If we crack fusion and solar energy, we enter an era of radical abundance where resource scarcity ceases to be a driver of human conflict.

(00:00:00) The following is a conversation with Demis Hassabis, his second time on the podcast. (00:00:06) He is the leader of Google DeepMind and is now a Nobel Prize winner. (00:00:12) Demis is one of the most brilliant and fascinating minds in the world today, working on understanding and building intelligence, and exploring the big mysteries of our universe. (00:00:26) This was truly an honor and a pleasure for me. (00:00:30) And now a quick few second mention of each sponsor. (00:00:32) Check them out in the description or at lexfriedman.com slash sponsors. (00:00:36) It's the best way to support this podcast. (00:00:39) We've got Hampton for connecting with founders and CEOs, Finn for AI customer service, Shopify for building e-commerce businesses, Element for daily electrolytes, and AG1 for your health. (00:00:53) Choose wisely, my friends. (00:00:54) And now on to the full ad reads. (00:00:56) I do try to make them interesting, but if you must skip, friends, (00:00:59) Please still check out our sponsors. (00:01:01) I enjoy their stuff. (00:01:01) Maybe you will too. (00:01:02) And also to get in touch with me for whatever reason, go to alexfreeman.com slash contact. (00:01:09) All right, let's go. (00:01:11) This episode is brought to you by Hampton, a private community for high growth founders and CEOs. (00:01:18) That's the interesting thing about starting a company and running a company, especially one that's growing really quickly, has to hire a lot, has to scale a lot. (00:01:27) It's perhaps a little bit counterintuitive, but for the founder, it can be deeply lonely. (00:01:31) I suppose that's one of the reasons they recommend to have a co-founder. (00:01:34) But even outside of that, there's just a deep loneliness with putting it all on the line, risking everything, knowing that the chances of success are low. (00:01:45) But if you do succeed, the gains are huge. (00:01:47) And you have your heart in it, you have your dreams in it, you believe in it. (00:01:52) but also there's a constant roller coaster of fear and doubt and hope and moments of triumph and moments of failure. (00:02:01) All those go back and forth and just it's a constant psychological turmoil. (00:02:05) Anyway, through all that, it's just nice to connect with other people that are going through the same thing. (00:02:10) And that's what Hampton is about. (00:02:12) It does a thing where every month, 8 founders face to face have real conversations about their daily struggles. (00:02:19) Groups are forming in a bunch of places, New York City, Austin, San Francisco, LA, Miami, Denver, and so on. (00:02:24) If you are a founder who's tired of carrying it all alone, visit joinhampton.com slash Lex to see if it's a fit for you. (00:02:34) That's joinhampton.com slash Lex. (00:02:39) This episode is also brought to you by Fin. (00:02:41) It's an AI agent for customer service. (00:02:44) So they are focused, laser focused on the customer service application and they want to do that better than anybody else in the world. (00:02:51) In fact, if you measure by the metric of the number of resolutions, so when you have the agent resolve the customer service issue, that's resolution, they have a 59% average resolution rate, which makes it the highest performing customer service agent on the market. (00:03:08) It's trusted by over 5,000 customer service leaders and even top AI companies, including Anthropic. (00:03:15) The way they design the system is it can continuously improve from the interactions, so you can continuously analyze, train, test, and deploy. (00:03:25) Also, probably important to say, they give you a 90-day money-back guarantee. (00:03:31) Go to fin.ai/lex to learn more about transforming your customer service and scaling your support team. (00:03:37) That's Finn.ai slash Lex. (00:03:41) This episode is also brought to you by Shopify, a platform designed for anyone to sell anywhere with a great looking online store. (00:03:47) Even I figured out how to create an online store at lexfreeman.com slash store. (00:03:53) I put up a few shirts. (00:03:55) Haven't done anything with it since because I'm not a serious person. (00:03:58) There's a lot of serious people that build real businesses on top of Shopify. (00:04:03) It's a platform that connects you. (00:04:06) with millions of people that want to buy stuff and gives you all the tools you need and all the integrations you need to do just that. (00:04:12) At scale, as we talked about with DHH, about the incredible beauty and power of Ruby on Rails that Shopify is powered by. (00:04:23) I have not yet built a serious sort of medium-scale project on Rails. (00:04:29) I need to. (00:04:30) It's just I need to actually find things that I need to do web dev type of stuff with to inspire myself to build something useful. (00:04:39) I don't want to build some weird variant of a to-do list, especially now with the help of LLMs, you can generate so much of the code. (00:04:45) So I need to figure out how to learn (00:04:50) a new framework, a new programming languages when LLMs can generate so much of it. (00:04:54) And I don't want to do it exclusively by vibe coding because I feel like that's not a way to learn fully a thing, but vibe coding does remove some of the friction of the learning. (00:05:05) So balancing that out is a tricky thing to do. (00:05:08) Anyway, that's about the programming language and the framework that powers Shopify. (00:05:13) But Shopify itself, (00:05:14) Connects buyers and sellers in an incredible scale that's awe-inspiring. (00:05:18) Sign up for a $1 per month trial period at shopify.com slash Lex. (00:05:23) That's all lowercase. (00:05:24) Go to shopify.com slash Lex to take your business to the next level today. (00:05:30) This episode is also brought to you by Element, my daily zero sugar and delicious electrolyte mix. (00:05:37) I've been traveling recently and I have a lot of Element packets with me and (00:05:43) I bring that and I bring bands, whatever you call them. (00:05:47) I don't know what they're called. (00:05:48) They're like rubber bands for like basic shoulder exercises. (00:05:52) So if I have to do a lot of either heavy lifting or heavy jiu-jitsu training, I like to warm up the shoulders really well because probably because I have issues with shoulders from many years of playing tennis and many years of doing stupidly. (00:06:05) Bench press. (00:06:06) Anyway, I think of element as a critical component of my workout routine. (00:06:10) Hydrate before, rehydrate after. (00:06:12) Fully embrace the deliciousness of watermelon salt flavor. (00:06:16) The flavor of champions, the one I recommend. (00:06:18) It's been quite a while since I tried the others. (00:06:20) They're all good, but for me, I'm a man of focus and dedication. (00:06:25) And I'm dedicated to watermelon salt. (00:06:28) I think they have actually, I saw a lemonade flavor. (00:06:30) I think a lot of people love lemonade. (00:06:33) So maybe that's your thing. (00:06:34) For me, I'm sticking to watermelon salt. (00:06:36) Get a free 8 calm sample pack with any purchase. (00:06:39) Try it at drinkelement.com slash flex. (00:06:44) This episode is also brought to you by AG1, an all-in-one daily drink to support better health and peak performance. (00:06:51) I travel with it. (00:06:52) makes me feel like I take a little piece of home with me. (00:06:55) I drink it at least once a day, very often twice a day. (00:06:59) And they keep innovating, they keep improving it. (00:07:02) They recently introduced the AG1 Next Gen. (00:07:05) improving every aspect, more vitamins and minerals, and upgraded probiotics. (00:07:09) It's funny how a morning routine can be the source of peace and happiness. (00:07:15) Because I find that if I check my phone at all in the first couple hours of the day, I get this weird anxiety that ultimately morphs into unhappiness. (00:07:27) And if I don't, I'm much more likely to sort of maintain that deep focus. (00:07:31) And a part of that early in the morning is some coffee or caffeinated drink, and then a few hours on is AG1. (00:07:38) And it's just many hours of deep focus in between. (00:07:42) It makes me feel happy, makes me feel at one with the universe, and it helps me get shit done. (00:07:49) Anyway, they'll give you one month's supply of fish oil when you sign up at drinkag1.com slash Lex. (00:07:57) This is the Lex Friedman Podcast. (00:07:59) To support it, please check out our sponsors in the description or at lexfriedman.com slash sponsors. (00:08:04) And consider subscribing, commenting, and sharing the podcast with folks who might find it interesting. (00:08:10) I promise to work extremely hard to always bring you nuanced and long-form conversations with a wide variety of interesting people from all walks of life. (00:08:20) And now, dear friends, here's Demis Hazalbas. (00:08:41) In your Nobel Prize lecture, you proposed what I think is a super interesting conjecture that, quote, any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm. (00:08:55) What kind of patterns of systems might be included in that? (00:09:00) Biology, chemistry, physics, maybe cosmology, neuroscience? (00:09:05) What are we talking about? (00:09:06) Sure. (00:09:07) look, I felt that it's sort of a tradition, I think, of Nobel Prize lectures that you're supposed to be a little bit provocative, and I wanted to follow that tradition. (00:09:14) What I was talking about there is if you take a step back and you look at all the work that we've done, especially with the AlphaX projects, so I'm thinking AlphaGo, of course, AlphaFold, what they really are is we're building models of very combinatorially high dimensional spaces that, you know, if you try to brute force a solution, find the best moving go, or find the exact shape of a protein, (00:09:37) And if you enumerated all the possibilities, there wouldn't be enough time in the, time of the universe. (00:09:42) So you have to do something much smarter. (00:09:44) And what we did in both cases was build models of those environments. (00:09:50) And that guided the search in a smart way. (00:09:53) And that makes it tractable. (00:09:54) So if you think about protein folding, which is obviously a natural system, you know, why should that be possible? (00:09:59) How does physics do that? (00:10:01) You know, proteins fold in milliseconds in our bodies. (00:10:03) So somehow physics solves this problem that we've now also solved computationally. (00:10:08) And I think the reason that's possible is that in nature, natural systems have structure because they were subject to evolutionary processes that shaped them. (00:10:19) And if that's true, (00:10:20) then you can maybe learn what that structure is. (00:10:24) This perspective, I think, is a really interesting one. (00:10:27) You've hinted at it, which is almost like crudely stated, anything that can be evolved can be efficiently modeled. (00:10:36) Do you think there's some truth to that? (00:10:37) Yeah, I sometimes call it survival of the stabilist or something like that, because, it's, of course, there's evolution for life, living things, but there's also, if you think about geological time, so the shape of mountains, that's been shaped by weathering processes, right, over thousands of years. (00:10:55) But then you can even take it cosmological, the orbits of planets, the shapes of asteroids, these have all been survived kind of processes that have acted on them many, many times. (00:11:05) So if that's true, (00:11:06) true, then there should be some sort of pattern that you can kind of reverse learn and a kind of manifold, really, that helps you search to the right solution, to the right shape, and actually allow you to predict things about it in an efficient way, because it's not a random pattern, right? (00:11:23) So it may not be possible for man-made things or abstract things like factorizing large numbers, because unless there's patterns in the number space, which there might be, but if there's not and it's uniform, then there's no pattern to learn. (00:11:36) There's (00:11:36) There's no model to learn that will help you search. (00:11:38) You have to do brute force. (00:11:39) So in that case, you maybe need a quantum computer, something like this. (00:11:43) But in most things in nature that we're interested in are not like that. (00:11:46) They have structure that evolved for a reason and survived over time. (00:11:52) And if that's true, I think that's potentially learnable by a neural network. (00:11:56) It's like nature is doing a search process, and it's so fascinating that it's in that search process, it's creating systems that can be efficiently modeled. (00:12:06) That's right. (00:12:07) Yeah. (00:12:07) So interesting. (00:12:08) So they can be efficiently rediscovered or recovered because nature's not random, right? (00:12:13) Everything that we see around us, including like the elements that are more stable, all of those things, they're subject to some kind of selection process, pressure. (00:12:22) Do you think, because you're also a fan of theoretical computer science and complexity, do you think we can come up with a kind of complexity class, like a complexity zoo type of class where maybe it's the set of learnable systems (00:12:35) The set of learnable natural systems (LNS). (00:12:40) This is a new class of systems that could be actually learnable by classical systems in this kind of way, natural systems that can be modeled efficiently. (00:12:52) I mean, I've always been fascinated by the P equals MP question and what is modelable by classical systems, i.e. (00:12:59) non-quantum systems, you know, Turing machines in effect. (00:13:03) And that's exactly what I'm working on actually in kind of my few moments of spare time with a few colleagues about is should there be, you know, maybe a new class of problem that is solvable by this type of neural network process and kind of mapped onto these natural systems. (00:13:18) So the things that exist in physics and have structure. (00:13:22) So I think that could be a very interesting new way of thinking about it. (00:13:26) And it sort of fits with the way I think about physics in general, which is that I think information is primary. (00:13:31) Information is the most sort of fundamental unit of the universe, more fundamental than energy and matter. (00:13:36) I think they can all be converted into each other. (00:13:38) But I think of the universe as a kind of informational system. (00:13:41) So when you think of the universe as an informational system, then the P equals NP question is a... (00:13:47) It's a physics question. (00:13:49) That's right. (00:13:49) And it's a question that can help us actually solve the entirety of this whole thing going on. (00:13:54) Yeah, I think it's one of the most fundamental questions, actually, if you think of physics as informational. (00:14:00) And the answer to that, I think, is going to be very enlightening. (00:14:04) More specific to the PNP question, again, some of the stuff we're saying is kind of crazy right now. (00:14:11) just like the Christian Anthonson Nobel Prize speech controversial thing that he said sounded crazy, and then you went and got a Nobel Prize for this with John Jumper, solved the problem. (00:14:21) So let me just stick to the P equals MP. (00:14:24) Do you think there's something in this thing we're talking about that could be shown if you can do something like polynomial time or constant time compute ahead of time and construct this gigantic model, then (00:14:41) you can solve some of these extremely difficult problems in a theoretical computer science kind of way. (00:14:46) Yeah, I think that there are actually a huge class of problems that could be couched in this way, the way we did AlphaGo and the way we did AlphaFold, where you model what the dynamics of the system is, the properties of that system, the environment that you're trying to understand. (00:15:03) And then that makes the search for the solution or the prediction of the next step efficient, basically polynomial time. (00:15:11) so tractable by a classical system, which a neural network is. (00:15:16) It runs on normal computers, right? (00:15:18) Classical computers, Turing machines in effect. (00:15:21) And I think it's one of the most interesting questions there is, is how far can that paradigm go? (00:15:27) You know, I think we've proven and the AI community in general that classical systems, Turing machines can go a lot further than we previously thought. (00:15:35) You know, they can do things like model the structures of proteins and play go to better than world champions. (00:15:41) level and a lot of people would have thought maybe 10, 20 years ago that was decades away or maybe you would need some sort of quantum machines to quantum systems to be able to do things like protein folding. (00:15:54) And so I think we haven't really (00:15:57) even sort of scratched the surface yet of what classical systems, so-called, could do. (00:16:03) And of course, AGI being built on a neural network system, on top of a neural network system, on top of a classical computer would be the ultimate expression of that. (00:16:12) And I think the limit, you know, what the bounds of that kind of system, what it can do, it's a very interesting question and directly speaks to the P equals NP question. (00:16:22) What do you think, again, hypothetical, might be outside of this? (00:16:26) maybe emergent phenomena, like if you look at cellular automata, some of the, you have extremely simple systems and then some complexity emerges. (00:16:35) Maybe that would be outside or even, would you guess even that might be amenable to efficient modeling by a classical machine? (00:16:44) Yeah, I think those systems would be right on the boundary, right? (00:16:47) So (00:16:48) I think most emergent systems, cellular automata, things like that, could be modelable by a classical system. (00:16:53) You just sort of do a forward simulation of it and it'd probably be efficient enough. (00:16:58) Of course, there's the question of things like chaotic systems where the initial conditions really matter and then you get to some uncorrelated end state. (00:17:07) Now, those could be difficult to model. (00:17:09) So I think these are kind of the open questions. (00:17:11) But I think when you step back and look at what we've done with the systems and the problems that we've solved, and then you look at things like (00:17:18) VO3 on like video generation, sort of rendering physics and lighting and things like that, you know, really in core fundamental things in physics. (00:17:29) It's pretty interesting. (00:17:30) I think it's telling us something quite fundamental about how the universe is structured, in my opinion. (00:17:35) So, you know, in a way, that's what I want to build AGI for is to help us as scientists answer these questions like P equals MP. (00:17:44) Yeah, I think we might be continuously surprised about what is modelable by (00:17:48) classical computers, I mean, AlphaFold3 on the interaction side is surprising that you can make any kind of progress on that direction. (00:17:57) AlphaGenome is surprising that you can map the genetic code to the function. (00:18:03) kind of playing with the emergent kind of phenomena, you think there's so many combinatorial options, and then here you go, you can find the kernel that is efficiently modeled. (00:18:11) Yes, because there's some structure, there's some landscape, you know, in the energy landscape or whatever it is that you can follow, some gradient you can follow, and of course what neural networks are very good at is following gradients. (00:18:22) And so if there's one to follow and you can specify the objective function correctly, you don't have to deal with all that complexity, which I think is how we maybe have naively thought about it for decades, those problems. (00:18:35) If you just enumerate all the possibilities, it looks totally intractable. (00:18:39) And there's many, many problems like that. (00:18:40) And then you think, well, it's like 10 to the 300 possible protein structures. (00:18:45) 10 to the 170 possible go positions. (00:18:49) All of these are way more than atoms in the universe. (00:18:51) So how could one possibly find the right solution or predict the next step? (00:18:56) But it turns out that it is possible. (00:18:59) And of course, reality in nature does do it, right? (00:19:01) Proteins do fold. (00:19:03) So that gives you confidence that there must be, if we understood how physics was doing that, in a sense, (00:19:09) then, and we could mimic that process, i.e. (00:19:12) model that process, it should be possible on our classical systems is basically what the conjecture is about. (00:19:18) And of course, there's nonlinear dynamical systems, highly nonlinear dynamical systems, everything involving fluid. (00:19:25) Yes, right. (00:19:26) You know, I recently had a conversation with Terence Tao, who mathematically contends with a very difficult aspect of systems that have some singularities in them that break the mathematics. (00:19:38) And it's just hard for us humans to make any kind of clean predictions about highly non-linear dynamical systems. (00:19:44) But again, to your point, we might be very surprised what classical learning systems might be able to do about even fluid. (00:19:51) Yes, exactly. (00:19:52) I mean, fluid dynamics, Navier-Stokes equations, these are traditionally thought of as very, very difficult, intractable kind of problems to do on classical systems. (00:20:00) They take enormous amounts of compute, you know, weather prediction systems, you know, these kind of things all involve fluid dynamics calculations. (00:20:07) And (00:20:08) But again, if you look at something like VO, our video generation model, it can model liquids quite well, surprisingly well, and materials, specular lighting. (00:20:18) I love the ones where, you know, there's people who generate videos where there's like clear liquids going through hydraulic presses and then it's being squeezed out. (00:20:26) I used to write (00:20:28) physics engines and graphics engines in my early days in gaming. (00:20:31) And I know it's just so painstakingly hard to build programs that can do that. (00:20:36) And yet somehow these systems are reverse engineering from just watching YouTube videos. (00:20:42) So presumably what's happening is it's extracting some underlying structure around how these materials behave. (00:20:50) So perhaps there is some kind of lower dimensional manifold that can be learned if we actually fully understood what's going on under the hood. (00:20:58) That's maybe (00:20:58) be, maybe true of most of reality. (00:21:01) I've been continuously, precisely by this aspect of VO3. (00:21:05) I think a lot of people highlight different aspects, including the comedic and the meme and all that kind of stuff. (00:21:10) And then the ultra realistic ability to capture humans in a really nice way that's compelling and feels close to reality. (00:21:19) And then combine that with native audio. (00:21:20) All of those are marvelous things about VO3, but the exactly the thing you're mentioning, which is the physics. (00:21:27) Yeah. (00:21:27) It's not perfect, but it's pretty damn good. (00:21:30) And then the really interesting scientific question is, what is it understanding about our world in order to be able to do that? (00:21:38) Because of the cynical take with diffusion models, there's no way it understands anything. (00:21:45) But it seems, I mean, I don't think you can generate that kind of video without understanding, and then our own philosophical notion of what it means to understand then is brought to the surface. (00:21:56) To what degree do you think VO3 understands our world? (00:21:59) I think to the extent that it can predict the next frames in a coherent way. (00:22:05) That is a form, of understanding, right? (00:22:08) Not in the anthropomorphic version of, it's not some kind of deep philosophical understanding of what's going on. (00:22:13) I don't think these systems have that, but they certainly have modeled enough of the dynamics. (00:22:19) put it that way, that they can pretty accurately generate whatever it is, 8 seconds of consistent video that by eye, at least at a glance, it's quite hard to distinguish what the issues are. (00:22:31) And imagine that in two or three more years time, that's the thing I'm thinking about and how incredible that they will look, given where we've come from, you know, the early versions of that one or two years ago. (00:22:42) And so the rate of progress is incredible. (00:22:46) And I think (00:22:48) I'm like you is like a lot of people love all of the stand-up comedians and that actually captures a lot of human dynamics very well and body language. (00:22:56) But actually, the thing I'm most impressed with and fascinated by is the physics behavior, the lighting and materials and liquids. (00:23:04) And it's pretty amazing that it can do that. (00:23:07) And I think that shows that it has some notion of at least intuitive physics, right? (00:23:14) How things are supposed to work intuitively, maybe the way (00:23:17) that a human child would understand physics, right? (00:23:20) As opposed to a PhD student really being able to unpack all the equations. (00:23:25) It's more of an intuitive physics understanding. (00:23:28) Well, that intuitive physics understanding, that's the base layer. (00:23:33) That's the thing people sometimes call a common sense. (00:23:36) It really understands something that I think that really surprised a lot of people. (00:23:39) It blows my mind that (00:23:41) I just didn't think it would be possible to generate that level of realism without understanding. (00:23:46) There's this notion that you can only understand the physical world by having an embodied AI system, a robot that interacts with that world. (00:23:54) That's the only way to construct an understanding of that world. (00:23:58) But VO3 is directly challenging that, it feels like. (00:24:03) And this is very interesting, even if you were to ask me 5, 10 years ago, I would have said, even though I was a must in all of this, I would have said, you probably need to understand intuitive physics, like if I push this off the table, this glass, it will maybe shatter, and the liquid will spill out, right? (00:24:19) So we know all of these things. (00:24:21) But I thought that, and there's a lot of theories in neuroscience, it's called action in perception, where, you need to act in the world to really truly perceive it in a deep way. (00:24:30) And there was a lot of theories about you'd need embodied intelligence or robotics or something, or maybe at least simulated action so that you would understand things like intuitive physics. (00:24:41) But it seems like (00:24:42) you can understand it through passive observation, which is pretty surprising to me. (00:24:46) And again, I think hints at something underlying about the nature of reality, in my opinion, beyond just the cool videos that it generates. (00:24:57) And of course, there's next stages is maybe even making those videos interactive. (00:25:01) So one can actually step into them and move around them, which would be really mind-blowing, especially given my games background. (00:25:09) So you can imagine. (00:25:11) And then I think, you know, we're starting to get towards what I would call a world model, a model of how the world works, the mechanics of the world, the physics of the world, and the things in that world. (00:25:21) And of course, that's what you would need for a true AGI system. (00:25:24) I have to talk to you about video games. (00:25:26) So you were being a bit trolly. (00:25:29) I think you're having more and more fun on Twitter on X, which is great to see. (00:25:33) So a guy named Jimmy Apples tweeted, let me play a video game of my VO3 videos already. (00:25:39) Google cooked so good, playable world models when, spelled W-E-N question mark. (00:25:45) And then you quote tweeted that with, now wouldn't that be something? (00:25:49) So how hard is it to build game worlds with AI? (00:25:53) Maybe can you look out into the future of video games 5, 10 years out? (00:25:59) What do you think that looks like? (00:26:01) Well, games were my first love really, and doing AI for games was the first thing I did professionally in my teenage years and was the first major AI systems that I built. (00:26:12) And I always want to scratch that itch one day and come back to that. (00:26:17) So, and I will do, I think. (00:26:19) And I think I'd sort of dream about, what would I have done back in the 90s if I'd had access to the kind of AI systems we have today? (00:26:26) And I think you could build absolutely mind-blowing games. (00:26:30) And I think the next stage is I always used to love making all the games I've made are open world games. (00:26:35) So they're games where there's a simulation and then there's AI characters and then the player (00:26:41) interacts with that simulation, and the simulation adapts to the way the player plays. (00:26:45) And I always thought they were the coolest games because, so games like Theme Park that I worked on where everybody's game experience would be unique to them, right? (00:26:53) Because you're kind of co-creating the game, right? (00:26:56) We set up the parameters, we set up initial conditions, and then you as the player immersed in it, and then you are co-creating it with the simulation. (00:27:05) But of course, it's very hard to program open-world games. (00:27:08) You know, you've got to be able to create content whichever direction the (00:27:11) player goes in and you want it to be compelling no matter what the player chooses. (00:27:16) And so it was always quite difficult to build things like cellular automata, actually, type of those kind of classical systems which created some emergent behavior. (00:27:24) But they're always a little bit fragile, a little bit limited. (00:27:27) Now we're maybe on the cusp in the next few years, 5, 10 years of having AI systems that can truly create around your imagination. (00:27:35) can now sort of dynamically change the story and storytell the narrative around and make it dramatic no matter what you end up choosing. (00:27:43) So it's like the ultimate choose your own adventure sort of game. (00:27:46) And, you know, I think maybe we're within reach if you think of a kind of interactive version of VO and then wind that forward 5 to 10 years and, you know, imagine how good it's going to be. (00:27:58) Yeah, so you said a lot of super interesting stuff there. (00:28:00) So one, the open world, (00:28:03) built into that is a deep personalization, the way you've described it. (00:28:07) So it's not just that it's open world, that you can open any door and there'll be something there. (00:28:12) It's that the choice of which door you open in an unconstrained way defines the worlds you see. (00:28:19) So some games try to do that to give you choice, but it's really just an illusion of choice because you only... (00:28:28) like Stanley Parable. (00:28:29) This is a game I recently played. (00:28:31) It's really, there's a couple of doors and it really just takes you down a narrative. (00:28:35) Stanley Parable is a great video game I recommend people play that kind of, in a meta way, mocks the illusion of choice and there's philosophical notions of free will and so on. (00:28:46) But I do, like one of my favorite games of Elder Scrolls is Daggerfall, I believe. (00:28:53) that they really played with a random generation of the dungeons. (00:28:58) Yeah. (00:28:59) Of if you can step in and they give you this feeling of an open world. (00:29:03) And there, you mentioned interactivity, you don't need to interact. (00:29:07) That's a first step because you don't need to interact that much. (00:29:10) You just, when you open the door, whatever you see is randomly generated for you. (00:29:15) And that's already an incredible experience because you might be the only person to ever see that. (00:29:20) Yeah. (00:29:21) Exactly. (00:29:22) And so, but what you'd like is a little bit better than just sort of a random generation, right? (00:29:26) So you'd like, and also better than a simple AB hard coded choice, right? (00:29:32) That's not really open world, right? (00:29:35) As you say, it's just giving you the illusion of choice. (00:29:37) What you want to be able to do is potentially anything in that game environment. (00:29:42) And I think the only way you can do that is to have generated systems, systems that will generate that on the fly. (00:29:49) Of course, you can't create infinite amounts of game assets, right? (00:29:52) It's expensive enough already how AAA games are made today. (00:29:56) And that was obvious to us back in the 90s when I was working on all these games. (00:30:00) I think maybe black and white was the game that I worked on, early stages of that, had still probably the best AI, learning AI in it. (00:30:08) was an early reinforcement learning system that you were (00:30:12) you were looking after this mythical creature and growing it and nurturing it. (00:30:15) And depending how you treated it, would treat the villagers in that world in the same way. (00:30:20) So if you were mean to it, would be mean. (00:30:21) If you were good, it would be protective. (00:30:23) And so it was really a reflection of the way you played it. (00:30:26) So actually, I've been working on sort of simulations and AI through the medium of games at the beginning of my career. (00:30:34) And really, the whole of what I do today is still a follow-on from (00:30:38) those early, more hard-coded ways of doing the AI to now, fully general learning systems that are trying to achieve the same thing. (00:30:46) Yeah, it's been interesting, hilarious, and fun to watch you and Elon obviously itching to create games because you're both gamers. (00:30:55) And one of the sad aspects of your incredible success in so many domains of science, like serious adult stuff, that you might (00:31:04) not have time to really create a game, you might end up creating the tooling that others will create the game. (00:31:10) You have to watch others create the thing you've always dreamed of. (00:31:16) Do you think it's possible you can somehow in your extremely busy schedule actually find time to create something like black and white? (00:31:24) An actual video game where, like, you could make the childhood dream become a reality. (00:31:32) there's two things where to think about that is maybe with vibe coding as it gets better and there's a possibility that I could, you know, one could do that actually in your spare time. (00:31:40) So I'm quite excited about that as that would be my project if I got the time to do some vibe coding. (00:31:46) I'm actually itching to do that. (00:31:48) And then the other thing is, maybe it's a sabbatical after AGI has been safely stewarded into the world and delivered into the world. (00:31:55) that, and then working on my physics theory, as we talked about at the beginning, those would be the two, my 2 post-AGI projects. (00:32:02) Let's call it that way. (00:32:03) I would love to see which post-AGI, which you choose. (00:32:08) solving the problem that some of the smartest people in human history contended with. (00:32:13) So P equals MP or creating a cool video. (00:32:18) Yeah, but in my world, they'd be related because it would be an open world simulated game as realistic as possible. (00:32:25) So, you know, what is the universe? (00:32:28) That's speaking to the same question, right? (00:32:30) MP equals MP. (00:32:31) I think all these things are related, at least in my mind. (00:32:33) I mean, in a really serious way. (00:32:36) I think video games sometimes are looked down upon. (00:32:39) That's just this fun side activity. (00:32:41) But especially as AI does more and more of the difficult, boring tasks, something we in modern world called work, video games is the thing in which we may find meaning, in which we may find what to do with our time. (00:32:59) You could create incredibly rich, meaningful (00:33:03) experiences. (00:33:04) Like that's what human life is. (00:33:05) And then in video games, you can create more sophisticated, more diverse ways of living. (00:33:14) Yeah, I think so. (00:33:16) I mean, those of us who love games and I still do is, you know, it's almost can let your imagination run wild, right? (00:33:25) Like I used to love games. (00:33:28) and working on games so much because it's the fusion, especially in the 90s and early 2000s, the sort of golden era and maybe the 80s of the games industry. (00:33:37) And it was all being discovered, new genres were being discovered. (00:33:40) We weren't just making games, we felt we were creating a new entertainment medium that never existed before, especially with these open world games and simulation games where you were co-create, you as the player were co-creating the story. (00:33:51) There's no other media, entertainment media, where you do that, where you as the audience actually co-create the story. (00:33:58) And of course, now with multiplayer games as well, it can be a very social activity and can explore all kinds of interesting worlds in that. (00:34:06) But on the other hand, it's very important to also enjoy and experience the physical world. (00:34:13) But the question is then, I think we're going to have to confront the question again of what is the fundamental nature of reality. (00:34:20) what is going to be the difference between these increasingly realistic simulations and multiplayer ones and emergent and what we do in the real world. (00:34:30) Yeah, there's clearly a huge amount of value to experiencing the real world, nature. (00:34:35) There's also a huge amount of value in experiencing other humans directly in person, the way we're sitting here today. (00:34:42) But we need to really scientifically, rigorously answer the question, why? (00:34:47) And which aspect of that can be mapped into the virtual world? (00:34:51) Exactly. (00:34:52) It's not enough to say, yeah, you should go touch grass and hang out in nature. (00:34:56) It's like, why exactly is that valuable? (00:34:59) Yes. (00:35:00) And I guess that's maybe the thing that's been haunting me, obsessing me from the beginning of my career. (00:35:05) If you think about all the different things I've done, that's they're all related in that way. (00:35:08) The simulation, nature of reality, and what is the bounds of, you know, what can be modeled. (00:35:15) Sorry for the ridiculous question, but so far, what is the greatest video game of all time? (00:35:18) What's up there? (00:35:19) Well, my favorite one of all time is Civilization. (00:35:22) I have to say. (00:35:23) That was the Civilization 1 and Civilization 2, my favorite games of all time. (00:35:29) I can only assume you've avoided the most recent one because it would probably, you would, that would be your sabbatical. (00:35:36) That would, you would disappear. (00:35:37) Yes, exactly. (00:35:38) They take a lot of time, these Civilization games. (00:35:40) So I've got to be careful with them. (00:35:43) fun question. (00:35:43) You and Elon seem to be somehow solid gamers. (00:35:48) Is there a connection between being great at gaming and being great leaders of AI companies? (00:35:54) I don't know. (00:35:55) It's an interesting one. (00:35:56) I mean, we both love games and it's interesting he wrote games as well to start off with. (00:36:01) It's probably, especially in the era I grew up in where home computers were just became a thing, you know, in the late 80s and 90s, especially in the UK, I had a Spectrum and then a Commodore Amiga 500, which is my (00:36:13) favorite computer ever and that's why I learn all my programming. (00:36:16) And of course it's a very fun thing to program is to program games. (00:36:20) So I think it's a great way to learn programming, probably still is. (00:36:24) And (00:36:26) And then, of course, I immediately took it in directions of AI and simulations, which so I was able to express my interest in games and my sort of wider scientific interests altogether. (00:36:38) And then the final thing I think that's great about games is it fuses artistic design, you know, art with the most cutting edge programming. (00:36:49) So again, in the 90s, all of the most interesting technical advances were happening in gaming, whether that was AI, graphics, (00:36:56) physics engines, hardware, even GPUs, of course, were designed for gaming originally. (00:37:01) So everything that was pushing computing forward in the 90s was due to gaming. (00:37:07) So interestingly, that was where the forefront of research was going on. (00:37:11) And it was this incredible fusion with art. (00:37:15) graphics, but also music, and just the whole new media of storytelling. (00:37:20) And I love that. (00:37:21) For me, it's this sort of multidisciplinary kind of effort is, again, something I've enjoyed my whole life. (00:37:27) I have to ask you, I almost forgot about one of the many, and I would say one of the most incredible things recently that somehow didn't yet get enough attention is Alpha Evolve. (00:37:38) We talked about evolution a little bit, but it's the (00:37:41) Google DeepMind system that evolves algorithms. (00:37:44) Are these kinds of evolution-like techniques promising as a component in a future superintelligent system? (00:37:49) So for people who don't know, it's kind of, I don't know if it's fair to say it's LLM guided. (00:37:56) evolution search. (00:37:58) So evolutionary algorithms are doing the search, and LLMs are telling you where. (00:38:03) Yes, exactly. (00:38:04) So LLMs are kind of proposing some possible solutions, and then you use evolutionary computing on top to find some novel part of the search space. (00:38:14) So actually, I think it's an example of very promising directions where you combine LLMs, or foundation models, with other computational techniques. (00:38:24) Evolutionary methods is one, but you could (00:38:26) also imagine Monte Carlo tree search, basically many types of search algorithms or reasoning algorithms sort of on top of or using the foundation models as a basis. (00:38:36) So I actually think there's quite a lot of interesting things to be discovered probably with these sort of hybrid systems, let's call them. (00:38:44) But not to romanticize evolution. (00:38:46) Yeah. (00:38:47) I'm only human. (00:38:48) But you think there's some value in whatever that mechanism is? (00:38:51) Because we already talked about natural systems. (00:38:53) Do you think (00:38:54) where there's a lot of low-hanging fruit of us understanding being able to model, being able to simulate evolution, and then using that, whatever we understand about that nature inspired mechanism to then do surge better and better and better. (00:39:11) Yes, so if you think about, again, breaking down the solar systems we've built to their really fundamental core, you've got the model of the underlying dynamics of the system, (00:39:23) And then, if you want to discover something new, something novel that hasn't been seen before, then you need some kind of search process on top to take you to a novel region of the search space, and... (00:39:37) You can do that in a number of ways. (00:39:38) Evolutionary computing is one. (00:39:40) With AlphaGo, we just use Monte Carlo tree search, right? (00:39:44) And that's what found Move 37, the new kind of never seen before strategy in Go. (00:39:50) And so that's how you can go beyond potentially what is already known. (00:39:53) So the model can model everything that you currently know about, right? (00:39:56) All the data that you currently have. (00:39:58) But then how do you go beyond that? (00:40:00) starts to speak about the ideas of creativity. (00:40:02) How can these systems create something new, discover something new? (00:40:06) Obviously, this is super relevant for scientific discovery or pushing science and medicine forward, which we want to do with these systems. (00:40:12) And you can actually bolt on some... (00:40:16) fairly simple search systems on top of these models and get you into a new region of space. (00:40:22) Of course, you also have to make sure that you're not searching that space totally randomly. (00:40:27) It would be too big. (00:40:28) So you have to have some objective function that you're trying to optimize and hill climb towards and that guides that search. (00:40:34) But there's some mechanism of evolution that are interesting. (00:40:37) maybe in the space of programs, but then the space of programs is an extremely important space because you can probably generalize to everything. (00:40:45) But for example, mutation, this is not just Monte Carlo tree search where it's like a search. (00:40:54) You could every once in a while combine things, combine things, alter like sub like a components of a thing. (00:41:00) So then, you know, what evolution is really good at is not just the natural selection. (00:41:06) It's (00:41:07) combining things and building increasingly complex hierarchical systems. (00:41:12) So that component is super interesting, especially like with Alpha evolving the space of programs. (00:41:17) Yeah, exactly. (00:41:18) So you can get a bit of an extra property out of evolutionary systems, which is some new emergent capability may come about. (00:41:25) But of course, like happened with life. (00:41:28) Interestingly, with naive sort of traditional evolution computing methods without LLMs and the modern AI, the problem with them, they were very well studied in the (00:41:37) and early 2000s and some promising results. (00:41:41) But the problem was they could never work out how to evolve new properties, new emergent properties. (00:41:46) You always had a sort of subset of the properties that you put into the system. (00:41:49) But maybe if we combine them with these foundation models, perhaps we can overcome that limitation. (00:41:55) Obviously, natural evolution clearly did, because it did evolve new capabilities, right? (00:42:00) So bacteria to where we are now. (00:42:03) So clearly that it must be possible with evolutionary systems to (00:42:07) generate new patterns, going back to the first thing we talked about, and new capabilities and emergent properties. (00:42:15) And maybe we're on the cusp of discovering how to do that. (00:42:19) Yeah, listen, Alpha Evolve is one of the coolest things I've ever seen. (00:42:22) I've, on my desk at home, you know, most of my time is spent behind that computer is just programming, and next (00:42:30) to the three screens is a skull of a tiktaalik, which is one of the early organisms that crawled out of the water onto land. (00:42:40) And I just kind of watch that little guy. (00:42:44) It's like, whatever the computation mechanism of evolution is, (00:42:49) It's quite incredible. (00:42:51) It's truly, truly incredible. (00:42:53) Now, whether that's exactly the thing we need to do to do our search, but never, dismiss the power of nature with what it did here. (00:43:00) Yeah. (00:43:01) And it's amazing, which is a relatively simple algorithm, right? (00:43:05) Effectively. (00:43:06) And it can generate all of this immense complexity emerges, obviously running over, you know, 4 billion years of time. (00:43:13) But it's, you know, you can think about that as, again, a search process that ran over (00:43:19) the physics substrate of the universe for a long amount of computational time. (00:43:24) But then it generated all this incredible, rich diversity. (00:43:28) So many questions I want to ask you. (00:43:30) So one, you do have a dream. (00:43:32) One of the natural systems you want to try to model is a cell. (00:43:37) That's a beautiful dream. (00:43:40) I could ask you about that. (00:43:41) I also just for that purpose, (00:43:43) on the AI scientist front just broadly. (00:43:46) So there's an essay from Daniel Cocotallo, Scott Alexander, and others that outlines steps along the way to get to ASI and has a lot of interesting ideas in it, one of which is including a superhuman coder and a superhuman AI researcher. (00:44:04) And in that, there's a term of research taste that's really interesting. (00:44:09) So in everything you've seen, do you think it's possible for AI systems to have research taste to help you in the way that AI co-scientist does to help steer human brilliant scientists and then potentially by itself to figure out what are the directions where you want to generate truly novel ideas? (00:44:34) Because that seems to be like a (00:44:36) really important component of how to do great science. (00:44:38) Yeah, I think that's going to be one of the hardest things to mimic or model is this idea of taste or judgment. (00:44:46) I think that's what separates the, you know, the great scientists from the good scientists, like all professional scientists are good technically, right? (00:44:53) Otherwise, it wouldn't have made it that far in academia and things like that. (00:44:57) But then do you have the taste to sort of sniff out what the right direction is, what the right experiment is, what the right question is? (00:45:04) So it's picking (00:45:05) the right question is the hardest part of science and making the right hypothesis. (00:45:11) And that's what today's systems definitely, they can't do. (00:45:15) So, I often say it's harder to come up with a conjecture, a really good conjecture, than it is to solve it. (00:45:21) So we may have systems soon that can solve pretty hard conjectures. (00:45:27) Maths Olympiad problems, Alpha proof last year, our system got a silver medal in that, really hard problems. (00:45:34) Maybe eventually we'll better solve a Millennium Prize kind of problem. (00:45:37) But could a system come up with a conjecture worthy of study that someone like Terence Tao would have gone, you know what? (00:45:44) That's a really deep question about the nature of maths or the nature of numbers or the nature of physics. (00:45:50) And that is a far harder type of creativity. (00:45:53) And we don't really know those systems (00:45:55) clearly can't do that, and we're not quite sure what that mechanism would be, this kind of leap of imagination, like Einstein had when he came up with, special relativity and then general relativity with the knowledge he had at the time. (00:46:07) As for conjecture, the... (00:46:11) You want to come up with a thing that's interesting. (00:46:14) It's amenable to proof. (00:46:15) Yes. (00:46:16) So like, it's easy to come up with a thing that's extremely difficult. (00:46:19) Yeah. (00:46:19) It's easy to come up with a thing that's extremely easy, but that at that very edge. (00:46:23) That sweet spot, right? (00:46:24) Of basically advancing the science and splitting the hypothesis space into two, ideally, right? (00:46:29) Whether if it's true or not true, you've learned something really useful. (00:46:33) And that's hard. (00:46:36) And making something that's also, (00:46:40) falsifiable and within sort of the technologies that you have, you currently have available. (00:46:45) So it's a very creative process, actually, highly creative process that I think just a kind of naive search on top of a model won't be enough for that. (00:46:55) Okay, the idea of splitting the hypothesis space in two is super interesting. (00:47:00) I've heard you say that there's basically no failure in, or failure is extremely valuable if it's done, if you construct the questions right, if you construct the experiments right, if you design them right, that failure or success are both useful. (00:47:14) So perhaps because it splits the hypothesis base in two, it's like a binary search. (00:47:18) That's right. (00:47:19) So when you do like, you know, real blue sky research, there's no such thing as failure, really, as long as you're picking experiments and hypotheses that meaningfully split the hypothesis space. (00:47:30) So, you know, and you learn something, you can learn something kind of equally valuable from an experiment that doesn't work. (00:47:36) That should tell you, if you've designed the experiment well, and your hypotheses are interesting, it should tell you a lot about where to go next. (00:47:43) And then you're effectively doing a search process and using that information in very helpful ways. (00:47:52) So to go to your dream of modeling a cell, what are the big challenges that lay ahead for us to make that happen? (00:48:01) We should maybe highlight that AlphaFold, I mean, there's just so many leaps. (00:48:05) So AlphaFold solved, if it's fair to say, protein folding, and there's so many incredible things we could talk about there, including the open sourcing. (00:48:14) everything you've released. (00:48:15) AlphaFold 3 is doing protein, RNA, DNA interactions, which is super complicated and fascinating. (00:48:23) It's amenable to modeling. (00:48:24) AlphaGenome predicts how small genetic changes, like if we think about single mutations, how they link to actual function. (00:48:33) So those are, it seems like it's creeping along. (00:48:37) Yes, to a sophisticated, to a much more complicated (00:48:40) things like a cell, but a cell has a lot of really complicated components. (00:48:45) Yeah. (00:48:45) So what I've tried to do throughout my career is I have these really grand dreams. (00:48:49) And then I try to, as you've noticed, and then I try to break, but I try to break them down. (00:48:53) It's easy to have a kind of a crazy ambitious dream, but the trick is how do you break it down into manageable, achievable interim steps that are meaningful and useful in their own right. (00:49:07) And so virtual cell, which is what I call the project of modeling a cell, (00:49:10) I've had this idea, of wanting to do that for maybe more like twenty-five years, and I used to talk with Paul Nurse, who is a bit of a mentor of mine in biology. (00:49:20) He runs the, you know, founded the Crick Institute and won the Nobel Prize in 2001. (00:49:25) It is, we've been talking about it since, before, in the 90s. (00:49:30) And I used to come back to it every five years. (00:49:33) It's like, what would you need to model of the full internals of a cell so that you could do experiments on the virtual cell and what those experiments, you know, in silico and those predictions would be useful for you to save you a lot of time in the wet lab, right? (00:49:47) That would be the dream. (00:49:47) Maybe you could 100X speed up experiments by doing most of it in silico, the search in silico, and then you do the validation step in the wet lab. (00:49:55) That would be, that's the dream. (00:49:57) And so, but maybe now, finally, so I was trying to build these components, AlphaFold being one, that would allow you eventually to model the full interaction, a full simulation of a cell. (00:50:10) And I'd probably start with a yeast cell. (00:50:13) And partly that's what Paul Nurse studied, because a yeast cell is like a full organism that's a single cell, right? (00:50:18) So it's the kind of simplest single cell organism. (00:50:21) And so it's not just a cell, it's a full organism. (00:50:23) And (00:50:25) And yeast is very well understood. (00:50:27) And so that would be a good candidate for a kind of full simulated model. (00:50:32) Now, AlphaFold is the solution to the kind of static picture of what does a protein look, 3D structure a protein look like, a static picture of it. (00:50:41) But we know that biology, all the interesting things happen with the dynamics, the interactions. (00:50:45) And that's what AlphaFold 3 is the first step towards, is modeling those interactions. (00:50:50) So first of all, pairwise, you know, proteins with proteins, proteins with RNA and DNA. (00:50:55) But then (00:50:55) Then the next step after that would be modeling maybe a whole pathway, maybe like the TOR pathway that's involved in cancer or something like this. (00:51:02) And then eventually you might be able to model, you know, a whole cell. (00:51:05) Also, there's another complexity here that stuff in a cell happens at different time scales. (00:51:11) Is that tricky? (00:51:12) Like the, you know, protein folding is, you know, super fast. (00:51:17) Yes. (00:51:19) I don't know all the biological mechanisms, but some of them take a long time. (00:51:22) And so that's a level. (00:51:23) So the levels of interaction has a different temporal scale that you have to be able to model. (00:51:28) So that would be hard. (00:51:29) So you'd probably need several simulated systems that can interact at these different temporal dynamics, or at least maybe it's like a hierarchical system. (00:51:37) So you can jump up or down the different temporal stages. (00:51:42) So can you avoid, I mean, one of the challenges here is (00:51:46) not avoid simulating, for example, the quantum mechanical aspects of any of this, right? (00:51:52) You want to not over model. (00:51:55) You can skip ahead to just model the really high level things that get you a really good estimate of what's going to happen. (00:52:02) Yes. (00:52:02) So you've got to make a decision when you're modeling any natural system, what is the cutoff level of the granularity that you're going to model it to, then it captures the dynamics that you're interested in. (00:52:11) So probably for a cell, I would hope that would be the protein level and that one wouldn't have to go down to the atomic level. (00:52:20) So, of course, that's where AlphaFold stock kicks in. (00:52:23) So that would be kind of the basis. (00:52:26) And then you'd build these higher level simulations that take those as building blocks and then you get the emergent behavior. (00:52:35) I apologize for the pothead questions ahead of time, but do you think we'll be able to simulate a model, the origin of life? (00:52:45) So being able to simulate the first from non-living organisms, the birth of a living organism. (00:52:54) I think that's one of the, of course, one of the deepest and most fascinating questions. (00:52:58) I love that area of biology, you know, (00:53:01) these people, like there's a great book by Nick Lane, one of the top, top experts in this area called The 10 Great Inventions of Evolution. (00:53:09) I think it's fantastic. (00:53:10) And it also speaks to what the great filters might be, you know, prior or are they ahead of us. (00:53:15) I think they're most likely in the past, if you read that book, of how unlikely to go, you know, have any life at all and then single cell to multi-cell seems an unbelievably big jump that took like a billion years, I think, on Earth to do, right? (00:53:28) So it shows you how hard it was. (00:53:29) Exterior, we're super happy for a very (00:53:31) for a very long time before they captured mitochondria somehow, right? (00:53:35) I don't see why not, why AI couldn't help with that. (00:53:38) Some kind of simulation again, it's again, it's a bit of a search process through a combinatorial space. (00:53:44) Here's like all the, you know, the chemical soup that you start with, the primordial soup that, you know, maybe was on Earth near these hot vents. (00:53:52) Here's some initial conditions. (00:53:53) Can you generate something that looks like a cell? (00:53:57) So perhaps that would be a next stage after the virtual cell project is, well, how could you (00:54:01) to actually something like that emerge from the chemical soup? (00:54:05) Well, I would love it if there was a move 37 for the origin of life. (00:54:09) I think that's one of the great mysteries. (00:54:12) I think ultimately what we'll figure out is their continuum. (00:54:14) There's no such thing as a line between non-living and living. (00:54:17) But if we can make that rigorous, that the very thing from the Big Bang to today has been the same process. (00:54:24) If we can break down that wall that we've constructed in our minds of the actual origin (00:54:29) of from non-living to living and it's not a line, that it's a continuum that connects physics and chemistry and biology. (00:54:37) There's no line. (00:54:38) I mean, this is my whole reason why I've worked on AI and AGI my whole life, because I think it can be the ultimate tool to help us answer these kind of questions. (00:54:45) And I don't really understand why. (00:54:49) the average person doesn't think, worry about this stuff more. (00:54:52) Like, how can we not have a good definition of life and not living and non-living and the nature of time and let alone consciousness and gravity and all these things? (00:55:03) It's just, and quantum mechanics weirdness, it's just, to me, it's, I've always had this sort of screaming at me in my face, the whole, I need that, it's getting louder. (00:55:12) You know, it's like, how, what is going on here? (00:55:14) You know, in, and I mean that in a deeper sense, like in the, (00:55:18) nature of reality, which has to be the ultimate question that would answer all of these things. (00:55:23) It's sort of crazy if you think about it. (00:55:24) We could stare at each other and all these living things all the time, we can inspect it in microscopes and take it apart almost down to the atomic level. (00:55:32) And yet we still can't answer that clearly in a simple way, that question of how do you define living? (00:55:39) It's kind of amazing. (00:55:40) Yeah. (00:55:40) Living, you can kind of talk your way out of thinking about, but like consciousness, (00:55:45) we have this very obviously subjective conscious experience, like we're at the center of our own world and it feels like something. (00:55:52) And then how are you not screaming at the mystery of it all? (00:55:57) I mean, but really humans have been contending with the mystery of the world around them for a long, time. (00:56:04) There's a lot of mysteries. (00:56:06) Like what's up with the sun and the rain? (00:56:10) what's that about? (00:56:11) And then last year we had a lot of rain and this year we don't have rain. (00:56:15) Like what do we do wrong? (00:56:17) Humans have been asking that question for a long time. (00:56:19) Exactly. (00:56:19) So we're quite, I guess we've developed a lot of mechanisms to cope with this, these deep mysteries that we can't fully, we can see, but we can't fully understand. (00:56:28) And we have to just get on with daily life. (00:56:30) And we keep ourselves busy, right? (00:56:33) In a way, do we keep ourselves distracted? (00:56:35) I mean, weather is one of the most important questions of human history. (00:56:39) We still, that's (00:56:40) the go-to small talk direction of the weather. (00:56:44) Especially in England. (00:56:45) And then it's, which is, famously, it's an extremely difficult system to model. (00:56:51) And even that system, Google DeepMind has made progress on. (00:56:57) Yes, we've created the best weather prediction systems in the world, and they're better than traditional fluid dynamics sort of systems that are usually calculated on massive supercomputers, takes days to calculate it. (00:57:10) We've managed to model a lot of the weather dynamics with neural network systems with our weather next system. (00:57:16) And again, it's interesting that those kinds of dynamics can be modeled, even though they're very complicated, almost bordering on chaotic systems in some cases. (00:57:25) A lot of the interesting aspects of that. (00:57:28) can be modeled by these neural network systems, including very recently we had, cyclone prediction of where, paths of hurricanes might go, of course, super useful, super important for the world. (00:57:38) And it's super important to do that very timely and very quickly and as well as accurately. (00:57:42) And I think it's a very promising direction, again, of, you know, simulating and so that you can run forward predictions and simulations of very complicated real world systems. (00:57:52) I should mention that I've got a chance in Texas. (00:57:56) To me, (00:57:57) a community of folks called the Storm Chasers. (00:57:59) And what's really incredible about them, I need to talk to them more, is they're extremely tech savvy. (00:58:04) Because what they have to do is they have to use models to predict where the storm is. (00:58:08) So it's this beautiful mix of crazy enough to go into the eye of the storm. (00:58:16) in order to protect your life and predict where the extreme events are going to be, they have to have increasingly sophisticated models of weather. (00:58:24) Yeah. (00:58:25) it's a beautiful balance of like being in it. (00:58:29) as living organisms and the cutting edge of science. (00:58:32) So they actually might be using DeepMind systems. (00:58:35) So that's... (00:58:36) Yeah, they are. (00:58:36) But hopefully they are. (00:58:37) And I'd love to join them on one of those checks. (00:58:39) They look amazing, right? (00:58:40) To actually experience it one time. (00:58:42) Exactly. (00:58:42) And then also to experience the correct prediction of where something will come and how it's going to evolve. (00:58:48) It's incredible. (00:58:49) Yeah. (00:58:50) You've estimated that we'll have AGI by 2030. (00:58:55) So there's interesting questions around that. (00:58:57) How will we actually know that we got there? (00:59:01) And what may be the move, quote, move 37 of AGI? (00:59:07) My estimate is sort of 50% chance in the next five years. (00:59:11) So, you know, by 2030, let's say. (00:59:14) And so I think there's a good chance that could happen. (00:59:17) Part of it is what is your definition of AGI? (00:59:19) Of course, people are arguing about that now. (00:59:22) And mine's quite a high bar and always has been of like, can we (00:59:25) match the cognitive functions that the brain has. (00:59:28) Right. (00:59:28) So we know our brains are pretty much general Turing machines, approximate. (00:59:33) And of course, we created incredible modern civilization with our minds. (00:59:38) So that also speaks to how general the brain is. (00:59:41) And for us to know we have a true AGI, we would have to make sure that it has all those capabilities. (00:59:47) It isn't kind of a jagged intelligence where some things it's really good at, like today's systems, but other things it's really (00:59:54) flaw that. (00:59:55) And that's what we currently have with today's systems. (00:59:57) They're not consistent. (00:59:58) So you'd want that consistency of intelligence across the board. (01:00:01) And then we have some missing, I think, capabilities, like sort of the true invention capabilities and creativity that we were talking about earlier. (01:00:09) So you'd want to see those. (01:00:11) How you test that, I think you just test it. (01:00:14) One way to do it would be kind of brute force test of 10s of thousands of cognitive tasks that, you know, we know that humans can do. (01:00:23) and maybe also make the system available to a few hundred of the world's top experts, Terence Towers of each subject area, and see if they can find, give them a month or two, and see if they can find an obvious flaw in the system. (01:00:38) And if they can't, then I think you're pretty, you know, you can be pretty confident we have a fully general system. (01:00:45) Maybe to push back a little bit, it seems like humans are really incredible (01:00:49) as the intelligence improves across all domains to take it for granted, like you mentioned, Terence Tao, these brilliant experts, they might quickly, in a span of weeks, take for granted all the incredible things it can do and then focus in, well, haha, right there. (01:01:07) You know, I consider myself, first of all, human. (01:01:13) I identify as human. (01:01:18) I, some people listen to me talk and they're like, that guy is not good at talking, the stuttering, the, so even humans have obvious across domains limits, even just outside of mathematics and physics and so on. (01:01:34) I wonder if it will take something like a Move 37, so on the positive side versus like a barrage of 10,000 cognitive tasks where it will be one or two where it's like, holy shit, this is special. (01:01:49) So I think there's the sort of blanket testing to just make sure you've got the consistency, but I think there are the sort of lighthouse moments like the Move 37 that I would be looking for. (01:02:01) So one would be inventing a new conjecture (01:02:04) or new hypothesis about physics like Einstein did. (01:02:08) So maybe you could even run the back test of that very rigorously, like have a cut off of knowledge cut off of 1900 and then give the system everything that was written up to 1900 and then see if it could come up with special relativity and general relativity, right, like Einstein did. (01:02:24) That would be an interesting test. (01:02:25) Another one would be, can it invent a game like Go? (01:02:29) Not just come up with Move 37, a new strategy, but can it invent a game that's as deep (01:02:34) as aesthetically beautiful, as elegant as go. (01:02:37) And those are the sorts of things I would be looking out for, and probably a system being able to do several of those things, right, for it to be very general. (01:02:47) not just one domain. (01:02:48) And so I think that would be the signs, at least that I would be looking for, that we've got a system that's AGI level. (01:02:54) And then maybe to fill that out, you would also check their consistency, make sure there's no holes in that system either. (01:03:01) Yeah, something like a new conjecture or scientific discovery, that would be a cool feeling. (01:03:06) Yeah, that would be amazing. (01:03:08) So it's not just helping us do that, but actually coming up with something brand new. (01:03:12) And you would be in the room for that. (01:03:14) And it would be like, (01:03:16) probably two or three months before announcing it. (01:03:20) And you would just be sitting there trying not to tweet. (01:03:24) Something like that. (01:03:25) Exactly. (01:03:25) It's like, what is this amazing new physics idea? (01:03:29) And then we would probably check it with world experts in that domain, right? (01:03:33) And validate it and kind of go through its workings. (01:03:36) And I guess it would be explaining its workings too. (01:03:40) Yeah, be an amazing moment. (01:03:42) Do you worry that we, as humans, even expert humans like you, might miss it? (01:03:46) Might miss? (01:03:47) Well, it may be pretty complicated. (01:03:48) So it could be, the analogy I give there is, I don't think it will be... (01:03:54) totally mysterious to the best human scientists. (01:03:57) But it may be a bit like, for example, in chess, if I was to talk to Garry Kasparov or Magnus Carlsen and play a game with them and they make a brilliant move, I might not be able to come up with that move, but they could explain why afterwards that move made sense. (01:04:11) And we would be able to understand it to some degree, not to the level they do, but if they were good at explaining, which is actually part of intelligence too, is being able to explain in a simple way that what you're thinking about. (01:04:23) I think that will be very possible for the best human scientists. (01:04:27) But I wonder, maybe you can educate me on the side of Go. (01:04:30) I wonder if there's moves from Agnes or Gary where they at first will dismiss it as a bad move. (01:04:36) Yeah, sure. (01:04:38) It could be. (01:04:38) But then afterwards, they'll figure out with their intuition that this why this works. (01:04:43) And then empirically, the nice thing about games is one of the great things about games is you can, it's a sort of scientific test. (01:04:49) Does it do you win the game or not win? (01:04:51) And then that tells you, okay, that move in the end was good. (01:04:55) That strategy was good. (01:04:57) And then you can go back and analyze that and explain even to yourself a little bit more why, explore around it. (01:05:03) And that's how chess analysis and things like that work. (01:05:06) So perhaps that's why my brain works (01:05:08) that because I've been doing that since I was four. (01:05:10) And you're trained, it's sort of hardcore training in that way. (01:05:14) But even now, like when I generate code, there is this kind of nuanced, fascinating contention that's happening where I might at first identify as a set of generated code as incorrect in some interesting, nuanced ways. (01:05:31) But then I'm always have to ask the question, is there a deeper insight here that I (01:05:37) I'm the one who's incorrect. (01:05:39) And that's going to, as the systems get more and more intelligent, you're going to have to contend with that. (01:05:44) It's like, what, do you, is this a bug or a feature what you just came up with? (01:05:48) Yeah, and they're going to be pretty complicated to do, but of course it will be, you can imagine also AI systems that are producing that code or whatever that is, and then human programmers looking at it, but also not unaided with the help of AI tools as well. (01:06:02) So it's going to be kind of an interesting, you know, maybe different AI tools to the ones (01:06:07) The more kind of monitoring tools are the ones that generated it. (01:06:10) So if we look at an AGI system, sorry to bring it back up, but Alpha Evolve, super cool. (01:06:17) So Alpha Evolve enables, on the programming side, something like recursive self-improvement, potentially. (01:06:25) Like what, if you can imagine what that AGI system, maybe not the first version, but a few versions. (01:06:32) Beyond that, what does that actually look like? (01:06:34) Do you think it will be simple? (01:06:35) Do you think it will be something like a self-improving program and a simple one? (01:06:40) I mean, potentially that's possible, I would say. (01:06:43) I'm not sure it's even desirable because that's a kind of like hard takeoff scenario. (01:06:47) But these current systems like Alpha Evolve, they have human in the loop deciding on various things. (01:06:54) They're separate hybrid systems that interact. (01:06:57) One could imagine eventually doing that end to end. (01:07:00) I don't see why that wouldn't be possible. (01:07:02) Right now, I think the systems are not good enough to do that in terms of coming up with the architecture of the code. (01:07:10) And again, it's a little bit reconnected to this idea of coming up with a new conjectural hypothesis. (01:07:15) They're good if you give them very specific instructions about what you're trying to do. (01:07:20) But if you give them a very vague, high-level instruction, that wouldn't work currently. (01:07:24) Like, and I think that's related to this idea of, like, invent a game as good as Go. (01:07:29) Right? (01:07:29) Imagine that was the prompt. (01:07:30) That's pretty underspecified. (01:07:32) And so the current systems wouldn't know, I think, what to do with that, how to narrow that down to something tractable. (01:07:38) And I think there's similar, like, look, just make a better version of yourself. (01:07:41) That's too unconstrained. (01:07:43) But we've done it in, you know, and as you know, with Alpha Evolve, like things like faster matrix multiplication. (01:07:49) So when you (01:07:50) hone it down to a very specific thing you want, it's very good at incrementally improving that. (01:07:55) But at the moment, these are more like incremental improvements, sort of small iterations. (01:08:00) Whereas if you wanted a big leap in understanding, you'd need a much larger advance. (01:08:08) Yeah, but it could also be sort of to push back against hard takeoff scenario. (01:08:12) It could be just a sequence of (01:08:16) incremental improvements like matrix multiplication like it has to sit there for days thinking how to incrementally improve a thing and that it does solve recursively and as you do more and more improvement it'll slow down right there'll be like a like the path to AGI won't be like a (01:08:34) it would be a gradual improvement over time. (01:08:37) Yes, if it was just incremental improvements, that's how it would look. (01:08:40) So the question is, could it come up with a new leap, like the Transformers architecture? (01:08:45) Could it have done that back in 2017 when we did it and Brain did it? (01:08:50) And it's not clear that these systems, something like AlphaVol wouldn't be able to do, make such a big leap. (01:08:56) So for sure, these systems are good. (01:08:58) We have systems, I think, that can do incremental hill climbing. (01:09:01) And that's a kind of bigger question about, is that all that's needed from here? (01:09:04) Or do we actually need one or two more big breakthroughs? (01:09:08) And can the same kind of systems provide the breakthroughs also? (01:09:13) So make it a bunch of S-curves. (01:09:15) Like incremental improvement, but also every once in a while, leaps. (01:09:19) Yeah, I don't think anyone has systems that can have shown unequivocally those big leaps. (01:09:25) Right, we have a lot of systems that do the hill climbing of the S-curve that you're currently on. (01:09:29) Yeah, and that would be the move 37 is a leap. (01:09:32) Yeah, I think it would be a leap. (01:09:34) Something like that. (01:09:36) Do you think the scaling laws are holding strong on pre-training, post-training, test-time, compute? (01:09:42) Do you, on the flip side of that, anticipate AI progress hitting a wall? (01:09:47) We certainly feel there's a lot more room just in the scaling, so actually all steps, pre-training, post-training, and... (01:09:55) inference time. (01:09:56) So there's sort of three scalings that are happening concurrently. (01:10:01) And we, again, there, it's about how innovative you can be. (01:10:05) And we, pride ourselves on having the broadest and deepest research bench. (01:10:11) We have amazing, incredible researchers and people like Noam Shazir who, came up with Transformers and Dave Silver, who led the AlphaGo project and so on. (01:10:22) And (01:10:23) it's that research base means that if some new breakthrough is required, like an AlphaGo or Transformers, I would back us to be the place that does that. (01:10:34) So I'm actually quite like it when the terrain gets harder, right? (01:10:36) Because then it veers more from just engineering to true research and, you know, research plus engineering, and that's our sweet spot. (01:10:45) And I think that's harder. (01:10:46) It's harder to invent things than to, you know, fast follow. (01:10:51) And (01:10:53) So, we don't know. (01:10:54) I would say it's kind of 50-50 whether new things are needed or whether the scaling the existing stuff is going to be enough. (01:11:02) And so in true kind of empirical fashion, we're pushing both of those as hard as possible. (01:11:07) The new blue sky ideas and, you know, maybe about half our resources on that. (01:11:11) And then scaling to the max, the current capabilities. (01:11:16) And we're still seeing some, you know, fantastic progress on each different version of Gemini. (01:11:23) That's interesting the way you put it in terms of the deep bench, that if progress towards AGI is more than just scaling compute, so the engineering side of the problem, and is more on the scientific side where there's breakthroughs needed, then you feel confident, DeepMind is well, Google DeepMind is well positioned to kick ass in that domain. (01:11:47) Well, I mean, if you look at the history of the last decade or 15 years, (01:11:52) It's been, maybe, I don't know, 80, 90% of the breakthroughs that underpins modern AI field today was from, originally Google Brain, Google Research, and DeepMind. (01:12:00) So, yeah, I would back that to continue, hopefully. (01:12:05) So, on the data side, are you concerned about running out of high-quality data, especially high-quality human data? (01:12:11) I'm not very worried about that, partly because I think there's enough data, and it's been proven to get the systems to be pretty good. (01:12:18) And this goes back to simulations again. (01:12:21) Do you have enough data to make simulations so that you can create more synthetic data that are from the right distribution? (01:12:29) Obviously, that's the key. (01:12:30) So you need enough real-world data in order to be able to create those kinds of data generators. (01:12:37) And I think that we're at that step at the moment. (01:12:40) Yeah, you've done a lot of incredible stuff on the side of science and biology, doing a lot with not so much data. (01:12:46) Yeah. (01:12:46) I mean, it's still a lot of data, but I guess enough takeoff. (01:12:50) Get that going. (01:12:50) Exactly. (01:12:51) So exactly. (01:12:53) How crucial is the scaling of compute to building AGI? (01:12:56) This is a question that's an engineering question. (01:12:59) It's almost a geopolitical question because it also integrated into that. (01:13:05) Is the supply chains and energy a thing that you care a lot about, which is potentially fusion? (01:13:11) So innovating on the side of energy also. (01:13:13) Do you think we're going to keep scaling compute? (01:13:16) I think so for several reasons. (01:13:17) I think compute, there's the amount of compute you have for training. (01:13:21) Often it needs to be co-located. (01:13:23) So actually even like bandwidth constraints between data centers can affect that. (01:13:28) So there's additional constraints even there. (01:13:31) And that's important for training, obviously, the largest models you can. (01:13:35) But there's also, because now AI systems are in products and being used by billions of people around the world, you need a ton of inference compute now. (01:13:45) And then on top of that, there's the thinking systems, the new paradigm of the last year, where they get smarter the longer amount of inference time you give them at test time. (01:13:55) So all of those things need a lot of compute. (01:13:58) And I don't really see that slowing down. (01:14:01) And as AI systems become better, they'll become more useful and there'll be more demand for them. (01:14:06) So both from the training side, the training side actually is only just one part of that. (01:14:10) It may even become the smaller part of what's needed in the overall compute that's required. (01:14:17) Yeah, that's one sort of almost meme-y kind of thing, which is like the success and the incredible aspects of VO3. (01:14:25) people kind of make fun of, the more successful it becomes, the servers are sweating. (01:14:30) Yes, exactly. (01:14:31) Yeah, exactly. (01:14:32) We did a little video of the servers frying eggs and things, and that's right. (01:14:38) And we're going to have to figure out how to do that. (01:14:41) There's a lot of interesting hardware innovations that we do. (01:14:43) As you know, we have our own TPU line and we're looking at like inference only things, inference only chips and how we can make those more efficient. (01:14:50) We're also very interested in building AI systems and we have done the help with energy usage. (01:14:55) So help data center energy like for the cooling systems be efficient, grid optimization, (01:15:03) and then eventually things like helping with plasma containment fusion reactors. (01:15:06) We've done lots of work on that with Commonwealth fusion and also one could imagine reactor design. (01:15:13) And then material design, I think, is one of the most exciting. (01:15:16) New types of solar material, solar panel material, room temperature superconductors has always been on my list of dream breakthroughs and optimal batteries. (01:15:25) And I think a solution to any, you know, one of those things would be absolutely revolutionary for, you know, climate and energy usage. (01:15:32) And we're probably close, you know, again, in the next five years to having AI systems that can materially help with those problems. (01:15:39) If you were to bet, sorry for the ridiculous question, what is the main source of energy in like 20, 30, 40 years? (01:15:47) Do you think it's going to be nuclear fusion? (01:15:49) I think fusion and solar are the two that I would bet on. (01:15:54) Solar, I mean, it's the fusion reactor in the sky, of course. (01:15:58) And I think really the problem there is batteries and transmission. (01:16:02) So, as well as more efficient, more and more efficient solar material, perhaps eventually, in space, these kind of Dyson sphere type ideas. (01:16:10) And fusion, I think, is definitely doable, it seems, if we have the right design of reactor and we can control the plasma and fast enough and so on. (01:16:21) And I think both of those things will actually get solved. (01:16:24) So we'll probably have at least, those are probably the two primary sources of renewable, clean, almost free, or perhaps free energy. (01:16:32) What a time to be alive. (01:16:34) If I traveled into the future with you 100 years from now, how much would you be surprised if we've passed a type one Kardashev scale civilization? (01:16:46) I would not be that surprised if there's like 100 year time scale from here. (01:16:50) I mean, I think it's pretty clear if we crack the energy problems in one of the ways we've just discussed, fusion or very efficient solar, then if energy is kind of free and renewable and clean, then (01:17:04) that solves a whole bunch of other problems. (01:17:06) So for example, the water access problem goes away because you can just use desalination. (01:17:11) We have the technology. (01:17:12) It's just too expensive. (01:17:13) So only, you know, fairly wealthy countries like Singapore and Israel and so on, like actually use it. (01:17:19) But if it was cheap, then, you know, all countries that have a coast could. (01:17:23) But also you'd have unlimited rocket fuel. (01:17:25) You could just separate seawater out into hydrogen and oxygen using energy and that's rocket fuel. (01:17:31) So combined with, you know, Elon's amazing (01:17:35) self-landing rockets, then it could be sort of like a bus service to space. (01:17:40) So that opens up incredible new resources and domains. (01:17:45) Asteroid mining, I think, will become a thing and maximum human flourishing to the stars. (01:17:49) That's what I dream about as well as like Carl Sagan's sort of idea of bringing consciousness to the universe, waking up the universe. (01:17:56) And I think human civilization will do that in the full sense of time if we get AI right and crack some of these problems with it. (01:18:05) I wonder what it would look like if you're just a tourist flying through space. (01:18:09) You would probably notice Earth, because if you solve the energy problem, you would see a lot of space rockets, probably. (01:18:16) So it would be like traffic here in London, but in space, just a lot of rockets. (01:18:22) And then you would probably see floating in space, some kind of source of energy, like solar. (01:18:29) potentially. (01:18:29) So Earth would just look more on the surface, more technological. (01:18:35) And then you would use the power of that energy then to preserve the natural, like the rainforest and all that kind of stuff. (01:18:41) Exactly. (01:18:42) Because for the first time in human history, we wouldn't be resource constrained. (01:18:48) And I (01:18:48) I think that could be an amazing new era for humanity where it's not zero sum, right? (01:18:54) I have this land, you don't have it. (01:18:56) Or if we take, you know, if the tigers have their forest, then the local villagers can't, what are they going to use? (01:19:03) I think that this will help a lot. (01:19:04) It won't solve all problems because there's still other human foibles that will still exist, but it will at least remove one, I think one of the big vectors, which is scarcity of resources, you know, including land and more materials and (01:19:18) Energy, and we know we should be, as sometimes call it, and others call it, about this kind of radical abundance era, where there's plenty of resources to go around. (01:19:27) Of course, the next big question is making sure that that's fairly, shared fairly. (01:19:33) and everyone in society benefits from that. (01:19:35) So there is something about human nature where I go, it's like Borat, like my neighbor, like you start trouble. (01:19:44) We do start conflicts. (01:19:47) And that's why games throughout, as I'm learning actually more and more, even in ancient history, serve the purpose of pushing people away from war, actually a hot war. (01:19:58) So maybe we can figure out