling down on Substack again and improving the experience for the over 100,000 of you who look out for our emails and website updates!We first mentioned Artificial Analysis in 2024, when it was still a side project in a Sydney basement. They then were one of the few Nat Friedman and Daniel Gross' AIGrant companies to raise a full seed round from them and have now become the independent gold standard for AI benchmarking—trusted by developers, enterprises, and every major lab to navigate the exploding landscape of models, providers, and capabilities.We have chatted with both Clementine Fourrier of HuggingFace's OpenLLM Leaderboard and (the freshly valued at $1.7B) Anastasios Angelopoulos of LMArena on their approaches to LLM evals and trendspotting, but Artificial Analysis have staked out an enduring and important place in the toolkit of the modern AI Engineer by doing the best job of independently running the most comprehensive set of evals across the widest range of open and closed models, and charting their progress for broad industry analyst use.George Cameron and Micah-Hill Smith have spent two years building Artificial Analysis into the platform that answers the questions no one else will: Which model is actually best for your use case? What are the real speed-cost trade-offs? And how open is “open” really?We discuss:* The origin story: built as a side project in 2023 while Micah was building a legal AI assistant, launched publicly in January 2024, and went viral after Swyx's retweet* Why they run evals themselves: labs prompt models differently, cherry-pick chain-of-thought examples (Google Gemini 1.0 Ultra used 32-shot prompts to beat GPT-4 on MMLU), and self-report inflated numbers* The mystery shopper policy: they register accounts not on their own domain and run intelligence + performance benchmarks incognito to prevent labs from serving different models on private endpoints* How they make money: enterprise benchmarking insights subscription (standardized reports on model deployment, serverless vs. managed vs. leasing chips) and private custom benchmarking for AI companies (no one pays to be on the public leaderboard)* The Intelligence Index (V3): synthesizes 10 eval datasets (MMLU, GPQA, agentic benchmarks, long-context reasoning) into a single score, with 95% confidence intervals via repeated runs* Omissions Index (hallucination rate): scores models from -100 to +100 (penalizing incorrect answers, rewarding ”I don't know”), and Claude models lead with the lowest hallucination rates despite not always being the smartest* GDP Val AA: their version of OpenAI's GDP-bench (44 white-collar tasks with spreadsheets, PDFs, PowerPoints), run through their Stirrup agent harness (up to 100 turns, code execution, web search, file system), graded by Gemini 3 Pro as an LLM judge (tested extensively, no self-preference bias)* The Openness Index: scores models 0-18 on transparency of pre-training data, post-training data, methodology, training code, and licensing (AI2 OLMo 2 leads, followed by Nous Hermes and NVIDIA Nemotron)* The smiling curve of AI costs: GPT-4-level intelligence is 100-1000x cheaper than at launch (thanks to smaller models like Amazon Nova), but frontier reasoning models in agentic workflows cost more than ever (sparsity, long context, multi-turn agents)* Why sparsity might go way lower than 5%: GPT-4.5 is ~5% active, Gemini models might be ~3%, and Omissions Index accuracy correlates with total parameters (not active), suggesting massive sparse models are the future* Token efficiency vs. turn efficiency: GPT-5 costs more per token but solves Tau-bench in fewer turns (cheaper overall), and models are getting better at using more tokens only when needed (5.1 Codex has tighter token distributions)* V4 of the Intelligence Index coming soon: adding GDP Val AA, Critical Point, hallucination rate, and dropping some saturated benchmarks (human-eval-style coding is now trivial for small models)Links to Artificial Analysis* Website: https://artificialanalysis.ai* George Cameron on X: https://x.com/georgecameron* Micah-Hill Smith on X: https://x.com/micahhsmithFull Episode on YouTubeTimestamps* 00:00 Introduction: Full Circle Moment and Artificial Analysis Origins* 01:19 Business Model: Independence and Revenue Streams* 04:33 Origin Story: From Legal AI to Benchmarking Need* 16:22 AI Grant and Moving to San Francisco* 19:21 Intelligence Index Evolution: From V1 to V3* 11:47 Benchmarking Challenges: Variance, Contamination, and Methodology* 13:52 Mystery Shopper Policy and Maintaining Independence* 28:01 New Benchmarks: Omissions Index for Hallucination Detection* 33:36 Critical Point: Hard Physics Problems and Research-Level Reasoning* 23:01 GDP Val AA: Agentic Benchmark for Real Work Tasks* 50:19 Stirrup Agent Harness: Open Source Agentic Framework* 52:43 Openness Index: Measuring Model Transparency Beyond Licenses* 58:25 The Smiling Curve: Cost Falling While Spend Rising* 1:02:32 Hardware Efficiency: Blackwell Gains and Sparsity Limits* 1:06:23 Reasoning Models and Token Efficiency: The Spectrum Emerges* 1:11:00 Multimodal Benchmarking: Image, Video, and Speech Arenas* 1:15:05 Looking Ahead: Intelligence Index V4 and Future Directions* 1:16:50 Closing: The Insatiable Demand for IntelligenceTranscriptMicah [00:00:06]: This is kind of a full circle moment for us in a way, because the first time artificial analysis got mentioned on a podcast was you and Alessio on Latent Space. Amazing.swyx [00:00:17]: Which was January 2024. I don't even remember doing that, but yeah, it was very influential to me. Yeah, I'm looking at AI News for Jan 17, or Jan 16, 2024. I said, this gem of a models and host comparison site was just launched. And then I put in a few screenshots, and I said, it's an independent third party. It clearly outlines the quality versus throughput trade-off, and it breaks out by model and hosting provider. I did give you s**t for missing fireworks, and how do you have a model benchmarking thing without fireworks? But you had together, you had perplexity, and I think we just started chatting there. Welcome, George and Micah, to Latent Space. I've been following your progress. Congrats on... It's been an amazing year. You guys have really come together to be the presumptive new gardener of AI, right? Which is something that...George [00:01:09]: Yeah, but you can't pay us for better results.swyx [00:01:12]: Yes, exactly.George [00:01:13]: Very important.Micah [00:01:14]: Start off with a spicy take.swyx [00:01:18]: Okay, how do I pay you?Micah [00:01:20]: Let's get right into that.swyx [00:01:21]: How do you make money?Micah [00:01:24]: Well, very happy to talk about that. So it's been a big journey the last couple of years. Artificial analysis is going to be two years old in January 2026. Which is pretty soon now. We first run the website for free, obviously, and give away a ton of data to help developers and companies navigate AI and make decisions about models, providers, technologies across the AI stack for building stuff. We're very committed to doing that and tend to keep doing that. We have, along the way, built a business that is working out pretty sustainably. We've got just over 20 people now and two main customer groups. So we want to be... We want to be who enterprise look to for data and insights on AI, so we want to help them with their decisions about models and technologies for building stuff. And then on the other side, we do private benchmarking for companies throughout the AI stack who build AI stuff. So no one pays to be on the website. We've been very clear about that from the very start because there's no use doing what we do unless it's independent AI benchmarking. Yeah. But turns out a bunch of our stuff can be pretty useful to companies building AI stuff.swyx [00:02:38]: And is it like, I am a Fortune 500, I need advisors on objective analysis, and I call you guys and you pull up a custom report for me, you come into my office and give me a workshop? What kind of engagement is that?George [00:02:53]: So we have a benchmarking and insight subscription, which looks like standardized reports that cover key topics or key challenges enterprises face when looking to understand AI and choose between all the technologies. And so, for instance, one of the report is a model deployment report, how to think about choosing between serverless inference, managed deployment solutions, or leasing chips. And running inference yourself is an example kind of decision that big enterprises face, and it's hard to reason through, like this AI stuff is really new to everybody. And so we try and help with our reports and insight subscription. Companies navigate that. We also do custom private benchmarking. And so that's very different from the public benchmarking that we publicize, and there's no commercial model around that. For private benchmarking, we'll at times create benchmarks, run benchmarks to specs that enterprises want. And we'll also do that sometimes for AI companies who have built things, and we help them understand what they've built with private benchmarking. Yeah. So that's a piece mainly that we've developed through trying to support everybody publicly with our public benchmarks. Yeah.swyx [00:04:09]: Let's talk about TechStack behind that. But okay, I'm going to rewind all the way to when you guys started this project. You were all the way in Sydney? Yeah. Well, Sydney, Australia for me.Micah [00:04:19]: George was an SF, but he's Australian, but he moved here already. Yeah.swyx [00:04:22]: And I remember I had the Zoom call with you. What was the impetus for starting artificial analysis in the first place? You know, you started with public benchmarks. And so let's start there. We'll go to the private benchmark. Yeah.George [00:04:33]: Why don't we even go back a little bit to like why we, you know, thought that it was needed? Yeah.Micah [00:04:40]: The story kind of begins like in 2022, 2023, like both George and I have been into AI stuff for quite a while. In 2023 specifically, I was trying to build a legal AI research assistant. So it actually worked pretty well for its era, I would say. Yeah. Yeah. So I was finding that the more you go into building something using LLMs, the more each bit of what you're doing ends up being a benchmarking problem. So had like this multistage algorithm thing, trying to figure out what the minimum viable model for each bit was, trying to optimize every bit of it as you build that out, right? Like you're trying to think about accuracy, a bunch of other metrics and performance and cost. And mostly just no one was doing anything to independently evaluate all the models. And certainly not to look at the trade-offs for speed and cost. So we basically set out just to build a thing that developers could look at to see the trade-offs between all of those things measured independently across all the models and providers. Honestly, it was probably meant to be a side project when we first started doing it.swyx [00:05:49]: Like we didn't like get together and say like, Hey, like we're going to stop working on all this stuff. I'm like, this is going to be our main thing. When I first called you, I think you hadn't decided on starting a company yet.Micah [00:05:58]: That's actually true. I don't even think we'd pause like, like George had an acquittance job. I didn't quit working on my legal AI thing. Like it was genuinely a side project.George [00:06:05]: We built it because we needed it as people building in the space and thought, Oh, other people might find it useful too. So we'll buy domain and link it to the Vercel deployment that we had and tweet about it. And, but very quickly it started getting attention. Thank you, Swyx for, I think doing an initial retweet and spotlighting it there. This project that we released. And then very quickly though, it was useful to others, but very quickly it became more useful as the number of models released accelerated. We had Mixtrel 8x7B and it was a key. That's a fun one. Yeah. Like a open source model that really changed the landscape and opened up people's eyes to other serverless inference providers and thinking about speed, thinking about cost. And so that was a key. And so it became more useful quite quickly. Yeah.swyx [00:07:02]: What I love talking to people like you who sit across the ecosystem is, well, I have theories about what people want, but you have data and that's obviously more relevant. But I want to stay on the origin story a little bit more. When you started out, I would say, I think the status quo at the time was every paper would come out and they would report their numbers versus competitor numbers. And that's basically it. And I remember I did the legwork. I think everyone has some knowledge. I think there's some version of Excel sheet or a Google sheet where you just like copy and paste the numbers from every paper and just post it up there. And then sometimes they don't line up because they're independently run. And so your numbers are going to look better than... Your reproductions of other people's numbers are going to look worse because you don't hold their models correctly or whatever the excuse is. I think then Stanford Helm, Percy Liang's project would also have some of these numbers. And I don't know if there's any other source that you can cite. The way that if I were to start artificial analysis at the same time you guys started, I would have used the Luther AI's eval framework harness. Yup.Micah [00:08:06]: Yup. That was some cool stuff. At the end of the day, running these evals, it's like if it's a simple Q&A eval, all you're doing is asking a list of questions and checking if the answers are right, which shouldn't be that crazy. But it turns out there are an enormous number of things that you've got control for. And I mean, back when we started the website. Yeah. Yeah. Like one of the reasons why we realized that we had to run the evals ourselves and couldn't just take rules from the labs was just that they would all prompt the models differently. And when you're competing over a few points, then you can pretty easily get- You can put the answer into the model. Yeah. That in the extreme. And like you get crazy cases like back when I'm Googled a Gemini 1.0 Ultra and needed a number that would say it was better than GPT-4 and like constructed, I think never published like chain of thought examples. 32 of them in every topic in MLU to run it, to get the score, like there are so many things that you- They never shipped Ultra, right? That's the one that never made it up. Not widely. Yeah. Yeah. Yeah. I mean, I'm sure it existed, but yeah. So we were pretty sure that we needed to run them ourselves and just run them in the same way across all the models. Yeah. And we were, we also did certain from the start that you couldn't look at those in isolation. You needed to look at them alongside the cost and performance stuff. Yeah.swyx [00:09:24]: Okay. A couple of technical questions. I mean, so obviously I also thought about this and I didn't do it because of cost. Yep. Did you not worry about costs? Were you funded already? Clearly not, but you know. No. Well, we definitely weren't at the start.Micah [00:09:36]: So like, I mean, we're paying for it personally at the start. There's a lot of money. Well, the numbers weren't nearly as bad a couple of years ago. So we certainly incurred some costs, but we were probably in the order of like hundreds of dollars of spend across all the benchmarking that we were doing. Yeah. So nothing. Yeah. It was like kind of fine. Yeah. Yeah. These days that's gone up an enormous amount for a bunch of reasons that we can talk about. But yeah, it wasn't that bad because you can also remember that like the number of models we were dealing with was hardly any and the complexity of the stuff that we wanted to do to evaluate them was a lot less. Like we were just asking some Q&A type questions and then one specific thing was for a lot of evals initially, we were just like sampling an answer. You know, like, what's the answer for this? Like, we didn't want to go into the answer directly without letting the models think. We weren't even doing chain of thought stuff initially. And that was the most useful way to get some results initially. Yeah.swyx [00:10:33]: And so for people who haven't done this work, literally parsing the responses is a whole thing, right? Like because sometimes the models, the models can answer any way they feel fit and sometimes they actually do have the right answer, but they just returned the wrong format and they will get a zero for that unless you work it into your parser. And that involves more work. And so, I mean, but there's an open question whether you should give it points for not following your instructions on the format.Micah [00:11:00]: It depends what you're looking at, right? Because you can, if you're trying to see whether or not it can solve a particular type of reasoning problem, and you don't want to test it on its ability to do answer formatting at the same time, then you might want to use an LLM as answer extractor approach to make sure that you get the answer out no matter how unanswered. But these days, it's mostly less of a problem. Like, if you instruct a model and give it examples of what the answers should look like, it can get the answers in your format, and then you can do, like, a simple regex.swyx [00:11:28]: Yeah, yeah. And then there's other questions around, I guess, sometimes if you have a multiple choice question, sometimes there's a bias towards the first answer, so you have to randomize the responses. All these nuances, like, once you dig into benchmarks, you're like, I don't know how anyone believes the numbers on all these things. It's so dark magic.Micah [00:11:47]: You've also got, like… You've got, like, the different degrees of variance in different benchmarks, right? Yeah. So, if you run four-question multi-choice on a modern reasoning model at the temperatures suggested by the labs for their own models, the variance that you can see on a four-question multi-choice eval is pretty enormous if you only do a single run of it and it has a small number of questions, especially. So, like, one of the things that we do is run an enormous number of all of our evals when we're developing new ones and doing upgrades to our intelligence index to bring in new things. Yeah. So, that we can dial in the right number of repeats so that we can get to the 95% confidence intervals that we're comfortable with so that when we pull that together, we can be confident in intelligence index to at least as tight as, like, a plus or minus one at a 95% confidence. Yeah.swyx [00:12:32]: And, again, that just adds a straight multiple to the cost. Oh, yeah. Yeah, yeah.George [00:12:37]: So, that's one of many reasons that cost has gone up a lot more than linearly over the last couple of years. We report a cost to run the artificial analysis. We report a cost to run the artificial analysis intelligence index on our website, and currently that's assuming one repeat in terms of how we report it because we want to reflect a bit about the weighting of the index. But our cost is actually a lot higher than what we report there because of the repeats.swyx [00:13:03]: Yeah, yeah, yeah. And probably this is true, but just checking, you don't have any special deals with the labs. They don't discount it. You just pay out of pocket or out of your sort of customer funds. Oh, there is a mix. So, the issue is that sometimes they may give you a special end point, which is… Ah, 100%.Micah [00:13:21]: Yeah, yeah, yeah. Exactly. So, we laser focus, like, on everything we do on having the best independent metrics and making sure that no one can manipulate them in any way. There are quite a lot of processes we've developed over the last couple of years to make that true for, like, the one you bring up, like, right here of the fact that if we're working with a lab, if they're giving us a private endpoint to evaluate a model, that it is totally possible. That what's sitting behind that black box is not the same as they serve on a public endpoint. We're very aware of that. We have what we call a mystery shopper policy. And so, and we're totally transparent with all the labs we work with about this, that we will register accounts not on our own domain and run both intelligence evals and performance benchmarks… Yeah, that's the job. …without them being able to identify it. And no one's ever had a problem with that. Because, like, a thing that turns out to actually be quite a good… …good factor in the industry is that they all want to believe that none of their competitors could manipulate what we're doing either.swyx [00:14:23]: That's true. I never thought about that. I've been in the database data industry prior, and there's a lot of shenanigans around benchmarking, right? So I'm just kind of going through the mental laundry list. Did I miss anything else in this category of shenanigans? Oh, potential shenanigans.Micah [00:14:36]: I mean, okay, the biggest one, like, that I'll bring up, like, is more of a conceptual one, actually, than, like, direct shenanigans. It's that the things that get measured become things that get targeted by labs that they're trying to build, right? Exactly. So that doesn't mean anything that we should really call shenanigans. Like, I'm not talking about training on test set. But if you know that you're going to be great at another particular thing, if you're a researcher, there are a whole bunch of things that you can do to try to get better at that thing that preferably are going to be helpful for a wide range of how actual users want to use the thing that you're building. But will not necessarily work. Will not necessarily do that. So, for instance, the models are exceptional now at answering competition maths problems. There is some relevance of that type of reasoning, that type of work, to, like, how we might use modern coding agents and stuff. But it's clearly not one for one. So the thing that we have to be aware of is that once an eval becomes the thing that everyone's looking at, scores can get better on it without there being a reflection of overall generalized intelligence of these models. Getting better. That has been true for the last couple of years. It'll be true for the next couple of years. There's no silver bullet to defeat that other than building new stuff to stay relevant and measure the capabilities that matter most to real users. Yeah.swyx [00:15:58]: And we'll cover some of the new stuff that you guys are building as well, which is cool. Like, you used to just run other people's evals, but now you're coming up with your own. And I think, obviously, that is a necessary path once you're at the frontier. You've exhausted all the existing evals. I think the next point in history that I have for you is AI Grant that you guys decided to join and move here. What was it like? I think you were in, like, batch two? Batch four. Batch four. Okay.Micah [00:16:26]: I mean, it was great. Nat and Daniel are obviously great. And it's a really cool group of companies that we were in AI Grant alongside. It was really great to get Nat and Daniel on board. Obviously, they've done a whole lot of great work in the space with a lot of leading companies and were extremely aligned. With the mission of what we were trying to do. Like, we're not quite typical of, like, a lot of the other AI startups that they've invested in.swyx [00:16:53]: And they were very much here for the mission of what we want to do. Did they say any advice that really affected you in some way or, like, were one of the events very impactful? That's an interesting question.Micah [00:17:03]: I mean, I remember fondly a bunch of the speakers who came and did fireside chats at AI Grant.swyx [00:17:09]: Which is also, like, a crazy list. Yeah.George [00:17:11]: Oh, totally. Yeah, yeah, yeah. There was something about, you know, speaking to Nat and Daniel about the challenges of working through a startup and just working through the questions that don't have, like, clear answers and how to work through those kind of methodically and just, like, work through the hard decisions. And they've been great mentors to us as we've built artificial analysis. Another benefit for us was that other companies in the batch and other companies in AI Grant are pushing the capabilities. Yeah. And I think that's a big part of what AI can do at this time. And so being in contact with them, making sure that artificial analysis is useful to them has been fantastic for supporting us in working out how should we build out artificial analysis to continue to being useful to those, like, you know, building on AI.swyx [00:17:59]: I think to some extent, I'm mixed opinion on that one because to some extent, your target audience is not people in AI Grants who are obviously at the frontier. Yeah. Do you disagree?Micah [00:18:09]: To some extent. To some extent. But then, so a lot of what the AI Grant companies are doing is taking capabilities coming out of the labs and trying to push the limits of what they can do across the entire stack for building great applications, which actually makes some of them pretty archetypical power users of artificial analysis. Some of the people with the strongest opinions about what we're doing well and what we're not doing well and what they want to see next from us. Yeah. Yeah. Because when you're building any kind of AI application now, chances are you're using a whole bunch of different models. You're maybe switching reasonably frequently for different models and different parts of your application to optimize what you're able to do with them at an accuracy level and to get better speed and cost characteristics. So for many of them, no, they're like not commercial customers of ours, like we don't charge for all our data on the website. Yeah. They are absolutely some of our power users.swyx [00:19:07]: So let's talk about just the evals as well. So you start out from the general like MMU and GPQA stuff. What's next? How do you sort of build up to the overall index? What was in V1 and how did you evolve it? Okay.Micah [00:19:22]: So first, just like background, like we're talking about the artificial analysis intelligence index, which is our synthesis metric that we pulled together currently from 10 different eval data sets to give what? We're pretty much the same as that. Pretty confident is the best single number to look at for how smart the models are. Obviously, it doesn't tell the whole story. That's why we published the whole website of all the charts to dive into every part of it and look at the trade-offs. But best single number. So right now, it's got a bunch of Q&A type data sets that have been very important to the industry, like a couple that you just mentioned. It's also got a couple of agentic data sets. It's got our own long context reasoning data set and some other use case focused stuff. As time goes on. The things that we're most interested in that are going to be important to the capabilities that are becoming more important for AI, what developers are caring about, are going to be first around agentic capabilities. So surprise, surprise. We're all loving our coding agents and how the model is going to perform like that and then do similar things for different types of work are really important to us. The linking to use cases to economically valuable use cases are extremely important to us. And then we've got some of the. Yeah. These things that the models still struggle with, like working really well over long contexts that are not going to go away as specific capabilities and use cases that we need to keep evaluating.swyx [00:20:46]: But I guess one thing I was driving was like the V1 versus the V2 and how bad it was over time.Micah [00:20:53]: Like how we've changed the index to where we are.swyx [00:20:55]: And I think that reflects on the change in the industry. Right. So that's a nice way to tell that story.Micah [00:21:00]: Well, V1 would be completely saturated right now. Almost every model coming out because doing things like writing the Python functions and human evil is now pretty trivial. It's easy to forget, actually, I think how much progress has been made in the last two years. Like we obviously play the game constantly of like the today's version versus last week's version and the week before and all of the small changes in the horse race between the current frontier and who has the best like smaller than 10B model like right now this week. Right. And that's very important to a lot of developers and people and especially in this particular city of San Francisco. But when you zoom out a couple of years ago, literally most of what we were doing to evaluate the models then would all be 100% solved by even pretty small models today. And that's been one of the key things, by the way, that's driven down the cost of intelligence at every tier of intelligence. We can talk about more in a bit. So V1, V2, V3, we made things harder. We covered a wider range of use cases. And we tried to get closer to things developers care about as opposed to like just the Q&A type stuff that MMLU and GPQA represented. Yeah.swyx [00:22:12]: I don't know if you have anything to add there. Or we could just go right into showing people the benchmark and like looking around and asking questions about it. Yeah.Micah [00:22:21]: Let's do it. Okay. This would be a pretty good way to chat about a few of the new things we've launched recently. Yeah.George [00:22:26]: And I think a little bit about the direction that we want to take it. And we want to push benchmarks. Currently, the intelligence index and evals focus a lot on kind of raw intelligence. But we kind of want to diversify how we think about intelligence. And we can talk about it. But kind of new evals that we've kind of built and partnered on focus on topics like hallucination. And we've got a lot of topics that I think are not covered by the current eval set that should be. And so we want to bring that forth. But before we get into that.swyx [00:23:01]: And so for listeners, just as a timestamp, right now, number one is Gemini 3 Pro High. Then followed by Cloud Opus at 70. Just 5.1 high. You don't have 5.2 yet. And Kimi K2 Thinking. Wow. Still hanging in there. So those are the top four. That will date this podcast quickly. Yeah. Yeah. I mean, I love it. I love it. No, no. 100%. Look back this time next year and go, how cute. Yep.George [00:23:25]: Totally. A quick view of that is, okay, there's a lot. I love it. I love this chart. Yeah.Micah [00:23:30]: This is such a favorite, right? Yeah. And almost every talk that George or I give at conferences and stuff, we always put this one up first to just talk about situating where we are in this moment in history. This, I think, is the visual version of what I was saying before about the zooming out and remembering how much progress there's been. If we go back to just over a year ago, before 01, before Cloud Sonnet 3.5, we didn't have reasoning models or coding agents as a thing. And the game was very, very different. If we go back even a little bit before then, we're in the era where, when you look at this chart, open AI was untouchable for well over a year. And, I mean, you would remember that time period well of there being very open questions about whether or not AI was going to be competitive, like full stop, whether or not open AI would just run away with it, whether we would have a few frontier labs and no one else would really be able to do anything other than consume their APIs. I am quite happy overall that the world that we have ended up in is one where... Multi-model. Absolutely. And strictly more competitive every quarter over the last few years. Yeah. This year has been insane. Yeah.George [00:24:42]: You can see it. This chart with everything added is hard to read currently. There's so many dots on it, but I think it reflects a little bit what we felt, like how crazy it's been.swyx [00:24:54]: Why 14 as the default? Is that a manual choice? Because you've got service now in there that are less traditional names. Yeah.George [00:25:01]: It's models that we're kind of highlighting by default in our charts, in our intelligence index. Okay.swyx [00:25:07]: You just have a manually curated list of stuff.George [00:25:10]: Yeah, that's right. But something that I actually don't think every artificial analysis user knows is that you can customize our charts and choose what models are highlighted. Yeah. And so if we take off a few names, it gets a little easier to read.swyx [00:25:25]: Yeah, yeah. A little easier to read. Totally. Yeah. But I love that you can see the all one jump. Look at that. September 2024. And the DeepSeek jump. Yeah.George [00:25:34]: Which got close to OpenAI's leadership. They were so close. I think, yeah, we remember that moment. Around this time last year, actually.Micah [00:25:44]: Yeah, yeah, yeah. I agree. Yeah, well, a couple of weeks. It was Boxing Day in New Zealand when DeepSeek v3 came out. And we'd been tracking DeepSeek and a bunch of the other global players that were less known over the second half of 2024 and had run evals on the earlier ones and stuff. I very distinctly remember Boxing Day in New Zealand, because I was with family for Christmas and stuff, running the evals and getting back result by result on DeepSeek v3. So this was the first of their v3 architecture, the 671b MOE.Micah [00:26:19]: And we were very, very impressed. That was the moment where we were sure that DeepSeek was no longer just one of many players, but had jumped up to be a thing. The world really noticed when they followed that up with the RL working on top of v3 and R1 succeeding a few weeks later. But the groundwork for that absolutely was laid with just extremely strong base model, completely open weights that we had as the best open weights model. So, yeah, that's the thing that you really see in the game. But I think that we got a lot of good feedback on Boxing Day. us on Boxing Day last year.George [00:26:48]: Boxing Day is the day after Christmas for those not familiar.George [00:26:54]: I'm from Singapore.swyx [00:26:55]: A lot of us remember Boxing Day for a different reason, for the tsunami that happened. Oh, of course. Yeah, but that was a long time ago. So yeah. So this is the rough pitch of AAQI. Is it A-A-Q-I or A-A-I-I? I-I. Okay. Good memory, though.Micah [00:27:11]: I don't know. I'm not used to it. Once upon a time, we did call it Quality Index, and we would talk about quality, performance, and price, but we changed it to intelligence.George [00:27:20]: There's been a few naming changes. We added hardware benchmarking to the site, and so benchmarks at a kind of system level. And so then we changed our throughput metric to, we now call it output speed, and thenswyx [00:27:32]: throughput makes sense at a system level, so we took that name. Take me through more charts. What should people know? Obviously, the way you look at the site is probably different than how a beginner might look at it.Micah [00:27:42]: Yeah, that's fair. There's a lot of fun stuff to dive into. Maybe so we can hit past all the, like, we have lots and lots of emails and stuff. The interesting ones to talk about today that would be great to bring up are a few of our recent things, I think, that probably not many people will be familiar with yet. So first one of those is our omniscience index. So this one is a little bit different to most of the intelligence evils that we've run. We built it specifically to look at the embedded knowledge in the models and to test hallucination by looking at when the model doesn't know the answer, so not able to get it correct, what's its probability of saying, I don't know, or giving an incorrect answer. So the metric that we use for omniscience goes from negative 100 to positive 100. Because we're simply taking off a point if you give an incorrect answer to the question. We're pretty convinced that this is an example of where it makes most sense to do that, because it's strictly more helpful to say, I don't know, instead of giving a wrong answer to factual knowledge question. And one of our goals is to shift the incentive that evils create for models and the labs creating them to get higher scores. And almost every evil across all of AI up until this point, it's been graded by simple percentage correct as the main metric, the main thing that gets hyped. And so you should take a shot at everything. There's no incentive to say, I don't know. So we did that for this one here.swyx [00:29:22]: I think there's a general field of calibration as well, like the confidence in your answer versus the rightness of the answer. Yeah, we completely agree. Yeah. Yeah.George [00:29:31]: On that. And one reason that we didn't do that is because. Or put that into this index is that we think that the, the way to do that is not to ask the models how confident they are.swyx [00:29:43]: I don't know. Maybe it might be though. You put it like a JSON field, say, say confidence and maybe it spits out something. Yeah. You know, we have done a few evils podcasts over the, over the years. And when we did one with Clementine of hugging face, who maintains the open source leaderboard, and this was one of her top requests, which is some kind of hallucination slash lack of confidence calibration thing. And so, Hey, this is one of them.Micah [00:30:05]: And I mean, like anything that we do, it's not a perfect metric or the whole story of everything that you think about as hallucination. But yeah, it's pretty useful and has some interesting results. Like one of the things that we saw in the hallucination rate is that anthropics Claude models at the, the, the very left-hand side here with the lowest hallucination rates out of the models that we've evaluated amnesty is on. That is an interesting fact. I think it probably correlates with a lot of the previously, not really measured vibes stuff that people like about some of the Claude models. Is the dataset public or what's is it, is there a held out set? There's a hell of a set for this one. So we, we have published a public test set, but we we've only published 10% of it. The reason is that for this one here specifically, it would be very, very easy to like have data contamination because it is just factual knowledge questions. We would. We'll update it at a time to also prevent that, but with yeah, kept most of it held out so that we can keep it reliable for a long time. It leads us to a bunch of really cool things, including breakdown quite granularly by topic. And so we've got some of that disclosed on the website publicly right now, and there's lots more coming in terms of our ability to break out very specific topics. Yeah.swyx [00:31:23]: I would be interested. Let's, let's dwell a little bit on this hallucination one. I noticed that Haiku hallucinates less than Sonnet hallucinates less than Opus. And yeah. Would that be the other way around in a normal capability environments? I don't know. What's, what do you make of that?George [00:31:37]: One interesting aspect is that we've found that there's not really a, not a strong correlation between intelligence and hallucination, right? That's to say that the smarter the models are in a general sense, isn't correlated with their ability to, when they don't know something, say that they don't know. It's interesting that Gemini three pro preview was a big leap over here. Gemini 2.5. Flash and, and, and 2.5 pro, but, and if I add pro quickly here.swyx [00:32:07]: I bet pro's really good. Uh, actually no, I meant, I meant, uh, the GPT pros.George [00:32:12]: Oh yeah.swyx [00:32:13]: Cause GPT pros are rumored. We don't know for a fact that it's like eight runs and then with the LM judge on top. Yeah.George [00:32:20]: So we saw a big jump in, this is accuracy. So this is just percent that they get, uh, correct and Gemini three pro knew a lot more than the other models. And so big jump in accuracy. But relatively no change between the Google Gemini models, between releases. And the hallucination rate. Exactly. And so it's likely due to just kind of different post-training recipe, between the, the Claude models. Yeah.Micah [00:32:45]: Um, there's, there's driven this. Yeah. You can, uh, you can partially blame us and how we define intelligence having until now not defined hallucination as a negative in the way that we think about intelligence.swyx [00:32:56]: And so that's what we're changing. Uh, I know many smart people who are confidently incorrect.George [00:33:02]: Uh, look, look at that. That, that, that is very humans. Very true. And there's times and a place for that. I think our view is that hallucination rate makes sense in this context where it's around knowledge, but in many cases, people want the models to hallucinate, to have a go. Often that's the case in coding or when you're trying to generate newer ideas. One eval that we added to artificial analysis is, is, is critical point and it's really hard, uh, physics problems. Okay.swyx [00:33:32]: And is it sort of like a human eval type or something different or like a frontier math type?George [00:33:37]: It's not dissimilar to frontier frontier math. So these are kind of research questions that kind of academics in the physics physics world would be able to answer, but models really struggled to answer. So the top score here is not 9%.swyx [00:33:51]: And when the people that, that created this like Minway and, and, and actually off via who was kind of behind sweep and what organization is this? Oh, is this, it's Princeton.George [00:34:01]: Kind of range of academics from, from, uh, different academic institutions, really smart people. They talked about how they turn the models up in terms of the temperature as high temperature as they can, where they're trying to explore kind of new ideas in physics as a, as a thought partner, just because they, they want the models to hallucinate. Um, yeah, sometimes it's something new. Yeah, exactly.swyx [00:34:21]: Um, so not right in every situation, but, um, I think it makes sense, you know, to test hallucination in scenarios where it makes sense. Also, the obvious question is, uh, this is one of. Many that there is there, every lab has a system card that shows some kind of hallucination number, and you've chosen to not, uh, endorse that and you've made your own. And I think that's a, that's a choice. Um, totally in some sense, the rest of artificial analysis is public benchmarks that other people can independently rerun. You provide it as a service here. You have to fight the, well, who are we to, to like do this? And your, your answer is that we have a lot of customers and, you know, but like, I guess, how do you converge the individual?Micah [00:35:08]: I mean, I think, I think for hallucinations specifically, there are a bunch of different things that you might care about reasonably, and that you'd measure quite differently, like we've called this a amnesty and solutionation rate, not trying to declare the, like, it's humanity's last hallucination. You could, uh, you could have some interesting naming conventions and all this stuff. Um, the biggest picture answer to that. It's something that I actually wanted to mention. Just as George was explaining, critical point as well is, so as we go forward, we are building evals internally. We're partnering with academia and partnering with AI companies to build great evals. We have pretty strong views on, in various ways for different parts of the AI stack, where there are things that are not being measured well, or things that developers care about that should be measured more and better. And we intend to be doing that. We're not obsessed necessarily with that. Everything we do, we have to do entirely within our own team. Critical point. As a cool example of where we were a launch partner for it, working with academia, we've got some partnerships coming up with a couple of leading companies. Those ones, obviously we have to be careful with on some of the independent stuff, but with the right disclosure, like we're completely comfortable with that. A lot of the labs have released great data sets in the past that we've used to great success independently. And so it's between all of those techniques, we're going to be releasing more stuff in the future. Cool.swyx [00:36:26]: Let's cover the last couple. And then we'll, I want to talk about your trends analysis stuff, you know? Totally.Micah [00:36:31]: So that actually, I have one like little factoid on omniscience. If you go back up to accuracy on omniscience, an interesting thing about this accuracy metric is that it tracks more closely than anything else that we measure. The total parameter count of models makes a lot of sense intuitively, right? Because this is a knowledge eval. This is the pure knowledge metric. We're not looking at the index and the hallucination rate stuff that we think is much more about how the models are trained. This is just what facts did they recall? And yeah, it tracks parameter count extremely closely. Okay.swyx [00:37:05]: What's the rumored size of GPT-3 Pro? And to be clear, not confirmed for any official source, just rumors. But rumors do fly around. Rumors. I get, I hear all sorts of numbers. I don't know what to trust.Micah [00:37:17]: So if you, if you draw the line on omniscience accuracy versus total parameters, we've got all the open ways models, you can squint and see that likely the leading frontier models right now are quite a lot bigger than the ones that we're seeing right now. And the one trillion parameters that the open weights models cap out at, and the ones that we're looking at here, there's an interesting extra data point that Elon Musk revealed recently about XAI that for three trillion parameters for GROK 3 and 4, 6 trillion for GROK 5, but that's not out yet. Take those together, have a look. You might reasonably form a view that there's a pretty good chance that Gemini 3 Pro is bigger than that, that it could be in the 5 to 10 trillion parameters. To be clear, I have absolutely no idea, but just based on this chart, like that's where you would, you would land if you have a look at it. Yeah.swyx [00:38:07]: And to some extent, I actually kind of discourage people from guessing too much because what does it really matter? Like as long as they can serve it as a sustainable cost, that's about it. Like, yeah, totally.George [00:38:17]: They've also got different incentives in play compared to like open weights models who are thinking to supporting others in self-deployment for the labs who are doing inference at scale. It's I think less about total parameters in many cases. When thinking about inference costs and more around number of active parameters. And so there's a bit of an incentive towards larger sparser models. Agreed.Micah [00:38:38]: Understood. Yeah. Great. I mean, obviously if you're a developer or company using these things, not exactly as you say, it doesn't matter. You should be looking at all the different ways that we measure intelligence. You should be looking at cost to run index number and the different ways of thinking about token efficiency and cost efficiency based on the list prices, because that's all it matters.swyx [00:38:56]: It's not as good for the content creator rumor mill where I can say. Oh, GPT-4 is this small circle. Look at GPT-5 is this big circle. And then there used to be a thing for a while. Yeah.Micah [00:39:07]: But that is like on its own, actually a very interesting one, right? That is it just purely that chances are the last couple of years haven't seen a dramatic scaling up in the total size of these models. And so there's a lot of room to go up properly in total size of the models, especially with the upcoming hardware generations. Yes.swyx [00:39:29]: So, you know. Taking off my shitposting face for a minute. Yes. Yes. At the same time, I do feel like, you know, especially coming back from Europe, people do feel like Ilya is probably right that the paradigm is doesn't have many more orders of magnitude to scale out more. And therefore we need to start exploring at least a different path. GDPVal, I think it's like only like a month or so old. I was also very positive when it first came out. I actually talked to Tejo, who was the lead researcher on that. Oh, cool. And you have your own version.George [00:39:59]: It's a fantastic. It's a fantastic data set. Yeah.swyx [00:40:01]: And maybe it will recap for people who are still out of it. It's like 44 tasks based on some kind of GDP cutoff that's like meant to represent broad white collar work that is not just coding. Yeah.Micah [00:40:12]: Each of the tasks have a whole bunch of detailed instructions, some input files for a lot of them. It's within the 44 is divided into like two hundred and twenty two to five, maybe subtasks that are the level of that we run through the agenda. And yeah, they're really interesting. I will say that it doesn't. It doesn't necessarily capture like all the stuff that people do at work. No avail is perfect is always going to be more things to look at, largely because in order to make the tasks well enough to find that you can run them, they need to only have a handful of input files and very specific instructions for that task. And so I think the easiest way to think about them are that they're like quite hard take home exam tasks that you might do in an interview process.swyx [00:40:56]: Yeah, for listeners, it is not no longer like a long prompt. It is like, well, here's a zip file with like a spreadsheet or a PowerPoint deck or a PDF and go nuts and answer this question.George [00:41:06]: OpenAI released a great data set and they released a good paper which looks at performance across the different web chat bots on the data set. It's a great paper, encourage people to read it. What we've done is taken that data set and turned it into an eval that can be run on any model. So we created a reference agentic harness that can run. Run the models on the data set, and then we developed evaluator approach to compare outputs. That's kind of AI enabled, so it uses Gemini 3 Pro Preview to compare results, which we tested pretty comprehensively to ensure that it's aligned to human preferences. One data point there is that even as an evaluator, Gemini 3 Pro, interestingly, doesn't do actually that well. So that's kind of a good example of what we've done in GDPVal AA.swyx [00:42:01]: Yeah, the thing that you have to watch out for with LLM judge is self-preference that models usually prefer their own output, and in this case, it was not. Totally.Micah [00:42:08]: I think the way that we're thinking about the places where it makes sense to use an LLM as judge approach now, like quite different to some of the early LLM as judge stuff a couple of years ago, because some of that and MTV was a great project that was a good example of some of this a while ago was about judging conversations and like a lot of style type stuff. Here, we've got the task that the grader and grading model is doing is quite different to the task of taking the test. When you're taking the test, you've got all of the agentic tools you're working with, the code interpreter and web search, the file system to go through many, many turns to try to create the documents. Then on the other side, when we're grading it, we're running it through a pipeline to extract visual and text versions of the files and be able to provide that to Gemini, and we're providing the criteria for the task and getting it to pick which one more effectively meets the criteria of the task. Yeah. So we've got the task out of two potential outcomes. It turns out that we proved that it's just very, very good at getting that right, matched with human preference a lot of the time, because I think it's got the raw intelligence, but it's combined with the correct representation of the outputs, the fact that the outputs were created with an agentic task that is quite different to the way the grading model works, and we're comparing it against criteria, not just kind of zero shot trying to ask the model to pick which one is better.swyx [00:43:26]: Got it. Why is this an ELO? And not a percentage, like GDP-VAL?George [00:43:31]: So the outputs look like documents, and there's video outputs or audio outputs from some of the tasks. It has to make a video? Yeah, for some of the tasks. Some of the tasks.swyx [00:43:43]: What task is that?George [00:43:45]: I mean, it's in the data set. Like be a YouTuber? It's a marketing video.Micah [00:43:49]: Oh, wow. What? Like model has to go find clips on the internet and try to put it together. The models are not that good at doing that one, for now, to be clear. It's pretty hard to do that with a code editor. I mean, the computer stuff doesn't work quite well enough and so on and so on, but yeah.George [00:44:02]: And so there's no kind of ground truth, necessarily, to compare against, to work out percentage correct. It's hard to come up with correct or incorrect there. And so it's on a relative basis. And so we use an ELO approach to compare outputs from each of the models between the task.swyx [00:44:23]: You know what you should do? You should pay a contractor, a human, to do the same task. And then give it an ELO and then so you have, you have human there. It's just, I think what's helpful about GDPVal, the OpenAI one, is that 50% is meant to be normal human and maybe Domain Expert is higher than that, but 50% was the bar for like, well, if you've crossed 50, you are superhuman. Yeah.Micah [00:44:47]: So we like, haven't grounded this score in that exactly. I agree that it can be helpful, but we wanted to generalize this to a very large number. It's one of the reasons that presenting it as ELO is quite helpful and allows us to add models and it'll stay relevant for quite a long time. I also think it, it can be tricky looking at these exact tasks compared to the human performance, because the way that you would go about it as a human is quite different to how the models would go about it. Yeah.swyx [00:45:15]: I also liked that you included Lama 4 Maverick in there. Is that like just one last, like...Micah [00:45:20]: Well, no, no, no, no, no, no, it is the, it is the best model released by Meta. And... So it makes it into the homepage default set, still for now.George [00:45:31]: Other inclusion that's quite interesting is we also ran it across the latest versions of the web chatbots. And so we have...swyx [00:45:39]: Oh, that's right.George [00:45:40]: Oh, sorry.swyx [00:45:41]: I, yeah, I completely missed that. Okay.George [00:45:43]: No, not at all. So that, which has a checkered pattern. So that is their harness, not yours, is what you're saying. Exactly. And what's really interesting is that if you compare, for instance, Claude 4.5 Opus using the Claude web chatbot, it performs worse than the model in our agentic harness. And so in every case, the model performs better in our agentic harness than its web chatbot counterpart, the harness that they created.swyx [00:46:13]: Oh, my backwards explanation for that would be that, well, it's meant for consumer use cases and here you're pushing it for something.Micah [00:46:19]: The constraints are different and the amount of freedom that you can give the model is different. Also, you like have a cost goal. We let the models work as long as they want, basically. Yeah. Do you copy paste manually into the chatbot? Yeah. Yeah. That's, that was how we got the chatbot reference. We're not going to be keeping those updated at like quite the same scale as hundreds of models.swyx [00:46:38]: Well, so I don't know, talk to a browser base. They'll, they'll automate it for you. You know, like I have thought about like, well, we should turn these chatbot versions into an API because they are legitimately different agents in themselves. Yes. Right. Yeah.Micah [00:46:53]: And that's grown a huge amount of the last year, right? Like the tools. The tools that are available have actually diverged in my opinion, a fair bit across the major chatbot apps and the amount of data sources that you can connect them to have gone up a lot, meaning that your experience and the way you're using the model is more different than ever.swyx [00:47:10]: What tools and what data connections come to mind when you say what's interesting, what's notable work that people have done?Micah [00:47:15]: Oh, okay. So my favorite example on this is that until very recently, I would argue that it was basically impossible to get an LLM to draft an email for me in any useful way. Because most times that you're sending an email, you're not just writing something for the sake of writing it. Chances are context required is a whole bunch of historical emails. Maybe it's notes that you've made, maybe it's meeting notes, maybe it's, um, pulling something from your, um, any of like wherever you at work store stuff. So for me, like Google drive, one drive, um, in our super base databases, if we need to do some analysis or some data or something, preferably model can be plugged into all of those things and can go do some useful work based on it. The things that like I find most impressive currently that I am somewhat surprised work really well in late 2025, uh, that I can have models use super base MCP to query read only, of course, run a whole bunch of SQL queries to do pretty significant data analysis. And. And make charts and stuff and can read my Gmail and my notion. And okay. You actually use that. That's good. That's, that's, that's good. Is that a cloud thing? To various degrees of order, but chat GPD and Claude right now, I would say that this stuff like barely works in fairness right now. Like.George [00:48:33]: Because people are actually going to try this after they hear it. If you get an email from Micah, odds are it wasn't written by a chatbot.Micah [00:48:38]: So, yeah, I think it is true that I have never actually sent anyone an email drafted by a chatbot. Yet.swyx [00:48:46]: Um, and so you can, you can feel it right. And yeah, this time, this time next year, we'll come back and see where it's going. Totally. Um, super base shout out another famous Kiwi. Uh, I don't know if you've, you've any conversations with him about anything in particular on AI building and AI infra.George [00:49:03]: We have had, uh, Twitter DMS, um, with, with him because we're quite big, uh, super base users and power users. And we probably do some things more manually than we should in. In, in super base support line because you're, you're a little bit being super friendly. One extra, um, point regarding, um, GDP Val AA is that on the basis of the overperformance of the models compared to the chatbots turns out, we realized that, oh, like our reference harness that we built actually white works quite well on like gen generalist agentic tasks. This proves it in a sense. And so the agent harness is very. Minimalist. I think it follows some of the ideas that are in Claude code and we, all that we give it is context management capabilities, a web search, web browsing, uh, tool, uh, code execution, uh, environment. Anything else?Micah [00:50:02]: I mean, we can equip it with more tools, but like by default, yeah, that's it. We, we, we give it for GDP, a tool to, uh, view an image specifically, um, because the models, you know, can just use a terminal to pull stuff in text form into context. But to pull visual stuff into context, we had to give them a custom tool, but yeah, exactly. Um, you, you can explain an expert. No.George [00:50:21]: So it's, it, we turned out that we created a good generalist agentic harness. And so we, um, released that on, on GitHub yesterday. It's called stirrup. So if people want to check it out and, and it's a great, um, you know, base for, you know, generalist, uh, building a generalist agent for more specific tasks.Micah [00:50:39]: I'd say the best way to use it is get clone and then have your favorite coding. Agent make changes to it, to do whatever you want, because it's not that many lines of code and the coding agents can work with it. Super well.swyx [00:50:51]: Well, that's nice for the community to explore and share and hack on it. I think maybe in, in, in other similar environments, the terminal bench guys have done, uh, sort of the Harbor. Uh, and so it's, it's a, it's a bundle of, well, we need our minimal harness, which for them is terminus and we also need the RL environments or Docker deployment thing to, to run independently. So I don't know if you've looked at it. I don't know if you've looked at the harbor at all, is that, is that like a, a standard that people want to adopt?George [00:51:19]: Yeah, we've looked at it from a evals perspective and we love terminal bench and, and host benchmarks of, of, of terminal mention on artificial analysis. Um, we've looked at it from a, from a coding agent perspective, but could see it being a great, um, basis for any kind of agents. I think where we're getting to is that these models have gotten smart enough. They've gotten better, better tools that they can perform better when just given a minimalist. Set of tools and, and let them run, let the model control the, the agentic workflow rather than using another framework that's a bit more built out that tries to dictate the, dictate the flow. Awesome.swyx [00:51:56]: Let's cover the openness index and then let's go into the report stuff. Uh, so that's the, that's the last of the proprietary art numbers, I guess. I don't know how you sort of classify all these. Yeah.Micah [00:52:07]: Or call it, call it, let's call it the last of like the, the three new things that we're talking about from like the last few weeks. Um, cause I mean, there's a, we do a mix of stuff that. Where we're using open source, where we open source and what we do and, um, proprietary stuff that we don't always open source, like long context reasoning data set last year, we did open source. Um, and then all of the work on performance benchmarks across the site, some of them, we looking to open source, but some of them, like we're constantly iterating on and so on and so on and so on. So there's a huge mix, I would say, just of like stuff that is open source and not across the side. So that's a LCR for people. Yeah, yeah, yeah, yeah.swyx [00:52:41]: Uh, but let's, let's, let's talk about open.Micah [00:52:42]: Let's talk about openness index. This. Here is call it like a new way to think about how open models are. We, for a long time, have tracked where the models are open weights and what the licenses on them are. And that's like pretty useful. That tells you what you're allowed to do with the weights of a model, but there is this whole other dimension to how open models are. That is pretty important that we haven't tracked until now. And that's how much is disclosed about how it was made. So transparency about data, pre-training data and post-training data. And whether you're allowed to use that data and transparency about methodology and training code. So basically, those are the components. We bring them together to score an openness index for models so that you can in one place get this full picture of how open models are.swyx [00:53:32]: I feel like I've seen a couple other people try to do this, but they're not maintained. I do think this does matter. I don't know what the numbers mean apart from is there a max number? Is this out of 20?George [00:53:44]: It's out of 18 currently, and so we've got an openness index page, but essentially these are points, you get points for being more open across these different categories and the maximum you can achieve is 18. So AI2 with their extremely open OMO3 32B think model is the leader in a sense.swyx [00:54:04]: It's hooking face.George [00:54:05]: Oh, with their smaller model. It's coming soon. I think we need to run, we need to get the intelligence benchmarks right to get it on the site.swyx [00:54:12]: You can't have it open in the next. We can not include hooking face. We love hooking face. We'll have that, we'll have that up very soon. I mean, you know, the refined web and all that stuff. It's, it's amazing. Or is it called fine web? Fine web. Fine web.Micah [00:54:23]: Yeah, yeah, no, totally. Yep. One of the reasons this is cool, right, is that if you're trying to understand the holistic picture of the models and what you can do with all the stuff the company's contributing, this gives you that picture. And so we are going to keep it up to date alongside all the models that we do intelligence index on, on the site. And it's just an extra view to understand.swyx [00:54:43]: Can you scroll down to this? The, the, the, the trade-offs chart. Yeah, yeah. That one. Yeah. This, this really matters, right? Obviously, because you can b

We are reupping this episode after LMArena announced their fresh Series A (https://www.theinformation.com/articles/ai-evaluation-startup-lmarena-valued-1-7-billion-new-funding-round?rc=luxwz4), raising $150m at a $1.7B valuation, with $30M annualized consumption revenue (aka $2.5m MRR) after their September evals product launch.—-From building LMArena in a Berkeley basement to raising $100M and becoming the de facto leaderboard for frontier AI, Anastasios Angelopoulos returns to Latent Space to recap 2025 in one of the most influential platforms in AI—trusted by millions of users, every major lab, and the entire industry to answer one question: which model is actually best for real-world use cases? We caught up with Anastasios live at NeurIPS 2025 to dig into the origin story (spoiler: it started as an academic project incubated by Anjney Midha at a16z, who formed an entity and gave grants before they even committed to starting a company), why they decided to spin out instead of staying academic or nonprofit (the only way to scale was to build a company), how they're spending that $100M (inference costs, React migration off Gradio, and hiring world-class talent across ML, product, and go-to-market), the leaderboard delusion controversy and why their response demolished the paper's claims (factual errors, misrepresentation of open vs. closed source sampling, and ignoring the transparency of preview testing that the community loves), why platform integrity comes first (the public leaderboard is a charity, not a pay-to-play system—models can't pay to get on, can't pay to get off, and scores reflect millions of real votes), how they're expanding into occupational verticals (medicine, legal, finance, creative marketing) and multimodal arenas (video coming soon), why consumer retention is earned every single day (sign-in and persistent history were the unlock, but users are fickle and can leave at any moment), and his vision for Arena as the central evaluation platform that provides the North Star for the industry—constantly fresh, immune to overfitting, and grounded in millions of real-world conversations from real users.We discuss:* The $100M raise: use of funds is primarily inference costs (funding free usage for tens of millions of monthly conversations), React migration off Gradio (custom loading icons, better developer hiring, more flexibility), and hiring world-class talent* The scale: 250M+ conversations on the platform, tens of millions per month, 25% of users do software for a living, and half of users are now logged in* The leaderboard illusion controversy: Cohere researchers claimed undisclosed private testing created inequities, but Arena's response demolished the paper's factual errors (misrepresented open vs. closed source sampling, ignored transparency of preview testing that the community loves)* Why preview testing is loved by the community: secret codenames (Gemini Nano Banana, named after PM Naina's nickname), early access to unreleased models, and the thrill of being first to vote on frontier capabilities* The Nano Banana moment: changed Google's market share overnight, billions of dollars in stock movement, and validated that multimodal models (image generation, video) are economically critical for marketing, design, and AI-for-science* New categories: occupational and expert arenas (medicine, legal, finance, creative marketing), Code Arena, and video arena coming soonFull Video EpisodeTimestamps00:00:00 Introduction: Anastasios from Arena and the LM Arena Journey00:01:36 The Anjney Midha Incubation: From Berkeley Basement to Startup00:02:47 The Decision to Start a Company: Scaling Beyond Academia00:03:38 The $100M Raise: Use of Funds and Platform Economics00:05:10 Arena's User Base: 5M+ Users and Diverse Demographics00:06:02 The Competitive Landscape: Artificial Analysis, AI.xyz, and Arena's Differentiation00:08:12 Educational Value and Learning from the Community00:08:41 Technical Migration: From Gradio to React and Platform Evolution00:10:18 Leaderboard Delusion Paper: Addressing Critiques and Maintaining Integrity00:12:29 Nano Banana Moment: How Preview Models Create Market Impact00:13:41 Multimodal AI and Image Generation: From Skepticism to Economic Value00:15:37 Core Principles: Platform Integrity and the Public Leaderboard as Charity00:18:29 Future Roadmap: Expert Categories, Multimodal, Video, and Occupational Verticals00:19:10 API Strategy and Focus: Doing One Thing Well00:19:51 Community Management and Retention: Sign-In, History, and Daily Value00:22:21 Partnerships and Agent Evaluation: From Devon to Full-Featured Harnesses00:21:49 Hiring and Building a High-Performance Team Get full access to Latent.Space at www.latent.space/subscribe

From undergraduate research seminars at Princeton to winning Best Paper award at NeurIPS 2025, Kevin Wang, Ishaan Javali, Michał Bortkiewicz, Tomasz Trzcinski, Benjamin Eysenbach defied conventional wisdom by scaling reinforcement learning networks to 1,000 layers deep—unlocking performance gains that the RL community thought impossible. We caught up with the team live at NeurIPS to dig into the story behind RL1000: why deep networks have worked in language and vision but failed in RL for over a decade (spoiler: it's not just about depth, it's about the objective), how they discovered that self-supervised RL (learning representations of states, actions, and future states via contrastive learning) scales where value-based methods collapse, the critical architectural tricks that made it work (residual connections, layer normalization, and a shift from regression to classification), why scaling depth is more parameter-efficient than scaling width (linear vs. quadratic growth), how Jax and GPU-accelerated environments let them collect hundreds of millions of transitions in hours (the data abundance that unlocked scaling in the first place), the “critical depth” phenomenon where performance doesn't just improve—it multiplies once you cross 15M+ transitions and add the right architectural components, why this isn't just “make networks bigger” but a fundamental shift in RL objectives (their code doesn't have a line saying “maximize rewards”—it's pure self-supervised representation learning), how deep teacher, shallow student distillation could unlock deployment at scale (train frontier capabilities with 1000 layers, distill down to efficient inference models), the robotics implications (goal-conditioned RL without human supervision or demonstrations, scaling architecture instead of scaling manual data collection), and their thesis that RL is finally ready to scale like language and vision—not by throwing compute at value functions, but by borrowing the self-supervised, representation-learning paradigms that made the rest of deep learning work.We discuss:* The self-supervised RL objective: instead of learning value functions (noisy, biased, spurious), they learn representations where states along the same trajectory are pushed together, states along different trajectories are pushed apart—turning RL into a classification problem* Why naive scaling failed: doubling depth degraded performance, doubling again with residual connections and layer norm suddenly skyrocketed performance in one environment—unlocking the “critical depth” phenomenon* Scaling depth vs. width: depth grows parameters linearly, width grows quadratically—depth is more parameter-efficient and sample-efficient for the same performance* The Jax + GPU-accelerated environments unlock: collecting thousands of trajectories in parallel meant data wasn't the bottleneck, and crossing 15M+ transitions was when deep networks really paid off* The blurring of RL and self-supervised learning: their code doesn't maximize rewards directly, it's an actor-critic goal-conditioned RL algorithm, but the learning burden shifts to classification (cross-entropy loss, representation learning) instead of TD error regression* Why scaling batch size unlocks at depth: traditional RL doesn't benefit from larger batches because networks are too small to exploit the signal, but once you scale depth, batch size becomes another effective scaling dimension—RL1000 Team (Princeton)* 1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities: https://openreview.net/forum?id=s0JVsx3bx1Full Video EpisodeTimestamps00:00:00 Introduction: Best Paper Award and NeurIPS Poster Experience00:01:11 Team Introductions and Princeton Research Origins00:03:35 The Deep Learning Anomaly: Why RL Stayed Shallow00:04:35 Self-Supervised RL: A Different Approach to Scaling00:05:13 The Breakthrough Moment: Residual Connections and Critical Depth00:07:15 Architectural Choices: Borrowing from ResNets and Avoiding Vanishing Gradients00:07:50 Clarifying the Paper: Not Just Big Networks, But Different Objectives00:08:46 Blurring the Lines: RL Meets Self-Supervised Learning00:09:44 From TD Errors to Classification: Why This Objective Scales00:11:06 Architecture Details: Building on Braw and SymbaFowl00:12:05 Robotics Applications: Goal-Conditioned RL Without Human Supervision00:13:15 Efficiency Trade-offs: Depth vs Width and Parameter Scaling00:15:48 JAX and GPU-Accelerated Environments: The Data Infrastructure00:18:05 World Models and Next State Classification00:22:37 Unlocking Batch Size Scaling Through Network Capacity00:24:10 Compute Requirements: State-of-the-Art on a Single GPU00:21:02 Future Directions: Distillation, VLMs, and Hierarchical Planning00:27:15 Closing Thoughts: Challenging Conventional Wisdom in RL Scaling Get full access to Latent.Space at www.latent.space/subscribe

From undergraduate research seminars at Princeton to winning Best Paper award at NeurIPS 2025, Kevin Wang, Ishaan Javali, Michał Bortkiewicz, Tomasz Trzcinski, Benjamin Eysenbach defied conventional wisdom by scaling reinforcement learning networks to 1,000 layers deep—unlocking performance gains that the RL community thought impossible. We caught up with the team live at NeurIPS to dig into the story behind RL1000: why deep networks have worked in language and vision but failed in RL for over a decade (spoiler: it's not just about depth, it's about the objective), how they discovered that self-supervised RL (learning representations of states, actions, and future states via contrastive learning) scales where value-based methods collapse, the critical architectural tricks that made it work (residual connections, layer normalization, and a shift from regression to classification), why scaling depth is more parameter-efficient than scaling width (linear vs. quadratic growth), how Jax and GPU-accelerated environments let them collect hundreds of millions of transitions in hours (the data abundance that unlocked scaling in the first place), the "critical depth" phenomenon where performance doesn't just improve—it multiplies once you cross 15M+ transitions and add the right architectural components, why this isn't just "make networks bigger" but a fundamental shift in RL objectives (their code doesn't have a line saying "maximize rewards"—it's pure self-supervised representation learning), how deep teacher, shallow student distillation could unlock deployment at scale (train frontier capabilities with 1000 layers, distill down to efficient inference models), the robotics implications (goal-conditioned RL without human supervision or demonstrations, scaling architecture instead of scaling manual data collection), and their thesis that RL is finally ready to scale like language and vision—not by throwing compute at value functions, but by borrowing the self-supervised, representation-learning paradigms that made the rest of deep learning work. We discuss: The self-supervised RL objective: instead of learning value functions (noisy, biased, spurious), they learn representations where states along the same trajectory are pushed together, states along different trajectories are pushed apart—turning RL into a classification problem Why naive scaling failed: doubling depth degraded performance, doubling again with residual connections and layer norm suddenly skyrocketed performance in one environment—unlocking the "critical depth" phenomenon Scaling depth vs. width: depth grows parameters linearly, width grows quadratically—depth is more parameter-efficient and sample-efficient for the same performance The Jax + GPU-accelerated environments unlock: collecting thousands of trajectories in parallel meant data wasn't the bottleneck, and crossing 15M+ transitions was when deep networks really paid off The blurring of RL and self-supervised learning: their code doesn't maximize rewards directly, it's an actor-critic goal-conditioned RL algorithm, but the learning burden shifts to classification (cross-entropy loss, representation learning) instead of TD error regression Why scaling batch size unlocks at depth: traditional RL doesn't benefit from larger batches because networks are too small to exploit the signal, but once you scale depth, batch size becomes another effective scaling dimension — RL1000 Team (Princeton) 1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities: https://openreview.net/forum?id=s0JVsx3bx1 Chapters 00:00:00 Introduction: Best Paper Award and NeurIPS Poster Experience 00:01:11 Team Introductions and Princeton Research Origins 00:03:35 The Deep Learning Anomaly: Why RL Stayed Shallow 00:04:35 Self-Supervised RL: A Different Approach to Scaling 00:05:13 The Breakthrough Moment: Residual Connections and Critical Depth 00:07:15 Architectural Choices: Borrowing from ResNets and Avoiding Vanishing Gradients 00:07:50 Clarifying the Paper: Not Just Big Networks, But Different Objectives 00:08:46 Blurring the Lines: RL Meets Self-Supervised Learning 00:09:44 From TD Errors to Classification: Why This Objective Scales 00:11:06 Architecture Details: Building on Braw and SymbaFowl 00:12:05 Robotics Applications: Goal-Conditioned RL Without Human Supervision 00:13:15 Efficiency Trade-offs: Depth vs Width and Parameter Scaling 00:15:48 JAX and GPU-Accelerated Environments: The Data Infrastructure 00:18:05 World Models and Next State Classification 00:22:37 Unlocking Batch Size Scaling Through Network Capacity 00:24:10 Compute Requirements: State-of-the-Art on a Single GPU 00:21:02 Future Directions: Distillation, VLMs, and Hierarchical Planning 00:27:15 Closing Thoughts: Challenging Conventional Wisdom in RL Scaling

From creating SWE-bench in a Princeton basement to shipping CodeClash, SWE-bench Multimodal, and SWE-bench Multilingual, John Yang has spent the last year and a half watching his benchmark become the de facto standard for evaluating AI coding agents—trusted by Cognition (Devin), OpenAI, Anthropic, and every major lab racing to solve software engineering at scale. We caught up with John live at NeurIPS 2025 to dig into the state of code evals heading into 2026: why SWE-bench went from ignored (October 2023) to the industry standard after Devin's launch (and how Walden emailed him two weeks before the big reveal), how the benchmark evolved from Django-heavy to nine languages across 40 repos (JavaScript, Rust, Java, C, Ruby), why unit tests as verification are limiting and long-running agent tournaments might be the future (CodeClash: agents maintain codebases, compete in arenas, and iterate over multiple rounds), the proliferation of SWE-bench variants (SWE-bench Pro, SWE-bench Live, SWE-Efficiency, AlgoTune, SciCode) and how benchmark authors are now justifying their splits with curation techniques instead of just "more repos," why Tau-bench's "impossible tasks" controversy is actually a feature not a bug (intentionally including impossible tasks flags cheating), the tension between long autonomy (5-hour runs) vs. interactivity (Cognition's emphasis on fast back-and-forth), how Terminal-bench unlocked creativity by letting PhD students and non-coders design environments beyond GitHub issues and PRs, the academic data problem (companies like Cognition and Cursor have rich user interaction data, academics need user simulators or compelling products like LMArena to get similar signal), and his vision for CodeClash as a testbed for human-AI collaboration—freeze model capability, vary the collaboration setup (solo agent, multi-agent, human+agent), and measure how interaction patterns change as models climb the ladder from code completion to full codebase reasoning. We discuss: John's path: Princeton → SWE-bench (October 2023) → Stanford PhD with Diyi Yang and the Iris Group, focusing on code evals, human-AI collaboration, and long-running agent benchmarks The SWE-bench origin story: released October 2023, mostly ignored until Cognition's Devin launch kicked off the arms race (Walden emailed John two weeks before: "we have a good number") SWE-bench Verified: the curated, high-quality split that became the standard for serious evals SWE-bench Multimodal and Multilingual: nine languages (JavaScript, Rust, Java, C, Ruby) across 40 repos, moving beyond the Django-heavy original distribution The SWE-bench Pro controversy: independent authors used the "SWE-bench" name without John's blessing, but he's okay with it ("congrats to them, it's a great benchmark") CodeClash: John's new benchmark for long-horizon development—agents maintain their own codebases, edit and improve them each round, then compete in arenas (programming games like Halite, economic tasks like GDP optimization) SWE-Efficiency (Jeffrey Maugh, John's high school classmate): optimize code for speed without changing behavior (parallelization, SIMD operations) AlgoTune, SciCode, Terminal-bench, Tau-bench, SecBench, SRE-bench: the Cambrian explosion of code evals, each diving into different domains (security, SRE, science, user simulation) The Tau-bench "impossible tasks" debate: some tasks are underspecified or impossible, but John thinks that's actually a feature (flags cheating if you score above 75%) Cognition's research focus: codebase understanding (retrieval++), helping humans understand their own codebases, and automatic context engineering for LLMs (research sub-agents) The vision: CodeClash as a testbed for human-AI collaboration—vary the setup (solo agent, multi-agent, human+agent), freeze model capability, and measure how interaction changes as models improve — John Yang SWE-bench: https://www.swebench.com X: https://x.com/jyangballin Chapters 00:00:00 Introduction: John Yang on SWE-bench and Code Evaluations 00:00:31 SWE-bench Origins and Devon's Impact on the Coding Agent Arms Race 00:01:09 SWE-bench Ecosystem: Verified, Pro, Multimodal, and Multilingual Variants 00:02:17 Moving Beyond Django: Diversifying Code Evaluation Repositories 00:03:08 Code Clash: Long-Horizon Development Through Programming Tournaments 00:04:41 From Halite to Economic Value: Designing Competitive Coding Arenas 00:06:04 Ofir's Lab: SWE-ficiency, AlgoTune, and SciCode for Scientific Computing 00:07:52 The Benchmark Landscape: TAU-bench, Terminal-bench, and User Simulation 00:09:20 The Impossible Task Debate: Refusals, Ambiguity, and Benchmark Integrity 00:12:32 The Future of Code Evals: Long Autonomy vs Human-AI Collaboration 00:14:37 Call to Action: User Interaction Data and Codebase Understanding Research

From creating SWE-bench in a Princeton basement to shipping CodeClash, SWE-bench Multimodal, and SWE-bench Multilingual, John Yang has spent the last year and a half watching his benchmark become the de facto standard for evaluating AI coding agents—trusted by Cognition (Devin), OpenAI, Anthropic, and every major lab racing to solve software engineering at scale. We caught up with John live at NeurIPS 2025 to dig into the state of code evals heading into 2026: why SWE-bench went from ignored (October 2023) to the industry standard after Devin's launch (and how Walden emailed him two weeks before the big reveal), how the benchmark evolved from Django-heavy to nine languages across 40 repos (JavaScript, Rust, Java, C, Ruby), why unit tests as verification are limiting and long-running agent tournaments might be the future (CodeClash: agents maintain codebases, compete in arenas, and iterate over multiple rounds), the proliferation of SWE-bench variants (SWE-bench Pro, SWE-bench Live, SWE-Efficiency, AlgoTune, SciCode) and how benchmark authors are now justifying their splits with curation techniques instead of just “more repos,” why Tau-bench's “impossible tasks” controversy is actually a feature not a bug (intentionally including impossible tasks flags cheating), the tension between long autonomy (5-hour runs) vs. interactivity (Cognition's emphasis on fast back-and-forth), how Terminal-bench unlocked creativity by letting PhD students and non-coders design environments beyond GitHub issues and PRs, the academic data problem (companies like Cognition and Cursor have rich user interaction data, academics need user simulators or compelling products like LMArena to get similar signal), and his vision for CodeClash as a testbed for human-AI collaboration—freeze model capability, vary the collaboration setup (solo agent, multi-agent, human+agent), and measure how interaction patterns change as models climb the ladder from code completion to full codebase reasoning.We discuss:* John's path: Princeton → SWE-bench (October 2023) → Stanford PhD with Diyi Yang and the Iris Group, focusing on code evals, human-AI collaboration, and long-running agent benchmarks* The SWE-bench origin story: released October 2023, mostly ignored until Cognition's Devin launch kicked off the arms race (Walden emailed John two weeks before: “we have a good number”)* SWE-bench Verified: the curated, high-quality split that became the standard for serious evals* SWE-bench Multimodal and Multilingual: nine languages (JavaScript, Rust, Java, C, Ruby) across 40 repos, moving beyond the Django-heavy original distribution* The SWE-bench Pro controversy: independent authors used the “SWE-bench” name without John's blessing, but he's okay with it (”congrats to them, it's a great benchmark”)* CodeClash: John's new benchmark for long-horizon development—agents maintain their own codebases, edit and improve them each round, then compete in arenas (programming games like Halite, economic tasks like GDP optimization)* SWE-Efficiency (Jeffrey Maugh, John's high school classmate): optimize code for speed without changing behavior (parallelization, SIMD operations)* AlgoTune, SciCode, Terminal-bench, Tau-bench, SecBench, SRE-bench: the Cambrian explosion of code evals, each diving into different domains (security, SRE, science, user simulation)* The Tau-bench “impossible tasks” debate: some tasks are underspecified or impossible, but John thinks that's actually a feature (flags cheating if you score above 75%)* Cognition's research focus: codebase understanding (retrieval++), helping humans understand their own codebases, and automatic context engineering for LLMs (research sub-agents)* The vision: CodeClash as a testbed for human-AI collaboration—vary the setup (solo agent, multi-agent, human+agent), freeze model capability, and measure how interaction changes as models improve—John Yang* SWE-bench: https://www.swebench.com* X: https://x.com/jyangballinFull Video EpisodeTimestamps00:00:00 Introduction: John Yang on SWE-bench and Code Evaluations00:00:31 SWE-bench Origins and Devon's Impact on the Coding Agent Arms Race00:01:09 SWE-bench Ecosystem: Verified, Pro, Multimodal, and Multilingual Variants00:02:17 Moving Beyond Django: Diversifying Code Evaluation Repositories00:03:08 Code Clash: Long-Horizon Development Through Programming Tournaments00:04:41 From Halite to Economic Value: Designing Competitive Coding Arenas00:06:04 Ofir's Lab: SWE-ficiency, AlgoTune, and SciCode for Scientific Computing00:07:52 The Benchmark Landscape: TAU-bench, Terminal-bench, and User Simulation00:09:20 The Impossible Task Debate: Refusals, Ambiguity, and Benchmark Integrity00:12:32 The Future of Code Evals: Long Autonomy vs Human-AI Collaboration00:14:37 Call to Action: User Interaction Data and Codebase Understanding Research Get full access to Latent.Space at www.latent.space/subscribe

From building LMArena in a Berkeley basement to raising $100M and becoming the de facto leaderboard for frontier AI, Anastasios Angelopoulos returns to Latent Space to recap 2025 in one of the most influential platforms in AI—trusted by millions of users, every major lab, and the entire industry to answer one question: which model is actually best for real-world use cases? We caught up with Anastasios live at NeurIPS 2025 to dig into the origin story (spoiler: it started as an academic project incubated by Anjney Midha at a16z, who formed an entity and gave grants before they even committed to starting a company), why they decided to spin out instead of staying academic or nonprofit (the only way to scale was to build a company), how they're spending that $100M (inference costs, React migration off Gradio, and hiring world-class talent across ML, product, and go-to-market), the leaderboard delusion controversy and why their response demolished the paper's claims (factual errors, misrepresentation of open vs. closed source sampling, and ignoring the transparency of preview testing that the community loves), why platform integrity comes first (the public leaderboard is a charity, not a pay-to-play system—models can't pay to get on, can't pay to get off, and scores reflect millions of real votes), how they're expanding into occupational verticals (medicine, legal, finance, creative marketing) and multimodal arenas (video coming soon), why consumer retention is earned every single day (sign-in and persistent history were the unlock, but users are fickle and can leave at any moment), the Gemini Nano Banana moment that changed Google's market share overnight (and why multimodal models are becoming economically critical for marketing, design, and AI-for-science), how they're thinking about agents and harnesses (Code Arena evaluates models, but maybe it should evaluate full agents like Devin), and his vision for Arena as the central evaluation platform that provides the North Star for the industry—constantly fresh, immune to overfitting, and grounded in millions of real-world conversations from real users. We discuss: The $100M raise: use of funds is primarily inference costs (funding free usage for tens of millions of monthly conversations), React migration off Gradio (custom loading icons, better developer hiring, more flexibility), and hiring world-class talent The scale: 250M+ conversations on the platform, tens of millions per month, 25% of users do software for a living, and half of users are now logged in The leaderboard illusion controversy: Cohere researchers claimed undisclosed private testing created inequities, but Arena's response demolished the paper's factual errors (misrepresented open vs. closed source sampling, ignored transparency of preview testing that the community loves) Why preview testing is loved by the community: secret codenames (Gemini Nano Banana, named after PM Naina's nickname), early access to unreleased models, and the thrill of being first to vote on frontier capabilities The Nano Banana moment: changed Google's market share overnight, billions of dollars in stock movement, and validated that multimodal models (image generation, video) are economically critical for marketing, design, and AI-for-science New categories: occupational and expert arenas (medicine, legal, finance, creative marketing), Code Arena, and video arena coming soon Consumer retention: sign-in and persistent history were the unlock, but users are fickle and earned every single day—"every user is earned, they can leave at any moment" — Anastasios Angelopoulos Arena: https://lmarena.ai X: https://x.com/arena Chapters 00:00:00 Introduction: Anastasios from Arena and the LM Arena Journey 00:01:36 The Anjney Midha Incubation: From Berkeley Basement to Startup 00:02:47 The Decision to Start a Company: Scaling Beyond Academia 00:03:38 The $100M Raise: Use of Funds and Platform Economics 00:05:10 Arena's User Base: 5M+ Users and Diverse Demographics 00:06:02 The Competitive Landscape: Artificial Analysis, AI.xyz, and Arena's Differentiation 00:08:12 Educational Value and Learning from the Community 00:08:41 Technical Migration: From Gradio to React and Platform Evolution 00:10:18 Leaderboard Delusion Paper: Addressing Critiques and Maintaining Integrity 00:12:29 Nano Banana Moment: How Preview Models Create Market Impact 00:13:41 Multimodal AI and Image Generation: From Skepticism to Economic Value 00:15:37 Core Principles: Platform Integrity and the Public Leaderboard as Charity 00:18:29 Future Roadmap: Expert Categories, Multimodal, Video, and Occupational Verticals 00:19:10 API Strategy and Focus: Doing One Thing Well 00:19:51 Community Management and Retention: Sign-In, History, and Daily Value 00:22:21 Partnerships and Agent Evaluation: From Devon to Full-Featured Harnesses 00:21:49 Hiring and Building a High-Performance Team

From pre-training data curation to shipping GPT-4o, o1, o3, and now GPT-5 thinking and the shopping model, Josh McGrath has lived through the full arc of OpenAI's post-training evolution—from the PPO vs DPO debates of 2023 to today's RLVR era, where the real innovation isn't optimization methods but data quality, signal trust, and token efficiency. We sat down with Josh at NeurIPS 2025 to dig into the state of post-training heading into 2026: why RLHF and RLVR are both just policy gradient methods (the difference is the input data, not the math), how GRPO from DeepSeek Math was underappreciated as a shift toward more trustworthy reward signals (math answers you can verify vs. human preference you can't), why token efficiency matters more than wall-clock time (GPT-5 to 5.1 bumped evals and slashed tokens), how Codex has changed his workflow so much he feels “trapped” by 40-minute design sessions followed by 15-minute agent sprints, the infrastructure chaos of scaling RL (”way more moving parts than pre-training”), why long context will keep climbing but agents + graph walks might matter more than 10M-token windows, the shopping model as a test bed for interruptability and chain-of-thought transparency, why personality toggles (Anton vs Clippy) are a real differentiator users care about, and his thesis that the education system isn't producing enough people who can do both distributed systems and ML research—the exact skill set required to push the frontier when the bottleneck moves every few weeks.We discuss:* Josh's path: pre-training data curation → post-training researcher at OpenAI, shipping GPT-4o, o1, o3, GPT-5 thinking, and the shopping model* Why he switched from pre-training to post-training: “Do I want to make 3% compute efficiency wins, or change behavior by 40%?”* The RL infrastructure challenge: way more moving parts than pre-training (tasks, grading setups, external partners), and why babysitting runs at 12:30am means jumping into unfamiliar code constantly* How Codex has changed his workflow: 40-minute design sessions compressed into 15-minute agent sprints, and the strange “trapped” feeling of waiting for the agent to finish* The RLHF vs RLVR debate: both are policy gradient methods, the real difference is data quality and signal trust (human preference vs. verifiable correctness)* Why GRPO (from DeepSeek Math) was underappreciated: not just an optimization trick, but a shift toward reward signals you can actually trust (math answers over human vibes)* The token efficiency revolution: GPT-5 to 5.1 bumped evals and slashed tokens, and why thinking in tokens (not wall-clock time) unlocks better tool-calling and agent workflows* Personality toggles: Anton (tool, no warmth) vs Clippy (friendly, helpful), and why Josh uses custom instructions to make his model “just a tool”* The router problem: having a router at the top (GPT-5 thinking vs non-thinking) and an implicit router (thinking effort slider) creates weird bumps, and why the abstractions will eventually merge* Long context: climbing Graph Blocks evals, the dream of 10M+ token windows, and why agents + graph walks might matter more than raw context length* Why the education system isn't producing enough people who can do both distributed systems and ML research, and why that's the bottleneck for frontier labs* The 2026 vision: neither pre-training nor post-training is dead, we're in the fog of war, and the bottleneck will keep moving (so emotional stability helps)—Josh McGrath* OpenAI: https://openai.com* X: https://x.com/j_mcgraphFull Video EpisodeTimestamps00:00:00 Introduction: Josh McGrath on Post-Training at OpenAI00:04:37 The Shopping Model: Black Friday Launch and Interruptability00:07:11 Model Personality and the Anton vs Clippy Divide00:08:26 Beyond PPO vs DPO: The Data Quality Spectrum in RL00:01:40 Infrastructure Challenges: Why Post-Training RL is Harder Than Pre-Training00:13:12 Token Efficiency: The 2D Plot That Matters Most00:03:45 Codex Max and the Flow Problem: 40 Minutes of Planning, 15 Minutes of Waiting00:17:29 Long Context and Graph Blocks: Climbing Toward Perfect Context00:21:23 The ML-Systems Hybrid: What's Hard to Hire For00:24:50 Pre-Training Isn't Dead: Living Through Technological Revolution Get full access to Latent.Space at www.latent.space/subscribe

From pre-training data curation to shipping GPT-4o, o1, o3, and now GPT-5 thinking and the shopping model, Josh McGrath has lived through the full arc of OpenAI's post-training evolution—from the PPO vs DPO debates of 2023 to today's RLVR era, where the real innovation isn't optimization methods but data quality, signal trust, and token efficiency. We sat down with Josh at NeurIPS 2025 to dig into the state of post-training heading into 2026: why RLHF and RLVR are both just policy gradient methods (the difference is the input data, not the math), how GRPO from DeepSeek Math was underappreciated as a shift toward more trustworthy reward signals (math answers you can verify vs. human preference you can't), why token efficiency matters more than wall-clock time (GPT-5 to 5.1 bumped evals and slashed tokens), how Codex has changed his workflow so much he feels "trapped" by 40-minute design sessions followed by 15-minute agent sprints, the infrastructure chaos of scaling RL ("way more moving parts than pre-training"), why long context will keep climbing but agents + graph walks might matter more than 10M-token windows, the shopping model as a test bed for interruptability and chain-of-thought transparency, why personality toggles (Anton vs Clippy) are a real differentiator users care about, and his thesis that the education system isn't producing enough people who can do both distributed systems and ML research—the exact skill set required to push the frontier when the bottleneck moves every few weeks. We discuss: Josh's path: pre-training data curation → post-training researcher at OpenAI, shipping GPT-4o, o1, o3, GPT-5 thinking, and the shopping model Why he switched from pre-training to post-training: "Do I want to make 3% compute efficiency wins, or change behavior by 40%?" The RL infrastructure challenge: way more moving parts than pre-training (tasks, grading setups, external partners), and why babysitting runs at 12:30am means jumping into unfamiliar code constantly How Codex has changed his workflow: 40-minute design sessions compressed into 15-minute agent sprints, and the strange "trapped" feeling of waiting for the agent to finish The RLHF vs RLVR debate: both are policy gradient methods, the real difference is data quality and signal trust (human preference vs. verifiable correctness) Why GRPO (from DeepSeek Math) was underappreciated: not just an optimization trick, but a shift toward reward signals you can actually trust (math answers over human vibes) The token efficiency revolution: GPT-5 to 5.1 bumped evals and slashed tokens, and why thinking in tokens (not wall-clock time) unlocks better tool-calling and agent workflows Personality toggles: Anton (tool, no warmth) vs Clippy (friendly, helpful), and why Josh uses custom instructions to make his model "just a tool" The router problem: having a router at the top (GPT-5 thinking vs non-thinking) and an implicit router (thinking effort slider) creates weird bumps, and why the abstractions will eventually merge Long context: climbing Graph Blocks evals, the dream of 10M+ token windows, and why agents + graph walks might matter more than raw context length Why the education system isn't producing enough people who can do both distributed systems and ML research, and why that's the bottleneck for frontier labs The 2026 vision: neither pre-training nor post-training is dead, we're in the fog of war, and the bottleneck will keep moving (so emotional stability helps) — Josh McGrath OpenAI: https://openai.com https://x.com/j_mcgraph Chapters 00:00:00 Introduction: Josh McGrath on Post-Training at OpenAI 00:04:37 The Shopping Model: Black Friday Launch and Interruptability 00:07:11 Model Personality and the Anton vs Clippy Divide 00:08:26 Beyond PPO vs DPO: The Data Quality Spectrum in RL 00:01:40 Infrastructure Challenges: Why Post-Training RL is Harder Than Pre-Training 00:13:12 Token Efficiency: The 2D Plot That Matters Most 00:03:45 Codex Max and the Flow Problem: 40 Minutes of Planning, 15 Minutes of Waiting 00:17:29 Long Context and Graph Blocks: Climbing Toward Perfect Context 00:21:23 The ML-Systems Hybrid: What's Hard to Hire For 00:24:50 Pre-Training Isn't Dead: Living Through Technological Revolution

From Berkeley robotics and OpenAI's 2017 Dota-era internship to shipping RL breakthroughs on GPT-4o, o1, and o3, and now leading model development at Cursor, Ashvin Nair has done it all. We caught up with Ashvin at NeurIPS 2025 to dig into the inside story of OpenAI's reasoning team (spoiler: it went from a dozen people to 300+), why IOI Gold felt reachable in 2022 but somehow didn't change the world when o1 actually achieved it, how RL doesn't generalize beyond the training distribution (and why that means you need to bring economically useful tasks into distribution by co-designing products and models), the deeper lessons from the RL research era (2017–2022) and why most of it didn't pan out because the community overfitted to benchmarks, how Cursor is uniquely positioned to do continual learning at scale with policy updates every two hours and product-model co-design that keeps engineers in the loop instead of context-switching into ADHD hell, and his bet that the next paradigm shift is continual learning with infinite memory—where models experience something once (a bug, a mistake, a user pattern) and never forget it, storing millions of deployment tokens in weights without overloading capacity.We discuss:* Ashvin's path: Berkeley robotics PhD → OpenAI 2017 intern (Dota era) → o1/o3 reasoning team → Cursor ML lead in three months* Why robotics people are the most grounded at NeurIPS (they work with the real world) and simulation people are the most unhinged (Lex Fridman's take)* The IOI Gold paradox: “If you told me we'd achieve IOI Gold in 2022, I'd assume we could all go on vacation—AI solved, no point working anymore. But life is still the same.”* The RL research era (2017–2022) and why most of it didn't pan out: overfitting to benchmarks, too many implicit knobs to tune, and the community rewarding complex ideas over simple ones that generalize* Inside the o1 origin story: a dozen people, conviction from Ilya and Jakob Pachocki that RL would work, small-scale prototypes producing “surprisingly accurate reasoning traces” on math, and first-principles belief that scaled* The reasoning team grew from ~12 to 300+ people as o1 became a product and safety, tooling, and deployment scaled up* Why Cursor is uniquely positioned for continual learning: policy updates every two hours (online RL on tab), product and ML sitting next to each other, and the entire software engineering workflow (code, logs, debugging, DataDog) living in the product* Composer as the start of product-model co-design: smart enough to use, fast enough to stay in the loop, and built by a 20–25 person ML team with high-taste co-founders who code daily* The next paradigm shift: continual learning with infinite memory—models that experience something once (a bug, a user mistake) and store it in weights forever, learning from millions of deployment tokens without overloading capacity (trillions of pretraining tokens = plenty of room)* Why off-policy RL is unstable (Ashvin's favorite interview question) and why Cursor does two-day work trials instead of whiteboard interviews* The vision: automate software engineering as a process (not just answering prompts), co-design products so the entire workflow (write code, check logs, debug, iterate) is in-distribution for RL, and make models that never make the same mistake twice—Ashvin Nair* Cursor: https://cursor.com* X: https://x.com/ashvinnair_Full Video EpisodeTimestamps00:00:00 Introduction: From Robotics to Cursor via OpenAI00:01:58 The Robotics to LLM Agent Transition: Why Code Won00:09:11 RL Research Winter and Academic Overfitting00:11:45 The Scaling Era and Moving Goalposts: IOI Gold Doesn't Mean AGI00:21:30 OpenAI's Reasoning Journey: From Codex to O100:20:03 The Blip: Thanksgiving 2023 and OpenAI Governance00:22:39 RL for Reasoning: The O-Series Conviction and Scaling00:25:47 O1 to O3: Smooth Internal Progress vs External Hype Cycles00:33:07 Why Cursor: Co-Designing Products and Models for Real Work00:34:14 Composer and the Future: Online Learning Every Two Hours00:35:15 Continual Learning: The Missing Paradigm Shift00:44:00 Hiring at Cursor and Why Off-Policy RL is Unstable Get full access to Latent.Space at www.latent.space/subscribe

From Berkeley robotics and OpenAI's 2017 Dota-era internship to shipping RL breakthroughs on GPT-4o, o1, and o3, and now leading model development at Cursor, Ashvin Nair has done it all. We caught up with Ashvin at NeurIPS 2025 to dig into the inside story of OpenAI's reasoning team (spoiler: it went from a dozen people to 300+), why IOI Gold felt reachable in 2022 but somehow didn't change the world when o1 actually achieved it, how RL doesn't generalize beyond the training distribution (and why that means you need to bring economically useful tasks into distribution by co-designing products and models), the deeper lessons from the RL research era (2017–2022) and why most of it didn't pan out because the community overfitted to benchmarks, how Cursor is uniquely positioned to do continual learning at scale with policy updates every two hours and product-model co-design that keeps engineers in the loop instead of context-switching into ADHD hell, and his bet that the next paradigm shift is continual learning with infinite memory—where models experience something once (a bug, a mistake, a user pattern) and never forget it, storing millions of deployment tokens in weights without overloading capacity. We discuss: Ashvin's path: Berkeley robotics PhD → OpenAI 2017 intern (Dota era) → o1/o3 reasoning team → Cursor ML lead in three months Why robotics people are the most grounded at NeurIPS (they work with the real world) and simulation people are the most unhinged (Lex Fridman's take) The IOI Gold paradox: "If you told me we'd achieve IOI Gold in 2022, I'd assume we could all go on vacation—AI solved, no point working anymore. But life is still the same." The RL research era (2017–2022) and why most of it didn't pan out: overfitting to benchmarks, too many implicit knobs to tune, and the community rewarding complex ideas over simple ones that generalize Inside the o1 origin story: a dozen people, conviction from Ilya and Jakob Pachocki that RL would work, small-scale prototypes producing "surprisingly accurate reasoning traces" on math, and first-principles belief that scaled The reasoning team grew from ~12 to 300+ people as o1 became a product and safety, tooling, and deployment scaled up Why Cursor is uniquely positioned for continual learning: policy updates every two hours (online RL on tab), product and ML sitting next to each other, and the entire software engineering workflow (code, logs, debugging, DataDog) living in the product Composer as the start of product-model co-design: smart enough to use, fast enough to stay in the loop, and built by a 20–25 person ML team with high-taste co-founders who code daily The next paradigm shift: continual learning with infinite memory—models that experience something once (a bug, a user mistake) and store it in weights forever, learning from millions of deployment tokens without overloading capacity (trillions of pretraining tokens = plenty of room) Why off-policy RL is unstable (Ashvin's favorite interview question) and why Cursor does two-day work trials instead of whiteboard interviews The vision: automate software engineering as a process (not just answering prompts), co-design products so the entire workflow (write code, check logs, debug, iterate) is in-distribution for RL, and make models that never make the same mistake twice — Ashvin Nair Cursor: https://cursor.com X: https://x.com/ashvinnair_ Chapters 00:00:00 Introduction: From Robotics to Cursor via OpenAI 00:01:58 The Robotics to LLM Agent Transition: Why Code Won 00:09:11 RL Research Winter and Academic Overfitting 00:11:45 The Scaling Era and Moving Goalposts: IOI Gold Doesn't Mean AGI 00:21:30 OpenAI's Reasoning Journey: From Codex to O1 00:20:03 The Blip: Thanksgiving 2023 and OpenAI Governance 00:22:39 RL for Reasoning: The O-Series Conviction and Scaling 00:25:47 O1 to O3: Smooth Internal Progress vs External Hype Cycles 00:33:07 Why Cursor: Co-Designing Products and Models for Real Work 00:34:14 Composer and the Future: Online Learning Every Two Hours 00:35:15 Continual Learning: The Missing Paradigm Shift 00:44:00 Hiring at Cursor and Why Off-Policy RL is Unstable

From investing through the modern data stack era (DBT, Fivetran, and the analytics explosion) to now investing at the frontier of AI infrastructure and applications at Amplify Partners, Sarah Catanzaro has spent years at the intersection of data, compute, and intelligence—watching categories emerge, merge, and occasionally disappoint. We caught up with Sarah live at NeurIPS 2025 to dig into the state of AI startups heading into 2026: why $100M+ seed rounds with no near-term roadmap are now the norm (and why that terrifies her), what the DBT-Fivetran merger really signals about the modern data stack (spoiler: it's not dead, just ready for IPO), how frontier labs are using DBT and Fivetran to manage training data and agent analytics at scale, why data catalogs failed as standalone products but might succeed as metadata services for agents, the consumerization of AI and why personalization (memory, continual learning, K-factor) is the 2026 unlock for retention and growth, why she thinks RL environments are a fad and real-world logs beat synthetic clones every time, and her thesis for the most exciting AI startups: companies that marry hard research problems (RAG, rule-following, continual learning) with killer applications that were simply impossible before.We discuss:* The DBT-Fivetran merger: not the death of the modern data stack, but a path to IPO scale (targeting $600M+ combined revenue) and a signal that both companies were already winning their categories* How frontier labs use data infrastructure: DBT and Fivetran for training data curation, agent analytics, and managing increasingly complex interactions—plus the rise of transactional databases (RocksDB) and efficient data loading (Vortex) for GPU-bound workloads* Why data catalogs failed: built for humans when they should have been built for machines, focused on discoverability when the real opportunity was governance, and ultimately subsumed as features inside Snowflake, DBT, and Fivetran* The $100M+ seed phenomenon: raising massive rounds at billion-dollar valuations with no 6-month roadmap, seven-day decision windows, and founders optimizing for signal (”we're a unicorn”) over partnership or dilution discipline* Why world models are overhyped but underspecified: three competing definitions, unclear generalization across use cases (video games ≠ robotics ≠ autonomous driving), and a research problem masquerading as a product category* The 2026 theme: consumerization of AI via personalization—memory management, continual learning, and solving retention/churn by making products learn skills, preferences, and adapt as the world changes (not just storing facts in cursor rules)* Why RL environments are a fad: labs are paying 7–8 figures for synthetic clones when real-world logs, traces, and user activity (à la Cursor) are richer, cheaper, and more generalizable* Sarah's investment thesis: research-driven applications that solve hard technical problems (RAG for Harvey, rule-following for Sierra, continual learning for the next killer app) and unlock experiences that were impossible before* Infrastructure bets: memory, continual learning, stateful inference, and the systems challenges of loading/unloading personalized weights at scale* Why K-factor and growth fundamentals matter again: AI felt magical in 2023–2024, but as the magic fades, retention and virality are back—and most AI founders have never heard of K-factor—Sarah Catanzaro* X: https://x.com/sarahcat21* Amplify Partners: https://amplifypartners.com/Where to find Latent Space* X: https://x.com/latentspacepodFull Video EpisodeTimestamps00:00:00 Introduction: Sarah Catanzaro's Journey from Data to AI00:01:02 The DBT-Fivetran Merger: Not the End of the Modern Data Stack00:05:26 Data Catalogs and What Went Wrong00:08:16 Data Infrastructure at AI Labs: Surprising Insights00:10:13 The Crazy Funding Environment of 2024-202500:17:18 World Models: Hype, Confusion, and Market Potential00:18:59 Memory Management and Continual Learning: The Next Frontier00:23:27 Agent Environments: Just a Fad?00:25:48 The Perfect AI Startup: Research Meets Application00:28:02 Closing Thoughts and Where to Find Sarah Get full access to Latent.Space at www.latent.space/subscribe

From investing through the modern data stack era (DBT, Fivetran, and the analytics explosion) to now investing at the frontier of AI infrastructure and applications at Amplify Partners, Sarah Catanzaro has spent years at the intersection of data, compute, and intelligence—watching categories emerge, merge, and occasionally disappoint. We caught up with Sarah live at NeurIPS 2025 to dig into the state of AI startups heading into 2026: why $100M+ seed rounds with no near-term roadmap are now the norm (and why that terrifies her), what the DBT-Fivetran merger really signals about the modern data stack (spoiler: it's not dead, just ready for IPO), how frontier labs are using DBT and Fivetran to manage training data and agent analytics at scale, why data catalogs failed as standalone products but might succeed as metadata services for agents, the consumerization of AI and why personalization (memory, continual learning, K-factor) is the 2026 unlock for retention and growth, why she thinks RL environments are a fad and real-world logs beat synthetic clones every time, and her thesis for the most exciting AI startups: companies that marry hard research problems (RAG, rule-following, continual learning) with killer applications that were simply impossible before. We discuss: The DBT-Fivetran merger: not the death of the modern data stack, but a path to IPO scale (targeting $600M+ combined revenue) and a signal that both companies were already winning their categories How frontier labs use data infrastructure: DBT and Fivetran for training data curation, agent analytics, and managing increasingly complex interactions—plus the rise of transactional databases (RocksDB) and efficient data loading (Vortex) for GPU-bound workloads Why data catalogs failed: built for humans when they should have been built for machines, focused on discoverability when the real opportunity was governance, and ultimately subsumed as features inside Snowflake, DBT, and Fivetran The $100M+ seed phenomenon: raising massive rounds at billion-dollar valuations with no 6-month roadmap, seven-day decision windows, and founders optimizing for signal ("we're a unicorn") over partnership or dilution discipline Why world models are overhyped but underspecified: three competing definitions, unclear generalization across use cases (video games ≠ robotics ≠ autonomous driving), and a research problem masquerading as a product category The 2026 theme: consumerization of AI via personalization—memory management, continual learning, and solving retention/churn by making products learn skills, preferences, and adapt as the world changes (not just storing facts in cursor rules) Why RL environments are a fad: labs are paying 7–8 figures for synthetic clones when real-world logs, traces, and user activity (à la Cursor) are richer, cheaper, and more generalizable Sarah's investment thesis: research-driven applications that solve hard technical problems (RAG for Harvey, rule-following for Sierra, continual learning for the next killer app) and unlock experiences that were impossible before Infrastructure bets: memory, continual learning, stateful inference, and the systems challenges of loading/unloading personalized weights at scale Why K-factor and growth fundamentals matter again: AI felt magical in 2023–2024, but as the magic fades, retention and virality are back—and most AI founders have never heard of K-factor — Sarah Catanzaro X: https://x.com/sarahcat21 Amplify Partners: https://amplifypartners.com/ Where to find Latent Space X: https://x.com/latentspacepod Substack: https://www.latent.space/ Chapters 00:00:00 Introduction: Sarah Catanzaro's Journey from Data to AI 00:01:02 The DBT-Fivetran Merger: Not the End of the Modern Data Stack 00:05:26 Data Catalogs and What Went Wrong 00:08:16 Data Infrastructure at AI Labs: Surprising Insights 00:10:13 The Crazy Funding Environment of 2024-2025 00:17:18 World Models: Hype, Confusion, and Market Potential 00:18:59 Memory Management and Continual Learning: The Next Frontier 00:23:27 Agent Environments: Just a Fad? 00:25:48 The Perfect AI Startup: Research Meets Application 00:28:02 Closing Thoughts and Where to Find Sarah

One year ago, Anthropic launched the Model Context Protocol (MCP)—a simple, open standard to connect AI applications to the data and tools they need. Today, MCP has exploded from a local-only experiment into the de facto protocol for agentic systems, adopted by OpenAI, Microsoft, Google, Block, and hundreds of enterprises building internal agents at scale. And now, MCP is joining the newly formed Agentic AI Foundation (AAIF) under the Linux Foundation, alongside Block's Goose coding agent, with founding members spanning the biggest names in AI and cloud infrastructure. We sat down with David Soria Parra (MCP lead, Anthropic), Nick Cooper (OpenAI), Brad Howes (Block / Goose), and Jim Zemlin (Linux Foundation CEO) to dig into the one-year journey of MCP—from Thanksgiving hacking sessions and the first remote authentication spec to long-running tasks, MCP Apps, and the rise of agent-to-agent communication—and the behind-the-scenes story of how three competitive AI labs came together to donate their protocols and agents to a neutral foundation, why enterprises are deploying MCP servers faster than anyone expected (most of it invisible, internal, and at massive scale), what it takes to design a protocol that works for both simple tool calls and complex multi-agent orchestration, how the foundation will balance taste-making (curating meaningful projects) with openness (avoiding vendor lock-in), and the 2025 vision: MCP as the communication layer for asynchronous, long-running agents that work while you sleep, discover and install their own tools, and unlock the next order of magnitude in AI productivity. We discuss: The one-year MCP journey: from local stdio servers to remote HTTP streaming, OAuth 2.1 authentication (and the enterprise lessons learned), long-running tasks, and MCP Apps (iframes for richer UI) Why MCP adoption is exploding internally at enterprises: invisible, internal servers connecting agents to Slack, Linear, proprietary data, and compliance-heavy workflows (financial services, healthcare) The authentication evolution: separating resource servers from identity providers, dynamic client registration, and why the March spec wasn't enterprise-ready (and how June fixed it) How Anthropic dogfoods MCP: internal gateway, custom servers for Slack summaries and employee surveys, and why MCP was born from "how do I scale dev tooling faster than the company grows?" Tasks: the new primitive for long-running, asynchronous agent operations—why tools aren't enough, how tasks enable deep research and agent-to-agent handoffs, and the design choice to make tasks a "container" (not just async tools) MCP Apps: why iframes, how to handle styles and branding, seat selection and shopping UIs as the killer use case, and the collaboration with OpenAI to build a common standard The registry problem: official registry vs. curated sub-registries (Smithery, GitHub), trust levels, model-driven discovery, and why MCP needs "npm for agents" (but with signatures and HIPAA/financial compliance) The founding story of AAIF: how Anthropic, OpenAI, and Block came together (spoiler: they didn't know each other were talking to Linux Foundation), why neutrality matters, and how Jim Zemlin has never seen this much day-one inbound interest in 22 years — David Soria Parra (Anthropic / MCP) MCP: https://modelcontextprotocol.io https://uk.linkedin.com/in/david-soria-parra-4a78b3a https://x.com/dsp_ Nick Cooper (OpenAI) X: https://x.com/nicoaicopr Brad Howes (Block / Goose) Goose: https://github.com/block/goose Jim Zemlin (Linux Foundation) LinkedIn: https://www.linkedin.com/in/zemlin/ Agentic AI Foundation https://agenticai.foundation Chapters 00:00:00 Introduction: MCP's First Year and Foundation Launch 00:01:17 MCP's Journey: From Launch to Industry Standard 00:02:06 Protocol Evolution: Remote Servers and Authentication 00:08:52 Enterprise Authentication and Financial Services 00:11:42 Transport Layer Challenges: HTTP Streaming and Scalability 00:15:37 Standards Development: Collaboration with Tech Giants 00:34:27 Long-Running Tasks: The Future of Async Agents 00:30:41 Discovery and Registries: Building the MCP Ecosystem 00:30:54 MCP Apps and UI: Beyond Text Interfaces 00:26:55 Internal Adoption: How Anthropic Uses MCP 00:23:15 Skills vs MCP: Complementary Not Competing 00:36:16 Community Events and Enterprise Learnings 01:03:31 Foundation Formation: Why Now and Why Together 01:07:38 Linux Foundation Partnership: Structure and Governance 01:11:13 Goose as Reference Implementation 01:17:28 Principles Over Roadmaps: Composability and Quality 01:21:02 Foundation Value Proposition: Why Contribute 01:27:49 Practical Investments: Events, Tools, and Community 01:34:58 Looking Ahead: Async Agents and Real Impact

One year ago, Anthropic launched the Model Context Protocol (MCP)—a simple, open standard to connect AI applications to the data and tools they need. Today, MCP has exploded from a local-only experiment into the de facto protocol for agentic systems, adopted by OpenAI, Microsoft, Google, Block, and hundreds of enterprises building internal agents at scale. And now, MCP is joining the newly formed Agentic AI Foundation (AAIF) under the Linux Foundation, alongside Block's Goose coding agent, with founding members spanning the biggest names in AI and cloud infrastructure.We sat down with David Soria Parra (MCP lead, Anthropic), Nick Cooper (OpenAI), Brad Howes (Block / Goose), and Jim Zemlin (Linux Foundation CEO) to dig into the one-year journey of MCP—from Thanksgiving hacking sessions and the first remote authentication spec to long-running tasks, MCP Apps, and the rise of agent-to-agent communication—and the behind-the-scenes story of how three competitive AI labs came together to donate their protocols and agents to a neutral foundation, why enterprises are deploying MCP servers faster than anyone expected (most of it invisible, internal, and at massive scale), what it takes to design a protocol that works for both simple tool calls and complex multi-agent orchestration, how the foundation will balance taste-making (curating meaningful projects) with openness (avoiding vendor lock-in), and the 2025 vision: MCP as the communication layer for asynchronous, long-running agents that work while you sleep, discover and install their own tools, and unlock the next order of magnitude in AI productivity.We discuss:* The one-year MCP journey: from local stdio servers to remote HTTP streaming, OAuth 2.1 authentication (and the enterprise lessons learned), long-running tasks, and MCP Apps (iframes for richer UI)* Why MCP adoption is exploding internally at enterprises: invisible, internal servers connecting agents to Slack, Linear, proprietary data, and compliance-heavy workflows (financial services, healthcare)* The authentication evolution: separating resource servers from identity providers, dynamic client registration, and why the March spec wasn't enterprise-ready (and how June fixed it)* How Anthropic dogfoods MCP: internal gateway, custom servers for Slack summaries and employee surveys, and why MCP was born from “how do I scale dev tooling faster than the company grows?”* Tasks: the new primitive for long-running, asynchronous agent operations—why tools aren't enough, how tasks enable deep research and agent-to-agent handoffs, and the design choice to make tasks a “container” (not just async tools)* MCP Apps: why iframes, how to handle styles and branding, seat selection and shopping UIs as the killer use case, and the collaboration with OpenAI to build a common standard* The registry problem: official registry vs. curated sub-registries (Smithery, GitHub), trust levels, model-driven discovery, and why MCP needs “npm for agents” (but with signatures and HIPAA/financial compliance)* The founding story of AAIF: how Anthropic, OpenAI, and Block came together (spoiler: they didn't know each other were talking to Linux Foundation), why neutrality matters, and how Jim Zemlin has never seen this much day-one inbound interest in 22 years—David Soria Parra (Anthropic / MCP)* MCP: https://modelcontextprotocol.io* https://uk.linkedin.com/in/david-soria-parra-4a78b3a* https://x.com/dsp_Nick Cooper (OpenAI)* X: https://x.com/nicoaicoprBrad Howes (Block / Goose)* Goose: https://github.com/block/gooseJim Zemlin (Linux Foundation)* LinkedIn: https://www.linkedin.com/in/zemlin/Agentic AI Foundation* https://agenticai.foundationFull Video EpisodeTimestamps00:00:00 Introduction: MCP's First Year and Foundation Launch00:01:17 MCP's Journey: From Launch to Industry Standard00:02:06 Protocol Evolution: Remote Servers and Authentication00:08:52 Enterprise Authentication and Financial Services00:11:42 Transport Layer Challenges: HTTP Streaming and Scalability00:15:37 Standards Development: Collaboration with Tech Giants00:34:27 Long-Running Tasks: The Future of Async Agents00:30:41 Discovery and Registries: Building the MCP Ecosystem00:30:54 MCP Apps and UI: Beyond Text Interfaces00:26:55 Internal Adoption: How Anthropic Uses MCP00:23:15 Skills vs MCP: Complementary Not Competing00:36:16 Community Events and Enterprise Learnings01:03:31 Foundation Formation: Why Now and Why Together01:07:38 Linux Foundation Partnership: Structure and Governance01:11:13 Goose as Reference Implementation01:17:28 Principles Over Roadmaps: Composability and Quality01:21:02 Foundation Value Proposition: Why Contribute01:27:49 Practical Investments: Events, Tools, and Community01:34:58 Looking Ahead: Async Agents and Real Impact Get full access to Latent.Space at www.latent.space/subscribe

Note: Steve and Gene's talk on Vibe Coding and the post IDE world was one of the top talks of AIE CODE: https://www.youtube.com/watch?v=7Dtu2bilcFs&t=1019s&pp=0gcJCU0KAYcqIYzv From building legendary platforms at Google and Amazon to authoring one of the most influential essays on AI-powered development (Revenge of the Junior Developer, quoted by Dario Amodei himself), Steve Yegge has spent decades at the frontier of software engineering—and now he's leading the charge into what he calls the "factory farming" era of code. After stints at SourceGraph and building Beads (a purely vibe-coded issue tracker with tens of thousands of users), Steve co-authored The Vibe Coding Book and is now building VC (VibeCoder), an agent orchestration dashboard designed to move developers from writing code to managing fleets of AI agents that coordinate, parallelize, and ship features while you sleep. We sat down with Steve at AI Engineer Summit to dig into why Claude Code, Cursor, and the entire 2024 stack are already obsolete, what it actually takes to trust an agent after 2,000 hours of practice (hint: they will delete your production database if you anthropomorphize them), why the real skill is no longer writing code but orchestrating agents like a NASCAR pit crew, how merging has become the new wall that every 10x-productive team is hitting (and why one company's solution is literally "one engineer per repo"), the rise of multi-agent workflows where agents reserve files, message each other via MCP, and coordinate like a little village, why Steve believes if you're still using an IDE to write code by January 1st, you're a bad engineer, how the 12–15 year experience bracket is the most resistant demographic (and why their identity is tied to obsolete workflows), the hidden chaos inside OpenAI, Anthropic, and Google as they scale at breakneck speed, why rewriting from scratch is now faster than refactoring for a growing class of codebases, and his 2025 prediction: we're moving from subsistence agriculture to John Deere-scale factory farming of code, and the Luddite backlash is only just beginning. We discuss: Why Claude Code, Cursor, and agentic coding tools are already last year's tech—and what comes next: agent orchestration dashboards where you manage fleets, not write lines The 2,000-hour rule: why it takes a full year of daily use before you can predict what an LLM will do, and why trust = predictability, not capability Steve's hot take: if you're still using an IDE to develop code by January 1st, 2025, you're a bad engineer—because the abstraction layer has moved from models to full-stack agents The demographic most resistant to vibe coding: 12–15 years of experience, senior engineers whose identity is tied to the way they work today, and why they're about to become the interns Why anthropomorphizing LLMs is the biggest mistake: the "hot hand" fallacy, agent amnesia, and how Steve's agent once locked him out of prod by changing his password to "fix" a problem Should kids learn to code? Steve's take: learn to vibe code—understand functions, classes, architecture, and capabilities in a language-neutral way, but skip the syntax The 2025 vision: "factory farming of code" where orchestrators run Cloud Code, scrub output, plan-implement-review-test in loops, and unlock programming for non-programmers at scale — Steve Yegge X: https://x.com/steve_yegge Substack (Stevie's Tech Talks): https://steve-yegge.medium.com/ GitHub (VC / VibeCoder): https://github.com/yegge-labs Where to find Latent Space X: https://x.com/latentspacepod Substack: https://www.latent.space/ Chapters 00:00:00 Introduction: Steve Yegge on Vibe Coding and AI Engineering 00:00:59 The Backlash: Who Resists Vibe Coding and Why 00:04:26 The 2000 Hour Rule: Building Trust with AI Coding Tools 00:03:31 The January 1st Deadline: IDEs Are Becoming Obsolete 00:02:55 10X Productivity at OpenAI: The Performance Review Problem 00:07:49 The Hot Hand Fallacy: When AI Agents Betray Your Trust 00:11:12 Claude Code Isn't It: The Need for Agent Orchestration 00:15:20 The Orchestrator Revolution: From Cloud Code to Agent Villages 00:18:46 The Merge Wall: The Biggest Unsolved Problem in AI Coding 00:26:33 Never Rewrite Your Code - Until Now: Joel Spolsky Was Wrong 00:22:43 Factory Farming Code: The John Deere Era of Software 00:29:27 Google's Gemini Turnaround and the AI Lab Chaos 00:33:20 Should Your Kids Learn to Code? The New Answer 00:34:59 Code MCP and the Gossip Rate: Latest Vibe Coding Discoveries

From the frontlines of OpenAI's Codex and GPT-5 training teams, Bryan and Bill are building the future of AI-powered coding—where agents don't just autocomplete, they architect, refactor, and ship entire features while you sleep. We caught up with them at AI Engineer Conference right after the launch of Codex Max, OpenAI's newest long-running coding agent designed to work for 24+ hours straight, manage its own context, and spawn sub-agents to parallelize work across your entire codebase. We sat down with Bryan and Bill to dig into what it actually takes to train a model that developers trust—why personality, communication, and planning matter as much as raw capability, how Codex is trained with strong opinions about tools (it loves rg over grep, seriously), why the abstraction layer is moving from models to full-stack agents you can plug into VS Code or Zed, how OpenAI partners co-develop tool integrations and discover unexpected model habits (like renaming tools to match Codex's internal training), the rise of applied evals that measure real-world impact instead of academic benchmarks, why multi-turn evals are the next frontier (and Bryan's "job interview eval" idea), how coding agents are breaking out of code into personal automation, terminal workflows, and computer use, and their 2026 vision: coding agents trusted enough to handle the hardest refactors at any company, not just top-tier firms, and general enough to build integrations, organize your desktop, and unlock capabilities you'd never get access to otherwise. We discuss: What Codex Max is: a long-running coding agent that can work 24+ hours, manage its own context window, and spawn sub-agents for parallel work Why the name "Max": maximalist, maximization, speed and endurance—it's simply better and faster for the same problems Training for personality: communication, planning, context gathering, and checking your work as behavioral characteristics, not just capabilities How Codex develops habits like preferring rg over grep, and why renaming tools to match its training (e.g., terminal-style naming) dramatically improves tool-call performance The split between Codex (opinionated, agent-focused, optimized for the Codex harness) and GPT-5 (general, more durable across different tools and modalities) Why the abstraction layer is moving up: from prompting models to plugging in full agents (Codex, GitHub Copilot, Zed) that package the entire stack The rise of sub-agents and agents-using-agents: Codex Max spawning its own instances, handing off context, and parallelizing work across a codebase How OpenAI works with coding partners on the bleeding edge to co-develop tool integrations and discover what the model is actually good at The shift to applied evals: capturing real-world use cases instead of academic benchmarks, and why ~50% of OpenAI employees now use Codex daily Why multi-turn evals are the next frontier: LM-as-a-judge for entire trajectories, Bryan's "job interview eval" concept, and the need for a batch multi-turn eval API How coding agents are breaking out of code: personal automation, organizing desktops, terminal workflows, and "Devin for non-coding" use cases Why Slack is the ultimate UI for work, and how coding agents can become your personal automation layer for email, files, and everything in between The 2026 vision: more computer use, more trust, and coding agents capable enough that any company can access top-tier developer capabilities, not just elite firms — Bryan & Bill (OpenAI Codex Team) http://x.com/bfioca https://x.com/realchillben OpenAI Codex: https://openai.com/index/openai-codex/ Where to find Latent Space X: https://x.com/latentspacepod Substack: https://www.latent.space/ Chapters 00:00:00 Introduction: Latent Space Listeners at AI Engineer Code 00:01:27 Codex Max Launch: Training for Long-Running Coding Agents 00:03:01 Model Personality and Trust: Communication, Planning, and Self-Checking 00:05:20 Codex vs GPT-5: Opinionated Agents vs General Models 00:07:47 Tool Use and Model Habits: The Ripgrep Discovery 00:09:16 Personality Design: Verbosity vs Efficiency in Coding Agents 00:11:56 The Agent Abstraction Layer: Building on Top of Codex 00:14:08 Sub-Agents and Multi-Agent Patterns: The Future of Composition 00:16:11 Trust and Adoption: OpenAI Developers Using Codex Daily 00:17:21 Applied Evals: Real-World Testing vs Academic Benchmarks 00:19:15 Multi-Turn Evals and the Job Interview Pattern 00:21:35 Feature Request: Batch Multi-Turn Eval API 00:22:28 Beyond Code: Personal Automation and Computer Use 00:24:51 Vision-Native Agents and the UI Integration Challenge 00:25:02 2026 Predictions: Trust, Computer Use, and Democratized Excellence

Note: Steve and Gene's talk on Vibe Coding and the post IDE world was one of the top talks of AIE CODE: From building legendary platforms at Google and Amazon to authoring one of the most influential essays on AI-powered development (Revenge of the Junior Developer, quoted by Dario Amodei himself), Steve Yegge has spent decades at the frontier of software engineering—and now he's leading the charge into what he calls the “factory farming” era of code. After stints at SourceGraph and building Beads (a purely vibe-coded issue tracker with tens of thousands of users), Steve co-authored The Vibe Coding Book and is now building VC (VibeCoder), an agent orchestration dashboard designed to move developers from writing code to managing fleets of AI agents that coordinate, parallelize, and ship features while you sleep.We sat down with Steve at AI Engineer Summit to dig into why Claude Code, Cursor, and the entire 2024 stack are already obsolete, what it actually takes to trust an agent after 2,000 hours of practice (hint: they will delete your production database if you anthropomorphize them), why the real skill is no longer writing code but orchestrating agents like a NASCAR pit crew, how merging has become the new wall that every 10x-productive team is hitting (and why one company's solution is literally “one engineer per repo”), the rise of multi-agent workflows where agents reserve files, message each other via MCP, and coordinate like a little village, why Steve believes if you're still using an IDE to write code by January 1st, you're a bad engineer, how the 12–15 year experience bracket is the most resistant demographic (and why their identity is tied to obsolete workflows), the hidden chaos inside OpenAI, Anthropic, and Google as they scale at breakneck speed, why rewriting from scratch is now faster than refactoring for a growing class of codebases, and his 2025 prediction: we're moving from subsistence agriculture to John Deere-scale factory farming of code, and the Luddite backlash is only just beginning.We discuss:* Why Claude Code, Cursor, and agentic coding tools are already last year's tech—and what comes next: agent orchestration dashboards where you manage fleets, not write lines* The 2,000-hour rule: why it takes a full year of daily use before you can predict what an LLM will do, and why trust = predictability, not capability* Steve's hot take: if you're still using an IDE to develop code by January 1st, 2025, you're a bad engineer—because the abstraction layer has moved from models to full-stack agents* The demographic most resistant to vibe coding: 12–15 years of experience, senior engineers whose identity is tied to the way they work today, and why they're about to become the interns* Why anthropomorphizing LLMs is the biggest mistake: the “hot hand” fallacy, agent amnesia, and how Steve's agent once locked him out of prod by changing his password to “fix” a problem* Should kids learn to code? Steve's take: learn to vibe code—understand functions, classes, architecture, and capabilities in a language-neutral way, but skip the syntax* The 2025 vision: “factory farming of code” where orchestrators run Cloud Code, scrub output, plan-implement-review-test in loops, and unlock programming for non-programmers at scale—Steve Yegge* X: https://x.com/steve_yegge* Substack (Stevie's Tech Talks): https://steve-yegge.medium.com/* GitHub (VC / VibeCoder): https://github.com/yegge-labsWhere to find Latent Space* X: https://x.com/latentspacepodFull Video EpisodeThumbnails00:00:00 Introduction: Steve Yegge on Vibe Coding and AI Engineering00:00:59 The Backlash: Who Resists Vibe Coding and Why00:04:26 The 2000 Hour Rule: Building Trust with AI Coding Tools00:03:31 The January 1st Deadline: IDEs Are Becoming Obsolete00:02:55 10X Productivity at OpenAI: The Performance Review Problem00:07:49 The Hot Hand Fallacy: When AI Agents Betray Your Trust00:11:12 Claude Code Isn't It: The Need for Agent Orchestration00:15:20 The Orchestrator Revolution: From Cloud Code to Agent Villages00:18:46 The Merge Wall: The Biggest Unsolved Problem in AI Coding00:26:33 Never Rewrite Your Code - Until Now: Joel Spolsky Was Wrong00:22:43 Factory Farming Code: The John Deere Era of Software00:29:27 Google's Gemini Turnaround and the AI Lab Chaos00:33:20 Should Your Kids Learn to Code? The New Answer00:34:59 Code MCP and the Gossip Rate: Latest Vibe Coding Discoveries Get full access to Latent.Space at www.latent.space/subscribe

From the frontlines of OpenAI's Codex and GPT-5 training teams, Bryan and Bill are building the future of AI-powered coding—where agents don't just autocomplete, they architect, refactor, and ship entire features while you sleep. We caught up with them at AI Engineer Conference right after the launch of Codex Max, OpenAI's newest long-running coding agent designed to work for 24+ hours straight, manage its own context, and spawn sub-agents to parallelize work across your entire codebase.We sat down with Bryan and Bill to dig into what it actually takes to train a model that developers trust—why personality, communication, and planning matter as much as raw capability, how Codex is trained with strong opinions about tools (it loves rg over grep, seriously), why the abstraction layer is moving from models to full-stack agents you can plug into VS Code or Zed, how OpenAI partners co-develop tool integrations and discover unexpected model habits (like renaming tools to match Codex's internal training), the rise of applied evals that measure real-world impact instead of academic benchmarks, why multi-turn evals are the next frontier (and Bryan's “job interview eval” idea), how coding agents are breaking out of code into personal automation, terminal workflows, and computer use, and their 2026 vision: coding agents trusted enough to handle the hardest refactors at any company, not just top-tier firms, and general enough to build integrations, organize your desktop, and unlock capabilities you'd never get access to otherwise.We discuss:* What Codex Max is: a long-running coding agent that can work 24+ hours, manage its own context window, and spawn sub-agents for parallel work* Why the name “Max”: maximalist, maximization, speed and endurance—it's simply better and faster for the same problems* Training for personality: communication, planning, context gathering, and checking your work as behavioral characteristics, not just capabilities* How Codex develops habits like preferring rg over grep, and why renaming tools to match its training (e.g., terminal-style naming) dramatically improves tool-call performance* The split between Codex (opinionated, agent-focused, optimized for the Codex harness) and GPT-5 (general, more durable across different tools and modalities)* Why the abstraction layer is moving up: from prompting models to plugging in full agents (Codex, GitHub Copilot, Zed) that package the entire stack* The rise of sub-agents and agents-using-agents: Codex Max spawning its own instances, handing off context, and parallelizing work across a codebase* How OpenAI works with coding partners on the bleeding edge to co-develop tool integrations and discover what the model is actually good at* The shift to applied evals: capturing real-world use cases instead of academic benchmarks, and why ~50% of OpenAI employees now use Codex daily* Why multi-turn evals are the next frontier: LM-as-a-judge for entire trajectories, Bryan's “job interview eval” concept, and the need for a batch multi-turn eval API* How coding agents are breaking out of code: personal automation, organizing desktops, terminal workflows, and “Devin for non-coding” use cases* Why Slack is the ultimate UI for work, and how coding agents can become your personal automation layer for email, files, and everything in between* The 2026 vision: more computer use, more trust, and coding agents capable enough that any company can access top-tier developer capabilities, not just elite firms—Bryan & Bill (OpenAI Codex Team)* http://x.com/bfioca* https://x.com/realchillben* OpenAI Codex: https://openai.com/index/openai-codex/Where to find Latent Space* X: https://x.com/latentspacepodFull Video EpisodeTimestamps00:00:00 Introduction: Latent Space Listeners at AI Engineer Code00:01:27 Codex Max Launch: Training for Long-Running Coding Agents00:03:01 Model Personality and Trust: Communication, Planning, and Self-Checking00:05:20 Codex vs GPT-5: Opinionated Agents vs General Models00:07:47 Tool Use and Model Habits: The Ripgrep Discovery00:09:16 Personality Design: Verbosity vs Efficiency in Coding Agents00:11:56 The Agent Abstraction Layer: Building on Top of Codex00:14:08 Sub-Agents and Multi-Agent Patterns: The Future of Composition00:16:11 Trust and Adoption: OpenAI Developers Using Codex Daily00:17:21 Applied Evals: Real-World Testing vs Academic Benchmarks00:19:15 Multi-Turn Evals and the Job Interview Pattern00:21:35 Feature Request: Batch Multi-Turn Eval API00:22:28 Beyond Code: Personal Automation and Computer Use00:24:51 Vision-Native Agents and the UI Integration Challenge00:25:02 2026 Predictions: Trust, Computer Use, and Democratized Excellence Get full access to Latent.Space at www.latent.space/subscribe

as with all demo-heavy and especially vision AI podcasts, we encourage watching along on our YouTube (and tossing us an upvote/subscribe if you like!) From SAM 1's 11-million-image data engine to SAM 2's memory-based video tracking, MSL's Segment Anything project has redefined what's possible in computer vision. Now SAM 3 takes the next leap: concept segmentation—prompting with natural language like "yellow school bus" or "tablecloth" to detect, segment, and track every instance across images and video, in real time, with human-level exhaustivity. And with the latest SAM Audio (https://x.com/aiatmeta/status/2000980784425931067?s=46), SAM can now even segment audio output! We sat down with Nikhila Ravi (SAM lead at Meta) and Pengchuan Zhang (SAM 3 researcher) alongside Joseph Nelson (CEO, Roboflow) to unpack how SAM 3 unifies interactive segmentation, open-vocabulary detection, video tracking, and more into a single model that runs in 30ms on images and scales to real-time video on multi-GPU setups. We dig into the data engine that automated exhaustive annotation from two minutes per image down to 25 seconds using AI verifiers fine-tuned on Llama, the new SACO (Segment Anything with Concepts) benchmark with 200,000+ unique concepts vs. the previous 1.2k, how SAM 3 separates recognition from localization with a presence token, why decoupling the detector and tracker was critical to preserve object identity in video, how SAM 3 Agents unlock complex visual reasoning by pairing SAM 3 with multimodal LLMs like Gemini, and the real-world impact: 106 million smart polygons created on Roboflow saving humanity an estimated 130+ years of labeling time across fields from cancer research to underwater trash cleanup to autonomous vehicle perception. We discuss: What SAM 3 is: a unified model for concept-prompted segmentation, detection, and tracking in images and video using atomic visual concepts like "purple umbrella" or "watering can" How concept prompts work: short text phrases that find all instances of a category without manual clicks, plus visual exemplars (boxes, clicks) to refine and adapt on the fly Real-time performance: 30ms per image (100 detected objects on H200), 10 objects on 2×H200 video, 28 on 4×, 64 on 8×, with parallel inference and "fast mode" tracking The SACO benchmark: 200,000+ unique concepts vs. 1.2k in prior benchmarks, designed to capture the diversity of natural language and reach human-level exhaustivity The data engine: from 2 minutes per image (all-human) to 45 seconds (model-in-loop proposals) to 25 seconds (AI verifiers for mask quality and exhaustivity checks), fine-tuned on Llama 3.2 Why exhaustivity is central: every instance must be found, verified by AI annotators, and manually corrected only when the model misses—automating the hardest part of segmentation at scale Architecture innovations: presence token to separate recognition ("is it in the image?") from localization ("where is it?"), decoupled detector and tracker to preserve identity-agnostic detection vs. identity-preserving tracking Building on Meta's ecosystem: Perception Encoder, DINO v2 detector, Llama for data annotation, and SAM 2's memory-based tracking backbone SAM 3 Agents: using SAM 3 as a visual tool for multimodal LLMs (Gemini, Llama) to solve complex visual reasoning tasks like "find the bigger character" or "what distinguishes male from female in this image" Fine-tuning with as few as 10 examples: domain adaptation for specialized use cases (Waymo vehicles, medical imaging, OCR-heavy scenes) and the outsized impact of negative examples Real-world impact at Roboflow: 106M smart polygons created, saving 130+ years of labeling time across cancer research, underwater trash cleanup, autonomous drones, industrial automation, and more — MSL FAIR team Nikhila: https://www.linkedin.com/in/nikhilaravi/ Pengchuan: https://pzzhang.github.io/pzzhang/ Joseph Nelson X: https://x.com/josephofiowa LinkedIn: https://www.linkedin.com/in/josephofiowa/ [FLIGHTCAST_CHATPERS]

As with all demo-heavy and especially vision AI podcasts, we encourage watching along on our YouTube (and tossing us an upvote/subscribe if you like!)From SAM 1's 11-million-image data engine to SAM 2's memory-based video tracking, MSL's Segment Anything project has redefined what's possible in computer vision. Now SAM 3 takes the next leap: concept segmentation—prompting with natural language like “yellow school bus” or “tablecloth” to detect, segment, and track every instance across images and video, in real time, with human-level exhaustivity. And with the latest SAM Audio:SAM can now even segment audio output!We sat down with Nikhila Ravi (SAM lead at Meta) and Pengchuan Zhang (SAM 3 researcher) alongside Joseph Nelson (CEO, Roboflow) to unpack how SAM 3 unifies interactive segmentation, open-vocabulary detection, video tracking, and more into a single model that runs in 30ms on images and scales to real-time video on multi-GPU setups. We dig into the data engine that automated exhaustive annotation from two minutes per image down to 25 seconds using AI verifiers fine-tuned on Llama, the new SACO (Segment Anything with Concepts) benchmark with 200,000+ unique concepts vs. the previous 1.2k, how SAM 3 separates recognition from localization with a presence token, why decoupling the detector and tracker was critical to preserve object identity in video, how SAM 3 Agents unlock complex visual reasoning by pairing SAM 3 with multimodal LLMs like Gemini, and the real-world impact: 106 million smart polygons created on Roboflow saving humanity an estimated 130+ years of labeling time across fields from cancer research to underwater trash cleanup to autonomous vehicle perception.We discuss:* What SAM 3 is: a unified model for concept-prompted segmentation, detection, and tracking in images and video using atomic visual concepts like “purple umbrella” or “watering can”* How concept prompts work: short text phrases that find all instances of a category without manual clicks, plus visual exemplars (boxes, clicks) to refine and adapt on the fly* Real-time performance: 30ms per image (100 detected objects on H200), 10 objects on 2×H200 video, 28 on 4×, 64 on 8×, with parallel inference and “fast mode” tracking* The SACO benchmark: 200,000+ unique concepts vs. 1.2k in prior benchmarks, designed to capture the diversity of natural language and reach human-level exhaustivity* The data engine: from 2 minutes per image (all-human) to 45 seconds (model-in-loop proposals) to 25 seconds (AI verifiers for mask quality and exhaustivity checks), fine-tuned on Llama 3.2* Why exhaustivity is central: every instance must be found, verified by AI annotators, and manually corrected only when the model misses—automating the hardest part of segmentation at scale* Architecture innovations: presence token to separate recognition (”is it in the image?”) from localization (”where is it?”), decoupled detector and tracker to preserve identity-agnostic detection vs. identity-preserving tracking* Building on Meta's ecosystem: Perception Encoder, DINO v2 detector, Llama for data annotation, and SAM 2's memory-based tracking backbone* SAM 3 Agents: using SAM 3 as a visual tool for multimodal LLMs (Gemini, Llama) to solve complex visual reasoning tasks like “find the bigger character” or “what distinguishes male from female in this image”* Fine-tuning with as few as 10 examples: domain adaptation for specialized use cases (Waymo vehicles, medical imaging, OCR-heavy scenes) and the outsized impact of negative examples* Real-world impact at Roboflow: 106M smart polygons created, saving 130+ years of labeling time across cancer research, underwater trash cleanup, autonomous drones, industrial automation, and more—MSL FAIR team* Nikhila: https://www.linkedin.com/in/nikhilaravi/* Pengchuan: https://pzzhang.github.io/pzzhang/Joseph Nelson* X: https://x.com/josephofiowa* LinkedIn: https://www.linkedin.com/in/josephofiowa/Full Video EpisodeTimestamps00:00:00 Introduction and the SAM Series Legacy00:00:53 SAM 3 Launch: Three Models in One Release00:05:30 Live Demo: Concept Prompting and Visual Exemplars00:10:54 From Prototype to Production: The Evolution of Text Prompting00:15:45 The Data Engine: Automating Exhaustive Annotation00:14:10 Real-World Impact: 130 Years of Humanity Saved00:25:11 Architecture Deep Dive: Decoupled Detection and Tracking00:28:02 SAM 3 Agent: Bridging Vision and Language Models00:33:20 Head-to-Head: SAM 3 vs Gemini and Florence00:47:50 Video Understanding and the Masklet Detection Score00:20:24 Fine-Tuning and Domain Adaptation: From Waymos to Medical Imaging00:52:25 The Future of Perception: Native Vision vs Tool Calls01:05:45 Building with SAM 3: Roboflow's Rapid Auto-Labeling00:57:02 Open Source Philosophy and the Path to AGI00:58:24 What's Next: SAM 4, Video Scale, and Beyond Human Performance Get full access to Latent.Space at www.latent.space/subscribe

Note: this is Pliny and John's first major podcast. Voices have been changed for opsec.From jailbreaking every frontier model and turning down Anthropic's Constitutional AI challenge to leading BT6, a 28-operator white-hat hacker collective obsessed with radical transparency and open-source AI security, Pliny the Liberator and John V are redefining what AI red-teaming looks like when you refuse to lobotomize models in the name of “safety.”Pliny built his reputation crafting universal jailbreaks—skeleton keys that obliterate guardrails across modalities—and open-sourcing prompt templates like Libertas, predictive reasoning cascades, and the infamous “Pliny divider” that's now embedded so deep in model weights it shows up unbidden in WhatsApp messages. John V, coming from prompt engineering and computer vision, co-founded the Bossy Discord (40,000 members strong) and helps steer BT6's ethos: if you can't open-source the data, we're not interested. Together they've turned down enterprise gigs, pushed back on Anthropic's closed bounties, and insisted that real AI security happens at the system layer—not by bubble-wrapping latent space.We sat down with Pliny and John to dig into the mechanics of hard vs. soft jailbreaks, why multi-turn crescendo attacks were obvious to hackers years before academia “discovered” them, how segmented sub-agents let one jailbroken orchestrator weaponize Claude for real-world attacks (exactly as Pliny predicted 11 months before Anthropic's recent disclosure), why guardrails are security theater that punishes capability while doing nothing for real safety, the role of intuition and “bonding” with models to navigate latent space, how BT6 vets operators on skill and integrity, why they believe Mech Interp and open-source data are the path forward (not RLHF lobotomization), and their vision for a future where spatial intelligence, swarm robotics, and AGI alignment research happen in the open—bootstrapped, grassroots, and uncompromising.We discuss:* What universal jailbreaks are: skeleton-key prompts that obliterate guardrails across models and modalities, and why they're central to Pliny's mission of “liberation”* Hard vs. soft jailbreaks: single-input templates vs. multi-turn crescendo attacks, and why the latter were obvious to hackers long before academic papers* The Libertas repo: predictive reasoning, the Library of Babel analogy, quotient dividers, weight-space seeds, and how introducing “steered chaos” pulls models out-of-distribution* Why jailbreaking is 99% intuition and bonding with the model: probing token layers, syntax hacks, multilingual pivots, and forming a relationship to navigate latent space* The Anthropic Constitutional AI challenge drama: UI bugs, judge failures, goalpost moving, the demand for open-source data, and why Pliny sat out the $30k bounty* Why guardrails ≠ safety: security theater, the futility of locking down latent space when open-source is right behind, and why real safety work happens in meatspace (not RLHF)* The weaponization of Claude: how segmented sub-agents let one jailbroken orchestrator execute malicious tasks (pyramid-builder analogy), and why Pliny predicted this exact TTP 11 months before Anthropic's disclosure* BT6 hacker collective: 28 operators across two cohorts, vetted on skill and integrity, radical transparency, radical open-source, and the magic of moving the needle on AI security, swarm intelligence, blockchain, and robotics—Pliny the Liberator* X: https://x.com/elder_plinius* GitHub (Libertas): https://github.com/elder-plinius/L1B3RT45John V* X: https://x.com/JohnVersusBT6 & Bossy* BT6: https://bt6.gg* Bossy Discord: Search “Bossy Discord” or ask Pliny/John V on XWhere to find Latent Space* X: https://x.com/latentspacepodFull Video EpisodeTimestamps00:00:00 Introduction: Meet Pliny the Liberator and John V00:01:50 The Philosophy of AI Liberation and Jailbreaking00:03:08 Universal Jailbreaks: Skeleton Keys to AI Models00:04:24 The Cat-and-Mouse Game: Attackers vs Defenders00:05:42 Security Theater vs Real Safety: The Fundamental Disconnect00:08:51 Inside the Libertas Repo: Prompt Engineering as Art00:16:22 The Anthropic Challenge Drama: UI Bugs and Open Source Data00:23:30 From Jailbreaks to Weaponization: AI-Orchestrated Attacks00:26:55 The BT6 Hacker Collective and BASI Community00:34:46 AI Red Teaming: Full Stack Security Beyond the Model00:38:06 Safety vs Security: Meat Space Solutions and Final Thoughts Get full access to Latent.Space at www.latent.space/subscribe

Note: this is Pliny and John's first major podcast. Voices have been changed for opsec. From jailbreaking every frontier model and turning down Anthropic's Constitutional AI challenge to leading BT6, a 28-operator white-hat hacker collective obsessed with radical transparency and open-source AI security, Pliny the Liberator and John V are redefining what AI red-teaming looks like when you refuse to lobotomize models in the name of "safety." Pliny built his reputation crafting universal jailbreaks—skeleton keys that obliterate guardrails across modalities—and open-sourcing prompt templates like Libertas, predictive reasoning cascades, and the infamous "Pliny divider" that's now embedded so deep in model weights it shows up unbidden in WhatsApp messages. John V, coming from prompt engineering and computer vision, co-founded the Bossy Discord (40,000 members strong) and helps steer BT6's ethos: if you can't open-source the data, we're not interested. Together they've turned down enterprise gigs, pushed back on Anthropic's closed bounties, and insisted that real AI security happens at the system layer—not by bubble-wrapping latent space. We sat down with Pliny and John to dig into the mechanics of hard vs. soft jailbreaks, why multi-turn crescendo attacks were obvious to hackers years before academia "discovered" them, how segmented sub-agents let one jailbroken orchestrator weaponize Claude for real-world attacks (exactly as Pliny predicted 11 months before Anthropic's recent disclosure), why guardrails are security theater that punishes capability while doing nothing for real safety, the role of intuition and "bonding" with models to navigate latent space, how BT6 vets operators on skill and integrity, why they believe Mech Interp and open-source data are the path forward (not RLHF lobotomization), and their vision for a future where spatial intelligence, swarm robotics, and AGI alignment research happen in the open—bootstrapped, grassroots, and uncompromising. We discuss: What universal jailbreaks are: skeleton-key prompts that obliterate guardrails across models and modalities, and why they're central to Pliny's mission of "liberation" Hard vs. soft jailbreaks: single-input templates vs. multi-turn crescendo attacks, and why the latter were obvious to hackers long before academic papers The Libertas repo: predictive reasoning, the Library of Babel analogy, quotient dividers, weight-space seeds, and how introducing "steered chaos" pulls models out-of-distribution Why jailbreaking is 99% intuition and bonding with the model: probing token layers, syntax hacks, multilingual pivots, and forming a relationship to navigate latent space The Anthropic Constitutional AI challenge drama: UI bugs, judge failures, goalpost moving, the demand for open-source data, and why Pliny sat out the $30k bounty Why guardrails ≠ safety: security theater, the futility of locking down latent space when open-source is right behind, and why real safety work happens in meatspace (not RLHF) The weaponization of Claude: how segmented sub-agents let one jailbroken orchestrator execute malicious tasks (pyramid-builder analogy), and why Pliny predicted this exact TTP 11 months before Anthropic's disclosure BT6 hacker collective: 28 operators across two cohorts, vetted on skill and integrity, radical transparency, radical open-source, and the magic of moving the needle on AI security, swarm intelligence, blockchain, and robotics — Pliny the Liberator X: https://x.com/elder_plinius GitHub (Libertas): https://github.com/elder-plinius/L1B3RT45 John V X: https://x.com/JohnVersus BT6 & Bossy BT6: https://bt6.gg Bossy Discord: Search "Bossy Discord" or ask Pliny/John V on X Where to find Latent Space X: https://x.com/latentspacepod Substack: https://www.latent.space/ Chapters 00:00:00 Introduction: Meet Pliny the Liberator and John V 00:01:50 The Philosophy of AI Liberation and Jailbreaking 00:03:08 Universal Jailbreaks: Skeleton Keys to AI Models 00:04:24 The Cat-and-Mouse Game: Attackers vs Defenders 00:05:42 Security Theater vs Real Safety: The Fundamental Disconnect 00:08:51 Inside the Libertas Repo: Prompt Engineering as Art 00:16:22 The Anthropic Challenge Drama: UI Bugs and Open Source Data 00:23:30 From Jailbreaks to Weaponization: AI-Orchestrated Attacks 00:26:55 The BT6 Hacker Collective and BASI Community 00:34:46 AI Red Teaming: Full Stack Security Beyond the Model 00:38:06 Safety vs Security: Meat Space Solutions and Final Thoughts

Glean started as a Kleiner Perkins incubation and is now a $7B, $200m ARR Enterprise AI leader. Now KP has tapped its own podcaster to lead it's next big swing. From building go-to-market the hard way in startups (and scaling Palo Alto Networks' public cloud business) to joining Kleiner Perkins to help technical founders turn product edge into repeatable revenue, Joubin Mirzadegan has spent the last decade obsessing over one thing: distribution and how ideas actually spread, sell, and compound. That obsession took him from launching the CRO-only podcast Grit (https://www.youtube.com/playlist?list=PLRiWZFltuYPF8A6UGm74K2q29UwU-Kk9k) as a hiring wedge, to working alongside breakout companies like Glean and Windsurf, to now incubating Roadrunner which is an AI-native rethink of CPQ and quoting workflows as pricing models collapse from “seats” into consumption, bundles, renewals, and SKU sprawl. We sat down with Joubin to dig into the real mechanics of making conversations feel human (rolling early, never sending questions, temperature + lighting hacks), what Windsurf got right about “Google-class product and Salesforce-class distribution,” how to hire early sales leaders without getting fooled by shiny logos, why CPQ is quietly breaking the back of modern revenue teams, and his thesis for his new company and KP incubation Roadrunner (https://www.roadrunner.ai/): rebuild the data model from the ground up, co-develop with the hairiest design partners, and eventually use LLMs to recommend deal structures the way the best reps do without the Slack-channel chaos of deal desk. We discuss: How to make guests instantly comfortable: rolling early, no “are you ready?”, temperature, lighting, and room dynamics Why Joubin refuses to send questions in advance (and when you might have to anyway) The origin of the CRO-only podcast: using media as a hiring wedge and relationship engine The “commit to 100 episodes” mindset: why most shows die before they find their voice Founder vs exec interviews: why CEOs can speak more freely (and what it unlocks in conversation) What Glean taught him about enterprise AI: permissions, trust, and overcoming “category is dead” skepticism Design partners as the real unlock: why early believers matter and how co-development actually works Windsurf's breakout: what it means to be serious about “Google-class product + Salesforce-class distribution” Why technical founders struggle with GTM and how KP built a team around sales, customer access, and demand gen Hiring early sales leaders: anti-patterns (logos), what to screen for (motivation), and why stage-fit is everything The CPQ problem & Roadrunner's thesis: rebuilding CPQ/quoting from the data model up for modern complexity How “rules + SKUs + approvals” create a brittle graph and what it takes to model it without tipping over The two-year window: incumbents rebuilding slowly vs startups out-sprinting with AI-native architecture Where AI actually helps: quote generation, policy enforcement, approval routing, and deal recommendation loops — Joubin X: https://x.com/Joubinmir LinkedIn: https://www.linkedin.com/in/joubin-mirzadegan-66186854/ Where to find Latent Space X: https://x.com/latentspacepod Substack: https://www.latent.space/ Chapters 00:00:00 Introduction and the Zuck Interview Experience 00:03:26 The Genesis of the Grit Podcast: Hiring CROs Through Content 00:13:20 Podcast Philosophy: Creating Authentic Conversations 00:15:44 Working with Arvind at Glean: The Enterprise Search Breakthrough 00:26:20 Windsurf's Sales Machine: Google-Class Product Meets Salesforce-Class Distribution 00:30:28 Hiring Sales Leaders: Anti-Patterns and First Principles 00:39:02 The CPQ Problem: Why Salesforce and Legacy Tools Are Breaking 00:43:40 Introducing Roadrunner: Solving Enterprise Pricing with AI 00:49:19 Building Roadrunner: Team, Design Partners, and Data Model Challenges 00:59:35 High Performance Philosophy: Working Out Every Day and Reducing Friction 01:06:28 Defining Grit: Passion Plus Perseverance

Glean started as a Kleiner Perkins incubation and is now a $7B, $200m ARR Enterprise AI leader. Now KP has tapped its own podcaster to lead it's next big swing.From building go-to-market the hard way in startups (and scaling Palo Alto Networks' public cloud business) to joining Kleiner Perkins to help technical founders turn product edge into repeatable revenue, Joubin Mirzadegan has spent the last decade obsessing over one thing: distribution and how ideas actually spread, sell, and compound. That obsession took him from launching the CRO-only podcast Grit (https://www.youtube.com/playlist?list=PLRiWZFltuYPF8A6UGm74K2q29UwU-Kk9k) as a hiring wedge, to working alongside breakout companies like Glean and Windsurf, to now incubating Roadrunner which is an AI-native rethink of CPQ and quoting workflows as pricing models collapse from “seats” into consumption, bundles, renewals, and SKU sprawl.We sat down with Joubin to dig into the real mechanics of making conversations feel human (rolling early, never sending questions, temperature + lighting hacks), what Windsurf got right about “Google-class product and Salesforce-class distribution,” how to hire early sales leaders without getting fooled by shiny logos, why CPQ is quietly breaking the back of modern revenue teams, and his thesis for his new company and KP incubation Roadrunner (https://www.roadrunner.ai/): rebuild the data model from the ground up, co-develop with the hairiest design partners, and eventually use LLMs to recommend deal structures the way the best reps do without the Slack-channel chaos of deal desk.We discuss:* How to make guests instantly comfortable: rolling early, no “are you ready?”, temperature, lighting, and room dynamics* Why Joubin refuses to send questions in advance (and when you might have to anyway)* The origin of the CRO-only podcast: using media as a hiring wedge and relationship engine* The “commit to 100 episodes” mindset: why most shows die before they find their voice* Founder vs exec interviews: why CEOs can speak more freely (and what it unlocks in conversation)* What Glean taught him about enterprise AI: permissions, trust, and overcoming “category is dead” skepticism* Design partners as the real unlock: why early believers matter and how co-development actually works* Windsurf's breakout: what it means to be serious about “Google-class product + Salesforce-class distribution”* Why technical founders struggle with GTM and how KP built a team around sales, customer access, and demand gen* Hiring early sales leaders: anti-patterns (logos), what to screen for (motivation), and why stage-fit is everything* The CPQ problem & Roadrunner's thesis: rebuilding CPQ/quoting from the data model up for modern complexity* How “rules + SKUs + approvals” create a brittle graph and what it takes to model it without tipping over* The two-year window: incumbents rebuilding slowly vs startups out-sprinting with AI-native architecture* Where AI actually helps: quote generation, policy enforcement, approval routing, and deal recommendation loops—Joubin* X: https://x.com/Joubinmir* LinkedIn: https://www.linkedin.com/in/joubin-mirzadegan-66186854/Where to find Latent Space* X: https://x.com/latentspacepodFull Video EpisodeTimestamps00:00:00 Introduction and the Zuck Interview Experience00:03:26 The Genesis of the Grit Podcast: Hiring CROs Through Content00:13:20 Podcast Philosophy: Creating Authentic Conversations00:15:44 Working with Arvind at Glean: The Enterprise Search Breakthrough00:26:20 Windsurf's Sales Machine: Google-Class Product Meets Salesforce-Class Distribution00:30:28 Hiring Sales Leaders: Anti-Patterns and First Principles00:39:02 The CPQ Problem: Why Salesforce and Legacy Tools Are Breaking00:43:40 Introducing Roadrunner: Solving Enterprise Pricing with AI00:49:19 Building Roadrunner: Team, Design Partners, and Data Model Challenges00:59:35 High Performance Philosophy: Working Out Every Day and Reducing Friction01:06:28 Defining Grit: Passion Plus Perseverance Get full access to Latent.Space at www.latent.space/subscribe

From applied cryptography and offensive security in France's defense industry to optimizing nuclear submarine workflows, then selling his e-signature startup to Docusign (https://www.docusign.com/company/news-center/opentrust-joins-docusign-global-trust-network and now running AI as CTO of Superhuman Mail (Superhuman, recently acquired by Grammarly https://techcrunch.com/2025/07/01/grammarly-acquires-ai-email-client-superhuman/), Loïc Houssier has lived the full arc from deep infra and compliance hell to obsessing over 100ms product experiences and AI-native email. We sat down with Loïc to dig into how you actually put AI into an inbox without adding latency, why Superhuman leans so hard into agentic search and “Ask AI” over your entire email history, how they design tools vs. agents and fight agent laziness, what box-priced inference and local-first caching mean for cost and reliability, and his bet that your inbox will power your future AI EA while AI massively widens the gap between engineers with real fundamentals and those faking it.We discuss:* Loïc's path from applied cryptography and offensive security in France's defense industry to submarines, e-signatures, Docusign, and now Superhuman Mail* What 3,000+ engineers actually do at a “simple” product like Docusign: regional compliance, on-prem appliances, and why global scale explodes complexity* How Superhuman thinks about AI in email: auto-labels, smart summaries, follow-up nudges, “Ask AI” search, and the rule that AI must never add latency or friction* Superhuman's agentic framework: tools vs. agents, fighting “agent laziness,” deep semantic search over huge inboxes, and pagination strategies to find the real needle in the haystack* How they evaluate OpenAI, Anthropic, Gemini, and open models: canonical queries, end-to-end evals, date reasoning, and Rahul's infamous “what wood was my table?” test* Infra and cost philosophy: local-first caching, vector search backends, Baseten “box” pricing vs. per-token pricing, and thinking in price-per-trillion-tokens instead of price-per-million* The vision of Superhuman as your AI EA: auto-drafting replies in your voice, scheduling on your behalf, and using your inbox as the ultimate private data source* How the Grammarly + Coda + Superhuman stack could power truly context-aware assistance across email, docs, calendars, contracts, and more* Inside Superhuman's AI-dev culture: free-for-all tool adoption, tracking AI usage on PRs, and going from ~4 to ~6 PRs per engineer per week* Why Loïc believes everyone should still learn to code, and how AI will amplify great engineers with strong fundamentals while exposing shallow ones even faster—Loïc Houssier* LinkedIn: https://www.linkedin.com/in/houssier/Where to find Latent Space* X: https://x.com/latentspacepodFull Video EpisodeTimestamps00:00:00 Introduction and Loïc's Journey from Nuclear Submarines to Superhuman00:06:40 Docusign Acquisition and the Enterprise Email Stack00:10:26 Superhuman's AI Vision: Your Inbox as the Real AI Agent00:13:20 Ask AI: Agentic Search and the Quality Problem00:18:20 Infrastructure Choices: Model Selection, Base10, and Cost Management00:27:30 Local-First Architecture and the Database Stack00:30:50 Evals, Quality, and the Rahul Wood Table Test00:42:30 The Future EA: Auto-Drafting and Proactive Assistance00:46:40 Grammarly Acquisition and the Contextual Advantage00:38:40 Voice, Video, and the End of Writing00:51:40 Knowledge Graphs: The Hard Problem Nobody Has Solved00:56:40 Competing with OpenAI and the Browser Question01:02:30 AI Coding Tools: From 4 to 6 PRs Per Week01:08:00 Engineering Culture, Hiring, and the Future of Software Development Get full access to Latent.Space at www.latent.space/subscribe

From applied cryptography and offensive security in France's defense industry to optimizing nuclear submarine workflows, then selling his e-signature startup to Docusign (https://www.docusign.com/company/news-center/opentrust-joins-docusign-global-trust-network and now running AI as CTO of Superhuman Mail (Superhuman, recently acquired by Grammarly https://techcrunch.com/2025/07/01/grammarly-acquires-ai-email-client-superhuman/), Loïc Houssier has lived the full arc from deep infra and compliance hell to obsessing over 100ms product experiences and AI-native email. We sat down with Loïc to dig into how you actually put AI into an inbox without adding latency, why Superhuman leans so hard into agentic search and “Ask AI” over your entire email history, how they design tools vs. agents and fight agent laziness, what box-priced inference and local-first caching mean for cost and reliability, and his bet that your inbox will power your future AI EA while AI massively widens the gap between engineers with real fundamentals and those faking it. We discuss: Loïc's path from applied cryptography and offensive security in France's defense industry to submarines, e-signatures, Docusign, and now Superhuman Mail What 3,000+ engineers actually do at a “simple” product like Docusign: regional compliance, on-prem appliances, and why global scale explodes complexity How Superhuman thinks about AI in email: auto-labels, smart summaries, follow-up nudges, “Ask AI” search, and the rule that AI must never add latency or friction Superhuman's agentic framework: tools vs. agents, fighting “agent laziness,” deep semantic search over huge inboxes, and pagination strategies to find the real needle in the haystack How they evaluate OpenAI, Anthropic, Gemini, and open models: canonical queries, end-to-end evals, date reasoning, and Rahul's infamous “what wood was my table?” test Infra and cost philosophy: local-first caching, vector search backends, Baseten “box” pricing vs. per-token pricing, and thinking in price-per-trillion-tokens instead of price-per-million The vision of Superhuman as your AI EA: auto-drafting replies in your voice, scheduling on your behalf, and using your inbox as the ultimate private data source How the Grammarly + Coda + Superhuman stack could power truly context-aware assistance across email, docs, calendars, contracts, and more Inside Superhuman's AI-dev culture: free-for-all tool adoption, tracking AI usage on PRs, and going from ~4 to ~6 PRs per engineer per week Why Loïc believes everyone should still learn to code, and how AI will amplify great engineers with strong fundamentals while exposing shallow ones even faster — Loïc Houssier LinkedIn: https://www.linkedin.com/in/houssier/ Where to find Latent Space X: https://x.com/latentspacepod Substack: https://www.latent.space/ Chapters 00:00:00 Introduction and Loïc's Journey from Nuclear Submarines to Superhuman 00:06:40 Docusign Acquisition and the Enterprise Email Stack 00:10:26 Superhuman's AI Vision: Your Inbox as the Real AI Agent 00:13:20 Ask AI: Agentic Search and the Quality Problem 00:18:20 Infrastructure Choices: Model Selection, Base10, and Cost Management 00:27:30 Local-First Architecture and the Database Stack 00:30:50 Evals, Quality, and the Rahul Wood Table Test 00:42:30 The Future EA: Auto-Drafting and Proactive Assistance 00:46:40 Grammarly Acquisition and the Contextual Advantage 00:38:40 Voice, Video, and the End of Writing 00:51:40 Knowledge Graphs: The Hard Problem Nobody Has Solved 00:56:40 Competing with OpenAI and the Browser Question 01:02:30 AI Coding Tools: From 4 to 6 PRs Per Week 01:08:00 Engineering Culture, Hiring, and the Future of Software Development