[0:00:00] The human brain is the only existence proof we have that general intelligence is even possible. What breakthroughs do we need on this path of like solving all diseases? If we have like unlimited compute, could we predict the future? So if you and I time travel to 2050,
[0:00:13] What does it look like? This is Demis Hassabis. He is leading the race to invent superintelligence. Demis committed his life to this 30 years ago when most people thought creating true AI was impossible. But Demis is in most people. He's a childhood chess champion, a neuroscientist, and as of last year, a Nobel Prize winner. Every chapter of his life has prepared him to create true artificial intelligence. And now we're closer than ever before. So in today's episode, I'm going to ask Demis questions he's never been asked before.
[0:00:42] Oh, wow. And here are his vision for the future so you can build the next big thing. That was really fun. Thanks so much for coming on the show. Thanks for doing this. Of course, yeah. To prepare for this one, I watched every single interview you've ever done. Gosh, okay. Which is awesome. And in a lot of them, you mentioned your scientific heroes and your mind. So I just wanted to say thank you. Oh, thank you. Thank you. That's very kind of you. I mean it. I want to start in your autobiography. Yep.
[0:01:09] There's so many different moments that kind of lead to this threadline of your life thinking about intelligence. And I feel like when you started DeepMind, or even further back studying AI, a lot of people didn't believe in it. What were your unconventional beliefs about the world that gave you so much conviction? Well, actually, to be honest with you, I just thought it was...
[0:01:27] A, one of the most fascinating problems you could spend your life working on. But really, it was my expression of doing science. So when I was a kid, I wanted to understand, I was fascinated by all the big questions, nature of reality, nature of consciousness, these things. I felt like they were kind of staring us in the face. And even the best scientists haven't made that much progress in answering these questions. And I felt that we maybe needed some help.
[0:01:50] One of the things that you've talked about is that if you can make an AGI system, then we can understand the differences between the human brain and an AI brain. What are the main differences you've noticed so far?
[0:02:13] Well, it's interesting. That's exactly right. So things like consciousness and even intelligence, like what are these things? Creativity, what are the processes involved in that? I think one of the difficulties, obviously we've made some progress by doing neuroscience and doing fMRI and studying our own brains and animal brains. And that's why I did my PhD in as well. And so we've made some progress, but we sort of lack a reference, a comparator that can say, oh, this sister, this thing, this entity is intelligent.
[0:02:42] I think AI also could be used for studying neuroscience too and as a comparator as well. So I think building AGI and then analyzing it will be one of the best ways to understand our own minds and the sort of deep mysteries of our own minds.
[0:03:01] Yesterday, you talked about how solving all disease is actually not that far away for us, like we're closer than we think. My cousin's getting a PhD right now and uses AlphaFold every day. Oh, amazing. And I feel like you guys have had so many AlphaFold moments. Yes. What breakthroughs do we need on this path of solving all diseases?
[0:03:19] I think it's what we're working on at Isomorphic Labs really, which is AlphaFold obviously was really helpful with like protein structure prediction, which is critical for one of the things that you need for drug discovery, but only it's a critical piece, but it's only one piece. So you can think of Isomorphic Labs, we're extending that into, obviously we're improving AlphaFold, but we're also extending it into biochemistry and chemistry to kind of understand like what compounds should you make? Where do they bind on the protein?
[0:03:49] And then also checking how your body reacts to those proteins. You know, how does it absorb them, those compounds? How does it absorb them? Are there any toxic side effects? So you're trying to minimize all those things and predict them and to make cleaner compounds and cleaner drugs. So there's sort of like a, I guess like half a dozen really big challenges there and that we're working on. And you can kind of think of like, we're going to try and put them all together. And then you have your drug discovery platform. What do you do for the proteins that are like unstable?
[0:04:17] Like Avofold works amazing for protein that has stable structures. Yes. What do you do for the ones that are unstable? Well, actually, so, I mean, the hypothesis is that those in, you know, disordered regions of those proteins, intrinsically disordered, they actually do.
[0:04:29] So we actually got to be able to predict.
[0:04:54] The dynamics of a protein, including sometimes disordered regions, in order to kind of, you know, design the right type of drug or the right type of antibody or whatever it is that whichever disease, you know, kind of profile we're looking at. And so we're trying to extend our model to be able to deal with that type of complexity. It's one of the big challenges of biology. There's so much complexity. Interesting. And so is it like we would use AlphaFold or some more technology to...
[0:05:22] Understand the proteins. And then that would cut down clinical trial time. Yes. So no. So the idea is to, what we're focusing on at the moment is just the drug discovery phase. It can still take, you know, many, many years, even a decade for to do that, where you go from the, you know, the target of interest, the biology understanding of that. And then you need a candidate compound that you want to take into clinical trials. So we're focusing on shortening that from years down to like months, maybe even one day weeks, which would be incredible.
[0:05:52] Seems sort of unthinkable right now, but I do think that's the same as thinking about protein structures back 10 years ago. Everyone thought that was impossible. And now here we are with all of the structures folded. And I think that's going to happen again with drug discovery. Then there's the second question of can we also speed up the clinical trials part? And that's hard. I mean, there's regulations. There's lots of things that are involved in that that aren't necessarily to do with the technology.
[0:06:17] I do actually think AI can help there too in terms of like stratifying patients and making sure they get the right test compounds, analyzing that data, maybe being more, if you're more accurate at predicting what the side effects were, then maybe you can jump more quickly through the dosage steps and things like that. So there's a lot of things in clinical trials that could probably be compressed as well once we had a better idea of the design of the drug compound in the first place.
[0:06:44] Something else that people also thought was impossible was the models understanding physics. And then yesterday you guys have Gemini. Yeah.
[0:06:51] And, okay, the coolest thing ever to me about this was that you didn't actually train it on physics, right? It just learned from videos? Yes, exactly. How did that work? Yeah, it's pretty mind-blowing, really. And, you know, the more videos you give it, also Gemini, of course, is under the hood, so it has understanding about the world, and it can label things and stuff like that. And Gemini, from the beginning, has been multimodal, so it's sort of really good at understanding scenes. And then we had to make it dynamic with Omni, and it's kind of mind-blowing that it does seem to pick up.
[0:07:20] Pretty accurate versions of intuitive physics. We are starting to create physics benchmarks where the future versions of Omni, OmniPro and so on are going to be like, see how well do we actually do on exactly marble runs and things falling down and gravity. But right now it's actually just amazing what it can do sort of almost in an emergent way, understand these different things. And it's the same thing with nano banana for image, and it means it opens up.
[0:07:49] Hopefully for creators like you as well, loads of possibilities where it's just incredibly easy to edit things because it understands what parts of the scene are. Yeah, I remember in a previous interview when you were asked about AGI, you said that Not A Banana and Veo were actually the closest thing we had to it because of this deep understanding. Yes. It's also interesting, AGI has become this term that there's a lot of discussion around.
[0:08:13] And I think your predictions on when we get to it are more conservative, but then what you think it will do is a lot more ambitious than other people. Yes, that's right. Can you bring me into your mind, how do you think about AGI and why does it matter in that sense? Yeah, well, I think it's, for me, it's a little bit of a neuroscience analogy, which is that the human brain is the only existence proof we have that general intelligence is even possible, right? So that's why we know it's possible because the human mind has invented all this amazing modern civilization that we have around us, including science.
[0:08:42] So it's incredibly general and that's why we're so adaptive. We can adapt to new technologies, we invent all these technologies, even our brains will evolve for hunter-gathering, you know.
[0:08:50] So that's kind of amazing. So I think we won't know we have a true general intelligence unless it has the capabilities that the brain has, because that's the only existence proof we have. And so that's a pretty high bar. And it means that it's probably higher bar than just being able to do some useful economic work or other things I've heard other people talk about. But I've been pretty consistent about that. The type of test I would have is like an Einstein test of, let's train one of these systems with a 1901 knowledge cutoff.
[0:09:20] Can it invent special relativity like Einstein did in 1905? And if it could, then what that means is...
[0:09:27] You could apply it to today's physics and say, okay, come up with the extensions of string theory or answer what's going on with dark matter to one of the hypotheses. And maybe it would be worth us investigating what it came up with. Because obviously, whatever theory it comes up with, it's going to take a lot of time and effort to test those theories, especially in physics. So you want to be pretty sure that those systems are actually capable of coming up with something sensible. And I've always said that coming up with a hypothesis,
[0:09:56] It's harder than solving an existing conjecture, even though that's impressive already. But actually coming up with some kind of problem that will be worthy of study and worthy of effort, that's even harder. Asking the right question is the hardest thing in science. And is that the true creativity of the AI system? Yeah, it's very related to what I would call true creativity, exactly. Can it come up with something truly novel and actually a leap forward rather than something incremental?
[0:10:23] So one of the ways to test that is the Einstein test. Are there other ways to test it? Yeah, you could. I mean, so another example I often give is with AlphaGo. So, of course, AlphaGo famously came up with Move 37 in game two of the match, right? And that was a novel strategy in Go, for sure. And it's changed the way Go is played. And it was amazing because humans have played Go for a couple of thousand years now. It's the oldest game ever invented, most complex game ever. And yet AlphaGo was able to invent new strategies.
[0:10:52] But the thing I say is that's not enough. What you'd actually want is something, a future version of AlphaGo to be able to invent Go, right? Invent a game as deep and as complex and as elegant and as beautiful as Go, not just come up with a strategy within the game. And so, again, I don't think today's systems are yet capable of doing that, but I think they will be able to in the future.
[0:11:15] Okay, this is like a wild question, but I'm just curious. I feel like as a person, there are times when I'm grappling with an idea, don't know what I'm doing, then I go to sleep, and I wake up the next day, and I feel like a thousand times better.
[0:11:25] Is there like a sleep on it mode for AI? Yeah, well, I think that, well, definitely that's been shown to be the case with the brain for sure. So they've done amazing sleep studies where, you know, people have a little nap even, and they've been given a problem, have a nap, and they're way better, statistically better than the people that didn't have the nap. And so your brain's definitely doing a bunch of work while you're sleeping, including things like memory replay. So replaying back things that you were pertinent during the day or recently, and then the rest.
[0:11:54] And then what I think that's about is incorporating the new knowledge into existing knowledge that you have in an elegant way. And then that obviously can then yield new insights, right? These aha moments you have when you wake up and say, oh, I've solved the problem. And people like Tesla very famously used to say he used to submit things to his subconscious so that they would solve it when he was asleep. I mean, he was an unbelievably brilliant guy, very eccentric, but a brilliant guy. And I think there may be
[0:12:23] What would be a need for something similar like a sleep mode or a consolidation mode is sometimes called of like, okay, how do you incorporate all the visual and input data you've seen today into like only a small fraction of that's actually useful. So it's kind of wasteful to store it all like we do today in the context window. How do you actually extract the useful bits and then incorporate that without overwriting your existing knowledge? Interesting.
[0:12:49] Okay, because your PhD, I think one of your main learnings is that memory and imagination are really interlinked. Like you looked at people that had like a damaged campus and then they had less ability to imagine a scene. That's right. In a lot of your video games, there was like this whole element of like simulation and like scene design. Yeah. How important is that to AI? Because I remember, I guess like with AI right now, we can simulate weather. Yes. But how do we simulate all the other things? Yeah.
[0:13:11] Exactly. So it's what we can simulate whether we can also, to some extent, I mean, eventually, I think we'll better simulate some biology. I love this idea of virtual cell, which is the simulation of a cell that would be accurate enough, you could kind of do virtual experiments, and it would tell you something really useful. And then there's things like material science and other things, even economics one day, I think simulations are going to be a really vital kind of component for us and AIs to understand the world. Because if you want to understand something very
[0:13:40] Thanks for watching.
[0:13:56] Just rerun lots of controlled experiments because it's basically a real world emergent system. So the only way you are going to be able to make a good decision is and gather data about that is to run lots of simulations. A bit like AlphaGo did with like Monte Carlo simulating out lots and lots of moves and lots and lots of plans and then back aggregating, okay, which one's the best one? I think you could think of if you had a lot of accurate simulations about these complex emergent systems like the weather or economics or biology, you'd be able to use sort of
[0:14:25] Can we not simulate the economy now?
[0:14:35] I don't think so. I think it's too complicated. I mean, people try, but the problem is economy is probably the most complex thing of all because it involves human beings, which are already very complex, and organisms like corporations, which are combinations of humans, and then those corporations are doing things with nation states. So it's sort of the most emergent, complex emergent system of all in a way. So I think it's pretty hard, as far as I know, of anyone making kind of direct simulations of that. But I think...
[0:15:05] You might be able to learn a simulation of that one day. And in fact, one of my favorite books, Asimov's Foundation Series, one of the characters in that was able to sort of predict the future by kind of aggregating human behavior in the aggregate. And then you could say something about the average situation. But obviously, I think no human mind's going to be good enough to be able to do that, but an AI might be able to.
[0:15:29] If we have unlimited compute and we were able to put in all the information of the world right now, could we predict the future, hypothetically? I think you might be able to predict the consequence of a decision now.
[0:15:39] Right. So that would be like a pretty dream thing. Yeah. I mean, I think if you think about the way economics is done at the moment, it's kind of very ad hoc, it feels to me. Like you sort of, okay, you know, your economy's doing this. You've got a few stats on it. And then you make these massive macro decisions that only later, you know, maybe five years later, you kind of like, oh, actually, maybe that decision wasn't very good. We caused a recession here or something there. And they're massive decisions and livelihoods depend on that.
[0:16:08] But there's no way of really testing that at the moment of the decision, right? To counterfactual is very difficult to do. So it's also hard for economists and politicians to improve. It's why it's called a social science because it's not a science because you can't repeat the experiment with the exact same conditions because it's in the real world and it's emergent. But with a very accurate simulator, you might be able to. That's so exciting.
[0:16:34] The other thing that really excites me and I actually talked about this with Sam Altman last week about like basically I think something that people don't think about with AI is that
[0:16:41] It's probably the thing that people talk to the most. And so you have like this really cool opportunity to reshape someone's worldview. Like what work choices you use with your AI. You can inform someone to be optimistic and work hard. How do you think about it? Like what are the important personality traits? Yeah, I mean, I think we have to be very careful with that because obviously, you know, it could go in bad directions. But I think that for the moment, we're building what I think of as really smart tools. So things that are extremely useful for a particular purpose that you want.
[0:17:10] So does that mean that everyone right now has like the same personality of...
[0:17:33] Gemini? Yes. I mean, there's personalization. So we're bringing that in. You saw that a lot yesterday in IO. And I think that's the direction that users want. So you don't have to keep telling it, what do you do? What's your family context? All of that. You just want to get advice about something or help with your planning or admin or whatever it is. So it's clearly going to make it more useful if it's personalized. But it still can be viewed as a personalized tool that just has a lot of...
[0:18:01] Prior information that you wanted to have so that you don't have to repeat yourself each time. Clearly, that's going to be useful.
[0:18:08] I feel like you study personality a lot. I'm surprised that that's not more exciting to you. I just feel like it's a huge... No, it's very exciting. I mean, I think if you think about the person... Well, personalization is super exciting, but also the persona of the... If you mean of the actual system itself. Yeah, we think about that a lot. And it's sort of implicit right now to train it with the reinforcement learning we do and the post-training about, okay, we have certain values and we want it to behave in a certain way and be helpful.
[0:18:37] And I think there's actually...
[0:18:58] More research that's needed there. So it's very interesting in terms of persona research and even things like the big five personality factors and other things like that. Maybe there's better personality models out there that could be applied to the AIs as well.
[0:19:13] Yeah, it's kind of cool that the work that you're doing now is going to unlock all these like new scientific fields. Like there probably could be another scientific branch where we like analyze personalities. Exactly. In a much more detailed way than we're able to before. I have found for sure that I think that's going to happen. I think that's the exciting thing right now when I talk to students is if they think creatively and imaginatively about the situation, there's so many, like even maybe even tiny new branches of science to be opened up.
[0:19:37] Agree. Yeah, we could like, could we create like an fMRI equivalent of understanding like the AI machine? Exactly. So I just think there's so many opportunities and potential there. So if you and I time travel to 2050, what does it look like? Bring it to your mind. Like, what's the dream? Wow, 2050. I mean, that's...
[0:19:55] That's a long time away given how fast things are improving. But what I hope on that kind of time scale is that we've got AGI safely over the line for humanity. We've worked out how to evolve economics so that everyone widely benefits from the increased resources and productivity this is going to bring. Hopefully, we're in a kind of post-scarcity world. And then the next obvious step to me is that humanity kind of...
[0:20:20] It goes to the stars and we have this maximum human flourishing. And by 2050, hopefully we should be having this interview on one of the moons of Jupiter and building Dyson spheres and really realizing, kind of waking up the universe with human consciousness in the way that Carl Sagan used to poetically talk about and scientists and writers like that. And I think, and there's lots of good sci-fi about this, the culture series by Hugh Banks and so on. So that would be incredible.
[0:20:49] Do you imagine that people will be...
[0:20:55] Using AI to like jet fuel their jobs and like still working on like traditional things? I think so. I think, I mean, 2050 is a long way out. But I think over the next 10 years, I feel like it's going to sort of, almost everyone has access to pretty much the most cutting edge technology, which is kind of amazing thing, really, right? You know, basically, almost only a few months behind was actually in the frontier labs. And I think the kind of next generation is going to be able to be they're going to
[0:21:24] And I'm really excited to see what they're going to do with these tools to kind of superpower themselves. You know, the sorts of things you'll be able to do will be incredible and would have taken teams of 10, 50, you know, to 20, 30, 50 people before. And so I think they should unlock a lot of creativity. So I think there's going to be a lot of change, but I think anytime there's a lot of change and disruption, that also means there's huge new opportunities for people who are kind of imaginative and creative enough and who lean in to what these tools are going to do.
[0:21:53] These new tools can do. So I think that's what's going to sort of happen over the next decade or so. Okay, my last question for you. You're very famous for like your 1 to 4 a.m. work session. And you said that kind of like during the day you're CEO and then at night it's like more of like research and stuff. Yes.
[0:22:09] What's like the most common thought in your head at that hour? Okay, well, this it sort of rotates a bit depending on how whether I've got an active project like an alpha fold thing going. So usually my most fun things would be to work on a science project or research project myself. But other times it's about thinking through, you know, how the next stage more philosophical issues around.
[0:22:30] How to get AI beneficial for the world? What kind of collaboration is needed between the leading labs? But also, how can we maybe get things like international standards and cooperation around AI? Because I think that's going to be vitally, you know, urgently needed in the next few years. Love it. Thanks so much for coming on. That was really fun. That was awesome. You're the best. That was so fun. Thanks for that. We covered a lot of ground.