Latent Space: The AI Engineer Podcast

Editor’s note: CuspAI raised a $100m Series A in September and is rumored to have reached a unicorn valuation. They have all-star advisors from Geoff Hinton to Yann Lecun and team of deep domain experts to tackle this next frontier in AI applications.
In this episode, Max Welling traces the thread connecting quantum gravity, equivariant neural networks, diffusion models, and climate-focused materials discovery (yes, there is one!!!).
We begin with a provocative framing: experiments as computation. Welling describes the idea of a “physics processing unit”—a world in which digital models and physical experiments work together, with nature itself acting as a kind of processor. It’s a grounded but ambitious vision of AI for science: not replacing chemists, but accelerating them.Along the way, we discuss:
* Why symmetry and equivariance matter in deep learning
* The tradeoff between scale and inductive bias
* The deep mathematical links between diffusion models and stochastic thermodynamics
* Why materials—not software—may be the real bottleneck for AI and the energy transition
* What it actually takes to build an AI-driven materials platform
Max reflects on moving from curiosity-driven theoretical physics (including work with Gerard ‘t Hooft) toward impact-driven research in climate and energy. The result is a conversation about convergence: physics and machine learning, digital models and laboratory experiments, long-term ambition and incremental progress.
Full Video Episode

Timestamps
* 00:00:00 – The Physics Processing Unit (PPU): Nature as the Ultimate Computer
* Max introduces the idea of a Physics Processing Unit — using real-world experiments as computation.
* 00:00:44 – From Quantum Gravity to AI for Materials
* Brandon frames Max’s career arc: VAE pioneer → equivariant GNNs → materials startup founder.
* 00:01:34 – Curiosity vs Impact: How His Motivation Evolved
* Max explains the shift from pure theoretical curiosity to climate-driven impact.
* 00:02:43 – Why CaspAI Exists: Technology as Climate Strategy
* Politics struggles; technology scales. Why materials innovation became the focus.
* 00:03:39 – The Thread: Physics → Symmetry → Machine Learning
* How gauge symmetry, group theory, and relativity informed equivariant neural networks.
* 00:06:52 – AI for Science Is Exploding (Not Emerging)
* The funding surge and why AI-for-Science feels like a new industrial era.
* 00:07:53 – Why Now? The Two Catalysts Behind AI for Science
* Protein folding, ML force fields, and the tipping point moment.
* 00:10:12 – How Engineers Can Enter AI for Science
* Practical pathways: curriculum, workshops, cross-disciplinary training.
* 00:11:28 – Why Materials Matter More Than Software
* The argument that everything—LLMs included—rests on materials innovation.
* 00:13:02 – Materials as a Search Engine
* The vision: automated exploration of chemical space like querying Google.
* 01:14:48 – Inside CuspAI: The Platform Architecture
* Generative models + multi-scale digital twin + experiment loop.
* 00:21:17 – Automating Chemistry: Human-in-the-Loop First
* Start manual → modular tools → agents → increasing autonomy.
* 00:25:04 – Moonshots vs Incremental Wins
* Balancing lighthouse materials with paid partnerships.
* 00:26:22 – Why Breakthroughs Will Still Require Humans
* Automation is vertical-specific and iterative.
* 00:29:01 – What Is Equivariance (In Plain English)?
* Symmetry in neural networks explained with the bottle example.
* 00:30:01 – Why Not Just Use Data Augmentation?
* The optimization trade-off between inductive bias and data scale.
* 00:31:55 – Generative AI Meets Stochastic Thermodynamics
* His upcoming book and the unification of diffusion models and physics.
* 00:33:44 – When the Book Drops (ICLR?)

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


This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe

This is a free preview of a paid episode. To hear more, visit www.latent.space

First 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 Mania
In 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 YouTube
Timestamps
00:00 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 Lessons

Transcript
[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 noise
Show 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/
Timestamps
00:00:00 – 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 Stack

Transcript
Latent.Space - Inside AI’s $10B+ Capital Flywheel — Martin Casado & Sarah Wang of a16z
[00:00:00] Welcome to Latent Space (Live from a16z) + Meet the Guests
[00:00:00] Alessio: Hey everyone. Welcome to the Latent Space podcast, live from a 16 z. Uh, this is Alessio founder Kernel Lance, and I’m joined by Twix, editor of Latent Space.
[00:00:08] swyx: Hey, hey, hey. Uh, and we’re so glad to be on with you guys. Also a top AI podcast, uh, Martin Cado and Sarah Wang. Welcome, very
[00:00:16] Martin Casado: happy to be here and welcome.
[00:00:17] swyx: Yes, uh, we love this office. We love what you’ve done with the place. Uh, the new logo is everywhere now. It’s, it’s still getting, takes a while to get used to, but it reminds me of like sort of a callback to a more ambitious age, which I think is kind of
[00:00:31] Martin Casado: definitely makes a statement.
[00:00:33] swyx: Yeah.
[00:00:34] Martin Casado: Not quite sure what that statement is, but it makes a statement.
[00:00:37] swyx: Uh, Martin, I go back with you to Netlify.
[00:00:40] Martin Casado: Yep.
[00:00:40] swyx: Uh, and, uh, you know, you create a software defined networking and all, all that stuff people can read up on your background. Yep. Sarah, I’m newer to you. Uh, you, you sort of started working together on AI infrastructure stuff.
[00:00:51] Sarah Wang: That’s right. Yeah. Seven, seven years ago now.
[00:00:53] Martin Casado: Best growth investor in the entire industry.
[00:00:55] swyx: Oh, say
[00:00:56] Martin Casado: more hands down there is, there is. [00:01:00] I mean, when it comes to AI companies, Sarah, I think has done the most kind of aggressive, um, investment thesis around AI models, right? So, worked for Nom Ja, Mira Ia, FEI Fey, and so just these frontier, kind of like large AI models.
[00:01:15] I think, you know, Sarah’s been the, the broadest investor. Is that fair?
[00:01:20] Venture vs. Growth in the Frontier Model Era
[00:01:20] Sarah Wang: No, I, well, I was gonna say, I think it’s been a really interesting tag, tag team actually just ‘cause the, a lot of these big C deals, not only are they raising a lot of money, um, it’s still a tech founder bet, which obviously is inherently early stage.
[00:01:33] But the resources,
[00:01:36] Martin Casado: so many, I
[00:01:36] Sarah Wang: was gonna say the resources one, they just grow really quickly. But then two, the resources that they need day one are kind of growth scale. So I, the hybrid tag team that we have is. Quite effective, I think,
[00:01:46] Martin Casado: what is growth these days? You know, you don’t wake up if it’s less than a billion or like, it’s, it’s actually, it’s actually very like, like no, it’s a very interesting time in investing because like, you know, take like the character around, right?
[00:01:59] These tend to [00:02:00] be like pre monetization, but the dollars are large enough that you need to have a larger fund and the analysis. You know, because you’ve got lots of users. ‘cause this stuff has such high demand requires, you know, more of a number sophistication. And so most of these deals, whether it’s US or other firms on these large model companies, are like this hybrid between venture growth.
[00:02:18] Sarah Wang: Yeah. Total. And I think, you know, stuff like BD for example, you wouldn’t usually need BD when you were seed stage trying to get market biz Devrel. Biz Devrel, exactly. Okay. But like now, sorry, I’m,
[00:02:27] swyx: I’m not familiar. What, what, what does biz Devrel mean for a venture fund? Because I know what biz Devrel means for a company.
[00:02:31] Sarah Wang: Yeah.
[00:02:32] Compute Deals, Strategics, and the ‘Circular Funding’ Question
[00:02:32] Sarah Wang: You know, so a, a good example is, I mean, we talk about buying compute, but there’s a huge negotiation involved there in terms of, okay, do you get equity for the compute? What, what sort of partner are you looking at? Is there a go-to market arm to that? Um, and these are just things on this scale, hundreds of millions, you know, maybe.
[00:02:50] Six months into the inception of a company, you just wouldn’t have to negotiate these deals before.
[00:02:54] Martin Casado: Yeah. These large rounds are very complex now. Like in the past, if you did a series A [00:03:00] or a series B, like whatever, you’re writing a 20 to a $60 million check and you call it a day. Now you normally have financial investors and strategic investors, and then the strategic portion always still goes with like these kind of large compute contracts, which can take months to do.
[00:03:13] And so it’s, it’s very different ties. I’ve been doing this for 10 years. It’s the, I’ve never seen anything like this.
[00:03:19] swyx: Yeah. Do you have worries about the circular funding from so disease strategics?
[00:03:24] Martin Casado: I mean, listen, as long as the demand is there, like the demand is there. Like the problem with the internet is the demand wasn’t there.
[00:03:29] swyx: Exactly. All right. This, this is like the, the whole pyramid scheme bubble thing, where like, as long as you mark to market on like the notional value of like, these deals, fine, but like once it starts to chip away, it really Well
[00:03:41] Martin Casado: no, like as, as, as, as long as there’s demand. I mean, you know, this, this is like a lot of these sound bites have already become kind of cliches, but they’re worth saying it.
[00:03:47] Right? Like during the internet days, like we were. Um, raising money to put fiber in the ground that wasn’t used. And that’s a problem, right? Because now you actually have a supply overhang.
[00:03:58] swyx: Mm-hmm.
[00:03:59] Martin Casado: And even in the, [00:04:00] the time of the, the internet, like the supply and, and bandwidth overhang, even as massive as it was in, as massive as the crash was only lasted about four years.
[00:04:09] But we don’t have a supply overhang. Like there’s no dark GPUs, right? I mean, and so, you know, circular or not, I mean, you know, if, if someone invests in a company that, um. You know, they’ll actually use the GPUs. And on the other side of it is the, is the ask for customer. So I I, I think it’s a different time.
[00:04:25] Sarah Wang: I think the other piece, maybe just to add onto this, and I’m gonna quote Martine in front of him, but this is probably also a unique time in that. For the first time, you can actually trace dollars to outcomes. Yeah, right. Provided that scaling laws are, are holding, um, and capabilities are actually moving forward.
[00:04:40] Because if you can put translate dollars into capabilities, uh, a capability improvement, there’s demand there to martine’s point. But if that somehow breaks, you know, obviously that’s an important assumption in this whole thing to make it work. But you know, instead of investing dollars into sales and marketing, you’re, you’re investing into r and d to get to the capability, um, you know, increase.
[00:04:59] And [00:05:00] that’s sort of been the demand driver because. Once there’s an unlock there, people are willing to pay for it.
[00:05:05] Alessio: Yeah.
[00:05:06] Blurring Lines: Models as Infra + Apps, and the New Fundraising Flywheel
[00:05:06] Alessio: Is there any difference in how you built the portfolio now that some of your growth companies are, like the infrastructure of the early stage companies, like, you know, OpenAI is now the same size as some of the cloud providers were early on.
[00:05:16] Like what does that look like? Like how much information can you feed off each other between the, the two?
[00:05:24] Martin Casado: There’s so many lines that are being crossed right now, or blurred. Right. So we already talked about venture and growth. Another one that’s being blurred is between infrastructure and apps, right? So like what is a model company?
[00:05:35] Mm-hmm. Like, it’s clearly infrastructure, right? Because it’s like, you know, it’s doing kind of core r and d. It’s a horizontal platform, but it’s also an app because it’s um, uh, touches the users directly. And then of course. You know, the, the, the growth of these is just so high. And so I actually think you’re just starting to see a, a, a new financing strategy emerge and, you know, we’ve had to adapt as a result of that.
[00:05:59] And [00:06:00] so there’s been a lot of changes. Um, you’re right that these companies become platform companies very quickly. You’ve got ecosystem build out. So none of this is necessarily new, but the timescales of which it’s happened is pretty phenomenal. And the way we’d normally cut lines before is blurred a little bit, but.
[00:06:16] But that, that, that said, I mean, a lot of it also just does feel like things that we’ve seen in the past, like cloud build out the internet build out as well.
[00:06:24] Sarah Wang: Yeah. Um, yeah, I think it’s interesting, uh, I don’t know if you guys would agree with this, but it feels like the emerging strategy is, and this builds off of your other question, um.
[00:06:33] You raise money for compute, you pour that or you, you pour the money into compute, you get some sort of breakthrough. You funnel the breakthrough into your vertically integrated application. That could be chat GBT, that could be cloud code, you know, whatever it is. You massively gain share and get users.
[00:06:49] Maybe you’re even subsidizing at that point. Um, depending on your strategy. You raise money at the peak momentum and then you repeat, rinse and repeat. Um, and so. And that wasn’t [00:07:00] true even two years ago, I think. Mm-hmm. And so it’s sort of to your, just tying it to fundraising strategy, right? There’s a, and hiring strategy.
[00:07:07] All of these are tied, I think the lines are blurring even more today where everyone is, and they, but of course these companies all have API businesses and so they’re these, these frenemy lines that are getting blurred in that a lot of, I mean, they have billions of dollars of API revenue, right? And so there are customers there.
[00:07:23] But they’re competing on the app layer.
[00:07:24] Martin Casado: Yeah. So this is a really, really important point. So I, I would say for sure, venture and growth, that line is blurry app and infrastructure. That line is blurry. Um, but I don’t think that that changes our practice so much. But like where the very open questions are like, does this layer in the same way.
[00:07:43] Compute traditionally has like during the cloud is like, you know, like whatever, somebody wins one layer, but then another whole set of companies wins another layer. But that might not, might not be the case here. It may be the case that you actually can’t verticalize on the token string. Like you can’t build an app like it, it necessarily goes down just because there are no [00:08:00] abstractions.
[00:08:00] So those are kinda the bigger existential questions we ask. Another thing that is very different this time than in the history of computer sciences is. In the past, if you raised money, then you basically had to wait for engineering to catch up. Which famously doesn’t scale like the mythical mammoth. It take a very long time.
[00:08:18] But like that’s not the case here. Like a model company can raise money and drop a model in a, in a year, and it’s better, right? And, and it does it with a team of 20 people or 10 people. So this type of like money entering a company and then producing something that has demand and growth right away and using that to raise more money is a very different capital flywheel than we’ve ever seen before.
[00:08:39] And I think everybody’s trying to understand what the consequences are. So I think it’s less about like. Big companies and growth and this, and more about these more systemic questions that we actually don’t have answers to.
[00:08:49] Alessio: Yeah, like at Kernel Labs, one of our ideas is like if you had unlimited money to spend productively to turn tokens into products, like the whole early stage [00:09:00] market is very different because today you’re investing X amount of capital to win a deal because of price structure and whatnot, and you’re kind of pot committing.
[00:09:07] Yeah. To a certain strategy for a certain amount of time. Yeah. But if you could like iteratively spin out companies and products and just throw, I, I wanna spend a million dollar of inference today and get a product out tomorrow.
[00:09:18] swyx: Yeah.
[00:09:19] Alessio: Like, we should get to the point where like the friction of like token to product is so low that you can do this and then you can change the Right, the early stage venture model to be much more iterative.
[00:09:30] And then every round is like either 100 k of inference or like a hundred million from a 16 Z. There’s no, there’s no like $8 million C round anymore. Right.
[00:09:38] When Frontier Labs Outspend the Entire App Ecosystem
[00:09:38] Martin Casado: But, but, but, but there’s a, there’s a, the, an industry structural question that we don’t know the answer to, which involves the frontier models, which is, let’s take.
[00:09:48] Anthropic it. Let’s say Anthropic has a state-of-the-art model that has some large percentage of market share. And let’s say that, uh, uh, uh, you know, uh, a company’s building smaller models [00:10:00] that, you know, use the bigger model in the background, open 4.5, but they add value on top of that. Now, if Anthropic can raise three times more.
[00:10:10] Every subsequent round, they probably can raise more money than the entire app ecosystem that’s built on top of it. And if that’s the case, they can expand beyond everything built on top of it. It’s like imagine like a star that’s just kind of expanding, so there could be a systemic. There could be a, a systemic situation where the soda models can raise so much money that they can out pay anybody that bills on top of ‘em, which would be something I don’t think we’ve ever seen before just because we were so bottlenecked in engineering, and this is a very open question.
[00:10:41] swyx: Yeah. It’s, it is almost like bitter lesson applied to the startup industry.
[00:10:45] Martin Casado: Yeah, a hundred percent. It literally becomes an issue of like raise capital, turn that directly into growth. Use that to raise three times more. Exactly. And if you can keep doing that, you literally can outspend any company that’s built the, not any company.
[00:10:57] You can outspend the aggregate of companies on top of [00:11:00] you and therefore you’ll necessarily take their share, which is crazy.
[00:11:02] swyx: Would you say that kind of happens in character? Is that the, the sort of postmortem on. What happened?
[00:11:10] Sarah Wang: Um,
[00:11:10] Martin Casado: no.
[00:11:12] Sarah Wang: Yeah, because I think so,
[00:11:13] swyx: I mean the actual postmortem is, he wanted to go back to Google.
[00:11:15] Exactly. But like
[00:11:18] Martin Casado: that’s another difference that
[00:11:19] Sarah Wang: you said
[00:11:21] Martin Casado: it. We should talk, we should actually talk about that.
[00:11:22] swyx: Yeah,
[00:11:22] Sarah Wang: that’s
[00:11:23] swyx: Go for it. Take it. Take,
[00:11:23] Sarah Wang: yeah.
[00:11:24] Character.AI, Founder Goals (AGI vs Product), and GPU Allocation Tradeoffs
[00:11:24] Sarah Wang: I was gonna say, I think, um. The, the, the character thing raises actually a different issue, which actually the Frontier Labs will face as well. So we’ll see how they handle it.
[00:11:34] But, um, so we invest in character in January, 2023, which feels like eons ago, I mean, three years ago. Feels like lifetimes ago. But, um, and then they, uh, did the IP licensing deal with Google in August, 2020. Uh, four. And so, um, you know, at the time, no, you know, he’s talked publicly about this, right? He wanted to Google wouldn’t let him put out products in the world.
[00:11:56] That’s obviously changed drastically. But, um, he went to go do [00:12:00] that. Um, but he had a product attached. The goal was, I mean, it’s Nome Shair, he wanted to get to a GI. That was always his personal goal. But, you know, I think through collecting data, right, and this sort of very human use case, that the character product.
[00:12:13] Originally was and still is, um, was one of the vehicles to do that. Um, I think the real reason that, you know. I if you think about the, the stress that any company feels before, um, you ultimately going one way or the other is sort of this a GI versus product. Um, and I think a lot of the big, I think, you know, opening eyes, feeling that, um, anthropic if they haven’t started, you know, felt it, certainly given the success of their products, they may start to feel that soon.
[00:12:39] And the real. I think there’s real trade-offs, right? It’s like how many, when you think about GPUs, that’s a limited resource. Where do you allocate the GPUs? Is it toward the product? Is it toward new re research? Right? Is it, or long-term research, is it toward, um, n you know, near to midterm research? And so, um, in a case where you’re resource constrained, um, [00:13:00] of course there’s this fundraising game you can play, right?
[00:13:01] But the fund, the market was very different back in 2023 too. Um. I think the best researchers in the world have this dilemma of, okay, I wanna go all in on a GI, but it’s the product usage revenue flywheel that keeps the revenue in the house to power all the GPUs to get to a GI. And so it does make, um, you know, I think it sets up an interesting dilemma for any startup that has trouble raising up until that level, right?
[00:13:27] And certainly if you don’t have that progress, you can’t continue this fly, you know, fundraising flywheel.
[00:13:32] Martin Casado: I would say that because, ‘cause we’re keeping track of all of the things that are different, right? Like, you know, venture growth and uh, app infra and one of the ones is definitely the personalities of the founders.
[00:13:45] It’s just very different this time I’ve been. Been doing this for a decade and I’ve been doing startups for 20 years. And so, um, I mean a lot of people start this to do a GI and we’ve never had like a unified North star that I recall in the same [00:14:00] way. Like people built companies to start companies in the past.
[00:14:02] Like that was what it was. Like I would create an internet company, I would create infrastructure company, like it’s kind of more engineering builders and this is kind of a different. You know, mentality. And some companies have harnessed that incredibly well because their direction is so obviously on the path to what somebody would consider a GI, but others have not.
[00:14:20] And so like there is always this tension with personnel. And so I think we’re seeing more kind of founder movement.
[00:14:27] Sarah Wang: Yeah.
[00:14:27] Martin Casado: You know, as a fraction of founders than we’ve ever seen. I mean, maybe since like, I don’t know the time of like Shockly and the trade DUR aid or something like that. Way back in the beginning of the industry, I, it’s a very, very.
[00:14:38] Unusual time of personnel.
[00:14:39] Sarah Wang: Totally.
[00:14:40] Talent Wars, Mega-Comp, and the Rise of Acquihire M&A
[00:14:40] Sarah Wang: And it, I think it’s exacerbated by the fact that talent wars, I mean, every industry has talent wars, but not at this magnitude, right? No. Yeah. Very rarely can you see someone get poached for $5 billion. That’s hard to compete with. And then secondly, if you’re a founder in ai, you could fart and it would be on the front page of, you know, the information these days.
[00:14:59] And so there’s [00:15:00] sort of this fishbowl effect that I think adds to the deep anxiety that, that these AI founders are feeling.
[00:15:06] Martin Casado: Hmm.
[00:15:06] swyx: Uh, yes. I mean, just on, uh, briefly comment on the founder, uh, the sort of. Talent wars thing. I feel like 2025 was just like a blip. Like I, I don’t know if we’ll see that again.
[00:15:17] ‘cause meta built the team. Like, I don’t know if, I think, I think they’re kind of done and like, who’s gonna pay more than meta? I, I don’t know.
[00:15:23] Martin Casado: I, I agree. So it feels so, it feel, it feels this way to me too. It’s like, it is like, basically Zuckerberg kind of came out swinging and then now he’s kind of back to building.
[00:15:30] Yeah,
[00:15:31] swyx: yeah. You know, you gotta like pay up to like assemble team to rush the job, whatever. But then now, now you like you, you made your choices and now they got a ship.
[00:15:38] Martin Casado: I mean, the, the o other side of that is like, you know, like we’re, we’re actually in the job hiring market. We’ve got 600 people here. I hire all the time.
[00:15:44] I’ve got three open recs if anybody’s interested, that’s listening to this for investor. Yeah, on, on the team, like on the investing side of the team, like, and, um, a lot of the people we talk to have acting, you know, active, um, offers for 10 million a year or something like that. And like, you know, and we pay really, [00:16:00] really well.
[00:16:00] And just to see what’s out on the market is really, is really remarkable. And so I would just say it’s actually, so you’re right, like the really flashy one, like I will get someone for, you know, a billion dollars, but like the inflated, um, uh, trickles down. Yeah, it is still very active today. I mean,
[00:16:18] Sarah Wang: yeah, you could be an L five and get an offer in the tens of millions.
[00:16:22] Okay. Yeah. Easily. Yeah. It’s so I think you’re right that it felt like a blip. I hope you’re right. Um, but I think it’s been, the steady state is now, I think got pulled up. Yeah. Yeah. I’ll pull up for
[00:16:31] Martin Casado: sure. Yeah.
[00:16:32] Alessio: Yeah. And I think that’s breaking the early stage founder math too. I think before a lot of people would be like, well, maybe I should just go be a founder instead of like getting paid.
[00:16:39] Yeah. 800 KA million at Google. But if I’m getting paid. Five, 6 million. That’s different but
[00:16:45] Martin Casado: on. But on the other hand, there’s more strategic money than we’ve ever seen historically, right? Mm-hmm. And so, yep. The economics, the, the, the, the calculus on the economics is very different in a number of ways. And, uh, it’s crazy.
[00:16:58] It’s cra it’s causing like a, [00:17:00] a, a, a ton of change in confusion in the market. Some very positive, sub negative, like, so for example, the other side of the, um. The co-founder, like, um, acquisition, you know, mark Zuckerberg poaching someone for a lot of money is like, we were actually seeing historic amount of m and a for basically acquihires, right?
[00:17:20] That you like, you know, really good outcomes from a venture perspective that are effective acquihires, right? So I would say it’s probably net positive from the investment standpoint, even though it seems from the headlines to be very disruptive in a negative way.
[00:17:33] Alessio: Yeah.
[00:17:33] What’s Underfunded: Boring Software, Robotics Skepticism, and Custom Silicon Economics
[00:17:33] Alessio: Um, let’s talk maybe about what’s not being invested in, like maybe some interesting ideas that you would see more people build or it, it seems in a way, you know, as ycs getting more popular, it’s like access getting more popular.
[00:17:47] There’s a startup school path that a lot of founders take and they know what’s hot in the VC circles and they know what gets funded. Uh, and there’s maybe not as much risk appetite for. Things outside of that. Um, I’m curious if you feel [00:18:00] like that’s true and what are maybe, uh, some of the areas, uh, that you think are under discussed?
[00:18:06] Martin Casado: I mean, I actually think that we’ve taken our eye off the ball in a lot of like, just traditional, you know, software companies. Um, so like, I mean. You know, I think right now there’s almost a barbell, like you’re like the hot thing on X, you’re deep tech.
[00:18:21] swyx: Mm-hmm.
[00:18:22] Martin Casado: Right. But I, you know, I feel like there’s just kind of a long, you know, list of like good.
[00:18:28] Good companies that will be around for a long time in very large markets. Say you’re building a database, you know, say you’re building, um, you know, kind of monitoring or logging or tooling or whatever. There’s some good companies out there right now, but like, they have a really hard time getting, um, the attention of investors.
[00:18:43] And it’s almost become a meme, right? Which is like, if you’re not basically growing from zero to a hundred in a year, you’re not interesting, which is just, is the silliest thing to say. I mean, think of yourself as like an introvert person, like, like your personal money, right? Mm-hmm. So. Your personal money, will you put it in the stock market at 7% or you put it in this company growing five x in a very large [00:19:00] market?
[00:19:00] Of course you can put it in the company five x. So it’s just like we say these stupid things, like if you’re not going from zero to a hundred, but like those, like who knows what the margins of those are mean. Clearly these are good investments. True for anybody, right? True. Like our LPs want whatever.
[00:19:12] Three x net over, you know, the life cycle of a fund, right? So a, a company in a big market growing five X is a great investment. We’d, everybody would be happy with these returns, but we’ve got this kind of mania on these, these strong growths. And so I would say that that’s probably the most underinvested sector.
[00:19:28] Right now.
[00:19:29] swyx: Boring software, boring enterprise software.
[00:19:31] Martin Casado: Traditional. Really good company.
[00:19:33] swyx: No, no AI here.
[00:19:34] Martin Casado: No. Like boring. Well, well, the AI of course is pulling them into use cases. Yeah, but that’s not what they’re, they’re not on the token path, right? Yeah. Let’s just say that like they’re software, but they’re not on the token path.
[00:19:41] Like these are like they’re great investments from any definition except for like random VC on Twitter saying VC on x, saying like, it’s not growing fast enough. What do you
[00:19:52] Sarah Wang: think? Yeah, maybe I’ll answer a slightly different. Question, but adjacent to what you asked, um, which is maybe an area that we’re not, uh, investing [00:20:00] right now that I think is a question and we’re spending a lot of time in regardless of whether we pull the trigger or not.
[00:20:05] Um, and it would probably be on the hardware side, actually. Robotics, right? And the robotics side. Robotics. Right. Which is, it’s, I don’t wanna say that it’s not getting funding ‘cause it’s clearly, uh, it’s, it’s sort of non-consensus to almost not invest in robotics at this point. But, um, we spent a lot of time in that space and I think for us, we just haven’t seen the chat GPT moment.
[00:20:22] Happen on the hardware side. Um, and the funding going into it feels like it’s already. Taking that for granted.
[00:20:30] Martin Casado: Yeah. Yeah. But we also went through the drone, you know, um, there’s a zip line right, right out there. What’s that? Oh yeah, there’s a zip line. Yeah. What the drone, what the av And like one of the takeaways is when it comes to hardware, um, most companies will end up verticalizing.
[00:20:46] Like if you’re. If you’re investing in a robot company for an A for agriculture, you’re investing in an ag company. ‘cause that’s the competition and that’s surprising. And that’s supply chain. And if you’re doing it for mining, that’s mining. And so the ad team does a lot of that type of stuff ‘cause they actually set up to [00:21:00] diligence that type of work.
[00:21:01] But for like horizontal technology investing, there’s very little when it comes to robots just because it’s so fit for, for purpose. And so we kinda like to look at software. Solutions or horizontal solutions like applied intuition. Clearly from the AV wave deep map, clearly from the AV wave, I would say scale AI was actually a horizontal one for That’s fair, you know, for robotics early on.
[00:21:23] And so that sort of thing we’re very, very interested. But the actual like robot interacting with the world is probably better for different team. Agree.
[00:21:30] Alessio: Yeah, I’m curious who these teams are supposed to be that invest in them. I feel like everybody’s like, yeah, robotics, it’s important and like people should invest in it.
[00:21:38] But then when you look at like the numbers, like the capital requirements early on versus like the moment of, okay, this is actually gonna work. Let’s keep investing. That seems really hard to predict in a way that is not,
[00:21:49] Martin Casado: I think co, CO two, kla, gc, I mean these are all invested in in Harvard companies. He just, you know, and [00:22:00] listen, I mean, it could work this time for sure.
[00:22:01] Right? I mean if Elon’s doing it, he’s like, right. Just, just the fact that Elon’s doing it means that there’s gonna be a lot of capital and a lot of attempts for a long period of time. So that alone maybe suggests that we should just be investing in robotics just ‘cause you have this North star who’s Elon with a humanoid and that’s gonna like basically willing into being an industry.
[00:22:17] Um, but we’ve just historically found like. We’re a huge believer that this is gonna happen. We just don’t feel like we’re in a good position to diligence these things. ‘cause again, robotics companies tend to be vertical. You really have to understand the market they’re being sold into. Like that’s like that competitive equilibrium with a human being is what’s important.
[00:22:34] It’s not like the core tech and like we’re kind of more horizontal core tech type investors. And this is Sarah and I. Yeah, the ad team is different. They can actually do these types of things.
[00:22:42] swyx: Uh, just to clarify, AD stands for
[00:22:44] Martin Casado: American Dynamism.
[00:22:45] swyx: Alright. Okay. Yeah, yeah, yeah. Uh, I actually, I do have a related question that, first of all, I wanna acknowledge also just on the, on the chip side.
[00:22:51] Yeah. I, I recall a podcast that where you were on, i, I, I think it was the a CC podcast, uh, about two or three years ago where you, where you suddenly said [00:23:00] something, which really stuck in my head about how at some point, at some point kind of scale it makes sense to. Build a custom aic Yes. For per run.
[00:23:07] Martin Casado: Yes.
[00:23:07] It’s crazy. Yeah.
[00:23:09] swyx: We’re here and I think you, you estimated 500 billion, uh, something.
[00:23:12] Martin Casado: No, no, no. A billion, a billion dollar training run of $1 billion training run. It makes sense to actually do a custom meic if you can do it in time. The question now is timelines. Yeah, but not money because just, just, just rough math.
[00:23:22] If it’s a billion dollar training. Then the inference for that model has to be over a billion, otherwise it won’t be solvent. So let’s assume it’s, if you could save 20%, which you could save much more than that with an ASIC 20%, that’s $200 million. You can tape out a chip for $200 million. Right? So now you can literally like justify economically, not timeline wise.
[00:23:41] That’s a different issue. An ASIC per model, which
[00:23:44] swyx: is because that, that’s how much we leave on the table every single time. We, we, we do like generic Nvidia.
[00:23:48] Martin Casado: Exactly. Exactly. No, it, it is actually much more than that. You could probably get, you know, a factor of two, which would be 500 million.
[00:23:54] swyx: Typical MFU would be like 50.
[00:23:55] Yeah, yeah. And that’s good.
[00:23:57] Martin Casado: Exactly. Yeah. Hundred
[00:23:57] swyx: percent. Um, so, so, yeah, and I mean, and I [00:24:00] just wanna acknowledge like, here we are in, in, in 2025 and opening eyes confirming like Broadcom and all the other like custom silicon deals, which is incredible. I, I think that, uh, you know, speaking about ad there’s, there’s a really like interesting tie in that obviously you guys are hit on, which is like these sort, this sort of like America first movement or like sort of re industrialized here.
[00:24:17] Yeah. Uh, move TSMC here, if that’s possible. Um, how much overlap is there from ad
[00:24:23] Martin Casado: Yeah.
[00:24:23] swyx: To, I guess, growth and, uh, investing in particularly like, you know, US AI companies that are strongly bounded by their compute.
[00:24:32] Martin Casado: Yeah. Yeah. So I mean, I, I would view, I would view AD as more as a market segmentation than like a mission, right?
[00:24:37] So the market segmentation is, it has kind of regulatory compliance issues or government, you know, sale or it deals with like hardware. I mean, they’re just set up to, to, to, to, to. To diligence those types of companies. So it’s a more of a market segmentation thing. I would say the entire firm. You know, which has been since it is been intercepted, you know, has geographical biases, right?
[00:24:58] I mean, for the longest time we’re like, you [00:25:00] know, bay Area is gonna be like, great, where the majority of the dollars go. Yeah. And, and listen, there, there’s actually a lot of compounding effects for having a geographic bias. Right. You know, everybody’s in the same place. You’ve got an ecosystem, you’re there, you’ve got presence, you’ve got a network.
[00:25:12] Um, and, uh, I mean, I would say the Bay area’s very much back. You know, like I, I remember during pre COVID, like it was like almost Crypto had kind of. Pulled startups away. Miami from the Bay Area. Miami, yeah. Yeah. New York was, you know, because it’s so close to finance, came up like Los Angeles had a moment ‘cause it was so close to consumer, but now it’s kind of come back here.
[00:25:29] And so I would say, you know, we tend to be very Bay area focused historically, even though of course we’ve asked all over the world. And then I would say like, if you take the ring out, you know, one more, it’s gonna be the US of course, because we know it very well. And then one more is gonna be getting us and its allies and Yeah.
[00:25:44] And it goes from there.
[00:25:45] Sarah Wang: Yeah,
[00:25:45] Martin Casado: sorry.
[00:25:46] Sarah Wang: No, no. I agree. I think from a, but I think from the intern that that’s sort of like where the companies are headquartered. Maybe your questions on supply chain and customer base. Uh, I, I would say our customers are, are, our companies are fairly international from that perspective.
[00:25:59] Like they’re selling [00:26:00] globally, right? They have global supply chains in some cases.
[00:26:03] Martin Casado: I would say also the stickiness is very different.
[00:26:05] Sarah Wang: Yeah.
[00:26:05] Martin Casado: Historically between venture and growth, like there’s so much company building in venture, so much so like hiring the next PM. Introducing the customer, like all of that stuff.
[00:26:15] Like of course we’re just gonna be stronger where we have our network and we’ve been doing business for 20 years. I’ve been in the Bay Area for 25 years, so clearly I’m just more effective here than I would be somewhere else. Um, where I think, I think for some of the later stage rounds, the companies don’t need that much help.
[00:26:30] They’re already kind of pretty mature historically, so like they can kind of be everywhere. So there’s kind of less of that stickiness. This is different in the AI time. I mean, Sarah is now the, uh, chief of staff of like half the AI companies in, uh, in the Bay Area right now. She’s like, ops Ninja Biz, Devrel, BizOps.
[00:26:48] swyx: Are, are you, are you finding much AI automation in your work? Like what, what is your stack.
[00:26:53] Sarah Wang: Oh my, in my personal stack.
[00:26:54] swyx: I mean, because like, uh, by the way, it’s the, the, the reason for this is it is triggering, uh, yeah. We, like, I’m hiring [00:27:00] ops, ops people. Um, a lot of ponders I know are also hiring ops people and I’m just, you know, it’s opportunity Since you’re, you’re also like basically helping out with ops with a lot of companies.
[00:27:09] What are people doing these days? Because it’s still very manual as far as I can tell.
[00:27:13] Sarah Wang: Hmm. Yeah. I think the things that we help with are pretty network based, um, in that. It’s sort of like, Hey, how do do I shortcut this process? Well, let’s connect you to the right person. So there’s not quite an AI workflow for that.
[00:27:26] I will say as a growth investor, Claude Cowork is pretty interesting. Yeah. Like for the first time, you can actually get one shot data analysis. Right. Which, you know, if you’re gonna do a customer database, analyze a cohort retention, right? That’s just stuff that you had to do by hand before. And our team, the other, it was like midnight and the three of us were playing with Claude Cowork.
[00:27:47] We gave it a raw file. Boom. Perfectly accurate. We checked the numbers. It was amazing. That was my like, aha moment. That sounds so boring. But you know, that’s, that’s the kind of thing that a growth investor is like, [00:28:00] you know, slaving away on late at night. Um, done in a few seconds.
[00:28:03] swyx: Yeah. You gotta wonder what the whole, like, philanthropic labs, which is like their new sort of products studio.
[00:28:10] Yeah. What would that be worth as an independent, uh, startup? You know, like a
[00:28:14] Martin Casado: lot.
[00:28:14] Sarah Wang: Yeah, true.
[00:28:16] swyx: Yeah. You
[00:28:16] Martin Casado: gotta hand it to them. They’ve been executing incredibly well.
[00:28:19] swyx: Yeah. I, I mean, to me, like, you know, philanthropic, like building on cloud code, I think, uh, it makes sense to me the, the real. Um, pedal to the metal, whatever the, the, the phrase is, is when they start coming after consumer with, uh, against OpenAI and like that is like red alert at Open ai.
[00:28:35] Oh, I
[00:28:35] Martin Casado: think they’ve been pretty clear. They’re enterprise focused.
[00:28:37] swyx: They have been, but like they’ve been free. Here’s
[00:28:40] Martin Casado: care publicly,
[00:28:40] swyx: it’s enterprise focused. It’s coding. Right. Yeah.
[00:28:43] AI Labs vs Startups: Disruption, Undercutting & the Innovator’s Dilemma
[00:28:43] swyx: And then, and, but here’s cloud, cloud, cowork, and, and here’s like, well, we, uh, they, apparently they’re running Instagram ads for Claudia.
[00:28:50] I, on, you know, for, for people on, I get them all the time. Right. And so, like,
[00:28:54] Martin Casado: uh,
[00:28:54] swyx: it, it’s kind of like this, the disruption thing of, uh, you know. Mo Open has been doing, [00:29:00] consumer been doing the, just pursuing general intelligence in every mo modality, and here’s a topic that only focus on this thing, but now they’re sort of undercutting and doing the whole innovator’s dilemma thing on like everything else.
[00:29:11] Martin Casado: It’s very
[00:29:11] swyx: interesting.
[00:29:12] Martin Casado: Yeah, I mean there’s, there’s a very open que so for me there’s like, do you know that meme where there’s like the guy in the path and there’s like a path this way? There’s a path this way. Like one which way Western man. Yeah. Yeah.
[00:29:23] Two Futures for AI: Infinite Market vs AGI Oligopoly
[00:29:23] Martin Casado: And for me, like, like all the entire industry kind of like hinges on like two potential futures.
[00:29:29] So in, in one potential future, um, the market is infinitely large. There’s perverse economies of scale. ‘cause as soon as you put a model out there, like it kind of sublimates and all the other models catch up and like, it’s just like software’s being rewritten and fractured all over the place and there’s tons of upside and it just grows.
[00:29:48] And then there’s another path which is like, well. Maybe these models actually generalize really well, and all you have to do is train them with three times more money. That’s all you have to [00:30:00] do, and it’ll just consume everything beyond it. And if that’s the case, like you end up with basically an oligopoly for everything, like, you know mm-hmm.
[00:30:06] Because they’re perfectly general and like, so this would be like the, the a GI path would be like, these are perfectly general. They can do everything. And this one is like, this is actually normal software. The universe is complicated. You’ve got, and nobody knows the answer.
[00:30:18] The Economics Reality Check: Gross Margins, Training Costs & Borrowing Against the Future
[00:30:18] Martin Casado: My belief is if you actually look at the numbers of these companies, so generally if you look at the numbers of these companies, if you look at like the amount they’re making and how much they, they spent training the last model, they’re gross margin positive.
[00:30:30] You’re like, oh, that’s really working. But if you look at like. The current training that they’re doing for the next model, their gross margin negative. So part of me thinks that a lot of ‘em are kind of borrowing against the future and that’s gonna have to slow down. It’s gonna catch up to them at some point in time, but we don’t really know.
[00:30:47] Sarah Wang: Yeah.
[00:30:47] Martin Casado: Does that make sense? Like, I mean, it could be, it could be the case that the only reason this is working is ‘cause they can raise that next round and they can train that next model. ‘cause these models have such a short. Life. And so at some point in time, like, you know, they won’t be able to [00:31:00] raise that next round for the next model and then things will kind of converge and fragment again.
[00:31:03] But right now it’s not.
[00:31:04] Sarah Wang: Totally. I think the other, by the way, just, um, a meta point. I think the other lesson from the last three years is, and we talk about this all the time ‘cause we’re on this. Twitter X bubble. Um, cool. But, you know, if you go back to, let’s say March, 2024, that period, it felt like a, I think an open source model with an, like a, you know, benchmark leading capability was sort of launching on a daily basis at that point.
[00:31:27] And, um, and so that, you know, that’s one period. Suddenly it’s sort of like open source takes over the world. There’s gonna be a plethora. It’s not an oligopoly, you know, if you fast, you know, if you, if you rewind time even before that GPT-4 was number one for. Nine months, 10 months. It’s a long time. Right.
[00:31:44] Um, and of course now we’re in this era where it feels like an oligopoly, um, maybe some very steady state shifts and, and you know, it could look like this in the future too, but it just, it’s so hard to call. And I think the thing that keeps, you know, us up at [00:32:00] night in, in a good way and bad way, is that the capability progress is actually not slowing down.
[00:32:06] And so until that happens, right, like you don’t know what’s gonna look like.
[00:32:09] Martin Casado: But I, I would, I would say for sure it’s not converged, like for sure, like the systemic capital flows have not converged, meaning right now it’s still borrowing against the future to subsidize growth currently, which you can do that for a period of time.
[00:32:23] But, but you know, at the end, at some point the market will rationalize that and just nobody knows what that will look like.
[00:32:29] Alessio: Yeah.
[00:32:29] Martin Casado: Or, or like the drop in price of compute will, will, will save them. Who knows?
[00:32:34] Alessio: Yeah. Yeah. I think the models need to ask them to, to specific tasks. You know? It’s like, okay, now Opus 4.5 might be a GI at some specific task, and now you can like depreciate the model over a longer time.
[00:32:45] I think now, now, right now there’s like no old model.
[00:32:47] Martin Casado: No, but let, but lemme just change that mental, that’s, that used to be my mental model. Lemme just change it a little bit.
[00:32:53] Capital as a Weapon vs Task Saturation: Where Real Enterprise Value Gets Built
[00:32:53] Martin Casado: If you can raise three times, if you can raise more than the aggregate of anybody that uses your models, that doesn’t even matter.
[00:32:59] It doesn’t [00:33:00] even matter. See what I’m saying? Like, yeah. Yeah. So, so I have an API Business. My API business is 60% margin, or 70% margin, or 80% margin is a high margin business. So I know what everybody is using. If I can raise more money than the aggregate of everybody that’s using it, I will consume them whether I’m a GI or not.
[00:33:14] And I will know if they’re using it ‘cause they’re using it. And like, unlike in the past where engineering stops me from doing that.
[00:33:21] Alessio: Mm-hmm.
[00:33:21] Martin Casado: It is very straightforward. You just train. So I also thought it was kind of like, you must ask the code a GI, general, general, general. But I think there’s also just a possibility that the, that the capital markets will just give them the, the, the ammunition to just go after everybody on top of ‘em.
[00:33:36] Sarah Wang: I, I do wonder though, to your point, um, if there’s a certain task that. Getting marginally better isn’t actually that much better. Like we’ve asked them to it, to, you know, we can call it a GI or whatever, you know, actually, Ali Goi talks about this, like we’re already at a GI for a lot of functions in the enterprise.
[00:33:50] Um. That’s probably those for those tasks, you probably could build very specific companies that focus on just getting as much value out of that task that isn’t [00:34:00] coming from the model itself. There’s probably a rich enterprise business to be built there. I mean, could be wrong on that, but there’s a lot of interesting examples.
[00:34:08] So, right, if you’re looking the legal profession or, or whatnot, and maybe that’s not a great one ‘cause the models are getting better on that front too, but just something where it’s a bit saturated, then the value comes from. Services. It comes from implementation, right? It comes from all these things that actually make it useful to the end customer.
[00:34:24] Martin Casado: Sorry, what am I, one more thing I think is, is underused in all of this is like, to what extent every task is a GI complete.
[00:34:31] Sarah Wang: Mm-hmm.
[00:34:32] Martin Casado: Yeah. I code every day. It’s so fun.
[00:34:35] Sarah Wang: That’s a core question. Yeah.
[00:34:36] Martin Casado: And like. When I’m talking to these models, it’s not just code. I mean, it’s everything, right? Like I, you know, like it’s,
[00:34:43] swyx: it’s healthcare.
[00:34:44] It’s,
[00:34:44] Martin Casado: I mean, it’s
[00:34:44] swyx: Mele,
[00:34:45] Martin Casado: but it’s every, it is exactly that. Like, yeah, that’s
[00:34:47] Sarah Wang: great support. Yeah.
[00:34:48] Martin Casado: It’s everything. Like I’m asking these models to, yeah, to understand compliance. I’m asking these models to go search the web. I’m asking these models to talk about things I know in the history, like it’s having a full conversation with me while I, I engineer, and so it could be [00:35:00] the case that like, mm-hmm.
[00:35:01] The most a, you know, a GI complete, like I’m not an a GI guy. Like I think that’s, you know, but like the most a GI complete model will is win independent of the task. And we don’t know the answer to that one either.
[00:35:11] swyx: Yeah.
[00:35:12] Martin Casado: But it seems to me that like, listen, codex in my experience is for sure better than Opus 4.5 for coding.
[00:35:18] Like it finds the hardest bugs that I work in with. Like, it is, you know. The smartest developers. I don’t work on it. It’s great. Um, but I think Opus 4.5 is actually very, it’s got a great bedside manner and it really, and it, it really matters if you’re building something very complex because like, it really, you know, like you’re, you’re, you’re a partner and a brainstorming partner for somebody.
[00:35:38] And I think we don’t discuss enough how every task kind of has that quality.
[00:35:42] swyx: Mm-hmm.
[00:35:43] Martin Casado: And what does that mean to like capital investment and like frontier models and Submodels? Yeah.
[00:35:47] Why “Coding Models” Keep Collapsing into Generalists (Reasoning vs Taste)
[00:35:47] Martin Casado: Like what happened to all the special coding models? Like, none of ‘em worked right. So
[00:35:51] Alessio: some of them, they didn’t even get released.
[00:35:53] Magical
[00:35:54] Martin Casado: Devrel. There’s a whole, there’s a whole host. We saw a bunch of them and like there’s this whole theory that like, there could be, and [00:36:00] I think one of the conclusions is, is like there’s no such thing as a coding model,
[00:36:04] Alessio: you know?
[00:36:04] Martin Casado: Like, that’s not a thing. Like you’re talking to another human being and it’s, it’s good at coding, but like it’s gotta be good at everything.
[00:36:10] swyx: Uh, minor disagree only because I, I’m pretty like, have pretty high confidence that basically open eye will always release a GPT five and a GT five codex. Like that’s the code’s. Yeah. The way I call it is one for raisin, one for Tiz. Um, and, and then like someone internal open, it was like, yeah, that’s a good way to frame it.
[00:36:32] Martin Casado: That’s so funny.
[00:36:33] swyx: Uh, but maybe it, maybe it collapses down to reason and that’s it. It’s not like a hundred dimensions doesn’t life. Yeah. It’s two dimensions. Yeah, yeah, yeah, yeah. Like and exactly. Beside manner versus coding. Yeah.
[00:36:43] Martin Casado: Yeah.
[00:36:44] swyx: It’s, yeah.
[00:36:46] Martin Casado: I, I think for, for any, it’s hilarious. For any, for anybody listening to this for, for, for, I mean, for you, like when, when you’re like coding or using these models for something like that.
[00:36:52] Like actually just like be aware of how much of the interaction has nothing to do with coding and it just turns out to be a large portion of it. And so like, you’re, I [00:37:00] think like, like the best Soto ish model. You know, it is going to remain very important no matter what the task is.
[00:37:06] swyx: Yeah.
[00:37:07] What He’s Actually Coding: Gaussian Splats, Spark.js & 3D Scene Rendering Demos
[00:37:07] swyx: Uh, speaking of coding, uh, I, I’m gonna be cheeky and ask like, what actually are you coding?
[00:37:11] Because obviously you, you could code anything and you are obviously a busy investor and a manager of the good. Giant team. Um, what are you calling?
[00:37:18] Martin Casado: I help, um, uh, FEFA at World Labs. Uh, it’s one of the investments and um, and they’re building a foundation model that creates 3D scenes.
[00:37:27] swyx: Yeah, we had it on the pod.
[00:37:28] Yeah. Yeah,
[00:37:28] Martin Casado: yeah. And so these 3D scenes are Gaussian splats, just by the way that kind of AI works. And so like, you can reconstruct a scene better with, with, with radiance feels than with meshes. ‘cause like they don’t really have topology. So, so they, they, they produce each. Beautiful, you know, 3D rendered scenes that are Gaussian splats, but the actual industry support for Gaussian splats isn’t great.
[00:37:50] It’s just never, you know, it’s always been meshes and like, things like unreal use meshes. And so I work on a open source library called Spark js, which is a. Uh, [00:38:00] a JavaScript rendering layer ready for Gaussian splats. And it’s just because, you know, um, you, you, you need that support and, and right now there’s kind of a three js moment that’s all meshes and so like, it’s become kind of the default in three Js ecosystem.
[00:38:13] As part of that to kind of exercise the library, I just build a whole bunch of cool demos. So if you see me on X, you see like all my demos and all the world building, but all of that is just to exercise this, this library that I work on. ‘cause it’s actually a very tough algorithmics problem to actually scale a library that much.
[00:38:29] And just so you know, this is ancient history now, but 30 years ago I paid for undergrad, you know, working on game engines in college in the late nineties. So I’ve got actually a back and it’s very old background, but I actually have a background in this and so a lot of it’s fun. You know, but, but the, the, the, the whole goal is just for this rendering library to, to,
[00:38:47] Sarah Wang: are you one of the most active contributors?
[00:38:49] The, their GitHub
[00:38:50] Martin Casado: spark? Yes.
[00:38:51] Sarah Wang: Yeah, yeah.
[00:38:51] Martin Casado: There’s only two of us there, so, yes. No, so by the way, so the, the pri The pri, yeah. Yeah. So the primary developer is a [00:39:00] guy named Andres Quist, who’s an absolute genius. He and I did our, our PhDs together. And so like, um, we studied for constant Quas together. It was almost like hanging out with an old friend, you know?
[00:39:09] And so like. So he, he’s the core, core guy. I did mostly kind of, you know, the side I run venture fund.
[00:39:14] swyx: It’s amazing. Like five years ago you would not have done any of this. And it brought you back
[00:39:19] Martin Casado: the act, the Activ energy, you’re still back. Energy was so high because you had to learn all the framework b******t.
[00:39:23] Man, I f*****g used to hate that. And so like, now I don’t have to deal with that. I can like focus on the algorithmics so I can focus on the scaling and I,
[00:39:29] swyx: yeah. Yeah.
[00:39:29] LLMs vs Spatial Intelligence + How to Value World Labs’ 3D Foundation Model
[00:39:29] swyx: And then, uh, I’ll observe one irony and then I’ll ask a serious investor question, uh, which is like, the irony is FFE actually doesn’t believe that LMS can lead us to spatial intelligence.
[00:39:37] And here you are using LMS to like help like achieve spatial intelligence. I just see, I see some like disconnect in there.
[00:39:45] Martin Casado: Yeah. Yeah. So I think, I think, you know, I think, I think what she would say is LLMs are great to help with coding.
[00:39:51] swyx: Yes.
[00:39:51] Martin Casado: But like, that’s very different than a model that actually like provides, they, they’ll never have the
[00:39:56] swyx: spatial inte
[00:39:56] Martin Casado: issues.
[00:39:56] And listen, our brains clearly listen, our brains, brains clearly have [00:40:00] both our, our brains clearly have a language reasoning section and they clearly have a spatial reasoning section. I mean, it’s just, you know, these are two pretty independent problems.
[00:40:07] swyx: Okay. And you, you, like, I, I would say that the, the one data point I recently had, uh, against it is the DeepMind, uh, IMO Gold, where, so, uh, typically the, the typical answer is that this is where you start going down the neuros symbolic path, right?
[00:40:21] Like one, uh, sort of very sort of abstract reasoning thing and one form, formal thing. Um, and that’s what. DeepMind had in 2024 with alpha proof, alpha geometry, and now they just use deep think and just extended thinking tokens. And it’s one model and it’s, and it’s in LM.
[00:40:36] Martin Casado: Yeah, yeah, yeah, yeah, yeah.
[00:40:37] swyx: And so that, that was my indication of like, maybe you don’t need a separate system.
[00:40:42] Martin Casado: Yeah. So, so let me step back. I mean, at the end of the day, at the end of the day, these things are like nodes in a graph with weights on them. Right. You know, like it can be modeled like if you, if you distill it down. But let me just talk about the two different substrates. Let’s, let me put you in a dark room.
[00:40:56] Like totally black room. And then let me just [00:41:00] describe how you exit it. Like to your left, there’s a table like duck below this thing, right? I mean like the chances that you’re gonna like not run into something are very low. Now let me like turn on the light and you actually see, and you can do distance and you know how far something away is and like where it is or whatever.
[00:41:17] Then you can do it, right? Like language is not the right primitives to describe. The universe because it’s not exact enough. So that’s all Faye, Faye is talking about. When it comes to like spatial reasoning, it’s like you actually have to know that this is three feet far, like that far away. It is curved.
[00:41:37] You have to understand, you know, the, like the actual movement through space.
[00:41:40] swyx: Yeah.
[00:41:40] Martin Casado: So I do, I listen, I do think at the end of these models are definitely converging as far as models, but there’s, there’s, there’s different representations of problems you’re solving. One is language. Which, you know, that would be like describing to somebody like what to do.
[00:41:51] And the other one is actually just showing them and the space reasoning is just showing them.
[00:41:55] swyx: Yeah, yeah, yeah. Right. Got it, got it. Uh, the, in the investor question was on, on, well labs [00:42:00] is, well, like, how do I value something like this? What, what, what work does the, do you do? I’m just like, Fefe is awesome.
[00:42:07] Justin’s awesome. And you know, the other two co-founder, co-founders, but like the, the, the tech, everyone’s building cool tech. But like, what’s the value of the tech? And this is the fundamental question
[00:42:16] Martin Casado: of, well, let, let, just like these, let me just maybe give you a rough sketch on the diffusion models. I actually love to hear Sarah because I’m a venture for, you know, so like, ventures always, always like kind of wild west type
[00:42:24] swyx: stuff.
[00:42:24] You, you, you, you paid a dream and she has to like, actually
[00:42:28] Martin Casado: I’m gonna say I’m gonna mar to reality, so I’m gonna say the venture for you. And she can be like, okay, you a little kid. Yeah. So like, so, so these diffusion models literally. Create something for, for almost nothing. And something that the, the world has found to be very valuable in the past, in our real markets, right?
[00:42:45] Like, like a 2D image. I mean, that’s been an entire market. People value them. It takes a human being a long time to create it, right? I mean, to create a, you know, a, to turn me into a whatever, like an image would cost a hundred bucks in an hour. The inference cost [00:43:00] us a hundredth of a penny, right? So we’ve seen this with speech in very successful companies.
[00:43:03] We’ve seen this with 2D image. We’ve seen this with movies. Right? Now, think about 3D scene. I mean, I mean, when’s Grand Theft Auto coming out? It’s been six, what? It’s been 10 years. I mean, how, how like, but hasn’t been 10 years.
[00:43:14] Alessio: Yeah.
[00:43:15] Martin Casado: How much would it cost to like, to reproduce this room in 3D? Right. If you, if you, if you hired somebody on fiber, like in, in any sort of quality, probably 4,000 to $10,000.
[00:43:24] And then if you had a professional, probably $30,000. So if you could generate the exact same thing from a 2D image, and we know that these are used and they’re using Unreal and they’re using Blend, or they’re using movies and they’re using video games and they’re using all. So if you could do that for.
[00:43:36] You know, less than a dollar, that’s four or five orders of magnitude cheaper. So you’re bringing the marginal cost of something that’s useful down by three orders of magnitude, which historically have created very large companies. So that would be like the venture kind of strategic dreaming map.
[00:43:49] swyx: Yeah.
[00:43:50] And, and for listeners, uh, you can do this yourself on your, on your own phone with like. Uh, the marble.
[00:43:55] Martin Casado: Yeah. Marble.
[00:43:55] swyx: Uh, or but also there’s many Nerf apps where you just go on your iPhone and, and do this.
[00:43:59] Martin Casado: Yeah. Yeah. [00:44:00] Yeah. And, and in the case of marble though, it would, what you do is you literally give it in.
[00:44:03] So most Nerf apps you like kind of run around and take a whole bunch of pictures and then you kind of reconstruct it.
[00:44:08] swyx: Yeah.
[00:44:08] Martin Casado: Um, things like marble, just that the whole generative 3D space will just take a 2D image and it’ll reconstruct all the like, like
[00:44:16] swyx: meaning it has to fill in. Uh,
[00:44:18] Martin Casado: stuff at the back of the table, under the table, the back, like, like the images, it doesn’t see.
[00:44:22] So the generator stuff is very different than reconstruction that it fills in the things that you can’t see.
[00:44:26] swyx: Yeah. Okay.
[00:44:26] Sarah Wang: So,
[00:44:27] Martin Casado: all right. So now the,
[00:44:28] Sarah Wang: no, no. I mean I love that
[00:44:29] Martin Casado: the adult
[00:44:29] Sarah Wang: perspective. Um, well, no, I was gonna say these are very much a tag team. So we, we started this pod with that, um, premise. And I think this is a perfect question to even build on that further.
[00:44:36] ‘cause it truly is, I mean, we’re tag teaming all of these together.
[00:44:39] Investing in Model Labs, Media Rumors, and the Cursor Playbook (Margins & Going Down-Stack)
[00:44:39] Sarah Wang: Um, but I think every investment fundamentally starts with the same. Maybe the same two premises. One is, at this point in time, we actually believe that there are. And of one founders for their particular craft, and they have to be demonstrated in their prior careers, right?
[00:44:56] So, uh, we’re not investing in every, you know, now the term is NEO [00:45:00] lab, but every foundation model, uh, any, any company, any founder trying to build a foundation model, we’re not, um, contrary to popular opinion, we’re not invested in all of them. Right. We have a very specific thesis. I don’t think people
[00:45:09] swyx: say that about you.
[00:45:10] No, they don’t. They don’t,
[00:45:12] Sarah Wang: they say that we’re big, we’re in everything. But, um, you know, if you think about ia, right? He’s at SSI, he’s sort of. Been behind almost every foundational breakthrough for the last 15 years. 15 years. Um, if you think about, you know, the Thinking machines team, right? Mira and John, right?
[00:45:27] John is the godfather of reinforcement learning. And so, um, I go through this because, you know, if you think about for each of the bets that we’ve made, it goes back to one of, to a very specific thesis about that person, the team they’ve assembled and what they’ve done in a prior life. Um, and you know, I, I think, you know, obviously we talked about talent wars.
[00:45:46] Um, we do think. At this particular moment in time, there are particular people that can move needles. Um, clearly, uh, other companies believe that too, otherwise they wouldn’t be willing to pay such crazy prices for single individuals. So that’s, that’s one. And then two, [00:46:00] we don’t think it’s a zero sum game, right?
[00:46:02] Like if that were true open AI or, or actually just deep mind would be number one and everything, right? There’s clear value. To specialization. It’s like 11 labs. There have been so, oh my God. Yeah. Many audio models that have hit the market, they’re still fricking number one, right? And so if you think about, and they’ve created a ton of value, um, for their customers, for their investors, you know, for their team.
[00:46:23] Um, and so if you think about those two put together, right? That’s sort of the foundation of our thesis when we back, uh, these foundation model, uh, companies. Um, of course. The valuations, you know, they sound astronomical when you think about current revenue, the numbers, um, you know, there’s, there’s sort of that I would, one, I would say that’s the market out there because they are raising larger dollars.
[00:46:47] They have compute needs, right? That’s 80% of around that they typically raise or typically of, of around that they raise. Um, but I think the thing that gets us excited about backing them is that the revenue growth has [00:47:00] typically followed the capability breakthrough. So you sort of ties back to that question of.
[00:47:04] The cyclical nature, like are you just funding it and then you raise more funding? Um, when there’s a real capability breakthrough, the demand is there. And so the revenue growth is much faster than we’ve ever seen. Once it’s turned on, there’s a company, I can’t share the name, um, but their product went GA in a few weeks.
[00:47:21] Tens of millions of revenue. Right. We have
[00:47:23] swyx: SaaS
[00:47:23] I’ve
[00:47:24] swyx: seen as myself. Yes,
[00:47:24] Sarah Wang: absolutely. We have SaaS. Absolutely. Companies that, you know, have been in business for seven years and they get to the same level seven years later. And the growth is, you know, eking to whatever it is. Um, and, and by the way, great companies not, not at all, um, diminishing what they’ve accomplished, but the fact is to get to that revenue growth that quickly.
[00:47:43] It’s not just the two companies that people talk about. It’s, it’s really a lot of these, you know, sort of. Every domain has a specialist, and we think if you can win that, you become very large, very quickly, and that’s actually played out in the numbers.
[00:47:56] swyx: Yeah. Uh, our, our viewers are going to, uh, so [00:48:00] first of all, thank you for that overall take.
[00:48:01] I think like it’s important to hear you guys’ perspective because the rest of us are just kind of looking at headlines and not knowing how to make sense of any of this. Um, we can mention like my, our listeners will roast us if we, if we mention thinky and not. Discuss what happened. Uh, I mean, obviously founder split happens, um, but like, I guess is the thesis unchanged is is like, um, you know, like what’s, what’s going on in thinking?
[00:48:25] Sarah Wang: Yeah. Um, we’re more excited than ever about them. Um, they have some things that. We’re not gonna do breaking news on a, a pod. Uh, you know, obviously they should share it themselves, but, um, they’ve, you know, I think when you bring a team of that caliber together, there’s special things that happen. And, um, I think 2026 is gonna be a big year for them.
[00:48:44] Um, obviously, you know, some of the themes that we talked about before, even with just the media news storm, like the whole, something happens and then it’s everywhere instantly. Um, you know, I think, uh. [00:49:00] That’s a, i, that’s a tough situation for any company to be in. Um, but to come out of that stronger than ever, I think that, you know, we’re, we’re more bullish about thinking than, um, you know, even before.
[00:49:12] And, um, obviously,
[00:49:13] swyx: and, and, and the story is tin, uh, is tinker. It’s our custom models are all. Um, yeah. Is that, is that what, is that what we’re aiming for?
[00:49:22] Sarah Wang: Yeah. And a bunch of stuff we, we can’t talk about here. Okay. Yeah. All right. Cool. Yeah, absolutely. But no, that team is cooking and, um, you know, I think, um, they’ll, they’ll be just fine from, uh, they’ll, they’ll recover from the events in January.
[00:49:34] swyx: Yeah.
[00:49:34] Martin Casado: I will say this is the furthest, so we have a very privileged position on the boards of these companies, and like I’ll say, I’ve never seen. The perception of the truth be further from the truth.
[00:49:48] swyx: Oh,
[00:49:48] Martin Casado: industry wide ever. Like I, I guarantee you, for any of these gossipy things, I guarantee you it’s way off.
[00:49:55] swyx: Okay.
[00:49:55] Martin Casado: Way, way off. Like, like the general sentiment and like, and what happens is like we’ve got this [00:50:00] crazy game of telephone right now where there’s always. Seeds of truth, but it gets so warped by the time, like we hear all the time rumors about stuff that we’re directly involved in. Like we’re literally on the board, you know, like we’re, we’re the one that did the thing.
[00:50:12] And by the time it gets so it’s gotten so warped and so twisted. I think this is like everybody’s excited. I. There’s a lot of focus. The shot on fried is so high that people just kind of will into being things that didn’t exist. Um, so I’m not, you know, I, you know, I don’t wanna comment specifically on the thinking machines, but like,
[00:50:31] swyx: it’s an important message to the general
[00:50:33] Martin Casado: audience.
[00:50:33] I, I’ll tell you, if you hear something IX like the chances that it’s. You know, it is accurate representing, but it’s saying to is very, very low.
[00:50:42] swyx: Yeah.
[00:50:43] Sarah Wang: I have never lost so much faith in the an, an non counts on Twitter that just seemed very confident in what they’re saying. Yeah,
[00:50:50] Martin Casado: no. Yeah.
[00:50:50] Sarah Wang: And couldn’t be further from the truth.
[00:50:52] I, I had a couple days stretch where I was like, oh my God, Twitter is mind poisoned and I. Love X. Yeah,
[00:50:56] Martin Casado: but we talk to each other all the time. ‘cause we actually know, ‘cause we’re there like, we’re [00:51:00] there singing these things and like, you know, Sarah will like text me, you know, like whatever. Like, it’s like ridiculous.
[00:51:06] So for us it’s like, it’s like this ridiculous. But the problem is, is we realize that things like things start taking on a life of their own and then people assume that they’re real and, and everything. And so I think it’s very tough for founders because, you know, it’s tough enough fighting the real battle.
[00:51:20] You know now. Absolutely. Now they’re fighting phantoms too. And so, you know, you know, more and more we’re just like, and I got this from the cursory guys, which I, I really appreciate Michael Troll. He’s like, listen, head’s down, focus on the business. Yeah. And, and he absolutely crushed
[00:51:35] swyx: it.
[00:51:35] Martin Casado: Yeah. Yeah. And I, I think that’s right.
[00:51:37] I all
[00:51:37] found
[00:51:37] Martin Casado: absolutely right now, ‘cause the noise is so hot.
[00:51:40] Sarah Wang: No, that team’s been back to business for, for weeks, the thinky team. So, yeah.
[00:51:43] swyx: Yeah. Well, thank you for acknowledging in that, uh, it, it is just, uh, the hot topic at the moment. Oh, we gotta, gotta address the elephant in the room. Um, uh, cursor, right?
[00:51:51] You obviously, you guys are big investors. Uh, 2025. I would say it’s cursor year. I mean. Maybe decade, but, uh, [00:52:00] uh, just like I, I think, you know, I, I just going back to the discussion about how a GI would just kind of consume everything. Yeah. S just like the one, like the kind of the shining example of like, here’s how you build application layer.
[00:52:10] That’s a wrapper.
[00:52:11] Martin Casado: Yeah.
[00:52:12] swyx: But extremely damn good one.
[00:52:14] Martin Casado: Yeah.
[00:52:14] swyx: Uh, and, uh, I guess just like the, the general. Analysis, I guess, of, of cursors development and what it means for everyone? Like is there a cursor in every industry to be built?
[00:52:24] Martin Casado: Yeah, so the, the interesting about cursors, they actually for, you know, a small fraction of the cost, a hundred of the costs or less.
[00:52:32] Developed an almost soda model, which for a period of time was the most popular coding model in the world. Right? Which is really crazy to think about. So I think they’re just kind of doing it in reverse, right? So there, there, there’s two approaches. You start with a foundation model and then you verticalize up, or you start with the app and all of the product data and you go down and they’re the ones that are doing that.
[00:52:55] I think any company that’s doing an app has to ask the margin question. Mm-hmm. Which is like, how, how [00:53:00] do I extract margin on, on, on the tokens that are going through? Like, everybody has to be on the token path and everybody has to ask that question. And I’ve just thought they’ve been incredibly thoughtful about it.
[00:53:09] And one reason is, is if you ask. You know, Michael, what type of company are they are a developer company for professional developers. That’s what they’re, they’re a Devrel tools company. They’re just focused on coding. And that’s a hu I mean, even if you didn’t do ai, that’s a ma. You know, they, they, they, um, they acquired graphite.
[00:53:25] I mean, like, you know, listen, we were investors in GitHub, like we know how big this market is. So that’s a massive market, even without becoming a model company. But they’ve also been quite successful in doing their own models. And so I think it just shows you that if you. Are focused, you have a large use case.
[00:53:40] There’s a huge opportunity not only to get the application, but to start building your own models. Are these gonna be the only models we use? Of course not. Um, but you know, they are in a great position to serve great models and they’ve demonstrated that.
[00:53:51] swyx: Yeah. My, my, uh, sort of, uh, thesis, which we’re not gonna have to go into here is actually I think a, um, what I’ve been calling Agent Labs, which are [00:54:00] people who build on top of, uh, all the other models.
[00:54:02] Martin Casado: Yeah.
[00:54:02] swyx: Um, will probably have a better time with the margins because they, they price against the end user hours spent, or like human labor, whereas models get commodity price per token.
[00:54:15] Sarah Wang: Yeah.
[00:54:15] swyx: And so margin wise. We know inference economics for, uh, uh, model labs, but agent labs, uh, the difference is the delta between token intelligence, which keeps going down, and human costs, which keep going up.
[00:54:28] Martin Casado: Yeah, yeah, yeah.
[00:54:28] swyx: And so margin should be higher.
[00:54:31] Martin Casado: They, they, they, they, they, they should be. The, the, the, the caveat to that is if the models go first party, right. Yeah. Yeah. What they can do is they can, they can, which is
[00:54:40] swyx: the, the composer dream.
[00:54:41] Martin Casado: Yes. Yeah. They can subsidize the, no, the models, they can subsidize themselves.
[00:54:46] Oh, cloud code, code, they can subsidize themselves and then they can charge the third party more, and it’s a very delicate. Yeah, it’s because you’re kind of competing with your own customers. And so, you know, we’ve seen this historically. We saw this with the cloud with EC2, like, so this is not unusual. We [00:55:00] saw this with the operating system.
[00:55:00] It’s not unusual, but it’s playing out very, very quickly.
[00:55:04] Alessio: Yeah. Thank you for joining us. That’s all the time we have today. This is such a pleasure. You’re welcome back anytime.
[00:55:09] swyx: And thank you for being so open and also like just leading the industry in so many areas. Uh, it’s uh, really inspiring to see. So
[00:55:16] Sarah Wang: thank you so much.
[00:55:17] Thank you much. Thank you for having us.
[00:55:17] swyx: Great. Thank you.


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 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 together
Show Notes:
* Gemma 3 Paper
* Gemma 3
* Gemini 2.5 Report
* Jeff Dean’s “Software Engineering Advice from
Building 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/jeffdean
Google
* https://google.com
* https://deepmind.google

Full Video Episode
Timestamps
00:00:04 — 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 thanks
Transcript
Alessio Fanelli [00:00:04]: Hey everyone, welcome to the Latent Space podcast. This is Alessio, founder of Kernel Labs, and I’m joined by Swyx, editor of Latent Space.
Shawn Wang [00:00:11]: Hello, hello. We’re here in the studio with Jeff Dean, chief AI scientist at Google. Welcome. Thanks for having me. It’s a bit surreal to have you in the studio. I’ve watched so many of your talks, and obviously your career has been super legendary. So, I mean, congrats. I think the first thing must be said, congrats on owning the Pareto Frontier.
Jeff Dean [00:00:30]: Thank you, thank you. Pareto Frontiers are good. It’s good to be out there.
Shawn Wang [00:00:34]: Yeah, I mean, I think it’s a combination of both. You have to own the Pareto Frontier. You have to have like frontier capability, but also efficiency, and then offer that range of models that people like to use. And, you know, some part of this was started because of your hardware work. Some part of that is your model work, and I’m sure there’s lots of secret sauce that you guys have worked on cumulatively. But, like, it’s really impressive to see it all come together in, like, this slittily advanced.
Jeff Dean [00:01:04]: Yeah, yeah. I mean, I think, as you say, it’s not just one thing. It’s like a whole bunch of things up and down the stack. And, you know, all of those really combine to help make UNOS able to make highly capable large models, as well as, you know, software techniques to get those large model capabilities into much smaller, lighter weight models that are, you know, much more cost effective and lower latency, but still, you know, quite capable for their size. Yeah.
Alessio Fanelli [00:01:31]: How much pressure do you have on, like, having the lower bound of the Pareto Frontier, too? I think, like, the new labs are always trying to push the top performance frontier because they need to raise more money and all of that. And you guys have billions of users. And I think initially when you worked on the CPU, you were thinking about, you know, if everybody that used Google, we use the voice model for, like, three minutes a day, they were like, you need to double your CPU number. Like, what’s that discussion today at Google? Like, how do you prioritize frontier versus, like, we have to do this? How do we actually need to deploy it if we build it?
Jeff Dean [00:02:03]: Yeah, I mean, I think we always want to have models that are at the frontier or pushing the frontier because I think that’s where you see what capabilities now exist that didn’t exist at the sort of slightly less capable last year’s version or last six months ago version. At the same time, you know, we know those are going to be really useful for a bunch of use cases, but they’re going to be a bit slower and a bit more expensive than people might like for a bunch of other broader models. So I think what we want to do is always have kind of a highly capable sort of affordable model that enables a whole bunch of, you know, lower latency use cases. People can use them for agentic coding much more readily and then have the high-end, you know, frontier model that is really useful for, you know, deep reasoning, you know, solving really complicated math problems, those kinds of things. And it’s not that. One or the other is useful. They’re both useful. So I think we’d like to do both. And also, you know, through distillation, which is a key technique for making the smaller models more capable, you know, you have to have the frontier model in order to then distill it into your smaller model. So it’s not like an either or choice. You sort of need that in order to actually get a highly capable, more modest size model. Yeah.
Alessio Fanelli [00:03:24]: I mean, you and Jeffrey came up with the solution in 2014.
Jeff Dean [00:03:28]: Don’t forget, L’Oreal Vinyls as well. Yeah, yeah.
Alessio Fanelli [00:03:30]: A long time ago. But like, I’m curious how you think about the cycle of these ideas, even like, you know, sparse models and, you know, how do you reevaluate them? How do you think about in the next generation of model, what is worth revisiting? Like, yeah, they’re just kind of like, you know, you worked on so many ideas that end up being influential, but like in the moment, they might not feel that way necessarily. Yeah.
Jeff Dean [00:03:52]: I mean, I think distillation was originally motivated because we were seeing that we had a very large image data set at the time, you know, 300 million images that we could train on. And we were seeing that if you create specialists for different subsets of those image categories, you know, this one’s going to be really good at sort of mammals, and this one’s going to be really good at sort of indoor room scenes or whatever, and you can cluster those categories and train on an enriched stream of data after you do pre-training on a much broader set of images. You get much better performance. If you then treat that whole set of maybe 50 models you’ve trained as a large ensemble, but that’s not a very practical thing to serve, right? So distillation really came about from the idea of, okay, what if we want to actually serve that and train all these independent sort of expert models and then squish it into something that actually fits in a form factor that you can actually serve? And that’s, you know, not that different from what we’re doing today. You know, often today we’re instead of having an ensemble of 50 models. We’re having a much larger scale model that we then distill into a much smaller scale model.
Shawn Wang [00:05:09]: Yeah. A part of me also wonders if distillation also has a story with the RL revolution. So let me maybe try to articulate what I mean by that, which is you can, RL basically spikes models in a certain part of the distribution. And then you have to sort of, well, you can spike models, but usually sometimes... It might be lossy in other areas and it’s kind of like an uneven technique, but you can probably distill it back and you can, I think that the sort of general dream is to be able to advance capabilities without regressing on anything else. And I think like that, that whole capability merging without loss, I feel like it’s like, you know, some part of that should be a distillation process, but I can’t quite articulate it. I haven’t seen much papers about it.
Jeff Dean [00:06:01]: Yeah, I mean, I tend to think of one of the key advantages of distillation is that you can have a much smaller model and you can have a very large, you know, training data set and you can get utility out of making many passes over that data set because you’re now getting the logits from the much larger model in order to sort of coax the right behavior out of the smaller model that you wouldn’t otherwise get with just the hard labels. And so, you know, I think that’s what we’ve observed. Is you can get, you know, very close to your largest model performance with distillation approaches. And that seems to be, you know, a nice sweet spot for a lot of people because it enables us to kind of, for multiple Gemini generations now, we’ve been able to make the sort of flash version of the next generation as good or even substantially better than the previous generations pro. And I think we’re going to keep trying to do that because that seems like a good trend to follow.
Shawn Wang [00:07:02]: So, Dara asked, so it was the original map was Flash Pro and Ultra. Are you just sitting on Ultra and distilling from that? Is that like the mother load?
Jeff Dean [00:07:12]: I mean, we have a lot of different kinds of models. Some are internal ones that are not necessarily meant to be released or served. Some are, you know, our pro scale model and we can distill from that as well into our Flash scale model. So I think, you know, it’s an important set of capabilities to have and also inference time scaling. It can also be a useful thing to improve the capabilities of the model.
Shawn Wang [00:07:35]: And yeah, yeah, cool. Yeah. And obviously, I think the economy of Flash is what led to the total dominance. I think the latest number is like 50 trillion tokens. I don’t know. I mean, obviously, it’s changing every day.
Jeff Dean [00:07:46]: Yeah, yeah. But, you know, by market share, hopefully up.
Shawn Wang [00:07:50]: No, I mean, there’s no I mean, there’s just the economics wise, like because Flash is so economical, like you can use it for everything. Like it’s in Gmail now. It’s in YouTube. Like it’s yeah. It’s in everything.
Jeff Dean [00:08:02]: We’re using it more in our search products of various AI mode reviews.
Shawn Wang [00:08:05]: Oh, my God. Flash past the AI mode. Oh, my God. Yeah, that’s yeah, I didn’t even think about that.
Jeff Dean [00:08:10]: I mean, I think one of the things that is quite nice about the Flash model is not only is it more affordable, it’s also a lower latency. And I think latency is actually a pretty important characteristic for these models because we’re going to want models to do much more complicated things that are going to involve, you know, generating many more tokens from when you ask the model to do so. So, you know, if you’re going to ask the model to do something until it actually finishes what you ask it to do, because you’re going to ask now, not just write me a for loop, but like write me a whole software package to do X or Y or Z. And so having low latency systems that can do that seems really important. And Flash is one direction, one way of doing that. You know, obviously our hardware platforms enable a bunch of interesting aspects of our, you know, serving stack as well, like TPUs, the interconnect between. Chips on the TPUs is actually quite, quite high performance and quite amenable to, for example, long context kind of attention operations, you know, having sparse models with lots of experts. These kinds of things really, really matter a lot in terms of how do you make them servable at scale.
Alessio Fanelli [00:09:19]: Yeah. Does it feel like there’s some breaking point for like the proto Flash distillation, kind of like one generation delayed? I almost think about almost like the capability as a. In certain tasks, like the pro model today is a saturated, some sort of task. So next generation, that same task will be saturated at the Flash price point. And I think for most of the things that people use models for at some point, the Flash model in two generation will be able to do basically everything. And how do you make it economical to like keep pushing the pro frontier when a lot of the population will be okay with the Flash model? I’m curious how you think about that.
Jeff Dean [00:09:59]: I mean, I think that’s true. If your distribution of what people are asking people, the models to do is stationary, right? But I think what often happens is as the models become more capable, people ask them to do more, right? So, I mean, I think this happens in my own usage. Like I used to try our models a year ago for some sort of coding task, and it was okay at some simpler things, but wouldn’t do work very well for more complicated things. And since then, we’ve improved dramatically on the more complicated coding tasks. And now I’ll ask it to do much more complicated things. And I think that’s true, not just of coding, but of, you know, now, you know, can you analyze all the, you know, renewable energy deployments in the world and give me a report on solar panel deployment or whatever. That’s a very complicated, you know, more complicated task than people would have asked a year ago. And so you are going to want more capable models to push the frontier in the absence of what people ask the models to do. And that also then gives us. Insight into, okay, where does the, where do things break down? How can we improve the model in these, these particular areas, uh, in order to sort of, um, make the next generation even better.
Alessio Fanelli [00:11:11]: Yeah. Are there any benchmarks or like test sets they use internally? Because it’s almost like the same benchmarks get reported every time. And it’s like, all right, it’s like 99 instead of 97. Like, how do you have to keep pushing the team internally to it? Or like, this is what we’re building towards. Yeah.
Jeff Dean [00:11:26]: I mean, I think. Benchmarks, particularly external ones that are publicly available. Have their utility, but they often kind of have a lifespan of utility where they’re introduced and maybe they’re quite hard for current models. You know, I, I like to think of the best kinds of benchmarks are ones where the initial scores are like 10 to 20 or 30%, maybe, but not higher. And then you can sort of work on improving that capability for, uh, whatever it is, the benchmark is trying to assess and get it up to like 80, 90%, whatever. I, I think once it hits kind of 95% or something, you get very diminishing returns from really focusing on that benchmark, cuz it’s sort of, it’s either the case that you’ve now achieved that capability, or there’s also the issue of leakage in public data or very related kind of data being, being in your training data. Um, so we have a bunch of held out internal benchmarks that we really look at where we know that wasn’t represented in the training data at all. There are capabilities that we want the model to have. Um, yeah. Yeah. Um, that it doesn’t have now, and then we can work on, you know, assessing, you know, how do we make the model better at these kinds of things? Is it, we need different kind of data to train on that’s more specialized for this particular kind of task. Do we need, um, you know, a bunch of, uh, you know, architectural improvements or some sort of, uh, model capability improvements, you know, what would help make that better?
Shawn Wang [00:12:53]: Is there, is there such an example that you, uh, a benchmark inspired in architectural improvement? Like, uh, I’m just kind of. Jumping on that because you just.
Jeff Dean [00:13:02]: Uh, I mean, I think some of the long context capability of the, of the Gemini models that came, I guess, first in 1.5 really were about looking at, okay, we want to have, um, you know,
Shawn Wang [00:13:15]: immediately everyone jumped to like completely green charts of like, everyone had, I was like, how did everyone crack this at the same time? Right. Yeah. Yeah.
Jeff Dean [00:13:23]: I mean, I think, um, and once you’re set, I mean, as you say that needed single needle and a half. Hey, stack benchmark is really saturated for at least context links up to 1, 2 and K or something. Don’t actually have, you know, much larger than 1, 2 and 8 K these days or two or something. We’re trying to push the frontier of 1 million or 2 million context, which is good because I think there are a lot of use cases where. Yeah. You know, putting a thousand pages of text or putting, you know, multiple hour long videos and the context and then actually being able to make use of that as useful. Try to, to explore the über graduation are fairly large. But the single needle in a haystack benchmark is sort of saturated. So you really want more complicated, sort of multi-needle or more realistic, take all this content and produce this kind of answer from a long context that sort of better assesses what it is people really want to do with long context. Which is not just, you know, can you tell me the product number for this particular thing?
Shawn Wang [00:14:31]: Yeah, it’s retrieval. It’s retrieval within machine learning. It’s interesting because I think the more meta level I’m trying to operate at here is you have a benchmark. You’re like, okay, I see the architectural thing I need to do in order to go fix that. But should you do it? Because sometimes that’s an inductive bias, basically. It’s what Jason Wei, who used to work at Google, would say. Exactly the kind of thing. Yeah, you’re going to win. Short term. Longer term, I don’t know if that’s going to scale. You might have to undo that.
Jeff Dean [00:15:01]: I mean, I like to sort of not focus on exactly what solution we’re going to derive, but what capability would you want? And I think we’re very convinced that, you know, long context is useful, but it’s way too short today. Right? Like, I think what you would really want is, can I attend to the internet while I answer my question? Right? But that’s not going to happen. I think that’s going to be solved by purely scaling the existing solutions, which are quadratic. So a million tokens kind of pushes what you can do. You’re not going to do that to a trillion tokens, let alone, you know, a billion tokens, let alone a trillion. But I think if you could give the illusion that you can attend to trillions of tokens, that would be amazing. You’d find all kinds of uses for that. You would have attend to the internet. You could attend to the pixels of YouTube and the sort of deeper representations that we can find. You could attend to the form for a single video, but across many videos, you know, on a personal Gemini level, you could attend to all of your personal state with your permission. So like your emails, your photos, your docs, your plane tickets you have. I think that would be really, really useful. And the question is, how do you get algorithmic improvements and system level improvements that get you to something where you actually can attend to trillions of tokens? Right. In a meaningful way. Yeah.
Shawn Wang [00:16:26]: But by the way, I think I did some math and it’s like, if you spoke all day, every day for eight hours a day, you only generate a maximum of like a hundred K tokens, which like very comfortably fits.
Jeff Dean [00:16:38]: Right. But if you then say, okay, I want to be able to understand everything people are putting on videos.
Shawn Wang [00:16:46]: Well, also, I think that the classic example is you start going beyond language into like proteins and whatever else is extremely information dense. Yeah. Yeah.
Jeff Dean [00:16:55]: I mean, I think one of the things about Gemini’s multimodal aspects is we’ve always wanted it to be multimodal from the start. And so, you know, that sometimes to people means text and images and video sort of human-like and audio, audio, human-like modalities. But I think it’s also really useful to have Gemini know about non-human modalities. Yeah. Like LIDAR sensor data from. Yes. Say, Waymo vehicles or. Like robots or, you know, various kinds of health modalities, x-rays and MRIs and imaging and genomics information. And I think there’s probably hundreds of modalities of data where you’d like the model to be able to at least be exposed to the fact that this is an interesting modality and has certain meaning in the world. Where even if you haven’t trained on all the LIDAR data or MRI data, you could have, because maybe that’s not, you know, it doesn’t make sense in terms of trade-offs of. You know, what you include in your main pre-training data mix, at least including a little bit of it is actually quite useful. Yeah. Because it sort of tempts the model that this is a thing.
Shawn Wang [00:18:04]: Yeah. Do you believe, I mean, since we’re on this topic and something I just get to ask you all the questions I always wanted to ask, which is fantastic. Like, are there some king modalities, like modalities that supersede all the other modalities? So a simple example was Vision can, on a pixel level, encode text. And DeepSeq had this DeepSeq CR paper that did that. Vision. And Vision has also been shown to maybe incorporate audio because you can do audio spectrograms and that’s, that’s also like a Vision capable thing. Like, so, so maybe Vision is just the king modality and like. Yeah.
Jeff Dean [00:18:36]: I mean, Vision and Motion are quite important things, right? Motion. Well, like video as opposed to static images, because I mean, there’s a reason evolution has evolved eyes like 23 independent ways, because it’s such a useful capability for sensing the world around you, which is really what we want these models to be. So I think the only thing that we can be able to do is interpret the things we’re seeing or the things we’re paying attention to and then help us in using that information to do things. Yeah.
Shawn Wang [00:19:05]: I think motion, you know, I still want to shout out, I think Gemini, still the only native video understanding model that’s out there. So I use it for YouTube all the time. Nice.
Jeff Dean [00:19:15]: Yeah. Yeah. I mean, it’s actually, I think people kind of are not necessarily aware of what the Gemini models can actually do. Yeah. Like I have an example I’ve used in one of my talks. It had like, it was like a YouTube highlight video of 18 memorable sports moments across the last 20 years or something. So it has like Michael Jordan hitting some jump shot at the end of the finals and, you know, some soccer goals and things like that. And you can literally just give it the video and say, can you please make me a table of what all these different events are? What when the date is when they happened? And a short description. And so you get like now an 18 row table of that information extracted from the video, which is, you know, not something most people think of as like a turn video into sequel like table.
Alessio Fanelli [00:20:11]: Has there been any discussion inside of Google of like, you mentioned tending to the whole internet, right? Google, it’s almost built because a human cannot tend to the whole internet and you need some sort of ranking to find what you need. Yep. That ranking is like much different for an LLM because you can expect a person to look at maybe the first five, six links in a Google search versus for an LLM. Should you expect to have 20 links that are highly relevant? Like how do you internally figure out, you know, how do we build the AI mode that is like maybe like much broader search and span versus like the more human one? Yeah.
Jeff Dean [00:20:47]: I mean, I think even pre-language model based work, you know, our ranking systems would be built to start. I mean, I think even pre-language model based work, you know, our ranking systems would be built to start. With a giant number of web pages in our index, many of them are not relevant. So you identify a subset of them that are relevant with very lightweight kinds of methods. You know, you’re down to like 30,000 documents or something. And then you gradually refine that to apply more and more sophisticated algorithms and more and more sophisticated sort of signals of various kinds in order to get down to ultimately what you show, which is, you know, the final 10 results or, you know, 10 results plus. Other kinds of information. And I think an LLM based system is not going to be that dissimilar, right? You’re going to attend to trillions of tokens, but you’re going to want to identify, you know, what are the 30,000 ish documents that are with the, you know, maybe 30 million interesting tokens. And then how do you go from that into what are the 117 documents I really should be paying attention to in order to carry out the tasks that the user has asked? And I think, you know, you can imagine systems where you have, you know, a lot of highly parallel processing to identify those initial 30,000 candidates, maybe with very lightweight kinds of models. Then you have some system that sort of helps you narrow down from 30,000 to the 117 with maybe a little bit more sophisticated model or set of models. And then maybe the final model is the thing that looks. So the 117 things that might be your most capable model. So I think it has to, it’s going to be some system like that, that is really enables you to give the illusion of attending to trillions of tokens. Sort of the way Google search gives you, you know, not the illusion, but you are searching the internet, but you’re finding, you know, a very small subset of things that are, that are relevant.
Shawn Wang [00:22:47]: Yeah. I often tell a lot of people that are not steeped in like Google search history that, well, you know, like Bert was. Like he was like basically immediately inside of Google search and that improves results a lot, right? Like I don’t, I don’t have any numbers off the top of my head, but like, I’m sure you guys, that’s obviously the most important numbers to Google. Yeah.
Jeff Dean [00:23:08]: I mean, I think going to an LLM based representation of text and words and so on enables you to get out of the explicit hard notion of, of particular words having to be on the page, but really getting at the notion of this topic of this page or this page. Paragraph is highly relevant to this query. Yeah.
Shawn Wang [00:23:28]: I don’t think people understand how much LLMs have taken over all these very high traffic system, very high traffic. Yeah. Like it’s Google, it’s YouTube. YouTube has this like semantics ID thing where it’s just like every token or every item in the vocab is a YouTube video or something that predicts the video using a code book, which is absurd to me for YouTube size.
Jeff Dean [00:23:50]: And then most recently GROK also for, for XAI, which is like, yeah. I mean, I’ll call out even before LLMs were used extensively in search, we put a lot of emphasis on softening the notion of what the user actually entered into the query.
Shawn Wang [00:24:06]: So do you have like a history of like, what’s the progression? Oh yeah.
Jeff Dean [00:24:09]: I mean, I actually gave a talk in, uh, I guess, uh, web search and data mining conference in 2009, uh, where we never actually published any papers about the origins of Google search, uh, sort of, but we went through sort of four or five or six. generations, four or five or six generations of, uh, redesigning of the search and retrieval system, uh, from about 1999 through 2004 or five. And that talk is really about that evolution. And one of the things that really happened in 2001 was we were sort of working to scale the system in multiple dimensions. So one is we wanted to make our index bigger, so we could retrieve from a larger index, which always helps your quality in general. Uh, because if you don’t have the page in your index, you’re going to not do well. Um, and then we also needed to scale our capacity because we were, our traffic was growing quite extensively. Um, and so we had, you know, a sharded system where you have more and more shards as the index grows, you have like 30 shards. And then if you want to double the index size, you make 60 shards so that you can bound the latency by which you respond for any particular user query. Um, and then as traffic grows, you add, you add more and more replicas of each of those. And so we eventually did the math that realized that in a data center where we had say 60 shards and, um, you know, 20 copies of each shard, we now had 1200 machines, uh, with disks. And we did the math and we’re like, Hey, one copy of that index would actually fit in memory across 1200 machines. So in 2001, we introduced, uh, we put our entire index in memory and what that enabled from a quality perspective was amazing. Um, and so we had more and more replicas of each of those. Before you had to be really careful about, you know, how many different terms you looked at for a query, because every one of them would involve a disk seek on every one of the 60 shards. And so you, as you make your index bigger, that becomes even more inefficient. But once you have the whole index in memory, it’s totally fine to have 50 terms you throw into the query from the user’s original three or four word query, because now you can add synonyms like restaurant and restaurants and cafe and, uh, you know, things like that. Uh, bistro and all these things. And you can suddenly start, uh, sort of really, uh, getting at the meaning of the word as opposed to the exact semantic form the user typed in. And that was, you know, 2001, very much pre LLM, but really it was about softening the, the strict definition of what the user typed in order to get at the meaning.
Alessio Fanelli [00:26:47]: What are like principles that you use to like design the systems, especially when you have, I mean, in 2001, the internet is like. Doubling, tripling every year in size is not like, uh, you know, and I think today you kind of see that with LLMs too, where like every year the jumps in size and like capabilities are just so big. Are there just any, you know, principles that you use to like, think about this? Yeah.
Jeff Dean [00:27:08]: I mean, I think, uh, you know, first, whenever you’re designing a system, you want to understand what are the sort of design parameters that are going to be most important in designing that, you know? So, you know, how many queries per second do you need to handle? How big is the internet? How big is the index you need to handle? How much data do you need to keep for every document in the index? How are you going to look at it when you retrieve things? Um, what happens if traffic were to double or triple, you know, will that system work well? And I think a good design principle is you’re going to want to design a system so that the most important characteristics could scale by like factors of five or 10, but probably not beyond that because often what happens is if you design a system for X. And something suddenly becomes a hundred X, that would enable a very different point in the design space that would not make sense at X. But all of a sudden at a hundred X makes total sense. So like going from a disk space index to a in memory index makes a lot of sense once you have enough traffic, because now you have enough replicas of the sort of state on disk that those machines now actually can hold, uh, you know, a full copy of the, uh, index and memory. Yeah. And that all of a sudden enabled. A completely different design that wouldn’t have been practical before. Yeah. Um, so I’m, I’m a big fan of thinking through designs in your head, just kind of playing with the design space a little before you actually do a lot of writing of code. But, you know, as you said, in the early days of Google, we were growing the index, uh, quite extensively. We were growing the update rate of the index. So the update rate actually is the parameter that changed the most. Surprising. So it used to be once a month.
Shawn Wang [00:28:55]: Yeah.
Jeff Dean [00:28:56]: And then we went to a system that could update any particular page in like sub one minute. Okay.
Shawn Wang [00:29:02]: Yeah. Because this is a competitive advantage, right?
Jeff Dean [00:29:04]: Because all of a sudden news related queries, you know, if you’re, if you’ve got last month’s news index, it’s not actually that useful for.
Shawn Wang [00:29:11]: News is a special beast. Was there any, like you could have split it onto a separate system.
Jeff Dean [00:29:15]: Well, we did. We launched a Google news product, but you also want news related queries that people type into the main index to also be sort of updated.
Shawn Wang [00:29:23]: So, yeah, it’s interesting. And then you have to like classify whether the page is, you have to decide which pages should be updated and what frequency. Oh yeah.
Jeff Dean [00:29:30]: There’s a whole like, uh, system behind the scenes that’s trying to decide update rates and importance of the pages. So even if the update rate seems low, you might still want to recrawl important pages quite often because, uh, the likelihood they change might be low, but the value of having updated is high.
Shawn Wang [00:29:50]: Yeah, yeah, yeah, yeah. Uh, well, you know, yeah. This, uh, you know, mention of latency and, and saving things to this reminds me of one of your classics, which I have to bring up, which is latency numbers. Every programmer should know, uh, was there a, was it just a, just a general story behind that? Did you like just write it down?
Jeff Dean [00:30:06]: I mean, this has like sort of eight or 10 different kinds of metrics that are like, how long does a cache mistake? How long does branch mispredict take? How long does a reference domain memory take? How long does it take to send, you know, a packet from the U S to the Netherlands or something? Um,
Shawn Wang [00:30:21]: why Netherlands, by the way, or is it, is that because of Chrome?
Jeff Dean [00:30:25]: Uh, we had a data center in the Netherlands, um, so, I mean, I think this gets to the point of being able to do the back of the envelope calculations. So these are sort of the raw ingredients of those, and you can use them to say, okay, well, if I need to design a system to do image search and thumb nailing or something of the result page, you know, how, what I do that I could pre-compute the image thumbnails. I could like. Try to thumbnail them on the fly from the larger images. What would that do? How much dis bandwidth than I need? How many des seeks would I do? Um, and you can sort of actually do thought experiments in, you know, 30 seconds or a minute with the sort of, uh, basic, uh, basic numbers at your fingertips. Uh, and then as you sort of build software using higher level libraries, you kind of want to develop the same intuitions for how long does it take to, you know, look up something in this particular kind of.
Shawn Wang [00:31:21]: I’ll see you next time.
Shawn Wang [00:31:51]: Which is a simple byte conversion. That’s nothing interesting. I wonder if you have any, if you were to update your...
Jeff Dean [00:31:58]: I mean, I think it’s really good to think about calculations you’re doing in a model, either for training or inference.
Jeff Dean [00:32:09]: Often a good way to view that is how much state will you need to bring in from memory, either like on-chip SRAM or HBM from the accelerator. Attached memory or DRAM or over the network. And then how expensive is that data motion relative to the cost of, say, an actual multiply in the matrix multiply unit? And that cost is actually really, really low, right? Because it’s order, depending on your precision, I think it’s like sub one picodule.
Shawn Wang [00:32:50]: Oh, okay. You measure it by energy. Yeah. Yeah.
Jeff Dean [00:32:52]: Yeah. I mean, it’s all going to be about energy and how do you make the most energy efficient system. And then moving data from the SRAM on the other side of the chip, not even off the off chip, but on the other side of the same chip can be, you know, a thousand picodules. Oh, yeah. And so all of a sudden, this is why your accelerators require batching. Because if you move, like, say, the parameter of a model from SRAM on the, on the chip into the multiplier unit, that’s going to cost you a thousand picodules. So you better make use of that, that thing that you moved many, many times with. So that’s where the batch dimension comes in. Because all of a sudden, you know, if you have a batch of 256 or something, that’s not so bad. But if you have a batch of one, that’s really not good.
Shawn Wang [00:33:40]: Yeah. Yeah. Right.
Jeff Dean [00:33:41]: Because then you paid a thousand picodules in order to do your one picodule multiply.
Shawn Wang [00:33:46]: I have never heard an energy-based analysis of batching.
Jeff Dean [00:33:50]: Yeah. I mean, that’s why people batch. Yeah. Ideally, you’d like to use batch size one because the latency would be great.
Shawn Wang [00:33:56]: The best latency.
Jeff Dean [00:33:56]: But the energy cost and the compute cost inefficiency that you get is quite large. So, yeah.
Shawn Wang [00:34:04]: Is there a similar trick like, like, like you did with, you know, putting everything in memory? Like, you know, I think obviously NVIDIA has caused a lot of waves with betting very hard on SRAM with Grok. I wonder if, like, that’s something that you already saw with, with the TPUs, right? Like that, that you had to. Uh, to serve at your scale, uh, you probably sort of saw that coming. Like what, what, what hardware, uh, innovations or insights were formed because of what you’re seeing there?
Jeff Dean [00:34:33]: Yeah. I mean, I think, you know, TPUs have this nice, uh, sort of regular structure of 2D or 3D meshes with a bunch of chips connected. Yeah. And each one of those has HBM attached. Um, I think for serving some kinds of models, uh, you know, you, you pay a lot higher cost. Uh, and time latency, um, bringing things in from HBM than you do bringing them in from, uh, SRAM on the chip. So if you have a small enough model, you can actually do model parallelism, spread it out over lots of chips and you actually get quite good throughput improvements and latency improvements from doing that. And so you’re now sort of striping your smallish scale model over say 16 or 64 chips. Uh, but as if you do that and it all fits in. In SRAM, uh, that can be a big win. So yeah, that’s not a surprise, but it is a good technique.
Alessio Fanelli [00:35:27]: Yeah. What about the TPU design? Like how much do you decide where the improvements have to go? So like, this is like a good example of like, is there a way to bring the thousand picojoules down to 50? Like, is it worth designing a new chip to do that? The extreme is like when people say, oh, you should burn the model on the ASIC and that’s kind of like the most extreme thing. How much of it? Is it worth doing an hardware when things change so quickly? Like what was the internal discussion? Yeah.
Jeff Dean [00:35:57]: I mean, we, we have a lot of interaction between say the TPU chip design architecture team and the sort of higher level modeling, uh, experts, because you really want to take advantage of being able to co-design what should future TPUs look like based on where we think the sort of ML research puck is going, uh, in some sense, because, uh, you know, as a hardware designer for ML and in particular, you’re trying to design a chip starting today and that design might take two years before it even lands in a data center. And then it has to sort of be a reasonable lifetime of the chip to take you three, four or five years. So you’re trying to predict two to six years out where, what ML computations will people want to run two to six years out in a very fast changing field. And so having people with interest. Interesting ML research ideas of things we think will start to work in that timeframe or will be more important in that timeframe, uh, really enables us to then get, you know, interesting hardware features put into, you know, TPU N plus two, where TPU N is what we have today.
Shawn Wang [00:37:10]: Oh, the cycle time is plus two.
Jeff Dean [00:37:12]: Roughly. Wow. Because, uh, I mean, sometimes you can squeeze some changes into N plus one, but, you know, bigger changes are going to require the chip. Yeah. Design be earlier in its lifetime design process. Um, so whenever we can do that, it’s generally good. And sometimes you can put in speculative features that maybe won’t cost you much chip area, but if it works out, it would make something, you know, 10 times as fast. And if it doesn’t work out, well, you burned a little bit of tiny amount of your chip area on that thing, but it’s not that big a deal. Uh, sometimes it’s a very big change and we want to be pretty sure this is going to work out. So we’ll do like lots of carefulness. Uh, ML experimentation to show us, uh, this is actually the, the way we want to go. Yeah.
Alessio Fanelli [00:37:58]: Is there a reverse of like, we already committed to this chip design so we can not take the model architecture that way because it doesn’t quite fit?
Jeff Dean [00:38:06]: Yeah. I mean, you, you definitely have things where you’re going to adapt what the model architecture looks like so that they’re efficient on the chips that you’re going to have for both training and inference of that, of that, uh, generation of model. So I think it kind of goes both ways. Um, you know, sometimes you can take advantage of, you know, lower precision things that are coming in a future generation. So you can, might train it at that lower precision, even if the current generation doesn’t quite do that. Mm.
Shawn Wang [00:38:40]: Yeah. How low can we go in precision?
Jeff Dean [00:38:43]: Because people are saying like ternary is like, uh, yeah, I mean, I’m a big fan of very low precision because I think that gets, that saves you a tremendous amount of time. Right. Because it’s picojoules per bit that you’re transferring and reducing the number of bits is a really good way to, to reduce that. Um, you know, I think people have gotten a lot of luck, uh, mileage out of having very low bit precision things, but then having scaling factors that apply to a whole bunch of, uh, those, those weights. Scaling. How does it, how does it, okay.
Shawn Wang [00:39:15]: Interesting. You, so low, low precision, but scaled up weights. Yeah. Huh. Yeah. Never considered that. Yeah. Interesting. Uh, w w while we’re on this topic, you know, I think there’s a lot of, um, uh, this, the concept of precision at all is weird when we’re sampling, you know, uh, we just, at the end of this, we’re going to have all these like chips that I’ll do like very good math. And then we’re just going to throw a random number generator at the start. So, I mean, there’s a movement towards, uh, energy based, uh, models and processors. I’m just curious if you’ve, obviously you’ve thought about it, but like, what’s your commentary?
Jeff Dean [00:39:50]: Yeah. I mean, I think. There’s a bunch of interesting trends though. Energy based models is one, you know, diffusion based models, which don’t sort of sequentially decode tokens is another, um, you know, speculative decoding is a way that you can get sort of an equivalent, very small.
Shawn Wang [00:40:06]: Draft.
Jeff Dean [00:40:07]: Batch factor, uh, for like you predict eight tokens out and that enables you to sort of increase the effective batch size of what you’re doing by a factor of eight, even, and then you maybe accept five or six of those tokens. So you get. A five, a five X improvement in the amortization of moving weights, uh, into the multipliers to do the prediction for the, the tokens. So these are all really good techniques and I think it’s really good to look at them from the lens of, uh, energy, real energy, not energy based models, um, and, and also latency and throughput, right? If you look at things from that lens, that sort of guides you to. Two solutions that are gonna be, uh, you know, better from, uh, you know, being able to serve larger models or, you know, equivalent size models more cheaply and with lower latency.
Shawn Wang [00:41:03]: Yeah. Well, I think, I think I, um, it’s appealing intellectually, uh, haven’t seen it like really hit the mainstream, but, um, I do think that, uh, there’s some poetry in the sense that, uh, you know, we don’t have to do, uh, a lot of shenanigans if like we fundamentally. Design it into the hardware. Yeah, yeah.
Jeff Dean [00:41:23]: I mean, I think there’s still a, there’s also sort of the more exotic things like analog based, uh, uh, computing substrates as opposed to digital ones. Uh, I’m, you know, I think those are super interesting cause they can be potentially low power. Uh, but I think you often end up wanting to interface that with digital systems and you end up losing a lot of the power advantages in the digital to analog and analog to digital conversions. You end up doing, uh, at the sort of boundaries. And periphery of that system. Um, I still think there’s a tremendous distance we can go from where we are today in terms of energy efficiency with sort of, uh, much better and specialized hardware for the models we care about.
Shawn Wang [00:42:05]: Yeah.
Alessio Fanelli [00:42:06]: Um, any other interesting research ideas that you’ve seen, or like maybe things that you cannot pursue a Google that you would be interested in seeing researchers take a step at, I guess you have a lot of researchers. Yeah, I guess you have enough, but our, our research.
Jeff Dean [00:42:21]: Our research portfolio is pretty broad. I would say, um, I mean, I think, uh, in terms of research directions, there’s a whole bunch of, uh, you know, open problems and how do you make these models reliable and able to do much longer, kind of, uh, more complex tasks that have lots of subtasks. How do you orchestrate, you know, maybe one model that’s using other models as tools in order to sort of build, uh, things that can accomplish, uh, you know, much more. Yeah. Significant pieces of work, uh, collectively, then you would ask a single model to do. Um, so that’s super interesting. How do you get more verifiable, uh, you know, how do you get RL to work for non-verifiable domains? I think it’s a pretty interesting open problem because I think that would broaden out the capabilities of the models, the improvements that you’re seeing in both math and coding. Uh, if we could apply those to other less verifiable domains, because we’ve come up with RL techniques that actually enable us to do that. Uh, effectively, that would, that would really make the models improve quite a lot. I think.
Alessio Fanelli [00:43:26]: I’m curious, like when we had Noam Brown on the podcast, he said, um, they already proved you can do it with deep research. Um, you kind of have it with AI mode in a way it’s not verifiable. I’m curious if there’s any thread that you think is interesting there. Like what is it? Both are like information retrieval of JSON. So I wonder if it’s like the retrieval is like the verifiable part. That you can score or what are like, yeah, yeah. How, how would you model that, that problem?
Jeff Dean [00:43:55]: Yeah. I mean, I think there are ways of having other models that can evaluate the results of what a first model did, maybe even retrieving. Can you have another model that says, is this things, are these things you retrieved relevant? Or can you rate these 2000 things you retrieved to assess which ones are the 50 most relevant or something? Um, I think those kinds of techniques are actually quite effective. Sometimes I can even be the same model, just prompted differently to be a, you know, a critic as opposed to a, uh, actual retrieval system. Yeah.
Shawn Wang [00:44:28]: Um, I do think like there, there is that, that weird cliff where like, it feels like we’ve done the easy stuff and then now it’s, but it always feels like that every year. It’s like, oh, like we know, we know, and the next part is super hard and nobody’s figured it out. And, uh, exactly with this RLVR thing where like everyone’s talking about, well, okay, how do we. the next stage of the non-verifiable stuff. And everyone’s like, I don’t know, you know, Ellen judge.
Jeff Dean [00:44:56]: I mean, I feel like the nice thing about this field is there’s lots and lots of smart people thinking about creative solutions to some of the problems that we all see. Uh, because I think everyone sort of sees that the models, you know, are great at some things and they fall down around the edges of those things and, and are not as capable as we’d like in those areas. And then coming up with good techniques and trying those. And seeing which ones actually make a difference is sort of what the whole research aspect of this field is, is pushing forward. And I think that’s why it’s super interesting. You know, if you think about two years ago, we were struggling with GSM, eight K problems, right? Like, you know, Fred has two rabbits. He gets three more rabbits. How many rabbits does he have? That’s a pretty far cry from the kinds of mathematics that the models can, and now you’re doing IMO and Erdos problems in pure language. Yeah. Yeah. Pure language. So that is a really, really amazing jump in capabilities in, you know, in a year and a half or something. And I think, um, for other areas, it’d be great if we could make that kind of leap. Uh, and you know, we don’t exactly see how to do it for some, some areas, but we do see it for some other areas and we’re going to work hard on making that better. Yeah.
Shawn Wang [00:46:13]: Yeah.
Alessio Fanelli [00:46:14]: Like YouTube thumbnail generation. That would be very helpful. We need that. That would be AGI. We need that.
Shawn Wang [00:46:20]: That would be. As far as content creators go.
Jeff Dean [00:46:22]: I guess I’m not a YouTube creator, so I don’t care that much about that problem, but I guess, uh, many people do.
Shawn Wang [00:46:27]: It does. Yeah. It doesn’t, it doesn’t matter. People do judge books by their covers as it turns out. Um, uh, just to draw a bit on the IMO goal. Um, I’m still not over the fact that a year ago we had alpha proof and alpha geometry and all those things. And then this year we were like, screw that we’ll just chuck it into Gemini. Yeah. What’s your reflection? Like, I think this, this question about. Like the merger of like symbolic systems and like, and, and LMS, uh, was a very much core belief. And then somewhere along the line, people would just said, Nope, we’ll just all do it in the LLM.
Jeff Dean [00:47:02]: Yeah. I mean, I think it makes a lot of sense to me because, you know, humans manipulate symbols, but we probably don’t have like a symbolic representation in our heads. Right. We have some distributed representation that is neural net, like in some way of lots of different neurons. And activation patterns firing when we see certain things and that enables us to reason and plan and, you know, do chains of thought and, you know, roll them back now that, that approach for solving the problem doesn’t seem like it’s going to work. I’m going to try this one. And, you know, in a lot of ways we’re emulating what we intuitively think, uh, is happening inside real brains in neural net based models. So it never made sense to me to have like completely separate. Uh, discrete, uh, symbolic things, and then a completely different way of, of, uh, you know, thinking about those things.
Shawn Wang [00:47:59]: Interesting. Yeah. Uh, I mean, it’s maybe seems obvious to you, but it wasn’t obvious to me a year ago. Yeah.
Jeff Dean [00:48:06]: I mean, I do think like that IMO with, you know, translating to lean and using lean and then the next year and also a specialized geometry model. And then this year switching to a single unified model. That is roughly the production model with a little bit more inference budget, uh, is actually, you know, quite good because it shows you that the capabilities of that general model have improved dramatically and, and now you don’t need the specialized model. This is actually sort of very similar to the 2013 to 16 era of machine learning, right? Like it used to be, people would train separate models for lots of different, each different problem, right? I have, I want to recognize street signs and something. So I train a street sign. Recognition recognition model, or I want to, you know, decode speech recognition. I have a speech model, right? I think now the era of unified models that do everything is really upon us. And the question is how well do those models generalize to new things they’ve never been asked to do and they’re getting better and better.
Shawn Wang [00:49:10]: And you don’t need domain experts. Like one of my, uh, so I interviewed ETA who was on, who was on that team. Uh, and he was like, yeah, I, I don’t know how they work. I don’t know where the IMO competition was held. I don’t know the rules of it. I just trained the models, the training models. Yeah. Yeah. And it’s kind of interesting that like people with these, this like universal skill set of just like machine learning, you just give them data and give them enough compute and they can kind of tackle any task, which is the bitter lesson, I guess. I don’t know. Yeah.
Jeff Dean [00:49:39]: I mean, I think, uh, general models, uh, will win out over specialized ones in most cases.
Shawn Wang [00:49:45]: Uh, so I want to push there a bit. I think there’s one hole here, which is like, uh. There’s this concept of like, uh, maybe capacity of a model, like abstractly a model can only contain the number of bits that it has. And, uh, and so it, you know, God knows like Gemini pro is like one to 10 trillion parameters. We don’t know, but, uh, the Gemma models, for example, right? Like a lot of people want like the open source local models that are like that, that, that, and, and, uh, they have some knowledge, which is not necessary, right? Like they can’t know everything like, like you have the. The luxury of you have the big model and big model should be able to capable of everything. But like when, when you’re distilling and you’re going down to the small models, you know, you’re actually memorizing things that are not useful. Yeah. And so like, how do we, I guess, do we want to extract that? Can we, can we divorce knowledge from reasoning, you know?
Jeff Dean [00:50:38]: Yeah. I mean, I think you do want the model to be most effective at reasoning if it can retrieve things, right? Because having the model devote precious parameter space. To remembering obscure facts that could be looked up is actually not the best use of that parameter space, right? Like you might prefer something that is more generally useful in more settings than this obscure fact that it has. Um, so I think that’s always attention at the same time. You also don’t want your model to be kind of completely detached from, you know, knowing stuff about the world, right? Like it’s probably useful to know how long the golden gate be. Bridges just as a general sense of like how long are bridges, right? And, uh, it should have that kind of knowledge. It maybe doesn’t need to know how long some teeny little bridge in some other more obscure part of the world is, but, uh, it does help it to have a fair bit of world knowledge and the bigger your model is, the more you can have. Uh, but I do think combining retrieval with sort of reasoning and making the model really good at doing multiple stages of retrieval. Yeah.
Shawn Wang [00:51:49]: And reasoning through the intermediate retrieval results is going to be a, a pretty effective way of making the model seem much more capable, because if you think about, say, a personal Gemini, yeah, right?
Jeff Dean [00:52:01]: Like we’re not going to train Gemini on my email. Probably we’d rather have a single model that, uh, we can then use and use being able to retrieve from my email as a tool and have the model reason about it and retrieve from my photos or whatever, uh, and then make use of that and have multiple. Um, you know, uh, stages of interaction. that makes sense.
Alessio Fanelli [00:52:24]: Do you think the vertical models are like, uh, interesting pursuit? Like when people are like, oh, we’re building the best healthcare LLM, we’re building the best law LLM, are those kind of like short-term stopgaps or?
Jeff Dean [00:52:37]: No, I mean, I think, I think vertical models are interesting. Like you want them to start from a pretty good base model, but then you can sort of, uh, sort of viewing them, view them as enriching the data. Data distribution for that particular vertical domain for healthcare, say, um, we’re probably not going to train or for say robotics. We’re probably not going to train Gemini on all possible robotics data. We, you could train it on because we want it to have a balanced set of capabilities. Um, so we’ll expose it to some robotics data, but if you’re trying to build a really, really good robotics model, you’re going to want to start with that and then train it on more robotics data. And then maybe that would. It’s multilingual translation capability, but improve its robotics capabilities. And we’re always making these kind of, uh, you know, trade-offs in the data mix that we train the base Gemini models on. You know, we’d love to include data from 200 more languages and as much data as we have for those languages, but that’s going to displace some other capabilities of the model. It won’t be as good at, um, you know, Pearl programming, you know, it’ll still be good at Python programming. Cause we’ll include it. Enough. Of that, but there’s other long tail computer languages or coding capabilities that it may suffer on or multi, uh, multimodal reasoning capabilities may suffer. Cause we didn’t get to expose it to as much data there, but it’s really good at multilingual things. So I, I think some combination of specialized models, maybe more modular models. So it’d be nice to have the capability to have those 200 languages, plus this awesome robotics model, plus this awesome healthcare, uh, module that all can be knitted together to work in concert and called upon in different circumstances. Right? Like if I have a health related thing, then it should enable using this health module in conjunction with the main base model to be even better at those kinds of things. Yeah.
Shawn Wang [00:54:36]: Installable knowledge. Yeah.
Jeff Dean [00:54:37]: Right.
Shawn Wang [00:54:38]: Just download as a, as a package.
Jeff Dean [00:54:39]: And some of that installable stuff can come from retrieval, but some of it probably should come from preloaded training on, you know, uh, a hundred billion tokens or a trillion tokens of health data. Yeah.
Shawn Wang [00:54:51]: And for listeners, I think, uh, I will highlight the Gemma three end paper where they, there was a little bit of that, I think. Yeah.
Alessio Fanelli [00:54:56]: Yeah. I guess the question is like, how many billions of tokens do you need to outpace the frontier model improvements? You know, it’s like, if I have to make this model better healthcare and the main. Gemini model is still improving. Do I need 50 billion tokens? Can I do it with a hundred, if I need a trillion healthcare tokens, it’s like, they’re probably not out there that you don’t have, you know, I think that’s really like the.
Jeff Dean [00:55:21]: Well, I mean, I think healthcare is a particularly challenging domain, so there’s a lot of healthcare data that, you know, we don’t have access to appropriately, but there’s a lot of, you know, uh, healthcare organizations that want to train models on their own data. That is not public healthcare data, uh, not public health. But public healthcare data. Um, so I think there are opportunities there to say, partner with a large healthcare organization and train models for their use that are going to be, you know, more bespoke, but probably, uh, might be better than a general model trained on say, public data. Yeah.
Shawn Wang [00:55:58]: Yeah. I, I believe, uh, by the way, also this is like somewhat related to the language conversation. Uh, I think one of your, your favorite examples was you can put a low resource language in the context and it just learns. Yeah.
Jeff Dean [00:56:09]: Oh, yeah, I think the example we used was Calamon, which is truly low resource because it’s only spoken by, I think 120 people in the world and there’s no written text.
Shawn Wang [00:56:20]: So, yeah. So you can just do it that way. Just put it in the context. Yeah. Yeah. But I think your whole data set in the context, right.
Jeff Dean [00:56:27]: If you, if you take a language like, uh, you know, Somali or something, there is a fair bit of Somali text in the world that, uh, or Ethiopian Amharic or something, um, you know, we probably. Yeah. Are not putting all the data from those languages into the Gemini based training. We put some of it, but if you put more of it, you’ll improve the capabilities of those models.
Shawn Wang [00:56:49]: Yeah.
Jeff Dean [00:56:49]: So, or of those languages.
Shawn Wang [00:56:52]: Uh, yeah, cool. Uh, it’s, uh, I have a side interest in linguistics. I, I, I did, uh, uh, a few classes back in college and like, uh, part of me, like if I was a linguist and I could have access to all these models, I would just be asking really fundamental questions about language itself. Yeah. Like, uh, one is th there’s one very obvious one, which is Sapir-Whorf, like how much does like the language that you speak affect your thinking, but then also there’s some languages where there’s just concepts that are not represented in other languages, but some others, many others that are just duplicates, right. Where, uh, there’s also another paper that people love called the platonic representation where, you know, like the, the, an image of a cup is, uh, if you say learn a model on that and you, you, you have a lot of texts with the word cup eventually maps to it, like roughly the same place. And so like that should apply to languages except where it doesn’t. And that’s actually like very interesting differences in what humanity has discovered as concepts that maybe English doesn’t have.
Shawn Wang [00:57:54]: I don’t know. It’s just like my, my rant on languages. Yeah.
Jeff Dean [00:57:58]: I mean, I, I did some work on a early model that fused together a language based model with you have, you know, nice word based representations and then an image model where you have. Trained it on image net like things. Yes. And then you fuse together the top layers of, uh, no, this is devise, uh, uh, the, you do a little bit more training to fuse together those representations. And what you found was that if you give a novel image that is not in any of the categories in the image model, it was trained on the model can often assigns kind of the right cat, the right label to that image. Um, so for example, um, I think, uh, telescope and, uh, binoculars were both in the training, uh, categories for the image model, but, um, microscope was not. Hmm. And so if you’re given an image of a microscope, it actually can come up with something that’s, uh, got the word microscope as the label that it assigns, even though it’s never actually seen an image labeled that.
Shawn Wang [00:59:01]: Oh, that’s nice. That’s kind of cool. Yeah.
Jeff Dean [00:59:04]: Um, so yeah.
Shawn Wang [00:59:07]: Useful. Uh, cool. Uh, I think. There, there’s more general, like broad questions, but like, I guess what, what do you, uh, wish you were asked more in, in, in general, like, you know, like you, you have such a broad scope. We’ve covered the hardware, we’ve covered the, the, the models research. Yeah.
Jeff Dean [00:59:22]: I mean, I think, uh, one thing that’s kind of interesting is, you know, I, I did a undergrad thesis on neural network, uh, training, uh, uh, parallel neural network training, uh, back in 1990 when I got exposed to neural nets and I always felt kind of, they were the right abstraction. Uh, but we just needed way more compute than we had then. Mm-hmm. So like the 32 processors in the department parallel computer, you know, could get you a, a little bit more interesting, uh, model, but not, not enough to solve real problems. And so starting in 2008 or nine, you know, the world started to have enough computing power through Moore’s law and, you know, larger, interesting data sets to train on to actually, you know, start training neural nets that could tackle real problems that people cared about. Yeah. Speech recognition. Vision, and eventually, uh, language. Um, and so, um, when I started working on neural nets at Google in, in late 2011, um, you know, I really just felt like we should scale up the size of neural networks we can train using, you know, large amounts of parallel computation. And so, uh, I actually, uh, revived some ideas from my undergrad thesis where I’d done both model parallel and data parallel, uh, training and I compared them. Uh, I, I called them. I’ve been doing this since I was eight. It was something different. There was like pattern partitioned and, you know, model partitioned or something.
Shawn Wang [01:00:43]: Well, I have to, is it, is it public? And we can go dig it up?
Jeff Dean [01:00:45]: Yeah, it’s on, it’s on the web. Okay, nice. Um, but, uh, you know, I think combining a lot of those techniques and really just trying to push on scaling things up over the last, you know, 15 years has been, you know, really important. And that means, you know, improvements in the hardware. So, you know, pushing on building specialized hardware like TPUs. Uh, it also means, you know, pushing on software, abstraction layers to let people express their ideas to the computer. Thank you for having me.
Jeff Dean [01:01:40]: Thank you for having me.
Shawn Wang [01:07:10]: If that’s something you would agree with at the time, or is there a different post-mortem?
Jeff Dean [01:07:15]: The brain marketplace for compute quotas.
Shawn Wang [01:07:18]: Compute quotas, where basically he was like, okay, David worked at OpenAI as VP Engine and then he worked at Google. He was like, fundamentally, OpenAI was willing to go all in, like, bet the farm on one thing, whereas Google was more democratic. Everyone had a quota. And I was like, okay, if you believe in scaling as an important thing, that’s an important organizational-wide decision to do.
Jeff Dean [01:07:41]: Yeah. Yeah, I mean, I think I would somewhat agree with that. I mean, I think I actually wrote a one-page memo saying we were being stupid by fragmenting our resources. So in particular, at the time, we had efforts within Google Research. And in the brain team in particular, on large language models. We also had efforts on multimodal models in other parts of brain and Google Research. And then Legacy DeepMind had efforts like Chinchilla models and Flamingo models. And so really, we were fragmenting not only our compute across those separate efforts, but also our best people and our best. And so I said, this is just stupid. Why don’t we combine things and have one effort to train an awesome single unified model that is multimodal from the start, that’s good at everything. And that was the origin of the Gemini effort.
Shawn Wang [01:08:52]: And my one-page memo worked, which is good. Did you have the name? Because also for those who don’t know, you named Gemini.
Jeff Dean [01:08:58]: I did. There was another name proposed. And I said, you know what? You know, it’s sort of like these two organizations really are like twins in some sense coming together. So I kind of like that. And then there’s also the NASA interpretation of the early Gemini project being an important thing on your way to the Apollo project. So it seemed like a good name. Twins coming together. Right.
Alessio Fanelli [01:09:27]: Yeah. Nice. I know we’re already running out of time, but I’m curious how you use AI. Today to code. So, I mean, you’re probably one of the most prolific engineers in the history of computer science. Um, I was reading on through the article about you and Sanjay’s friendship and how you work together. And you have one quote about, you need to find someone that you’re going to pair program with who’s compatible with your way of thinking so that the two of you together are a complimentary force. And I was thinking about how you think about coding agents and this, like, how do you shape a coding agents to be compatible with your way of thinking? Like. How would you rate the tools today? Like, where should things go? Yeah.
Jeff Dean [01:10:07]: I mean, first, I think the coding tools are, you know, getting vastly better compared to where they were a year or two, two years ago. So now you can actually rely on them to do more complex things that you as a, as a software engineer want to accomplish. And you can sort of delegate, you know, pretty complex things to these tools. And I think one really nice aspect about the, uh, interaction between, uh, uh, human, uh, software engineer and, uh, uh, coding model that they’re working with is your way of talking to that, uh, coding model actually sort of, uh, dictates how it interacts with you, right? Like you could ask it, please write a bunch of good tests for this. You could ask it, please help me brainstorm. Performance ideas and your way of doing that is going to shape how the model responds, what kinds of problems it tackles, you know, how much do you want the model to go off and do things that are larger and more independent versus interact with it, uh, more to make sure that you’re shaping the right kinds of, of things. And I think it’s not the case that any one style is the right thing for everything, right? Like some kinds of problems you actually want, uh, maybe a more frequent interaction style with a model. And other ones, you’re just like, yeah, please just go write this because I, I know I need this thing. I can specify it well enough, um, and go off and do it and come back when you’re done. And so I do think there’s going to be more of a style of having lots of independent, uh, software agents off doing things on your behalf and figuring out the right sort of human computer interaction model and UI and so on for when should it interrupt you and say, Hey, I need a little more guidance here, or I’ve done this thing. Now what, now what should I do? Um, I think we, we’re not at the end all answer to that question. And as the models get better, that, uh, set of decisions you put into how the interaction should happen may, may change, right? Like if you, if you have a team of 50 interns, how would you manage that if they were people? And I think it’s not, do you want 50 interns? You might, if they’re really good, right?
Shawn Wang [01:12:23]: It’s a lot of management. But it’s a lot of, uh.
Jeff Dean [01:12:25]: Uh, yeah. I mean, I think that is probably within the realm of possibilities that lots of people could have 50 interns. Yeah. And so how would you actually deal with that as a person, right? Like you would probably want them to form small sub teams, so you don’t have to interact with 50 of them. You can interact with five, five of those teams and they’re off doing things on your behalf, but I don’t know exactly what the, how this is going to unfold.
Alessio Fanelli [01:12:52]: Hmm. Yeah. How do you think about bringing people? Like the pair programming is always helpful to like get net new ideas in the distribution, so to speak. It feels as we have more of these coding agents, write the code, it’s hard to bring other people into the problem. So you go to like, you know, you have 50 interns, right? And then you want to go to Noam Shazier be like, Hey, no, I’m, I want to like pair on this thing. But now there’s like this huge amount of work that has been done in parallel that you need to catch him up on. Right. And I’m curious, like if people are going to be in a way more isolated in their teams, where it’s. It’s like, okay, there’s so much context in these 50 interns that it’s just hard for me to like relay everything back to you.
Jeff Dean [01:13:33]: Maybe. I mean, on the other hand, like imagine a classical software organization without any AI assisted tools, right. You would have, you know, 50 people doing stuff and their interaction style is going to be naturally very hierarchical because, um, you know, these 50 people are going to be working on this part of the system and not. Not interact that much with these other people over here. But if you have, you know, five people each managing 50 virtual agents, you know, they might be able to actually have much higher bandwidth communication among the five people, uh, then you would have among five people who are also trying to coordinate, you know, a 50 person software team. Each.
Alessio Fanelli [01:14:15]: So how, how do you, I’m curious how you change your just working rhythm, you know, like you spend more time ahead with people going through SPACs and design. Goals. Like,
Jeff Dean [01:14:26]: um, I mean, I do think it’s interesting that, you know, whenever people were taught how to write software, they were taught that it’s really important to write specifications super clearly, but no one really believed that. Like it was like, yeah, whatever. I don’t need to do that. I’m going to really, I don’t know. I mean, writing the English language specification was never kind of an artifact that was really paid a lot of attention to. I mean, it was important, but it wasn’t sort of the thing. That drove the actual creative process quite as much as if you specify what software you want the agent to write for you, you’d better be pretty darn careful of and how you specify that because that’s going to dictate the quality of the output, right? Like if you, if you don’t cover that it needs to handle this kind of thing, or that this is a super important corner case, or that, you know, you really care about the performance of this part of it, you know, it may, uh, not do what you want. Yeah. And the better you get at interacting with these models. And I think one of the ways people will get better is they will get really good at crisply specifying things rather than leaving things to ambiguity. And that is actually probably not a bad thing. It’s not a bad skill to have, regardless of whether you’re a software engineer or a, you know, trying to do some other kind of, uh, task, you know, being able to crisply specify what it is you want. It’s going to be really important. Yeah.
Shawn Wang [01:15:52]: My, my joke is, um, you know, good. Yeah. I think one thing is in, uh, indistinguishable from sufficiently advanced executive communication, like it’s like writing an internal memo, like weigh your words very carefully and also I think very important to be multimodal, right? I think, uh, one thing that, uh, anti-gravity from, from Google also did was like, just come out the gate to very, very strong multimodal, including videos, and that’s the highest bandwidth communication prompt that you can give to the model, which is fantastic. Yeah.
Alessio Fanelli [01:16:20]: How do you collect things that you often you will have in your mind? So you have this amazing, like performance sense thing that you’ve heard about how to look for performance improvements. And is there a lot more value in like people writing these like generic things down so that they can then put them back as like potential retrieval artifacts for the model? Like, or do I have like the edge cases is like a good example, right? It’s like, if you’re building systems, you already have in your mind, specific edge cases, depending on it. But now you have to like, every time repeat it. Like, are you having people spend a lot more time writing? Are you finding out more generic things to bring back?
Jeff Dean [01:16:56]: Or, um, I mean, I do think well-written guides of, of how to do good software engineering are going to be useful because they can be used as input to models or, you know, read by other developers so that their prompts are, you know, more clear about what the, the underlying software system should, should be doing. Um, you know, I think it may not be that you need to create a custom one. For every situation, if you have general guides and put those into, you know, the context of a coding agent, that, that can be helpful. Like in, you can imagine one for distributed systems, you could say, okay, think about failures of these kinds of things. And these are some techniques you can deal with failures. You know, you can have, uh, you know, Paxos like replication, or, you know, you can, uh, send the request to two places and tolerate failure because you only need one of them to come back. You know, a little. Description of 20 techniques like that in building distributed systems, probably would go a long way to having a coding agent be able to sort of cobble up more reliable and robust distributed systems.
Shawn Wang [01:18:07]: Yeah. Yeah. I wonder when Gemini will be able to build Spanner, right?
Alessio Fanelli [01:18:12]: Probably already has the code inside, you know?
Alessio Fanelli [01:18:16]: Yeah. That, I mean, that’s a good example, right? When you have like, you know, the cap theorem and it’s like, well, this is like truth and you cannot break that. And then you build something that broke it.
Shawn Wang [01:18:26]: Like, I’m curious, like models in a way are like, would he say he broke it? Did you, would you say you broke cap theorem? Really? Yeah. Okay. All right. I mean, under local assumptions. Yeah. Under some, some, yeah. And they’re like, you know, good clocks. Yeah. Yeah.
Alessio Fanelli [01:18:41]: It’s like some, sometimes you don’t have to like always follow what is known to be true. Right. And I, I think models in a way, like if you tell them something, they’re like really buy into that, you know? Um, yeah. So yeah, just more. Thinking than any answer on how to fix it.
Jeff Dean [01:18:57]: Yeah, my, my, uh, you know, it’s just on this, like, like big prompting and, and, uh, iteration, you know, I think that coming back to your latency point, um, I always, I always try to one, one AB test or experiment or benchmark or research I would like is what is the performance difference between, let’s say three dumb fast model calls with human alignment because the human will correct human alignment, being human looks at the first one and produces a new prompt.
Shawn Wang [01:19:23]: For the second one. Correct. Okay. As opposed to like, you spec it out, you know, it’s been a long time writing as a pro a big, big fat prompt, and then you have a very smart model. Do it right. Right. You know, cause, uh, really is, is, uh, our lacks in performance, uh, an issue of like, well, you just haven’t specified well enough. There’s no universe in which I can produce what you want because you just haven’t told me. Right.
Jeff Dean [01:19:44]: It’s underspecified. So I could produce 10 different things and only one of them is the thing you wanted. Yeah.
Shawn Wang [01:19:49]: And the multi-turn taking with a flash model is enough. Yeah.
Jeff Dean [01:19:54]: Yeah, I’m, I’m a big believer in pushing on latency because I think being able to have really low latency interactions with a system you’re using is just much more delightful than something that is, you know, 10 times as slow or 20 times as slow. And I think, you know, in the future we’ll see models that are, and, and underlying software and hardware systems that are 20X lower latency than what we have today, 50X lower latency. And that’s going to be really, really important for systems. That need to do a lot of stuff, uh, between your interactions.
Shawn Wang [01:20:27]: Yeah. Yeah. There, there’s two extremes, right? And then meanwhile, you also have DeepThink, which is all the way on the other side. Right.
Jeff Dean [01:20:33]: But you would use DeepThink all the time if it weren’t for cost and latency, right? If, if you could have that capability in a model because the latency improvement was 20X, uh, in the underlying hardware and system and costs, you know, there’s no reason you wouldn’t want that.
Shawn Wang [01:20:50]: Yeah.
Jeff Dean [01:20:52]: But at the same time, then you’d probably have a model. That is even better. That would take you 20X longer, even on that new hardware. Yeah.
Shawn Wang [01:21:00]: Uh, you know, there, there’s, uh, the Pareto curve keeps climbing. Um, yeah, onward and outward, onward and outward. Yeah. Should we ask him for predictions to, to go? I don’t know if you have any predictions that you, that you like to keep, you know, like, uh, one, one way to do this is you have your tests whenever a new model comes out that you run, uh, what’s something that you’re, you’re not quite happy with yet. That you think we’ll get done soon.
Jeff Dean [01:21:29]: Um, let me make two predictions that are not quite in that vein. Yeah. So I think a personalized model that knows you and knows all your state and is able to retrieve over all state you have access to, that you opt into is going to be incredibly useful compared to a more generic model that doesn’t have access to that. So like, can something attend to everything I’ve ever seen? Yeah. Every email, every photo, every. Yeah. Video I’ve watched, that’s going to be really useful. Uh, I think, uh, more and more specialized hardware is going to enable much lower latency models and much more capable models for affordable prices, uh, than say the current, current status quo. Uh, that’s going to be also quite important. Yeah.
Shawn Wang [01:22:16]: When you say much lower latency, uh, people usually talk in tokens per second. Is that a term that is okay? Okay. Uh, you know, we’re at, let’s say a hundred. Now we can go to a thousand. Is it meaningful to go 10,000? Yes. Really? Okay. Absolutely. Right. Yeah. Because of chain of thought and chain of thought reasoning.
Jeff Dean [01:22:36]: I mean, you could think, you know, uh, many more tokens, you could do many more parallel rollouts. You could generate way more code, uh, and check that the code is cracked with a chain of thought reasoning. So I think, you know, being able to do that at 10,000 tokens per second would be awesome. Yeah.
Shawn Wang [01:22:52]: At 10,000 tokens per second, you are no longer reading code. Yeah. Like you will just generate it. You’ll, I’m not reading it.
Jeff Dean [01:22:58]: Well, remember, it may not, it may not end up with 10,000 tokens of code. Yeah. It may be a thousand tokens of code that with 9,000 tokens of reasoning behind it, which would actually be probably much better code to read. Yeah.
Alessio Fanelli [01:23:11]: Yeah. If I had more time, I would have written a shorter letter. Yeah. Yeah. Yeah. Um, awesome. Jeff, this was amazing. Thanks for taking the time. Thank you.
Jeff Dean [01:23:20]: It’s been fun. Thanks for having me.


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