[0:00:01] Because I am super excited for our next conversation with Google DeepMind's head and Nobel laureate, Demis Hassabis. Demis, thank you so much for being here. One of the things everyone tries to do at Davos is figure out where things are going, what's coming next. And in AI, things change so fast. One thing you've said is that we're moving from pattern matching systems to ones that can reason and plan and learn more like humans. When you look ahead,
[0:00:35] As far as I guess one can see these days, two or three model generations, what do you think AI systems will be doing that they aren't doing today that would really surprise people? What do you see on the horizon?
[0:00:46] So I think there's a few things that, I mean, the systems today do a lot of impressive things. We've all used them and there's some things that they can't do. There's flaws that still need to be addressed. I think in the kind of areas that we're looking at are things like continual learning. Can they continue to learn out in the world after you finish training them? That would be extraordinarily useful for being useful agents or useful in the workplace. They need to be able to do that. Also things like long-term planning, true creativity.
[0:01:15] None of those kind of capabilities are there yet, and maybe scaling up existing techniques will get there, or perhaps one or two new breakthroughs will be needed. And you mentioned kind of the nerdy term is recursive self-learning models. You know, a lot of people have said, you know, that's one of the next things. So these are models that, you know, today we get new models every couple months, so we feel like they're learning, but the models themselves aren't learning. And that is a direction that's seen with a lot of promise, but also a lot of risk,
[0:01:45] There's less control when it's learning out there in the wild. How do you see that balance? Is that one of the things that worries you at all, or you feel like we can have pretty good control? Well, we kind of pioneered a lot of these systems many years back with our AlphaGo, AlphaZero series of work. It was in the limited domain of games where these systems would learn to play Go or chess by playing many, many games against themselves.
[0:02:11] Different earlier versions of themselves and then self-improve. And of course, that's incredibly efficient. So the example I always used to give, you start with basically the rules of chess in the morning. By lunchtime, it's better than a master. And then by the evening, in one day, less than 24 hours, it's better than the world champion level. So it's quite extraordinary to see something like that improvement curve in real time. Now, obviously...
[0:02:35] The real world's way messier, way more complicated than a game. So the question is, can those kinds of techniques be transferred over to real world useful domains? And it seems like, my guess would be that in coding and mathematics specifically, that may be possible because you can verify the answer.
[0:02:53] So that's the thing about games, maths and coding is when the system proposes an idea or move or conjecture, you can kind of validate, it can validate that and decide if that's the right decision. So that could be useful in itself, but it also comes with risks. If you close the loop on self-improvement, so there's really no human in the loop, obviously today, even with the coding assistance we have, you still need human coder to make the decisions and the architecture decisions. Then you also have to worry about the risks of these things.
[0:03:22] The capabilities proliferating in a way that you didn't want. And what's the hard limit preventing us from having self-learning models today? As you mentioned, AlphaGo was a self-learning model, but for a super limited domain. As you look to incorporate this attribute, what makes it hard today? Is it, as you mentioned, the fact that only certain domains can you really learn and prove?
[0:03:45] Yeah, there's two issues, and we're working on both of them. There's first of all what we sometimes call a world model. So if you're going to plan in silico, sort of plan in the mind of the AI system, like here's an idea for a move, what's going to happen in 30 moves time? Obviously in a game it's trivial because you know the rules of the game, so the transition matrix is clear. But in the real world, it's not.
[0:04:11] We don't have a full world model. So for example, if you want to plan, say you're controlling a robot, and you want to plan some long-term plan about how to get to here to the Congress Center, that involves a lot of different planning that you have to do at different levels. And it would be easier if you had a realistic simulation of the world, and you could plan in that simulation.
[0:04:33] Before taking any action in the real world that's maybe irreversible. And the robot slips on the ice or something. So I think we don't have the right world models yet. And then the second part is what I mentioned earlier is verifying that plan. Are you in a good end state? Even if that plan worked.
[0:04:50] Is that a good end state? Obviously, for a game, that's also very easy because you can evaluate the position, count up the number of pieces, this kind of thing, see if you won or lost the game at the end. But most real-world things are not that easily verifiable. I want to shift gears a little bit to another topic that's coming up a lot in conjunction with AI, especially on the consumer side, which is advertising.
[0:05:14] Again, it opens up a lot of possibilities. When OpenAI recently announced they're planning to bring ads, their argument is we want to deliver more services to more people in the world if they don't have to pay and we can use advertising. Of course, that's what's funded the entire consumer internet. There's a lot of ways to do advertising. One of the things I think about and worry about is these machines are becoming so good at persuasion.
[0:05:42] And right now, if you look at Gemini, Today's Chat, GPT, the others, their only stated interest and seemingly is to help you to answer the question. As soon as you introduce advertisers, you have another customer. And I'm curious, how do you think about advertising? It would let you get Gemini to more people, but it comes with these risks. How do you think about it? Yeah, no, it's a great question. Of course, we're thinking very carefully about this, like others in the industry.
[0:06:12] You know, a little bit surprised they've moved so early into that because I think, as you say, I mean, look, ads, there's nothing wrong with ads. I think, as you say, they funded much of the consumer internet, but, and have done well, they can be useful to the user in terms of finding things that they actually, you know, that are useful to them and they actually need. But I think in the realm of assistants, if you think of the chatbot as an assistant that's meant to be helpful, and ideally, in my mind, as they become kind of more powerful, they kind of technology.
[0:06:41] It works for you as the individual. That's what I'd like to see with these systems. Then there is a question about how does ads fit into that model? As you say, you want to have trust in your assistant. So how does that all work? And I think no one's really got a full answer to that yet. So we're kind of thinking about it, brainstorming it, but we don't have any current plans to do it ourselves. But we're certainly going to monitor the situation carefully and how users respond to that.
[0:07:09] And to be clear, that's different than in search where you may have AI powered.
[0:07:14] Yes, but there it's a completely different use case because you've already, just like how it's always worked with search, we know what your intent is, basically, and so we can be helpful there. And that's a very different construct, I think. I think the new thing is the assistant idea. If you think of the chatbots eventually becoming very useful digital, universal digital assistants that know a lot about your life and help you with many aspects of your life, I think that's very different from the search use case. So I think there this has to be thought through very carefully.
[0:07:43] And your boss, Sundar, makes a little money off selling ads. Do you feel some pressure there? Do you get pressure? Or is he like, let's keep free, let's keep subscription, and let's figure it out before we put ads in? Well, that's the good thing about what we're doing is we don't feel any immediate pressure to have to make knee-jerk decisions like that.
[0:08:04] That's been the history of what we've done at Google DeepMind is be very scientific and rigorous and thoughtful about each step that we take, be that the technology itself or the product. And so I think there are very complex questions around this. And I think it can be done right, but it can also be done in a way that's not good. And I think in the end, what we want to do is be the most useful we can be to our users. That's what we're focused on first and foremost. That's what I'm focused on at the moment is building the best technology that powers an incredibly used
[0:08:34] And I feel like I ask you this question every time I'm fortunate enough to get to sit down with you, but it seems like over the last years, the capabilities of these AI systems are growing and the talk about trust and safety and responsibility keeps getting pushed a little further, at least from the public viewpoint. And is it because the risks are better understood or because the competitive pressure has just...
[0:08:59] Push that safety talk to the background. Well, look, speaking for ourselves, I think it's always forefront of our minds, but you've got to, it's not a theoretical construct anymore of like, you know.
[0:09:11] AGI rise. What's the theoretical safety arguments about that? I mean, we do discuss that as well. But it's actually a practical thing right now of, you know, these systems are out in the world. Billions of people are using them, including Gemini. And what are the practical guardrails and how should those work? And actually, I think we're learning a lot.
[0:09:32] Doing this in this incremental way when the stakes are kind of relatively low compared to what they will be when these systems become more autonomous and we get closer towards AGI eventually. So I think it's actually great that we are, this is a kind of proving ground for how to, at least today, make sure that these systems behave in a way that's useful to the world. But if we're really all being honest and looking at the whole industry, I'm curious, do you really think that the guiding factor right now is
[0:10:02] Let's do this safely or the competitive dynamic? Well, look, it's ferocious, the competition. I mean, I don't think there's anything ever been like it. And in terms of how capable the organizations are that are doing this at the frontier and the types of researchers and entrepreneurs and leaders that are involved in it, I don't think there's been that big a concentration of talent and resources ever before, as far as I can remember anyway, in technology. So that's all true. And you have to have the best technology.
[0:10:31] That's kind of table stakes in terms of the models. But actually, the commercial pressures are, and if you think about things like enterprise, are actually driving the right type of behavior, I think. And I think you'll start seeing that more obviously.
[0:10:46] This year and next year, as agent-based systems become more prevalent. If you're a big enterprise with a bunch of your own customers, and many of you in the room are, then you're going to want to know that the provider you're getting your AI from, they've got certain guarantees around how that AI is going to behave with your customer data and your customers and in your own networks and products. So I think that...
[0:11:08] Demand is going to actually push the AI frontier labs to be more responsible in terms of and think through these kind of guarantees and monitoring systems and safety and cybersecurity. Some of it's not to do with the AI even, like just cybersecurity around prompt injection of these AI models and things like that. So I think there's actually going to be a lot of commercial pressure to get that right. And I think that's a good thing because that would be a good training run for us for when.
[0:11:38] Bigger stakes come along with AGI. And kind of along those lines, as these systems get more general, more powerful, and some of this market will get to decide anyway, but I'm curious your thoughts. Do you think it's better to be in a world where there are a small number of companies at the frontier, or is that concentration itself more of a systemic risk, and it's better to have a lot of different competition at the frontier? I think there's pros and cons with it.
[0:12:06] You know, it looks like there's going to be several Frontier Labs and but maybe not, you know.
[0:12:12] Maybe half a dozen, right? So I think that's good because it gives lots of choice and it's obviously great for consumers and it's great for other businesses because it's going to drive prices down, all the usual things that are good about competition, drive progress faster. But obviously, as you know very well, it can also mean this intense commercial pressure and competitive pressure. We've got to all remember, all of us that are sort of leading frontier labs, that there's a bigger picture at stake of safety overall.
[0:12:41] And stewarding AGI safely into the world for the benefit of everyone. That is the overriding priority, even above all these commercial pressures, intense though those are. And I certainly have that in the back of my mind. And we hope to be good role models in that. But it's complicated balancing those different things together. And having a larger number of leading providers will make it more difficult to maybe collaborate on things like safety.
[0:13:08] And obviously AI has been the talk of Davos. It's one of the most talked about topics in general. But at the same time, things are moving so fast. Faster, I think, than a lot of humans can adapt, a lot of institutions. Do you feel like institutions, governments, businesses are adapting fast enough? Especially as we prepare for AGI. It seems like some people are still thinking of AGI as a...
[0:13:32] Far off issue. How well do you think our institutions and as individuals? No, I don't think we're ready. And unfortunately, and that's, you know, why I come to places like this, because I think there needs to be more dialogue, difficult though it is between governments and technology companies, but also wider society too. I've always said that, you know, that I've been talking about this for many years now. I do think we have.
[0:13:54] I mean, the good news is I think we have a little bit more time than, say, maybe some of my peers and colleagues say. They have very short timelines to AGI. Mine are still five to 10 years. But that's still not a lot of time if you think about what's coming down the line, how transformative it is. And so I think maybe as we get closer and...
[0:14:12] We have more agentic autonomous systems. It will become clearer to more of the world that we need to get our act together quickly on things like international cooperation, at least some kind of minimal safety standards, these kinds of things. It would be great if there was more dialogue about that, as well as what happens economically. So one more challenging question, then I have a fun thing to end on. But if we look back at this moment, where we are right now in five years, what do you think the decision that the AI community is likely to regret the most? Is it moving?
[0:14:41] Too fast? Or is it not taking where we are seriously enough? Well, it depends which part of the AI community you mean. I mean, I think, look.
[0:14:51] It's amazing to see the progress. Let me just sort of finish with that. We're working on things like cure for diseases with Isomorphic, our spin-out building on AlphaFold. We're advancing material science, fusion, quantum computing with AI. There's some unbelievable things happening that we actually as a society want as fast as possible. I think the way we're going to deal with climate, for example, is through more technology that is new materials, new types of energy sources that...
[0:15:16] Optimal batteries that will, in part, or in the main, be helped by AI to design those things. I genuinely believe that's the way we're going to deal with climate change. And so we want those types of technologies as quickly as possible. So it's fantastic to see the progress. But it does mean we have less time to get our institutions ready and also to do the safety work that's required. So there is this sort of dichotomy.
[0:15:41] Most of us in the frontier labs, I would say, at least the ones I'm close to, understand that we've got to put more and more resources into the sort of security and safety domain. But it's not just a question of the technology companies doing that on their own. In fact, society doesn't, you know, that's not enough. It needs to be broader society that's involved in that. It can't just be the technology companies. So before you go, I want to bring it back to chess, which I know you love.
[0:16:05] I did some reporting because it's one of the only things that AI systems don't do yet is go out there and report. I heard, I want you to tell me if this is true and what you've learned from it. So I heard that when you were in school, you had some mates and they actually were the ones who won the chess tournaments for their age level because your parents...
[0:16:24] Put you in the chess of people several years older, and you didn't win. You finished second or third, but they wanted you to be challenged. First, is that true? Yes, it is true. So after I was about six years old, I won the London Under-8 Championships when I was six. So this is with eight-year-olds. And then after that, I started playing 14-year-olds, and then around about seven or eight, playing adults. So it was definitely the right thing to improve my chess.
[0:16:51] Now, looking back on it, I wouldn't have minded winning a few more medals on the way. So, you know, I think if I had my time again, I would have kind of had a bit of a balance. But they were trying to, you know, always put me at the most challenging end of what I could deal with. Well, in the end, I think you got a medal that a lot of other people come in. Maybe it worked out in the end, exactly. Please give a very warm thank you to Demis Hassabis. Thank you.