[0:00:00] Welcome to City Arts and Lectures, a season of talks and onstage conversations with some of the most celebrated writers, artists, and thinkers of our day, recorded before an audience at the Sidney Goldstein Theater in San Francisco. This week, our guests are Sir Demis Hassabis and Sebastian Malaby. In 2024, Hassabis won a Nobel Prize for developing an AI tool to solve one of biology's greatest challenges.
[0:00:56] Prediction of the 3D structure of proteins. He's currently the head of DeepMind, an AI research lab at Google. On April 20th, Hassabis and his biographer, Sebastian Malaby, talked to journalist Emily Chang about Hassabis' work and the future of AI. Join us now for a conversation with Sir Demis Hassabis and Sebastian Malaby. Good to see you.
[0:01:30] All right. So you submitted to quite a few interviews and background conversations for this book. And I wonder, what is it like reading about yourself? Were there moments of, yes, he nailed it. And we really got to talk about chapter seven. This is working for me. Well, it's...
[0:01:53] I'm a pretty private person, so it's pretty tricky reading about oneself in a book. But I think Sebastian's done the best possible job that anyone could have. So I'm really pleased that we did this together in a way.
[0:02:08] And it's been a pleasure, you know, the number of hours we spent together. That wasn't contractual. It was our choice to do that. So I think we just actually got on very well on a broad range of subjects. So it was actually pretty fun to spend that time with Sebastian for the book. And I was surprised how deeply technical you got. Surprised and delighted and also challenged. You open the book talking about the Manhattan Project.
[0:02:36] What were you trying to signal to the reader right away in that moment? Well, maybe Demis doesn't like this, but perhaps he is the modern Robert Oppenheimer, right? I mean, you know, the idea that in the 20th century you had this technology, the creation of which was extraordinary, an amazing scientific achievement, but at the same time it had these big political consequences, big risk.
[0:03:03] Even existential risk. And I remember, before I even began writing the book, that parallel struck me as interesting. I was a little nervous about raising it directly with people involved in the sector, not just EMIS, but other people I talked to.
[0:03:20] And I found that actually they thought about it more than I had. They were very conscious of the parallel. They had thought about the parallel. It was on their minds. And I remember particularly one time, I don't know if you remember this or not, but we used to meet in a pub in London. And there was a kind of a staircase at the back of the pub. You go up the stairs.
[0:03:43] And then there's this empty room which nobody ever went to. So we just had it to ourselves. We talked for two hours. And one day we got to the point in the series of conversations where I said, OK, so Demis, you always wanted to build AI from when you were in your teens. And then you finally raised money from Peter Thiel to open up a company devoted to artificial general intelligence. And you open the office and it's in Russell Square in London, this historic square. What did it feel like?
[0:04:12] And I know from experience as a writer that usually you ask people about what it felt like, and it was 15 years ago, the answer you get is, it was cool.
[0:04:22] That's all anyone says because they can't really remember. But Demis is an amazing storyteller. So he said, okay, Sebastian, so I was up in the attic, the top level of this building. It was quite a small office and a bit hot and stuffy in the summer because there's not much ventilation. So I had to come downstairs for a breather sometimes. So you come down the stairs, of course, there's no elevator. You go ding, ding, ding, ding, ding, ding, ding down the stairs. And there in front of you is Russell Square and there are trees and you breathe.
[0:04:46] Like this, and you feel much better. And then down the street, there is the London Mathematical Society, about four doors down, and that's fantastic, because that's where Turing gave his lectures on the origins of computer science and AI, and now we're completing that story. And if you go round the corner, there's a level crossing, a pedestrian crossing, it goes black-white, black-white, black-white. And in the 1930s, the Hungarian nuclear physicist, Zeilard, was crossing...
[0:05:11] And as he crossed black, white, black, white, he had the idea for the nuclear chain reaction, which led to the Manhattan Project. And now we are doing the modern equivalent, and it's perfect. I'm not sure I said it was perfect. Okay, whatever. It was definitely on our minds from the beginning. It was on your mind. You added that piece of history to Russell Square. You do point out that you were just yards away from this crossing where the nuclear chain reaction was discovered.
[0:05:40] What has that comparison to the Manhattan Project meant to you as a builder of AI? I think the analogies can be taken. I mean, it's a very different technology, and I think it's a lot more complicated in some ways because there's so many more good use cases. Obviously, that's why we're building AGI than I think Nuclear, certainly the Manhattan Project had. But it's of that level of consequence, and it was obvious to us, even though we were in a little attic in Russell Square,
[0:06:10] Nobody cared about AI back in 2010. It might be very difficult to raise any money. But we knew we were working on something that would change the world and indeed has changed the world. And that was the beginning of the atomic age. And at some point, someone will name what this age is. Maybe it's the intelligence age. And I think it's of that. It was clear to us, the few of us working on it, that it was going to be that significant.
[0:06:37] Now, my favorite part of the book as a mom was reading about your upbringing. And obviously, I knew you were a chess prodigy and everything, but like, no, your dad was taking you to chess tournaments from the age of five and you were practically living out of your car to go to these tournaments. Was that like a gift or was that traumatic? Like, you know, how do you look back on that now? Should parents be taking note or chilling out?
[0:07:02] I'm not sure I'd recommend it for the parents in the audience. I think it can go many different ways from there. What I will tell you is I had a very unusual childhood as the book chronicles. Put it this way, I've never had a summer holiday, ever.
[0:07:22] Ever? No, ever. Because they're very big in Europe, as I've been told. So what we used to do when I was a very young kid was every summer holiday, we would go around the country and then eventually the world, playing chess tournaments back to back from the day the school term ended, actually sometimes during school term, and then all the way to the start of the next school term. And then when I was older, in my teenage years,
[0:07:49] I started working professionally from 14, 15 years old at computer games companies. So I was also doing literally from the day my exams finished to the start of the next school year, working at Bullfog and other places, trying to...
[0:08:06] You know, I learned professional engineering. I was working on some AI. I was fascinated by game design. So I just got into this habit. So I actually don't know what it would be like to just, you know, be calm for a while and enjoy. And I have some regrets about that, of course. You know, it would have been lovely to have traveled around South America, you know, backpacking with friends, you know, for some holiday. I miss, you know, obviously I missed out on those things.
[0:08:35] You could still do that now. Maybe, but it wouldn't be quite the same now. So it wouldn't be quite as fun to do it at that time when you're just sort of discovering things. But I don't really have any regrets because I don't think normal things, the things I wanted to achieve could have been done unless I was...
[0:08:54] Doing things that extreme. Maybe we took it too far, but you can't tell at the time. So my aim was to try and understand as much of the world around us as possible. And AI was one of the expressions of doing that. Building AI is the ultimate tool to do that. And there was sort of no time to lose, really, I felt. And so I was pleased to try to be productive with every moment that was available.
[0:09:19] Sebastian, you describe Demis as being almost alien, quote unquote, in how he thinks. What do you mean by that? And secondarily, Demis, do you think that is a compliment? I think in the book, the alien word is actually a quote from Demis. And I think what you meant was that, you know, when you were young and at school in North London, you know,
[0:09:43] Sometimes you dropped out for a year because you were doing so much chess that you just taught yourself the school material in your bedroom at home. You obviously weren't a normal kid because you were just a chess prodigy already and that was evident. I don't know whether racially that's part of it. Your mother is Singaporean Chinese and that was a bit unusual in North London back then.
[0:10:06] Your name is difficult to pronounce. I find, by the way, that even here in the US, as I've gone around talking about this book, the number of people who, I think I'm right, it's Demis Hassabis, but they say Hassabis. And I say, no, no, you have to put emphasis on the right syllable. And so for various reasons, it was tough to fit in when you were an exceptional person. Compliment or no? I mean, I guess if you said it about yourself.
[0:10:34] Well, I don't know. I mean, it's definitely, I think there's a lot of truth to that. So I definitely didn't fit in in the normal way. But having said that, I didn't feel I was like an outsider or isolated either. I just felt I had a very strong sense of the mission from the very beginning, I guess, which was what was guiding me. So let's talk about the mission.
[0:10:56] Make sure I get this right, but at age 18, you turn down a million dollars to stay in school. After Cambridge, you start a gaming company. You pivot to a PhD in neuroscience, which you never really studied. At that point, AI could barely do anything. How did you know this was worth betting your life on? Well, in some senses, I didn't know that it would all work out in the way it has, although that's what we were...
[0:11:24] We're sort of predicting, even at the start of DeepMind, we thought it would be a 20-year mission, and I think it's roughly going to be that, right? From 2010, I would say 2030, we're sort of on track. But it sort of didn't matter to me in a way because my mission was to try and...
[0:11:42] I guess since a kid, since I can remember, is trying to understand the world around us and the big questions around us. And I don't, I remember distinctly feeling, I don't understand why more people aren't worried about all of these things. You know, what is time? What is consciousness? What is gravity, quantum mechanics? And of course you can read, and I did read all the physicists and read everything there is to read about these things. But you just sort of conclude actually, although there's amazing people, many of my scientific heroes like Feynman and so on have worked on these problems.
[0:12:11] I don't actually know very much about them. And that's still true today. It's definitely true 40 years ago. And so that was pretty shocking to me, actually, in a way. There were all these brilliant, brilliant people, beyond brilliant, all my heroes, and yet still, we didn't know anything about these things we sort of swim in all the time, like time. And that was quite...
[0:12:32] Troubling to me, as well as exciting. So I felt that that was my mission, was to, I've just got this insatiable curiosity, I suppose. But I'm also quite strategic, probably from my chest. So I was sort of thinking, how can we really do this properly? And that's what led me to computers and then AI very early on, because it was obvious to me that if that could be done, and there didn't seem any reason, specific law of physics preventing it,
[0:13:02] Turing and people like that had shown that, then this would be the most profound invention, the most useful tool, like a meta tool to do everything else, all the other sciences with, and also an unbelievably fascinating artifact in of itself. So what could be more interesting or more important to work on? And I made that, it was obvious to me that conclusion.
[0:13:28] When I was 14, 15, I guess. And then everything I've chosen to do after that was trying to find things and do things and find experiences that would help ultimately build a company like DeepMind and do what we've done. So that would explain why I then did some neuroscience and I was doing games, but I was doing AI in games. That's how I was funding my AI research at the time was under the guise of game design. And I didn't feel that, you know, why didn't I start DeepMind earlier? Well...
[0:13:57] I mean, it was hard enough starting it in 2010, but let alone if, you know, but every five years I would think about, well, is it the time to start a dedicated AI company? And only around 2009, 8, 9, did I decide actually we had enough ingredients to do it. So for me, it's never been a question of would it work and what if it didn't work? It's just...
[0:14:17] Sebastian, are you more or less scared of AI than when you were when you started writing this book? A bit more. Lovely. The reason is that I always,
[0:14:47] I assumed and understood that, of course, machines would be more intelligent than humans. But I thought they would have no motive to attack humans. Humans have evolved over centuries to millennia to want to survive, to pass on our DNA, to not be eaten by the lion. And machines just don't have that, right? I mean, as Marc Andreessen once said,
[0:15:13] The AI isn't going to attack you because it's a machine. It's not going to attack you any more than your toaster will attack you. So I went around with this comforting thought for quite a long time. And one day I went to visit Jeff Hinton, the sort of academic father of deep learning. And I sat in his kitchen in Toronto for a couple of hours because he's a real doomer. And I said, well, what's your PD? What's your probability that these machines are going to attack humans and really wipe most of us out? And he said, I don't know, 50%.
[0:15:42] I was like, whoa. But that wasn't what really got my attention most. What got my attention was I said to him, look, we evolved to survive, machines don't, so they're not going to attack us. And he said, okay, imagine you have a powerful AI. Imagine that there's an enemy AI, maybe in another enemy country, and it's going to attack your AI. So now you need to give your own AI instructions to defend itself if you see the attack coming. And so basically you've given it a survival instinct. Now do you feel happy, Sebastian?
[0:16:12] And I had to admit I was a little bit. So that was the thing that made me a bit more unsettled. What do you think, 50% or less? No, I don't have, I mean, I don't subscribe to these ideas of a PDoom or this sort of thing because I think it implies a precision that just isn't there. So the way I talk about it is there's, no, I mean, well, we've got to do some work to quantify that risk much better. And I don't think it's helpful to throw out 20%, 50%.
[0:16:41] What I say is it's non-zero. So I don't agree with the Pollyanna sort of view of like there's nothing to worry about here that people like Yann LeCun or maybe Andreessen would say. That's clearly, that makes no sense. And I'm also an optimist in the sense that I think we'll be able to deal with the issues if we give ourselves, and ourselves meaning humanity, enough time and we collaborate enough on doing that.
[0:17:09] So that I'm very sure. But obviously, that's not a given either. And the two things I worry about, which I think are non-zero, are misuse and misalignment. So misuse, which can happen with any powerful technology at dual purpose, is basically bad actors repurposing general parts of technology that were supposed to be built for good things, like we're trying to do with AI, for harmful ends.
[0:17:35] So that's what number one. Number two is if as these systems get more and more autonomous, and we're now sort of in the beginnings of the agentic era, which a lot of us have been talking about for a while, but now everyone can see it happening. As the systems get more autonomous, as we get closer to AGI, how can you make sure that they're going to do the things that you thought they were going to do?
[0:17:59] That you sort of gave them the goals or the objectives and some guardrails. And how do you make sure they stay within those guardrails and they don't, for example, misinterpret them? Which is what the whole Asimov's entire career and books were about, was that all the obvious things you think are going to be easy, like an off switch or three laws, they don't work because there's room for misinterpretation in the wrong context.
[0:18:26] So I worry about, so those are two real, very difficult challenges. One's more societal and social, the other one and governance, and the other one is technical. And they're both very difficult, but I think they're doable. And I think there's evidence that, let's take misalignment, the current systems seem a bit more steerable.
[0:18:47] than we might have imagined 10 years ago. Even the chatbots and the assistants that we're building, you know, they're pretty useful and they sort of roughly stick on some guardrails, which is a little bit surprising that they're that sort of, you know, relatively simple methods.
[0:19:04] RLHF and other methods like that are able to hone them into something that is useful. Of course, there's question marks about how deep that goes and all these sorts of things which need to be urgently answered. So I'm in the sort of middle camp really of...
[0:19:23] There's definitely something to worry about, and it's very serious, but there's a lot of uncertainty. So in that sort of situation, the only correct approach, in my opinion, is kind of cautious optimism. So I think we obviously want all the amazing things that AGI can do, advanced science.
[0:19:41] Come up with cures for diseases, medicine, all the things I've tried to work on personally and why I started this whole journey. But we need to also be cognizant and mitigate the downsides. And there's no simple answer to that. And it's a real.
[0:19:55] There are real risks. But I think they, you know, I'm a huge believer in human ingenuity. And I think we will be able to solve those problems, both of those problems I mentioned, given enough time and enough rigor using scientific method and other things. So let's go back a bit to how we got here. We're 10 years since AlphaGo, when DeepMind beat the world's leading Go player. And actually, Sebastian, the way that you describe it, and maybe this was through you, Demis.
[0:20:25] The moves the machine was making were not just to win, but they were beautiful moves. Did that moment for you, Demis, change how you think about machine intelligence? And if so, how? It didn't change it. That's why we built AlphaGo and the way we built it. But we definitely crossed the threshold then. And I think the whole...
[0:20:46] The people were paying attention and understood that. So I think that's sort of the beginning of the modern AI era. Of course, there was deep learning before and transformers were still to come to be invented a couple of years after that. But I think the goal was always the pinnacle of...
[0:21:02] Let's call it sort of traditional AI. And games and AI have been sort of the testing ground games, I guess, for AI since Turing's day. All of the pioneers in AI tried to write chess programs and those sorts of things. But as we know, Go is much more complicated. It doesn't lend itself to rules and heuristics. So we always thought, Dave Silva and I were leading that project, that if one were to solve Go and crack Go,
[0:21:29] Then you must have done it in a way that would be very interesting for general intelligence. Because it couldn't be done in the sort of narrow chess computer, deep blue way, brute force way. And that's what it transpired with the AlphaGo match. Not only did we win, but AlphaGo famously made this move 37 in game two, which all the gold experts around the world were stunned by how original it was and how good that strategy was. And I was waiting for that moment to...
[0:21:57] Sort of think that the day we got back from Seoul, we basically started the AlphaFold project in earnest, and now AI for Science division, because I was waiting for that moment to show that, okay, now we have technology and algorithmic ideas that were...
[0:22:14] That were powerful enough to actually tackle real world challenges. Which is of course what we were building all of this for. I love games and it's really fun and a huge challenge to crack those things. But really they were not an ends in themselves. They were a means to an end. And the end was to do things like advanced science and productivity. And build things like AlphaFold. You could think of it as the first infinity machine, right? This was just an enormous search space with Go. There's a 19x19 board.
[0:22:43] If you think about that, it means that the first move, there's 361 options, second one 360, you multiply that out, you get to an almost infinitely big number of possible permutations of Go. And if you can build one machine that makes sense of that infinity of different combinations, that's a repeatable trick. Exactly.
[0:23:05] You talk about this as the sort of alpha moment for AI. Do you think that was the moment like 10 years ago where we should have realized, hey, this could be really big? Maybe we would be more prepared for the moment that we're in now? Well, I think it had a bit of that effect on me when I was thinking about what book to write next. I finished my previous one at the beginning of 2022, and I was looking around for the next topic.
[0:23:30] The fact that you had by then both AlphaGo in 2016 and AlphaFold, the protein folding program in 2020, you had these two instances of this almost infinitely large search space, and that had been solved by AI. And it felt as though this was already exciting, but it was going to be on an exponential curve upwards. And so I actually went to see Demis and pitch him on the idea of...
[0:23:56] Handing over large numbers of his hours of work, which are precious to him, to meet me in a pub or something. I pitched this in November of 2022, just a little bit before Chachapiti came out. Which is incredible timing, by the way. So there's a big message from the book. Listen up, everybody. The big message is... Would you still have done it? It's better to be lucky than smart, right? Because I got very lucky with the timing.
[0:24:24] You won the Nobel Prize with AlphaFold. Congratulations. I don't know if that I was saying, I don't know if that ever gets old. I mean, hey. Yeah, I mean, pretty amazing. Solving one of biology's biggest puzzles. What is the next AlphaFold or AlphaGo moment? Like, when is that next breakthrough? And what mysteries are you hoping to solve?
[0:24:53] We're working a lot of things that I think could be the next kind of AlphaFold moment. I think the thing I'm most excited about is the work we're doing with the spin-out company Isomorphic Labs that we spun out basically after AlphaFold was done, so in 2021, to accelerate drug discovery by at least 10x.
[0:25:12] So it takes an average of 10 years to go from a target to a potential drug compound. And I think we can compress that to months, maybe weeks, even days, using various AI systems. Obviously, AlphaFold and knowing the shape of the protein in question is one critical piece, but it's only one small piece of the whole drug discovery.
[0:25:35] So you can think of it, we need to build another half dozen alpha folds adjacent to what alpha fold was doing. So mostly actually we're working in biochemistry and chemistry. So you know the shape of the protein, you know where you've got which bit of the protein you need to attach to to affect what it does, its function. And then we're building these systems to kind of design the right, the perfect compound to bind to exactly the right place with the right strength and critically not to bind to anything else.
[0:26:05] And we're getting to the point now, and it's going really well, and it's very exciting work, where we can...
[0:26:13] Basically perform a virtual screen of a compound we're proposing to every protein in the body in a few seconds. So you can imagine how useful that is towards accelerating drug discovery. So the idea is, borrowing again from Sebastian's term, this infinity machine, again, it's using that same idea. There's almost infinite possibilities of potential drug compounds.
[0:26:38] And there's a needle in a haystack situation. There's a few that might be suitable. And you've got to find that somehow. And what we try to do is do the search in Silico. Because obviously it will be thousands of times faster than trying to do it in the wet lab. And you just reserve the wet lab step for the validation step. Because you still have to obviously validate your predictions in patients and make sure they're safe and all of those things and the predictions are correct. But really you want to just do that at the very last step.
[0:27:08] Rather than trying to have that
[0:27:11] Slow process all the way through the discovery. So that's the idea. It's going really well. So that's in biology. So we're deepening our investment and our ideas in the biology space. But I do think also AI is going to have amazing impact on other areas of science. Like material science. We've got exciting work going on there. Some mathematics and physics. And eventually almost every branch of science.
[0:27:40] So going, and by the way, the history that you tell in this book is really incredible. I said part thriller, part drama, a little bit of comedy. You know, early 2010s, Larry Page and Elon Musk are basically fighting over DeepMind. You sell to Google in 2014. OpenAI doesn't even exist yet. OpenAI started in 2015. You know, chat GPT then explodes.
[0:28:09] goes nuclear, pun intended. Did DeepMind miss something fundamental? How did that happen? You mean at the point when ChatGPT came out? You know, the decision to not, and you'll have a chance to respond to that, to not go the LLM route, to go the LLM route, and to, you know, build ChatGPT. Well, I think there are various reasons why DeepMind didn't do that. You know...
[0:28:35] One of them is a little bit what you've just been hearing from Demis, that there are other things you can do in AI, which is super exciting to do with medicine and biology. And so there are other things to do, right? Another, and I think this is what I remember being really, really interested by when we discussed it, is that there are different sort of ideas about what intelligence actually is. If full machine intelligence, how important is it for the...
[0:29:04] Intelligence to learn from trial and error in the real world, to be grounded in the real world, to have direct experience with the real world. And I remember sitting in the pub with Demis, and he picked up a glass of water, because we were not actually drinking beer, I want to make it clear.
[0:29:18] Coffee. So he had this glass of water and he said, look, I'm grounded in the real world. I know because of my experience, if I pick it up, I can feel the weight. If I drop it on the floor, it's going to make a mess. If I drop it on concrete, it's going to shatter. These things about the physical world are stuff that you learn by being in the physical world. And if you do this sort of experiment, if you downloaded the entirety of Wikipedia into a computer and it trained on that,
[0:29:43] Would it really understand what it felt like to pick up a glass of water? And maybe it wouldn't. And also, the breadth of human experience is so vast that can you really teach a machine to be intelligent like a human is just from the crystallized knowledge that people have put into writing? Don't you have to be more like a robotic?
[0:30:08] Intelligence or perhaps operating in a simulation, which could be digital, but at least you're moving around some simulation of the physical world. And I think, you know, as Demis explained to me, there's this field in neuroscience called action in perception, which argues that, you know, language is a set of symbols. So if you have a word like cat and you have a picture of a cat, you can relate one to the other. But without that groundedness, you're not really understanding what a cat
[0:30:37] Actually is. And so this was sort of a theory about intelligence, what intelligence is, what it would take to get there. And, you know, ex ante, in other words, in 2017, you know, five years before ChatGPD, it wasn't obvious that language was going to unlock as much as it turned out to do. And so I think that, you know, in some sense, you know, your neuroscience PhD, which, as I explain in my book,
[0:31:05] was very helpful in the early systems like Atari. There's Atari games which were solved in the first big DeepMind experiment. And things like how to think about memory, how memory interacts with experience. Intuitions from neuroscience were super important in designing that system in a way that made it work.
[0:31:29] When you go a few years later and you're faced with this question about language or intuition from neuroscience actually might have misled you. Or at least put you in a position where you had a hypothesis which was reasonable, ex ante, but turned out to be wrong, ex post. And so that's why DeepMind was slow at that point, but then it caught up. And in some ways, in science and even in business, you're always going to make mistakes. There's no such thing.
[0:31:57] Excellence does not consist in just avoiding mistakes. Excellent consists in recovering from the mistakes. Yeah, how would you assess this diagnosis there? Yeah, I mean, look, I wouldn't quite characterize exactly. There's a few different things, first of all, I think that are getting conflated. So one is whether you're talking about the product, ChatGPT versus LLMs. So actually, we and the other leading labs had very capable LLMs.
[0:32:27] So in fact, a lot of the scaling laws, they're called Chinchilla scaling laws, which in Chinchilla was our internal name for one of our language models. And it was a whole branch of maybe about a third of the group and the company was working on that. So one could argue it should be more, but it wasn't. We weren't doing LLMs.
[0:32:44] And they're a roughly comparable power. And actually, Google Brain, which was a separate, obviously, research group at the time, they had their LLMs. They invented transformers. And actually, there was even a product internally at Google, not at DeepMind, but at the time, DeepMind was a separate alphabet company that was doing research. So that's why we were doing AlphaFold, those kinds of things.
[0:33:07] We didn't put a product out because I think it's a case of actually being a bit too close to the technology. So when you're building it and you're immersed in it all the time, you can see all the flaws in it, right? So it was hallucinating, it was making mistakes. How could this really, especially a company like Google that's known for, what is Google known for? Correct information.
[0:33:29] That's literally its mission statement, organizing the world's information, and it's an amazing mission, and it's a very compatible one with AI. But you can see why there would be a dilemma there, and I wasn't involved in that decision, but you can see why there would be some dilemma, which would not be the same dilemma for a startup, of course, an ambitious startup. That's the whole point of startups. They haven't got a reputation they need to worry about, and they have to just try and risk things.
[0:33:56] And Fair Play to OpenAI, they took that risk. But I will also say, from the product side, they did not know that it was going to be this viral success. Otherwise, they would have called it something a bit more catchy, I think. It's a very programmer name. You can see, it's like a programmer's going, you know what, I'm going to chuck this out next week, and I'll just put chat in front of the technology, GPT. It's not, you know, it's so...
[0:34:23] And I think Sam and others have said that themselves, that they were not expecting it to go a million users or whatever it was in a week. So that was a wake-up call to all of us. And I'll come back to the technical part of it in a second about language and what it means about language. But from the product perspective, the earthquake was, oh, actually, we've all been doing this in the lab.
[0:34:45] And we're fascinated by the technology. And I would have said, and people were saying, maybe it needs a couple more years to cook and to mature. But in fact, if you just chuck it over the fence, it looks like...
[0:34:57] People will find a good use for it, you know, whether that's, you know, summarizing or just sort of using it as a bouncing board for ideas. Maybe it doesn't matter if it's hallucinating. Perhaps it's even good. So it turns out that obviously people are very ingenious and they found value in it.
[0:35:17] Even with the early versions of these chatbots, of course, now we've got these amazing things that are pretty comprehensive, still not AGI, but very comprehensive. And so that was the surprise. Okay, it's like, oh, it's ready to come out of the lab now, these ideas, and be used in the real world. And people will find it valuable, even though it has these flaws and these problems. So that was not, I don't think anyone understood that till it happened.
[0:35:43] And then we all understood it. And then obviously we then had to rush ourselves to have an answer on the product side. Now, going back to the science of it, the research, I think a lot of us, and me included, did think that language would be much harder to tackle.
[0:36:04] than it turned out to be. And part of that was actually not so much my neuroscience background, it was my upbringing in the 90s at Cambridge and then later MIT, so Cambridge, UK, and then Cambridge, Massachusetts, both Cambridges. They were both actually the seats of, in the UK and US respectively, of traditional AI. You can call it expert systems or logic networks. And that's what they both taught. And I was kind of steeped in that.
[0:36:31] And I never believed it at either place, right? So even in my undergrad at Cambridge, and I checked again my postdoc at MIT, and I was like, you know what, it's obvious why this doesn't work. Because if you're in logic land or abstract land, which language is, or at least it was thought to be.
[0:36:48] Then how do you know what these symbols really mean in the real world, which is what Sebastian was alluding to is symbol grounding. And that was the, if you ask me, what were the main top three problems in the 90s and the approaches that people took, very smart people at MIT and other places took to try and solve AI, why did they fail? Symbol grounding would have been top of my list.
[0:37:11] Because actually you're just building these abstract maths entities that have no ability to actually really understand the real world when they encounter it. So the classic example would be in these databases you'd have, you could program in a dog has four legs, okay? And that would be some predicate logic in the database. But then as soon as it saw a picture of a dog, how does it know that symbol dog maps to that messy image?
[0:37:36] And then even if you hand code that, then what about all the other dogs you could show it? And what is a leg anyway? And it all just starts to unravel. So that's when the opposite approach is like, it's got to learn from experience like humans do. And this is where reinforcement learning comes in, deep learning. And I think we were correct about that. But what was interesting was, and we had, by the way, all of the systems we built, we were ahead of our time because AlphaGo and all the game systems, they were agent-based systems.
[0:38:03] So we already had agents. We were working on agents from the very beginning at DeepMind.
[0:38:07] That's what we started working on, these active systems that would try and solve things in their environments. Of course, they were limited to games. Now we have general agents, which is what we're seeing, which is much harder and was not obvious if it could generalize. But part of that was we thought that systems would have to learn, just like animals and humans do, from real-world experience. And that's how we get the grounding of what do these symbols mean, whether they're math symbols or language symbols with the real world. Now, why didn't it turn out like that?
[0:38:36] And that was a big school of thought because I think linguists probably were wrong. So if you read all the famous linguists, Chomsky, all these people, which I did, and Pinker, all these folks, and debated with them, there's sort of...
[0:38:53] It turns out that language is not predicate logic, I don't think. It's actually got a lot of grounding in it. So if you probably, if you're a machine, like our AI systems, and you can look at everything that's ever been written on the internet, you'll see some subtle correlations and causations in the language structure. I think linguists probably need to catch up and look at that, maybe by analysing AI systems today. So that's one thing. The second thing is that what I didn't quite understand before all of this happened was that...
[0:39:21] With these systems, you don't just show them the internet. That's the first thing. Here's all the words ever written. Predict the next token. We all know that's how, predict the next word. But we also do this fine-tuning before they can become chatbots where we give them feedback, reinforcement learning feedback, where they say something and you kind of thumbs up, thumbs down it, or you correct what it said. And of course, it's humans doing that. Now AI systems are also doing that, but originally it was human feedback raters doing that.
[0:39:51] And we're obviously grounded agents, grounded systems, of course, by definition. We live in the real world. We learn it from experience. And it's slow and it's very low bandwidth, but we are providing, through our feedback, grounding to this.
[0:40:05] Otherwise would be an abstract system if we left it on its own. And so then the corrections that causes, you know, actually object next to other object. That's what this means. Here's the picture. And that's what it means. A bit like you would teach a child. And it's slow and it's a bit inefficient in terms of bit rate, but it is introducing grounding to your language system. So it's sort of, so language itself may be more grounded than we realize. It's not just an abstraction. And then we're also giving grounding through the feedback process.
[0:40:34] If the future is world models, if that's really the next step, which you can describe it, but it understands cause and effect in a more human-like way, not just predicting the next word, will we look back on this chat GPT-LLM moment as a detour? Or is that what progress looks like? I think that we work on world models. Others people do as well. People mean slightly different things by it. So our one is called Genie. And the key thing...
[0:41:02] that I believe is it's built on top of the foundation models. So not instead of. So Genie's built on Gemini technology and the other things that we have, Nanobot on, all these other things that understand visuals and so on. And then we build this sort of generative capacity, interactive experiences that you can have on top of that. Maybe more is needed to understand causality, long-term planning. These are unsolved problems. But I would be...
[0:41:30] Super surprised if that was a new separate track and it turned out that LLM foundation models were dead end. To me, the question is, the empirical question, I don't think we know the answer to this yet, is will they scale all the way to AGI on themselves? So, i.e., will they be the whole answer? Basically, that's it, but much bigger, much more scale with some incremental improvements.
[0:41:55] Are there one or two extra breakthroughs needed of the size of an AlphaGo or a Deep or a Transformers, something of that level? But I think it will be on top of what we have with the foundation models. So I think the only question, is it the only system or is it a key part of the system? So the idea that these enormous investments in data centers are going to be redundant, that's not right. Sebastian, it's hard to, you know, not a half second passes with...
[0:42:25] Without something changes. When you look at the competitive landscape, who's ahead, who's behind? Are we asking the right question? Are we all racing to the same finish line? So you're right, it changes a lot. When I finished writing the manuscript, which was towards the end of 2025, it was pretty clear from the independent assessments, these leaderboards, that Gemini 3.0 was the best model out there.
[0:42:54] I think in 2026, you know, Anthropic and Claude have had a moment. The agentic system they produced in January was widely adopted. And more recently, Mythos has created a massive stir. And, you know, that hasn't been released, but it seems from those who have assessed it, notably the UK government AI Safety Institute, that this is actually a sort of state of the art. It's a breakthrough in...
[0:43:22] What you can do in terms of both diagnosis of weaknesses in cyber systems and also then attacking them and exposing the weaknesses. So I'd say that Anthropic has had a good three, four months. But the leadership does change. I've written back in January an essay in the New York Times.
[0:43:47] Saying that I feel like OpenAI is running an experiment, not merely in frontier technology, but also in the depth of the global capital market, because the amount of burn is just off the charts. I mean, there was one internal document that got leaked, which suggested they needed $660 billion between now and 2030 to make up for all the investments they need to do, which won't be covered by revenue, even if revenue is ramping up very, very fast.
[0:44:16] The investments are so ginormous that you need 660 billion. That's an unprecedented, I mean, it's an almost unfathomable amount of money. So unless they make a lot of changes and cut back their ambitions, which to be fair, they have done recently. They've canned their video generation model, which was essentially a money sink for them. I felt like there were 50% chance of getting busted in the next 18 months. Wow. Okay. I mean, do you want to comment on that?
[0:44:46] Yeah, well, look, I'll comment more generally. Do you live in a constant state of paranoia about what you're going to wake up to? It's not paranoia because you know it's happening. There's nothing paranoid about it. It's just the reality. But it's the most ferocious competition probably there's ever been in capitalism. I can't think, I mean...
[0:45:09] Maybe people with longer memories than me can experience something. But I think in living memory of the tech industry, certainly the people I know involved in all the dot-com stuff and all of that, they say this is way more competitive. You know, I like sports, football specifically. I think it's like competing, you know, the champion leagues level with all the smartest minds and entrepreneurs and biggest, you know, tech titans all.
[0:45:33] Now understanding what I think we understood 20 years ago, that this is it, this is the biggest game, biggest prize, however you want to look at it, biggest technology. So that makes sense now, and it's attracting all the best talent. And part of me loves that, right? It's what I live for, competition, and I don't really feel alive, actually, in some ways, without that.
[0:45:58] Probably it's fortunate I have that part of me, which again comes to my chest probably and my youth, my training in these extreme competitive environments and learning to try and enjoy it as a way to dealing with the pressure, I suppose. But the problem is that in the back of my mind, I'm really thinking this is actually not actually the real...
[0:46:21] It's sort of a parochial problem. If you just step back for a moment from the frontier and it's ferocious and it's also exciting and scary and all of those things. And there's, you know, month by month.
[0:46:35] The three leading labs jockey each other, and it depends on which dimension you're talking about, because there's code, but then we're still better on image and multimodal, but code seems more important right now. So there's all these things, and I try to block out a lot of that noise, to be honest, because otherwise you'd go crazy. You just got to focus on what you're doing best, double down.
[0:46:53] Correct things if they're needed. This is fairly standard scientific process, right? And be confident in the talent we have and the breadth of research we've got and the pedigree we have. And we're fixing the things that are deficient in today. But that's a sort of ongoing thing. But what's in the back of my mind really is that we need to rise above all of this competition. And what's at stake is the future of humanity. I mean, it sounds a bit grandiose, but I think that is the case.
[0:47:20] And there's something much more important than the sort of capitalist race there is. And by the way, of course, on top of that, there's the geopolitical race going on. So you've got a kind of double level of very intense competition between superpowers and then between some of the leading companies. So it's a very complicated situation. But in the back of my mind, I've always got, which is why I can't...
[0:47:43] Fully enjoy just the competitive aspect is because actually there's something much more important than competition here is to get this right for the world and for the benefit of everyone, which is what I set out to do. And I still hope to be able to influence it in that way where we can, you know, I think we need to have some international standards, some coordination, at least on safety and sharing of ideas between the leading labs. And, you know, the lab leaders, we all know each other well, and I get on pretty much with all of them.
[0:48:13] We try and, you know, I think people understand what's at stake. The question is, what can be done about it? And I think we need to spend more time on that in the next few, you know, I was going to say years, but actually months, and think about that beyond and maybe set aside some of the ferocious competition. Sebastian, you were just in China. Is cooperation on this realistic?
[0:48:41] Is competition realistic? Cooperation. Yeah, cooperation. Yeah, so, sorry. Competition is a given. Cooperation is the question. Yeah, I went to China for eight days in March because the Chinese are always the fastest about everything, including publishing books, and so the book came out first in China.
[0:49:01] Even though they had to translate it. Then they wanted pictures of Demis. Then they wanted me to write a handwritten note to readers. Then there had to be like three prefaces from different Chinese experts. They did all this stuff. They were still a month earlier. So I go there and, you know, it is quite fun in China because they haven't learned to treat authors with the contempt that they deserve. So wherever you go, you know, autograph, autograph, and, you know, there's one guy kind of...
[0:49:28] Filming you and the other one is doing stills. And there's a poster which is sort of four feet high. And there's one of Demis, of course, but then there's one of me too, which is kind of surprising. And so I had a good time in China. But the serious thing about the trip was I did have a chance to meet with a lot of the senior AI academic researchers and the chief executives of some two or three of the big companies, tech companies. And what surprised me is that most of them spontaneously talked about safety.
[0:49:57] Without necessarily bringing it up. And while I was there, the agent OpenClaude was a big sort of excitement. Lots of ordinary Chinese were downloading it. And nearly all the people I talked to were more on the expert level or the company leader level. So this is crazy. You know, Claude is some piece of open source code. You don't know what it's going to do with your data if you put it in your laptop. You know, the government should really step in and do something about safety. And then lo and behold, a day or two later, the government did step in.
[0:50:26] And at least restrict the use of OpenClaude in government computer systems or state-owned enterprises. So I think the Chinese, you know, are perfectly rational about this. They see that AI has big upsides, that AI also has some dangers. They want to manage the dangers, which is perfectly rational. For the moment, you know, accelerating and competing is the dominant thing.
[0:50:55] And the West is doing. They're not going to slow down by themselves. I think both sides are caught in this prisoner's dilemma where we need to have conversations to break the race dynamic. And it may feel impossible in the current China-American standoff. The US has an export embargo on cutting-edge chips to try and slow them down. That certainly hasn't worked. But there's this environment of confrontation. But I would just point out that in the Cold War,
[0:51:25] It turned out to be possible both to have fierce competition between the Soviet Union and the West and to have collaboration on making nuclear technology safe. And in fact, 1956 was the year both when the IAEA was negotiated.
[0:51:41] The treaty for that was negotiated at the International Atomic Energy Agency, which was going to keep track of nuclear material and try to make sure it wasn't used for weapons. So this was actually concluded the same day, I think, that the Soviet Union invaded Hungary, leading to... So the Hungarian uprising started, leading to the Soviets rolling in with their tanks about two weeks later. And, by the way, about two weeks later, there was the Suez Crisis,
[0:52:11] During which the Soviets threatened to attack the West with nuclear weapons because of the provocation of Suez. So in other words, right in the depths of the competition and the scare of the Cold War, the International Atomic Energy Agency was created. And it's the same with the Non-Proliferation Treaty in the 60s, that you had the Cuban Missile Crisis on the one hand in 1962, but you had the NPT in 1968.
[0:52:39] So I think we need to get over this idea that China is the enemy. We can't talk to them. We're competing so we can never cooperate. In fact, you can compete and cooperate, and history shows us that. Google is a major investor in Anthropic. What is your relationship like with Dario, and what do you make of some of the decisions that have been made about not working with the U.S. government on certain things, releasing certain models that are, you know...
[0:53:11] Potentially too powerful to companies before the general public. And how do you think about the stakes at this level? We also have, you know, I've known Dario since we were both in academia and we get on very well.
[0:53:29] I probably know one of the best of all the other lab leaders. We have a lot of similar views on some of the safety issues, but also some of the potential. I think he's talked about a lot of the work we've done with AI for science has been an inspiration for him, which is very kind of him to say. And he's a scientist too, actually a neuroscientist as well, as one of the things, I think he's did physics too. So we actually have quite a similar...
[0:53:51] So there's a lot we share in common. I don't know the specifics of what he's been discussing. I think he's been working a lot longer with government. I do think we have a duty to help with democratically elected governments. I think we want them to get the benefit of our technologies. But we have to look at the exact specifics of what they're used for. And obviously, as everyone knows,
[0:54:21] My view we discussed is to try and use it. What I'm focused on and the things we focus our technologies on is advancing areas like science and medicine, and that's what we're building our tools for. That's what they're good for. That's what we sort of measure them against. That's what we train them against. And obviously general productivity with things like the Gemini app, which I think is getting more and more useful by the day. So those are the zones of areas that we are expert in and we try and build our tools for.
[0:54:51] But the world is becoming, as we all know, it's a pretty unstable place and the geopolitics are complicated and we're not immune to that. And areas like cyber, I think we can really help with cyber defense because if you think about Google, I think it's already probably the best at cyber defense of all the hyperscalers. We've just acquired Wiz, who are an amazing startup company in Israel that are amazing at.
[0:55:19] Cyber defense. And then we also have frontier lab capabilities on the AI side. So I think there's those kinds of areas where we can help a lot. I also think bio is going to be the next issue with these general models, biosecurity. Obviously, we're trying to cure.
[0:55:35] Cancer and cure diseases, but a bad actor could misuse these tools. Or we could have a natural event or lab leak, like we saw with COVID, that could cause a global crisis that maybe AI, I hope, can help with in a pretty emergency situation. So I think these are all the areas where I think we can be very helpful to the government and also national security issues like that. And so that's what I focus on and we focus on.
[0:56:04] Jario is one of the louder voices on risk, Elon Musk as well. AI has a PR problem, it seems to me. People are terrified, they're scared, they're confused. There's real-world violence happening. What should they be afraid of and what shouldn't they be afraid of? Well, I think people in the leadership position should need to be careful about how they talk about things because I think there's a lot of...
[0:56:33] Ways of discussing AI that...
[0:56:35] There's so much uncertainty in my view of what's going to happen next. That's the thing that we should be focusing on is reducing that uncertainty and then mapping out mitigations, quantifying the risk correctly in a sort of sober and scientific way, rigorous way, and then coming up with creative solutions to mitigate those risks. That's how I would couch things. That's what I've tried to do with my work. Also, I think as an industry, we probably need to do better.
[0:57:04] to focus more of our resources on unequivocal good use cases.
[0:57:10] We've tried to do that with AlphaFold and now Isomorphic Labs and things like this, but we're just one actor. It would be great to see a lot more of that happening and a lot more of that delivery, not just talk about those things. And so I think that's part of perhaps the image issue that it has. It's interesting, it's a US kind of specific problem, because when I came back from India, there was a recent AI summit in India, hosted by Prime Minister Modi, and there's an...
[0:57:39] Incredible enthusiasm in some of the other parts of the world, like the youth of India, for the opportunities and the potential that AI could unlock for them in being entrepreneurs or for science and academia. So it was very interesting to see the differences between the different countries and how they thought about this. So it does have an image problem, and I think we need to think about collectively, the lab leaders here, how to talk about this. And I think part of it is also...
[0:58:07] Getting together, I think there's a real need for some sort of conversation around standards or certification of these systems to...
[0:58:15] We assure the public that it isn't just an unfettered race to the edge, right? That no shortcuts are being taken and that defectors from that process or would-be bad actors are not rewarded for doing that because that's the prisoner's dilemma problem, right? If shortcuts are successful, then why would people not take those shortcuts? So I think...
[0:58:39] I think the time is coming rapidly where this has to be discussed very seriously. And I plan to be part of those discussions. And I think that will help to mitigate some of the, I think, reasonable fears that the general public have. I was going to say, the thing that worries me about the way the mythos news has, you know, that whole issue with Anthropic.
[0:59:04] Creating a short list of companies that will get its technology first so that they can build cyber defenses is that it's not really tenable for more than a short period to have one company, a five-year-old startup in California, deciding which companies with an internet presence out there get the privilege of defending themselves and which are kind of cast out to survive in the outer darkness. I mean, that's just not a tenable position for a private company to be in.
[0:59:33] And the alternative that's being mooted is that the U.S. government comes in and takes over the model or sort of exerts control over it and takes over that list. But actually, even that isn't at all comforting. We don't really want the government making kind of case-by-case decisions on who gets to defend themselves and who doesn't, especially since this is not just a U.S. problem. This is a global problem. There's going to be, within probably six to nine months,
[1:00:03] A company that has AI that can do a similar thing because that seems to be the trend. They are very fast followers. So the point is that we've learned from the last 10 years of AI history two things. First is companies cannot deliver governance, right? Both Google...
[1:00:21] And on the other hand, OpenAI experimented with various governance mechanisms to try to make AI safe. But this is a problem that's bigger than a company, and ultimately that doesn't work. And nor does it work to just do it on the basis of one nation. And it was encouraging before, you know, in 23 and 24, various AI safety institutes were set up.
[1:00:42] The US had one, Britain had one, Japan had one, Singapore had one, the EU had one. And then even actually in the late 24, there was a summit in San Francisco of all of these different AI safety institutes who were coming together to kind of be a bit like a proto-IAEA, you know, back to that nuclear analogy. And then it kind of just got pushed aside. And we're in full-on race from 2025 onwards.
[1:01:08] And I think we need to go back. It's not impossible to do it. I mean, 2024 showed that you could make progress in that direction. But it's been pushed to the side, and we need to get back to where we were. All right, we're going to try to get to a couple audience questions before we go. I like the first one. It takes me back to the last chapter of the book, which I just think is so lovely, because it's almost like you're making this...
[1:01:30] Beautiful discovery about what Demis is really after and the mysteries of the world and the universe that he really wants to solve, which, of course, we've barely scratched the surface of here. How many hours did you guys talk? I lost count after 30. Yeah, yeah, exactly. So if you didn't solve it, then we're not going to do it now. But what happens after AGI?
[1:01:52] Well, if everything goes well and the dream of it was to, and I think this is still, the future is still to be written by what happens in the next couple of years, is we have this sort of maximum human flourishing. I think there's been, there's quite a few great science fiction.
[1:02:10] I can imagine we solve all disease in the next decade or two. We come up with some wonder materials that may be room temperature superconductors, optimal batteries, perhaps we solve fusion and better magnets. So we have renewable, free, clean energy.
[1:02:36] Which would then solve a whole bunch of other problems like water access. We could create as much rocket fuel as we want by catalysing sea water. So suddenly that should open up space and mining asteroids, these kinds of things. Sounds pretty science fiction right now, but I think in the next 20 plus years that will happen. And then we're in a post-scarcity world. So then the question comes, can we share the proceeds of that?
[1:03:04] In a way that everyone benefits. So it's more of an economics problem at that point, assuming we got the technical things right. And then there'll be philosophical issues, which I think will need great new philosophers to sort of discuss things like purpose and meaning. And I think it will be a new era of maximum human flourishing, but it will be a new era of the human condition. So I think that's what lies in front of us.
[1:03:30] I think it could be a very exciting future, an amazing future, but there are many other possible paths too. So on that note, I'm a mom, you're a dad, you're a dad as well. This is another question from the audience. What jobs would you tell college kids to prepare for today? I would add all kids, what should they be doing?
[1:03:54] Well, what I say, I mean, I get asked that obviously a lot when I give lectures at universities and sometimes schools is I think I look back to my childhood in a way of where the dawn of the computer age and video games and things that I think the adults of the day probably didn't understand. And I knew there was magic there. And I just immersed myself.
[1:04:15] Fully in that world, along with learning the usual technical things of maths and STEM subjects and so on. And I would probably recommend the same thing again here, which is immerse yourselves in these amazing new AI tools. I think the frontier is moving so fast. There's a huge overhang of, I mean, we're in an amazing situation where the general public is basically getting hold.
[1:04:37] And if you're a kid, somewhere, actually anywhere around the world, you're getting hold of a frontier model that's only maybe three or six months later than what's actually being done at the companies. So it's kind of incredible, really, the pace of...
[1:04:53] That dissemination of that technology is almost unheard of, I would say. And so what that means is there's enormous opportunities probably applying that technology to all sorts of other areas that we've barely scratched the surface of. So I have a very strong feeling that's the case.
[1:05:08] And I think over the next 10 years, it's hard to look out more than 10. I think the kids of today, if they use the tools in the right way and use it to kind of enhance the way they work and learn and the way their brain works, they're almost going to become superpowered in terms of what they'll be able to do, whether that's writing games or creating films or whatever it is that they might want to apply it to, or even doing science.
[1:05:31] If they're native, you know, kids of today will be first generation that grows up AI native, like I did computer native. And though, you know, I'm sure the best of them will do unbelievable things. Sebastian, in this post-scarcity future, do we journalists and authors still have jobs? What do you think? Do we still need books? So right now, as you know, the systems aren't that great at writing anything longer than an email. But, you know, presumably that's going to change.
[1:06:01] So then what happens? I suppose the part of my job which involves talking to other humans and understanding them, understanding their motivation, what makes them tick and why they're doing something, that probably remains because that's a human-to-human interaction and no matter how charming an LLM might be, I might have a slight edge on that. But then I wouldn't write so much because the system would do some of that for me and it would be a partnership maybe. I guess that's how I see it.
[1:06:31] So, you know, in closing, at the end of the book, Demis, you're quoted as saying that you personally, you're in a paradoxical moment. On the one hand, your quest is on schedule. The timing has lined up with what you expected. On the other, the landscape is not necessarily what you had imagined. You thought maybe you'd be able to do this in a single independent lab, but here we are in a global race. What are the knowns on the other side of this?
[1:07:01] And what are the unknowns? Well, I suppose when I was thinking about this 20, 30 years ago, I was hoping when we would get to this moment like this a few years out from AGI, there would be time to, first of all, kind of understand and maybe even savor the technology that we're building and kind of really appreciate, try to both appreciate and analyze.
[1:07:25] What it's capable of doing in, I suppose, a kind of philosophical and scientific way and then collaborate. I was thinking of a kind of CERN model where, you know, the best scientists in the world are helping each other and checking each other's work and coming up with ideas that can be stress tested by the other great scientists in a way that was sort of fitting for the moment that we're about to enter.
[1:07:51] It's not like that, and I'm also very practical as well as being a scientist, and that's maybe my engineering side, is we've got to deal with the world as it is today, and I think what I'm going to try and do is my job to try and help with the influence that I have, make it more considered, more scientific, more rigorous, and more thoughtful as we get to these final points towards this destination. And then...
[1:08:20] We're going to also need to start thinking now what's going to happen after that, even if that goes well, which was one of your earlier questions. And that itself needs a lot of work now. And actually, of course, not just from technologists. It needs to be economists, social scientists.
[1:08:36] And academia. Everyone should be thinking about, and I try and encourage that, some of the consequences for their own domain expertise about what should happen next. And do we need a new form of economics? There's lots of questions that I have if some of those things turn out to be true, like the post-scarcity world.
[1:08:56] And so I'm going to try what I can do from the platform I have to try and influence it in that more thoughtful, considered manner, but still very optimistic. Obviously, I'm still optimistic now. I think that's the last sentence of the book that Sebastian chose to be the book, the last sentence of that things will go well from here, but only if we give ourselves the time and space as humanity to get it right. Like, I just sort of think this is for the next thousand years. Why rush?
[1:09:26] What's the difference between the five or 10 years if it's for the next thousand years? And that's the sort of feeling I hope to convey to my peers and my competitors. And I think I'm still optimistic. People are realizing that something needs to be done and we need a plan for that. And I hope to be able to help with that. In closing.
[1:09:49] When are you going to take your first summer holiday? There are some chatbots that can help you plan these things. Yes. I think, you know, post-AGI now. So I'm going to have to wait a few more years. 2030 and beyond. There was a moment when you went on a honeymoon. Yes. But you took a very large stack of neuroscience papers with you. Yes. Yes, exactly. But it will be a hell of a gap year when it happens.
[1:10:18] You know, I'll be ready for it. All right. Demis Hassabis and Sebastian Malaby, thank you so much for doing this. So thank you all for coming here tonight. You've been listening to Sir Demis Hassabis and Sebastian Malaby in conversation with journalist Emily Chang. This program was recorded at the Sydney Goldstein Theatre in San Francisco on April 20th, 2026. City Arts and Lectures programs are recorded live in San Francisco and you can be there.
[1:10:53] Jose Andres, Ben Rhodes, and Laverne Cox are just a handful of our upcoming guests. To find out more, visit cityarts.net. These broadcasts are produced by City Arts and Lectures in association with KQED Public Radio, San Francisco. Executive producers are Kate Goldstein-Briar and Holly Mulder-Wallin. Alexandra Blackman is Director of Design and Communications. Assistance provided by Mark Spiro.
[1:11:29] Sydney Goldstein Technical Director, Stephen Eckert. Our post-production director is Nina Thorson. The recording engineer is Jane Heaven. Theme music composed and performed by Pat Gleason. City Arts and Lectures founding producer is Sydney Goldstein. City Arts and Lectures programs are supported by grants for the arts of the San Francisco Hotel Tax Fund. Additional funding provided by the Bernard Osher Foundation, Barry Traub, and
[1:12:00] The Friends of City Arts and Lectures. Support for recording and post-production of City Arts and Lectures is provided by Robert Mailer Anderson and Nicola Miner. To attend a live program or for a list of upcoming guests, visit our website at cityarts.net. That's cityarts.net.