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

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

From building Medal into a 12M-user game clipping platform with 3.8B highlight moments to turning down a reported $500M offer from OpenAI (https://www.theinformation.com/articles/openai-offered-pay-500-million-startup-videogame-data) and raising a $134M seed from Khosla (https://techcrunch.com/2025/10/16/general-intuition-lands-134m-seed-to-teach-agents-spatial-reasoning-using-video-game-clips/) to spin out General Intuition, Pim is betting that world models trained on peak human gameplay are the next frontier after LLMs.We sat down with Pim to dig into why game highlights are “episodic memory for simulation” (and how Medal's privacy-first action labels became a world-model goldmine https://medal.tv/blog/posts/enabling-state-of-the-art-security-and-protections-on-medals-new-apm-and-controller-overlay-features), what it takes to build fully vision-based agents that just see frames and output actions in real time, how General Intuition transfers from games to real-world video and then into robotics, why world models and LLMs are complementary rather than rivals, what founders with proprietary datasets should know before selling or licensing to labs, and his bet that spatial-temporal foundation models will power 80% of future atoms-to-atoms interactions in both simulation and the real world.We discuss:* How Medal's 3.8B action-labeled highlight clips became a privacy-preserving goldmine for world models* Building fully vision-based agents that only see frames and output actions yet play like (and sometimes better than) humans* Transferring from arcade-style games to realistic games to real-world video using the same perception–action recipe* Why world models need actions, memory, and partial observability (smoke, occlusion, camera shake) vs. “just” pretty video generation* Distilling giant policies into tiny real-time models that still navigate, hide, and peek corners like real players* Pim's path from RuneScape private servers, Tourette's, and reverse engineering to leading a frontier world-model lab* How data-rich founders should think about valuing their datasets, negotiating with big labs, and deciding when to go independent* GI's first customers: replacing brittle behavior trees in games, engines, and controller-based robots with a “frames in, actions out” API* Using Medal clips as “episodic memory of simulation” to move from imitation learning to RL via world models and negative events* The 2030 vision: spatial–temporal foundation models that power the majority of atoms-to-atoms interactions in simulation and the real world—Pim* X: https://x.com/PimDeWitte* LinkedIn: https://www.linkedin.com/in/pimdw/Where to find Latent Space* X: https://x.com/latentspacepodFull Video EpisodeTimestamps00:00:00 Introduction and Medal's Gaming Data Advantage00:02:08 Exclusive Demo: Vision-Based Gaming Agents00:06:17 Action Prediction and Real-World Video Transfer00:08:41 World Models: Interactive Video Generation00:13:42 From Runescape to AI: Pim's Founder Journey00:16:45 The Research Foundations: Diamond, Genie, and SEMA00:33:03 Vinod Khosla's Largest Seed Bet Since OpenAI00:35:04 Data Moats and Why GI Stayed Independent00:38:42 Self-Teaching AI Fundamentals: The Francois Fleuret Course00:40:28 Defining World Models vs Video Generation00:41:52 Why Simulation Complexity Favors World Models00:43:30 World Labs, Yann LeCun, and the Spatial Intelligence Race00:50:08 Business Model: APIs, Agents, and Game Developer Partnerships00:58:57 From Imitation Learning to RL: Making Clips Playable01:00:15 Open Research, Academic Partnerships, and Hiring01:02:09 2030 Vision: 80 Percent of Atoms-to-Atoms AI Interactions Get full access to Latent.Space at www.latent.space/subscribe

From building Medal into a 12M-user game clipping platform with 3.8B highlight moments to turning down a reported $500M offer from OpenAI (https://www.theinformation.com/articles/openai-offered-pay-500-million-startup-videogame-data) and raising a $134M seed from Khosla (https://techcrunch.com/2025/10/16/general-intuition-lands-134m-seed-to-teach-agents-spatial-reasoning-using-video-game-clips/) to spin out General Intuition, Pim is betting that world models trained on peak human gameplay are the next frontier after LLMs. We sat down with Pim to dig into why game highlights are “episodic memory for simulation” (and how Medal's privacy-first action labels became a world-model goldmine https://medal.tv/blog/posts/enabling-state-of-the-art-security-and-protections-on-medals-new-apm-and-controller-overlay-features), what it takes to build fully vision-based agents that just see frames and output actions in real time, how General Intuition transfers from games to real-world video and then into robotics, why world models and LLMs are complementary rather than rivals, what founders with proprietary datasets should know before selling or licensing to labs, and his bet that spatial-temporal foundation models will power 80% of future atoms-to-atoms interactions in both simulation and the real world. We discuss: How Medal's 3.8B action-labeled highlight clips became a privacy-preserving goldmine for world models Building fully vision-based agents that only see frames and output actions yet play like (and sometimes better than) humans Transferring from arcade-style games to realistic games to real-world video using the same perception–action recipe Why world models need actions, memory, and partial observability (smoke, occlusion, camera shake) vs. “just” pretty video generation Distilling giant policies into tiny real-time models that still navigate, hide, and peek corners like real players Pim's path from RuneScape private servers, Tourette's, and reverse engineering to leading a frontier world-model lab How data-rich founders should think about valuing their datasets, negotiating with big labs, and deciding when to go independent GI's first customers: replacing brittle behavior trees in games, engines, and controller-based robots with a “frames in, actions out” API Using Medal clips as “episodic memory of simulation” to move from imitation learning to RL via world models and negative events The 2030 vision: spatial–temporal foundation models that power the majority of atoms-to-atoms interactions in simulation and the real world — Pim X: https://x.com/PimDeWitte LinkedIn: https://www.linkedin.com/in/pimdw/ Where to find Latent Space X: https://x.com/latentspacepod Substack: https://www.latent.space/ Chapters 00:00:00 Introduction and Medal's Gaming Data Advantage 00:02:08 Exclusive Demo: Vision-Based Gaming Agents 00:06:17 Action Prediction and Real-World Video Transfer 00:08:41 World Models: Interactive Video Generation 00:13:42 From Runescape to AI: Pim's Founder Journey 00:16:45 The Research Foundations: Diamond, Genie, and SEMA 00:33:03 Vinod Khosla's Largest Seed Bet Since OpenAI 00:35:04 Data Moats and Why GI Stayed Independent 00:38:42 Self-Teaching AI Fundamentals: The Francois Fleuret Course 00:40:28 Defining World Models vs Video Generation 00:41:52 Why Simulation Complexity Favors World Models 00:43:30 World Labs, Yann LeCun, and the Spatial Intelligence Race 00:50:08 Business Model: APIs, Agents, and Game Developer Partnerships 00:58:57 From Imitation Learning to RL: Making Clips Playable 01:00:15 Open Research, Academic Partnerships, and Hiring 01:02:09 2030 Vision: 80 Percent of Atoms-to-Atoms AI Interactions

Fei-Fei Li and Justin Johnson are cofounders of World Labs, who have recently launched Marble (https://marble.worldlabs.ai/), a new kind of generative “world model” that can create editable 3D environments from text, images, and other spatial inputs. Marble lets creators generate persistent 3D worlds, precisely control cameras, and interactively edit scenes, making it a powerful tool for games, film, VR, robotics simulation, and more. In this episode, Fei-Fei and Justin share how their journey from ImageNet and Stanford research led to World Labs, why spatial intelligence is the next frontier after LLMs, and how world models could change how machines see, understand, and build in 3D. We discuss: The massive compute scaling from AlexNet to today and why world models and spatial data are the most compelling way to “soak up” modern GPU clusters compared to language alone. What Marble actually is: a generative model of 3D worlds that turns text and images into editable scenes using Gaussian splats, supports precise camera control and recording, and runs interactively on phones, laptops, and VR headsets. Fei-fei's essay (https://drfeifei.substack.com/p/from-words-to-worlds-spatial-intelligence) on spatial intelligence as a distinct form of intelligence from language: from picking up a mug to inferring the 3D structure of DNA, and why language is a lossy, low-bandwidth channel for describing the rich 3D/4D world we live in. Whether current models “understand” physics or just fit patterns: the gap between predicting orbits and discovering F=ma, and how attaching physical properties to splats and distilling physics engines into neural networks could lead to genuine causal reasoning. The changing role of academia in AI, why Fei-Fei worries more about under-resourced universities than “open vs closed,” and how initiatives like national AI compute clouds and open benchmarks can rebalance the ecosystem. Why transformers are fundamentally set models, not sequence models, and how that perspective opens up new architectures for world models, especially as hardware shifts from single GPUs to massive distributed clusters. Real use cases for Marble today: previsualization and VFX, game environments, virtual production, interior and architectural design (including kitchen remodels), and generating synthetic simulation worlds for training embodied agents and robots. How spatial intelligence and language intelligence will work together in multimodal systems, and why the goal isn't to throw away LLMs but to complement them with rich, embodied models of the world. Fei-Fei and Justin's long-term vision for spatial intelligence: from creative tools for artists and game devs to broader applications in science, medicine, and real-world decision-making. — Fei-Fei Li X: https://x.com/drfeifei LinkedIn: https://www.linkedin.com/in/fei-fei-li-4541247 Justin Johnson X: https://x.com/jcjohnss LinkedIn: https://www.linkedin.com/in/justin-johnson-41b43664 Where to find Latent Space X: https://x.com/latentspacepod Substack: https://www.latent.space/ Chapters 00:00:00 Introduction and the Fei-Fei Li & Justin Johnson Partnership 00:02:00 From ImageNet to World Models: The Evolution of Computer Vision 00:12:42 Dense Captioning and Early Vision-Language Work 00:19:57 Spatial Intelligence: Beyond Language Models 00:28:46 Introducing Marble: World Labs' First Spatial Intelligence Model 00:33:21 Gaussian Splats and the Technical Architecture of Marble 00:22:10 Physics, Dynamics, and the Future of World Models 00:41:09 Multimodality and the Interplay of Language and Space 00:37:37 Use Cases: From Creative Industries to Robotics and Embodied AI 00:56:58 Hiring, Research Directions, and the Future of World Labs

Fei-Fei Li and Justin Johnson are cofounders of World Labs, who have recently launched Marble (https://marble.worldlabs.ai/), a new kind of generative “world model” that can create editable 3D environments from text, images, and other spatial inputs. Marble lets creators generate persistent 3D worlds, precisely control cameras, and interactively edit scenes, making it a powerful tool for games, film, VR, robotics simulation, and more. In this episode, Fei-Fei and Justin share how their journey from ImageNet and Stanford research led to World Labs, why spatial intelligence is the next frontier after LLMs, and how world models could change how machines see, understand, and build in 3D.We discuss:* The massive compute scaling from AlexNet to today and why world models and spatial data are the most compelling way to “soak up” modern GPU clusters compared to language alone.* What Marble actually is: a generative model of 3D worlds that turns text and images into editable scenes using Gaussian splats, supports precise camera control and recording, and runs interactively on phones, laptops, and VR headsets.* Fei-fei's essay:on spatial intelligence as a distinct form of intelligence from language: from picking up a mug to inferring the 3D structure of DNA, and why language is a lossy, low-bandwidth channel for describing the rich 3D/4D world we live in.* Whether current models “understand” physics or just fit patterns: the gap between predicting orbits and discovering F=ma, and how attaching physical properties to splats and distilling physics engines into neural networks could lead to genuine causal reasoning.* The changing role of academia in AI, why Fei-Fei worries more about under-resourced universities than “open vs closed,” and how initiatives like national AI compute clouds and open benchmarks can rebalance the ecosystem.* Why transformers are fundamentally set models, not sequence models, and how that perspective opens up new architectures for world models, especially as hardware shifts from single GPUs to massive distributed clusters.* Real use cases for Marble today: previsualization and VFX, game environments, virtual production, interior and architectural design (including kitchen remodels), and generating synthetic simulation worlds for training embodied agents and robots.* How spatial intelligence and language intelligence will work together in multimodal systems, and why the goal isn't to throw away LLMs but to complement them with rich, embodied models of the world.* Fei-Fei and Justin's long-term vision for spatial intelligence: from creative tools for artists and game devs to broader applications in science, medicine, and real-world decision-making.—Fei-Fei Li* X: https://x.com/drfeifei* LinkedIn: https://www.linkedin.com/in/fei-fei-li-4541247Justin Johnson* X: https://x.com/jcjohnss* LinkedIn: https://www.linkedin.com/in/justin-johnson-41b43664Where to find Latent Space* X: https://x.com/latentspacepodFull Video EpisodeTimestamps00:00:00 Introduction and the Fei-Fei Li & Justin Johnson Partnership00:02:00 From ImageNet to World Models: The Evolution of Computer Vision00:12:42 Dense Captioning and Early Vision-Language Work00:19:57 Spatial Intelligence: Beyond Language Models00:28:46 Introducing Marble: World Labs' First Spatial Intelligence Model00:33:21 Gaussian Splats and the Technical Architecture of Marble00:22:10 Physics, Dynamics, and the Future of World Models00:41:09 Multimodality and the Interplay of Language and Space00:37:37 Use Cases: From Creative Industries to Robotics and Embodied AI00:56:58 Hiring, Research Directions, and the Future of World Labs Get full access to Latent.Space at www.latent.space/subscribe

Alex Lieberman and Arman Hezarkani, co-founders of Tenex, reveal how they're revolutionizing software consulting by compensating AI engineers for output rather than hours—enabling some engineers to earn over $1 million annually while delivering 10x productivity gains. Their company represents a fundamental rethinking of knowledge work compensation in the age of AI agents, where traditional hourly billing models perversely incentivize slower work even as AI tools enable unprecedented speed. The Genesis: From 90% Downsizing to 10x Output The story behind 10X begins with Arman's previous company, Parthian, where he was forced to downsize his engineering team by 90%. Rather than collapse, Arman re-architected the entire product and engineering process to be AI-first—and discovered that production-ready software output increased 10x despite the massive headcount reduction. This counterintuitive result exposed a fundamental misalignment: engineers compensated by the hour are disincentivized from leveraging AI to work faster, even when the technology enables dramatic productivity gains. Alex, who had invested in Parthian, initially didn't believe the numbers until Arman walked him through why LLMs have made such a profound impact specifically on engineering as knowledge work. The Economic Model: Story Points Over Hours 10X's core innovation is compensating engineers based on story points—units of completed, quality output—rather than hours worked. This creates direct economic incentives for engineers to adopt every new AI tool, optimize their workflows, and maximize throughput. The company expects multiple engineers to earn over $1 million in cash compensation next year purely from story point earnings. To prevent gaming the system, they hire for two profiles: engineers who are "long-term selfish" (understanding that inflating story points will destroy client relationships) and those who genuinely love writing code and working with smart people. They also employ technical strategists incentivized on client retention (NRR) who serve as the final quality gate before any engineering plan reaches a client. Impressive Builds: From Retail AI to App Store Hits The results speak for themselves. In one project, 10X built a computer vision system for retail cameras that provides heat maps, queue detection, shelf stocking analysis, and theft detection—creating early prototypes in just two weeks for work that previously took quarters. They built Snapback Sports' mobile trivia app in one month, which hit 20th globally on the App Store. In a sales context, an engineer spent four hours building a working prototype of a fitness influencer's AI health coach app after the prospect initially said no—immediately moving 10X to the top of their vendor list. These examples demonstrate how AI-enabled speed fundamentally changes sales motions and product development timelines. The Interview Process: Unreasonably Difficult Take-Homes Despite concerns that AI would make take-home assessments obsolete, 10X still uses them—but makes them "unreasonably difficult." About 50% of candidates don't even respond, but those who complete the challenge demonstrate the caliber needed. The interview process is remarkably short: two calls before the take-home, review, then one or two final meetings—completable in as little as a week. A signature question: "If you had infinite resources to build an AI that could replace either of us on this call, what would be the first major bottleneck?" The sophisticated answer isn't just "model intelligence" or "context length"—it's controlling entropy, the accumulating error rate that derails autonomous agents over time. The Limiting Factor: Human Capital, Not Technology Despite being an AI-first company, 10X's primary constraint is human capital—finding and hiring enough exceptional engineers fast enough, then matching them with the right processes to maintain delivery quality as they scale. The company has ambitions beyond consulting to build their own technology, but for the foreseeable future, recruiting remains the bottleneck. This reveals an important insight about the AI era: even as technology enables unprecedented leverage, the constraint shifts to finding people who can harness that leverage effectively. Chapters 00:00:00 Introduction and Meeting the 10X Co-founders 00:01:29 The 10X Moment: From Hourly Billing to Output-Based Compensation 00:04:44 The Economic Model Behind 10X 00:05:42 Story Points and Measuring Engineering Output 00:08:41 Impressive Client Projects and Rapid Prototyping 00:12:22 The 10X Tech Stack: TypeScript and High Structure 00:13:21 AI Coding Tools: The Daily Evolution 00:15:05 Human Capital as the Limiting Factor 00:16:02 The Unreasonably Difficult Interview Process 00:17:14 Entropy and Context Engineering: The Future of AI Agents 00:23:28 The MCP Debate and AI Industry Sociology 00:26:01 Consulting, Digital Transformation, and Conference Insights

Alex Lieberman and Arman Hezarkani, co-founders of Tenex, reveal how they're revolutionizing software consulting by compensating AI engineers for output rather than hours—enabling some engineers to earn over $1 million annually while delivering 10x productivity gains. Their company represents a fundamental rethinking of knowledge work compensation in the age of AI agents, where traditional hourly billing models perversely incentivize slower work even as AI tools enable unprecedented speed.The Genesis: From 90% Downsizing to 10x Output The story behind 10X begins with Arman's previous company, Parthian, where he was forced to downsize his engineering team by 90%. Rather than collapse, Arman re-architected the entire product and engineering process to be AI-first—and discovered that production-ready software output increased 10x despite the massive headcount reduction. This counterintuitive result exposed a fundamental misalignment: engineers compensated by the hour are disincentivized from leveraging AI to work faster, even when the technology enables dramatic productivity gains. Alex, who had invested in Parthian, initially didn't believe the numbers until Arman walked him through why LLMs have made such a profound impact specifically on engineering as knowledge work.The Economic Model: Story Points Over Hours 10X's core innovation is compensating engineers based on story points—units of completed, quality output—rather than hours worked. This creates direct economic incentives for engineers to adopt every new AI tool, optimize their workflows, and maximize throughput. The company expects multiple engineers to earn over $1 million in cash compensation next year purely from story point earnings. To prevent gaming the system, they hire for two profiles: engineers who are “long-term selfish” (understanding that inflating story points will destroy client relationships) and those who genuinely love writing code and working with smart people. They also employ technical strategists incentivized on client retention (NRR) who serve as the final quality gate before any engineering plan reaches a client.Impressive Builds: From Retail AI to App Store Hits The results speak for themselves. In one project, 10X built a computer vision system for retail cameras that provides heat maps, queue detection, shelf stocking analysis, and theft detection—creating early prototypes in just two weeks for work that previously took quarters. They built Snapback Sports' mobile trivia app in one month, which hit 20th globally on the App Store. In a sales context, an engineer spent four hours building a working prototype of a fitness influencer's AI health coach app after the prospect initially said no—immediately moving 10X to the top of their vendor list. These examples demonstrate how AI-enabled speed fundamentally changes sales motions and product development timelines.The Interview Process: Unreasonably Difficult Take-Homes Despite concerns that AI would make take-home assessments obsolete, 10X still uses them—but makes them “unreasonably difficult.” About 50% of candidates don't even respond, but those who complete the challenge demonstrate the caliber needed. The interview process is remarkably short: two calls before the take-home, review, then one or two final meetings—completable in as little as a week. A signature question: “If you had infinite resources to build an AI that could replace either of us on this call, what would be the first major bottleneck?” The sophisticated answer isn't just “model intelligence” or “context length”—it's controlling entropy, the accumulating error rate that derails autonomous agents over time.The Limiting Factor: Human Capital, Not Technology Despite being an AI-first company, 10X's primary constraint is human capital—finding and hiring enough exceptional engineers fast enough, then matching them with the right processes to maintain delivery quality as they scale. The company has ambitions beyond consulting to build their own technology, but for the foreseeable future, recruiting remains the bottleneck. This reveals an important insight about the AI era: even as technology enables unprecedented leverage, the constraint shifts to finding people who can harness that leverage effectively.Full Video EpisodeTimestamps00:00:00 Introduction and Meeting the 10X Co-founders00:01:29 The 10X Moment: From Hourly Billing to Output-Based Compensation00:04:44 The Economic Model Behind 10X00:05:42 Story Points and Measuring Engineering Output00:08:41 Impressive Client Projects and Rapid Prototyping00:12:22 The 10X Tech Stack: TypeScript and High Structure00:13:21 AI Coding Tools: The Daily Evolution00:15:05 Human Capital as the Limiting Factor00:16:02 The Unreasonably Difficult Interview Process00:17:14 Entropy and Context Engineering: The Future of AI Agents00:23:28 The MCP Debate and AI Industry Sociology00:26:01 Consulting, Digital Transformation, and Conference Insights Get full access to Latent.Space at www.latent.space/subscribe

Deedy Das, Partner at Menlo Ventures, returns to Latent Space to discuss his journey from Glean to venture capital, the explosive rise of Anthropic, and how AI is reshaping enterprise software and coding. From investing in Anthropic early on when they had no revenue to managing the $100M Ontology Fund, Das shares insider perspectives on the fastest-growing software company in history and what's next for AI infrastructure, research investing, and the future of engineering. We cover Glean's rise from “boring” enterprise search to a $7B AI-native company, Anthropic's meteoric rise, the strategic decisions behind products like Claude Code, and why market share in enterprise AI is shifting dramatically. Das explains his investment thesis on research companies like Goodfire, Prime Intellect, and OpenRouter and how the Anthology Fund is quietly seeding the next wave of AI infra, research, and devtools.

Deedy Das, Partner at Menlo Ventures, returns to Latent Space to discuss his journey from Glean to venture capital, the explosive rise of Anthropic, and how AI is reshaping enterprise software and coding. From investing in Anthropic early on when they had no revenue to managing the $100M Ontology Fund, Das shares insider perspectives on the fastest-growing software company in history and what's next for AI infrastructure, research investing, and the future of engineering.We cover Glean's rise from “boring” enterprise search to a $7B AI-native company, Anthropic's meteoric rise, the strategic decisions behind products like Claude Code, and why market share in enterprise AI is shifting dramatically. Das explains his investment thesis on research companies like Goodfire, Prime Intellect, and OpenRouter and how the Anthology Fund is quietly seeding the next wave of AI infra, research, and devtools.Full Video EpisodeTimestamps* 00:00:00 Introduction and Deedy's Return to Latent Space* 00:01:20 Glean's Journey: From Boring Enterprise Search to Valuation* 00:15:37 Anthropic's Meteoric Rise and Market Share Dynamics* 00:17:50 Claude Artifacts and Product Innovation* 00:41:20 The Anthology Fund: Investing in the Anthropic Ecosystem* 00:48:01 Goodfire and Mechanistic Interpretability* 00:51:25 Prime Intellect and Distributed AI Training* 00:53:40 OpenRouter: Building the AI Model Gateway* 01:13:36 The Stargate Project and Infrastructure Arms Race* 01:18:14 The Future of Software Engineering and AI Coding Get full access to Latent.Space at www.latent.space/subscribe

Jared Palmer, SVP at GitHub and VP of CoreAI at Microsoft, joins Latent Space for an in-depth look at the evolution of coding agents and modern developer tools. Recently joining after leading AI initiatives at Vercel, Palmer shares firsthand insights from behind the scenes at GitHub Universe, including the launch of Agent HQ which is a new collaboration hub for coding agents and developers. This episode traces Palmer's journey from building Copilot inspired tools to pioneering the focused Next.js coding agent, v0, and explores how platform constraints fostered rapid experimentation and a breakout success in AI-powered frontend development. Palmer explains the unique advantages of GitHub's massive developer network, the challenges of scaling agent-based workflows, and why integrating seamless AI into developer experiences is now a top priority for both Microsoft and GitHub.