Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold

Dwarkesh Patel: I wouldn't be surprised if we had AGI-like systems within the next decade.

Demis Hassabis: It was pretty surprising to almost everyone, including the people who first worked on the scaling hypotheses, how far it's gone. In a way, I look at the large models today and I think they're almost unreasonably effective for what they are. It's an empirical question whether that will hit an asymptote or a brick wall. I think no one knows.

Dwarkesh Patel: When you think about superhuman intelligence, is it still controlled by a private company?

Demis Hassabis: As Gemini becomes more multimodal and we start ingesting audiovisual data as well as text data, I do think our systems are going to start to understand the physics of the real world better. The world's about to become very exciting, I think, in the next few years as we start getting used to the idea of what true multimodality means.

Dwarkesh Patel: Okay, today it is a true honor to speak with Demis Hassabis, who is the CEO of DeepMind. Demis, welcome to the podcast.

Demis Hassabis: Thanks for having me.

Dwarkesh Patel: First question, given your neuroscience background, how do you think about intelligence? Specifically, do you think it's like one higher level general reasoning circuit, or do you think it's thousands of independent subscales and heuristics?

Demis Hassabis: Well, it's interesting because intelligence is so broad and what we use it for is so generally applicable. I think that suggests that there must be some sort of high level common algorithmic themes around how the brain processes the world around us. Of course, there are specialized parts of the brain that do specific things, but I think there are probably some underlying principles that underpin all of that.

Dwarkesh Patel: How do you make sense of the fact that in these LLMs though, when you give them a lot of data in any specific domain, they tend to get asymmetrically better in that domain? Wouldn't we expect a sort of general improvement across all the different areas?

Demis Hassabis: Well, first of all, I think you do sometimes get surprising improvement in other domains when you improve in a specific domain. For example, when these large models improve at coding, that can actually improve their general reasoning. So there is some evidence of transfer, although I think we would like a lot more evidence of that. But also, that's how the human brain learns too. If we experience and practice a lot of things like chess or creative writing, we also tend to specialize and get better at that specific thing, even though we're using general learning techniques to get good at that domain.

Dwarkesh Patel: What's been the most surprising example of this kind of transfer for you? Like you see language and code or images and text?

Demis Hassabis: I think probably, I mean, I'm hoping we're going to see a lot more of this kind of transfer. But things like getting better at coding and math and generally improving your reasoning, that is how it works with us as human learners. But I think it's interesting seeing that in these artificial systems.

Dwarkesh Patel: Can you see the sort of mechanistic way in which, let's say in the language and code example, there's like, I found the place in a neural network that's getting better with both the language and the code, or is that too far down the weeds?

Demis Hassabis: I don't think our analysis techniques are quite sophisticated enough to be able to hone in on that. I think that's actually one of the areas that a lot more research needs to be done on, kind of mechanistic analysis of the representations that these systems build up. I sometimes like to call it virtual brain analytics in a way. It's a bit like doing fMRI or single cell recording from a real brain. What's the analogous analysis techniques for these artificial minds? There's a lot of great work going on on this sort of stuff. People like Chris Olah, I really like his work. A lot of computational neuroscience techniques, I think, could be brought to bear on analyzing these current systems we're building. In fact, I try to encourage a lot of my computational neuroscience friends to start thinking in that direction and applying their know-how to the large models.

Dwarkesh Patel: What do other AI researchers not understand about human intelligence that you have some sort of insight on given your neuroscience background?

Demis Hassabis: I think neuroscience has added a lot. If you look at the last 10 or 20 years that we've been at it, and I've been thinking about this for 30 plus years. In the earlier days of this new wave of AI, I think neuroscience was providing a lot of interesting directional clues. Things like reinforcement learning, combining that with deep learning, some of our pioneering work we did there, things like experience replay, even the notion of attention, which has become super important. A lot of those original inspirations come from some understanding about how the brain works. Not the exact specifics. One's an engineered system, the other one's a natural system. So it's not so much about a one-to-one mapping of a specific algorithm. It's more inspirational direction, maybe some ideas for architecture or algorithmic ideas or representational ideas. Because the brain's an existence proof that general intelligence is possible at all. The history of human endeavors has been that once you know something's possible, it's easier to push hard in that direction because you know it's a question of effort then and a question of when, not if. That allows you to make progress a lot more quickly. So I think neuroscience has inspired a lot of the thinking, at least in a soft way, behind where we are today. As for going forwards, I think there's still a lot of interesting things to be resolved around planning and how does the brain construct the right world models. I studied, for example, how the brain does imagination, or you can think of it as mental simulation. So how do we create very rich visual spatial simulations of the world in order for us to plan better?

Dwarkesh Patel: I'm curious how you think that will interface with LLMs. Obviously, DeepMind is at the frontier and has been for many years with systems like AlphaGo and so forth, having these agents who can think through different steps to get to an end outcome. Will this be a path for LLMs to have this sort of tree search kind of thing on top of them? How do you think about this?

Demis Hassabis: I think that's a super promising direction in my opinion. We've got to carry on improving the large models and we've got to carry on making them more and more accurate predictors of the world, in effect making them more and more reliable world models. That's clearly a necessary but probably not sufficient component of an AGI system. On top of that, we're working on things like AlphaZero, like planning mechanisms that make use of that model in order to make concrete plans to achieve certain goals in the world. Perhaps chain thought together or lines of reasoning together, and maybe use search to explore massive spaces of possibility. I think that's kind of missing from our current large models.

Dwarkesh Patel: How do you get past the immense amount of compute that these approaches tend to require? Even the AlphaGo system was a pretty expensive system because you had to run an LLM on each node of the tree. How do you anticipate that'll get made more efficient?

Demis Hassabis: Well, one thing is Moore's law tends to help if every year more computation comes in. But we focus a lot on efficient, sample efficient methods and reusing existing data, things like experience replay, and also looking at more efficient ways. The better your world model is, the more efficient your search can be. One example I always give with AlphaZero, our system to play Go and chess and any game, is that it's stronger than world champion level at all these games. And it uses a lot less search than a brute force method. Deep Blue, say, to play chess, would maybe look at millions of possible moves for every decision it's going to make. AlphaZero and AlphaGo looked at around tens of thousands of possible positions in order to make a decision about what to move next. But a human world champion probably only looks at a few hundred moves, even the top ones, in order to make their very good decision about what to play next. That suggests that the brute force systems don't have any real model other than the heuristics about the game. AlphaZero has quite a decent model, but the top human players have a much richer, much more accurate model of Go or chess. That allows them to make world-class decisions on a very small amount of search. So there's a trade-off there. If you improve the models, then your search can be more efficient and therefore you can get further with your search.

Dwarkesh Patel: I have two questions based on that. The first being with AlphaGo, you had a very concrete win condition of, at the end of the day, do I win this game or not? And you can reinforce on that. When you're just thinking of an LLM putting out thought, do you think there will be this kind of ability to discriminate in the end whether that was a good thing to reward or not?

Demis Hassabis: Of course, that's why we pioneered and DeepMind's famous for using games as a proving ground, partly because it's efficient to research in that domain. But the other reason is it's extremely easy to specify a reward function. Winning the game or improving the score is built into most games. So that is one of the challenges of real world systems: how does one define the right objective function, the right reward function, and the right goals and specify them in a general way, but specific enough to point the system in the right direction? For real world problems, that can be a lot harder. But actually, if you think about it, even scientific problems, there are usually ways that you can specify the goal that you're after.

Dwarkesh Patel: When you think about human intelligence, humans are just super sample efficient. Einstein coming up with relativity, right? There's just thousands of possible permutations of the equations. Do you think it's also this sort of sense of different heuristics of, I'm going to try this approach instead of this? Or is it a totally different way of approaching coming up with that solution than what AlphaGo does to plan the next move?

Demis Hassabis: Well, look, I think it's different because our brains are not built for doing Monte Carlo tree search. That's just not the way our organic brains would work. So I think that in order to compensate for that, people like Einstein have used their intuition. They use their knowledge and their experience to build extremely accurate models of physics, including these mental simulations. If you read about Einstein and how he came up with things, he used to visualize and really feel what these physical systems should be like, not just the mathematics of it, but have a really intuitive feel for what they would be like in reality. That allowed him to think these very outlandish thoughts at the time. So I think it's the sophistication of the world models that we're building. If you imagine your world model can get you to a certain node in a tree that you're searching, and then you just do a little bit of search around that leaf node, that gets you to these original places. But if your model and your judgment on that model is very good, then you can pick which leaf nodes you should expand with search much more accurately. Therefore, overall, you do a lot less search. There's no way that any human could do a brute force search over any significant space.

Dwarkesh Patel: A big open question right now is whether RL will allow these models to do the self-play synthetic data to get over the data bottleneck. It sounds like you're optimistic about this.

Demis Hassabis: Yeah, I'm very optimistic about that. First of all, there's still a lot more data that can be used, especially if one views multimodal and video and these kinds of things. Society is adding more data all the time to the Internet and things like that. But I think there's a lot of scope for creating synthetic data. We're looking at it in different ways, partly through simulation, using very realistic games environments to generate realistic data, but also self-play. That's where systems interact with each other or converse with each other. That worked very well for us with AlphaGo and AlphaZero, where we got the systems to play against each other and learn from each other's mistakes and build up a knowledge base that way. I think there are some good analogies for that, though it's a little more complicated to build a general kind of world data.

Dwarkesh Patel: How do you get to the point where these models, the synthetic data they're outputting, the self-play they're doing, is not just more of what they've already got in their data set, but is something they haven't seen before? To actually improve the abilities.

Demis Hassabis: There, I think there's a whole science needed. I think we're still in the nascent stage of this, of data curation and data analysis. Actually analyzing the holes that you have in your data distribution. This is important for things like fairness and bias to remove that from the system, to try and really make sure that your data set is representative of the distribution you're trying to learn. There are many tricks one can use, like over weighting or replaying certain parts of the data. Or you could imagine if you identify some gap in your data set, that's where you put your synthetic generation capabilities to work.

Dwarkesh Patel: Nowadays people are paying attention to the RL stuff that DeepMind did many years before. What are the sort of early research directions or something that was done way back in the past, but people just haven't been paying attention to that you think will be a big deal? Like there was a time where people weren't paying attention to scaling. What's the thing now where it's totally underrated?

Demis Hassabis: The history of the last couple of decades has been things coming in and out of fashion. I do feel like a while ago, maybe five plus years ago, when we were pioneering with AlphaGo and before that, DQN, where it was the first system that worked on Atari, our first big system more than 10 years ago now, that scaled up Q learning and reinforcement learning techniques to combine that with deep learning to create deep reinforcement learning. And then use that to scale up to master some pretty complex tasks, like playing Atari games just from the pixels. I do actually think a lot of those ideas need to come back in again. And as we talked about earlier, combine it with the new advances in large models and large multimodal models, which is obviously very exciting as well. So I do think there's a lot of potential for combining some of those older ideas together with the newer ones.

Dwarkesh Patel: Is there any potential for AGI to eventually come from just a pure RL approach? Like the way we're talking about it, it sounds like the LLM will form the gripe fryer and then this sort of tree search will go on top of that. Or is it a possibility it's just completely out of the dark?

Demis Hassabis: I think theoretically there's no reason why you couldn't go full AlphaZero on it. There are some people here at DeepMind, at Google DeepMind, and in the RL community who work on that, fully assuming no priors, no data, and just build all knowledge from scratch. I think that's valuable because those ideas and those algorithms should also work when you have some knowledge too. But having said that, my betting would be the quickest way to get to AGI and the most likely plausible way is to use all the knowledge that's existing in the world right now on things like the web, that we've collected, and we have these scalable algorithms like Transformers that are capable of ingesting all of that information. I don't see why you wouldn't start with a model as a kind of prior or to build on and to make predictions that helps bootstrap your learning. I just think it doesn't make sense not to make use of that. So my betting would be that the final AGI system will have these large multi-models as part of the overall solution, but probably won't be enough on their own. You will need this additional planning search on top.

Dwarkesh Patel: This sounds like the answer to the question I'm about to ask, which is, as somebody who's been in this field for a long time and seen different trends come and go, what do you think that the strong version of the scaling hypothesis gets right and what does it get wrong? It's just the idea that you throw enough compute at a wide enough distribution of data and you get intelligence.

Demis Hassabis: Look, my view is this is kind of an empirical question right now. I think it was pretty surprising to almost everyone, including the people who first worked on the scaling hypotheses, how far it's gone. In a way, I look at the large models today and I think they're almost unreasonably effective for what they are. I think it's pretty surprising some of the properties that emerge. It's clearly, in my opinion, got some form of concepts and abstractions. If we were talking five plus years ago, I would have said to you, maybe we need an additional algorithmic breakthrough in order to do that, like maybe more like the brain works. I think that's still true if we want explicit abstract concepts, neat concepts, but it seems that these systems can implicitly learn that. Another really interesting, I think, unexpected thing was that these systems have some sort of grounding. Even though they don't experience the world multimodally, or at least until more recently when we have the multimodal models. That's surprising that the amount of information and models that can be built up just from language. I have some hypotheses about why that is. I think we get some grounding through the RLHF feedback systems, because the human raters are by definition grounded people. We're grounded in reality. So our feedback is also grounded. Perhaps there's some grounding coming in through there. And also maybe language contains more grounding, if you're able to ingest all of it, than we perhaps thought or linguists perhaps thought before. So there are some very interesting philosophical questions. I think people haven't even really scratched the surface of yet, looking at the advances that have been made. It's quite interesting to think about where it's going to go next. But in terms of your question about large models, I think we've got to push scaling as hard as we can. That's what we're doing here. It's an empirical question whether that will hit an asymptote or a brick wall. Different people argue about that. But actually, I think we should just test it. I think no one knows. In the meantime, we should also double down on innovation and invention. This is something that Google Research and DeepMind and Google Brain have pioneered many things over the last decade. That's our bread and butter. You can think of half our effort as to do with scaling and half our effort as to do with inventing the next architectures, the next algorithms that will be needed, knowing that you've got this scaled larger and larger model coming along the lines. My betting right now, but it's a loose betting, is that you would need both. But I think you've got to push both of them as hard as possible. We're in a lucky position that we can do that.

Dwarkesh Patel: I want to ask more about the grounding. You can imagine two things that might change, which would make the grounding more difficult. One is that as these models get smarter, they're going to be able to operate in domains where we just can't generate enough human labels because we're not smart enough. If it does a million line pull request, how do we tell it this is within the constraints of our morality and the end goal we wanted and this isn't? The other is, it sounds like you're saying more of the compute so far has been doing next token prediction and in some sense it's a guardrail because you have to talk as a human would talk and think as a human would think. Now, if additional compute is going to come in the form of reinforcement learning, where you just get to the end objective and we can't really trace how you got there, when you combine those two, how worried are you that the grounding goes away?

Demis Hassabis: Well, look, I think if the grounding is not properly grounded, the system won't be able to achieve those goals properly. I think you sort of have to have the grounding, or at least some of it, in order for a system to actually achieve goals in the real world. I do actually think that as these systems and things like Gemini are becoming more multimodal, and we start ingesting things like video and audiovisual data as well as text data, and the system starts correlating those things together, I think that is a form of proper grounding. So I do think our systems are going to start to understand the physics of the real world better. One could imagine the active version of that is being in a very realistic simulation or game environment where you're starting to learn about what your actions do in the world and how that affects the world state itself, but also what next learning episode you're getting. These RL agents we've always been working on and pioneered like AlphaZero and AlphaGo, they actually affect their active learners. What they decide to do next affects what the next learning piece of data or experience they're going to get. There's this very interesting feedback loop. Of course, if we ever want to be good at things like robotics, we're going to have to understand how to act in the real world.

Dwarkesh Patel: So there's a grounding in terms of will the capabilities be able to proceed? Will they be enough in touch with reality to be able to do the things we want? And there's another sense of grounding: we've gotten lucky in the sense that since they're trained on human thought, they maybe think like a human. To what extent does that stay true when more of the compute for training comes from just did you get the right outcome and not guardrailed by are you proceeding on the next token as a human would? Maybe the broader question I'll pose to you is, and this is what I asked Shane as well, what would it take to align a system that's smarter than the human, maybe thinks in alien concepts, and you can't really monitor the million line pull request because you can't really understand the whole thing?

Demis Hassabis: Shane and I and many others here, we've had that at the forefront of our minds since before we started DeepMind. We planned for success. In 2010, no one was thinking about AI, let alone AGI, but we already knew that if we could make progress with these systems and these ideas, the technology being created would be unbelievably transformative. So we were already thinking 20 years ago about what the consequences of that would be, both positive and negative. The positive direction is amazing science, things like AlphaFold, incredible breakthroughs in health and science and math and scientific discovery. But then also, we've got to make sure these systems are understandable and controllable. This would be a whole discussion in itself. But there are many ideas that people have from much more stringent eval systems. I think we don't have good enough evaluations and benchmarks for things like, can the system deceive you? Can it exfiltrate its own code, sort of undesirable behaviors? Then there's ideas of actually using AI, maybe narrow AIs, not general learning ones, but systems that are specialized for a domain to help us as the human scientists analyze and summarize what the more general system is doing. Kind of narrow AI tools. I think there's a lot of promise in creating hardened sandboxes or simulations so that they're hardened with cybersecurity arrangements around the simulation, both to keep the AI in, but also as cybersecurity to keep hackers out. Then you could experiment a lot more freely within that sandbox domain. I think a lot of these ideas, and there are many others, including the analysis stuff we talked about earlier, where can we analyze and understand what the concepts are that this system is building, what the representations are, so maybe they're not so alien to us and we can actually keep track of the kind of knowledge that it's building.

Dwarkesh Patel: So being back up to it, I'm curious what your timelines are. Shane said his modal outcome is 2028. I think that's maybe his median. What is yours?

Demis Hassabis: I don't have prescribed specific numbers to it because I think there's so many unknowns and uncertainties and human ingenuity and endeavor comes up with surprises all the time. That could meaningfully move the timelines. But I will say that when we started DeepMind back in 2010, we thought of it as a 20-year project. Actually, I think we're on track, which is kind of amazing for 20-year projects because usually they're always 20 years away. That's the joke about whatever it is, quantum, AI, take your pick. But I think we're on track. So I wouldn't be surprised if we had AGI-like systems within the next decade.

Dwarkesh Patel: Do you buy the model that once you have an AGI, you have a system that basically speeds up further AI research? Maybe not in an overnight sense, but over the course of months and years you have much faster progress than you would have otherwise?

Demis Hassabis: I think that's potentially possible. I think it partly depends what we decide as a society to use the first AGI systems or even proto-AGI systems for. Even the current LLMs seem to be pretty good at coding. We have systems like AlphaCode, we also have theorem-proving systems. One could imagine combining these ideas together and making them a lot better. Then I could imagine these systems being quite good at designing and helping us build future versions of themselves. But we also have to think about the safety implications of that, of course.

Dwarkesh Patel: I'm curious what you think about that. I'm not saying this is happening this year or anything, but eventually you'll be developing a model where during the process of development, you think there's some chance that once this is fully developed, it'll be capable of an intelligence explosion dynamic. What would have to be true of that model at that point where you're like, I've seen these specific evals, I understand its internal thinking enough and its future thinking that I'm comfortable continuing development of the system?

Dwarkesh Patel: Well, look, we need a lot more understanding of the systems than we do today before I would be even confident of explaining to you what we would need to tick box there. So I think actually what we've got to do in the next few years, in the time we have before those systems start arriving, is come up with the right evaluations and metrics and maybe ideally formal proofs. But it's going to be hard for these types of systems, but at least empirical bounds around what these systems can do. And that's why I think about things like deception as being quite root node traits that you don't want. Because if you're confident that your system is exposing what it actually thinks, then that opens up possibilities of using the system itself to explain aspects of itself to you.

The way I think about that actually is like if I was to play a game of chess against Gary Kasparov, which I played in the past, or Magnus Carlsen, the amazing chess players of our time, I wouldn't be able to come up with a move that they could, but they could explain to me why they came up with that move and I could understand it post hoc. That's the sort of thing one could imagine with these systems—for them to explain it to us and even maybe the proofs behind why they're thinking something, certainly in a mathematical context.

Demis Hassabis: Got it.

Dwarkesh Patel: Do you have a sense of what the converse answer would be? What would have to be true where tomorrow morning you're like, "Oh man, I didn't anticipate this. We got to stop Gemini 2 training"? What would specifically trigger that?

Demis Hassabis: Yeah, I could imagine that. This is where things like the sandbox simulations come in. I would hope we're experimenting in a safe, secure environment. And then something happens in it where something very unexpected occurs—a new unexpected capability, or something that we explicitly told the system we didn't want, that it did, but then lied about. These are the kinds of things where one would want to dig in carefully with the systems that are around today, which are not dangerous in my opinion today, but in a few years they might have potential. You would ideally pause and really get to the bottom of why it was doing those things before one continued.

Dwarkesh Patel: Going back to Gemini, I'm curious what the bottlenecks were in the development. Why not make it immediately one order of magnitude bigger if scaling works?

Demis Hassabis: Well, first of all, there are practical limits—how much compute can you actually fit in one data center? You're bumping up against very interesting distributed computing challenges. Fortunately, we have some of the best people in the world on those challenges, cross data center training, all these kinds of things. Very interesting challenges, hardware challenges. We have our TPUs that we're building and designing all the time, as well as using GPUs. And then you also have to consider that the scaling laws don't just work by magic. You still need to scale up the hyperparameters, and various innovations are going in all the time with each new scale. It's not just about repeating the same recipe. At each new scale, you have to adjust the recipe, and that's a bit of an art form in a way. You have to almost get new data points. If you try to extrapolate your predictions several orders of magnitude out, sometimes they don't hold anymore because new capabilities can be step functions. Some things just hold and other things don't. So often you do need those intermediate data points to correct some of your hyperparameter optimization so that the scaling law continues to be true. One order of magnitude is probably the maximum you want to do between each era.

Dwarkesh Patel: That's so fascinating. In the GPT-4 technical report, they say they were able to predict the training loss tens of thousands of times less compute than GPT-4—they could see the curve. But the point you're making is that the actual capabilities that loss implies may not be so clear. Downstream capabilities sometimes don't follow from that.

Demis Hassabis: You can often predict that the core metrics like training loss will hold, but then it doesn't actually translate into MMLU or math or some other actual capability you care about. They're not necessarily linear all the time. There are nonlinear effects there.

Dwarkesh Patel: What was the biggest surprise to you during the development of Gemini? Something like this happening?

Demis Hassabis: Well, I wouldn't say there was one big surprise, but it was very interesting trying to train things at that size and learning about all sorts of things—from organization to how to babysit such a system and track it. I think getting a better understanding of the metrics you're optimizing versus the final capabilities that you want is still not a perfectly understood mapping, but it's one we're getting better and better at.

Dwarkesh Patel: There's a perception that maybe other labs are more compute efficient than DeepMind has been with Gemini. What do you make of that perception?

Demis Hassabis: I don't think that's the case. Gemini 1 used roughly the same amount of compute, maybe slightly more, than what was rumored for GPT-4. I don't know exactly what was used, so I think it was in the same ballpark. I think we're very efficient with our compute, and we use our compute for many things. It's not just about scaling. Going back to earlier points about innovation and ideas, a new invention is only useful if it can also scale. So you need quite a lot of compute to do new invention because you've got to test many things at at least some reasonable scale and make sure they work at that scale. Some new ideas may not work at a toy scale but do work at a larger scale, and those are the more valuable ones. If you think about that exploration process, you need quite a lot of compute. The good news is we're pretty lucky at Google that this year we're going to have the most compute by far of any research lab, and we hope to make very efficient use of that in terms of scaling, capability, and new inventions.

Dwarkesh Patel: What's been the biggest surprise to you if you go back to yourself in 2010 when you were starting DeepMind, in terms of what AI progress has looked like? Did you anticipate back then that it would in some large sense amount to spending billions of dollars into these models, or did you have a different sense of what it would look like?

Demis Hassabis: We thought that. I know you've interviewed my colleague Shane, and he always thought in terms of compute curves and comparing roughly to the brain—how many neurons and synapses there are, very loosely. We're actually in that kind of regime now, roughly in the right order of magnitude of the number of synapses in the brain and the compute that we have. But more fundamentally, we always bet on generality and learning. Those were always at the core of any technique we would use. That's why we triangulated on reinforcement learning, search, and deep learning as three types of algorithms that would scale and be very general without requiring a lot of handcrafted human priors, which we thought was the failure mode of efforts to build AI in the 90s. Places like MIT with logic-based systems, expert systems, masses of hand-coded human information—that turned out to be wrong or too rigid. We wanted to move away from that. We spotted that trend early, used games as our proving ground, and did very well. I think that inspired others. AlphaGo was a big moment for inspiring many to think that these systems were ready to scale. Then with the advent of Transformers, invented by our colleagues at Google Research and Brain, that was the type of deep learning that allowed us to ingest masses of information, and that turbocharged where we are today. I think that's all part of the same lineage. We couldn't have predicted every twist and turn, but the general direction was the right one.

Dwarkesh Patel: It's fascinating because if you read your old papers or Shane's old papers—Shane's thesis in 2009 said the way we would test for AI is if you can compress Wikipedia, and that's literally the loss function of our LLMs. Or your own paper in 2016 before Transformers, where you were comparing neuroscience and AI and said attention is what is needed. We had these things called out. We had some early attention papers like Neural Turing Machines, but they weren't as elegant as Transformers. Transformers was the nicer and more general architecture.

When you extrapolate all this out forward and think about superhuman intelligence, what does that landscape look like to you? Is it still controlled by a private company? What should the governance of that look like concretely?

Demis Hassabis: I would love—I think this is so consequential, this technology. It's much bigger than any one company or even industry in general. It has to be a big collaboration with many stakeholders from civil society, academia, and government. The good news is that with the popularity of recent chatbot systems, that has woken up many other parts of society to the fact that this is coming and what it will be like to interact with these systems. That's great. It's opened up lots of doors for very good conversations. An example was the safety summit in the UK hosted a few months ago, which I thought was a big success in getting this international dialogue going. The whole of society needs to be involved in deciding what we want to deploy these models for, how we want to use them, and what we don't want to use them for. We need to get some international consensus around that. We also need to make sure the benefits of these systems benefit everyone, for the good of everyone in society. That's why I push so hard on things like AI for science. With things like our spin-out Isomorphic, we're going to start curing terrible diseases with AI and accelerate drug discovery. Amazing things like climate change and other big challenges that face humanity—I'm optimistic we can solve them because we've got this incredibly powerful tool coming along. Ideally, we would have a big consensus around that and a big discussion, almost like at the UN level if possible.

Dwarkesh Patel: One interesting thing is if you look at these systems, you chat with them and they're immensely powerful and intelligent. But it's interesting the extent to which they haven't automated large sections of the economy yet. Five years ago, if I showed you Gemini, you'd be like, "Wow, this is totally coming for a lot of things." How do you account for that? What's going on where it hasn't had the broader impact?

Demis Hassabis: I think that just shows we're still at the beginning of this new era. For these systems, there are some interesting use cases where you can use chatbot systems to summarize stuff, do some simple writing, maybe boilerplate type writing. But that's only a small part of what we all do every day. For more general use cases, we still need new capabilities—things like planning and search, but also personalization and memory, episodic memory. Not just long context windows, but actually remembering what we spoke about 100 conversations ago. Once those start coming in, I'm really looking forward to things like recommendation systems that help me find better, more enriching material, whether that's books, films, or music. I would use that type of system every day. I think we're just scratching the surface of what these AI assistants could actually do for us in our general everyday lives and in our work context. They're not reliable enough yet to do things like science with them. But once we fix factuality and grounding and other things, I think they could become like the world's best research assistant for you as a scientist or clinician.

Dwarkesh Patel: I want to ask about memory. You had this fascinating paper in 2007 where you talk about the links between memory and imagination and how they are in some sense very similar. People often claim that these models are just memorizing. How do you think about that claim? Is memorization all you need because in some deep sense that's compression? What's your intuition here?

Demis Hassabis: At the limit, one could try to memorize everything, but it wouldn't generalize out of your distribution. I think these systems are clearly—the early criticisms were that they were just regurgitating and memorizing. But clearly the new era, the Gemini and GPT-4 type era, they are definitely generalizing to new constructs. In my thesis and that paper that started that area of imagination in neuroscience, I was showing that human memory is a reconstructive process. It's not a videotape. We put it together from components that seem familiar to us, the ensemble. That made me think that imagination might be the same thing, except you're using the same semantic components but putting them together in a way your brain thinks is novel for a particular purpose like planning. I do think that kind of idea is still probably missing from our current systems—this pulling together different parts of your world model to simulate something new that helps with your planning, which is what I would call imagination.

Dwarkesh Patel: Now you guys have the best models in the world with the Gemini models. Do you plan on putting out some sort of framework like the other two major AI labs have—once we see these specific capabilities, unless we have these specific safeguards, we're not going to continue development or ship the product?

Demis Hassabis: Yes, we have lots of internal checks and balances already, but we're going to start publishing. Watch this space—we're working on a whole bunch of blog posts and technical papers that we'll be putting out in the next few months along the lines of things like responsible scaling laws. We have those implicitly internally and various safety councils, people like Shane chairing them. But it's time for us to talk about that more publicly, so we'll be doing that throughout the course of the year.

Dwarkesh Patel: That's great to hear. Another thing I'm curious about is the risk of rogue actors, foreign agents, being able to steal the weights and fine-tune them to do crazy things. How do you think about securing the weights to make sure something like this doesn't happen? Making sure a very key group of people have access to them?

Demis Hassabis: There are two parts to this. One is security, one is open source. Security is super key—normal cybersecurity type things. We're lucky at Google DeepMind to be behind Google's firewall and cloud protection, which is best in class in the world corporately. Behind that, we have specific DeepMind protections within our code base, so it's a double layer of protection. I feel pretty good about that. You can never be complacent, but it's already best in the world in terms of cyber defenses. We've got to carry on improving that. Things like hardened sandboxes could be a way of doing that. Maybe there are specifically secure data centers or hardware solutions to this too that we're thinking about. In the next three, four, five years, we would also want air gaps and various other things known in the security community. I think all frontier labs should be doing that because otherwise nation states and rogue nation states and other dangerous actors would have a lot of incentive to steal things like the weights.

Open source is another interesting question. We're huge proponents of open source and open science. We've published thousands of papers. AlphaFold, Transformers, AlphaGo—all of these we put out into the world, published and open source. Most recently, our weather prediction system. But when it comes to the core technology, the foundational technology that is very general purpose, the question I would have for open source proponents is: how does one stop bad actors—individuals or rogue states—from taking those same open source systems and repurposing them for harmful ends? We have to answer that question. I haven't heard a compelling, clear answer to that from proponents of open sourcing everything. There has to be some balance there, though obviously it's a complex question of what that is.

Dwarkesh Patel: I feel like tech doesn't get the credit it deserves for funding hundreds of billions of dollars worth of R&D. Obviously we have DeepMind with systems like AlphaFold. But when we talk about securing the weights, maybe right now it's not something that's going to cause the end of the world. But as these systems get better, the worry is that a foreign agent gets access to them. Presumably right now there are dozens to hundreds of researchers who have access to the weights. What's the plan for getting to a situation where if you need access, it's an extremely strenuous process and nobody individual can really take them out?

Demis Hassabis: One has to balance that with allowing for collaboration and speed of progress. Another interesting thing is you want brilliant independent researchers from academia or the UK AI Safety Institute and US one to be able to red team these systems, so one has to expose them to a certain extent, though that's not necessarily the weights. We have a lot of processes in place about making sure only those who need access have access. Right now we're still in the early days of those kinds of systems being at risk. As these systems become more powerful, more general, and more capable, one has to look at the access question.

Dwarkesh Patel: Some of these other labs have specialized in different things relative to safety—Anthropic with interpretability, for example. Do you have a sense of where you guys might have an edge so that now that you have the frontier model, you're going to scale up safety and put out the best frontier research?

Demis Hassabis: We helped pioneer RLHF and other things that can be used for performance but also for safety. A lot of the self-play ideas and these kinds of things could also be used potentially to automate a lot of the boundary conditions you have with new systems. Part of the issue is that with these very general systems, there's so much surface area to cover about how they behave. I think we are going to need some automated testing. With things like simulations and games—very realistic environments, virtual environments—we have a long history in that and using those kinds of systems for building AI algorithms. I think we can leverage all of that history. Around Google, we're very lucky to have some of the world's best cybersecurity experts and hardware designers, so I think we can bring that to bear for security and safety as well.

Dwarkesh Patel: Let's talk about Gemini. Now you guys have the best model in the world. The default way to interact with these systems has been through chat so far. Now that we have multimodal and all these new capabilities, how do you anticipate that changing? Or do you think that'll still be the case?

Demis Hassabis: I think we're just at the beginning of understanding what a full multimodal model system might be like to interact with. It'll be quite different to what we're used to today with chatbots. The next versions of this over the next year, 18 months, maybe we'll have some contextual understanding around the environment around you through a camera or a phone. I could imagine that as the next awesome glasses, the next step. We'll start becoming more fluid in understanding—let's sample from a video, let's use voice. Maybe eventually things like touch, and if you think about robotics, other types of sensors. I think the world's about to become very exciting in the next few years as we start getting used to the idea of what true multimodality means.

Dwarkesh Patel: On the robotics subject, Ilya said when he was on the podcast that the reason OpenAI gave up on robotics was because they didn't have enough data in that domain, at least at the time they were pursuing it. You guys have put out different things like RoboTransformer and other things. Do you think that's still a bottleneck for robotics progress, or will we see progress in the world of atoms as well?

Demis Hassabis: We're very excited about our progress with things like Gato and RT2, the robotic transformer. We've always liked robotics and we've had amazing research in that, and we still have that going now because we like the fact that it's a data-poor regime. That pushes us on very interesting research directions that we think are going to be useful anyway—sampling efficiency, data efficiency in general, transfer learning, learning from simulation and transferring that to reality. All of these are very interesting general challenges we would like to solve. So we've always pushed hard on that. I think Ilya's right that it is more challenging because of the data problem. But I think we're starting to see the beginnings of these large models being transferable to the robotics regime—learning in the general domain, language domain, and then just treating tokens like Gato does as any type of token. The token could be an action, a word, part of an image, a pixel. That's what I think true multimodality is. To begin with, it's harder to train a system like that than a straightforward text language system. But going back to our early conversation about transfer learning, you start seeing that a true multimodal system benefits from the other modalities. You get better at language because you now understand a little bit about video. It's harder to get going, but ultimately we'll have a more general, more capable system like that.

Dwarkesh Patel: Whatever happened to Gato? That was super fascinating—you could play games and also do video and text.

Demis Hassabis: We're still working on those kinds of systems. Those ideas we're trying to build into our future generations of Gemini to be able to do all of those things. Robotics transformers and things like that are kind of follow-ups to that.

Dwarkesh Patel: We see asymmetric progress towards the domains in which the self-play kinds of things we're talking about will be especially powerful—math and code. You can use these things to do really cool novel things. Will they just be superhuman coders but in other ways still worse than humans? How do you think about that?

Demis Hassabis: We're making great progress with math, theorem proving, and coding. But it's still interesting if one looks at creativity in general and scientific endeavor. I think we're getting to the stage where our systems could help the best human scientists make their breakthroughs quicker—almost triage the search space in some ways, or find a solution like AlphaFold does with a protein structure. But they're not at the level where they can create the hypothesis themselves or ask the right question. As any top scientist will tell you, that's the hardest part of science—asking the right question, boiling down the space to what's the critical question to go after, and formulating that problem in the right way to attack it. That's not something our systems have any idea how to do. But they are suitable for searching large combinatorial spaces if one can specify the problem in that way with a clear objective function. That's very useful for many of the problems we deal with today, but not the most high-level creative problems.

Dwarkesh Patel: DeepMind has published all kinds of interesting stuff speeding up science in different areas. How do you think about that in the context of if you think AGI is going to happen in the next 10, 20 years? Why not just wait for the AGI to do it for you? Why build these domain-specific solutions?

Demis Hassabis: We don't know how long AGI is going to be. We always used to say back when we started DeepMind that we don't have to wait for AGI in order to bring incredible benefits to the world. My personal passion has been AI for science and health. You can see that with AlphaFold and all of our various Nature papers in different domains, material science work, and so on. There's lots of exciting directions and impact in the world through products too. It's a huge opportunity we have as part of Google—they've got dozens of billion-user products that we can immediately ship our advances into, and billions of people can improve their daily lives. It enriches and enhances their daily lives. From the point of view of AGI specifically, it battle-tests your ideas.

Demis Hassabis: So you don't want to be in a sort of research bunker where you just theoretically push some things forward. But then your internal metrics start deviating from real world things that people would care about, or real world impact. You get a lot of direct feedback from these real world applications that tells you whether your systems really are scaling, or if we need to be more data efficient or sample efficient. Most real world challenges require that. It keeps you honest and pushes you, nudging and steering your research directions to make sure they're on the right path. So I think it's fantastic. And of course, the world benefits from that, and society benefits from that, many years before AGI arrives.

Dwarkesh Patel: The development of Gemini is super interesting because it comes right at the heels of merging these different organizations, Brain and DeepMind. I'm curious: what have been the challenges there? What have been the synergies? It's been successful in the sense that you have the best model in the world now. What's that been like?

Demis Hassabis: It's been fantastic over the last year. Of course, it's been challenging to do that, like any big integration. But you're talking about two world-class organizations with long, storied histories of inventing many important things, from deep reinforcement learning to transformers. It's very exciting pulling all of that together and collaborating much more closely. We always used to collaborate, but more on a project-by-project basis versus a much deeper, broader collaboration like we have now. Gemini is the first fruit of that collaboration—the name Gemini implies twins. A lot of other things are made more efficient, like pooling compute resources together and ideas and engineering. At the stage we're at now, where there's huge amounts of world-class engineering that has to go on to build the frontier systems, it makes sense to coordinate that more closely.

Dwarkesh Patel: You and Shane started DeepMind partly because you were concerned about safety. You saw AGI coming as a live possibility. Do you think the people who were formerly part of Brain—the other half of Google DeepMind now—approach it in the same way? Have there been cultural differences there in terms of that question?

Demis Hassabis: Overall, I think one of the reasons we joined forces with Google back in 2014 is that the entirety of Google and Alphabet, not just Brain and DeepMind, take these questions of responsibility very seriously. Our mantra is to try and be bold and responsible with these systems. I'm obviously a huge techno-optimist, but I want us to be cautious given the transformative power of what we're bringing into the world collectively. I think it's going to be one of the most important technologies humanity will ever invent. We've got to put all our efforts into getting this right and be thoughtful and humble about what we know and don't know about the systems that are coming and the uncertainties around that.

In my view, the only sensible approach when you have huge uncertainty is to be cautiously optimistic and use the scientific method to try and have as much foresight and understanding about what's coming down the line and the consequences before it happens. You don't want to be live A-B testing out in the world with these very consequential systems, because unintended consequences may be quite severe. I want us to move away as a field from the "move fast and break things" attitude. That may have served the valley very well in the past and created important innovations, but in this case, we want to be bold with the positive things it can do and make sure we realize things like medicine and science, while being responsible and thoughtful about mitigating the risks.

Dwarkesh Patel: That's why the responsible scaling process seems like a very good empirical way to pre-commit to these kinds of things.

Demis Hassabis: Yes, exactly.

Dwarkesh Patel: I'm curious if you have a sense of, for example, when you're doing these evaluations, if it turns out your next model could help a layperson build a pandemic-class bioweapon or something, how you would think about making sure those weights are secure so that doesn't get out. And second, what would have to be true for you to be comfortable deploying that system? How would you make sure that capability isn't exposed?

Demis Hassabis: The secure model part, I think we've covered with cybersecurity—making sure that's well classed and you're monitoring all those things. If a capability was discovered like that through red teaming or external testing by government institutes or academia or independent testers, then we would have to fix that loophole depending on what it was. That might require a different kind of constitution, different guardrails, more RLHF to avoid it, or removing some training data. There could be a number of mitigations. The first part is making sure you detect it ahead of time—that's about the right evaluations, right benchmarking, and right testing. Then the question is how one would fix that before you deployed it. It would need to be fixed before it was deployed generally, for sure, if that was an exposure surface.

Dwarkesh Patel: Final question. You've been thinking in terms of the end goal of AGI at a time when other people thought it was ridiculous in 2010. Now that we're seeing this slow takeoff where we're actually seeing generalization and intelligence, what is the psychology of seeing this? Have you just priced it into your world model so it's not new news for you? Or is it like actually seeing it live, you're like, "Wow, something's really changed"? What does it feel like?

Demis Hassabis: For me, yes, it's already priced into my world model of how things were going to go, at least from the technology side. But I didn't necessarily anticipate that the general public would be that interested this early in the sequence. If say ChatGPT and chatbots hadn't gotten the interest they ended up getting—which I think was quite surprising to everyone, that people were ready to use these things even though they were lacking in certain directions, impressive though they are—then we would have produced more specialized systems built off the main track, like AlphaFold and AlphaGo and our scientific work. I think the general public would have only paid attention later down the road, in a few years time, when we have more generally useful assistant-type systems. So that's been interesting. It's created a different type of environment that we're now all operating in as a field. It's a little bit more chaotic because there's so many more things going on, so much VC money going into it, and everyone's almost losing their minds over it.

The thing I worry about is making sure that as a field, we act responsibly and thoughtfully and scientifically about this, and use the scientific method to approach it in an optimistic but careful way. I've always believed that's the right approach for something like AI. I just hope that doesn't get lost in this huge rush.

Dwarkesh Patel: I think that's a great place to close. Demis, thanks so much. Thank you for your time and for coming on the podcast.

Demis Hassabis: Thanks, it's been a real pleasure.

Dwarkesh Patel: Hey, everybody. I hope you enjoyed that episode. As always, the most helpful thing you can do is to share the podcast. Send it to people you think might enjoy it. Put it on Twitter, your group chats, etc. It just splits the world. I appreciate you listening. I'll see you next time. Cheers.