Continual learning, long-term reasoning, some aspects of memory, these are still unsolved. I think all of these are going to be required for AGI. Depending on what your AGI timeline is, mine's like 2030 or something like this, then if you start off on a deep tech journey today, you have to just consider AGI appearing in the middle of that journey. It's not bad necessarily, but you have to take that into account. You have to have an active system that can actively solve problems for you to get to AGI. So agents are that path and I think we're just getting going.
Demis Hassabis has had one of the most unusual careers in tech. He was a chess prodigy as a kid, then designed his first hit video game Theme Park at 17. He then went back to school, got a PhD in cognitive neuroscience, published foundational work on how memory and imagination work in the brain, and then in 2010 co-founded DeepMind with one mission: solve intelligence. And I think they've done it. Since then, his lab has gone on to do things most people thought were decades away. AlphaGo beat a world champion at Go. AlphaFold cracked protein structure prediction, a 50-year grand challenge in biology, and they gave it away for free to every scientist on Earth. That work won him the Nobel Prize in Chemistry last year. Today, Demis leads Google DeepMind where he's building Gemini and pushing toward the same goal he set when he was a teenager: artificial general intelligence. Please welcome Demis.
So you've been thinking about AGI longer than almost anyone. When you look at the current paradigm, large-scale pre-training, RLHF, chain of thought, how much of the final architecture for AGI do you think we already have and what's fundamentally missing right now?
Well, first of all, thank you Gary for that great introduction and it's great to be here. Thanks for welcoming me here. It's an amazing space actually. I'm going to have to come back here often. Very inspiring that you get to work in this space. So the question is, I think the components that you just mentioned, I'm pretty sure will be part of the final architecture for AGI. So I think they've come such a long way now and we've proven out so many things about what they can do. I can't see a world in which we will sort of realize in a couple of years this was a dead end. That doesn't make sense to me. But there still might be one or two things missing on top of what we already know works. So continual learning, long-term reasoning, some aspects of memory, these are still unsolved. And how to get the systems to be more consistent across the board. I think all of these are going to be required for AGI. Now it might be that the existing techniques can just scale up to that with some innovation and some incremental innovation. But it could be that there's still one or two big ideas left that need to be cracked. I don't think it's more than one or two if there are out there. And I think my betting is about 50/50 if that's the case. So of course at DeepMind, at Google DeepMind, we work on both those things.
I guess that's, I mean working with a bunch of agentic systems, the wildest thing to me is to what degree it's the same weights over and over. So this idea of continual learning is so interesting because like right now we're sort of cobbling it together with duct tape, you know, these dream cycles at night and things like that.
Yeah, it's pretty cool the dream cycles. And we used to think about this with consolidation with episodic memories. Actually, that's what I studied for my PhD, is how the hippocampus works and integrates new knowledge gracefully into the existing knowledge base. So the brain does that amazingly well. It does it during sleep, especially things like REM sleep, replaying back episodes that are important so that you can learn from it. In fact, our very first Atari program, DQN, one of the ways it was able to master Atari games was by doing experience replay. So we sort of borrowed that from neuroscience and replayed successful trajectories many times. That's way back in 2013 now, in the dark ages of AI. It was a really important thing and I agree with you, we're kind of using duct tape right now. So like shove it all in the context window, but this seems a bit unsatisfying, right? And actually even though we're working on machines, not biological brains, and so potentially you could have millions or tens of millions context window or memory and it can be perfect, there's still a cost to looking it up and finding the right thing that's actually relevant for the specific decision you've got to make right now. And that's non-trivial, that cost, even if you can potentially store it all. I think there's actually a lot of room for innovation in areas like memory.
Yeah, I mean the wild thing is it feels like a million token context ones is actually bigger than, I mean it's plenty big honestly. You can do so much.
Well, it's plenty big for most things that it should be used for. I mean, if you think about the context window as sort of equivalent to working memory, humans have, we have like a few digits, it's like a dozen digits maybe, average of seven. We got million or 10 million context windows. But the problem is that we're trying to store everything in that. Things that are not important, things that are wrong. It's pretty brute force currently and that doesn't seem right. And then the problem is if you're now trying to process live video and you're just going to naively record all the tokens, then actually a million tokens isn't that much. It's only like 20 minutes. So actually you need more if you want something that's going to understand what's going on in your life over maybe a month or two.
DeepMind has historically leaned into reinforcement learning and search, AlphaGo, AlphaZero, and MuZero. How much of that philosophy is actually embedded in how you're building Gemini today? Is RL still underrated?
Yeah, I think potentially it is. It sort of goes in waves. We've worked on agents since the beginning of DeepMind. In fact, that's what we said we were working on. So all of the Atari work and AlphaGo most specifically, they're agent systems. And what we meant by that is systems that are able to accomplish goals on their own and make active decisions and make plans. And so of course we were doing it in the domain of games to make it tractable and then doing increasingly complex games, things like StarCraft after AlphaGo, AlphaStar. So we basically did all the games that were out there. And then of course the question is, can you generalize those models to be world models or models of language, not just models of simple games or even complex games? And that's what the last few years has been about. But really, you can think of a lot of the things we're doing today, all the leading models with thinking modes and chain of thought reasoning, as aspects of what was sort of pioneered with AlphaGo coming back now. And I actually think there's a lot of work we did back then that is relevant today and we're sort of relooking at some of those old ideas at scale today in a more general way, including things like Monte Carlo research and other ways of augmenting the RL on top of the reinforcement learning we're ready to do today. And I think a lot of those ideas both from AlphaGo and AlphaZero are really, really relevant to where we are with today's foundation models. And I think a lot of that is what we're going to see of the advances the next few years.
One question I would have, like obviously today you need bigger and bigger models to be smarter and smarter, but then we're also seeing distillation working and then smaller models can be like quite a bit faster. I think you guys have incredible Flash models that are like, you're finding that they're 95% as good as the frontier and at like one-tenth the price. Is that right?
I think that's one of our core strengths. I mean you have to build the biggest models to have the frontier capabilities. But I think one of our biggest strengths has been distilling and packing that power into smaller and smaller models very quickly. Obviously we invented the kind of distillation process and people like Jeff and Oriol and others, and we're still world experts in that. And we also have a huge need to do it because we've got to serve the biggest probably AI surfaces there are. Obviously there's search with AI Overviews and AI Mode, then there's the Gemini app, and now increasingly every single product at Google, Maps and YouTube and so on, has some aspect of Gemini or Gemini-related technology in it. And so that's billions of users, more than a dozen billion-user products, and they have to be served extremely fast, extremely efficiently and cheaply, and with low latency. So that gives us a really important incentive to make these Flash and even smaller models, Flashlight models, extremely efficient. And hopefully that ends up then being really useful for many of the workloads that all of you use.
I'm curious about how much smarter these smaller models can actually be. Like are there limits to the distillation process? Like could a 50B or 400B model be as smart as like a Mythos for today?
Yeah, I don't see any, I don't think we've got to any kind of, or at least none of us know yet if we've got to any kind of informational limit. I mean maybe at some point that will be the case where there's just an information density that we can't get beyond. But I think for now, the assumption we make is that a year later after one of our leading, you know, Pro models or frontier models goes out, half a year later, a year later, you'll have them in the really tiny almost edge models. And you also see some of that goodness in our Gemma models, which hopefully you're all enjoying, our Gemma 4 models, which I think are really amazing power for their sizes. So again, that uses a lot of these distillation techniques and the idea of how to make things really efficient in these very small models. So I don't really see any limit yet in terms of like some kind of theoretical limit. I think we're still pretty far off of that.
That's amazing. I mean that is really good because one of the weirder things that we're seeing right now is like engineers can do like 500 to a thousand times the amount of work that they were doing like six months ago. I guess I mean the people in this room, there are people who are doing about like a thousand times the work that a Google engineer from the 2000s was doing.
I think it's very exciting. I mean I think the small models have many uses. One is obviously cost, but the speed can allow, you know, if you think about coding even or other things, you can iterate a lot faster also, especially if you're collaborating with the system. I think there's a lot of need for having fast systems that maybe are not quite frontier, like you said, like 95%, 90%, but that's plenty good enough and actually you gain back more than the 10% on the iteration speed. So and then the other big thing I think is running these things on the edge, again for efficiency reasons but also for privacy and security reasons too. If you think about different devices that you might run these systems on that process very personal information, you can also think about robotics as well. Robots in your house, I think you're going to want very efficient, very powerful local models which maybe are orchestrated with some bigger models, frontier models, in the cloud, but you only delegate to that in certain circumstances. And perhaps you process all of the audio, a visual feed, let's say, locally, and that stays local. I could imagine that would be a very good sort of end state.
YC Starter School is back. We're hand-selecting the most promising builders in the world and flying them out to San Francisco for July 25th and 26th to discuss the cutting edge of tech. Apply now for a spot. Okay, back to the video. Going back to context and memory, models currently stateless, but you know, continue, like what would the developer experience even be like for someone who's using a continual learning model? Like any idea like how you'd steer it?
I think it's really interesting. I think that's one of the, not having continual learning currently is one of the things holding back agents from doing full tasks. You know, I think they're really useful for aspects of tasks right now. And you can patch them together and do some really cool things, but they don't adapt well with the context that you're in. And I think that's the missing piece from them being really kind of fire and forget and they'll figure it out themselves. You know, I think they need to be able to learn about the specific context that you're going to put them in. So I think we have to crack that to get full general intelligence.
Where are we on reasoning? So models can do really impressive chain of thought now, but they still fail on things a smart undergrad wouldn't. What specifically needs to change and what progress do you expect in reasoning?
There's a lot of innovation left in thinking paradigms. I would say again, I think we're doing fairly simplistic things, fairly brute force. One could imagine, I think there's a lot of scope for example in monitoring the chain of thought, maybe interjecting midway through a thought process. I often get the impression with our systems and our competitor systems that they're almost overthinking, they're almost getting into sort of loops of things. Like one thing I sometimes like to do is play chess against Gemini, and you know, all the leading foundation models are pretty poor at games, which is quite interesting. It's very cool to kind of look at the thinking traces because obviously these can be well understood. I can tell quite quickly if it's going off on a tangent and it's very provable what the thinking is doing, whether it's useful or not. And so what we see is that sometimes it will consider a move, it will realize it's a blunder, but it can't find anything better, so it kind of goes back to that move and does it anyway. So you just shouldn't be seeing that happening in a very precise reasoning system. So there's just sort of huge gaps I think still, but it may only be one or two tweaks that are required to fix those kind of gaps just to be clear. But I think that's pretty obvious there are there. And that's why you get this kind of jagged intelligence. You know, on the one hand, it can solve gold medal problems in IMO, which is super hard, but on the other hand, as we've all seen, it can still make basic elementary maths errors if you pose the question in a certain way, right? Or elementary reasoning errors. So there's just something to me about the almost an introspection about its own thought process that I feel like there's something maybe missing there.
Agents are really big. Some would say they're hyped. I personally think they're just getting started. It's totally insane. What does DeepMind's internal research tell you about where agent capabilities actually are right now versus, you know, sort of the hype out there?
I think we are, I agree with you. I think we're just at the beginning. You have to have an active system that can actively solve problems for you to get to AGI. That was always clear to us. So agents are that path and I think we're just getting going. I think all of us are getting used to how do we best work and you're leading the way in a lot of this in your own personal experiments. I'm sure many of you are doing that. I think how do you incorporate it into your workflow in a way that isn't just sort of a nice to have but actually starting to do fundamental things. My impression is at the moment we're all experimenting on lots of things, but we're only in the maybe the last couple of months starting to find the really valuable places and the technology is probably only getting good enough for that to be the case, right? That it's not a kind of toy nice demonstration, but actually really adding value to your time and efficiency. I often wonder, I see a lot of people working on like setting off, you know, dozens of agents for like 40 hours, but I'm not sure I've seen the output yet of that quite justify that level of input going in, but I think it will come. So, I still think we're in the experimentation phase. We haven't seen a AAA game that tops the app store charts that was sort of vibe coded yet, right? I've seen and I've programmed and I'm sure many we've all done little nice demonstrations and it's like amazing I can do a prototype of Theme Park in half an hour now which took me six months back when I was 17. It's kind of mind-blowing and I wish I got this feeling if I spent the whole summer working on it you could make something really incredible but it still needs craft and human sort of soul into it and taste. I think that's something that you have to make sure you still bring that to whatever it is you're building. And I think it still shows like it's not quite there yet because why haven't we seen a kid making a hit game that sells 10 million copies, right? That should be possible given the effort that's gone in. So something's still somehow missing. Maybe it's to do with the process or maybe it's to do with the tools. I'm not quite sure. You all will probably know better than me because I'm sure you're all experimenting on that. I haven't seen the result yet which I would expect once this is really delivering that full value which I think will come in the next six to 12 months.
Some of it is like how much of it will be autonomous versus, I mean I don't think we'd see autonomous first. We would actually probably see people in this room operating at a thousand x and then...
That's what you should see first and then many of you, there'll be like games companies or other types of companies that have built some kind of bestselling app, bestselling game using these tools. That's what you should see first and then more of that will get automated.
I mean some of it is like there's a human in there and then the human doesn't want to say that the agents did it yet.
I think part of it might be though that if we want to discuss like creativity, what I often say about that is like if we look at the things we've done like AlphaGo, so obviously very famously you'll all know about the move 37 in game 2. And for me I was waiting for a moment like that to start the science projects like AlphaFold. So we started AlphaFold like the day we got back from Seoul, which is 10 years ago now. I'm going to Korea after this to celebrate the 10-year anniversary of AlphaGo. But it's not enough to come up with move 37. Like that's pretty cool, very useful. But can it invent Go? That's what I want, a system that can invent Go if you give it a high-level description, like a game you can learn the rules of in five minutes, but it takes many lifetimes to master. It's beautiful aesthetically. But you can play it in a few hours in an afternoon. So maybe you could imagine that would be the high-level description I would give. And then I'd want the return, the thing I get back is Go, right? And clearly today's systems I think can't do that. So the question is why and I think there's something still missing there.
