Google DeepMind CEO Demis Hassabis: AI's Next Breakthroughs, AGI Timeline, Google's AI Glasses Bet

So don't store everything, just store the important things. That would be a lot more efficient. That's what the brain does. And better long-term reasoning and planning. Now, it remains to be seen whether just sort of scaling up existing ideas and technologies will be enough to do that, or we need one or two more really big, insightful innovations. I'm probably, if you were to push me, I would be in the latter camp. But I think no matter what camp you're in, we're going to need large foundation models as the key component of the final AGI systems. Of that, I'm sure. So I'm not a subscriber to someone like Jan LeCun who thinks that there's sort of some kind of dead end.

I think the only debate in my mind is, are they a key component or the only component? So I think it's between those two options. And for me, this is one advantage we have of having such a deep and rich research bench. We can go after both of those things at maximum with maximum force, both scaling up the current paradigms and ideas. And when I say scaling up, that also involves innovation, by the way, pre-training especially, I think we're very strong on. And then really new blue sky ideas for new architectures and things, you know, the kinds of things we've invented over the last 10 years as Google and DeepMind, you know, of course, including Transformers.

Can something with a lot of hard-coded stuff ever be considered AGI? No, I think, well, depends what you mean by a lot. I think that I'm very interested in hybrid systems is what I would call them or neurosymbolics. Sometimes people call them, you know, AlphaFold, AlphaGo are examples of that. to some of our most important work. combines neural networks and deep learning with things like Monte Carlo Tree Search. So I think that could be possible, and there's some very interesting work we're doing using the LLMs with things like evolutionary methods, AlphaEvolve, to actually go and discover new knowledge. You may need something beyond what the existing methods do. But I think learning is a critical part of AGI. It's actually almost a defining feature. When we say general we mean general learning can it can it learn uh new knowledge and can it learn across any domain that's the general part so for me learning is synonymous with intelligence and always has been okay so if learning is synonymous with intelligence and these models still don't have the ability to continually learn no uh like i said earlier it has goldfish brain yeah it can search the internet and it can be like i figured this out yeah but it doesn't change the model

it's just, we'll forget it after the session. Do you have a theory as to how the continual learning problem can be solved? And do you want to share it with us all day? I can give you some clues. We are working very hard on it. We've done some work on, you know, I think the best work on this in the past with things like AlphaZero, you know, that learn from scratch, versions of AlphaGo, AlphaGo Zero also learned on top of the knowledge it already had. So we've done it in much narrower domains. Games are obviously a lot easier than the messy real world. So it remains to be seen if those kinds of techniques will really scale and generalize to the real world and actual real world problems. But at least the methods we know can do some pretty impressive things. And so now the question is, can we blend that, at least in my mind, with these big foundation models? And so of course, the foundation models are learning

Demis Hassabisduring training, but we would love them to learn, you know, out in the wild and including things like personalization. I think that's going to happen. And I feel like that's a critical part

of building a great assistant is that it understands you and it works for you as technology that works for you. And we've released our first versions of that just last week. Personal intelligence is the sort of first baby steps towards that. But I think to have it, you want to do it more than just having your data in the context window. That's you want to have something a bit deeper than that, which is, as you say, actually changes the model over time. That's what ideally you would have. And that technique has not been cracked yet. We've brought up AGI a couple of times. So let me put this to you, because I was speaking with Sam Altman towards the end of the year. And I asked him, I was like, you seem to be saying two things. We're not at AGI yet, but every time he talks about what GPT models can do, it seems like it fits his definition. And he said that AGI is underdefined. What he wishes everybody could agree to was that we've sort of whooshed by AGI and we moved towards super intelligence. Do you agree with that? I'm sure he does wish that. No, absolutely not. I don't think AGI should be sort of turned into a marketing term for commercial gain. I think there has always

Demis Hassabisbeen a scientific definition of that. My definition of that is a system that can exhibit all the cognitive capabilities humans can. And I mean all. So that means, you know, the kind of highest

levels of human creativity that we always celebrate, the scientists and the artists that we admire. So it means, you know, not just solving a maths equation or a conjecture, but coming up with a breakthrough conjecture. That's much harder. You know, not solving something in physics or some bit of chemistry, some problem, even like alpha folds, you know, protein folding, but actually coming up with a new theory of physics, something like, you know, like Einstein did with general relativity, right? Can a system come up with that? Because of course we can do that. The smartest humans with their brain architect, our human brain architectures have been able to do that in history.

And the same on the art side, you know, not just create a pastiche of what's known, but actually be. Picasso or Mozart and create a completely new genre of art that we'd never seen before right and today's systems in my opinion are nowhere near that um doesn't matter how many you know Erdos problems you solve which for some reason you know I mean you know that's it's good that we're doing those things but I think it's far far from what uh you know a true invention or someone like a Ramanujan would have been able to do and you need to and you need to have a system that can potentially do that across all these domains and then on top of that I'd add in physical intelligence, because of course, you know, we can play sports and control our bodies and to amazing levels, the elite sports people that are walking around, you know, here today in Davos. And we're still way off of that on robotics, as another example. So I think an AGI system would have to be able to do all of those things to really fulfill the original sort of goal of the AI field. And I think, you know, we're five to 10 years away from that.

I think the argument would be that if something can do all those things, it would be considered superintelligence but no you think agi is a good no of course not because the individual those individual humans could we can come up with new theories that einstein did feinman did all the all the greats that more my scientific heroes they were able to do that it's rare but it's possible with the human brain architecture so superintelligence is another concept that's worth talking about but that would be things that can really go beyond what human intelligence can do we can't think in 14 dimensions or you know plug in weather satellites into our brains Not yet, anyway.

And so those are truly beyond human or superhuman. And that's a whole other debate to have, but once we get to AGI. I was listening to you recently, and something you said really surprised me. You were asked on the Google DeepMind podcast, which is a great listen, if you have a system today that is close to AGI, I thought it might be Gemini 3. You named Nano Banana. Yes. The image generator. Yes. What? Well, you know, sometimes you have to have these fun names and have fun with those. But how is an image generator close to AGI? Oh, well, of course, look, let's take image generators, but also let's talk about our video generator, VO, which is the state of the art in video generation. I think that's even more interesting from an AGI perspective. You know, you can think of a video model that can generate you 10 seconds, 20 seconds of a realistic scene.

it's sort of a model of the physical world. Intuitive physics, we'd sometimes call it in physics land. And it's sort of intuitively understood how liquids and objects behave in the world. And that's, and obviously one way to exhibit understanding is to be able to generate it, at least to the human eye being accurate enough to be satisfying to the human eye. Obviously, it's not completely accurate from a physics point of view, and we're going to improve that but it's it's it steps towards having uh this idea of a world model a system that can understand the world and the mechanics and the causality of the world and of course that would be i think essential for a gi because that would allow these systems to plan long-term plan in the real world um over perhaps very long time horizons which of course we as humans can do you know um i'll spend four years getting a degree so that i have more qualifications so that in 10 years i'll have a better job. These are very long-term plans that we all do quite effortlessly. And at the moment, without these systems, we still don't know how to do. We can do short-term plans over one timescale. But I think you need these kind of world models. And I think if you imagine robotics,

that's exactly what you want for robotics, is robots planning in the real world, being able to imagine many trajectories from the current situation they're in, in order to complete some task. That's exactly what you'd want. And then finally, from our point of view, and this is why we worked with Gemini as being multimodal from the beginning, able to deal with video, image, and eventually converge that all into one model, that's our plan, is that it'll be very useful for a universal assistant as well. So let's talk product a little bit. I watched the documentary, The Thinking Game, along with 300 million other people.

There was something kind of interesting that happened there. Throughout the documentary, yourself and some colleagues kept pointing your phone at things and asking an assistant, Alpha, what was going on. And I was yelling at the computer, as I usually do, and said, this guy needs glasses. Like, he needs smart glasses to be able to do it. The phone is the wrong form factor. What is your vision for AI glasses and when is the rollout happening? I think you're exactly right. And that was our conclusion. It's very obvious when you sort of dog food these things and internally that, as you saw from the film, we were holding up, you know, you're holding up your phone to get it to tell you about the real world.

And it's amazing it works, but it's clearly not the right form factor for a lot of things you want to do. you know cooking or you're roaming around the city and asking for directions or recommendations um or even helping the you know partially sighted there's a huge i think use case there to help uh with those types of uh situations and um and for that i think you need something that's hands free and the obvious thing is for those of us anyway that wear glasses like me is that is to put it on glasses but there may well be other devices too i'm not sure that glasses is the final form factor but it's definitely it's obviously a clear next form factor and of course at google and an alphabet we have a long history with glasses and maybe we're a bit too early in the past but i think the my analysis of it and talking to the people working on that project was a couple of things that the form factor was a bit too chunky and clunky and the battery life and these kind of things which are now more or less solved but i think the thing it was missing was

Demis Hassabiskiller app and i think the killer app is universal digital assistant that's with you helping you in your everyday life and they're available to you on any surface on your computer on your browser

on your phone but also on you know devices like glasses when you're uh you know walking around the city and i think it needs to be kind of seamless um and kind of knows uh each of those contexts and understands each of those contexts around you and i think we're close now especially with gemini 3 i feel we finally got uh ai that is maybe powerful enough to make that a reality and we're you know it's one of the most exciting projects we're working on i would say and it's one of the things i'm personally working on is is making smart glasses really work uh and we hope to we've we've done some great partnerships with warby parker and gentle monster and samsung to build these next generation glasses. And you should start seeing that, you know, maybe by the summer. Yeah. Warby Parker did have a filing that said that these glasses are coming out pretty soon this year. Yeah. And the prototype design, it depends how, you know, when prototype phase, it depends how quickly that advances. But I think it's going to happen very soon. And I think it will be, you know, a category, a new category defining technology. Given your personal involvement is it safe to say that this is a pretty important initiative yeah it's one well

yes but it's i mean i you know i like to it's it's not just as important obviously i like spending my own time on important things but i like to be at the push the most cutting edge thing and and and and and that's often the hardest thing and picking interim goals and giving confidence to the team and and also just sort of understanding if the timing's right and over the years i've been doing this the many you know the decades now um you know i've got quite good at doing that so i try to uh be at the the most cutting edge parts of um i feel i can make the most difference there so things like glasses robotics i'm spending time on and world models right okay so timing's right for glasses let's talk about ads sure it's the timing right for ads um let me set it yes okay um there's been some news that jim and i might include ads there's been some news that some of your competitors.

I'd include ads. The funniest thing I saw about that on social media was someone who said these people are nowhere close to AGI. It's not going to be this world disrupting technology if the business model is advertising. Yeah, well, it's interesting. I think those are tells on, you know, I think actions speak louder than words going back to the original conversation we were having with, you know, Sam and others claiming AGI is around the corner. Why would you bother with ads then? So that is, I think, a reasonable question to ask. But I think, look, from our point of view, we have no plans at the moment to do ads if you're talking about the Gemini app, right, specifically. I think we are going, obviously, we're going to watch very carefully what, you know, the outcome of what ChatGPT is saying they're going to do. I think has to be handled very carefully because the dichotomy i see is that um uh uh if you if you

Demis Hassabiswant an assistant that works for you what is the most important thing trust okay so trust and security and privacy um because you want to share potentially your life with that assistant then you

want to be confident that it's it's working on your you know behalf and uh and and with your best interests and so you know you've got to be careful i think there are ways one could do it but you'll be careful that it doesn't the advertising model doesn't bleed into that and confuse the the user uh as to what you know what is this assistant recommending you and i think um you know that's going to be interesting challenge in that space and that's what's not to do and sudar in a recent earnings call said there are some ideas within google of the right way to approach this sure how do you approach advertising well um you know that's still we're still brainstorming that but i think it's i think it's um there are also you know very interesting ways when if you think about glasses devices there are other revenue models out there okay um so you know it's going to be interesting to see i don't think we've made any strong conclusions on that but it's an area that uh needs very careful thought just to get a definitive answer from you i think you've given it but i'm just going to do it one more time i read before we met google has told advertisers in recent days from last year that it plans to bring ads to its AI chatbot Gemini in 2026?

Nope. We have no current plans. That's all I can say. All right. That's pretty definitive. Yep. All right. Let's just keep going through some of your competitors. Anthropic. Claude Code. Yeah. Claude Code and Claude Cowork have caused a tremendous amount of buzz. Yeah. It is amazing to see what some people have done. I saw a post from an ex-Amazon executive who said that he built a custom crm in a weekend or actually a day and a half let's call it a weekend um what do you think about it yeah and do you do you plan to have an answer to it it's very exciting um and i think you know kudos to to to anthropic i think they built a very good model there with claw code we're very happy with the the current coding capabilities of gemini 3 it's very good at certain things like front end work i've been using it over the christmas to prototype type games. So it's amazing. It's getting me back into programming. I love the whole vibe coding wave that's happening. I think it will open up the whole productivity space to designers, creatives, artists that maybe would have had to work with teams about access to teams of programmers.

Now they can probably do a lot more just on their own. I think that's going to be amazing once that's sort of out in the world in a more general way to create lots of new creative opportunities. We're very happy with our work on code. We've got more to do there. We've just released Antigravity, our own IDE, which is very, very popular. We can't actually serve all the demand that we're seeing there. And we're pushing very hard on coding and tool use performance of Gemini. But it's one thing that I think Anthropic have fully focused on. They don't make image models, multimodal models, world models, they just do.

coding and language models. And they're very, very good at that. And we're pleased to be partnering on that on the one hand, and also it gives us something to push for to improve with our own models. Let's just talk broadly about the AI industry business. I have a theory for how this could all fall apart, and I want to run it by you. So it's a three-step process. The first is that large language model training runs produce limited returns uh the second is that there are flash models like gemini flash that run ai computing as cheap as search and then step three is that the massive infrastructure commitments uh that have been made become somewhat useless given those two factors and there is a cascading collapse that happens is that a legitimate worry um i think it's a plausible possible scenario. I don't think it's the likely one, in my opinion. I mean, in my mind, there's no doubt. AI's gone already proved out enough, I would say, and our work, I think, in things like science and AlphaFold and drug discovery that it's here to stay.

It's not like tomorrow, like, oh, we found that AI doesn't work. We've blasted way past that. So I think it's clearly going to be the most transformative technology in human history. There's maybe a question mark about timelines. Is it two years or five years? I I mean, either way, it's very soon for something this transformative. And I think we're still in the nascent era of actually figuring out how to make use of it and deploy it because the technology is improving so fast. I think there's a huge capability overhang, actually, of what even today's models can do that maybe even us as building those things don't fully know.

So I think there's just a vast amount of product opportunities that we see. And I think we're, as Google only, just starting to scratch the surface now of actually natively sort of plugging these things in to our amazing existing products, let alone building the new ones. You know, AI Inbox, we've just started trialing. I mean, who wants to do email admin? I mean, wouldn't we all love that to just go away? That's my number one pain point for my work, King Day. And there's so many examples like that just waiting to be addressed, I think. you know, agents in browsers, helping out with YouTube. Obviously, we're now powering search with it. So I think there's enormous opportunities. And if you're talking about the AI bubble, if that's the question. I was trying to not ask the AI bubble. Well, I think it's fine. I mean, it seems like that's the question. I'm very happy to answer it. Because I think, look,

Demis Hassabismy view is it's not binary. Are we in a bubble, not in a bubble? I think parts of the AI industry probably are. And other parts, I think it remains to be seen. So I think some of the things are,

you know, when you see seed rounds of tens of billions of dollars of companies that basically have no product or research, it's just some people coming together. That seems a bit unsustainable to me in a normal market, a bit frothy. On the other hand, you know, we're businesses like us. We have massive underlying businesses and products that it's very obvious how AI would increase the efficiency or the productivity of using those products. And then it remains to be seen how popular the monetization of these new ai native products like chatbots glasses all of these things um we'll have to see i think there will be enormous markets but they're yet to be proven out but from my perspective you know running google deep mind is um my job is to make sure that whatever happens with an air bubble if it if it if it bursts or if there isn't one and it continues we win either way right and i think we're incredibly well positioned as alphabet uh in either case um and you know doubling down on existing businesses in the one case or being at the forefront and the front here uh in in the in the bull case going back to thinking game speaking of the way that this will impact the economy i started to feel bad for the opponents of your

technology um lisa doll okay uh demoralized sure uh this guy mana who played starcraft beat your bot but realize that it's basically over for you. Versus machines. Now, we're all up against this in some way, as this stuff makes its way into knowledge work. I thought you were meaning our AI competitors. Them, I'm okay with. I don't feel sad about that. No, no. Sorry. Relentless progress of AI. You mean the gamers. The gamers. Yeah, sure. You made me feel bad for gamers. Well, you know. But I want to ask about this. You know, we're going to have the same situation with knowledge work. that these models that performed admirably against the world's best StarCraft and Go players are now starting to do our work.

And are we going to end up in the same position? Well, look, given you brought up games as an example, let's look at what's happened in games. So chess, we've had chess computers that are better since I was a teenager than Garry Kasparov in the 90s, right? They weren't general AI systems, but they were deep blue. Chess is more popular than ever. No one's interested in seeing computers playing computers. We're interested in Magnus Carlsen playing, you know, the top, the other top chess players in the world. Interestingly in Go, the best South Go player in the world is a South Korean. And he was about 15, I think, when AlphaGo match happened.

He's in his mid-20s now.

Demis HassabisAnd he's by far the strongest player there's ever been by the ELO ratings because he's learned natively young enough.

He was, you know, he's the first generation, you could say, that's learned with AlphaGo knowledge in the knowledge pool. and um you know he may actually be stronger than alpha go was back then so i think and we all still enjoy starcraft and all the other all the other computer games we enjoy human endeavor i think it's a bit more a bit similar to like we still love the 100 meters uh olympic race um even though we have vehicles that can go way faster than usain bolt but you know we we don't you know that's that's different thing right and so i think we have infinite capacity to adapt and um and uh and and sort of evolve uh with our technologies because why is that because we have we are general intelligences um that's the thing about it is we are agi systems we are obviously we're not artificial we're general systems and it's and and we are capable of inventing science and uh we're tool making uh animals that's what separates us humans from from the other animals is we're able to make tools all around modern civilization including computers and of course ai being the ultimate expression of computers that all has come from our human minds which were evolved for you know hunter-gathering lifestyle so it's kind of amazing we were able and it shows how general we

are that we're able to get to the modern civilization we see around us today and we're talking about things like ai and you know science and physics and all these things right and i think we'll adapt again but there is an important question actually beyond the economics one about jobs and those things is purpose and meaning because we all get a lot of our purpose and meaning from the jobs we do i certainly do from the science i do so how does what happens when a lot of that is automated um i think you know that that's why i've been calling for you know i think we knew new great philosophers actually and it will be a change to the human condition but i don't think it necessarily has to be worse i think we've it's like the industrial revolution maybe 10x of that but we'll have to adapt again and i think we'll find new um uh meaning and things and we do a lot of things already today that are not just for economic gain you know art extreme sports ex-polar exploration many of these things um and maybe we'll have much more sophisticated esoteric versions of those things in the future. Okay, two minutes left. I have two questions. I don't know if we're going to get to both of them. Let me ask the one that I want to know the answer most

about. In a recent interview, you said that you have a theory that information is the most fundamental unit of the universe, not energy, not matter, information. Yeah. How? Well, look, I think if you look at energy, I mean, I don't know if we're going to cover this in two minutes, but in energy, energy, energy and... I think a lot of people sort of think of them as isomorphic with information. But I think information is really the right way to understand the universe. So if you think of biology and living systems, we're information systems that are resisting entropy, right? We're trying to retain our structure, retain our information in the face of, you know, a randomness that's happening around us.

And I think you can look at that, you know, in a larger physics scale. So almost not just biology, but things like, you know, mountains and planets and asteroids, they've all been subject to some kind of selection pressure, not Darwinian evolution, but some kind of external pressure. And the fact that they've been stable over a long amount of time means that that information is kind of stable and meaningful. So I think one could view the world in terms of its complexity, information complexity. And I think a lot of what we're doing with our, the reason I'm thinking about all of that is because of things like AlphaGo and AlphaFold, especially AlphaFold, where, you know, we solved all the protein structures that are kind of known to science.

and how we've done that well because only a certain number of those in the kind of almost infinite possibilities of protein structures are stable and and those are the ones you've got to find so you've got to understand that topology that information topology and follow it and then suddenly these problems that seem to be intractable because you know how can you find the needle in

Demis Hassabisthe haystack actually become very tractable if you understand the energy landscape or the information landscape around that and that's how i think eventually we'll solve most diseases come

with new drugs new materials new superconductors with the help of ai helping us navigate that information landscape uh dennis before we go i just want to wrap with this uh well maybe maybe quickly this first one and then a big question at the end in the thinking game speaking of um health and ai there's this moment where there's a discussion in the lab about whether to release the results of alpha folds and you kind of sit there adamantly and you're like why are we going through a process release it release it now talk a little bit about um the lesson from there yeah well look we we started alpha fold to crack a unbelievably tough scientific challenge 50 year grand challenge uh of protein folding and protein structure prediction and the reason we worked on

Demis Hassabisthat and the reason we've put so much effort into it is we sort of thought it was a root node problem if we could solve it and put that out in the world it could be it could do amazing untold impact on

things like human health and understanding of biology but we as a team no matter how talented or hard hard working we are we would only be able to scratch a surf a small tiny amount of that potential on our own it's clear so um in that case and in this case it was obviously the right thing to do to maximize the benefit to the world here to put it out there to the scientific massive scientific community uh to build on top of and use alpha fold and it's been incredibly gratifying to to see three million researchers around the world use it in their important research. I think in future, almost every single drug that's discovered from now on will probably have used AlphaFold at some point in that process, which is amazing for us.

And really, that's what we do, all the work we do for. I also read that moment, you tell me if I'm wrong, as something of a metaphor, small, passionate AI division, kind of yelling in a big company, get this out, cut the red tape. Yeah, potentially. But look, I mean, we've had amazing support from the beginning from Google. And the reason that we, you know, we joined forces with Google back in 2014 is Google itself is a scientific, research, engineering, technical-led company, always has been and has that at its core. And that's why, you know, I think that we have the scientific method and the scientific approach, that thoughtful approach, that rigorous approach in everything we do.

So, of course, they're going to love something like AlphaFold. Okay, here's the big question at the end. you built AlphaGo trained the computer to play Go on human knowledge and then once it mastered the human In level playing, you kind of like let it loose with a program called AlphaZero, and it started doing things that you could never even imagine and making new circuits in ways that surprised you. Eventually, maybe, there will come a time where LLMs or some version of them reach a mastery of human knowledge in the same way. What is going to happen when you then let that loose and it does the same, potentially does the same thing as AlphaZero?

Yeah, I think it'd be very exciting. I mean, that's what to me would be the AGI moment is, you know, then it will discover a new superconductor, room temperature superconductor that's possible in the laws of physics, but we just haven't found that needle in the haystack or a new source of energy, a new way to build optimal batteries. I think all of those things will become possible and indeed not just possible. I think they will happen once we get to a system that's, first of all, got to, you know, human level knowledge. And then there'll be some techniques. Maybe it will have to help invent some of those techniques, but kind of like AlphaZero that will allow it to go beyond into new uncharted territory.

That idea of it like plugging weather system into its brain, like it's going to be on that. Yeah, exactly. All right. Exciting times. Demis, thanks for coming on the show. Thank you. Thanks, everybody. Thank you so much.