Une nouvelle ère de découvertes : l’IA et les frontières de la science avec Demis Hassabis

[MUSIC PLAYING] [APPLAUSE] MIKE ALLEN: I'm Mike Allen, co-founder of Axios. Thank you to the Google I/O team for making these conversations possible. Demis Hassabis started as a chess savant and a gamer, now co-founder and CEO of Google DeepMind, where he's on the front lines of both the scientific race and the business race for AI. DeepMind's landmark AI breakthroughs include AlphaGo and AlphaFold, which solved the 50-year grand challenge of protein structure. For cracking that code, Demis was awarded the Nobel Prize in chemistry. "The Economist" last month put Demis on the cover as a Greek god, complete with tunic, in a pantheon of the Titans, who are shaping all of our future. He's even been knighted the mind behind DeepMind. Sir Demis Hassabis, welcome to Dialogues. [CROWD CHEERING] [MUSIC PLAYING] DEMIS HASSABIS: Hey, Mike. Good to see you. MIKE ALLEN: Thank you so much for doing this. DEMIS HASSABIS: Thanks. Yeah, thanks for doing it. MIKE ALLEN: So, Demis, yesterday, at your keynote, where you got, by the way, big applause and even whoops for announcing that the Gemini for Science, which is one of your favorites-- but you left us with the ultimate cliffhanger. You said, when we look back at this time, I think we will realize we were standing in the foothills of the singularity. DEMIS HASSABIS: Mm-hmm. MIKE ALLEN: Take us one layer deeper. What do you think is going to happen? What do you hope is going to happen? DEMIS HASSABIS: Yeah, I thought people seemed to pick up on that line specifically. MIKE ALLEN: Ya think? [LAUGHTER] DEMIS HASSABIS: Yes, so it was kind of expected. But it's great to be able to expand on it a little bit. So I'm talking a lot about, obviously, AGI being on the horizon. I think we are only a few years away now from the full version of that. And the singularity is a kind of good word, I think, to describe the era that is precipitated by a technology like AGI arriving on the scene. And partly, it's because it's going to be so transformative, I think the most important invention ever, that it's hard to make many predictions beyond that horizon because it will change so much. And I think, this year, with the agents, agentic systems that we're all seeing and using, I think we can start feeling it now. And that's what I was sort of meaning by that. When we look back, I think, in five, 10 years' time, I think we'll see it as like, oh, yes, 2026, '27 is when it was starting. MIKE ALLEN: Yeah, and I love that we're starting to feel it. DEMIS HASSABIS: Yeah. MIKE ALLEN: Expand on that a little bit. How do you feel it? DEMIS HASSABIS: Well, I think all of you-- I mean, I've been feeling it just by using our own coding systems and coding agents and just vibe coding things, even little game prototypes in my spare one or two hours in the small hours of the morning, which would have taken six months before. And I just think that kind of acceleration is very interesting to see. I think we're seeing it in areas of science, obviously with AlphaFold, which was a few years ago now, but almost every branch of science AI being applied to. And I'm also seeing the speed at which we're making advances in things like Isomorphic Labs for drug discovery. So I think taking that all together and looking at the current progress of the foundation models and the chatbot systems and things like Omni, all of that, taken together, is just astounding progress, I would say. MIKE ALLEN: Yeah, and based on the markers you set over the years, I feel like we're kind of right on schedule. DEMIS HASSABIS: Yeah. MIKE ALLEN: You say, a few years. DEMIS HASSABIS: Yeah. MIKE ALLEN: What is your latest marker? DEMIS HASSABIS: Well, I feel like it's-- I've been saying, recently, around 2030, plus or minus a year, I think is a reasonable estimate, from what I'm seeing now. MIKE ALLEN: So 2029 to 2031. DEMIS HASSABIS: Mm-hmm. MIKE ALLEN: So DeepMind's original mission statement was solve intelligence. How are we doing? DEMIS HASSABIS: Well, I think we're doing pretty well. So our original statement was, step one, solve intelligence. Step two, use it to solve everything else. And I think that was pretty prescient. That's what we felt AGI would be is a general purpose technology that could help. I was thinking more about science and medicine, specifically, in increasing our understanding of how the world works around us. That's always been my passion, as you know. And I think we are on the cusp of that now. MIKE ALLEN: In the lifetime of the builders and searchers in this room and watching us around the world, do you think intelligence will be solved? DEMIS HASSABIS: Yes, well, I mean, look, solved, it's just a neat term. What does it actually mean? But what I meant by that was building AGI. And I think we're a few years away. So that would be the first step to me. MIKE ALLEN: And with AGI, different people talk about it different ways. DEMIS HASSABIS: Yeah. MIKE ALLEN: But I think you're one of the people that thinks that there will be a moment and that we will know. DEMIS HASSABIS: Well, no, I think that's not the case. I think it will be-- it's going to be a gradual improvement. I think that's what we're seeing. Probably a pretty fast improvement. But I have a particular test in mind, which I call kind of the Einstein test. So imagine you had a knowledge cutoff of one of these leading frontier models of 1901. Would it be able to come up with his incredible sort of insights and discoveries that he did in 1905, including special relativity? Could a system come up with those insights and breakthroughs? And I think today systems clearly can't. But I think I don't see why, in the future, they won't be able to. So that's the creativity or leap of insight that I'm still looking for from these systems for it to be full AGI. MIKE ALLEN: Lots of buzz among tech writers in Press Park, right here at Google I/O, about Google having its groove back. Brian Chen of "The New York Times" posted yesterday how Google is starting to win the AI race. DEMIS HASSABIS: Mm-hmm. MIKE ALLEN: If you win, it will be, why? What is the Google edge? DEMIS HASSABIS: Mm-hmm. MIKE ALLEN: And what changed? DEMIS HASSABIS: Look, I think it's sort of weird, this sort of idea of, we're back, and then we're winning, and then we're not back. And the thing is that, at least the three frontier labs and actually the whole of the frontier, it's the most ferocious competition there's ever been, I would say, in tech history, maybe in corporate history. So there's that going on. Very capable, brilliant people in all the organizations, each trying to push, as hard as they can, on the research and the breakthroughs. I think we're doing really well for what we need to do, both in terms of the underlying technologies. We have a kind of broader research bench than I think the other labs have. We've always had that. I think that's what gives us our innovative capacity to invent the next breakthroughs. You see, now, new models that are coming out-- our 3.5 Flash Pro coming soon. We're very happy with the progress that we're making on the underlying models. We also have all our generative media models, Omni being the latest. So that's going well. Where I think things are going extremely well is on the consumer side. So I think that's what the really big leap has been, for us, in the last year, is streamlining our tech stack, actually almost rewriting it from the ground up, to be AI first and now agent first. That's a big change for a incredible engineering organization, like Google, that serves billions of users every day on multiple product surfaces on our incredible engineering stack. But that has to adapt so that you can ship the model, the goodness of the latest models, directly into all of those products at massive, massive scale. And I think we've got that down really well now, better than any of the other, shall we say, big tech companies. And I think you're starting to see that in the progress we're making with the adoption of Gemini app-- 900 million monthly users now, incredible growth there that we're seeing-- and also things like AI Mode in Search, an unbelievably popular and incredibly useful, I think, version of AI for information seeking. So I think if you look across our surfaces and what we're doing with our models, especially the small, efficient models that are still very powerful-- so that's really important for our internal use cases because we've got to serve billions of users. But it's also very important for some of our big enterprise customers as well. So I think when you take a look at that, at aggregate, and then you also can see the advantage we have being full stack, all the way from the chips, to the data centers, to cloud, to frontier lab, to billion-user products that have AI embedded in them, I think we're the only organization in the world that has that full stack. And we're continually trying to optimize that feedback loop so that they all feed into each other. MIKE ALLEN: Do you think you will win? DEMIS HASSABIS: Well, I think we're going to be one of the big winners of the era. MIKE ALLEN: How many do you think there will be? DEMIS HASSABIS: Well, I don't know. I don't really see it like that, in terms of win-- win to what, exactly? So I think the AI era is going to be immense. It's going to be unbelievable amount of new opportunities created, new economic progress made. And we've got to make sure that we're one of the key players in that new post-AGI world. And I think we're incredibly well positioned for that. And as to how many players there'll be, my guess is a handful, given how powerful the technology is. But actually, stepping back from that, what I always say is, although we're ferociously competing in this kind of race, all the leaders of the frontier labs need to think about-- and they are thinking about-- the bigger picture of the responsibility to the world. As we bring this unbelievable technology into the world, we want it to be for the benefit of everyone and the benefit of humanity. And it's something bigger than all of us and the competition is remembering we've got to take these final steps in a careful, thoughtful way and make sure that we're building it with safety and security in mind, as well as enabling all of the amazing, good use cases that we hope to see. MIKE ALLEN: Demis, picking up on your word "opportunity"-- and the 5,000 students and professional developers here on campus are living proof of that. DEMIS HASSABIS: Yeah. MIKE ALLEN: And yet a lot of America doesn't see it, feel it. Front page of today's "Wall Street Journal," the AI lash building. Right? American rebellion. Eric Schmidt getting booed. Why is that happening, and what can you do about it? DEMIS HASSABIS: Well, I think a lot of the ways that some of, let's say, my peers are talking about this I don't really agree with. I don't really agree with the style of communication but also some of the substance too. So what I would say, the way I'm looking at it is, there's going to be no doubt huge amount of change. So something this transformative, I've been on record saying, if you want to try and quantify it, it's going to be 10 times the impact of the Industrial Revolution, maybe 10 times faster-- so over a decade, instead of a century. And that might be an underestimate. Probably is. MIKE ALLEN: Wait. 10 times the impact and 10 times faster? DEMIS HASSABIS: Yeah, so you can think of it as 100x the Industrial Revolution. MIKE ALLEN: Yes, yes. DEMIS HASSABIS: Maybe an underestimate. So that's huge. But what I say about that, the future is not written. It's very uncertain. Are there more jobs going to be created? Are there different ones? We don't know. Nobody knows. None of these experts, so-called experts, know. And the future is to be written, actually, by the next generation. I'm actually very excited when I talk to students and when I go to places-- like, we just came back earlier this year from the India AI Summit. And the youth of India are so excited about what they can do, what these tools will enable them to do-- build companies, build applications, compete at the global scale. So this technology is actually democratizing opportunities for what you can build. What would have taken a team of 10, 50 experts to build, maybe a smart kid in their garage can build it themselves. So when I think back to the era I grew up in, with computers and home computers-- and then there was the internet and the amount of opportunities that then created for kids of the day, students of the day, who grew up native with that technology, to actually go on and build the future that we see today. And I think that will happen again with the next generation that's going to grow up AI native. And I think, for the next 10 years at least-- and it's hard to see beyond that-- they're going to be super powered. As long as they're technically trained enough and then they bring creativity and taste and judgment, these kinds of skills, to the table, I think that's going to be the next era. So I'm actually very optimistic about human ingenuity to find amazing ways to use-- and we're seeing that all the time-- to use these new tools. MIKE ALLEN: Your hunch would be net more jobs, or you can't even say that? DEMIS HASSABIS: I think there will be changes to the job market, as always there is, with any new revolution. But I don't see why there wouldn't be more opportunities. Maybe it will be more like creating startups, one- or two-person startups. More entrepreneurial skills will be needed because you have these incredible creative tools. So there's going to be changes. But I think, overall, it will be better. And then the other thing I would say on the backlash is that I think it's incumbent on the field, our field, the AI field and industry, to show the unequivocal benefits more clearly and more concretely. And that's what we try to do with our AI for Science work and applying AI to advancing science and medicine. So that's why I've spent my whole career working on using these AI tools for things like AlphaFold, as you mentioned, and drug discovery. And I think there's so many areas of science that AI can be applied to. And we're trying to do that-- material science, mathematics, energy sources. So that's what I think we need to show more of. And I hope our other peer organizations do more of as well. MIKE ALLEN: Now, Demis, taking your point about style of communication, your other fellow Greek gods on the cover of "The Economist," you're proudly a scientist. You're not a doomer. DEMIS HASSABIS: Mm-hmm. MIKE ALLEN: You're not an accelerationist. Why do some AI leaders talk about the negatives of AI? How does that help them? DEMIS HASSABIS: Well, I'm in the middle. So I call myself a cautious optimist. MIKE ALLEN: Cautious optimist. DEMIS HASSABIS: So I'm optimistic. Obviously, I'm optimistic-- otherwise, I wouldn't have been doing this for 30 years-- that the benefits are going to far outweigh the negatives with something like AI, as long as we take our time, we give ourselves the time, as society, to get this next few years right. We're talking about the next hundreds, maybe thousand years of human flourishing, if we get this right. MIKE ALLEN: And all of us are living that. That's amazing. DEMIS HASSABIS: But there's a lot of challenges to get that right. So we've got to give ourselves the time to do it, rather than just racing to just get the next thing out. And I think that's where we've got to put our attention to and come together to discuss the bigger impacts that all of this work is going to produce. MIKE ALLEN: We have a treat, a clip from "The Thinking Game," a remarkable documentary about Demis and DeepMind, filmed over five years. We have a fascinating scene from your old offices in London. This was not rehearsed or recreated. Totally serendipitous. Cameras just happened to be there. Let's watch. [VIDEO PLAYBACK] - Now we had a tool that can be used practically by scientists. - These people asking us, I've got this protein involved in malaria or some infectious disease. We don't know the structure. Can we use AlphaFold to solve it? - We can easily predict all known sequences in a month. - All known sequences in a month? - Yeah, easily. - Mm-hmm. - A billion, 2 billion. And they are-- - So why don't we just do that? Yeah, we should just do the lot. - Well, I mean, now you can really-- - That's way better. Why don't we just do that? - Well, so that's one of the options. - Right. - There's this-- - We should just run-- that's a great idea. We should just run every protein in existence and then release that. Why didn't someone suggests this before? Of course, that's what we should do. Why are we thinking about making a service and then people submit their protein? We just fold everything and then give it to everyone in the world. Who knows how many discoveries will be made from that? [END PLAYBACK] [APPLAUSE] MIKE ALLEN: So a quote for the ages, an iconic moment. We saw the DeepMind working, Demis, when you said, so why don't we do that? So why don't we just do that? Tell us what was in your DeepMind and what was happening. [LAUGHTER] DEMIS HASSABIS: So look, this is the-- that particular example-- and actually, I'll just get back to your last question as well about acceleration. This is the type acceleration that we want in my opinion. So you couldn't get any more accelerations, in a sense, than what I spent my whole career doing. I think there are so many challenges facing society today, from disease to energy, our environmental ecosystems, climate change. We need technology like AI to come and help us solve these other problems. And the number one, in my opinion, as I said yesterday, is applying AI to improving human health. With AlphaFold, protein folding and predicting protein structures is one of the critical pieces of developing drugs to help with these diseases-- only one piece but a critical piece. And what was happening there is I just sort of worked out because AlphaFold was-- literally, in that meeting-- because AlphaFold was so accurate. But it's also so fast. So it could fold a protein in a matter of seconds. And then I knew there were a couple hundred million proteins known to science. And just the back of envelope calculation, I could feel that if we put a certain amount of computers on it, we could fold all of them in a year. So there was no need to set up the normal way of doing this, which is like a server and have people submit their genetic sequences, the amino asset sequences of the protein they cared about. And then you'd go away and fold it for a week and give them back the structure. We could just put all of it out there, for free, into the world, on a huge database run by the European Bioinformatics Institute in Cambridge, and just let everybody around the world, all the researchers around the world, access whatever protein they wanted, almost at the speed of a Google Search. So that, to us, felt like-- it was obvious, to me, that was the biggest impact we could do with this breakthrough. So I think that's an example of great acceleration. But on the other hand, I'm not in the camp of the Pollyanna camp of, well, there's no risk here. Just let it rip. There's no danger. Of course, when we're bringing in something this momentous, as something like AGI-- so a new era we're going into, a new-- we have to think that through, very carefully. And there are going to be challenges and risks. And there'll be ways which we can try and mitigate that but, again, only if we give ourselves enough time, collectively, to do that. MIKE ALLEN: The fact that cameras followed you for five years, the fact that you're willing to let them in, you recognize that you're on the frontier of something that will shape humanity. It's hard to see ourselves as others see us. What did you learn from the documentary, the process, or actually watching it? DEMIS HASSABIS: Well, I actually hate seeing myself on camera, so it was a hard thing to agree to. But it's the same film crew that did the "AlphaGo" movie-- if some of you have seen that, for the AlphaGo match. And they came along right at the last minute to do that. And I thought it was a really moving documentary recording of quite an important moment in sort of modern AI history, the big match we had against Lee Sedol. We just celebrated the 10-year anniversary of that, which is kind of unbelievable, really. In some ways, it feels like a century has gone by. In other ways, it feels like yesterday. And then they proposed to, basically, be a kind of fly on the wall. And we trusted them with that. And they happened to be around at key moments like that one. It was kind of amazing. They happened to be in that meeting because then they come over, every now and again, and record a few things. And one of the reasons we agreed to do that is we felt that this is a societal-level transformation, this technology. We felt that people would want to know something about the people behind it, what their motivations were, why were they doing this, what do we think the good that can come from it. And I hope that documentary shows that. It seems to have gone down well and I think does portray-- they've done a good job of portraying what our motivations and passions are, things like AI for Science. MIKE ALLEN: Demis, a quick tour of your mind and world-- simulations and robotics. What's going on and why are simulations such a key milestone? DEMIS HASSABIS: I think simulations are really important. I've been sort of enamored with simulations as long as I've had with AI. In fact, the first real professional thing I programmed was "Theme Park," this game I wrote in the '90s, which was both AI, had AI in it, and it was a simulation game, one of the first for both. It came out in 1994. And so I've always seen them as quite complementary, so both in the sense that you can use AI to learn simulations about complex phenomena-- so for example, our WeatherNext system that we showcased yesterday that helps predict the paths of hurricanes and other extreme weather events, much more quickly and more accurately than traditional systems. But also, it can provide data for AI systems. So for example, we've been using Genie, our interactive 3D model, to test Waymo cars in very unusual one-in-a-billion-type situations that you would never see in the real world. So for example, what if you're driving down a road and there's a forest fire around you, or a biplane has done an emergency landing on the motorway, or an elephant's appeared in the road? These kind of things, which you're never going to see in your everyday training data, but you want to know, how is your system going to react? And so these types of simulations are very useful for testing that. And if your system doesn't do well there, you can use it to generate more data that you can learn from. MIKE ALLEN: OK. This is going to be your favorite question. I know you love to talk about the scientific method. AI for science and health-- and you've talked about human health as the ultimate use case for AI. DEMIS HASSABIS: Yes, so I think that the AI is the ultimate tool for science in the limit. It's why I started this whole journey was to try, and as a scientist, understand better the world around us. And I felt that our minds, we could use an incredible tool to help even the smartest experts more quickly navigate all the complexity of the data that we collect and try and find connections and insights and structure in that data. And what better tool than AI to do that? And I think that's coming to fruition now. And it's just a very, very exciting time to see the pace of science pick up in all sorts of fields-- many fields where we've been stuck for 30 or 40 years, like we were with protein folding. And I think we're going to see a lot of those breakthroughs over the next decade or so. MIKE ALLEN: AI in the real world-- what are some of the ways that people are already feeling it and is part of the AI lash that they don't realize it? DEMIS HASSABIS: Well, I mean, we're using AI all the time in pretty much every product that we use. So everything, from our point of view, Maps to Gmail has now been enhanced and is in the middle of being upgraded with AI. What I hope to see is with the next generation of assistants-- so you saw us talk about Gemini Spark yesterday but also the glasses, the smart glasses-- is, eventually, this will give us back time. That's what I hope. I hope that the smarter our AI assistants are, the less admin we need to do, the less time we actually need to spend with a lot of the technologies that are mundane and actually give us time back to focus on what we think is important, whether that's creativity, our families, and just connecting with other humans. That's, I think, in the next few years, what I hope will happen. But we can still stay connected and get the information that we need to do our jobs or that interests us but more conveniently at the time that we want because we'll have an AI assistant that's helping us manage all of that. MIKE ALLEN: If you look at all the keynote announcements yesterday here at Google I/O, the keynote put a bow around a lot of it. But it was also a lot to take in. Ina Fried at Axios, had a great headline about AI everywhere. Hit pause. And what's something from these announcements that we should remember, focus on? DEMIS HASSABIS: I think the main thing is that the agentic era is really here. And I think what we were really excited about this I/O is how we're integrating that across the Google ecosystem. So with Workspace, with all of the things that people use every day and love, how is that going to interact, securely, with an agentic system? I think we've seen the potential of agents, but I think we're just scratching the surface of what they can be, from coding to all the way to managing your admin load. I'm very excited about all of those things. And then just, obviously, the potential of these world models to help in the real world, with things like-- hopefully, we'll be talking about this next year-- some big advances in areas like robotics. MIKE ALLEN: And responsibility-- how do you think about that in your work? Does it inhibit? Does it accelerate? How do you think about it? DEMIS HASSABIS: I think it's a critical-- it's not really about inhibiting or accelerate. It's a critical part of everything we do. When you're working on this kind of technology, which we've been for a long time, and you're thinking about the implications, both, obviously, the enormous, amazing benefits that we're hoping for-- but that comes with-- it's a dual-purpose technology. Those things can be also repurposed by bad actors for harmful ends. So those two things weigh heavily on us, or should do, that are at the forefront in the frontier, and understand both the negative and the positive consequences of what this technology can do. MIKE ALLEN: And the last stop on our tour, future vision. Tell us what you know is behind the curtain and what we should be watching for in, let's say, coming months. DEMIS HASSABIS: Well, the big thing that everyone's working at the moment on is this idea of self-improvement. So of course, we're seeing a sort of a slow version of that, with coding agents helping our own engineers get better and more productive with coding. The question is, is like, what's the next level of that? And I think coding and mathematics are quite special cases because the outputs are verifiable. You can verify whether they're correct. And you can also generate as much synthetic data as you want. So I think there's evidence that this is a kind of compounding flywheel. So I think that's a very interesting space that all the leading labs are investing heavily on. And then I think the other area is just to build the reliability and robustness of these systems and consistency, really, of how these systems work, tool use, and instruction following, and make sure that they're reliable over long-term plans. So I think, again, that's going to be the next step change for these things. And part of that is making sure that they're robust and safe and secure. I think the more powerful these agents get and the more we delegate tasks to them, the more we have to make sure, from a cybersecurity standpoint, they're doing exactly what we're expecting them to do-- and we can monitor that-- and they're being done securely. MIKE ALLEN: OK. So that's a good frame way to think about step changes. And you pointed to consistency, right? What would our step change buckets be for, say, 2026? DEMIS HASSABIS: Well, consistency, we shouldn't be seeing-- as we get closer to something like AGI, we shouldn't be seeing systems that are brilliant at certain things when posed in certain ways, but they're still tripped up. If you present them with a simpler problem, say, in mathematics, but you present it in a certain way that it hasn't-- it's a little bit more out of distribution. And I think that shouldn't be the case for an AGI system, a general intelligence. That's not the case with human intelligence. If you're good at one sort of game or one kind of a task and you give that person a strictly simpler problem in the same sort of domain, they'll be able to solve that problem too. But there won't be this kind of jagged intelligence. I think another big area that's missing and we're focusing on, too, is continual learning. How do you have systems that learn after their training, after they've been released out into the wild? How can they adapt and learn further? And that's also related to personalization and making sure that the systems are kind of personalized to the extent that you want, as the user, to your particular context. MIKE ALLEN: Last one on the future. A year from now, at Google I/O 2027, what do you think we'll be talking about? What will be the centerpiece? DEMIS HASSABIS: I think there'll be a lot more agents in the real world, kind of embedded in our workflows. I think we're starting to see that now. I think, over the next year, that will mature. So I think that will be a huge thing. And I hope, also, there'll be big advances in AI in the real world. So I'm hoping we'll see some big leaps forward in applying these kinds of models that understand the world, Gemini robotics, the SOTA robotics model out there, right now, but applied to many different form factors in robotics. MIKE ALLEN: And AI in the real world-- I think something you've talked about as an example, Hurricane Melissa-- DEMIS HASSABIS: Mm-hmm. MIKE ALLEN: --and how AI can be used in weather forecasting? DEMIS HASSABIS: Yes, so that was one of the examples we showed yesterday of an incredible way that we can use AI and simulation. So weather, I think it's a very complicated, dynamic system, involves solving a lot of equations in the traditional way, Navier-Stokes equations, fluid dynamics. So these kinds of traditional systems take weeks to run. And they're kind of limited in how far ahead they can look. And for a situation like an extreme weather event, like a hurricane, you want to know-- and as we showed in the video yesterday, even a day ahead of time is very important for emergency services and reducing the damage to human lives, and also, property that these extreme weather events come with. So I think that's just one example of the big strides we're making with applying AI to learn simulations from available data, rather than directly programming the mechanics of a simulation by hand. MIKE ALLEN: So your advice to your younger self-- in an interview that you gave in connection when you were awarded the Nobel Prize, you said believe in your ideas, even though they can seem quite outlandish. And you said you, personally, had done that. How has AI made that even more true? You said having confidence in your ideas, from the very earliest stage, makes all the difference. DEMIS HASSABIS: Well, I think you need to have confidence in your ideas on anything big if you're going to stand a chance of achieving it. That's very clear. But I also think that it's not really just about confidence. It's also your passion and your belief in how important that that work is. So when I talk about AI, for example, obviously, we're here today where AI is the biggest thing that the world's talking about. And I think that we always imagined, back when we started DeepMind in 2010 and before then, that that could happen. But I would have also been still working on AI now, whether or not that had happened-- so if it was still a more of a niche, you could say, academic subject-- because I think it was the most fascinating problem and most important problem to be working on. It's worked out. The timing was right. We thought, back then, that it was a kind of 20-year mission. We're sort of exactly on schedule. But I would have been doing it anyway. So it's a little bit beyond confidence. It's a sense of almost obsession with that problem and it being the right problem and the most interesting problem, at least for my mind and the way that my mind works and the passions that I had. That was my expression of trying to advance science was to build this amazing tool that was also an interesting scientific artifact, in and of itself. So what more an interesting thing could one spend your life and career on? That's how I felt 30 years ago when I first started on all of this. MIKE ALLEN: Yeah, so, Demis, how you chose that road-- and historians will study this, but Sebastian Mallaby is out with a great book, "The Infinity Machine." And he said that you spent 30-plus hours, more than 30 hours with him at a London pub. I will say that you might be the first biographee in history to spend 30 hours and not regret it. But this book boils down your choice to-- you chose to use your mind, your skills to understand the universe. Tell us that road taken and roads that you didn't take. DEMIS HASSABIS: Yeah, I should say, we weren't drinking in the pub. It was just coffee for 30 hours. MIKE ALLEN: Irish? DEMIS HASSABIS: So no, it went fast because we actually got on really well. And I mean, I actually wasn't counting the hours, but he suggested that. And we had a very fun discussion around-- very wide ranging on all sorts of topics. And for me, I started off on this journey because I've always been fascinated by the big questions in science, the nature of reality, the nature of consciousness. And I've always been sort of almost puzzled, actually, that when we go through life-- and most of us-- we don't really stop to really ponder those big questions, even though they're staring us in the face. Like, what is time? We don't know what time is. We have some kind of hand-wavy idea about it. It's a direction of entropy and it's one way. But we don't really know what it is, which is kind of astounding to me, let alone gravity, quantum mechanics, all of these mysteries. And so I've always been fascinated by that since I can remember. And normally, when you're fascinated by those questions, you go into physics. So physics was my favorite subject at school, but I started reading biographies of all the greats. My favorites are, like, Feynman and, obviously, Einstein, Weinberg, all of these types of brilliant physicists. And what I realized is that-- or at least for me, post-war, post-1970s, we hadn't really made much progress, in my opinion, on-- we made a lot of progress at the beginning of the 1900s, like Einstein and Heisenberg and so on. But then, towards this latter part of the 20th century, we hadn't made that much progress. And I was kind of wondering why this was. And I felt that it was like, maybe we're reaching the limit of what the human mind can comprehend and understand, even the most brilliant minds, like people like Feynman, without the aid of some new tools. So that was my real feeling. That was my kind of takeaway, I suppose, from reading some of those biographies. And then combining that with reading a lot of science fiction-- Asimov's "Foundation" series, Iain Banks's "Culture" series, and then philosophy books like "Godel, Escher, Bach"-- which was very formative for me, reading that book-- it all kind of came together in that, well, we could try and directly answer some of these questions about, What's the nature of intelligence and consciousness? by building AI and comparing it to the human mind, the human brain. But also, it would be a tool, itself, that could be used for science and to advance the frontiers of what we know. So it was sort of obvious to me, 30-plus years ago, that that would be the most interesting and perhaps most valuable path forward. MIKE ALLEN: Inspired by the point you made about biographies, who's a scientist, living, working today, who doesn't work for Google, who either should get more attention, or you admire, or future biographies will be written about? DEMIS HASSABIS: I mean, there's so many great scientists out there. I don't know if you're asking me who's sort of underrated. But I do feel like a lot of the big biology scientists-- I mean, they're well known, right? They've won Nobel prizes, like Paul Nurse and Jennifer Doudna. These are amazing scientists inventing CRISPR, these kinds of technologies. And I think the next generation, I can't wait to see what they're going to do with the AI tools and make their own way and make their own big breakthroughs, with the aid of these amazing tools that we have. MIKE ALLEN: Demis, we're going to finish on a useful note for the builders and future builders in this room, watching around the world and will watch this conversation in perpetuity. You spend most of your time in London. But when you come to the Bay Area, you feel a palpable builder energy, which is exciting for all of us. When we visit, a lot of people here are living it, creating that. But you also have a warning for all of us. You detect a freneticism. DEMIS HASSABIS: Yeah, it's amazing energy. I mean, I spend about half my time out here and half in London. It's always great to plug into the energy of the builder and entrepreneurial ecosystem out here. Obviously, it's second to none. And it's incredible to see that enthusiasm and that embracing of the opportunities. But I would also say I do also detect, as you say, this kind of frenetic rush as well. And I think that that's not conducive to thinking deeply about hard topics for a long amount of time. So I think a lot of the next advances are going to be in deep tech using AI, maybe applying to another area of hard science. I think there's a lot of opportunities there. And you're going to have to go at that for many, many years before there's a sort of result. So I think having some sort of quieter, reflective moments where you can think very deeply about what you're planning, not just rushing in the next thing that you see, the kind of-- I think a lot of the times, speed is favored over the direction. And making sure you have the direction right is often more important than just the speed. Speed is important. Velocity is important. But if you're running 100 miles an hour in the wrong direction, that's actually worse than standing still and taking a moment to think about getting the direction right. So I often think there's an obsession, a bit of an obsession, out here with just velocity. And it's critical. So I'm not saying that's not important. But more critical is the direction of that velocity vector. MIKE ALLEN: Very useful. Thank you. So at Axios, we always finish with one fun thing. So I wanted to ask you about longevity escape velocity. The idea is, if you and I can live five more years, say, we'll live forever. How much stock do you put in that notion? And what are you doing to make sure that you make it? DEMIS HASSABIS: I don't really worry about that kind of question. I'm very pragmatic when it comes to these things. So what I'm trying to do with Isomorphic and what we try to do with AlphaFold is use these technologies to cure terrible diseases. That's the number one thing I can think of using AI for. Now, when it comes to aging, there's, I know, lots of aging researchers and scientists. There are lots of companies trying to work on this. There's lots of theories about what it is. Is it the culmination of all diseases? Is it a disease in itself? Does it need to be reprogrammed? There's a lot of theories out there. I think that we should start with just by solving existing terrible diseases we know about. That will help us live longer and more healthy lives. And then let's see what's remaining. And then we'll take it from there. MIKE ALLEN: And what is your personal health obsession? So sometimes you have a whoop, right, or what is-- DEMIS HASSABIS: Yeah, sometimes. Yeah. No, well, I try to stay as healthy as possible in the very few spare minutes I don't have. I don't sleep enough. That's one of the main issues I've got to try and address. MIKE ALLEN: What's your average sleep? DEMIS HASSABIS: Well, I try to get six hours, but sometimes I fail. [LAUGHS] MIKE ALLEN: Thank you to Google I/O. And, Demis Hassabis, thanks for an incredible conversation. DEMIS HASSABIS: Thanks, Mike. Thank you. Thank you. [APPLAUSE] [MUSIC PLAYING]