Podcasts about Latency

Multiplayer games are among the hardest software systems to build, requiring developers to synchronize state across unreliable networks while maintaining fairness, performance, and a responsive player experience. Latency, cheating, server costs, and debugging distributed game logic all introduce complexity that single-player games never encounter. Dome Keeper is a minimalist tower defense game with roguelike elements The post Developing Multiplayer Games in Godot appeared first on Software Engineering Daily.

Timestamps: 0:00 Windows Low Latency Profile 1:26 YouTube AI Music Training 2:36 Anthropic Releases Mythos 4:47 QUICK BITS INTRO 4:54 Instagram Algorithm Adjustments 5:17 Donut Lab Battery Investigation 5:51 Google Gemini Live Translate 6:17 Hospital Using Palantir System to Save Lives 6:51 Miniature Surgical Robot NEWS SOURCES: https://lmg.gg/Iz3qB Learn more about your ad choices. Visit megaphone.fm/adchoices

Multiplayer games are among the hardest software systems to build, requiring developers to synchronize state across unreliable networks while maintaining fairness, performance, and a responsive player experience. Latency, cheating, server costs, and debugging distributed game logic all introduce complexity that single-player games never encounter. Dome Keeper is a minimalist tower defense game with roguelike elements The post Developing Multiplayer Games in Godot appeared first on Software Engineering Daily.

In this episode, we explore the OneOdio Studio Max 2 headphones, a collaboration with producer KSHMR designed for DJs and studio professionals.  These headphones feature 9-millisecond ultra-low latency monitoring through their RapidWill+ 3.0 technology, addressing the sync issues that typically plague wireless audio in live performance settings. The 45mm drivers deliver Hi-Res certified sound across a 20Hz-40kHz range, with support for high-bitrate codecs including Sony LDAC at 990 kbps and OPUS at 400 kbps.  Battery life reaches 120 hours in standard Bluetooth mode, with a fast-charge feature providing 9 hours of playback from just 5 minutes of charging. The headphones offer four connectivity modes, including both wireless options via Bluetooth 6.0 and wired connections through dual jacks (3.5mm and 6.35mm) with a 3-meter coiled cable. The 180-degree swivel earcups enable single-ear monitoring for DJ work, while the OneOdio app provides EQ customization and a 60ms gaming mode. At $199.99, these headphones aim to serve both studio production and live performance needs, with 1,000 units available in a signed limited edition that includes a KSHMR sound pack and premium EVA case. Follow AndroidGuys(X) Twitter: https://www.twitter.com/androidguysInstagram: https://www.instagram.com/androidguysTikTok: https://www.tiktok.com/@androidguysofficialYouTube: https://www.youtube.com/@AndroidGuyscomOfficialWebsite: http://www.androidguys.comFollow Scott WebsterInstagram: https://www.instagram.com/scottwebsterFollow Luke GaulInstagram: https://www.instagram.com/lukegaul

AI has already and will continue to impact the salon industry. One area that is getting a lot of attention is the front desk and the role of AI in streamlining the client booking process, capturing missed calls and directly having a positive impact on salon revenue.  My guest today is Universe Walker, a hairdresser, salon owner, and the founder of Beauty Desk AI, whose AI receptionist Eve answers calls and texts, books appointments, handles escalations, and never puts anyone on hold. We dig into how she built it and why online booking alone isn't enough.If you've ever lost a client because nobody answered the phone, this episode will hit home. You'll walk away understanding exactly what an AI receptionist can and can't do, how it fits alongside your front desk team, and whether an AI receptionist could be the solution your front desk needs.EPISODE TIMESTAMPS[00:00] Introduction: AI is changing how salons handle every incoming call[01:06] Meet Universe Walker, salon owner, hairdresser, and founder of Beauty Desk AI[03:26] The problem Universe set out to solve after COVID changed everything[05:06] Why 75% of salons no longer have a front desk, and what that costs[07:00] What Eve actually does: voice, SMS, bookings, and the all-in-one dashboard[10:12] How Eve integrates with your existing booking platform[12:09] Latency, the "army of Eves," and why conversations feel surprisingly human[16:06] The 200-call experiment: what happened when nobody answered the phone[21:20] How Eve is trained to understand your salon's specific services and language[26:12] Real examples of Eve's empathy, from nervous kids to tsunami evacuations[29:00] How salons with a full reception team use Eve as a backup[36:41] The business insights Eve unlocks that no booking platform can give youWant MORE to help you GROW?

Guthrie Cooper (Senior Group Product Manager, AI & Robotics) and Nidhi Sharma (Global Head of Engineering AI & Incubation) from Just Eat Takeaway.com join the MLOps.community to pull back the curtain on how one of Europe's largest food delivery platforms is running an internal innovation engine. From autonomous delivery robots to agentic AI voice assistants, they share what it actually takes to build like a startup inside a 40,000-person company.Inside Just Eat's AI Lab: Voice Agents & Agentic Commerce // MLOps Podcast #377 with Just Eat Takeaway.com's Guthrie Cooper (Senior Group Product Manager, AI & Robotics) and Nidhi Sharma (Global Head of Engineering AI & Incubation)

AI is moving fast, but after my conversation with Alex Bouzari, Co-Founder and CEO at DDN, at Google Cloud Next '26, one thing became clear.The bottleneck is no longer the model.It is the infrastructure behind it. Alex broke it down in a very real way. Today's AI systems are powerful, but the way data moves through them is still inefficient. You train these large models, but when it comes to actually running them at scale, things slow down. Latency increases, costs go up, and performance becomes unpredictable.That is what is broken.He shared how this shows up in real scenarios. When enterprises deploy AI, especially with large models, they struggle with speed and consistency. It is not that the model cannot perform, it is that the infrastructure cannot keep up with the demand.At Next, DDN focused on solving exactly this. Building what Alex called a new foundation for AI, designed for high-performance workloads where data access and speed matter just as much as the model itself.One concept that stood out was KV cache.It sounds technical, but the idea is simple. Instead of recomputing everything every time a model runs, you reuse key pieces of information. That reduces latency and makes systems faster and more efficient. In large-scale AI systems, that becomes a big deal.The bigger shift here is clear.We are moving from experimenting with AI to operationalizing it at scale. And that means infrastructure is becoming the deciding factor.What makes DDN different is their focus on this layer. Not just enabling AI, but making sure it actually performs in real-world environments.My takeaway. The future of AI will not just be defined by better models. It will be defined by better infrastructure.#data #ai #ddn #infrastructure #googlecloudnext #api #google #theravitshow

In today's episode, we dive deep into Solana's PropAMM landscape and the battle shaping onchain market structure.We discuss:What PropAMMs are and why they've taken over Solana spot liquidityHow gas economics shift costs from makers to takersLatency games, flashing, and spoofing tacticsExecution quality vs oracle pricingWhy onchain price discovery is still laggingBinance vs PropAMMs vs true decentralizationThe role of Solana's scheduler, priority fees, and compute unitsWhy PropAMMs dominate spot but struggle in perpsThe case for continuous order booksDoubleZero and private fiber networksBuilding Blink: a neutral, price-discovery-first venueThe future of fair and verifiable crypto exchangesAnd much more—enjoy!—Timestamps:(00:00) Introduction(01:07) What is Blink?(04:07) Crypto market structure problems(10:46) Binance vs decentralization spectrum(22:10) PropAMM gas advantages(32:37) Latency and order flow(39:26) Spoofing and flashing tactics(1:03:27) Solana PropAMM efficiency(1:04:25) Blink tech stack(1:06:13) Spot vs perps limits(1:07:26) Continuous order books(1:10:33) Outro—Content links:https://dune.com/the_defi_report/prop-ammshttps://x.com/moonshiesty/status/2019803058750824710?s=20https://x.com/moonshiesty/status/2041266918015803661?s=46&t=CgOOIjEz_uRbAZFFRLsgAghttps://x.com/jump_/status/2044494127367135386?s=20https://0x.org/post/propamm-shenaniganshttps://www.helius.dev/blog/solanas-proprietary-amm-revolution—NEW: Join the Indexed Pod group chat:https://t.me/+Jmox7c6mB8AzOWU0And our new website is live:https://indexedpod.com—Follow the guest:https://x.com/moonshiestyFollow the co-hosts:https://x.com/hildobbyhttps://x.com/0xBoxerhttps://x.com/sui414Follow the Indexed Podcast:https://x.com/indexed_pod—The Indexed Podcast discusses hot topics, trendy metrics and chart crimes in the crypto industry, with a new episode every 1st and 3rd Thursday of the month, brought to you by wizards @hildobby @0xBoxer @sui414.Subscribe/follow the show and leave a comment to help us grow the show!—DISCLAIMER: All information presented here should not be relied upon as legal, financial, investment, tax or even life advice. The views expressed in the podcast are not representative of hosts' employers views. We are acting independently of our respective professional roles.

Special discounts up for AIE Melbourne (LS discount) and AIE World's Fair (group discounts up to 25% - CFPs still open for Autoresearch and Vertical AI) Cya there!Abridge did not start as an “GPT wrapper”. It was founded in 2018, years before the Cambrian explosion of AI application layer companies. OpenAI launched ChatGPT publicly on November 30, 2022 and by then, Abridge had already spent years doing the unglamorous work of building trust for one of the highest context, most important workflows in healthcare: the conversation between a patient and a clinician.Abridge's original wedge was clinical documentation. Listen to the visit, generate the note, reduce the clerical burden, and let clinicians spend more time with patients instead of the EHR. By focusing on how doctors actually document, how health systems actually buy, how EHR integration actually works, how clinicians verify outputs, and how missing context during a visit turns into downstream friction across billing, prior authorization, quality, and follow-up, the adoption of LLMs became a force multiplier on a workflow already optimized for sensitive context gathering.The company has scaled fast: Abridge says it is projected to support 80M+ patient-clinician conversations this year across 250 large and complex U.S. health systems, with support for 28+ languages and 50+ specialties. It raised $300M at a $5.3B valuation in June 2025, after a $250M round earlier that year.Today, Janie Lee and Chaitanya “Chai” Asawa of Abridge join us for another crossover pod with Redpoint's Jacob Effron (who is on the board of Abridge) to dive into how Abridge is building the clinical intelligence layer for healthcare starting with ambient documentation, then expanding into clinical decision support, prior authorization, payer/provider/pharma workflows, and eventually real-time agents that act before, during, and after the patient conversation. We go inside the product, data, infra, evals, workflow, privacy, and org design choices behind bringing AI into one of the highest-stakes enterprise environments from 100M+ medical conversations and specialty-specific evals to real-time alerts, EHR integration, de-identification, clinician-scientist teams, and why healthcare may solve some of the hardest AI problems first.We discuss:* Why Abridge started with clinical documentation, “pajama time,” and saving clinicians 10–20 hours a week* The transition from ambient scribe to clinical intelligence layer: save time, save money, and save lives* Why conversations between patients and clinicians may be the most important workflow in healthcare (patient visit summary feature)* Chai's “healthcare-coded Glean” framing: context is king, but healthcare raises the stakes on safety, evals, and rollout* Why Abridge wants AI to feel like “air conditioning”: always in the background, but only interrupting when it truly matters* The prior authorization example: turning a denied MRI weeks later into real-time guidance while the patient is still in the room* Why payer policies, EHR data, medical literature, and hospital-specific guidelines make the problem hard, and also create the moat* How Abridge thinks about ambient form factors: mobile, desktop, in-room devices, nursing workflows, multimodality, and future AR* The multi-sided healthcare customer: CMIOs, CFOs, CIOs, clinicians, patients, payers, and pharma* The hardest AI problem at Abridge: high-quality, low-latency, low-cost real-time support in a high-stakes clinical setting* When Abridge uses frontier models vs proprietary models, and why its unique data from medical conversations matters* Why “every agent is a coding agent underneath,” and how the EHR can be thought of as a filesystem for healthcare agents* How Abridge approaches personalization across individual doctors, specialties, and health systems* Why “AI slop” is AI without context, and how edits, memories, and clinician preferences create a data flywheel* Abridge's eval stack: LFDs, LLM judges, in-house clinicians, third-party evaluators, specialty-specific evals, and progressive rollout* HIPAA, PHI, de-identification, one-way anonymization, customer contracts, and learning from healthcare data safely* What changes when you operate at 100M+ conversations: reliability, cost, post-training, model routing, and infrastructure optimization* Why the same clinical conversation can serve doctors, patients, payers, pharma, and future clinical-trial workflows* How Abridge works with EHRs, and why deep interoperability is table stakes for clinician adoption* Why healthcare AI has regulatory tailwinds, why 80/20 does not work here, and why high-stakes domains may drive AI forward* Why Abridge embeds “clinician scientists” into product and eval teams* What Chai learned from Glean about search, quality, and durable AI infrastructure* Why the future of AI infra may look like context layers, event-driven systems, Kafka, Temporal, sockets, CRDTs, and tools built for humans* Why Janie changed her mind on “PRDs are dead,” and why crisp written clarity matters more in complex AI products* How Abridge uses Claude Code, Cursor, and coding agents internallyAbridge:* Website: https://www.abridge.com/* X: https://x.com/AbridgeHQJanie Lee:* LinkedIn: https://www.linkedin.com/in/janiejleeChaitanya “Chai” Asawa:* LinkedIn: https://www.linkedin.com/in/casawaTimestamps00:00:00 Introduction and what Abridge does00:02:05 From ambient documentation to clinical intelligence00:04:04 Clinical decision support and context as king00:06:57 Alert fatigue, proactive intelligence, and prior authorization00:12:36 Ambient AI form factors and healthcare customers00:16:59 The hardest AI problems in healthcare00:18:26 Frontier models, proprietary data, and model strategy00:21:07 The EHR as a filesystem for agents00:24:03 Personalization, memory, and clinician preferences00:30:40 Evals, LLM judges, and progressive rollout00:36:47 HIPAA, de-identification, and privacy00:39:21 100M conversations and operating at scale00:44:10 EHR integration and the clinical intelligence layer00:46:39 Healthcare regulation, latency, and high-stakes AI00:50:11 Clinician scientists and long-tail quality00:53:04 Lessons from Glean and durable AI infrastructure00:57:03 The future of agentic healthcare workflows00:57:34 PRDs, product clarity, and building serious AI products01:03:11 AI coding tools at Abridge01:04:06 OutroTranscriptIntroduction: Abridge, Clinical Intelligence, and the Latent Space x Unsupervised Learning CrossoverSwyx [00:00:00]: Okay. This is a special crossover Latent Space Unsupervised Learning pod.Jacob [00:00:07]: Very excited to do this.Jacob [00:00:08]: At this point, we get together once a year.Swyx [00:00:10]: Once a yearJacob [00:00:11]: And this is a fun occasion to get to do it on.Swyx [00:00:13]: I really wanted to talk to Abridge but I felt very underqualified because healthcare is not something we cover very intensely. It just so happens that Redpoint's our big investors and supporters of Abridge.Jacob [00:00:27]: Anytime you want to have a portfolio company on your podcastJacob [00:00:29]: Please, by all means.Swyx [00:00:31]: So we'll introduce our guests. Chai and Janie, welcome to the pod.Janie [00:00:34]: Thanks for having us.Chai [00:00:35]: Thank you.Janie [00:00:35]: We're excited to be here.Chai [00:00:36]: Thank you.Swyx [00:00:36]: So for listeners, what do you guys do, just to situate you guys in the company?Janie [00:00:42]: Abridge is a clinical intelligence layer for health systems. We really started with documentation and building for clinicians and as we think about reducing the burden that clinicians have, they're spending 10 to 20 hours a week on documentation. There's a massive doctor shortage in the country. We also think that conversations between patients and clinicians are probably the most important workflow in healthcare. It's where care is given and received but if you think about the 20% of our GDP that goes towards healthcare, almost everything is a derivative of that conversation, whether it's the claim, the payment, the actual diagnosis given, the treatment. And we've started with a conversation to reduce the burden for doctors on documentation but we're really excited about the path ahead as we become this broader clinical intelligence layer.Chai [00:01:34]: I'm Chai. I work on clinical decision support at Abridge.Swyx [00:01:37]: Yes.Chai [00:01:37]: And so as Janie said, we're uniquely situated where we started off with the clinical note. What I'm really excited about and where we're expanding towards is what are all the things you can do before the conversation, during the conversation and after the conversation if you did have access to all the context about patients, payer guidelines, medical literature and put that together and to serve, how healthcare could look fundamentally different.Swyx [00:02:01]: And that's the context engine that you guys have?Chai [00:02:04]: Yes.Swyx [00:02:04]: Is that what it's called? Okay.Swyx [00:02:05]: So historically, as I understand it, the company started in 2018. A lot of people would be familiar with the AI voice notes form factor that doctors would be “Well, do you consent to being recorded?” It replaces handwriting and what have you. But it sounds like more recently there's been a big transition in the company. Tell me about the broader transition.From Documentation to Clinical Intelligence: Save Time, Save Money, Save LivesJanie [00:02:26]: So from a transition perspective, we really think about our journey as The first act was: how do we help save time? And that's where a lot of that original product was.Swyx [00:02:37]: By the way, one of those interesting statsSwyx [00:02:39]: On your landing page was, doctors spend time after hours.Janie [00:02:43]: They call it pajama time.Swyx [00:02:44]: Why is that pajama time?Janie [00:02:46]: Doctors after work in their pajamasSwyx [00:02:48]: In their pajamas. OhJanie [00:02:49]: At home are just writing and catching up on their notes every day.Janie [00:02:53]: Some of our favorite customer love stories, we have a Slack channel called Love Stories. We have clinicians telling us, “Abridge has helped us, from retiring early or we're now finally able toJanie [00:03:06]: go home and eat dinner with our kids for the first time.”Chai [00:03:08]: Save the marriage in some cases.Swyx [00:03:10]: One of the quotes was “We're not divorcing anymore.”Swyx [00:03:12]: I'm asking, “Why?”Swyx [00:03:14]: Because they're working too much.Janie [00:03:16]: But, in terms of where we're going and where we're expanding, we really think about our second and third acts around how do we help health systems save and make more money. Health systems are operating with record-low operating margins. It's getting harder and harder to serve patients and they have regulatory, some tailwinds but also a lot of headwinds coming their way and AI is ripe for helping on the saving and make-more-money piece. And then ultimately, how do we help save lives? The fact that our software and our product is open millions of times a week before, during and after a patient walks in the room, gives us massive opportunity with products like clinical decision support, which Chai is building but so many others to improve patient outcomes and probably one of the most important workflows and problems to be going after right now.From Glean to Healthcare: Context Is KingJacob [00:04:04]: One thing that's interesting, Chai, is you came over to Abridge from Glean and clinical decision support, which for our listeners is, in the context of a visit, helping a doctor figure out the right type of care. It's really a search problem in many ways, going through lots of different data sources. Very analogous to your previous role as one of the earliest engineers over at Glean. I'm sure a lot of our listeners are curious what's similar about the problems that you're going after now and what feels different, now that you're in healthcare.Chai [00:04:33]: Very similar. Taking a step back, with every wave, there's a lot of very similar patterns that happen across different products. A lot of social networking products look the same. A lot of credit-based products look the same. And we're seeing that very similar in the agent era with many companies, of course, in Redpoint's portfolio and so forth. And the key insight between both companies is that you have amazing models but context is king. Context is what puts them to work. So I see it in a lot of ways, a lot of similarities in this is a healthcare-coded version of Glean but the differences are really interesting. A couple things that come to mind. First and foremost, the rigor of the setting we're in. The downside risk is extremely high here in healthcare. It can be fatal in some cases. You prescribe something that the patient is allergic to for example. Whereas at Glean, it's “Oh, you got the question wrong.” It wasn't the end of the world in most cases. And so what does that mean? That shapes our evaluation strategy, both offline evaluation, progressive rollout and there's a lot more we could go into there. Second thing that comes to mind is, vertical versus horizontal. In both cases, there's a large variance but when Glean is, it's a much more horizontal company, there's a variance of personas, companies that you're working with. We also have a variance of personas, different types of specialties, different hospital systems. But the variance is a little more narrow. So from a product perspective, you're able to focus far more, especially when you have a maturing technology and you're building new products that never existed before. It lets you go after them much more easily and especially in healthcare where so many problems were solved with labor and process, that it's extremely ripe for AI to keep helping augment and enable. And the final thing that's really interesting, Abridge specifically compared to many other companies in the AI area, is the modality we started with where we're ambient and we're always listening in the background. And many more AI products will go that way but it's how we started. And that's the greatest form of AI we can create, AI that's seamless. You're not looking at your screen. It's always there. It's always helping you out and being proactive. The Jarvis vision that, every hackathon I went to over the past decade, there was always a Jarvis competitor. But Abridge very much started from the opportunity and continues to go that way.Ambient AI and Alert Fatigue: When Should the Product Interrupt?Jacob [00:06:57]: One thing that is super interesting then from a product perspective is you have this always-on seamless in the background and then you have to decide when you break the wall almost and say, “Hey, clinician, you might not have thought about X,” or whatever it is that you want to do. And in healthcare traditionally there's been this idea of alert fatigue and a million pop-ups and then a doctor just ignores all of them. It's probably a pattern that a lot of builders are thinking through now. How do you think about the right way to intervene or to pop up in a doctor visit?Janie [00:07:26]: It's such a good question. Alerts are notorious in healthcare specifically. Over 90% of alerts are ignored. The first and most important thing is context is everything, as Chai alluded to and I also think about how do we go from being reactive alerting to really proactive intelligence at the point at which it matters most. One thing we like to say is we want our product to feel like air conditioning. It should be in the background just making things better and if there is something that has great clinical risk and we're acutely aware that intervening now and not later is incredibly important, we should decide to act. But if you think about proactive versus reactive, instead of alerting a clinician during a visit when they're with their patient having a pretty serious and sensitive conversation, how do we prep a clinician before they walk into the room with that patient? And so historically, clinicians might have to manually go through charts with a patient that they've had over the course of months or years and they'll try to suss out what are the things they should be doing. You can imagine a world with Abridge. We'll summarize all of the most recent context for you, tell you based on the reason for a visit the patient is coming in for the types of things you should be discussing. And so you're going into that conversation prepped rather than walking in cold to that patient visit and then having this product interrupt you five or 10 times throughout the visit. And there might be times where it's really important to interrupt. We have a product called Prior Authorization and so this is when you may go into a doctor's office with knee pain. They'll prescribe you an MRI and so many of us have had this experience before, where in four weeks you'll get a call saying, “Hey, Sean, that MRI that you were prescribed wasn't approved and why don't you come back in? We'll figure it out.” In a world with Abridge, we might choose to quietly but still alert a doctor in that visit. And alert is probably not even the word we would want to use. Before a patient leaves, we would want to tell the doctor, “Hey, Doctor, before Sean leaves, you should ask him, has he had physical therapy and has his pain lasted for more than six weeks? Because the Aetna plan that he's on in California requires six things. We've already confirmed four of them have been met ‘cause we have all the context. But these two last criteria, if you can address with Sean before he leaves the room, we could guarantee that your MRI is approved before you leave.” And so when you think about clinical usefulness, impact to the patient, there are instances in which if we can catch a doctor while the patient is still in the room, as we think about save time, save money, save lives, we get to check all of those boxes. But when doctors have 15 minutes between visits, we have to be really thoughtful about when it matters.Prior Authorization: Reducing Latency in CareChai [00:10:23]: There's this interesting product opportunity AI has is reducing latency in the world. For example, prior authorization is an example of where care gets delayed and so great AI can reduce that. And the problem with alerts before partially is a technical problem: the quality of your alerts really matters. They're going to get ignored if you get alerts that... Similarly in engineering, where they're noisy alerts that you can't act on. But if you can make really high-quality alerts with both the context, as Janie said, and really high-quality models, then you can create a whole other game.Janie [00:10:53]: And I really like that experience because it starts to tease apart, what makes this so hard and unique. One, to make that prior authorization example possible, think about all the data that you need to have. You need to integrate with the electronic health record to know all of the patient context. Do we have access to your previous labs, previous imaging? And then to match you and to know that you're on Aetna, we have to collect all of the different payer policies and they vary by state. Some of these payer policies live on websites. Some of them live in unstructured 50-page PDF files.Jacob [00:11:31]: I thought this episode wasJacob [00:11:31]: To make sure we didn't scare people from healthcare.Janie [00:11:34]: But when you think about the things that make it hard, it also gives you the moat.Janie [00:11:39]: And then the second is the AI and the model quality we need to be able to hang our hat on. And so the bar, similarly when I worked at Opendoor, I worked on pricing models. Every outlier wiped out the margins of 30 and so similarly here in healthcare, the bar for accuracy is so high. And then I'd say the last is workflow is everything. If insurance companies deploy AI, it typically happens too late and this is when you have the notorious comical examples of AI just fighting each other when it's too late. But if we can pull forward the use of both the AI but also the ability to solve problems when the patient's in the room, you can start to collapse what typically takes weeks or months after your visit, ideally down to minutes or real-time. And it's where healthcare is both very difficult but also extremely rewarding if you can crack it.Product Form Factors: Mobile, Desktop, In-Room Devices, and ARSwyx [00:12:36]: Just to get some baseline on the form factors, because I've seen some videos on your website and stuff. You guys talk a lot about ambient AI. Is it primarily on the phone? Is there any other form factor that people get Abridge in? Is there an Abridge room setup where it's always on? I don't know.Jacob [00:12:55]: An Abridge podcast studio.Janie [00:12:58]: Primary form factor is mobile and desktop. UsuallyJanie [00:13:00]: Clinicians are walking in and out of rooms with mobile but at the end of the day, when they're closing out their notes or wanting to prep for the day ahead, they might use desktop. We have been having a lot of really interesting partnership conversations with a lot of these in-room device companies as you think about the power of multimodality and even more data, as you think about all of what is not captured today. It is fascinating to think about, especially even as we go into building and scaling our nursing product. It's one where nurses constantly, as they're walking in to check in on a patient for two minutes or maybe even 30 seconds,Janie [00:13:43]: Starting an Abridge experience is probably going to take longer than the visit. And so what can we do with in-room devices that are always on starts to raise really interesting and fun product questions.Swyx [00:13:54]: I was thinking, the way in tech companies we have all these Google MeetSwyx [00:13:58]: And other things, we might as well set up entire rooms with just Abridge tech.Chai [00:14:02]: Very much. AR glasses and related form factors are also relevant: how do we bring the information to the clinician in real-time without a screen, while still letting them focus on the patient?Swyx [00:14:18]: Do you think they want that? I'm skeptical of AR, but I'm curious what you've tried.Chai [00:14:26]: Admittedly, it's not a near-term product roadmapChai [00:14:29]: By any means. I'm being far-fetched.Jacob [00:14:31]: There's some sick AR stuff for surgeries.Swyx [00:14:33]: Really?Jacob [00:14:33]: When people are trying to visualize, you're about to make an incision but you want to see, what the cut might look or what the body might look like inside and they can layer in imaging.Swyx [00:14:43]: That's cool.Chai [00:14:45]: At some point in the future.Janie [00:14:46]: But there are a lot of our largest customers and at the largest health systems integrating already and so even as we think about building into it, unlocks a lot of product capabilities.Swyx [00:14:57]: And just to establish the terminology. Sorry, and I know I'm asking basic questions somewhat for myself but also for the audience who might beHealth Systems, Buyers, Clinicians, Patients, and PayersSwyx [00:15:05]: Less integrated. When you say health systems, it's like the Johns Hopkins, the Kaiser Permanentes.Janie [00:15:09]: Mayos, the Kaisers of the world.Swyx [00:15:10]: These are your customers, right? And the outcome that you deliver for them is happier doctors, reduced cost of processing, reduced mistakes. It's weird in a sense that I feel like there's also, a secondary customer, the customer of the customer and I don't know if you — do you think about it that way?Janie [00:15:28]: The other interesting and complex part of building product is we have our buyers, who are the chief medical information officersJanie [00:15:39]: The chief financial officers, the CIOs of these large health systems. Our users today are clinicians but if you think about who downstream is impacted, it's patients. And so as we build, with every product in mind, we think about who we're building for, who the secondary user is and what does that mean either in terms of experience, security compliance, ROI that we have to make tangible. And so like you said, time savings is one of them. But for CFOs, they care a lot more than just time savings. We have to show for every dollar you put into Abridge, because you have more compliant documentation or because you have fewer queries coming from your billing team, we save or add real dollars to your bottom line or top line, are things that we're constantly thinking about because of the dynamic across all three sets of users.Chai [00:16:32]: There's a whole other axis too with the payers and pharmaChai [00:16:35]: as well. Connecting all these three big stakeholders in healthcare isSwyx [00:16:39]: Do the payers ever see your data? Sorry, the payers meaning the insurers, right?Chai [00:16:44]: Yes.Swyx [00:16:44]: They also see Abridge data?Chai [00:16:47]: NoSwyx [00:16:47]: Like the direct integration to you guysChai [00:16:48]: They wouldn't see the raw Abridge data but when you're working together on something like prior authorization, whatever information they need, we'd communicate to them.Jacob [00:16:59]: That's cool. I would love to dig into the AI side. You still have a lot of problems on the AI side. And so maybe to start at the highest level, what's one of the hardest problems you have to solve in AI at Abridge today?The Hardest AI Problems: Quality, Latency, and CostChai [00:17:11]: To make things simple, let's take, building off the prior auth example. So one thing Janie talked about is okay, this data is all over the place and there's this combinatorial explosion of procedures, payer policies and even sometimes different health systems. There can be some cross-product of all of these different considerations you have to take into account. But what's really hard about this problem is doing it real-time in the conversation. So, in any AI product, usually the three KPIs you care about are quality, latency and cost. Now, what we're saying is we want you to do this real-time in the conversation, guiding the clinician. How do we do it in a way that does not break the bank? But we're using — But we also need very intelligent models because you're working with this cross-product of data and this, all this context layer as well. So you need high intelligence and high-quality because you don't want the alert fatigue but you also need to be fast and cost-effective. And so that's where a lot of clever engineering goes. It's okay, without getting into all the details here, can you model these policies in some intermediate representation or other things that you can do that can make this problem tractable? And of course, the Pareto frontier is always changing but we are also trying to do this now.Model Strategy: Third-Party Models, Proprietary Data, and Medical ConversationsJacob [00:18:26]: What implications has that had for what you take off-the-shelf and say, “ what? We don't need to be world-class at X. We'll just take this from the model providers or from some infrastructure player,” and what you're “No, this is where we spend most of our time focused on”?Chai [00:18:38]: This is, the fun challenge in AI?Jacob [00:18:42]: It changes every three months? SoChai [00:18:42]: Of course, with the shifting landscape, we try to be extremely thoughtful on predicting the trends of where third-party models are going and where we can uniquely go. And, sometimes when you talk about AI models, we're the models are just going to get infinitely better. But I don't think... It may be in the grandness of time you could say that but, within every month, every quarter, there's specific ways they're getting better. They're training on a lot more, coding data to be better coding agents, for example. And soChai [00:19:14]: We have to think about where are the things that won't — unique data that we're uniquely training on or to step back a little, where is a proprietary model bringing advantage to us is if it can give higher quality or lower cost and latency for similar quality, very similar to many other companies. And when we can do that is when we have proprietary data. So, for example, we have on the order of eighty million or hundreds of millions now getting close to of medical conversations.Jacob [00:19:44]: It's insane.Chai [00:19:45]: This is a unique data set. And this data set, it's very interesting because this data set is effectively a large part of the trace between the patient and the provider. That's where the quote-unquote debugging happens in healthcare. We have these traces at scale, as in as, our CEOs even called it, an exhaust that comes out of our product. And so when you have these traces, that's how you can train better agents on certain use cases, whether it's your transcription diarization use cases or so on or like note generation models and we can do that much cheaper and faster. But we're always also working with these third-party model providers. We closely collaborate with them and that's how we predict where the trends are going. The thing that I think about a lot is that, I know that the model providers are going to train much more on agentic workflows and so forth, so that's great, so that you have a better agentic harness. But the other thing that's interesting is that the model providers, because a large class of the consumer model providers is healthcare queries, that they might, optimize to train a lot of healthcare data to encode the knowledge in its weights. And this is just a great thing for us as well, where the off-the-shelf models can keep bett-getting better at general healthcare information, such that what our strategy is, we have a constellation of models, we can use something for this, that and, we only care about, at the end of the day, the best product experience.EHR as File System: Agentic Workflows and Real-Time InterfacesJacob [00:21:07]: And, you have, overall capabilities improving. I'm curious, as these models get better, is there something you look at and you're “, three months ago, we really couldn't do that but God, the the latest models really allow us to do it”?Chai [00:21:19]: So here's something interesting that I've, been toying with. So all models are... This wasn't super obvious a year ago but now it's become clear and clear that almost every agent is a coding agent underneath the hood? So you give it whatever file system, it can write its own code and so forth. So when you think about within healthcare and the use case that we have, you can think of the EHR effectively like a file system. It's just — it's a storage of all this information. It's a lot of information there that cannot fit into the context window, at least of today's models and you want to use that context effectively for all these product use cases we're talking about. And so if you have better agents that can, manipulate data, read that data, treat it as a file system as we see they're going and we know model companies are investing this way, then that very directly benefits us.Swyx [00:22:09]: Yeah. Okay, cool. Again, just establishing basic things. But we're going back to the model stuff. I'm really interested in double-clicking more on the real-time, element, which is pretty important for both of you. Is it — Is real-time just batches of every one minute, every five minutes? Is that how we do it? Or is there some more native, genuinely real-time in the sense that OpenAI has a real-time API or Gemini has a real-time API?Chai [00:22:35]: Yeah. Yeah. So today it is more on the on the batch basis but there's interestingChai [00:22:41]: Prototypes that we have that we're still not fully, full time, voice in text out or in that sense. But, can you trigger your models, your agents or agentic workflows, depending on the right times in the conversation?Chai [00:22:58]: And so you can imagine, different techniques to bring this latency down and, you want to bring the feedback loop down as much as you can. And so a lot of clever engineering there without fully... Maybe one day we'll do full voice in and text out, train a model to do something like that.Swyx [00:23:15]: You do — People don't want voice in voice out?Chai [00:23:18]: Now we aren't creating experiences that are, during the conversation, inter — It's almost likeSwyx [00:23:25]: Might be too disruptiveChai [00:23:26]: Too disruptive until, who knows, maybe eventually you could have full voice agents once we — the quality and we improve the comfort of the technology. But right now gra — that change is much more gradual and it's more text focus, text out.Janie [00:23:42]: And so much of currently what our product is trying to do is allow a clinician to focus on their patient and maybe at some point but right now patients, clinicians don't want a third voice, at least in a literal voice in that room. And so how do we be there with all the contacts and information ready at hand when there's the right moment?Personalization: Individual Doctors, Specialties, and Health SystemsJacob [00:24:03]: Jenny, one thing I'm curious about is how you think about, personalization in the product. I imagine, every doctor is a special snowflake in their own way, has their own way they like to do things. There are probably a bunch of different approaches you could take to doing that, both within the model layer itself but then also just with clever prompting or engineering. How do youJacob [00:24:20]: Deliver on that?Janie [00:24:21]: It's such a good question. Personalization is massive for us. We think about personalization at three levels. The first is at the individual, the second is at the specialty level and then the third is at the health system or the organization level. To your point, there are a lot of individual preferences. You-When a note is produced, it almost is a reflection that is so deeply personal of a doctor's work and how they give care. And so do they have preferences on things like style? They might want bullets versus paragraphs, really concise versus comprehensive. They also might have phrases that they really like to use or the templates that they want every note to be structured. And, we see it in our feedback all the time. We want two spaces in between sentences or I refuse to use this tool. And so that's something that we've had to build in. And the tricky part is how do you make sure that stylistic preferences don't interrupt accuracy and quality and that's something that we've really had to refine and hone over time. Second is at the specialty level. A cardiologist note or workflow is going to look very different from a dermatologist workflow.Jacob [00:25:32]: I assume cardiology notes are the highest stakes for you guys, given your CEO is a cardiologist.Jacob [00:25:36]: It's “Oh my God, make sure we get this one.”Janie [00:25:37]: Shiv, our CEO, is still a practicing cardiologist. He rounds once a month. And so, first call when we want just quick and easy user feedback too.Janie [00:25:46]: But, specialties require a lot of personalization, both in terms of what does the product look and so we make sure that as new users onboard, we catch that and the product proportionally reflects that. But also on the back end, evals at the specialty level, they are hard-earned to calibrate and get. What does a really great dermatology note look like? What makes it complete? What makes it compliant and billable is very different than a primary care doctor. And so it's not just about what does the product experience look but on the back end tuning and really deepening our understanding for the specialists. What does great output look like? And that's, a problem that we need to calibrate internally, externally, online, offline but, takes lots of cycles but is necessary in a high-stakes environment. And then at the health system level, for products like clinical decision support, you have health systems who've spent years or decades refining their best practices and they want to know, “Hey, we love your clinical decision support product but how do we embed our own hospital guidelines into them to inform clinicians before, during or after a visit what brest — best practices should look like?” And as you think about, deepening moats as well, when health systems, trust us with that data, allow us to productize it and directly into the clinical workflow, makes us a really great partner to health systems who want to build something that truly meets their needs, their practicing guidelines.AI Slop, Memory, and Product Data FlywheelsChai [00:27:23]: And I want to add onto that. The for the clinical documentation problem, it's very similar to AI writing that doesn't feel like your own and then we call that slop. But the way I describe one framing of slop is like AI without context. But we have all that context and both the clinicians, can have it and can guide it. And so part of the other interesting exhaust for us is, memory is, one of these new systems recordsChai [00:27:49]: Almost.Janie [00:27:50]: And we also have all the edits people make on our product and when you think about a data flywheel and how we get better over time becomes really powerful as a mechanism to just going deeper in personalization.Jacob [00:28:04]: It's interesting. I love this idea of working with systems on the guidelines they built up over a long time. I feel like so many of the best AI app companies today are... The question is: How do you take the expertise that a law firm or a bank has built up over many years and then add that as context and also a special sauce over, a an AI tool? And so seems like y'all are really doing that very effectively.Janie [00:28:24]: We're now starting to have our customers ask, “What are other customers doing?”Janie [00:28:28]: “And how are they doing it?”Janie [00:28:30]: And as we think about having visibility across such a large set of care being delivered right now, a really interesting place we could also partner.Swyx [00:28:40]: I'm just curious. I — This may be a nothing question but, how different are health system guidelines from each other? Don't they all converge to the same thing? And if not, where do they differ?Chai [00:28:52]: At a really high level, they're going to talk about very similar things but the difference is probably in some more of the details. “Oh, you should refer to specialists only when XYZ conditions are met,” or so forth and maybe different organizations have different practices and guidelines around that. But high level, talking about similar things but the details are what, of course, that shapes the context and the decisions you make.Swyx [00:29:15]: And this all goes into the context engine and it might affect the notes but maybe not.Chai [00:29:21]: The — For these local pathways, we're definitely thinking about it a little more for our clinical decision support product.Chai [00:29:26]: So yeah.Swyx [00:29:27]: Which is your stuff, yeah.Swyx [00:29:28]: And then the memory which you raised, let's just tell us more about that. What have you tried in memory? What's the structure of the memory? What works? What doesn't work?Chai [00:29:38]: There's, of course, many different ways you could do memory, where it's okay, can you bake it into the model weights or can you do it in some external store? For us, what's interesting is, of course, when you think the models are rapidly changing, whether it's in-house or third-party, baking into the model weights, sometimes you worry that it could be a little throwaway. And so, how do you... You need to find a way that you decompose the problem, the preferences from the underlying models and so forth. The thing we're right now most both that's easiest to start with and we're excited about is having, a separate store for memory, where you have, for example, a memory sub-agent that's, working in the background, figuring out what are the important parts of the clinician's actions that we want to remember for the long term. And then you can also imagine, other things where in the — you have background jobs that are running that are collating these, memories similar to Sleep, of course and what other pattern, patterns products do as well. Learning over all these action, all the action data we have, again, note edits, the conversations they did and the actual transcripts.Evals: LFD, LLM Judges, and Clinical SafetyJacob [00:30:40]: What about evals? How in the world do you... It is such a complex product surface area. We would love to hear you riff on that and also how has that evolved? I'm sure you've gotten better at it, so any learnings along the way.Janie [00:30:50]: From an evals perspective, we, from day one when we build any new product or feature, we think about, what does good look like? And there are table stakes things like clinical safety but then you start to get deeper into what does good quality look like. And when you go into something like our core product, there's stuff like style and completeness and there's things like does this note become something that can be billable, which is very high stakes for a health system. We have a number of ways in which we get confidence for this. We have, internal in-house clinicians who do what we call an LFD process to give us our very first pass at is this or isn't this a good enough output, look at the effing data.Jacob [00:31:41]: LFD?Chai [00:31:42]: That's why I was smiling. I was “Is Janie going to mention what it stands for?”Jacob [00:31:46]: I was not... There's like a million acronyms.Jacob [00:31:48]: How am I supposed to know that I don't? So “Oh yeah, of course, an LFD.”Swyx [00:31:51]: I've never heard of LFDs.Chai [00:31:53]: It's a bridge for sure.Janie [00:31:55]: I got through three days and then I had to ask someone.Janie [00:31:58]: I thought it was just me that didn't knowJanie [00:32:01]: It's our internal process.Swyx [00:32:02]: But look at the data as a meme in ML, ‘cause you tend to not look at it. You just want to look at number go up.Chai [00:32:06]: Exactly.Swyx [00:32:07]: But yes.Janie [00:32:08]: But so, we make sure we look at the data and then as we think about all of the components of good output, we, one, create LLM judges across all of these and we make sure with annotated data and either internal or external evaluators, we feel like these judges are calibrated. And then depending on the stakes, we also work with in-house and third-party evaluators across all of these before we ship any big change. And the goal is, in terms of evolution, how do you go from this process taking months, down to weeks, down to days? Some of it is, a true science and ML problem. A lot of it's also just, hard operational work. Have you planned ahead in terms of what you need? Have you really optimized the capacity that you need across all of the different specialties you need? Have you gotten a really good sense of which third parties are great to work with for what use cases? This takes a lot of domain, expertise and, lots of mistakes and errors in figuring that out. And so as much of it is an ML problem, so much of it has also been operational gains that are hugely important, where domain-specific expertise is everything.Specialty-Level Evaluation and Progressive RolloutsJacob [00:33:23]: But it's funny, ‘cause I feel like people talk about healthcare like it's one giant market and the reality isJacob [00:33:26]: It's, dozens and dozens of sub-markets. And so it feels like in your evals you have to build that up across the board, probably.Swyx [00:33:34]: And is specialization the primary cardinality at... That's the word that comes to mind.Janie [00:33:40]: Sometimes, depending on the product or the use case. And so if we're making a note improvement or feature for a particular specialty, definitely but we have products that are for nurses. We have products that, are really aimed at making the document or the output a lot more billable. And so we'll want to work with coding teams and not necessary clinicians. And so likeJacob [00:34:05]: Coding meaning healthcare coding.Janie [00:34:06]: Yes. Yes.Jacob [00:34:07]: NotChai [00:34:07]: Yes. I see you.Swyx [00:34:07]: Other kinds.Janie [00:34:09]: But is this output proportional to the work that was delivered? Is there sufficient documentation to justify the amount that a health system may end up charging? And so, specialty sometimes but also domain, very different across all of the different products that we're working for. And building out that network is, not easy and is where a lot of our operational investments have gone into.Chai [00:34:35]: And I view a lot of analogies to self-driving cars here, where, part of it is we really want progressive rollout of features to test in the real world is this useful? Is this going to work? One big difference compared to past lives is before I'd build a product, maybe I'd alpha it and then I'd like GA it the next week, ‘cause I'm “Go, move fast, ship,” and whatnot. But the mentality is like you... I want to make contact with the reality as quick as possible but I want a progressive rollout. Because as much as I get as large of an offline eval set, I want the distribution of that to match real-life distribution. And over time, by rolling out early, similar to Waymo has a tagline, “The world's most experienced driver,” another thing that can, at least linearly increase for us is, both the size of our evaluation offline and online, that and it all feeds back.Janie [00:35:25]: Something that's been earned over time, speaking of evolution, is just the trust we've gotten with customers. Historically, a lot of these health systems, when they bring on new vendors, their release cycles are quarters, sometimes twice a year. We've gotten our customers onto monthly release cycles, which is pretty fast for health systems but what is more exciting over the last, call it, few quarters, has been, a subset of our customers have said, “We want to innovate with you. We trust you,” and we have a pretty, decent chunk of our customers who say, “We'll develop with you outside of these monthly release cycles. We have a higher tolerance. We know that the stakes are very high but we want to be the first ones using these products, giving you feedback.” And so for a pretty substantial set of our customers, we've been able to convince them to be able to ship, in this gradual way before GA. Something we talk about a lot internally is, trust is earned in drops, earned in buckets and so we still can't do what I used to do when I worked at Loom. We had 30 million users. I'd just be, rolling out experiments left and. The bar is still quite high for iterative rollout but because of the trust we've earned, we're able to learn at pretty high volume very quickly.Privacy, HIPAA, and De-IdentificationSwyx [00:36:45]: Your scale is still pretty huge.Swyx [00:36:47]: One thing I want to... We were going to go into scale? In a sec. One thing I wanted to call up, follow up on evals, which, again, just coming from a generalist engineer point of view, just thinking through what would people be scared of in doing this, the privacy and HIPAAJacob [00:37:00]: Elements of this. I have zero experience in that. What do you have to do? What is surprisingly not that bad?Chai [00:37:06]: So one thing that's really important here from a compliance perspective is very much that any of the data we use needs to be de-identified, any real-world data we use as a basis of online eval sets we're learning from. And so you have to — And there's, very clear, government guidelines, what counts as PHI. And so we've even have built models that can take, for example, a clinical transcript and remove all the key PHI indicators and so you have a scrubbed/de-identified version. And then once you... And so one thing that's important is first you've got to get confidence in that model in the first place? And prove that out. Because, now you have, multiple probabilistic systems on top of each other.Chai [00:37:46]: But once you have that, then you can train on it use it for evaluation and so forth, provided one of the cool things also that you can do from a business side is the right data contracting as well with your partners.Jacob [00:37:57]: Is the anonymization one way? Once it's done, you cannot undo it? Or is there someoneChai [00:38:01]: YesJacob [00:38:02]: Who holds the master key that can... Yeah, okay. So it's one way.Chai [00:38:05]: It's one way. Yeah.Jacob [00:38:06]: That's how it works. I just wanted to... Because, there's a lot of this, learning from feedback and everything that, you would want to debug more but you can't because you just physically don't allow yourself to.Janie [00:38:17]: Some of it's also written in our customer contracts in terms of who can or can't access PHI data, how long do we retain it,Jacob [00:38:27]: Very goodJanie [00:38:27]: Before it gets de-identified. And so we have a pretty high bar for who can access that PHI data, just to make sure that we always respect our customer data and privacy. But that's something that we partner with our customers on too, to make sure that as we want full, as close to precision as possible in that qualityJanie [00:38:48]: We can still use it.Jacob [00:38:50]: But it'll be fascinating to see how that space evolves? Because you think about, I used to work at a company that, did a lot of healthcare data in the cancer space and if you asked, the average cancer patient, “Hey, do you want people, do you want other patients to be able to learn-”Chai [00:39:03]: Take it.Jacob [00:39:03]: “... Learn from your experience?”Chai [00:39:04]: Take it all.Jacob [00:39:05]: They're “Please.”Jacob [00:39:06]: “I'd love, nothing more than for other people to be able to learn fromJacob [00:39:10]: The experience that I had.” And so in the past it was a lot harder to do that learning. But with this technology, that might really be practical and so it'll be fascinating to see how that continues to evolve.Chai [00:39:21]: There's so much in our data set of 100 million conversations.Chai [00:39:26]: You can imagine things like insights that you can give to the clinician. How could you, oh, how could you have reacted to this? In coaching or insights around, which treatments are effective or, like... Because you have this, again, this data source that was never captured before but that's, where, intuition or experience is created from, going back to this idea that the conversation is the agent of truth.Operating at Scale: Reliability, Cost, and Token EfficiencyJacob [00:39:46]: Back to the 100 million conversations, I feel like you have this insane scale that maybe only a few other AI app companies have and everyone else dreams of. So not everyone has had to confront this yet but maybe just talk about some of the challenges of operating at that scale and what, our listeners have to look forward to if they ever get to this level of scale.Chai [00:40:05]: At large and larger in scale, so of course there's a general, infrastructure reliability. When you... In any given startup, you're building the plane while it's flying. So there's some notion of that. But what gets interesting on the AI and ML side for sure is this, as you get at more and more scale, so one, you have the data to first and foremost do this. But, you start thinking about costs or infrastructure in a whole different way at scale versus, a prototype.Chai [00:40:34]: You can use the most expensive model, you can burn as many tokens as you want but when you're doing 100 million conversationsJacob [00:40:41]: Token max on leaderboards are less upsetting than that context.Chai [00:40:45]: . When you're doing that and so that comes for we have the data and we also have the team that's able to post-train based on this and you can optimize for efficiency, especially in areas where you believe that maybe a lot of the quality headroom is less so and you don't expect the other off-the-shelf models to go that way, such that you want to do, efficiency maximization, in terms of compute and tokens.Jacob [00:41:08]: I feel like you guys live in the future in some way where most use cases today are really just in use case discovery mode, where it's “God, I really hope I can find something that can get to scale,” and so you're always going to use the most powerful model. And then the few things that do get to this level of scale, you start to do those optimizations.Chai [00:41:22]: It's a natural trajectory where it's like zero-to-one, we're not talking about any of these optimizations.Chai [00:41:26]: But when maybe we're in the one-to-100 or so forth, then we're in optimization mode and, what works out really well is you've got all this data from zero-to-one that lets you do this.What Comes Next: The Conversation as the Shared Healthcare PlatformJacob [00:41:36]: That's fascinating. I feel like one thing that's so interesting about the Abridge footprint is that you're in the doctor-patient visit in real-time. I always like to say, there's like probably 50 years' worth of product you could build on top of that. What gets each of you, I don't know, what are you most excited about building, either in the short term or medium term or even, long down the line?Janie [00:41:53]: Something that I get really excited about is that the same conversation can serve so many stakeholders. If you think about the conversation, a doctor needs to know what is the documentation, how do I make sure that this fully represent the care I gave? A patient needs to know, “What the heck just happened? This was really overwhelming. What are my next steps?” A payer needs to know, was this the proper and appropriate care given? A pharma company might want to know why isn't this drug being properly used or is there a good candidate for this clinical trial that I'm about to run? And where I get excited is that our product and our platform and our infrastructure can be the same product across all of those things and start to what's today, separate, very expensive, complex systems that serve each one of these stakeholders in very different ways, start to collapse all of that into a singular platform that enables not just more efficiency across the board but also better outcomes for everyone. And, all of us experience healthcare in probably very painful ways and knowing that there is a world in which we can simplify a lot is really exciting to me and it all starts with the conversation.Chai [00:43:15]: It's interesting. Of it very similar to going back to the KPIs that any AI product cares about. How do you increase quality of care? How do you reduce latency to care? And how do you reduce costs? Which is a huge, in healthcareJacob [00:43:28]: They call it the triple aim in healthcare.Chai [00:43:30]: But very similar to building AI products and the thing that really excites me is when we talk about that latency piece, we talked about one example earlier of prior authorization, can you reduce the latency to care? But you can imagine so much more. Oh, as soon as the lab value gets updated, do you have like a background agent that, kicks off and uses all the context to be “Oh, hey, the patient should do this next,” for example. And of flagging that to the clinician who's always in the loop but reducing that latency, to care. And then you can imagine this is much further down the road but it's like even connecting that to the direct patient and the consumer. And so how can you, how can you build a bridge to all of these things?EHR Partnerships and the Clinical Intelligence LayerJacob [00:44:10]: Very cool. The connections piece is just an ever-growing thing. And one of the key partners is the EHR and I wonder what that relationship is like. Will they, look at this as, something that is valuable enough that they want to own someday?Janie [00:44:29]: Our partnerships with the EHR is, we know that we have to be extremely close partners with all the EHRs who we partner with. Being able to not only pull and push all of the data into the right places is, not only table stakes, if we can't do that, health systems don't want to use us. The second and the reality of today is clinicians spend a lot of their days in the EHR. So much of what allowed us to win in the largest health systems was pretty direct and, very close partnerships with some of the largest electronic health records that allowed us to pull and push data with APIs that weren't ready out of the box. And clinicians want to save clicks. Anytime we introduce a new product that, adds two clicks for them in their day, they're “We're not going to use it.”Janie [00:45:21]: They have 15-minute back-to-back appointments with their patients. They're spending, hours during pajama time doing documentation. Every second and every minute counts and so we really think about being deeply integrated into the EHR as also table stakes to getting real usage and adoption. And anything that we build or introduce, we really talk about earn the right internally a lot, which is we have to provide so much value or save so much time that people will use us. But those are the two things that are close to us, is we know that the product won't be used unless it is deeply interoperable.Chai [00:46:01]: And strategically, to your point, it's like what does EHR want to own versus us? EHRs are really focused on the clinical workflows and so forth but some of the things that we're talking about here, I do these traditionally are outside of the domain where it's oh, connecting pairs and providers together with provider policies or the clinical trial matching, as Janie brought up. And so these are, entirely — we position ourselves as building this entirely new intelligence, clinical intelligence layer across, again, providers, pharma and, payers.Chai [00:46:33]: And so that's a it's a whole different ballgame that we try to playChai [00:46:36]: In combination with them.Jacob [00:46:37]: But it's like a different layer of scope.Healthcare AI Regulation, Technical Depth, and What Changed Their MindsJacob [00:46:39]: I'm curious, you are both relatively newcomers to healthcare. People have these, there's lots of futuristic healthcare AI takes of “Oh, everything will look different.”, now that you've been in healthcare for a bit, you live at the edge of AI, what have you, changed your mind on around this, as you think about what healthcare looks like in ten, 20 years? Any updates to your mental model from the time being close to the problems?Chai [00:47:02]: One thing that IChai [00:47:04]: Was hesitant about before and it's a common thing when I'm trying to recruit engineers that people ask me around, is definitely oh, healthcare, heavily regulated space. And it is, rightfully so. You want to keep, the patients at the end of the day safe. But one of the interesting things that, is a that surprised me how much it is coming to the company is there's a lot of really favorable regulatory tailwinds as well. Where you think about, government really wants interoperability between all these systems that we talked about and so agents can access this information. The government just in January, the FDA released updated guidance on clinical decision support, what I work on in such a way that they used to have guidance from like 2022 that required you to have, mention all these options and do all these other things but it's a very forward and forward-looking way. And so for me, what's been really cool to work on is this, there's this very special moment both in AI in general, we all know that but there's a special moment also regulatory in healthcare as well.Janie [00:48:05]: One thing I would call out is for the very reasons things are higher stakes or, potentially considered more difficult in healthcare, it's where some of the hardest AI problems will get solved first, just because the bar is so high. When I first joined, I was “Oh, this is where we'll be on the tail end of where, all of the AI innovation will be able to be applied.” But when you think about, zero error evals or multi-step workflows that have really low tolerance, a lot of the innovation will happen here just because we have to or else we can't ship.Jacob [00:48:42]: ‘Cause like in other domains, you'd much rather just solve the 80%-is-good-enough problems firstJanie [00:48:46]: 80/20 doesn't work hereChai [00:48:48]: And building off that, traditionally, there was a bit of stigma that, oh, healthcare companies are not that interesting from a technical perspective or I've seen that or faced that myself. But these are really hard and fun problems from a pure technical perspective beyond just the impact. How do you bring the latency of this thing down and make it really high-quality?Reducing Latency: Clinical Workflows, Agents, and Implementation RealityJacob [00:49:07]: How do you bring the latency of things down?Chai [00:49:10]: Yeah. Yeah. Yeah. So okay, let's answer the latency question. And maybe hopefully not too redundant with some of the things I've said earlier but some part of it is with any latency, you have to like what is, what is really your bottleneck. In a lot of workflows, it's sometimes it's the model itself. And so that's where like our data flywheel, our post-training team and so forth come in so that can you make the models far more efficient. So that's one aspect of latency. But there's whole other aspects of latency where it's okay, on top of that, if you use a constellation of different models, can you use — can you first use like a — it's like thinking fast and slow. Can you use a cheap, fast model that triages and hands it off to a larger model where you get more intelligence and so forth and so all theseChai [00:49:56]: Clever tricks to make it work.Chai [00:49:58]: And by the way, we are totally — we also realize that the parameter frontier is changing and so these tricks will — may not get us to where we want to be in five years but we need to if we want to build a useful product right now.Jacob [00:50:11]: Should we go to the quick-fire or you want to ask more about Abridge? We can stuff everything that's not Abridge into the quick-fireSwyx [00:50:16]: I don't mind. I was — I feel like Janie was on the topic of more long tail stuff, which isSwyx [00:50:21]: Not the eighty/twenty thing and that really matters. And I'll —, if you have any tips or cool stories or just general approaches that have worked for you that's interesting to dig into.Janie [00:50:32]: One of them is even just how we staff our teams looks different than a traditional software engineering team, I'd say.Swyx [00:50:40]: Let's go.Clinician Scientists, Edge Cases, and Evals at ScaleJanie [00:50:41]: We have a bunch of folks with different roles who are clinicians and so we have this role called the clinician scientist and I heard one of our leaders refer to them as mutants recently. But they are people who've had clinical backgrounds, so MDs typically, who are also deeply technical, somewhere, on the spectrum of like a full stack engineer all the way to like extremely scrappy prompter. But having each of these people embedded within our teams instantly raises the bar for everything that we build because not only are they determining, is this product clinically useful but they're deeply embedded in our whole evals process. And so when we talk about LFDs, when we talk about what is our actual evaluation criteria, you don't want Chai or me creating what those are because we don't have clinical background. But is probably unique to Abridge but has been game changing. And when you think about where the puck is going, you have people build with clinical backgrounds who are technical and where AI tools are going, they just becomeJanie [00:51:53]: More and more, critical and like the killers of the team. And so that's one. And then the second is just the scale at which we do evals to catch that long tail up front before anything ever gets into production is something that we've pretty much like really started to fine-tune, both from a scale but when do we know we need to get several hundred versus several thousand offline responses, what helps us make that quick decision and make this less of an art and as much of a science as possible. But that's also been something we've had to tune over time.Swyx [00:52:27]: And you have partners who opted in to give you those evals.Janie [00:52:31]: So we work either internally or with third-party for offline evals and then we have customers who also agree to give us, whether it's like thumbs up, thumbs down to like choose this or that, a lot of data to get us to what is as close to fully confident as possible.Swyx [00:52:51]: The term that comes to mind isSwyx [00:52:53]: Like active learning on things where you're weak. I feel like it's a lost artSwyx [00:52:58]: Is a lot of the polish that comes into doing something like this.Janie [00:53:02]: Really.Chai [00:53:03]: Hundred percent.Lessons from Glean: Technical Foundations and AI App InfrastructureJacob [00:53:04]: Maybe, on a totally unrelated note, Chai, you had a very, storied run at Glean b

So, both along the fault, we like those sensors really close the fault because that gives us really those first motions that are coming out of the earthquake.这两个传感器都紧贴断层，我们很喜欢紧贴断层的传感器，因为这样能让我们真正捕捉到地震最初的震动情况。It tells us the earthquake has started. So, once that earthquake started, seismic waves start travelling away from where that earthquake starts.这样我们就知道地震发生了。一旦地震开始发生，地震波就会从地震起始处向外传播。'P waves' or primary waves are the first waves that arrive. They're faster and you feel sort of a jolt coming up beneath you.“P波”或“初至波”最先到达观测点，传播速度很快，你会感觉到脚下有种震动感。The 'S waves' would travel along behind those and start shaking you back and forth.“S波”会紧随其后，让你来回摇晃。So, we're in Pasadena and we're at one of the seismic stations. At the bottom of this vault, this is what it is.我们正在帕萨迪纳的一个地震监测站。在这个小仓库的地下，这就是传感器。So, and it moves relative to the earth and basically whatever happens in here, gets turned into an electrical signal and boom.这个传感器会随地面一同移动，基本上，无论地表发生什么运动，它都能将其转化为电信号。It takes some very, very precise observation about what's happening.要记录这些运动需要非常精密的监测。We were one of the early adopters of the ShakeAlert system. We implemented it in January of 2017.我们是早期采用ShakeAlert系统的机构之一。我们于2017年1月启用了该系统。And it gives us precious seconds to anticipate an earthquake that may be coming to Los Angeles area and that will enable us to put an alert out to our operators.它能让我们争取到宝贵的几秒钟时间来预判可能会袭击洛杉矶地区的地震，还能让我们向工作人员发出警报。Part of the protocol is to slow down the train as it is moving in advance of an earthquake.协议的一部分是让列车在地震发生前减速行驶。So, the slower you're moving, the more safely you can operate the vehicle and stopping the vehicle is the best course of action for a large scale event.开得越慢就越安全，在这种大型灾难面前，停车当然是最好的选择。For very large earthquakes, the ShakeAlert early warning system is going to use seismometers.对于影响范围很广的地震，ShakeAlert预警系统将开启地震仪，But with the really big earthquakes, we're going to need information from these stations但如果是真的非常剧烈的地震，我们就需要这些监测站的信息，because what's going to happen is that if there's a really big earthquake, these stations are going to move very quickly in a very short amount of time.因为如果真的发生大地震，这些监测站会在极短时间内快速移动。And that information is critical to understand the biggest of earthquakes. We're talking earthquakes that are magnitude 8 and above.这些信息对于最大程度了解地震是至关重要的。我们谈论的是8级及以上的地震。Big earthquakes in places like Iran and Indonesia and they could stand to have their own earthquake early warning systems.伊朗和印度尼西亚等地都发生过大地震，他们可以有自己的地震预警系统。And this has actually become a really interesting international community. We're all learning from each other.这实际上已经成为一个非常有趣的国际社区。我们都在互相学习。One thing we want to do and the big currency of ShakeAlert is time. Time is really important to us.我们想要进一步完善的，同时也是ShakeAlert的重要价值所在，就是争取时间。时间对我们来说真的很重要。Currently, the time that it takes to get messages out is taking way too long and we want to reduce what we call 'latencies' as much as possible.目前，信息传递所需的时间过长，我们希望尽可能减少所谓的“延迟”。So, we're working with a whole array of partners, the telecommunications industry.我们正在与整个电信行业的一系列合作伙伴合作。We're working with mass notification providers, to get those messages out as quickly as possible.我们正在与大规模通知服务提供商合作，以便尽快发布这些消息。

Rafael (Head of Innovation, iFood) and Daniel (Data and AI Manager, iFood) pull back the curtain on ILO-Agent — iFood's conversational AI ordering system built for 200 million users across Latin America. Recorded live at AI House Amsterdam, this conversation goes deep into the engineering and product decisions behind building recommendation systems and agentic AI, and why the speed of your AI's response might actually be destroying user trust.The Latency Goldilocks Zone Explained // MLOps Podcast #376 with iFood's Rafael Borger (Head of Innovation) and Daniel Wolbert (Data and AI Manager)

This episode is brought to you by the MLflow team. Check out more information at MLflow.org.What does it actually take to build voice AI at a billion-interaction scale? This episode features an ex-Amazon voice AI engineer who built customer support systems handling 2 billion+ interactions — now working on next-gen voice agent platforms. Anurag digs deep into the real engineering tradeoffs, design patterns, and use cases that separate production-grade voice agents from demos.Voice Agent Use Cases // MLOps Podcast #372 with Anurag Beniwal, Member of the Technical Staff at ElevenLabs

It's another Coca Thursday, people! Yes, I complain about the Mets again. Yes, I root for BAD teams! But, let's talk about contracts. How do they get negotiated? Where does the price come from? Please fix the latency issue in streaming. If you know anything about this issue, contact me! How do the minor leagues work? Is the major league team instructing them on what they can do with MLB players? MLS vs Vancouver! Are the Whitecaps leaving for Las Vegas? Learn more about your ad choices. Visit podcastchoices.com/adchoices

It's another Coca Thursday, people! Yes, I complain about the Mets again. Yes, I root for BAD teams! But, let's talk about contracts. How do they get negotiated? Where does the price come from? Please fix the latency issue in streaming. If you know anything about this issue, contact me! How do the minor leagues work? Is the major league team instructing them on what they can do with MLB players? MLS vs Vancouver! Are the Whitecaps leaving for Las Vegas? Learn more about your ad choices. Visit podcastchoices.com/adchoices

From building Applied Intuition from YC-era autonomy tooling into a $15B physical AI company, Qasar Younis and Peter Ludwig have spent the last decade living through the full arc of autonomy: from simulation and data infrastructure for robotaxi companies, to operating systems for safety-critical machines, to deploying AI onto cars, trucks, mining equipment, construction vehicles, agriculture, defense systems, and driverless L4 trucks running in Japan today. They join us to explain why “physical AI” is not just LLMs on wheels, why the real bottleneck is no longer model intelligence but deployment onto constrained hardware, and why the future of autonomy may look less like one-off demos and more like Android for every moving machine.We discuss:* Applied Intuition's mission: building physical AI for a safer, more prosperous world, powering cars, trucks, construction and mining equipment, agriculture, defense, and other moving machines* Why physical AI is different from screen-based AI: learned systems can make mistakes in chat or coding, but safety-critical machines like driverless trucks, autonomous vehicles, and robots need much higher reliability* The evolution from autonomy tooling to a broad physical AI platform: starting with simulation and data infrastructure for robotaxi companies, then expanding into 30+ products across simulation, operating systems, autonomy, and AI models* Why tooling companies came back into fashion: Qasar on why developer tooling looked unfashionable in 2016, why Applied Intuition still bet on it, and how the AI boom made workflows and tools central again* The three core buckets of Applied Intuition's technology: simulation and RL infrastructure, true operating systems for vehicles and machines, and fundamental AI models for autonomy and world understanding* Why vehicles need a real AI operating system: real-time control, sensor streaming, latency, memory management, fail-safes, reliable updates, and why “bricking a car” is much worse than bricking an iPad* Physical machines as “phones before Android and iOS”: Peter explains why today's vehicle and machine software stack is fragmented across many operating systems, and why Applied Intuition wants to consolidate the platform layer* Coding agents inside Applied Intuition: Cursor, Claude Code, internal adoption leaderboards, and how AI tools are changing engineering workflows even in embedded systems and safety-critical software* Verification and validation for physical AI: why evals get harder as models improve, how end-to-end autonomy changes simulation requirements, and why neural simulation has to be fast and cheap enough to make RL practical* From deterministic tests to statistical safety: why autonomy validation is shifting from binary pass/fail requirements toward “how many nines” of reliability and mean time between failures* Cruise, Waymo, and public trust: Qasar and Peter discuss why autonomy failures are not just technical issues, how companies interact with regulators, and why Waymo is setting a high bar for the industry* Simulation vs. reality: why no simulator perfectly represents the real world, how sim-to-real validation works, and why real-world testing will never disappear* World models for physical AI: hydroplaning, construction equipment, visual cues, cause-and-effect learning, and where world models help versus where they are not enough* Onboard vs. offboard AI: why data-center models can be huge and slow, but onboard vehicle models need millisecond-level latency, low power, small size, and distillation-like efficiency* Why physical AI is not constrained by model intelligence alone: the hard part is deploying models onto real hardware, under safety, latency, power, cost, and reliability constraints* Legacy autonomy vs. intelligent autonomy: RTK GPS in mining and agriculture, why hand-coded path-following worked for decades, and why modern systems need perception and dynamic intelligence* Planning for physical systems: how “plan mode” applies to robotaxis, mining, defense, and multi-step physical tasks where actions change the state of the world* Why robotics demos are not production: the brittle last 1%, humanoid reliability, DARPA Grand Challenge-style prize policy, and the advanced engineering gap between research and deployment* Applied Intuition's hard-earned lessons: after nearly a decade, Peter says they can look at a robotics demo and predict the next 20 problems the company will hit* Qasar's advice to founders: constrain the commercial problem, avoid copying mature-company strategies too early, and remember that compounding technology only matters if you survive long enough to see it compound* Why 2014 YC advice may not apply in 2026: capital markets, AI company dynamics, and the difference between building in stealth with a deep network versus building as a new founder today* What Applied is hiring for: operating systems, autonomy, dev tooling, model performance, evals, safety-critical systems, hardware/software boundaries, and engineers with deep curiosity about how things workApplied Intuition:* YouTube: https://www.youtube.com/@AppliedIntuitionInc* X: https://x.com/AppliedInt* LinkedIn: https://www.linkedin.com/company/applied-intuition-incQasar Younis:* X: https://x.com/qasar* LinkedIn: https://www.linkedin.com/in/qasar/Peter Ludwig:* LinkedIn: https://www.linkedin.com/in/peterwludwig/Timestamps00:00:00 Introduction: Applied Intuition, Physical AI, and 10 Years of Building00:01:37 Physical AI vs. Screen AI: Why Safety-Critical Changes Everything00:02:51 The Origin Story: Tooling, YC, and the Scale AI Comparison00:05:41 The Three Buckets: Simulation, Operating Systems, and Autonomy Models00:11:10 Hardware, Sensors, and the LiDAR Question00:14:26 The Operating System Layer: Why Vehicles Are Like Pre-Android Phones00:19:13 Customers, Licensing, and the Better-Together Stack00:21:19 AI Coding Adoption: Cursor, Claude Code, and the Bimodal Engineer00:26:41 Verifiable Rewards, Evals, and Neural Simulation00:31:04 Statistical Validation, Regulators, and the Cruise Lesson00:40:25 World Models, Hydroplaning, and Cause-Effect Learning00:43:34 Onboard vs. Offboard: Latency, Embedded ML, and Distillation00:50:57 Plan Mode for Physical Systems and Next-Token Prediction Universally00:53:04 Productionization: The 20 Problems Every Robotics Demo Will Hit00:58:00 Founder Advice: Constraints, Compounding Tech, and Mature-Company Mimicry01:05:41 Hiring Philosophy: Hardware/Software Boundary and Engineering Mindset01:08:50 General Motors Institute, Education, and the Curiosity MindsetTranscriptIntroduction: Applied Intuition, Physical AI, and 10 Years of BuildingAlessio [00:00:00]: 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.Swyx [00:00:10]: And today we're very honored to have the founders of Applied Intuition, Qasar and Peter. Welcome.Qasar [00:00:17]: You guys really know how to turn it on to podcast mode. That was, you guys are real pros at this.Qasar [00:00:23]: They were just joking around right before this, and then they flipped it pretty quick.Alessio [00:00:29]: Oh, yeah, it's good to have you guys. Maybe you just wanna introduce yourself so people know the voice on the mic and they'll know what they're hearing.Peter [00:00:33]: Oh, sure. Yeah, I'm Peter Ludwig. I'm the co-founder and CTO of Applied Intuition.Qasar [00:00:38]: And my name is Qasar Younis. I am the CEO and co-founder with Peter.Alessio [00:00:42]: Nice. Can you guys give the high-level overview of what Applied Intuition is? And I was reading through some of the Congress files, when you went out there, Peter, and eighteen of the top twenty global non-Chinese automakers, you two guys, you have customers in agriculture, defense, construction. I think most people have heard of Applied Intuition tied to YC when it was first started, and then you were kinda in stealth for a long time, so maybe just give people the high-level overview of what it is today, and then we'll dive into the different pieces.Peter [00:01:10]: Yeah. So at Applied Intuition, our mission is to build physical AI for a safer, more prosperous world. And so we work on physical AI for all different types of moving systems, everything from cars to trucks to construction and mining equipment, to defense technologies. And we're a true technology company, so we build and sell the technology, and we sell it to the companies that make the machines. We sell it to the government, really anyone that wants to buy a technology to make machines smart.Physical AI vs. Screen AI: Why Safety-Critical Changes EverythingQasar [00:01:38]: Yeah. And I think in the broader AI landscape, a lot of the focus, rightfully so in the last, three years has been on large language models, and so everything fits in a screen. Like, whether it's code complete products or things like that. And what's different about us is we're deploying intelligence onto a lot of things that don't have screens. they're physical machines. There are sometimes screens within the cabin or for example of a car or a truck or something like that, but most of the value we provide is putting intelligence that is in safety critical environments. So that those two words are really important because learn systems can make mistakes if you're asking for, like, some, so something like, “Tell me about these podcast hostsQasar [00:02:28]: that I'm about to go meet.” But you can't do that obviously when you run, like, as an example, we run driverless trucks in Japan right now, as we speak. We can't have errors. Those are L4 trucks. Yeah.Alessio [00:02:40]: Yeah. Was that always the mission? I remember initially, I think people put you and Scale AI very similarly for some things about being kinda like on the data infrastructure side of things. What was the evolution of the company?The Origin Story: Tooling, YC, and the Scale AI ComparisonPeter [00:02:51]: Well, from the very beginning, we always wanted to, really be a technology company that helped generally push forward the industrial sector. And so we started off working in autonomy. Our very first customers were robotaxi companies. And we started off doing a lot of work in simulation and data infrastructure. And then over the years, we've expanded our portfolios. Now we have, over thirty products, and it's a pretty broad technology play within the landscape of physical AI.Qasar [00:03:19]: Yeah, I think the Scale reason is because we're all YC Universe companies. But it was a very different company. Scale, was, is more of a services company, data labeling company fundamentally. We started and still are, do a lot of tooling. So like, you think developer tooling is now in vogue again, thanks to the AI boom. But honestly, ten years ago, it was out of vogue. It w Like, doing a tooling company in 2016, 2017 was not, like, the thing to do because, I don't know if you remember, the VCs generally, their views was that toolings are They're just workflows, and workflows ultimately are not really interesting. And we've gone and come, full circle with that. But when we started the company, our kind of it's kinda like in the periphery of what the company wants to be. It was like, from our earliest days, like, we wanna deploy software on physical machines, like on cars and on trucks and things like that. And obviously, we didn't know that the transformer boom was gonna happen. We didn't know that autonomy systems would become end-to-end. Those things we didn't know. And why that's important when autonomy systems become end-to-end, it is just now those models can be generalized to, multiple form factors. And so back nine, ten years ago, tooling was a great way, and still is a great way to, build the technology and sell technology to our end customers, a lot of them who wanna build this stuff themselves. And so we just offer like a spectrum of solutions from you can just use like one part of a development suite of tools all the way to buying the full thing. The way to think about the company, or at least the way we think about the company is, as Peter said, a technology provider. It's kinda like, what NVIDIA does or what an AMD, but we just don't do chips.Qasar [00:05:06]: We don't do silicon. But we're a technology provider fundamentally. And I think even, we used to joke when we started the company, like, we're not the guys to build, like, Instagram. Like that was just towards That's not our That's just not us in a most fundamental way. IAlessio [00:05:20]: You have thoughts.Qasar [00:05:21]: Yes.Qasar [00:05:22]: Well, it's, it's I mean, I think it's just like what And I mean, we worked on Maps and stuff, Google Maps. Consumer products are extremely difficult for a lot of different reasons. It just, I think doesn't scratch the itch. I think we're like Michigan guys who are kind of more of that traditional engineering kind of a realm, or lineage. we used to jokeThe Three Buckets: Simulation, Operating Systems, and Autonomy ModelsPeter [00:05:41]: I gotta say, though, what was clear ten years ago was that there was so much more that was possible with software and AI in vehiclesPeter [00:05:47]: and that was generally the space that we started in ten years ago.Peter [00:05:51]: And the precise path that we've taken over the years, I think we've been strategic, and we've adjusted to make sure that we're actually building stuff that's valuable to the market. And like, the technology has changed so much. Like our own technology stack has completely changed, I would say, roughly every two years. And so now we've probably done, let's say, four complete evolutions of our own technology stack. And I sort of see that cadence roughly keeping up.Peter [00:06:13]: And so the way even we think about engineering is almost on this two-year horizon, we're preparing ourselves that, hey, like, we wanna invest the appropriate amount, but then also be very dynamic as the research gets published and as our research team figures out new advancements and adapting to that.Qasar [00:06:27]: Yeah. One thing that has been consistent is the type of people we've, we've recruited. It's engineers who are fall into the sometimes very traditional, like, GoogleQasar [00:06:38]: -gen suite, but way different from, other companies. We are hiring folks who really know the intersection of hardware and software, who know really low-level systems. Obviously, traditional ML researchers and folks who've, actually, put ML systems into production. That's been pretty consistent. I think that, like, you look at the mix of our engineering, eighty-three percent of the company is engineering, so it's, like, a giant list.Qasar [00:07:05]: A lot of engineers.Alessio [00:07:06]: Which, by the way, a thousand engineersQasar [00:07:07]: Yeah. A thousand engineers.Alessio [00:07:08]: that's on your website, so I imagine it's up to date.Qasar [00:07:11]: It is, it is up to date, yes. Yes.Alessio [00:07:12]: okay. And then forty-plus founders.Qasar [00:07:15]: Yeah. We would tend to also, This was more luck than strategy. But we've recruited a lot of ex-founders. It's been a great place for founders, YC and non, ‘cause obviously I know a lot of the YC folks. It's kind of like we recruit a lot of Google people.Qasar [00:07:33]: For them to exercise both their technical and non-technical skills because, we're, we're, we're on the applied side. We have a research team that we do fundamental research, we publish, and we've, we've had great traction there. But fundamentally, the business wants to take this intelligence and deploy it into production and there's, like, a certain type of person that's more interested in that.Alessio [00:07:54]: Yeah. You mentioned the tech stack, Peter, so I just wanted to give you some rein to just go into it. I'm interested in where Wayve Nutrition, starts and ends in some sense, what won't you do? What, do you do that's common among all the verticals that you cover?Peter [00:08:10]: There's a few buckets of work that we do, and we've been at this for almost ten years now, so the technology's pretty broad. But we got startedQasar [00:08:17]: Yeah, with a thousand engineers, like, you could work on lots of things.Peter [00:08:19]: There's lots of stuff, yeah, espe-especially with AI tools to help.Peter [00:08:22]: So we got our start in simulation and simulation tooling and infrastructure. And so generally, if you're trying to build a very complex software system that involves moving machines, you need to test that, and the best way to test it is it's a combination of virtual developments, a simulation, and then also obviously real world testing.Peter [00:08:39]: And then there's a very careful process of that correlation between the simulation results and the real world results and ensuring that the simulator is in fact accurate to that. Simulation's a very deep topic.Peter [00:08:49]: We have a whole suite of products in that, and we could talk for many hours about that specifically. But that is one part of what we do as a company. Reinforcement learning as a subpart of that is also super critical. I think a lot of the a lot of the best advancements happening in a lot of these AI systems right now in some way relate to reinforcement learning, and with now we have lots of compute, and you can do tons of interesting things for reinforcement learning. The second bucket of work that we do is on operating systems technology. true operating systems. Like, think about, schedulers and memory management and middleware and message passing and highly reliable networking and data links. Like, the reality is, if you want to deploy AI onto vehicles, you need a really good operating system. And when we were getting deeper into that space, there wasn't really anything that we were happy with.Peter [00:09:39]: Like, things existed, absolutely, and we were using what was available in the market, and as an engineering organization, we roughly realized these things aren't great. We think we can do this better, and so let's, let's build something. And that was then the that was the moment of inspiration that started our operating systems business, which is now a very real business for us. And in order to write and run great AI, you need a great operating system, and so that-that's what got us into that. And then the third bucket that we work on, it's, it's true fundamental AI technology. Models, we do a lot of work in, as mentioned, the foundational research, but then the also the world models and the actual autonomy models that are running on these physical machines, and that's across cars, trucks, mining, construction, agriculture, and defense, and so that's both land, air, and sea.Qasar [00:10:31]: And also, a smaller subsector of that third bucket is the interaction of humans with those machines.Qasar [00:10:38]: So that's a multimodal, experience. Historically, if you're moving a dirt mover or any of these machines, there are, like, buttons you press, whether they're actual physical tactile buttons or something like a touch screen. That's just That fundamentally is changing to where you're just talking to the machine and the machine and you're teaming with the machine.Alessio [00:10:58]: Voice?Qasar [00:10:59]: Yeah, voice, absolutely, yeah.Alessio [00:11:00]: Oh.Qasar [00:11:00]: And also the machine just being aware of who is in the cabin, what their state is. you can think from a safety systems perspective, the most simple version of this is, like, the driver is tired, right? They're, they're if you get those alerts when you're driving your car and saysHardware, Sensors, and the LiDAR QuestionQasar [00:11:15]: -maybe take a coffee break, that take that times, a couple of order of magnitudes up. But this concept of teaming man and machine is important. When you think about running agents or just running, different instances of, Claude and doing work for you in the background, you can take that analogy out, almost copy and paste and put it into, like, a farm, where you have a farmer who's running a number of machines. So where they interact with the machine is where there's maybe a critical decision or a disengagement or something like that, but generally speaking, the agent on the physical machine is running and making decisions on the behalf of the farmer until there's something maybe critical. And that's also what we work on. So that's not pure autonomy. It's a little bit of a mix, but it falls under, autonomy. In the automotive sense, that's typically defined in SAE levels as an L2++ systemQasar [00:12:05]: -with a human in the loop. But just take that idea, to other verticals.Alessio [00:12:09]: Yeah. You've not mentioned hardware at all, like sensors or obviously we you mentioned you don't do chips. I think even in AV there's, like, a big, cameras versus lidars. Like, what are, like, in your space maybe some of those design decisions that you made, and are they driven by the OEM's ability to put things on the machinery? And like, how much influence do you guys have on co-designing those?Peter [00:12:32]: Yeah. So we don't make sensors. Like, we're, we're not a manufacturer. Obviously, we use a lot of sensors in our autonomy products. in terms of what actually goes on the vehicles, we have a preferred set of sensors that we, let's say fully support, and then our customers, they can sort of choose from those. And obviously if there's a very strong opinion on supporting something else, we'll add that to the platform as well. And the lidar question is at this point sort of the age-old,Peter [00:12:59]: topic in autonomy, and the state of the industry right now is lidar is hands down a useful sensor, specifically for data collection and the R&D phase of autonomy development. if you see, for example, a Tesla R&D vehicle, it actually has lidar on itPeter [00:13:17]: to this day, right? In the Bay Area we see these. you'll see, like, Model Ys or Cybercab that have lidars on them just driving around. So it's, it's useful because it gives you per pixel depth information. So if you can pair a lidar with a camerand you can say that, well, this camera's looking this direction, this lidar's looking this direction, and now for each pixel of the camera I can see how far away is that pixel. you can actually then use that as a part of your model training, and then the that depth information then becomes a learned, a learned state of the camera data. And then when you're doing the production system, you can now remove the lidarPeter [00:13:52]: and now you can actually get depth with just the camera. And so that difference between, like, a highly sensored R&D vehicle and then the down-costed production vehicle, we use that across our whole portfolio of products. And of course the end goal is you want super low cost and super reliable.Peter [00:14:08]: And then in certain use cases you have some more, bespoke things. Like in defense as an example, you do things at night oftentimes, and so you care about sensors like infrared, more so than And you don't, you don't wanna be putting energy out, so you don't wanna use lidar or radar.Peter [00:14:23]: but you still need to be able to see at nighttime. So yeah, we work the whole gamut.The Operating System Layer: Why Vehicles Are Like Pre-Android PhonesAlessio [00:14:27]: Cool. So that's kinda like on the hardware level. Then on the OS level, how does that look like? What is, like, unique? my drive- I drive a Tesla. Whenever I drive some other car that has a screen, it always sucks.Alessio [00:14:38]: It's on, like, cheap Android tablet. It's like, it's laggy and all of that. What does the OS of, like, the autonomy future look like?Peter [00:14:46]: When most people, it's really what you just described. When you think about operating system in a vehicle, you're thinking about the HMI, right? The human machine interface, and absolutely that's a an important part of it, but that's actually only one thin layer on top. So when we talk about operating systems for, like, AI in vehicles, there's many layers that go deep into the CPU critical realm and embedded systems, and you're talking about the real time control ofPeter [00:15:13]: let's say the electric motors or the engine and the actuators, and you have different redundancies for different, let's say, the steering actuation in the vehicle. And all of these things, need very core support in the in the operating system. And then of course for autonomy you have real time sensor data that's streaming in, and the latencies there are really important, right? If you try to Imagine you try to run Microsoft WindowsPeter [00:15:35]: like streaming your sensor data in or controlling the vehicle. Like, the latencies are gonna be absurd. Like, you can never do that. And so what's special about what we do is we really have this system level thinking, right? So we're looking at, we care about every performance characteristics of the entire system, and then we also, because we're doing a lot of the software or all of that software, we can fine-tune and control all of those things. So we can very carefully tune in the latencies for every aspect of the system. We can carefully tune in the memory management. We can have the right, fail-safes and fallbacks, for different things. ‘Cause you have to account for what if, what if there is a critical failure? What if there's a cosmic ray that flipsPeter [00:16:14]: a bit in the middle of the processor that causes some, malfunction? And you have to have a fail-safe to all of that, and so the core operating system is a part of that. And then the one last thing, which is a lot less exciting but is, actually a very big topic, is reliability of updates.Peter [00:16:30]: so the I have a Tesla and you get updates fairly frequently, right?Peter [00:16:36]: Once a month. Most companies that are making vehiclesPeter [00:16:40]: are basically never doing updates, and they're And even if they are doing updates, they're usually only updating maybe one module. Maybe they're updating the HMI module. But they're not able to update, let's say, the CPU critical parts of the system.Peter [00:16:51]: You have to go into the dealer for that. And so with our operating system now we can actually enable highly reliable updates of any system in the vehicle, and that's way easier said than done. Like, there's lots of technical, technically deep stuff, in the tech stack to do that in a way that you're not going to accidentally brick a vehicle.Peter [00:17:08]: And right? If, imagine yourAlessio [00:17:10]: That would be bad.Alessio [00:17:11]: Bad.Peter [00:17:11]: Bricking a car is a very expensivePeter [00:17:13]: and honestly, like across the industry maybe one of the most just pure impactful things that we've done is we've just, we're, we're now enabling the industry to actually do software updates.Alessio [00:17:22]: Just to clarify as well, who is the customer for this? Like, I assume a lot of hardware manufacturers have their own firmware, and I'm sure some of them would just have you write it for them because you're experts. And others would have their own. Like, who pays for this? Who invites you into the house? Is it, is it the end user, or is it, is it the manufacturer?Peter [00:17:41]: Yeah. So let me make an analogy firstly on the on the fragmentation of software. So physical machines today are more akin to the state of the phone market before Android and iOS existed, right? So I worked on Android at Google by the way many years ago, and part of the reason that Larry at Google decided to get into Android was they wanted to run Google products on a bunch of phones, and they bought all of these phones from the industry, and it turned out they had like 50 different operating systems on these phones. And it was virtually impossiblePeter [00:18:17]: for Google to make their app run on all 50 devices equally well. And so the solution was, well, actually what if, what if they created-A really great operating system and made it attractive to all of these phone makers, and that was sort of the genesis for what Android was and why Android existed. It was a way for Google to get their products onto really wide diversity of devices. The state of the physical, industry right now, it's a little bit like that. Like, there's yes, these companies have firmware, but they have so many different operating systems, it's so fragmented, and to actually get a modern AI application to run on these vehicles, you actually, you first have to consolidate the operating system, and so that's, that's why we've done that. And then, your specific question was who are our customers? It's, it's, generally it's the companies that are making these machines.Peter [00:19:06]: And we're, we're, we're selling our technology to them to really simplify the architecture and then enable these AI applications to run on them.Customers, Licensing, and the Better-Together StackSwyx [00:19:13]: How much is reusable across? Like, do you have, like, one OS that is just configured for everything, or is there some more customization that is needed?Peter [00:19:22]: Yeah, highly reusable. So the fundamental technology is quite universal, right? So things that we do have to think about though are, like, chipset support. And so if you're, if you're coding, let's say, an LLM and you have start with an assumption that, “Hey, oh, I'm gonna, I'm gonna use CUDA, and I'm gonna run this, on an NVIDIA chip,” then you don't really have to think about the hardware in that sense. Like, you're just, “Okay, I'm just I'm in the CUDA/NVIDIA ecosystem, and I'm, I'm going to use that.” But the hardware, especially in safety critical systems, it's a lot more diverse. There's not one or one or two players. There's a bunch of different chipsets that we have to support. And so our operating system doesn't just run on, like, the equivalent of X86. It has to, it has to run on a number of different architectures from chips from a bunch of different companies. But again, we've been working on this for a long time now, so we have, we have support for all of those chipsets. And then when you want to then run the AI applications, we can then do that reliably across now a variety of providers.Qasar [00:20:19]: And I think that is, like, heavily inspired by Android, right? Android has a huge suite of testing and it's a reliable operating system that runs on thousands of devices. And we think we can, we can do the same in all these physical moving machines, with the difference that we're really in a safety critical realm. Android isn't.Alessio [00:20:40]: So on Android, I don't need to use Gmail, I can use Superhuman. Like, what about your machinery? Like, can people bring somebody else's automation to it, or is it kinda like all-in-one?Qasar [00:20:50]: You have to use us. No. Yeah. we're If, Yeah. Yeah, it's totally open. Yeah.Peter [00:20:56]: Yeah. our philosophy is that we are a technology company, and so we license our technology to customers to use how they want. And so if a customer wants to If they wanna license our autonomy tech and our operating system, then great, we'll license those. If they just wanna license the operating system and then use different autonomy tech, that's fine also, and we have great documentation andSwyx [00:21:17]: Or if they wanna use developer tooling.Peter [00:21:18]: Yeah, exactly.AI Coding Adoption: Cursor, Claude Code, and the Bimodal EngineerSwyx [00:21:19]: It's, like, a better together if, obviously, if you, if they work together. Is it all C++ I assume is with different compile targets?Peter [00:21:27]: We use a lot of C++.Peter [00:21:28]: Rust is sort of a hot, the new hot kid on the blockPeter [00:21:32]: for a bunch of things as well. But yeah, the lower level you get, especially when you get to real-time constraints, you hit C++ at some point, and at some point maybe you work your way into assembly when needed.Swyx [00:21:44]: Oh, damn.Alessio [00:21:46]: I'm curious about the coding agent adoption, just, like, since you're mentioning more esoteric languages. Like, what's the adoption internally? What have you learned?Peter [00:21:55]: Yeah. We use everything. So Cursor was, I think the hottest tool in the company for a good while. Now Claude Code, I think has taken the reign on that. We have a internal leader, leaderboard that we use just to sort of encourage adoptionPeter [00:22:09]: with-within the company. And yeah, it's, they're phenomenally useful. it's, Honestly, we take inspiration from some of those tools also in how we're adapting some of that mindset of thinking to the physical realm. Like if it's so easy to build an app for this or that thing that lives just on a screen, we can We're taking now a lot of the same ideas and applying that to, “Okay, well, if you wanted a physical machine to do something, how easy can we make that, using our own tooling and platform as well?”Alessio [00:22:40]: Are you changing any of, like, the OS architecture, kinda like the way you expose services to, like, be more AI friendly or?Peter [00:22:48]: Yeah, absolutely. The in the early days of our tools infrastructure work, it was a lot about, You had engineers that were experts in certain topics, but the things that you're dealing with, they're oftentimes more mathematical or more abstract, where actually GUI tools are very useful for certain things. Like as an example, we have a product we call Sensor Studio, which is, it helps you design the sensor suite for your autonomous vehicle, whether, again, it could be a car, it could be a drone, could be a mining equipment, could be a robot. And you place sensors in different places. You There's different, There's a library. You can understand what are the trade-offs that you're making in the design of that system, and that was, like, a very, a very GUI intensive, thing ‘cause it's a little more like a CAD tool in that senseSwyx [00:23:37]: YepPeter [00:23:37]: if you've seen CAD tools. Nowadays, though, right, we expose all of the underlying APIs for that and now using, AI agents, you can actually configure a sensor suite with just text and likely reach a better result than you could've through the GUI in the past, and we're taking that thinking now through the whole product portfolio.Swyx [00:23:57]: Another thing I was thinking about is just in terms of, like, AI, adoption, does it change your hiring at least a little bit, or how do you, how do you sort of manage engineers, differently?Peter [00:24:08]: Yeah. absolutely, it does. we, I think like every company in the Valley right now, are evolving our hiring practicesPeter [00:24:16]: because the skills required to be effective are changing so fast, right? you used to really select for just rote implementation ability and now it is more the AI engineer skill set, right? Where it's like, yeah, how to implement, but actually-Just banging out code is no longer the core job, right? It's, it's actually knowing what questions to ask, knowing how to tie, how to tie together these different AI tools. And so the interviews that we give now I think are way harder than they've ever been.Peter [00:24:46]: But we also allow, right, selective use of AI tools to solve the problems. And I think in that you start to see more of a bimodal distribution of engineers, right? You start to see like wow, there's, there's this subset of people that they really get it. Like they're, they're all in and they've, they've clearly invested the hours needed to learn these tools and how to be effective.Peter [00:25:09]: And then there's sort of the group of people that haven't done that, and that the productivity gap is just enormous. And so we're, we're trying to obviously select for the people that are really into this.Qasar [00:25:20]: I first wrote the my AI engineer piece three years ago, and when I first wrote about it, I was like, “Actually, not everyone should be an AI engineer,” ‘cause I think there's a there's an extremist stance where well, every software is an engineer is an AI engineer. And my actual example of people who should not be adopting AI was embedded systems and operating systems, and database people. Are they adopting AI?Peter [00:25:41]: I think it's the classic bitter lesson, topic, which is the Six months ago I would've said the same thing, but it's, it's becoming super useful for every domain.Qasar [00:25:53]: I'm sure.Peter [00:25:54]: Right? Like,Peter [00:25:56]: there was, I think six months ago, or maybe a year ago, if you tried to use, let's say the latest Claude model for writing shaders, GPU shaders, the results were probably underwhelming. And if you use the latest model now to do that kind of task, you're a little bit blown away, like, “Wow, that actually worked. That's amazing.” And we see the same thing in the embedded realm. No question though, especially when you get into safety critical systems, the human validation isPeter [00:26:25]: is 100% key. Like I You're not gonna trust your life to a an AI written software that's, that's not been very carefully, checked by humans. And so I think now the really the challenge is about that appropriate level of human validation for these safety critical systems.Verifiable Rewards, Evals, and Neural SimulationAlessio [00:26:41]: How do you think about, yeah, touching on the simulation side, I think verifiable reward and reinforcement learning is, like, the hottest thing. What have you done internally to build around that? And like, what gives you What makes you sleep at night? Like, if somebody's like, just web coding something or likeAlessio [00:26:57]: wants to try something new, you have like a good enough system. Because I think the opposite is also true, is like if it's super easy to write anythingAlessio [00:27:04]: then it puts a lot of work on like the verifiableAlessio [00:27:07]: side of it. Like, what does that look like for people?Peter [00:27:10]: Yeah. So verifiability, a broader bucket of like evaluations, right? Like how do you evaluate the results that you're, you're getting? I think this is probably the hardest problem right now, because the As the models get better, it can be harder and harder to find the faults on the system.Peter [00:27:29]: And so like the problem of doing proper eval to find those faults, like that problem also keeps getting harder as the models get better. But it's no less important than it's ever been, right? You still there are still going to be edge cases that are not met and whatnot. And so it's, it's a big area of investment for us. On the reinforcement learning topic, the key thing is there's all these new requirements that come to be in the latest generation of these technologies. So for example, end-to-end is the big thing right now in autonomy and physical AI, which is you can now train these models that can effectively take sensor data in and then put control signals out, and get really good results out of that. But the way that you train and improve those models is really different from the previous generations. And so to do reinforcement learning on an end-to-end model, you now need to actually simulate all the sensor data, right? So then this becomes a we call our, work in this neural simulation, but it'sPeter [00:28:26]: think of it like a hybrid of Gaussian, splatting and diffusion methods, and where you really care about performance. Like performance is everything. If you can't do enough simulation fast enough and cheap enough, you actually can't get results that are worthwhile, in the end. It also gets to a lot of our work in embedded systems, which is like performance critical work, and that performance optimization, performance criticality, it carries over to a lot of the model training work. because, like, the only way to make it affordable is it has to be really fast.Qasar [00:28:58]: I think it's worth a few minutes talking about our own, evolving thoughts on verification and validation withinQasar [00:29:05]: kind of, traditional simulators, which are, you can think of like vehicle dynamics or something like that, which you're just taking textbooks and taking those formulasQasar [00:29:13]: and putting them into software, to like now this neural sim/world model universe. I think that's an interesting topic.Peter [00:29:20]: Yeah. So in more traditional development, right, you oftentimes would have, more black-and-white answers to questions.Peter [00:29:28]: And so the in Europe as an example, there's, a regulatory, system, it's called Euro NCAP. It's the European New Car Assessment Program, and as part of that, the vehicles have to pass a bunch of tests, and those tests actually, include, safety systems. So automatic emergency braking for a child that runs in front of a carPeter [00:29:51]: or let's say an occluded child that runs out and you hit it. And so you have You end up with sort of these binary answers of like, well, did the car under test pass this specific test? And there's a very well-known set of test casesPeter [00:30:05]: that the vehicle has to pass. And that was how the industry worked, let's say, until 10-ish years ago. But what's changed now is with these models, everything is statistics, right? Like you no longer have a black-and-white answer, but it's like, well, how many orders of magnitude or how many nines of reliability can I get in the system, and how can I, how can I prove that to be true? And the big unlock honestly for physical AI as an industry is that these models are just becoming much more reliable. Right? Things like things actually work a lot better. It's like the number of nines you can get out of these systems are now good enough that it actually becomes cost effective to really deploy these things. And so the big shift in, so verification and validation has been from a little bit more of a Again the past it was strictly requirements, and are you meeting or not? And now it's more of a statistical, verification and validation case where it's all about how many nines of reliability and meantime between failures, that sort of thing.Statistical Validation, Regulators, and the Cruise LessonSwyx [00:31:04]: And is the target audience regulators or even the customers are yeah, if you I imagine the customers are bought in, and it's mostly regulators that need to be satisfied.Peter [00:31:15]: We do work with the US government, we do work of course with the European governments and the government of Japan, and the government is not like an AI lab by any means.Peter [00:31:25]: So Swyx [00:31:26]: They just care about the outcome.Peter [00:31:27]: They care about the outcome.Peter [00:31:28]: And so we do education, in that regard, and like so sort of teaching about, “Hey, this is how we think validation should be done, and this is an approach that we think is reasonable,” and how to think about like when is a driverless system actually safe enough to go on the roads and that sort of thing. But I wouldn't say that the government is asking for it. It's like we're more teaching the government in that, in that sense. It's honestly, it's more so for our own, our own comfort, right? Like, we want to build very safe systems, and then of course our customers care deeply about that as well. But in that context we're also typically educating our customers.Qasar [00:32:01]: Yeah. Our first, our first core value is on round safety. So I think we can't underline enough that, us also verifying and validating that the systems that we're deploying are safe to us is probably as important as, like, some regulator or a customer saying,Swyx [00:32:19]: Of course. Okay. Yeah.Swyx [00:32:20]: You have to satisfy yourselves.Peter [00:32:22]: As I say, as a whole across the world, regulation oftentimes it's like a almost lowest common denominator. But like, you really have to substantially exceed what the regulators are expecting to make good products.Swyx [00:32:33]: Yeah. One thing I often talk about, I think and I try to make this relatable to the audience also, is Cruise, where they had an accident that basically ended the company. I wonder if people overreact to single incidents, because incidents are going to happen regardless, right? ‘Cause it's a statistical thing, but as long I don't know if regulators understand that, you cannot extrapolate from a single incident, but we do because that's all we have to go on. And your sample sizes are necessarily gonna be lower than, I don't knowSwyx [00:33:00]: consumer driving.Qasar [00:33:01]: Yeah. I think the Cruise example wasn't a technology failure. there was The real, compounding issue there was just how did the company talk to the regulators and what was their kind of behavior, and I think that became more of the issue. If you look,Peter [00:33:19]: It isn't It definitely was a technology failure, but it was made much worse by theSwyx [00:33:23]: Put the car back on the woman.Qasar [00:33:25]: Yeah. And let me put it another way. There is a version where Cruise still exists.Swyx [00:33:29]: right. Right.Qasar [00:33:30]: Right. It'sSwyx [00:33:30]: It was like the last strawQasar [00:33:31]: ItSwyx [00:33:31]: in like a long chain ofSwyx [00:33:33]: like issues.Qasar [00:33:33]: So do you feel like ATG had that horrific accident or someone actually dying, because, that was a homeless person crossing the street? So yeah, I think we can't understate enough that ultimately, like, statistical validation of something, that's one part of it, but it's not the only part of it. Like, consumer and let's say, mainstream adoption of these technologies is also gonna be part of that conversation. I think companies like Waymo are doing a lot of service positively to the industry in the sense of they're, they're setting a high benchmark and they're showing, kind of in a very responsible way how to, how to deal with these. There have been Waymo incidences as well. They've just not been as significant as the Cruise one that you mentioned. But yeah, so I think you'll just continue to see that. I think probably the long term question is really gonna be, again, around Like it is very clear humans are way worse drivers statistically.Qasar [00:34:29]: Like, there's no, there's no debate. And so at what point But we're emotional animals.Swyx [00:34:34]: Yeah. So my thing is, like, we have to get to a point as a society where we accept horrific accidents that would never happen by a human because statistically we understand that it is safer overall. In the same way that planes, they're safer, than I think they're the safest mode of transport that we have.Qasar [00:34:50]: Yeah. it's more dangerous to drive to the airport than it is to get on a flight.Qasar [00:34:53]: So if you're everQasar [00:34:54]: if you're ever getting nervous about getting on a plane, just think “I just gotta get to the airport.”Swyx [00:34:58]: Yes, we're flying.Qasar [00:34:59]: If I get to the airportQasar [00:35:00]: I'll be good.Swyx [00:35:00]: But then it's, planes also concentrate the tail risk if planesQasar [00:35:03]: Yeah. AndPeter [00:35:04]: And I was, I don't think we honestly have to worry about there ever being, accidents from these systems that are like much worse than what humans would cause, ‘cause humans do terrible things.Peter [00:35:14]: Like, people fall asleep at the wheel all the time.Swyx [00:35:16]: I have.Swyx [00:35:17]: Like, I'll call, I've been a drowsy driver.Peter [00:35:19]: Kinda drunk drivers, and that'sPeter [00:35:20]: that's the extreme end of the example. But these AI systems, you have redundancies, you have fallbacks. Like, there's many things have to go wrong for there to actually be a something catastrophic because there's, there's so many, fallbacks that these systems have.Alessio [00:35:36]: your simulation is like so vast because there's so many use cases. What are, like, maybe things that worked in a simulation and then you put it out and it's like, “F**k, this isAlessio [00:35:45]: this just did not work at all?”Peter [00:35:47]: Yes.Alessio [00:35:47]: IsPeter [00:35:47]: That's maybe a bit of a misconception, about simulation there. So let me go a little bit, more technical on this. So at first go, no simulation is going to represent the real world. There's always a process of this, sim to real matchingPeter [00:36:02]: where you actually, you need the real world feedback to basically feed into the parameters that are being used in the simulator, and you have to do that, it's like this validation flow, a number of times until you can get some confidence that, like I think the simulator is now accurately representingPeter [00:36:19]: what's gonna happen in the real world. Now, if you have a situation where you've done that full validation and you thought that it was accurate and then there's something different, those are much trickier cases, and that's, that absolutely can happen, but really I think the validation process is a really important part. You can never skip the simulation validation process, like where you're actually ensuring that, hey, the actual, my sim to real gap here is small enough that I can trust these simulation results. And there's, there's so many fun things that you can do when you get into it. Like, I'll, I'll give one fun example that came up recently is like in these humanoid robotics, systemsOverheating actuators is a real problem, right? So obviously phenomenal demos. IPeter [00:37:01]: The most amazingAlessio [00:37:02]: For 10 minutes.Peter [00:37:03]: The most amazing I can get. I love, I love watching robots do acrobatics like everybody but the these systems actually overheat, right? If, like, And one of the ways you can use simulation though is you can actually have that, the temperature of those actuators be one of the parameters that's representedPeter [00:37:18]: in the simulation. And if you're doing reinforcement learning over a certain task, then the robot can actually adjust its motions in the simulation to account for the fact that, oh, it knows that as it's moving, it's actually beginning to overheat this motor. But if you didn't have that parameter of, let's say, the heat of that motor represented in the simulation initially, then your RL policy might It will disregard that. And now you run that on the robot and the robot will overheat and fail.Alessio [00:37:43]: I guess the question is, like, how do you have all of these parameters taken care of while also understanding the deployment environment? Like, temperature is like a great example, right? WellAlessio [00:37:53]: why did you make my robot worse when it runs in like a freezer?Alessio [00:37:57]: So it actually shouldn't worry about that. it's like, yeah, how do you design these simulations?Peter [00:38:02]: This is honestly the This is what makes simulation so hard, right? it's because you Simulation is fundamentally about you're trying to optimize the development of a system, right? Like, how can I build this system faster and better and cheaper and what are all the levers that I have to actually accomplish that? And because simulation's just a software program, you can, you can change it a lot more easily than you can hardware systems. And then what's particularly awesome about the let's say, world models and using that as a part of simulation is now the simulation doesn't just scale with, let's say, adding new math equations inPeter [00:38:36]: but we can actually scale the simulation environment now with additional real world data and that also unlocks a whole new field of robotics.Qasar [00:38:46]: There is a meniscus line where you cross where still doing real world testing is better. there's, in this, sim-to-real gap, you can reproduce reality at exceedingly expensive costs and this So nothing is free. So really you have to you're finding that line where you're getting great performance, you're getting great feedback, whether it's on the training side or on the eval side, but it's way cheaper than doing it in the real world. At some point it, that doesn't make sense. And so even, from our earliest days in autonomy, our view was you're still gonna do real world testing. You There's, there's not, there's not this, magical land where you're not gonna do that. And maybe even like a more nuanced version of this in like traditional software development is, most of your testing for software in a vehicle, 95% of that can be like traditional CI/CD kind of, flows that you would have in traditional web development. But once you have Now you, let's say you have a truck. Well, you can do like 4% of those in like a rig which has all the components, the electrical and electronics of a truck, but doesn't have, it doesn't have the tires and it doesn't have the And then you have the 1%, which is actually the vehicle. There's something There's a similar analogy in terms of using simulation for intelligent systems. You can do a lot in a simulator, but in using world models, but ultimately it's, it's physical AI. So you're gonna deploy it on physical machines andQasar [00:40:17]: the freezer example comes to, comes to light.Alessio [00:40:20]: The world model thing has been to me the hardest thing toAlessio [00:40:22]: wrap my head around. Like we have Faith Eliyon on the podcast.World Models, Hydroplaning, and Cause-Effect LearningQasar [00:40:25]: We've been doing a small series with like another Intuition company, General Intuition as well.Qasar [00:40:31]: yeah, and I mean, lots of, lots of coverage on NeRFs and yes.Alessio [00:40:34]: Yeah. It feels like we talk with about, the heliocentric system, right? It's like in a world model, if you just feed visual data, the model might learn that the sun spins around the Earth. It makes sense, right? And it's like, well, not really. And I think what are like some of these other things that like hydroplaning is one thing I think about, is like can a world model understand hydroplaning and like what amount of water like causes it to happen? And it's like, yeah, to me it's like I don't understand how you guys do it. I guess it's like the real thing is like when you're doing both cars and the highway in Japan versus the excavator in a mine in,Qasar [00:41:13]: ArizonaAlessio [00:41:13]: wherever you're Arizona, wherever you're deploying them.Alessio [00:41:15]: How much of it are you relying on the world models to like generate the simulations for you and then try and close the gap after versus like giving the world models as a tool to your engineers to like curate the simulations if that makes sense?Peter [00:41:28]: Yeah, totally. So yeah, I can say at a pure engineering level, I think if you're hoping to do real world deploys and you're purely relying on a world model approach, you probably won't get to something that works, before you go bankrupt. So there is just a very practical mindset of like, world models are amazing and they're extremely useful for a lot of use cases, but there are a lot of other things that you need to do to actually get something started and something deployed and working. most fundamentally, world models are all about It's understanding the world, but also understanding what's going to happen. It's like the cause-effect relationship.Peter [00:42:01]: Right? And so like it, right, if you have a take some sort of construction tool, and that construction tool is gonna be doing some work on the Earth in some way, it's gonna be moving earth, the world model needs to understand that cause-effect relationship. Like, okay, when I, when I take this material from here and put it over there and now I have things that are over here and not over there anymore and that cause-effect, relationship. data obviously is a is a big problem. The hydroplaningPeter [00:42:26]: one is actually a really great example because it's actually quite non-obvious sometimes. Right? It's like, well, it's, it's raining and well this road, has, let's say the appropriate curvature to it so the water is running off the road and cars are driving faster here and then you approach a road that's very flat and water is now puddling on that road and all of a sudden cars are driving slower because when they were driving faster they were starting to lose control. And there are a lot of visual nuance, very nuanced visual cues in the scene and so I do think in the world model concept there's a good chance that the model actually would learn that you should just drive slower when these visual cues exist, and that's obviously the beautiful-The beauty of, these kinds of models where they just, they learn these non-obvious things.Swyx [00:43:14]: It doesn't need to know about hydroplaning to know that it needs to drive slower.Peter [00:43:17]: Yes.Swyx [00:43:17]: I guess it's Yeah. I wanna ask questions about, also deploying models. I presume, like, you use a lot of these world models for training data and simulation, but what about deploying it onto the systems in production? Presumably you have you have, like, GPUs on deviceOnboard vs. Offboard: Latency, Embedded ML, and DistillationSwyx [00:43:36]: but they're I keep saying on device. What's the what's the right term for that?Peter [00:43:40]: On machine.Swyx [00:43:41]: On machine.Peter [00:43:41]: Or embedded, yeah.Swyx [00:43:42]: Yeah. What is the embedded world like? because for people who are not used to that world, this is very alien.Peter [00:43:49]: Yeah. So it's actually We call it onboard and off board.Peter [00:43:52]: So like, onboard software and off board software.Peter [00:43:54]: And the great thing about off board software is you don't have to care about time, and you can run really large models, right? So you can, you can say, “Well, this model, I don't care if it takes one second for it to give me a result or 10 seconds for it to give me a result, because we have time.” And the models can be really big, and they can run, in a data center or on a on a huge GPU and you can obviously have distribute to compute, et cetera. But onboard you don't have any of those benefits. You're like, “Well, I need I have this many milliseconds where I need an answer from this model.” And so a lot more of the energy then is about, think of it more like distillation and it's like truly efficiency and like, literally every fraction of a millisecond counts. And you can't have a situation where the model takes too long because then the vehicle can't actually function.Peter [00:44:42]: And so you can, you can still use a lot of the same techniques, and the models themselves you can think of as like a derivative of larger models that you can run offline, and then you're, you're trying to just get a model that is still performs really well but it's, it's a it's smaller, small enough version that you can then run on this embedded system where you care about latency and power.Qasar [00:45:03]: Yeah. And I think like, the broader point I think which, maybe is not obvious but it's worth saying is in physical AI world, we're not really constrained right now by, like, the intelligence of the models. It's actually what Peter's talking about, it's actually deploying them inSwyx [00:45:19]: The hardware they give you.Qasar [00:45:21]: Yeah. On the hardware you give you.Qasar [00:45:22]: And so And there's just a reality is of safety critical systems. So those end up being the your limiting factorsQasar [00:45:29]: rather than, let's say, a limiting factor for, a foundation model companyQasar [00:45:34]: is gonna be just capital maybe or researchers.Qasar [00:45:38]: So we're, we're in that way dealing with, for us as people who kind of come in that realm with like a very interesting Those constraints force creativity.Swyx [00:45:47]: And I imagine, nobody was deploying or giving you the hardware for transformers back in 2018, whatever, but now they are. What's the evolution like? just peel back the curtains a little bit.Peter [00:45:59]: Yeah. Transformers first off, I think the paper was originally published in 2017.Swyx [00:46:02]: 2017.Swyx [00:46:02]: So there's no time.Peter [00:46:04]: And ISwyx [00:46:05]: But I'm just saying I guess I'm saying, like, embedded ML systems usually, like, a lot less parameters, a lot less compute, and now, like, orders of magnitude more.Peter [00:46:14]: Yeah. absolutely. what I was gonna say though was I think in the in the original paper in 2017, maybe it's in the last paragraph, somewhere in the paper they talk about, like, “Oh, by the way, this technique might be useful for, like, images and videos as well.”Peter [00:46:30]: These last subjects.Peter [00:46:31]: And it took a few years for that impact to really hit. But like, now, we're seeing transformers are everywhere.Swyx [00:46:39]: Yeah. Vision transformers.Peter [00:46:40]: And then then the compute just keeps getting better and better. But you do have this fundamental trade-off, right? It's like you have power, you have cost, and performance and like, getting the right, getting the right mix of those things in an embedded package that can also be, like, shaken and baked in all thePeter [00:47:00]: conditions that these things have to have to operate in. But yeah, I think that they're only going to keep getting better and so we also try to plan our strategy understanding that, we know the rate of improvements of these systems.Swyx [00:47:11]: Yeah. So like, Google just released the Gemma 2B modelSwyx [00:47:15]: that effective 2B model. Is that useful to you guys or is that too big?Peter [00:47:18]: You can run that model on an embedded system, definitely.Peter [00:47:21]: the So yes, it's, it's useful in that regard. The bigger question is, like, what do you use it for in an embedded system? Like, you actually need to customize it quite a bit to make it useful for something. But yeah, you could run a two billion parameter model, definitely.Swyx [00:47:35]: It also interesting, like, what percent is a custom ML model that only does that thing versus a generalist LLMSwyx [00:47:41]: which probably is not that useful actually for your context.Peter [00:47:46]: Like, you, like, you can imagine different use cases, right?Peter [00:47:48]: So theSwyx [00:47:49]: The voice stuff, yes.Peter [00:47:49]: Yeah, the voice test. Totally, yes.Peter [00:47:51]: So for the actual, autonomy elements, that's 100% in-house. We do every bit of that, the data simulation, the model, everything. But when you get into the more generic use cases like voice or voice assistant kind of thing, that's where these more generalist models like Gemma actually can be quite, can be quite useful.Swyx [00:48:09]: Yeah. And then there's also obviously a trade-off between, like, what percent must you do on machine, versus just call home.Peter [00:48:16]: Yeah. It's all about latency.Swyx [00:48:17]: Latency.Peter [00:48:17]: It's all about latency. Yeah.Swyx [00:48:18]: Yeah. Well, like, I think actually in a lot of contexts, especially in the US, you can just have a connection to the web.Qasar [00:48:26]: Yeah. I think though most of our universe is everything has to be fairly, embedded and local because just the nature of Even in the US there's a lot of likeSwyx [00:48:39]: PatchinessQasar [00:48:40]: don't haveQasar [00:48:41]: have coverage, right? And if you look at, like, the old world of autonomy within mining, which is, like, long before transformers and kind of, neural networks, in the like CNN and kind of a universe, they were really just hand-coded, systems. They were just like, this machine is gonna run to that place with thisPeter [00:49:03]: That was our GPS, like very accurate GPS.Qasar [00:49:05]: Yeah. And so that worked, and that worked for 20 years, so why would we actually need to use transformers or kind of more modern end-to-end systems? Mainly because you can only really run a path and run backwards. That provided a lot of value, but m-Not as much as you get when the machine is actually intelligent. It's, it's seeing, it's perceiving, it's acting in a dynamic world.Alessio [00:49:28]: I looked up RTK, real-time kinematic, one to two-centimeter accuracy.Qasar [00:49:32]: Yeah. Fantastic. But the and fantastic in faraway lands where there's not gonna be cell phone coverage.Peter [00:49:39]: Yeah, so it's widely used on the legacy mining and agricultural autonomy systems today. So like, for example, a combine that can be precise within one or two centimeters as it's driving down the field, they use RTK.Qasar [00:49:53]: Yes.Peter [00:49:53]: But it's, it's expensive.Qasar [00:49:54]: Yeah. And it's, it's, it's autonomy, but it's not intelligent in the way that I think all of usQasar [00:49:58]: if in twenty-six we'd be talking about intelligence.Alessio [00:50:00]: In one of your blog posts, you mentioned research on large scale transformers that are similar to those doing modern generative AI. What are, like, the big differences other than, “You're absolutely right. I should steer the car, so you probably wanna remove that?”Peter [00:50:14]: We have a diversified bet strategy internally, and the reason we've done that is because we operate in now a bunch of industries, a bunch of geographies, and each of the approaches has, obviously a different risk to them.Peter [00:50:27]: And so like, we're not going to put all of our eggs in a single basket for a single approach because that approach may no

Welcome to How Humans Heal. In this episode, I want to address a question I hear all the time: does HPV ever really go away, or is it dormant and going to reappear again? I'm Dr. Doni Wilson, and I'm so glad you're here, because there is so much misinformation about high-risk HPV that the confusion is creating more risk for women and men when it comes to HPV-related cancer. I'm here to break those myths and give you the facts, so you can make real decisions to move forward and protect yourself. I'm sharing this based on over 26 years of clinical experience as a naturopathic doctor and midwife, helping thousands of women clear HPV to negative — and keep it negative for long periods of time, including cases that have been negative for more than 15 years at this point. I'm here to help you! LINKS FROM THE EPISODE:   Sign up for Dr. Doni's 5-Day HPV Workshop: https://doctordoni.com/HPV-workshop/   Schedule A Chat With Dr. Doni: https://intakeq.com/new/hhsnib/vuaovx    Read the full episode notes and find more information: https://doctordoni.com/blog/podcasts/ MORE RESOURCES FROM DR. DONI:   Quick links to social media, free guides and programs, and more: https://doctordoni.com/links     Disclosure: Some of the links in this post are product links and affiliate links and if you go through them to make a purchase I will earn a commission at no cost to you. Keep in mind that I link these companies and their products because of their quality and not because of the commission I receive from your purchases. The decision is yours, and whether or not you decide to buy something is completely up to you.  

Computer scientist Keith Winstein is an expert in how computers communicate. Computer networks create what he calls shared fictions – abstract realities, like a website or a Zoom call, that exist only because the computers on either end agree to act as if they are real. Unfortunately, today's networks lack a shared notion of a “computation,” which hurts market efficiency in cloud computing and frustrates efforts to hold tech companies accountable for the results of their algorithms. As computational power becomes concentrated in a smaller number of companies, Winstein advocates for a shared language of “computational truths,” defining computations precisely so results are reproducible and auditable. His research group hopes this will lead to greater transparency and accountability in the cloud and, ultimately, to greater confidence in the computations that companies do every day on our behalf. The truth matters, Winstein tells host Russ Altman on this episode of Stanford Engineering's The Future of Everything podcast. Have a question for Russ? Send it our way in writing or via voice memo, and it might be featured on an upcoming episode. Please introduce yourself, let us know where you're listening from, and share your question. You can send questions to thefutureofeverything@stanford.edu. Episode Reference Links: Stanford Profile: Keith Winstein Connect With Us: Episode Transcripts >>> The Future of Everything Website Connect with Russ >>> Threads / Bluesky / Mastodon Connect with School of Engineering >>> Twitter/X / Instagram / LinkedIn / Facebook Chapters: (00:00:00) Introduction Russ Altman introduces guest Keith Winstein, a professor of computer science and electrical engineering at Stanford University (00:02:56) Why Choose Networking The appeal of the shared digital “fictions” created by connected computers. (00:04:22) The Internet's Impact The broader societal implications of networking technologies. (00:05:35) Computational Truth The concept of tracking how data is produced and verified. (00:09:18) Misaligned Cloud Computing How “pay for effort” models create inefficiencies in cloud systems. (00:13:51) Determining Computational Truth The need for verifiable computation that produces consistent results. (00:18:19) Computations & Accountability How identifying computations could improve trust in systems. (00:20:56) Collaborating Online Why latency challenges make online performance collaboration difficult. (00:24:38) Real-Time Performance Systems Creating a custom system for musicians to perform together online. (00:28:00) Latency vs. Bandwidth Why faster internet speeds don't necessarily reduce delay. (00:30:43) Eliminating Latency How buffering layers in software create unnecessary delay. (00:32:41) Balancing Audio Quality & Delay The different trade-offs for musicians, actors, and audiences. (00:34:20) Rethinking Computer Science Education The need to bring playfulness and interactivity back into learning. (00:35:46) The Xylophone-Based Class Teaching computation through real-time sound and music. (00:38:34) Future In a Minute Rapid-fire Q&A: optimism, truth in computing, and innovation. (00:41:01) Conclusion Connect With Us:Episode Transcripts >>> The Future of Everything WebsiteConnect with Russ >>> Threads / Bluesky / MastodonConnect with School of Engineering >>>Twitter/X / Instagram / LinkedIn / Facebook Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.

In this talk, Nasser Qadri, AI Engineering Manager at Google, shares his unique career journey—from a PhD in Politics and International Relations to leading high-stakes AI initiatives. We explore the evolution of the AI Engineer role, the critical intersection of social science and machine learning, and how to build robust agentic workflows with engineering rigor.You'll learn about:- Moving beyond simple API calls to implementing full-stack engineering principles and "Agent Ops."- How a background in qualitative research and statistics provides a unique "moral compass" for building ethical AI.- A strategic roadmap for transitioning from non-traditional backgrounds into elite AI engineering roles.- Using design thinking and personal "pain points" to drive meaningful technical innovation.- Why traditional ML and model distillation will remain vital as we move from generalist LLMs to specialized, high-speed agents.- How to navigate the complex landscape of AI frameworks and build depth in your technical stack.TIMECODES:00:00 Transitioning from Social Science to Software Engineering07:45 Applying Statistical Rigor to Generative AI Evaluation12:10 Balancing Research Mindsets with Engineering Speed16:30 Managing Non-Deterministic Systems and Model Creativity20:15 Comparing AI Roles in Big Tech vs Startups24:40 Learning by Building: Solving Personal Pain Points31:50 Mental Frameworks for Problem Finders and Solvers36:15 Human-Centered Design in the Age of LLMs42:05 Beyond API Calls: Software Engineering Rigor for Agents45:50 Orchestration and the Rise of Agent Ops51:30 Depth vs Breadth in AI Framework Selection56:10 The Future of Latency and Traditional ML Integration1:01:20 When to Prioritize Model Distillation and Fine-Tuning1:02:10 Closing Thoughts and Future OutlookThis conversation is designed for software engineers, data scientists, and career-switchers looking to transition into the Generative AI space. It is particularly valuable for technical leaders in large organizations and startups who need to balance rapid AI prototyping with long-term system reliability.Connect with Nasser- Linkedin - https://www.linkedin.com/in/nasserq/Connect with DataTalks.Club:- Join the community - https://datatalks.club/slack.html- Subscribe to our Google calendar to have all our events in your calendar - https://calendar.google.com/calendar/r?cid=ZjhxaWRqbnEwamhzY3A4ODA5azFlZ2hzNjBAZ3JvdXAuY2FsZW5kYXIuZ29vZ2xlLmNvbQ- Check other upcoming events - https://lu.ma/dtc-events- GitHub: https://github.com/DataTalksClub- LinkedIn - https://www.linkedin.com/company/datatalks-club/ - Twitter - https://twitter.com/DataTalksClub - Website - https://datatalks.club/

Maher Hanafi is an engineering leader who went from zero AI experience to self-hosting LLMs at enterprise scale — managing GPU costs, optimizing inference with TensorRT LLM, and building an AI platform for HR tech. In this conversation, he breaks down exactly how his team cut latency by 70%, reduced GPU spend through counterintuitive scaling strategies, and navigated the messy reality of taking AI from proof-of-concept to production.How We Cut LLM Latency 70% With TensorRT in Production // MLOps Podcast #369 with Maher Hanafi, SVP of Engineering at Betterworks Key topics covered:The AI Iceberg — Why the invisible work behind AI (performance, latency, throughput, cost, accuracy) is harder than building the features themselvesGPU Cost Optimization — How upgrading to more expensive GPUs actually saved money by reducing total runtime hoursTensorRT LLM Deep Dive — Rewiring neural networks to match GPU architecture for 50-70% latency reductionCold Start Solutions — Using AWS FSx, baking models into container images, and cutting minutes off spin-up timesKV Cache & In-Flight Batching — Why using one model per GPU with maximum KV cache beats cramming multiple models togetherScheduled & Dynamic Scaling — Pattern-based scaling for HR tech workloads (nights, weekends, end-of-quarter spikes)Verticalized AI Platform — Building horizontal AI infrastructure that serves multiple HR product verticalsAI Engineering Lab — How junior vs. senior engineers adopted AI coding tools differently, and the cultural shift that followedAgentic Coding in Practice — Navigating AI coding agent costs, quality control, and redefining the SDLCChinese Models & Compliance — Why enterprise customers block DeepSeek/Qwen and the geopolitics of model training dataThis episode is for engineering leaders building AI in production, MLOps engineers optimizing GPU infrastructure, and anyone navigating the gap between AI demos and enterprise-scale deployment.Links & Resources:TensorRT LLM: https://github.com/NVIDIA/TensorRT-LLMNVIDIA Run: ai Model Streamer (cold start optimization): https://developer.nvidia.com/blog/reducing-cold-start-latency-for-llm-inference-with-nvidia-runai-model-streamer/vLLM vs TensorRT-LLM comparison: https://northflank.com/blog/vllm-vs-tensorrt-llm-and-how-to-run-themTimestamps: [00:00] Optimizing GPU Usage and Latency[00:21] Learning AI as Leadership[04:34] AI Cost Centers[13:56] Throughput and Infrastructure Efficiency[18:10] Scaling and Unit Economics[24:14] Championing AI ROI[36:11] Queue to Value Engine[41:30] Failed Product Features[46:12] Agentic Engineering Costs[58:49] AI Self-Hosting in Engineering[1:04:40] Wrap up

You asked, we answered. We take on the burning questions one by one.Before returning next week with regular episodes of the show.Thank you again for your support chooms!The title track link is hereMore info can be found here:⁠⁠⁠⁠ ⁠⁠⁠⁠linktr.ee/NoLatency⁠⁠⁠⁠⁠⁠⁠⁠Check out out Patreon!⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Patreon.com/nolatency⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Even more information and MERCH is on our website!Get a 10% DISCOUNT on the ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Official Rtalsorian Store⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Using Our Promocode: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠NOLATENCY⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠BlueSky & Twitter: @nolatencypodInstagram: @nolatencypodFind @SkullorJade,⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ @Miss_Magitek⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ and⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ @Binary_Dragon⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠,⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠  @retrodatv⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ on twitch, for live D&D Vampire the Masquerade and more.#cyberpunkred #actualplay #ttrpg #radioplay #scifi #cyberpunk #drama #comedy #Questions #answers #trivia

Send us Fan MailMustafa Poonawala is a globally recognized leader in medical device and diagnostics innovation, known for his ability to translate strategy into execution across R&D, clinical operations, and portfolio management. Over a career spanning more than two decades, he has built and led world-class engineering and program teams, guided products from early development through regulatory approval, and driven large-scale organizational transformation in highly regulated environments.Currently, Mustafa is the CEO of DynaMill Research, a specialized Clinical Research Organization focused on helping diagnostics companies dramatically reduce cycle times and improve cost predictability. DynaMill's approach blends agile program management, end-to-end digital clinical workflows, predictive enrollment strategies, and deep partnerships with multi-site clinical networks. The goal is simple but ambitious: help diagnostic innovations reach patients faster without sacrificing rigor or quality.In parallel, Mustafa is Managing Partner at Steps Program Management, where he has spent nearly a decade advising organizations on agile transformation, PMO maturity, and portfolio optimization—particularly within medical device R&D. His work emphasizes lean, value-driven processes, difficult prioritization, and delivery predictability, all grounded in real-world execution rather than theory.Previously, Mustafa held senior leadership roles at BD, Hospira, OBS Medical, and Boston Scientific. His experience spans implantable and disposable devices, complex electromechanical systems, software and cybersecurity for safety-critical systems, and large-scale diagnostics portfolios exceeding billions of dollars in revenue. With a PhD in Software Engineering focused on safety-critical systems, Mustafa brings a rare blend of deep technical rigor, business acumen, and servant leadership to every challenge he tackles.LINKS:Guest LinkedIn: https://www.linkedin.com/in/mustafapGuest website: https://dynamillcro.comAaron Moncur, host Subscribe to the show to get notified so you don't miss new episodes every Friday.The Being An Engineer podcast is brought to you by Pipeline Design & Engineering. Pipeline partners with medical & other device engineering teams who need turnkey equipment like cycle test machines, custom test fixtures, automation equipment, assembly jigs, inspection stations and more. You can find us at www.teampipeline.usWatch the show on YouTube: www.youtube.com/@TeamPipelineus 

Rising demand for AI computing power is reshaping data centre strategies worldwide. Sungai Samak Estate presents a rare opportunity with five strategically located plots supporting low-latency AI infrastructure, renewable energy integration, and water-secure operations, positioning Malaysia as a competitive hub for next-generation data centre development. Sungai Samak Estate City: Kuala Lumpur Address: 2 Jalan Sempurna off Jalan Gombak Website: https://sgsamak.com

SojuTalk is back at it again as we discuss releases from BTS, Yuna, Latency! As always, the Crew keeps you up to date with all the recent Kpop News/Events. And you know we gonna get hype as we declare this week's Spice King and give our State of the Nation!!! Links ◆Email - sojutalkpodcast@gmail.com ◆Discord - discord.gg/3rb74x4 ◆Patreon - patreon.com/sojutalk Timestamps ◆Intro - 0:00 ◆Big New Releases - 2:30 ◆Swim Discussion - 13:03 ◆SojuScore - 39:56 ◆Show Winners - 46:20 ◆News - 48:30 ◆Arirang Album Discussion - 1:26:58

In this episode of Run the Numbers, CJ sits down with Mike Wilcox, CFO of Blockchain.com, to unpack the economics of crypto exchanges. They discuss how platforms serve both retail traders and institutional clients, the different ways exchanges generate revenue, and the tension between blockchain's radical transparency and the valuable first-party data exchanges control.—SPONSORS:Brex is an intelligent finance platform that combines corporate cards, built-in expense management, and AI agents to eliminate manual finance work. By automating expense reviews and reconciliations, Brex gives CFOs more time for the high-impact work that drives growth. Join 35,000+ companies like Anthropic, Coinbase, and DoorDash at https://www.brex.com/metricsMetronome is real-time billing built for modern software companies. Metronome turns raw usage events into accurate invoices, gives customers bills they actually understand, and keeps finance, product, and engineering perfectly in sync. That's why category-defining companies like OpenAI and Anthropic trust Metronome to power usage-based pricing and enterprise contracts at scale. Focus on your product — not your billing. Learn more and get started at https://www.metronome.comRightRev is an automated revenue recognition platform built for modern pricing models like usage-based pricing, bundles, and mid-cycle upgrades. RightRev lets companies scale monetization without slowing down close or compliance. For RevRec that keeps growth moving, visit https://www.rightrev.comRillet is an AI-native ERP built for modern finance teams that want to close faster without fighting legacy systems. Designed to support complex revenue recognition, multi-entity operations, and real-time reporting, Rillet helps teams achieve a true zero-day close—with some customers closing in hours, not days. If you're scaling on an ERP that wasn't built in the 90s, book a demo at https://www.rillet.com/cjTabs is an AI-native revenue platform that unifies billing, collections, and revenue recognition for companies running usage-based or complex contracts. By bringing together ERP, CRM, and real product usage data into a single system of record, Tabs eliminates manual reconciliations and speeds up close and cash collection. Companies like Cortex, Statsig, and Cursor trust Tabs to scale revenue efficiently. Learn more at https://www.tabs.com/runAbacum is a modern FP&A platform built by former CFOs to replace slow, consultant-heavy planning tools. With self-service integrations and AI-powered workflows for forecasting, variance analysis, and scenario modeling, Abacum helps finance teams scale without becoming software admins. Trusted by teams at Strava, Replit, and JG Wentworth—learn more at https://www.abacum.ai—LINKS: Mostly Talent: https://mostlymetrics.typeform.com/to/cLTxtAsNMike: https://www.linkedin.com/in/mike-wilcox-65078a12/https://www.blockchain.com/CJ: https://www.linkedin.com/in/cj-gustafson-13140948/Mostly metrics: https://www.mostlymetrics.com—TIMESTAMPS:Here's the trimmed version:0:00 Preview and intro2:12 Tradfi to crypto transition3:49 Blockchain.com origin5:45 CFO as business partner6:01 Finance team backgrounds7:02 Banking relationships8:51 On ramps and off ramps8:51 Retail vs. institutional10:53 Sponsors — Brex | Metronome | RightRev14:09 B2C to B2B motion16:04 Shared infrastructure18:31 Go-to-market differences19:00 Brand equity and low CAC20:06 Education as top-of-funnel21:13 Institutional vs. retail volatility22:37 Exchange vs. brokerage model23:56 How brokerages make money24:06 Sponsors — Rillet | Tabs | Abacum27:31 Setting take rates29:13 Distribution flywheel30:13 Data as a moat31:13 Nigeria market playbook31:44 Crypto balance sheet33:25 Duration matching34:37 Transaction-level risk36:12 Latency arms race37:28 Stablecoins and CFOs38:13 Risk vectors40:01 Annual planning41:51 Lightning round43:00 Finance software stack43:31 Advice to younger self44:09 Credits

M

What happens when the real bottleneck in artificial intelligence is no longer training models, but actually running them at scale? In this episode of Tech Talks Daily, I sit down with Satyam Srivastava from d-Matrix to explore a shift that is quietly reshaping the entire AI infrastructure landscape. While much of the early AI race focused on training ever larger models, the next phase of AI adoption is increasingly defined by inference. That is the moment when trained models are deployed and used to generate real-world results millions of times a day. Satyam brings a unique perspective shaped by years of experience in signal processing, machine learning, and hardware architecture, including time spent at NVIDIA and Intel working on graphics, media technologies, and AI systems. Now at d-Matrix, he is helping design next-generation computing architectures focused on one of the biggest challenges facing the AI industry today: efficiently running large language models without overwhelming data centers with unsustainable power and infrastructure demands. During our conversation, we explored why the industry underestimated the infrastructure implications of inference at scale. While training large models grabs headlines, the real operational pressure often comes later when those models must serve millions of queries in real time. That shift places enormous strain on memory bandwidth, energy consumption, and data movement inside modern data centers. Satyam explains how d-Matrix identified this challenge years before generative AI exploded into the mainstream. Instead of focusing on training hardware like many AI startups at the time, the company concentrated on inference efficiency. That decision is becoming increasingly relevant as organizations begin to realize that simply adding more GPUs to data centers is not a sustainable long-term strategy. We also discuss the growing power constraints surrounding AI infrastructure, and why efficiency-driven design may be the only realistic path forward. With electricity supply, cooling capacity, and semiconductor availability all becoming limiting factors, the industry is being forced to rethink how AI systems are architected. Custom silicon, purpose-built accelerators, and heterogeneous computing environments are now emerging as key pieces of the puzzle. The conversation also touches on the geopolitical and economic importance of AI semiconductor leadership, and why the relationship between frontier AI labs, infrastructure providers, and chip designers is becoming increasingly strategic. As governments and companies compete to maintain technological leadership, the question of who controls the hardware powering AI may prove just as important as the models themselves. Looking ahead, Satyam shares his perspective on how the role of engineers will evolve as AI infrastructure becomes more specialized and energy-aware. Foundational engineering skills remain essential, but the next generation of engineers will also need to think in terms of entire systems, combining software, hardware, and AI tools to build more efficient computing environments. As AI continues to move from research labs into everyday products and services, are organizations prepared for the infrastructure shift that comes with an inference-driven future? And could efficiency, rather than raw computing power, become the defining metric of the next phase of the AI race?

Igal Raichelgauz, Founder & CEO, Autobrains joined Grayson Brulte on The Road to Autonomy podcast to discuss the company's strategic partnership with VinFast and the development of an affordable, scalable robo-car.The operational backbone of Autobrains' strategy is a Thinking AI approach that utilizes an agentic architecture rather than traditional monolithic models. By using a library of specific skills that can be added incrementally, the system scales from basic safety features to full autonomy without requiring massive data retraining or excessive computational power.In the field, Autobrains is rigorously applying its technology to the VinFast VF 8 and VF 9 models, proving the system's robustness in some of the world's most complex driving environments, such as the congested streets of Hanoi, Vietnam. Autobrains utilizes a vision-only approach that mimics human perception to navigate urban traffic, heavy rain, and high-speed highways.Autobrains' Physical AI ecosystem also includes an air to road localization system, which uses compressed satellite imagery signatures to provide 10-centimeter positioning accuracy. Allowing the vehicle to localize itself globally and understand lane boundaries or construction sites without relying on expensive, high-maintenance HD maps.Looking ahead, Igal envisions a future where autonomous driving reaches a mass-market inflection point within the next five years. This evolution aims to fundamentally transform the industry by delivering a fully autonomous robo-car at a $30,000 price point, enabling every vehicle to become a revenue-generating asset that increases safety and gives time back to the consumer.Episode Chapters00:00 How the VinFast Deal Came Together03:16 Skills-Based Agentic AI Architecture 07:16 Six Cameras, 360° Coverage, Low Compute 09:37 Air-to-Road: Satellite Imagery Replaces HD Maps12:40 Robo-car Vision 15:10 The $30K Fully Autonomous Car 20:20 The Thinking Layer24:22 20 Teraflops, Sub-20ms Latency, Edge Computing 27:58 No Lidar: The Vision-Only Thesis 28:59 The Future of Autobrains--------About The Road to AutonomyThe Road to Autonomy is the definitive media brand covering the Autonomy Economy™. Through our podcasts, newsletter, and proprietary market intelligence, we set the narrative for institutional investors, industry executives, and policymakers navigating the convergence of automation, autonomy, and economic growth.Join institutional investors and industry leaders who read This Week in The Autonomy Economy every Sunday. Each edition delivers exclusive insight and commentary on the autonomy economy, helping you stay ahead of what's next. Subscribe today for free: https://www.roadtoautonomy.com/ae/See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Jacob is so disgusted by a plate of deviled eggs that he has to throw them out before the show can begin. | Jay shares the text messages of the gym equipment company that refuses to make good on his purchase. | Kid Rock goes on Fox News to make excuses for the latency issue that hampered his halftime performance. No one remembers that he performed the actual Super Bowl in 2004. | Bobby reveals that his wife once waxed Britney Spears. *To hear the full show to go www.siriusxm.com/bonfire to learn more! FOLLOW THE CREW ON SOCIAL MEDIA: @thebonfiresxm @louisjohnson @christinemevans @bigjayoakerson @robertkellylive @louwitzkee @jjbwolf Subscribe to SiriusXM Podcasts+ to listen to new episodes of The Bonfire ad-free and a whole week early.  Start a free trial now on Apple Podcasts or by visiting siriusxm.com/podcastsplus. Hosted by Simplecast, an AdsWizz company. See https://pcm.adswizz.com for information about our collection and use of personal data for advertising.

Enjoyed our podcast? Shoot us a text and let us know—because great conversations never end at the last word!This week on TezTalks Radio, host Brandon Langston speaks with Yann Régis-Gianas, Head of Engineering at Nomadic Labs, about what Tezos X is — and more importantly, what it changes for the people actually using Tezos.Rather than focusing on abstract architecture, this conversation centers on experience. What does latency really mean? What is instant confirmation in practical terms? And when these pieces come together, how different does Tezos feel?

The Boys return with some wrapping up after a needed time off, but from Foundation to the Animorphs with some TTRPG lessons in between including lessons where you mislead to your players. All of that and more in the factually correct episode of Gayme Boys.

Analysts Don Kellogg and Roger Entner share insights from the recent T-Mobile Capital Markets Day update event, covering changes in the retail store strategy, new AI translation features, and updated reporting metrics.00:00 Episode intro 00:24 T-Mobile Capital Markets Day event overview 01:39 Pursuing convergence through fiber and FWA 02:37 Live AI translation could be a game-changer 04:03 Telecom must transcend the “dumb pipe” 05:08 Latency is key going forward 06:01 How AI actually impacts user experience 08:20 Key reporting metrics are changing 10:28 Episode wrap-upTags: telecom, telecommunications, wireless, prepaid, postpaid, cellular phone, Don Kellogg, Roger Entner, AI, T-Mobile, retail, FWA, fiber, convergence, USI, translation, network, Verizon, AT&T, latency, Apple, Android, churn, Srini Gopalan

Vodafone Ireland has once again been recognised as Ireland's leading mobile network, winning Best Mobile Internet Performance 2025 and Best 5G Network 2025 at this year's nPerf Awards, marking an exceptional fourth consecutive year in the top position. These latest wins follow years of continuous network improvement, with enhanced customer experience seen throughout the country. This result builds on last year's achievement, where Vodafone was awarded Best Mobile Internet Performance and Best 5G Provider for 2024. The 2025 results reinforce Vodafone's continued leadership in reliability, speed, and customer experience. This accolade further strengthens Vodafone Ireland's position as the country's most trusted mobile network, having also been independently recognised by umlaut as Ireland's Best Mobile Network for an unparalleled ten consecutive years. According to nPerf's independent analysis, based on thousands of real-world tests carried out by mobile users across Ireland, Vodafone achieved a total score of 98 361nPoints, outperforming all competitors across critical measures of network quality. The company excelled in multiple categories and was the best-performing provider across four – Upload Speed, Latency, Web Browsing, and YouTube Streaming – showcasing reliable performance in the 5G sector. Vodafone continues to advance its nationwide mobile network with €100 million invested annually to expand coverage, boost capacity, and deliver best-in-class mobile experiences. This sustained investment is driving significant modernisation across the network, including accelerated 4G and 5G rollout, enhanced resilience, and technology upgrades designed to meet Ireland's growing demand for fast, reliable connectivity. This investment forms part of Vodafone Ireland's wider €500 million five?year network investment cycle, which is delivering extensive upgrades and new site developments nationwide. As part of this long-term programme, Vodafone is delivering network enhancements across the country — recent examples in the last number of weeks include Carlow Town, Clane, Greystones, Mountcollins in Co. Limerick and the Mount Falcon region in Co. Mayo. These reflect just a snapshot of the wider, ongoing investment aimed at boosting coverage, increasing capacity and strengthening reliability for customers wherever they live, work and travel. In parallel, Vodafone continues to support Ireland's growing data demand, with annual mobile data traffic rising sharply at 20% year-on-year, ensuring customers can rely on a resilient, future proof network wherever they live, work, and travel. Sheila Kavanagh, Director of Networks, Vodafone Ireland, said: "We are incredibly proud to secure the nPerf award for the fourth year in a row. This result reflects the dedication of our teams and the impact of our continued investment, including €100 million each year — to deliver a fast, reliable and resilient network for customers across Ireland. As data usage grows and customer expectations evolve, we remain focused on building a future proof network that supports homes, businesses and communities, from major cities to remote rural areas. Our ambition is simple: to provide an outstanding, high quality experience every single day." Recognising Vodafone's continued success in 2025 Sébastien de Rosbo, Managing Director nPerf, stated: "Vodafone Ireland once again achieved the highest overall performance across our 2025 benchmarks. Based on large-scale, real-world testing carried out by users nationwide, Vodafone delivered consistently strong results across key indicators including mobile internet performance and 5G experience. This fourth consecutive win reflects the strength and consistency of Vodafone's network for customers across Ireland." See more stories here.

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]:

Guest: Dr. Sharon Lewin is the Director of the Peter Doherty Institute for Infection and Immunity, where her team studies HIV. She talks about the current landscape in HIV research and treatments, and how new therapies could target latent viral reservoirs. Featured Products and Resources: Register now for IMMUNOLOGY2026! Make the Easy Choice. Try EasySep to Win! The Immunology Science Round Up Immunosurveillance in the Skin: A neuro-epithelial axis can tune regional immunosurveillance against melanoma. B Cells in Aging: B cells contributed to the age-related reduction of naive CD4 T cells. The Gut–Brain Axis in Parkinson’s: Muscularis macrophages, housekeepers of intestinal homeostasis, modulate α-synuclein pathology and neurodegeneration in models of Parkinson’s disease. How IL-2 Signaling Regulates Inflammation: IL-2 signaling promotes the generation of IL-10pos age-associated B cells, with implications for autoimmunity and inflammation. Image courtesy of Dr. Sharon Lewin Subscribe to our newsletter! Never miss updates about new episodes. Subscribe

Topics covered in this episode: Command Book App uvx.sh: Install Python tools without uv or Python Ending 15 years of subprocess polling monty: A minimal, secure Python interpreter written in Rust for use by AI Extras Joke Watch on YouTube About the show Sponsored by us! Support our work through: Our courses at Talk Python Training The Complete pytest Course Patreon Supporters Connect with the hosts Michael: @mkennedy@fosstodon.org / @mkennedy.codes (bsky) Brian: @brianokken@fosstodon.org / @brianokken.bsky.social Show: @pythonbytes@fosstodon.org / @pythonbytes.fm (bsky) Join us on YouTube at pythonbytes.fm/live to be part of the audience. Usually Monday at 10am PT. Older video versions available there too. Finally, if you want an artisanal, hand-crafted digest of every week of the show notes in email form? Add your name and email to our friends of the show list, we'll never share it. Michael #1: Command Book App New app from Michael Command Book App is a native macOS app for developers, data scientists, AI enthusiasts and more. This is a tool I've been using lately to help build Talk Python, Python Bytes, Talk Python Training, and many more applications. It's a bit like advanced terminal commands or complex shell aliases, but hosted outside of your terminal. This leaves the terminal there for interactive commands, exploration, short actions. Command Book manages commands like "tail this log while I'm developing the app", "Run the dev web server with true auto-reload", and even "Run MongoDB in Docker with exactly the settings I need" I'd love it if you gave it a look, shared it with your team, and send me feedback. Has a free version and paid version. Build with Swift and Swift UI Check it out at https://commandbookapp.com Brian #2: uvx.sh: Install Python tools without uv or Python Tim Hopper Michael #3: Ending 15 years of subprocess polling by Giampaolo Rodola The standard library's subprocess module has relied on a busy-loop polling approach since the timeout parameter was added to Popen.wait() in Python 3.3, around 15 years ago The problem with busy-polling CPU wake-ups: even with exponential backoff (starting at 0.1ms, capping at 40ms), the system constantly wakes up to check process status, wasting CPU cycles and draining batteries. Latency: there's always a gap between when a process actually terminates and when you detect it. Scalability: monitoring many processes simultaneously magnifies all of the above. + L1/L2 CPU cache invalidations It's interesting to note that waiting via poll() (or kqueue()) puts the process into the exact same sleeping state as a plain time.sleep() call. From the kernel's perspective, both are interruptible sleeps. Here is the merged PR for this change. Brian #4: monty: A minimal, secure Python interpreter written in Rust for use by AI Samuel Colvin and others at Pydantic Still experimental “Monty avoids the cost, latency, complexity and general faff of using a full container based sandbox for running LLM generated code. “ “Instead, it lets you safely run Python code written by an LLM embedded in your agent, with startup times measured in single digit microseconds not hundreds of milliseconds.” Extras Brian: Expertise is the art of ignoring - Kevin Renskers You don't need to master the language. You need to master your slice. Learning everything up front is wasted effort. Experience changes what you pay attention to. I hate fish - Rands (Michael Lopp) Really about productivity systems And a nice process for dealing with email Michael: Talk Python now has a CLI New essay: It's not vibe coding - Agentic engineering GitHub is having a day Python 3.14.3 and 3.13.12 are available Wall Street just lost $285 billion because of 13 markdown files Joke: Silence, current side project!

We welcome back Andrew Sillifant, Solution Director at Pure Storage, for a deep dive into the concept of data gravity. We start with the traditional 2010 definition coined by Dave McCrory—that data accumulates, making it harder to move, and forcing dependent systems to cluster nearby. However, Andrew presents his core thesis, arguing that this foundational principle is no longer sufficient in a world of exploding complexity. Our conversation emphasizes the need to re-examine data gravity through a modern lens, acknowledging the massive shift to cloud computing and the proliferation of interconnected systems over the last decade. Andrew introduces five crucial dimensions that now describe data's impact: Volume, redefined by context and classification; Dependency, now accelerated by API calls, integration points, and AI agents; Criticality, which includes regulations, security, and implicit SLAs; Velocity, measured by how many functions data is used for; and Latency, complicated by geographic requirements that skew response times. These dimensions highlight how non-physical constraints, like egress fees and data sovereignty laws, create artificial friction that compounds the problem beyond sheer data size. Our discussion concludes with a new framework of five sources of data gravity that IT leaders must address: Technical Gravity (the physical component and mobility), Economic Gravity (the costs of hosting and moving data, like egress fees), Regulatory Gravity (compliance and legal restrictions), Institutional Gravity (the dependency on a small number of people who know how to manage old systems), and Measurement Gravity (budgeting and decision-making risks). Finally, Andrew connects these challenges to Pure Storage, noting how platform features like deduplication and continuous innovation are actively working to lessen the effects of data gravity for customers. To learn more, visit https://blog.purestorage.com/purely-technical/the-economics-of-data-gravity/ Check out the new Pure Storage digital customer community to join the conversation with peers and Pure experts: https://purecommunity.purestorage.com/ 00:00 Intro and Welcome 01:05 Andrew Observations About the USA 04:19 Defining Data Gravity 07:30 Challenges Caused By Data Gravity 09:01 Real World Data Gravity Examples 17:15 Data Gravity Impact Vectors 33:02 New Dimensions of Data Gravity 40:30 Where Pure Helps with Data Gravity

From Pepcom at CES 2026 in Las Vegas, Paul Tan, COO for Keychron, about the new Q Ultra Series keyboards, their best yet. Highlights include an 8K wireless polling rate with ultra-low latency, exceptional battery life measured in months, a premium CNC-milled aluminum build, upgraded switches, and browser-based configuration that eliminates software downloads that allow for deep customization.  Show Notes: Chapters: Links: Guests: Support:      Become a MacVoices Patron on Patreon     http://patreon.com/macvoices      Enjoy this episode? Make a one-time donation with PayPal Connect:      Web:     http://macvoices.com      Twitter:     http://www.twitter.com/chuckjoiner     http://www.twitter.com/macvoices      Mastodon:     https://mastodon.cloud/@chuckjoiner      Facebook:     http://www.facebook.com/chuck.joiner      MacVoices Page on Facebook:     http://www.facebook.com/macvoices/      MacVoices Group on Facebook:     http://www.facebook.com/groups/macvoice      LinkedIn:     https://www.linkedin.com/in/chuckjoiner/      Instagram:     https://www.instagram.com/chuckjoiner/ Subscribe:      Audio in iTunes     Video in iTunes      Subscribe manually via iTunes or any podcatcher:      Audio: http://www.macvoices.com/rss/macvoicesrss      Video: http://www.macvoices.com/rss/macvoicesvideorss

From Pepcom at CES 2026 in Las Vegas, Paul Tan, COO for Keychron, about the new Q Ultra Series keyboards, their best yet. Highlights include an 8K wireless polling rate with ultra-low latency, exceptional battery life measured in months, a premium CNC-milled aluminum build, upgraded switches, and browser-based configuration that eliminates software downloads that allow for deep customization.  Show Notes: Chapters: Links: Guests: Support:      Become a MacVoices Patron on Patreon      http://patreon.com/macvoices      Enjoy this episode? Make a one-time donation with PayPal Connect:      Web:      http://macvoices.com      Twitter:      http://www.twitter.com/chuckjoiner      http://www.twitter.com/macvoices      Mastodon:      https://mastodon.cloud/@chuckjoiner      Facebook:      http://www.facebook.com/chuck.joiner      MacVoices Page on Facebook:      http://www.facebook.com/macvoices/      MacVoices Group on Facebook:      http://www.facebook.com/groups/macvoice      LinkedIn:      https://www.linkedin.com/in/chuckjoiner/      Instagram:      https://www.instagram.com/chuckjoiner/ Subscribe:      Audio in iTunes      Video in iTunes      Subscribe manually via iTunes or any podcatcher:      Audio: http://www.macvoices.com/rss/macvoicesrss      Video: http://www.macvoices.com/rss/macvoicesvideorss

Snow in the forecast, bass in the bloodstream. We kick things off with a little storm prep and then slide straight into three new releases that light up different corners of the pop universe: a girl band debut with real instruments and real restraint, a solo star leaning into neon funk, and a multinational rookie squad ringing the alarm with swagger.First, LATENCY's “It Was Love” trades flash for feel. Five members play and sing, shaping a soft, mid-tempo band sound that grows from quiet reflection to gentle resolve. Clean guitars, steady percussion, and a tasteful piano-synth color create space for melodies that stick without shouting. By the time the guitar solo lifts around the midpoint, the song has earned its momentum—intimate, bittersweet, and clear-eyed.Then Kento Nakajima flips the room with “XTC”. Think syncopated 80s synths, jazzy guitar licks, and a rhythm that snaps you into motion. Kenty's vocal control is the secret weapon, gliding from sultry lows to sharp falsetto as the track dances between playful invitation and seductive tension. It's a confident solo statement that blends disco shimmer with modern polish, perfect for late-night drives or living room dance breaks.Finally, meet ALPHA DRIVE ONE and their adrenaline-spiked “FREAK ALARM.” The intro hums like a clock before exploding into 80s hip-hop charge—reverb bass, crisp drums, and a chorus that stretches like a siren. The lyrics double as a manifesto: they're the “new alien,” inviting you to own your edge and step into the circle. It's a debut with grit and charm, made to wake up your playlist.If you want a soundtrack for the week—reflection, heat, and high-voltage confidence—this one's for you. Listen, save your favorite moments, and tell us which chorus won your day.LATENCY Instagram X YouTube It Was LoveKento Nakajima Instagram X YouTube XTCALPHA DRIVE ONE Instagram X YouTube FREAK ALARMSupport the showPlease help Music Elixir by rating, reviewing, and sharing the episode. We appreciate your support!Follow us on:TwitterInstagram BlueskyIf have questions, comments, or requests click on our form:Music Elixir FormDJ Panic Blog:OK ASIA

Cursive Cursing. A Cat Named Beef. Getting Twitchy. Let's Test the Show's Latency. Sour creme on your chalupa is not a euphemism. Emails dash jugs of pee. A Baja Better Time. You Need A Grommet. Imma seal up my hole. The hoity toity mall. Shrodenger's seating. Panky and the brian. But I would have liked to watch you struggle a little bit. Flower Child Something. Belter Passwords with Tom and more on this episode of The Morning Stream. Hosted on Acast. See acast.com/privacy for more information.

Cursive Cursing. A Cat Named Beef. Getting Twitchy. Let's Test the Show's Latency. Sour creme on your chalupa is not a euphemism. Emails dash jugs of pee. A Baja Better Time. You Need A Grommet. Imma seal up my hole. The hoity toity mall. Shrodenger's seating. Panky and the brian. But I would have liked to watch you struggle a little bit. Flower Child Something. Belter Passwords with Tom and more on this episode of The Morning Stream. Hosted on Acast. See acast.com/privacy for more information.

As constraints on energy, water, and permitting collide with exploding demand for AI and compute, a once-fringe idea is moving rapidly toward the center of the conversation: putting data centers in space. Starcloud believes orbital infrastructure isn't science fiction—it's a necessary extension of the global compute stack if scaling is going to continue at anything close to its current pace.Founded by Philip Johnston, Starcloud is building space-based compute systems designed to compete on cost, performance, and scale with terrestrial data centers. The company has already flown a data center–grade GPU in orbit and is now working toward larger, commercially viable systems that could reshape where and how AI is powered. We discuss:How energy and permitting constraints are reshaping the future of computeWhy space-based data centers may be economically inevitable, not optionalWhat Starcloud proved by running an H100 GPU in orbitHow launch costs, watts-per-kilogram, and chip longevity define the real economicsThe national security implications of who controls future compute capacity • Chapters •00:00 - Intro00:50 - The issue with data centers02:20 - Explosion of the data center debates04:58 - Philip's 5GW data center rendering and early conceptions of data centers in space at YC08:16 - Proving people wrong11:17 - The team at Starcloud today12:29 - Competing against SpaceX's data center14:42 - Sam Altman's beef with Starlink16:52 - Economics of Orbital vs Terrestrial Data Centers by Andrew McCallip21:33 - Where are we putting these things?23:50 - Latency in space25:59 - Political side of building data centers28:36 - Starcloud 130:16 - Space based processors30:51 - Shakespeare in space32:00 - Hardening an Nvidia H100 against radiation and making chips in space economical34:43 - Cooling systems in space36:01 - How Starcloud is thinking about replacing failed GPUs38:46 - The mission for Starcloud 240:05 - Competitors outside of SpaceX40:49 - Getting to economical launch costs44:35 - Will the next great wars be over water and power for data centers?46:25 - What keeps Philip up at night?47:11 - What keeps Mo up at night? • Show notes •Starcloud's website — https://www.starcloud.com/Philip's socials — https://x.com/PhilipJohnstonMo's socials — https://x.com/itsmoislamPayload's socials — https://twitter.com/payloadspace / https://www.linkedin.com/company/payloadspaceIgnition's socials — https://twitter.com/ignitionnuclear /  https://www.linkedin.com/company/ignition-nuclear/Tectonic's socials — https://twitter.com/tectonicdefense / https://www.linkedin.com/company/tectonicdefense/Valley of Depth archive — Listen: https://pod.payloadspace.com/ • About us •Valley of Depth is a podcast about the technologies that matter — and the people building them. Brought to you by Arkaea Media, the team behind Payload (space), Ignition (nuclear energy), and Tectonic (defense tech), this show goes beyond headlines and hype. We talk to founders, investors, government officials, and military leaders shaping the future of national security and deep tech. From breakthrough science to strategic policy, we dive into the high-stakes decisions behind the world's hardest technologies.Payload: www.payloadspace.comTectonic: www.tectonicdefense.comIgnition: www.ignition-news.com

"How do I create the 'wow' moment for my end user? And slowness never creates the wow moment." PolarGrid founder and CEO Rade Kovacevic believes GenAI video and voice will be killer apps once they can function in real-time. Enabling real-time GenAI requires uncorking the inference bottleneck that the hyperscalers have helped build. The BetaKit podcast is presented by Fasken Emerging Tech, supporting trailblazing startups, venture capital funds and acquirers of high-growth tech companies for over 30 years. If you're curious about the health of Canada's tech M&A scene, you've got to check out Exit InSights. It's a first-of-its-kind report from Fasken's Emerging Technology & Venture Capital Group that analyses private M&A activity among VC-backed and high-growth tech companies. You'll learn how buyers and sellers are maximizing value, minimizing risk, and navigating one of the most vibrant tech ecosystems out there. Download your free copy of the report.

In this episode, Sahaj Garg, CTO of wispr.ai, joins SE Radio host Robert Blumen to talk about the challenges of building low-latency AI applications. They discuss latency's effect on consumer behavior as well as interactive applications. The conversation explores how to measure latency and how scale impacts it. Then Sahaj and Robert shift to themes around AI, including whether "AI" means LLMs or something broader, as they look at latency requirements and challenges around subtypes of AI applications. The final part of the episode explores techniques for managing latency in AI: speed vs accuracy trade-offs; speed vs cost; latency vs cost; choosing the right model; reducing quantization; distillation; and guessing + validating.

Healthcare organizations are navigating modernization under intense regulatory, security, and resource constraints. This episode explores how the Microsoft technology stack shows up differently in healthcare. The conversation breaks down hybrid cloud realities, Azure managed services, security and compliance, business resiliency, disaster recovery, and cost optimization, all grounded in real healthcare use cases. The episode also explores at how organizations can measure ROI beyond cost savings, connecting Microsoft investments to patient care, clinician experience, and operational resilience. Speakers: Jennifer Johnson, Director of Healthcare at Connection David Carey and Kevin Paiva, Senior Field Solution Architects at Connection Show Notes: 00:10 Welcome and session overview 01:40 Why healthcare cloud adoption is different 02:10 Defining hybrid cloud in healthcare 03:00 Why hybrid is now the default model 03:55 Latency myths and performance realities 04:45 Which workloads belong on-prem vs. in the cloud 05:45 SaaS, staffing pressure, and infrastructure complexity 06:30 Azure managed services and Connection's approach 07:45 Co-managed Azure vs. fully outsourced models 08:30 Why Azure over other hyperscalers 09:20 Azure security, HIPAA, and Zero Trust 10:30 Azure Health Data Services 11:45 Business continuity vs. business resiliency 14:10 What healthcare leaders worry about most today 15:00 Disaster recovery and Azure Expert MSP 16:30 Post-pandemic resource constraints 17:30 Application sprawl, security, and identity management 18:50 Cost containment and ROI in healthcare IT 21:15 The teams behind Connection's Microsoft practice 24:45 Final takeaways and next steps

In this episode of the Crazy Wisdom podcast, host Stewart Alsop interviews Marcin Dymczyk, CPO and co-founder of SevenSense Robotics, exploring the fascinating world of advanced robotics and AI. Their conversation covers the evolution from traditional "standard" robotics with predetermined pathways to advanced robotics that incorporates perception, reasoning, and adaptability - essentially the AGI of physical robotics. Dymczyk explains how his company builds "the eyes and brains of mobile robots" using camera-based autonomy algorithms, drawing parallels between robot sensing systems and human vision, inner ear balance, and proprioception. The discussion ranges from the technical challenges of sensor fusion and world models to broader topics including robotics regulation across different countries, the role of federalism in innovation, and how recent geopolitical changes are driving localized high-tech development, particularly in defense applications. They also touch on the democratization of robotics for small businesses and the philosophical implications of increasingly sophisticated AI systems operating in physical environments. To learn more about SevenSense, visit www.sevensense.ai.Check out this GPT we trained on the conversationTimestamps00:00 Introduction to Robotics and Personal Journey05:27 The Evolution of Robotics: From Standard to Advanced09:56 The Future of Robotics: AI and Automation12:09 The Role of Edge Computing in Robotics17:40 FPGA and AI: The Future of Robotics Processing21:54 Sensing the World: How Robots Perceive Their Environment29:01 Learning from the Physical World: Insights from Robotics33:21 The Intersection of Robotics and Manufacturing35:01 Journey into Robotics: Education and Passion36:41 Practical Robotics Projects for Beginners39:06 Understanding Particle Filters in Robotics40:37 World Models: The Future of AI and Robotics41:51 The Black Box Dilemma in AI and Robotics44:27 Safety and Interpretability in Autonomous Systems49:16 Regulatory Challenges in Robotics and AI51:19 Global Perspectives on Robotics Regulation54:43 The Future of Robotics in Emerging Markets57:38 The Role of Engineers in Modern WarfareKey Insights1. Advanced robotics transcends traditional programming through perception and intelligence. Dymczyk distinguishes between standard robotics that follows rigid, predefined pathways and advanced robotics that incorporates perception and reasoning. This evolution enables robots to make autonomous decisions about navigation and task execution, similar to how humans adapt to unexpected situations rather than following predetermined scripts.2. Camera-based sensing systems mirror human biological navigation. SevenSense Robotics builds "eyes and brains" for mobile robots using multiple cameras (up to eight), IMUs (accelerometers/gyroscopes), and wheel encoders that parallel human vision, inner ear balance, and proprioception. This redundant sensing approach allows robots to navigate even when one system fails, such as operating in dark environments where visual sensors are compromised.3. Edge computing dominates industrial robotics due to connectivity and security constraints. Many industrial applications operate in environments with poor connectivity (like underground grocery stores) or require on-premise solutions for confidentiality. This necessitates powerful local processing capabilities rather than cloud-dependent AI, particularly in automotive factories where data security about new models is paramount.4. Safety regulations create mandatory "kill switches" that bypass AI decision-making. European and US regulatory bodies require deterministic safety systems that can instantly stop robots regardless of AI reasoning. These systems operate like human reflexes, providing immediate responses to obstacles while the main AI brain handles complex navigation and planning tasks.5. Modern robotics development benefits from increasingly affordable optical sensors. The democratization of 3D cameras, laser range finders, and miniature range measurement chips (costing just a few dollars from distributors like DigiKey) enables rapid prototyping and innovation that was previously limited to well-funded research institutions.6. Geopolitical shifts are driving localized high-tech development, particularly in defense applications. The changing role of US global leadership and lessons from Ukraine's drone warfare are motivating countries like Poland to develop indigenous robotics capabilities. Small engineering teams can now create battlefield-effective technology using consumer drones equipped with advanced sensors.7. The future of robotics lies in natural language programming for non-experts. Dymczyk envisions a transformation where small business owners can instruct robots using conversational language rather than complex programming, similar to how AI coding assistants now enable non-programmers to build applications through natural language prompts.

De Rosalia à Little Simz, en passant par Bad Bunny, Sophian Fanen balaye son année 2025. En cette fin d'année, Sophian vous fait plaisir et sélectionne ses obsessions favorites. Rosalía, Berghain, tiré de l'album Lux (Columbia Records, 2025) Bad Bunny, Weltita (feat. Chuwi), tiré de l'album Debí Tirar Más Fotos (Rimas Entertainment)  Theodora, Ils me rient tous au nez, tiré de l'album Mega BBL (Boss Lady, 2025)  Tarta Relena, Si veriash a la rana, tiré de l'album És pregunta (Latency, 2024) Ale hop & Titi Bakorta, Bonne année, tiré de l'album Mapambazuko (Nyege Nyege Tapes, 2025)  Andrea Laszlo de Simone, Quando, tiré de l'album Una Lunghissima Ombra (Ekler/Hamburger Records, 2025)  Blaiz Fayah, Maureen et DJ Glad, Money Pull Up, tiré de l'album Shatta Ting (Creepy Music, 2025)  Little Simz, Lion (feat. Obongjayar), tiré de l'album Lotus (Awal, 2025)  Miki, Jtm encore, tiré de l'album Industry Plant (Structure, 2025)  Xania Monet, How Was I Supposed to Know?, tiré de l'album Unfolded (TMJ/Hallwood, 2025) 

In this week's episode of Hands-On Tech, Lance asks Mikah Sargent about the pros and cons of using powerline ethernet adapters, and Mikah shares his strong thoughts on these devices. Don't forget to send in your questions for Mikah to answer during the show! hot@twit.tv Host: Mikah Sargent Download or subscribe to Hands-On Tech at https://twit.tv/shows/hands-on-tech Want access to the ad-free video and exclusive features? Become a member of Club TWiT today! https://twit.tv/clubtwit Club TWiT members can discuss this episode and leave feedback in the Club TWiT Discord.

In this week's episode of Hands-On Tech, Lance asks Mikah Sargent about the pros and cons of using powerline ethernet adapters, and Mikah shares his strong thoughts on these devices. Don't forget to send in your questions for Mikah to answer during the show! hot@twit.tv Host: Mikah Sargent Download or subscribe to Hands-On Tech at https://twit.tv/shows/hands-on-tech Want access to the ad-free video and exclusive features? Become a member of Club TWiT today! https://twit.tv/clubtwit Club TWiT members can discuss this episode and leave feedback in the Club TWiT Discord.

In this week's episode of Hands-On Tech, Lance asks Mikah Sargent about the pros and cons of using powerline ethernet adapters, and Mikah shares his strong thoughts on these devices. Don't forget to send in your questions for Mikah to answer during the show! hot@twit.tv Host: Mikah Sargent Download or subscribe to Hands-On Tech at https://twit.tv/shows/hands-on-tech Want access to the ad-free video and exclusive features? Become a member of Club TWiT today! https://twit.tv/clubtwit Club TWiT members can discuss this episode and leave feedback in the Club TWiT Discord.