Latent Space: The AI Engineer Podcast

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

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

Tickets for AIEi Miami and AIE Europe are live, with first wave speakers announced!From pioneering software-defined networking to backing many of the most aggressive AI model companies of this cycle, Martin Casado and Sarah Wang sit at the center of the capital, compute, and talent arms race reshaping the tech industry. As partners at a16z investing across infrastructure and growth, they’ve watched venture and growth blur, model labs turn dollars into capability at unprecedented speed, and startups raise nine-figure rounds before monetization.Martin and Sarah join us to unpack the new financing playbook for AI: why today’s rounds are really compute contracts in disguise, how the “raise → train → ship → raise bigger” flywheel works, and whether foundation model companies can outspend the entire app ecosystem built on top of them. They also share what’s underhyped (boring enterprise software), what’s overheated (talent wars and compensation spirals), and the two radically different futures they see for AI’s market structure.We discuss:* Martin’s “two futures” fork: infinite fragmentation and new software categories vs. a small oligopoly of general models that consume everything above them* The capital flywheel: how model labs translate funding directly into capability gains, then into revenue growth measured in weeks, not years* Why venture and growth have merged: $100M–$1B hybrid rounds, strategic investors, compute negotiations, and complex deal structures* The AGI vs. product tension: allocating scarce GPUs between long-term research and near-term revenue flywheels* Whether frontier labs can out-raise and outspend the entire app ecosystem built on top of their APIs* Why today’s talent wars ($10M+ comp packages, $B acqui-hires) are breaking early-stage founder math* Cursor as a case study: building up from the app layer while training down into your own models* Why “boring” enterprise software may be the most underinvested opportunity in the AI mania* Hardware and robotics: why the ChatGPT moment hasn’t yet arrived for robots and what would need to change* World Labs and generative 3D: bringing the marginal cost of 3D scene creation down by orders of magnitude* Why public AI discourse is often wildly disconnected from boardroom reality and how founders should navigate the noiseShow Notes:* “Where Value Will Accrue in AI: Martin Casado & Sarah Wang” - a16z show* “Jack Altman & Martin Casado on the Future of Venture Capital”* World Labs—Martin Casado• LinkedIn: https://www.linkedin.com/in/martincasado/• X: https://x.com/martin_casadoSarah Wang• LinkedIn: https://www.linkedin.com/in/sarah-wang-59b96a7• X: https://x.com/sarahdingwanga16z• https://a16z.com/Timestamps00:00:00 – Intro: Live from a16z00:01:20 – The New AI Funding Model: Venture + Growth Collide00:03:19 – Circular Funding, Demand & “No Dark GPUs”00:05:24 – Infrastructure vs Apps: The Lines Blur00:06:24 – The Capital Flywheel: Raise → Train → Ship → Raise Bigger00:09:39 – Can Frontier Labs Outspend the Entire App Ecosystem?00:11:24 – Character AI & The AGI vs Product Dilemma00:14:39 – Talent Wars, $10M Engineers & Founder Anxiety00:17:33 – What’s Underinvested? The Case for “Boring” Software00:19:29 – Robotics, Hardware & Why It’s Hard to Win00:22:42 – Custom ASICs & The $1B Training Run Economics00:24:23 – American Dynamism, Geography & AI Power Centers00:26:48 – How AI Is Changing the Investor Workflow (Claude Cowork)00:29:12 – Two Futures of AI: Infinite Expansion or Oligopoly?00:32:48 – If You Can Raise More Than Your Ecosystem, You Win00:34:27 – Are All Tasks AGI-Complete? Coding as the Test Case00:38:55 – Cursor & The Power of the App Layer00:44:05 – World Labs, Spatial Intelligence & 3D Foundation Models00:47:20 – Thinking Machines, Founder Drama & Media Narratives00:52:30 – Where Long-Term Power Accrues in the AI StackTranscriptLatent.Space - Inside AI’s $10B+ Capital Flywheel — Martin Casado & Sarah Wang of a16z[00:00:00] Welcome to Latent Space (Live from a16z) + Meet the Guests[00:00:00] Alessio: Hey everyone. Welcome to the Latent Space podcast, live from a 16 z. Uh, this is Alessio founder Kernel Lance, and I’m joined by Twix, editor of Latent Space.[00:00:08] swyx: Hey, hey, hey. Uh, and we’re so glad to be on with you guys. Also a top AI podcast, uh, Martin Cado and Sarah Wang. Welcome, very[00:00:16] Martin Casado: happy to be here and welcome.[00:00:17] swyx: Yes, uh, we love this office. We love what you’ve done with the place. Uh, the new logo is everywhere now. It’s, it’s still getting, takes a while to get used to, but it reminds me of like sort of a callback to a more ambitious age, which I think is kind of[00:00:31] Martin Casado: definitely makes a statement.[00:00:33] swyx: Yeah.[00:00:34] Martin Casado: Not quite sure what that statement is, but it makes a statement.[00:00:37] swyx: Uh, Martin, I go back with you to Netlify.[00:00:40] Martin Casado: Yep.[00:00:40] swyx: Uh, and, uh, you know, you create a software defined networking and all, all that stuff people can read up on your background. Yep. Sarah, I’m newer to you. Uh, you, you sort of started working together on AI infrastructure stuff.[00:00:51] Sarah Wang: That’s right. Yeah. Seven, seven years ago now.[00:00:53] Martin Casado: Best growth investor in the entire industry.[00:00:55] swyx: Oh, say[00:00:56] Martin Casado: more hands down there is, there is. [00:01:00] I mean, when it comes to AI companies, Sarah, I think has done the most kind of aggressive, um, investment thesis around AI models, right? So, worked for Nom Ja, Mira Ia, FEI Fey, and so just these frontier, kind of like large AI models.[00:01:15] I think, you know, Sarah’s been the, the broadest investor. Is that fair?[00:01:20] Venture vs. Growth in the Frontier Model Era[00:01:20] Sarah Wang: No, I, well, I was gonna say, I think it’s been a really interesting tag, tag team actually just ‘cause the, a lot of these big C deals, not only are they raising a lot of money, um, it’s still a tech founder bet, which obviously is inherently early stage.[00:01:33] But the resources,[00:01:36] Martin Casado: so many, I[00:01:36] Sarah Wang: was gonna say the resources one, they just grow really quickly. But then two, the resources that they need day one are kind of growth scale. So I, the hybrid tag team that we have is. Quite effective, I think,[00:01:46] Martin Casado: what is growth these days? You know, you don’t wake up if it’s less than a billion or like, it’s, it’s actually, it’s actually very like, like no, it’s a very interesting time in investing because like, you know, take like the character around, right?[00:01:59] These tend to [00:02:00] be like pre monetization, but the dollars are large enough that you need to have a larger fund and the analysis. You know, because you’ve got lots of users. ‘cause this stuff has such high demand requires, you know, more of a number sophistication. And so most of these deals, whether it’s US or other firms on these large model companies, are like this hybrid between venture growth.[00:02:18] Sarah Wang: Yeah. Total. And I think, you know, stuff like BD for example, you wouldn’t usually need BD when you were seed stage trying to get market biz Devrel. Biz Devrel, exactly. Okay. But like now, sorry, I’m,[00:02:27] swyx: I’m not familiar. What, what, what does biz Devrel mean for a venture fund? Because I know what biz Devrel means for a company.[00:02:31] Sarah Wang: Yeah.[00:02:32] Compute Deals, Strategics, and the ‘Circular Funding’ Question[00:02:32] Sarah Wang: You know, so a, a good example is, I mean, we talk about buying compute, but there’s a huge negotiation involved there in terms of, okay, do you get equity for the compute? What, what sort of partner are you looking at? Is there a go-to market arm to that? Um, and these are just things on this scale, hundreds of millions, you know, maybe.[00:02:50] Six months into the inception of a company, you just wouldn’t have to negotiate these deals before.[00:02:54] Martin Casado: Yeah. These large rounds are very complex now. Like in the past, if you did a series A [00:03:00] or a series B, like whatever, you’re writing a 20 to a $60 million check and you call it a day. Now you normally have financial investors and strategic investors, and then the strategic portion always still goes with like these kind of large compute contracts, which can take months to do.[00:03:13] And so it’s, it’s very different ties. I’ve been doing this for 10 years. It’s the, I’ve never seen anything like this.[00:03:19] swyx: Yeah. Do you have worries about the circular funding from so disease strategics?[00:03:24] Martin Casado: I mean, listen, as long as the demand is there, like the demand is there. Like the problem with the internet is the demand wasn’t there.[00:03:29] swyx: Exactly. All right. This, this is like the, the whole pyramid scheme bubble thing, where like, as long as you mark to market on like the notional value of like, these deals, fine, but like once it starts to chip away, it really Well[00:03:41] Martin Casado: no, like as, as, as, as long as there’s demand. I mean, you know, this, this is like a lot of these sound bites have already become kind of cliches, but they’re worth saying it.[00:03:47] Right? Like during the internet days, like we were. Um, raising money to put fiber in the ground that wasn’t used. And that’s a problem, right? Because now you actually have a supply overhang.[00:03:58] swyx: Mm-hmm.[00:03:59] Martin Casado: And even in the, [00:04:00] the time of the, the internet, like the supply and, and bandwidth overhang, even as massive as it was in, as massive as the crash was only lasted about four years.[00:04:09] But we don’t have a supply overhang. Like there’s no dark GPUs, right? I mean, and so, you know, circular or not, I mean, you know, if, if someone invests in a company that, um. You know, they’ll actually use the GPUs. And on the other side of it is the, is the ask for customer. So I I,

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 t

This podcast features Gabriele Corso and Jeremy Wohlwend, co-founders of Boltz and authors of the Boltz Manifesto, discussing the rapid evolution of structural biology models from AlphaFold to their own open-source suite, Boltz-1 and Boltz-2. The central thesis is that while single-chain protein structure prediction is largely “solved” through evolutionary hints, the next frontier lies in modeling complex interactions (protein-ligand, protein-protein) and generative protein design, which Boltz aims to democratize via open-source foundations and scalable infrastructure.Full Video PodOn YouTube!Timestamps* 00:00 Introduction to Benchmarking and the “Solved” Protein Problem* 06:48 Evolutionary Hints and Co-evolution in Structure Prediction* 10:00 The Importance of Protein Function and Disease States* 15:31 Transitioning from AlphaFold 2 to AlphaFold 3 Capabilities* 19:48 Generative Modeling vs. Regression in Structural Biology* 25:00 The “Bitter Lesson” and Specialized AI Architectures* 29:14 Development Anecdotes: Training Boltz-1 on a Budget* 32:00 Validation Strategies and the Protein Data Bank (PDB)* 37:26 The Mission of Boltz: Democratizing Access and Open Source* 41:43 Building a Self-Sustaining Research Community* 44:40 Boltz-2 Advancements: Affinity Prediction and Design* 51:03 BoltzGen: Merging Structure and Sequence Prediction* 55:18 Large-Scale Wet Lab Validation Results* 01:02:44 Boltz Lab Product Launch: Agents and Infrastructure* 01:13:06 Future Directions: Developpability and the “Virtual Cell”* 01:17:35 Interacting with Skeptical Medicinal ChemistsKey SummaryEvolution of Structure Prediction & Evolutionary Hints* Co-evolutionary Landscapes: The speakers explain that breakthrough progress in single-chain protein prediction relied on decoding evolutionary correlations where mutations in one position necessitate mutations in another to conserve 3D structure.* Structure vs. Folding: They differentiate between structure prediction (getting the final answer) and folding (the kinetic process of reaching that state), noting that the field is still quite poor at modeling the latter.* Physics vs. Statistics: RJ posits that while models use evolutionary statistics to find the right “valley” in the energy landscape, they likely possess a “light understanding” of physics to refine the local minimum.The Shift to Generative Architectures* Generative Modeling: A key leap in AlphaFold 3 and Boltz-1 was moving from regression (predicting one static coordinate) to a generative diffusion approach that samples from a posterior distribution.* Handling Uncertainty: This shift allows models to represent multiple conformational states and avoid the “averaging” effect seen in regression models when the ground truth is ambiguous.* Specialized Architectures: Despite the “bitter lesson” of general-purpose transformers, the speakers argue that equivariant architectures remain vastly superior for biological data due to the inherent 3D geometric constraints of molecules.Boltz-2 and Generative Protein Design* Unified Encoding: Boltz-2 (and BoltzGen) treats structure and sequence prediction as a single task by encoding amino acid identities into the atomic composition of the predicted structure.* Design Specifics: Instead of a sequence, users feed the model blank tokens and a high-level “spec” (e.g., an antibody framework), and the model decodes both the 3D structure and the corresponding amino acids.* Affinity Prediction: While model confidence is a common metric, Boltz-2 focuses on affinity prediction—quantifying exactly how tightly a designed binder will stick to its target.Real-World Validation and Productization* Generalized Validation: To prove the model isn’t just “regurgitating” known data, Boltz tested its designs on 9 targets with zero known interactions in the PDB, achieving nanomolar binders for two-thirds of them.* Boltz Lab Infrastructure: The newly launched Boltz Lab platform provides “agents” for protein and small molecule design, optimized to run 10x faster than open-source versions through proprietary GPU kernels.* Human-in-the-Loop: The platform is designed to convert skeptical medicinal chemists by allowing them to run parallel screens and use their intuition to filter model outputs.TranscriptRJ [00:05:35]: But the goal remains to, like, you know, really challenge the models, like, how well do these models generalize? And, you know, we’ve seen in some of the latest CASP competitions, like, while we’ve become really, really good at proteins, especially monomeric proteins, you know, other modalities still remain pretty difficult. So it’s really essential, you know, in the field that there are, like, these efforts to gather, you know, benchmarks that are challenging. So it keeps us in line, you know, about what the models can do or not.Gabriel [00:06:26]: Yeah, it’s interesting you say that, like, in some sense, CASP, you know, at CASP 14, a problem was solved and, like, pretty comprehensively, right? But at the same time, it was really only the beginning. So you can say, like, what was the specific problem you would argue was solved? And then, like, you know, what is remaining, which is probably quite open.RJ [00:06:48]: I think we’ll steer away from the term solved, because we have many friends in the community who get pretty upset at that word. And I think, you know, fairly so. But the problem that was, you know, that a lot of progress was made on was the ability to predict the structure of single chain proteins. So proteins can, like, be composed of many chains. And single chain proteins are, you know, just a single sequence of amino acids. And one of the reasons that we’ve been able to make such progress is also because we take a lot of hints from evolution. So the way the models work is that, you know, they sort of decode a lot of hints. That comes from evolutionary landscapes. So if you have, like, you know, some protein in an animal, and you go find the similar protein across, like, you know, different organisms, you might find different mutations in them. And as it turns out, if you take a lot of the sequences together, and you analyze them, you see that some positions in the sequence tend to evolve at the same time as other positions in the sequence, sort of this, like, correlation between different positions. And it turns out that that is typically a hint that these two positions are close in three dimension. So part of the, you know, part of the breakthrough has been, like, our ability to also decode that very, very effectively. But what it implies also is that in absence of that co-evolutionary landscape, the models don’t quite perform as well. And so, you know, I think when that information is available, maybe one could say, you know, the problem is, like, somewhat solved. From the perspective of structure prediction, when it isn’t, it’s much more challenging. And I think it’s also worth also differentiating the, sometimes we confound a little bit, structure prediction and folding. Folding is the more complex process of actually understanding, like, how it goes from, like, this disordered state into, like, a structured, like, state. And that I don’t think we’ve made that much progress on. But the idea of, like, yeah, going straight to the answer, we’ve become pretty good at.Brandon [00:08:49]: So there’s this protein that is, like, just a long chain and it folds up. Yeah. And so we’re good at getting from that long chain in whatever form it was originally to the thing. But we don’t know how it necessarily gets to that state. And there might be intermediate states that it’s in sometimes that we’re not aware of.RJ [00:09:10]: That’s right. And that relates also to, like, you know, our general ability to model, like, the different, you know, proteins are not static. They move, they take different shapes based on their energy states. And I think we are, also not that good at understanding the different states that the protein can be in and at what frequency, what probability. So I think the two problems are quite related in some ways. Still a lot to solve. But I think it was very surprising at the time, you know, that even with these evolutionary hints that we were able to, you know, to make such dramatic progress.Brandon [00:09:45]: So I want to ask, why does the intermediate states matter? But first, I kind of want to understand, why do we care? What proteins are shaped like?Gabriel [00:09:54]: Yeah, I mean, the proteins are kind of the machines of our body. You know, the way that all the processes that we have in our cells, you know, work is typically through proteins, sometimes other molecules, sort of intermediate interactions. And through that interactions, we have all sorts of cell functions. And so when we try to understand, you know, a lot of biology, how our body works, how disease work. So we often try to boil it down to, okay, what is going right in case of, you know, our normal biological function and what is going wrong in case of the disease state. And we boil it down to kind of, you know, proteins and kind of other molecules and their interaction. And so when we try predicting the structure of proteins, it’s critical to, you know, have an understanding of kind of those interactions. It’s a bit like seeing the difference between... Having kind of a list of parts that you would put it in a car and seeing kind of the car in its final form, you know, seeing the car really helps you understand what it does. On the other hand, kind of going to your question of, you know, why do we care about, you know, how the protein falls or, you know, how the car is made to some extent is that, you know, sometimes when something goes wrong, you know, there are, you know, cases of, you know, proteins misfolding. In some diseases and so on, if we don’t understand this folding process, we don’t really know how to intervene.RJ [00:11:30]: There’s this nice line in the, I think it’s in the Alpha Fold 2 manuscrip

From Palantir and Two Sigma to building Goodfire into the poster-child for actionable mechanistic interpretability, Mark Bissell (Member of Technical Staff) and Myra Deng (Head of Product) are trying to turn “peeking inside the model” into a repeatable production workflow by shipping APIs, landing real enterprise deployments, and now scaling the bet with a recent $150M Series B funding round at a $1.25B valuation.In this episode, we go far beyond the usual “SAEs are cool” take. We talk about Goodfire’s core bet: that the AI lifecycle is still fundamentally broken because the only reliable control we have is data and we post-train, RLHF, and fine-tune by “slurping supervision through a straw,” hoping the model picks up the right behaviors while quietly absorbing the wrong ones. Goodfire’s answer is to build a bi-directional interface between humans and models: read what’s happening inside, edit it surgically, and eventually use interpretability during training so customization isn’t just brute-force guesswork.Mark and Myra walk through what that looks like when you stop treating interpretability like a lab demo and start treating it like infrastructure: lightweight probes that add near-zero latency, token-level safety filters that can run at inference time, and interpretability workflows that survive messy constraints (multilingual inputs, synthetic→real transfer, regulated domains, no access to sensitive data). We also get a live window into what “frontier-scale interp” means operationally (i.e. steering a trillion-parameter model in real time by targeting internal features) plus why the same tooling generalizes cleanly from language models to genomics, medical imaging, and “pixel-space” world models.We discuss:* Myra + Mark’s path: Palantir (health systems, forward-deployed engineering) → Goodfire early team; Two Sigma → Head of Product, translating frontier interpretability research into a platform and real-world deployments* What “interpretability” actually means in practice: not just post-hoc poking, but a broader “science of deep learning” approach across the full AI lifecycle (data curation → post-training → internal representations → model design)* Why post-training is the first big wedge: “surgical edits” for unintended behaviors likereward hacking, sycophancy, noise learned during customization plus the dream of targeted unlearning and bias removal without wrecking capabilities* SAEs vs probes in the real world: why SAE feature spaces sometimes underperform classifiers trained on raw activations for downstream detection tasks (hallucination, harmful intent, PII), and what that implies about “clean concept spaces”* Rakuten in production: deploying interpretability-based token-level PII detection at inference time to prevent routing private data to downstream providers plus the gnarly constraints: no training on real customer PII, synthetic→real transfer, English + Japanese, and tokenization quirks* Why interp can be operationally cheaper than LLM-judge guardrails: probes are lightweight, low-latency, and don’t require hosting a second large model in the loop* Real-time steering at frontier scale: a demo of steering Kimi K2 (~1T params) live and finding features via SAE pipelines, auto-labeling via LLMs, and toggling a “Gen-Z slang” feature across multiple layers without breaking tool use* Hallucinations as an internal signal: the case that models have latent uncertainty / “user-pleasing” circuitry you can detect and potentially mitigate more directly than black-box methods* Steering vs prompting: the emerging view that activation steering and in-context learning are more closely connected than people think, including work mapping between the two (even for jailbreak-style behaviors)* Interpretability for science: using the same tooling across domains (genomics, medical imaging, materials) to debug spurious correlations and extract new knowledge up to and including early biomarker discovery work with major partners* World models + “pixel-space” interpretability: why vision/video models make concepts easier to see, how that accelerates the feedback loop, and why robotics/world-model partners are especially interesting design partners* The north star: moving from “data in, weights out” to intentional model design where experts can impart goals and constraints directly, not just via reward signals and brute-force post-training—Goodfire AI* Website: https://goodfire.ai* LinkedIn: https://www.linkedin.com/company/goodfire-ai/* X: https://x.com/GoodfireAIMyra Deng* Website: https://myradeng.com/* LinkedIn: https://www.linkedin.com/in/myra-deng/* X: https://x.com/myra_dengMark Bissell* LinkedIn: https://www.linkedin.com/in/mark-bissell/* X: https://x.com/MarkMBissellFull Video EpisodeTimestamps00:00:00 Introduction00:00:05 Introduction to the Latent Space Podcast and Guests from Goodfire00:00:29 What is Goodfire? Mission and Focus on Interpretability00:01:01 Goodfire’s Practical Approach to Interpretability00:01:37 Goodfire’s Series B Fundraise Announcement00:02:04 Backgrounds of Mark and Myra from Goodfire00:02:51 Team Structure and Roles at Goodfire00:05:13 What is Interpretability? Definitions and Techniques00:05:30 Understanding Errors00:07:29 Post-training vs. Pre-training Interpretability Applications00:08:51 Using Interpretability to Remove Unwanted Behaviors00:10:09 Grokking, Double Descent, and Generalization in Models00:10:15 404 Not Found Explained00:12:06 Subliminal Learning and Hidden Biases in Models00:14:07 How Goodfire Chooses Research Directions and Projects00:15:00 Troubleshooting Errors00:16:04 Limitations of SAEs and Probes in Interpretability00:18:14 Rakuten Case Study: Production Deployment of Interpretability00:20:45 Conclusion00:21:12 Efficiency Benefits of Interpretability Techniques00:21:26 Live Demo: Real-Time Steering in a Trillion Parameter Model00:25:15 How Steering Features are Identified and Labeled00:26:51 Detecting and Mitigating Hallucinations Using Interpretability00:31:20 Equivalence of Activation Steering and Prompting00:34:06 Comparing Steering with Fine-Tuning and LoRA Techniques00:36:04 Model Design and the Future of Intentional AI Development00:38:09 Getting Started in Mechinterp: Resources, Programs, and Open Problems00:40:51 Industry Applications and the Rise of Mechinterp in Practice00:41:39 Interpretability for Code Models and Real-World Usage00:43:07 Making Steering Useful for More Than Stylistic Edits00:46:17 Applying Interpretability to Healthcare and Scientific Discovery00:49:15 Why Interpretability is Crucial in High-Stakes Domains like Healthcare00:52:03 Call for Design Partners Across Domains00:54:18 Interest in World Models and Visual Interpretability00:57:22 Sci-Fi Inspiration: Ted Chiang and Interpretability01:00:14 Interpretability, Safety, and Alignment Perspectives01:04:27 Weak-to-Strong Generalization and Future Alignment Challenges01:05:38 Final Thoughts and Hiring/Collaboration Opportunities at GoodfireTranscriptShawn Wang [00:00:05]: So welcome to the Latent Space pod. We’re back in the studio with our special MechInterp co-host, Vibhu. Welcome. Mochi, Mochi’s special co-host. And Mochi, the mechanistic interpretability doggo. We have with us Mark and Myra from Goodfire. Welcome. Thanks for having us on. Maybe we can sort of introduce Goodfire and then introduce you guys. How do you introduce Goodfire today?Myra Deng [00:00:29]: Yeah, it’s a great question. So Goodfire, we like to say, is an AI research lab that focuses on using interpretability to understand, learn from, and design AI models. And we really believe that interpretability will unlock the new generation, next frontier of safe and powerful AI models. That’s our description right now, and I’m excited to dive more into the work we’re doing to make that happen.Shawn Wang [00:00:55]: Yeah. And there’s always like the official description. Is there an understatement? Is there an unofficial one that sort of resonates more with a different audience?Mark Bissell [00:01:01]: Well, being an AI research lab that’s focused on interpretability, there’s obviously a lot of people have a lot that they think about when they think of interpretability. And I think we have a pretty broad definition of what that means and the types of places that can be applied. And in particular, applying it in production scenarios, in high stakes industries, and really taking it sort of from the research world into the real world. Which, you know. It’s a new field, so that hasn’t been done all that much. And we’re excited about actually seeing that sort of put into practice.Shawn Wang [00:01:37]: Yeah, I would say it wasn’t too long ago that Anthopic was like still putting out like toy models or superposition and that kind of stuff. And I wouldn’t have pegged it to be this far along. When you and I talked at NeurIPS, you were talking a little bit about your production use cases and your customers. And then not to bury the lead, today we’re also announcing the fundraise, your Series B. $150 million. $150 million at a 1.25B valuation. Congrats, Unicorn.Mark Bissell [00:02:02]: Thank you. Yeah, no, things move fast.Shawn Wang [00:02:04]: We were talking to you in December and already some big updates since then. Let’s dive, I guess, into a bit of your backgrounds as well. Mark, you were at Palantir working on health stuff, which is really interesting because the Goodfire has some interesting like health use cases. I don’t know how related they are in practice.Mark Bissell [00:02:22]: Yeah, not super related, but I don’t know. It was helpful context to know what it’s like. Just to work. Just to work with health systems and generally in that domain. Yeah.Shawn Wang [00:02:32]: And Mara, you were at Two Sigma, which actually I was also at Two Sigma back in the day. Wow, nice.Myra Deng [00:02:37]: Did we overlap at all?Shawn Wang [00:02:38]: No, this is when I was briefly a software engineer before I became a sort of developer

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

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

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

Happy New Year! You may have noticed that in 2025 we had moved toward YouTube as our primary podcasting platform. As we’ll explain in the next State of Latent Space post, we’ll be doubling down on Substack again and improving the experience for the over 100,000 of you who look out for our emails and website updates!We first mentioned Artificial Analysis in 2024, when it was still a side project in a Sydney basement. They then were one of the few Nat Friedman and Daniel Gross’ AIGrant companies to raise a full seed round from them and have now become the independent gold standard for AI benchmarking—trusted by developers, enterprises, and every major lab to navigate the exploding landscape of models, providers, and capabilities.We have chatted with both Clementine Fourrier of HuggingFace’s OpenLLM Leaderboard and (the freshly valued at $1.7B) Anastasios Angelopoulos of LMArena on their approaches to LLM evals and trendspotting, but Artificial Analysis have staked out an enduring and important place in the toolkit of the modern AI Engineer by doing the best job of independently running the most comprehensive set of evals across the widest range of open and closed models, and charting their progress for broad industry analyst use.George Cameron and Micah-Hill Smith have spent two years building Artificial Analysis into the platform that answers the questions no one else will: Which model is actually best for your use case? What are the real speed-cost trade-offs? And how open is “open” really?We discuss:* The origin story: built as a side project in 2023 while Micah was building a legal AI assistant, launched publicly in January 2024, and went viral after Swyx’s retweet* Why they run evals themselves: labs prompt models differently, cherry-pick chain-of-thought examples (Google Gemini 1.0 Ultra used 32-shot prompts to beat GPT-4 on MMLU), and self-report inflated numbers* The mystery shopper policy: they register accounts not on their own domain and run intelligence + performance benchmarks incognito to prevent labs from serving different models on private endpoints* How they make money: enterprise benchmarking insights subscription (standardized reports on model deployment, serverless vs. managed vs. leasing chips) and private custom benchmarking for AI companies (no one pays to be on the public leaderboard)* The Intelligence Index (V3): synthesizes 10 eval datasets (MMLU, GPQA, agentic benchmarks, long-context reasoning) into a single score, with 95% confidence intervals via repeated runs* Omissions Index (hallucination rate): scores models from -100 to +100 (penalizing incorrect answers, rewarding \”I don’t know\”), and Claude models lead with the lowest hallucination rates despite not always being the smartest* GDP Val AA: their version of OpenAI’s GDP-bench (44 white-collar tasks with spreadsheets, PDFs, PowerPoints), run through their Stirrup agent harness (up to 100 turns, code execution, web search, file system), graded by Gemini 3 Pro as an LLM judge (tested extensively, no self-preference bias)* The Openness Index: scores models 0-18 on transparency of pre-training data, post-training data, methodology, training code, and licensing (AI2 OLMo 2 leads, followed by Nous Hermes and NVIDIA Nemotron)* The smiling curve of AI costs: GPT-4-level intelligence is 100-1000x cheaper than at launch (thanks to smaller models like Amazon Nova), but frontier reasoning models in agentic workflows cost more than ever (sparsity, long context, multi-turn agents)* Why sparsity might go way lower than 5%: GPT-4.5 is ~5% active, Gemini models might be ~3%, and Omissions Index accuracy correlates with total parameters (not active), suggesting massive sparse models are the future* Token efficiency vs. turn efficiency: GPT-5 costs more per token but solves Tau-bench in fewer turns (cheaper overall), and models are getting better at using more tokens only when needed (5.1 Codex has tighter token distributions)* V4 of the Intelligence Index coming soon: adding GDP Val AA, Critical Point, hallucination rate, and dropping some saturated benchmarks (human-eval-style coding is now trivial for small models)Links to Artificial Analysis* Website: https://artificialanalysis.ai* George Cameron on X: https://x.com/georgecameron* Micah-Hill Smith on X: https://x.com/micahhsmithFull Episode on YouTubeTimestamps* 00:00 Introduction: Full Circle Moment and Artificial Analysis Origins* 01:19 Business Model: Independence and Revenue Streams* 04:33 Origin Story: From Legal AI to Benchmarking Need* 16:22 AI Grant and Moving to San Francisco* 19:21 Intelligence Index Evolution: From V1 to V3* 11:47 Benchmarking Challenges: Variance, Contamination, and Methodology* 13:52 Mystery Shopper Policy and Maintaining Independence* 28:01 New Benchmarks: Omissions Index for Hallucination Detection* 33:36 Critical Point: Hard Physics Problems and Research-Level Reasoning* 23:01 GDP Val AA: Agentic Benchmark for Real Work Tasks* 50:19 Stirrup Agent Harness: Open Source Agentic Framework* 52:43 Openness Index: Measuring Model Transparency Beyond Licenses* 58:25 The Smiling Curve: Cost Falling While Spend Rising* 1:02:32 Hardware Efficiency: Blackwell Gains and Sparsity Limits* 1:06:23 Reasoning Models and Token Efficiency: The Spectrum Emerges* 1:11:00 Multimodal Benchmarking: Image, Video, and Speech Arenas* 1:15:05 Looking Ahead: Intelligence Index V4 and Future Directions* 1:16:50 Closing: The Insatiable Demand for IntelligenceTranscriptMicah [00:00:06]: This is kind of a full circle moment for us in a way, because the first time artificial analysis got mentioned on a podcast was you and Alessio on Latent Space. Amazing.swyx [00:00:17]: Which was January 2024. I don’t even remember doing that, but yeah, it was very influential to me. Yeah, I’m looking at AI News for Jan 17, or Jan 16, 2024. I said, this gem of a models and host comparison site was just launched. And then I put in a few screenshots, and I said, it’s an independent third party. It clearly outlines the quality versus throughput trade-off, and it breaks out by model and hosting provider. I did give you s**t for missing fireworks, and how do you have a model benchmarking thing without fireworks? But you had together, you had perplexity, and I think we just started chatting there. Welcome, George and Micah, to Latent Space. I’ve been following your progress. Congrats on... It’s been an amazing year. You guys have really come together to be the presumptive new gardener of AI, right? Which is something that...George [00:01:09]: Yeah, but you can’t pay us for better results.swyx [00:01:12]: Yes, exactly.George [00:01:13]: Very important.Micah [00:01:14]: Start off with a spicy take.swyx [00:01:18]: Okay, how do I pay you?Micah [00:01:20]: Let’s get right into that.swyx [00:01:21]: How do you make money?Micah [00:01:24]: Well, very happy to talk about that. So it’s been a big journey the last couple of years. Artificial analysis is going to be two years old in January 2026. Which is pretty soon now. We first run the website for free, obviously, and give away a ton of data to help developers and companies navigate AI and make decisions about models, providers, technologies across the AI stack for building stuff. We’re very committed to doing that and tend to keep doing that. We have, along the way, built a business that is working out pretty sustainably. We’ve got just over 20 people now and two main customer groups. So we want to be... We want to be who enterprise look to for data and insights on AI, so we want to help them with their decisions about models and technologies for building stuff. And then on the other side, we do private benchmarking for companies throughout the AI stack who build AI stuff. So no one pays to be on the website. We’ve been very clear about that from the very start because there’s no use doing what we do unless it’s independent AI benchmarking. Yeah. But turns out a bunch of our stuff can be pretty useful to companies building AI stuff.swyx [00:02:38]: And is it like, I am a Fortune 500, I need advisors on objective analysis, and I call you guys and you pull up a custom report for me, you come into my office and give me a workshop? What kind of engagement is that?George [00:02:53]: So we have a benchmarking and insight subscription, which looks like standardized reports that cover key topics or key challenges enterprises face when looking to understand AI and choose between all the technologies. And so, for instance, one of the report is a model deployment report, how to think about choosing between serverless inference, managed deployment solutions, or leasing chips. And running inference yourself is an example kind of decision that big enterprises face, and it’s hard to reason through, like this AI stuff is really new to everybody. And so we try and help with our reports and insight subscription. Companies navigate that. We also do custom private benchmarking. And so that’s very different from the public benchmarking that we publicize, and there’s no commercial model around that. For private benchmarking, we’ll at times create benchmarks, run benchmarks to specs that enterprises want. And we’ll also do that sometimes for AI companies who have built things, and we help them understand what they’ve built with private benchmarking. Yeah. So that’s a piece mainly that we’ve developed through trying to support everybody publicly with our public benchmarks. Yeah.swyx [00:04:09]: Let’s talk about TechStack behind that. But okay, I’m going to rewind all the way to when you guys started this project. You were all the way in Sydney? Yeah. Well, Sydney, Australia for me.Micah [00:04:19]: George was an SF, but he’s Australian, but he moved here already. Yeah.swyx [00:04:22]: And I remember I had the Zoom call with you. What was the impetus for starting artificial analysis in the first place? You know, you started with public benchmarks. And so let’s start there. We’ll go to the privat

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

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

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

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

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

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

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

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

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

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