How to Build the Future: Demis Hassabis

0:00Continual 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, you know, 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.

0:39Demis 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.

1:29That 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 Hassapes.

1:53So 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, thanks, 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 will 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.

2:28So 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.

3:11but 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 eGentic 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, you know, right now we're sort of cobbling it together with duct tape. Yes. These dream cycles at night and things like that.

3:47Yeah. It's pretty cool, the dream cycles. And we used to think about this with consolidation with episodic memory. Actually, that's what I studied for my PhD is how the hippocampus works and integrates, you know, 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.

4:27You know, 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. This seems a bit unsatisfying, right? And actually, even though we're working on machines, not biological brains, and so potentially you could have, you know, millions or tens of millions size 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.

5:08I 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. It's plenty big for most things that it should be used for. I mean, if you think about the context windows sort of equivalent to working memory, you know, humans have, we have like a few digits, you know, it's like a dozen digits, maybe, you know, average of seven, we got million or, you know, 10 million context windows. But the problem is, is that we're trying to store everything in that, you know, things that aren't and not important things that are wrong, it's pretty brute force currently.

5:46and that doesn't seem right. And then the problem is if you're now trying to try and 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.

6:25It sort of goes in ebbs and 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?

7:12And 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 re-looking at some of those old ideas at scale today in a more general way, including things like Monte Carlo Tree Search and other ways of doing, 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.

7:54And 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, 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 quite a bit faster. I think you guys have incredible flash models that 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 is, 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.

8:36Obviously, we invented the kind of distillation process and people like Jeff and Oriel 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 and this Gemini app. And now increasingly, every single product at Google has, you know, 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.

9:20So 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 for. I'm curious about how much smarter these smaller models can actually be. Are there limits to the distillation process? Could a 50B or 400B model be as smart as a mythos for today? Yeah, 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.

10:00But I think for now, the assumption we make is that a year later after one of our leading 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 didn't really see any limit yet in terms of like some kind of theoretical limit.

10:37I think we're still pretty far off of that. That's amazing. I mean, that is really good. Yes. You know, one of the weirder things that we're seeing right now is like engineers can do like 500 to 1000 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 X the work that like Steve Yagi talks about this. It's like a thousand X 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.

11:14Also, 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 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.

11:53I think you're going to want very efficient, very powerful local models, which may be 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 visual feed, let's say, locally, and that stays local. I could imagine that would be a very good sort of end state. YC Startup 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 what would the developer experience even be like for someone who's using a continual learning model?

12:44Any idea 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 for 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.

13:21So 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 the 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.

14:03They'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 it's that all the leading foundation models are pretty poor at games which is quite interesting it's very uh uh cool to kind of look at the thinking traces because obviously these can be a well understood you know i can tell quite quickly if it's going off on a tangent and it's very sort of provable what the what the the thinking is doing whether it's useful or not and so what we see is that you know sometimes it will it will it will consider a move it will realize it's a blunder, but it can't find anything better.

14:39So it kind of goes back to that move and does it anyway. So, you know, 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 math errors if you pose the question in a certain way, right?

15:16So, 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 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.

15:48That 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, you know, we're experimenting on lots of things, but we're only in maybe the last couple of months starting to find the really valuable places.

16:18And 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 that 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 triple a 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 a 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 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, you know, human sort of soul into it and taste.

17:20I 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 probably know better than me because I'm sure you're all experimenting on that. But 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.

17:57Some 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 a thousand X and then... That's what you should see first. And then many of you, you know, There'll be games companies or other types of companies that have built some kind of best-selling app, best-selling 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 we want to discuss creativity.

18:37creativity what i often say about that is like if we look at the things we've done like alpha go so obviously very famously you'll all know about the move 37 in game two and for me i was waiting for a moment like that to start the science projects like alpha fold so we started alpha fold like the day we got back from seoul which is 10 years ago now i'm going to career after this to celebrate the 10-year anniversary of alpha go 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.

19:19It'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, um, clearly today's systems, I think can't do that. So the question is why? Um, and I think there's something still missing there. Well, someone in this room might, might make it. Then the answer would be, there's nothing missing. It just was the way we were using the systems. And that might actually be the answer. It might be that today's systems are capable of that with a brilliant enough creative person using it and providing that impetus, that's the soul of the project and being able to So probably being au fait enough with the tools to almost be at one with the tools.

20:07I could imagine that would be happening if you experimented with the tools all day and all night, like probably many of you are doing. And you combine that with proper deep creativity, something more incredible could be done. Switching gears to open source, I mean, or open weights. I mean, the recent release of Gemma, you're making highly capable, open and accessible ones that can actually run locally. What do you think that means for, will AI be something that is in the hands of the users instead of primarily in the cloud? And does that change who gets to build with these models? We're huge proponents of, in general, of open source and open science.

20:46And you mentioned AlphaFold at the beginning. You know, we put that all out there for free and all of our science work, even still today, we publish in, you know, the big journals. We wanted to create world leading models for their sizes. And so that's what hopefully we've done with Gemma. And we're, you know, very committed to that path. And hopefully you all experiment and build and enjoy using Gemma. I think it's been like 40 million downloads now and just in, you know, two and a half weeks. So we're really excited about that. And I also think it's important for there to be Western stacks on open source.

21:19You know, obviously, a lot of the Chinese models are excellent, and they're currently leading in open source. And we think Gemma is very competitive for its sizes in all those respects. And for us, I mean, there is a question of resources, talent and compute. Like nobody has enough spare compute to just make two, you know, frontier models at maximum size, right, with different attributes. So that's pretty difficult. But also for now, what we've decided is that our edge models, the things we want to use for Android and glasses and robotics, it's best that they're open models because they're vulnerable anyway once you put them out on the surfaces.

21:59so they might as well be actually fully open, right? So we've sort of made a decision to kind of unify that at the kind of, we call it nano size level. And so that actually works for us strategically as well. And, you know, we hope as many people as possible build on it. And of course, we'll be building on that too. Earlier, before we came on, I got to show you a demo of my version of Samantha from her, which is harrowing for me to try to demo something to you. And it worked, which is amazing. Gemini was built multimodal and I spent a lot of time with a bunch of the models. And I mean, the depth of the context and the tool use with speech directly to model.

22:41Yeah. There is nothing like bar none, like the best one. Yeah. Yeah. I think I think that's a sort of still a slightly underappreciated aspect of of of the Gemini series is we started it being multimodal from the start. But that made it a little bit more difficult actually to begin with because then just focusing on text, for example. But we believe we're going to gain from that in the long run. And I think we're seeing that now for things like world model building. So stuff like Genie that we build on top of Gemini. I think it's going to be really important for things like robotics. So this is why Gemini Robotics, which many of you probably played around with.

23:17I think it's going to be built on multimodal foundation models, the robotics models. and we think we have a sort of competitive advantage with Gemini being so strong at multimodal. We're using it increasingly in things like Waymo but also if you imagine devices and assistants that digital assistants that come with you into the real world, maybe on your phone or glasses or some other device, it needs to understand the physical world around you and intuitive physics and the physical context you're in And that's what our systems are extremely good at. And I think you found that's why you've enjoyed using it in your setup.

23:54We're planning to continue on that. And I think we're far and away the strongest models on those types of problems. So the cost of inference is dropping fast. What becomes possible when inference is essentially free? And how does that change what your team is actually optimizing for? Yeah, I'm not sure inference will ever be essentially free. I mean, there's sort of Jevon's paradox and other things about like, I think we'll just end up using, all of us will end up using whatever we can get our hands on. And you could imagine millions of agents, swarms of agents working together on things. That's one way to use the inference.

24:31Or you could imagine single agents or smaller groups of agents thinking in multiple directions and then ensembling that. So we're experimenting with all these things. Probably many of you are. All of that will use up any inference, I think, that's available. I mean, one day, maybe it can be almost cost zero. Certainly the energy, if we solve fusion or superconductors or optimal batteries or some set of those things, which I think we will do with material science, energy costs will be essentially zero. But there'll still be the physical creation of the chips and other things. There'll be some bottleneck, at least for the next few decades, I think.

25:10And so if that's the case, they'll still be rationing on the inference side. They'll still have to use it, I think, efficiently. Yeah. Well, luckily, the smaller models are getting smarter and smarter, which is fantastic. We got a lot of bio and biotech founders in the audience. I can see a few. Alpha 3 took us beyond proteins to a broad spectrum of biomolecules. How close are we to modeling full cellular systems? Or is that still a fundamentally harder problem in a class of its own? Well, Isomorphic Labs, which we spun out from DeepMind after we did AlphaFold 2, which is going amazingly well, it's trying to build out not just AlphaFold, it's just one piece of the drug discovery process, as many of you know, but we're trying to do the adjacent biochemistry and chemistry to design the right compounds with the right properties and so on.

26:02We'll have some big announcements very soon to talk about on that front. I think that's going really well. Eventually, you want a whole virtual cell. So I've talked about this in many of my science talks about a full working simulation of a cell that you can perturb. And then the outputs of that would be close enough to experimental that it's useful. You could skip out a lot of the search steps and generate lots of synthetic data to train other models that then would predict things about real cells. And I think we're about 10 years away, probably, from something like a virtual cell, like a full virtual cell.