Dwarkesh PodcastDemis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold
EVERY SPOKEN WORD
140 min read · 27,719 words- 0:00 – 5:56
Nature of intelligence
- DHDemis Hassabis
... so I wouldn't be surprised if we had AGI-like systems within the next decade. It was pretty surprising to almost everyone, including the people, uh, who first worked on the scaling hypotheses that 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.
- DPDwarkesh Patel
When you think about, uh, superhuman intelligence, is it, like, con- still controlled by a private company?
- DHDemis Hassabis
A- as Gemini are becoming 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 what true multimodality means.
- DPDwarkesh Patel
Okay. Today, it is a true honor to speak with Demis Hassabis, who is the CEO of DeepMind. Demis, welcome to the podcast.
- DHDemis Hassabis
Thanks for having me.
- DPDwarkesh 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 subskills and heuristics?
- DHDemis Hassabis
Well, it's interesting because intelligence is so, uh, uh, broad and, um, you know, what we use it for is- is so sort of generally applicable. I think that suggests that, you know, there must be some sort of high level, uh, uh, common things in- i- you know, common kind of algorithmic themes, I think, around how the brain processes the world around us. So, um, of course, there- there then there are specialized parts of the brain that- that do specific things, um, but I think there are probably some underlying principles that underpin all of that.
- DPDwarkesh Patel
Yeah. 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, uh, asymmetrically better in that domain? Uh, w- uh, wouldn't we expect a sort of, like, general improvement across all the- all the different areas as well?
- DHDemis Hassabis
Well, I think you... First of all, I think you do actually sometimes get surprising improvement in other domains when you improve in a specific domain. So for example, uh, when these, uh, large models sort of improve at coding, that can actually improve their general reasoning. So there- there is some evidence of some transfer, although I think we would p- we would like a lot more evidence of that. Um, but also, you know, that's how, uh, the human brain learns too, is if we experience and practice a lo- a lot of things like chess or, you know, writing, creative writing, whatever that is, we also tend to specialize and get better at that specific thing, even though we're using, uh, sort of general learning techniques and general learning systems in order to, uh, sp- you know, to get good at that domain.
- DPDwarkesh Patel
Yeah. Well, what's been the most surprising example of this kind of tr- uh, transfer for you? Like, uh, where you see language and code or images and text? What- what's-
- DHDemis Hassabis
Yeah. I think probably, um... I mean, I'm hoping we're gonna see a lot more of this trying to transfer, but u- but I think, uh, things like getting better at coding, uh, and math and generally improving your reasoning, um, that is how it works with us as- as human learners, but, uh, I think it's interesting seeing that in the- in these- in these, uh, artificial systems.
- DPDwarkesh Patel
And can you see the sort of mechanistic way in which, uh, l- let's say in the language and code example, there is like, "I have found a place in a neural network that's getting better with both the language and the coders." Is it- is it that too f- too far down the weeds?
- DHDemis Hassabis
Yeah. Well, w- I don't think our analyst- analysis techniques are quite sophisticated enough to be able to hone in on that. Um, I think that's actually one of the areas that, um, a lot more research needs to be done on kind of mechanistic analysis of the representations that these systems build up and, um, you know, I sometimes like to call it virtual brain analytics in a way.
- DPDwarkesh Patel
(laughs)
- DHDemis Hassabis
It's a bit like doing, uh, fMRI or, uh, single cell recording from, uh, from a real brain. Uh, what's the analogous sort of analysis techniques for these artificial minds? And, um, there's a lot of great work going on on this sort of stuff. People like Chris Olah. Uh, I really like his work and a lot of computational neuroscience techniques, I think, could be brought to bear, uh, on, uh, analyzing these current systems we're building. In fact, I try to encourage a lot of my computational neuroscience friends to- to- to start thinking in that direction and applying their knowhow, um, uh, to the- to the- to the large models.
- DPDwarkesh Patel
Yeah. What do- what do other AI researchers not understand about human intelligence that you've- y- you, uh, that you have some sort of, like, insight on, given your neuroscience background?
- DHDemis Hassabis
I- I- I think, um, neuroscience has added a lot. Uh, if you look at the last sort of 10, 20 years that- that we've been at it at least, and- and y- I've been thinking about this for 30 plus years, um, I think in the earlier days of this sort of new wave of AI, I think neuroscience was providing a lot of interesting directional clues. So things like reinforcement learning, combining that with deep learning, you know, some of our pioneering work we did there, things like experience replay, um, even the notion of attention, which has become super important.
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
Um, a lot of those, uh, o- original sort of inspirations come from some understanding about how the brain works. Not the exact specifics, of course. You know, 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 kind of inspirational direction, maybe some ideas for architecture or algorithmic ideas or representational ideas. Um, and because you know, you know, the brain's in existence proof that general intelligence is possible at all, I think, um, you know, the- the history of human endeavors has been the... 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, uh, and sort of a question of when, not if.
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
Um, and that allows you to, you know, I think make progress a lot more quickly. So I think neuroscience has- has had a lot of, um, uh... has inspired a lot of the thinking, uh, at least in a soft way, uh, behind where we are today.Um, but as for, you know, going forwards, um, I think that there's still a lot of interesting, um, things to be resolved around planning and, um, how does the brain construct the right world models? Um, y- you know, I studied, for example, uh, how the brain does imagination, or you can think of it as, uh, mental simulation. So, how do we create, you know, very rich, visual-spatial simulations of the world in order for us to plan better?
- 5:56 – 16:31
RL atop LLMs
- DHDemis Hassabis
- DPDwarkesh Patel
Yeah. Actually, I'm curious how you think that will sort of interface with LLMs. So obviously, DeepMind is at the frontier and has been for many years, uh, you know, with systems like AlphaZero and so forth-
- DHDemis Hassabis
Yeah.
- DPDwarkesh Patel
... of having these agents who can, like, think through different steps to get to an end outcome.
- DHDemis Hassabis
Yeah.
- DPDwarkesh Patel
Um, uh, will this just be... is the path for LLMs to have this sort of, uh, t- tree search kind of thing on top of them? How- how do you think about this?
- DHDemis Hassabis
I think that's a super promising direction in my opinion. So, you know, we've got to carry on improving, uh, the large models, and we've got to carry on, um, basically making them more and more accurate predictors of the world. So in effect, making them more and more reliable world models. That's clearly a necessary, but I would say probably not sufficient, component of an AGI system. Um, and then on top of that I would... you know, we're working on things like AlphaZero, like planning mechanisms on top that make use of that model in order to make concrete plans to achieve certain goals in the world, um, and- and perhaps sort of chain, you know, uh, uh, chain thought together or lines of reasoning together, and maybe use search to k- kind of explore massive spaces of possibility. I think that's kinda missing from our current large models.
- DPDwarkesh Patel
Um, how do you get past the sort of, uh, uh, i- immense amount of compute that these approaches tend to require? So even the AlphaGo, uh, system was, uh, you know, a pretty expensive system, um, 'cause you had to do this sort of, uh, l- running LL- LLM on each node of the tree. Uh, h- how- how do you anticipate that'll get more, made more efficient?
- DHDemis Hassabis
Well, I mean, one thing is Moore's Law tends to k- it tends to- it tends to- it tends to help, uh, if- if, you know, o- over every- every year, of course, um, um, more computation comes in. But, um, we focus a lot on efficient, you know, sample-efficient methods and- and- and reusing, uh, existing data, things like experience replay, um, and also just looking at, uh, more efficient ways. I mean, the better your world model is, the more efficient your search can be. So, one example I always give with AlphaZero, our system to play Go and chess and, you know, any game is that, um, it's stronger than world champion level, human world champion level at- at all these games, um, and it uses a lot less search than a brute force method, um, like Deep Blue, say, to play chess. Deep Blue, uh, uh, uh, one of the traditional Stockfish or Deep Blue, um, uh, systems would maybe look at millions of, uh, possible moves for every decision it's gonna make. AlphaZero, uh, uh, and- and AlphaGo made, you know, looked at around ten- tens of thousands of, um, possible positions in order to make a decision about what to move next. But a human grandmaster, a human world champion, uh, probably only looks at a few hundreds of moves, even the top ones, in order to make their very, uh, good decision about what to play next.
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
So that suggests that obviously the brute force systems don't have any real model other than heuristics about the game. Uh, AlphaZero has quite a decent m- uh, uh, model, but the wor- but the human... you know, human top- human players have a much richer, much more accurate model than of Go or chess, so that allows them to make, you know, world-class decisions on a very small amount of search.
- DPDwarkesh Patel
Mm-hmm.
- DHDemis Hassabis
So I think there's still... there's a sort of trade-off there. Like, you know, if you improve the models, then I think your search can be more efficient, and therefore you can get further with your search.
- DPDwarkesh Patel
Yeah. I- I have two questions based on that, uh, the first being, with AlphaGo you had, um, a very conquered win condition of, you know, at the end of the day, do I win this game of Go or not? And you can reinforce on that. Uh, w- when you're just thinking of, like, an LLM putting out thought, what will, uh... do you think there will be this kind of ability to discriminate in the end whether that was, like, a good- a good thing to reward or not?
- DHDemis Hassabis
Well, of course, that's why we, you know, we pioneered and- and DeepMind's sort of famous for using games as a proving ground, um, partly because obviously it's efficient to research in that domain, but the other reason is obviously it's ver- it's, you know, extremely easy to specify a reward function, winning the game or improving the score, something like that, sort of built into most games. So that is the- the- that is the- one of the challenges of real world systems, is how does one define, uh, the right objective function, the right reward function, um, and the right goals? Um, and specify them in a- in- in, you know, in a general way but that's specific enough and- and- and actually points the system in the right direction? And, um, for real world problems, that can be a lot harder, but actually if you think about it, uh, a- and even scientific problems, uh, there are usually ways that you can specify the goal that you're after.
- DPDwarkesh Patel
Mm. And then when you think about human intelligence, you were just saying, well, you know, h- the humans thinking about these thoughts are just super sample efficient, um, how... Einstein coming up with relativity, right? There's just, like, thousands of possible permutations of the equations. Do you think it's also this sort of sense of, like, different heuristics of, like, I'm gonna try out this approach instead of this? Or is it a totally different way of approaching... coming up with that solution, uh, than w- you know, what AlphaGo does to plan the next move?
- DHDemis Hassabis
Yeah. Well, look, I think it's different because there's... our brains are not built for tr- doing Monte Carlo tree search, right?
- DPDwarkesh Patel
Yeah, yeah.
- DHDemis Hassabis
Um, it's- it's- it's just not the way, uh, uh, our organic brains would work. Um, so I think that i- in order to compensate for that, you know, people like Einstein have come up, you know, their brains have, using their intuition, and, you know, we can maybe come to what intuition is, but they use their sort of knowledge and their experience to build extremely, uh, you know, in Einstein's case, extremely accurate models of physics, including these sort of mental simulations. I think if you read about Einstein and how he came up with things, he used to visualize and sort of, uh, really kind of, um, uh, 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. And that allowed him to think these- these- these sort of very outlandish thoughts at the time. Um, so I think that it's- it's- it's the sophistication of the world models that we're building which then, you know, if you imagine your world model can get you to a certain node in a- in a tree that you're searching, and then you just do a little-... little bit of search around that node, that leaf node, and that gets you to these original places. But obviously, if your model is, and your judgment on that model is, is very, very good, then you can pick which leaf nodes you should sort of expand with search much more accurately. So therefore, overall, you do a lot less search. I mean, there's no way that, you know, any human could s- could do a kind of brute force search over any- any kind of significant space.
- DPDwarkesh Patel
Yeah, yeah, yeah. Um, a- a- a big sort of open question right now is whether, uh, RL will allow these models to do the self-play synthetic data to get over the da- data bottleneck. It sounds like you're optimistic about this. Uh, what do you think of this?
- DHDemis Hassabis
Yeah, I'm very optimistic about that. I mean, I think, uh... Well, first of all, there's still a lot more data I think that can be used, especially if one views like multimodal and video and these kind of things, and, uh, obviously, you know, society's adding more data all the time. Um, but I think, uh, uh, to the internet and things like that. But I think that, uh, there's a lot of scope for creating synthetic data. Um, we- we're looking at different ways, um, partly through simulation, u- using, uh, game, very realistic games environments, for example, to generate, uh, realistic data, um, but also, um, self-play. So that's where, um, systems, um, uh, interact with each other or- or converse with each other, um, and in the sense of, you know, what? Very well for us with AlphaGo and AlphaZero where we got the systems to play against each other and actually learn from each other's mistakes and- and build up a knowledge base that way. And I think there are some good analogies for that, it's a little bit more complicated, but to- to- to build a general kind of world data.
- DPDwarkesh Patel
Mm-hmm. How do you get to the point where these models, the, um, the sort of synthetic data they're outputting and the self-play they're doing, uh, is- is not just more of what they've already got in their dataset, but is something they haven't seen before? It w- do, you know what I mean? To actually improve the abilities.
- DHDemis Hassabis
Yeah, so there I think there's a whole, uh, science needed and- and I think we're still in the nascent stage of this, of data curation and data analysis, so actually analyzing, uh, the holes that you have in your data distribution, uh, and this is important for things like fairness and bias and other stuff, to remove that from the system is to- is to try and really make sure that your dataset is representative of the distribution you're trying to learn. And, uh, and, you know, there are many tricks there one can use, like overweighting or replaying certain parts of the data, or you could imagine if you identify some- some gap in your dataset, that's where you put your synthetic generation capabilities to work on.
- DPDwarkesh Patel
Yep. So n- and that, you know, but nowadays people are paying attention to, uh, the- the- the RL stuff that Alpha- uh, DeepMind did many years before, what are the sort of, uh, either 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, right? Like, there was a time wh- where people weren't paying attention to scaling.
- DHDemis Hassabis
Yeah.
- DPDwarkesh Patel
What's the thing now where it's, like, totally underrated?
- DHDemis Hassabis
Well actually, I think that, you know, there- there's the- the history of the sort of last couple of decades has been things coming in and out of fashion, right?
- DPDwarkesh Patel
Yeah, yeah.
- DHDemis Hassabis
And- and I- I do feel like, um, a while ago when, you know, maybe five plus years ago when we were pioneering with AlphaGo and before that DQN where it was the first system with, you know, that worked on Atari but our f- our first big system really more than 10 years ago now that scaled up Q-learning and reinforcement learning techniques to deal, you know, combine that with deep learning, uh, to create deep reinforcement learning and then, uh, use that to scale up to complete some, you know, master some pretty complex tasks like playing Atari games just from the pixels. And, uh, I do actually think a lot of those ideas, um, 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, uh, uh, some of those older ideas together with the newer ones.
- DPDwarkesh Patel
Uh, is there any potential for something to come, uh, the AGI to eventually come from just a pure RL approach? Like, the- the way we're talking about it, it sounds like there'll be, uh, the LLM will help form the b- g- right prior and then this sort of research will go on top of that.
- 16:31 – 24:13
Scaling and alignment
- DHDemis Hassabis
- DPDwarkesh Patel
Okay. The- this sounds like the answer to the question I'm about to ask which is, um, what- what is you- uh, as somebody who's been in this field for a long time and seen different trends come and go, what do you think that strong version of the scaling hypothesis gets right and what does it get wrong? Just the idea that you just throw enough compute at a wide enough distribution of data and you get intelligence.
- DHDemis Hassabis
Yeah, look, I- my- my view is this isn't kind of an empirical question right now.
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
So I think it was pretty surprising to almost everyone including the people, uh, you know, who- who first worked on the scaling hypothesis that how far it's gone. In a way I- I mean, I sort of look at, uh, the large models today and I think they're almost unreasonably effective for what they are, you know? Um, I think it's pretty surprising some of the properties that emerge, things like, you know, it's clearly in my opinion got some form of concepts and abstractions and some things like that, and I think if we were talking five plus years ago I would've said to you maybe we need an additional algorithmic breakthrough, uh, in order to- to do that like, um, you know, maybe more like the brain works. And- and I think that's still true if we want explicit abstract concepts, neat concepts, but it seems-... is that these systems can implicitly learn that. Another really interesting, I think, uh, unexpected thing was that these systems have some sort of grounding. Um, you know, even though they don't experience the world multimodally, or at least until more recently when we have the multimodal models. And, uh, that's surprising that- that the amount of information that can be, uh, uh, and- and models that can be built up just from language. And I think that I have hy- some hypotheses about why that is, um, I think we get some grounding through the RLHF feedback systems because obviously the human raters are- are by definition grounded, uh, uh, uh, uh, grounded people. We're grounded, right, in the, in reality, so our feedback's also grounded. So perhaps there's some grounding coming in through there and also maybe language contains more grounding, you know, if you're in the, if you, if you're able to ingest all of it-
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
... than we, than we perhaps thought or linguists per- perhaps thought before.
- DPDwarkesh Patel
Right.
- DHDemis Hassabis
So there's actually some very interesting philosophical questions-
- DPDwarkesh Patel
Yeah, totally.
- DHDemis Hassabis
... that I think we haven't, we- w- people haven't even really scratched the surface of yet, uh, uh, that- that looking at the advances that have been made, um, you know, it's quite interesting to think about where it's gonna go next. But in terms of your question of, like, the, you know, large models, I think we've got to push scaling as- as hard as we can and that's what we're doing here, and, you know, it's an empirical question whether that will hit an asymptote or a brick wall, and there are, you know, different people argue about that. But actually I think we should just test it. I think no one knows, um, and but in the meantime we should also double down on innovation and invention, and this is something that- that- that- that the Google Research and DeepMind and Google Brain have- have- have- have, you know, we pioneered many, many things over the last decade, that's something that's our bread and butter.
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
And, um, you know, you can think of half our effort as to do with scaling and half our efforts to- in- do with inventing the next architectures, the next algorithms that will be needed, um, knowing that you've got this scaled larger and larger model coming along the lines.
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
So I- I- I- I... My- my betting right now, but it's a loose betting, is that you would need both, um, but I think, you know, it- it's- you got to push both of them as hard as possible and we're in a lucky position that we can do that.
- DPDwarkesh Patel
Yeah. I want to ask more about the grounding. So 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, um, operate in domains where we just can't generate enough human labels just because we're not smart enough, right? So if it does like a million line pull request, you know, how- how do we d- tell it like, "This is, this is within the constraints of our morality and the end goal we wanted and this isn't"? And the other is, it sounds like you're saying more of the compute so far we've been doing, you know, next token prediction and in some sense it's a guardrail because you're t- you have to talk as a human would talk and think as a human would think.
- DHDemis Hassabis
Yeah.
- DPDwarkesh Patel
Now if additional compute is going to come in the form of, uh, reinforcement learning where just get to the eng- objective, uh, we can't really trace how you got there.
- DHDemis Hassabis
Yeah.
- DPDwarkesh Patel
Um, when you combine those two, how worried are you that the sort of grounding goes away?
- DHDemis Hassabis
Well, look, I- I think, um, uh, i- if the grounding, you know, if it's not properly grounded the system won't be able to achieve those goals properly, right? I think... So I think in a sense 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. Um, I do actually think that a- as these systems and- and things like Gemini are becoming more multimodal, um, and we start ingesting things like video and- and- and- and- and- and, you know, audiovisual data as well as text data and then, you know, the system starts correlating those things together, um, I do s- I think that is a form of- of proper grounding actually. So- so I do think our systems are going to start to understand, you know, the physics of the real world better. And then one could imagine the active version of that as being in a very realistic simulation or game environment where you're starting to learn about what your actions do in the world and, um, and how that affects, uh, uh, uh, uh, uh the world itself, the world stay itself, but also what next learning episode you're getting.
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
So, you know, these RL agents we- we've always been working on and pioneered like AlphaZero and AlphaGo, um, they actually affect... they're active learners. What they decide to do next affects what, uh, their next learning, uh, piece of data or experience they're going to get. So there's this very interesting sort of feedback loop, and of course if we ever want to be good at things like robotics, we're gonna have to understand how to act in the real world.
- DPDwarkesh Patel
Mm. Yeah. So there's a grounding in terms of will the capabilities be able to proceed or will they be like enough in touch with reality to be able to, like, do the things we want? And there's another sense of grounding of, um, we've gotten lucky in the sense that since they're trained on human thought, they like maybe think like a human.
- DHDemis Hassabis
Yeah.
- DPDwarkesh Patel
To what extent does that stay true when more of the compute for training comes from just, did you get the right ou- outcome? And not guardrailed by like are you like proceeding on the next token as a human would?
- DHDemis Hassabis
Yeah.
- DPDwarkesh Patel
Maybe the broader question I'll like pose to you is, um, uh, and this is what I asked Shane as well, wh- wha- what would it take to align a system that's smarter than a human, maybe thinks in alien concepts, uh, and you can't like really monitor the million line pull request because it's, uh, you can't really understand the whole thing?
- DHDemis Hassabis
Yeah, I mean-
- DPDwarkesh Patel
And you can't give labels.
- 24:13 – 28:42
Timelines and intelligence explosion
- DHDemis Hassabis
that it's building.
- DPDwarkesh Patel
Yeah. Yeah. Um, stepping back a bit, I'm curious what your timelines are. So Shane said he- his like, I think modal outcome is 2028. I think that was maybe his median.
- DHDemis Hassabis
Yeah.
- DPDwarkesh Patel
Uh, what, what is yours?
- DHDemis Hassabis
Yeah. Well, I, you know, I, I, I, I, I don't have prescribed kind of specific numbers to it, because I think there's, there's so many unknowns and uncertainties, and, and, and, um, you know, human ingenuity and endeavor comes up with surprises all the time. So, that could meaningfully move the, the, the, the, the timelines. But I will say that when we started DeepMind back in 2010, you know, we thought of it as a 20-year project. And, and actually I think we're on track-
- DPDwarkesh Patel
(laughs)
- DHDemis Hassabis
... which is kind of-
- DPDwarkesh Patel
That's a very good, yeah.
- DHDemis Hassabis
... amazing for a 20-year project, 'cause usually they're always 20 years away.
- DPDwarkesh Patel
Sure, sure. Yeah. Yeah.
- DHDemis Hassabis
Right? So that's the joke about, you know, whatever it is-
- DPDwarkesh Patel
Fusion.
- DHDemis Hassabis
... fusion, quantum, AI, you know, take your pick. And, um, but I think we, you know, I think we're on track, so I wouldn't be surprised if we had, uh, AGI-like systems, uh, within the next decade.
- DPDwarkesh Patel
Mm-hmm. And do you buy the model that once you have an AGI, you can ha- you have a system that basically speeds up further AI research? Maybe not like in a overnight sense, but you know, over the course of months and years, you have much faster progress than you would've otherwise had?
- DHDemis Hassabis
I, I, I think that's potentially possible. Um, I, I think it partly depends what we, uh, decide, we as society decide to use the first AGI, nascent AGI systems, or even proto-AGI systems for. Um, so, uh, you know, even the current LLMs, uh, seem to be pretty good at coding. So, uh, and, you know, we have systems like AlphaCode. We've also got th- theorem proving systems. So, one could imagine, uh, combining these ideas together and, and, and, and, and making them a lot better, and then I, I, I could imagine these systems being quite good at, at, at, at designing and helping f- uh, us build future versions of themselves. Um, but we also have to think about the safety implications of that, of course.
- DPDwarkesh Patel
Yeah. I'm curious what you think about that. So, I mean, I'm not saying this is happening this year or anything, but eventually you'll be developing a model where during the process of developing you think, you know, there's some chance that once this is fully developed, it'll be capable of like an intelligence explosion like Dynamic. Um, what would have to be true of that model at that point where you're like, "I, I, you know, I've seen these specific evals. I've like, I've, I've like understand its internal thinking enough, and like its future thinking that I'm comfortable continuing development of the system"?
- DHDemis Hassabis
Yeah. Well, look, we need, um, we need a lot more understanding of the systems than we do today before I would be even confident of even 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, is, is come up with the right, uh, evaluations and metrics, and maybe ideally formal proofs. But, you know, it's gonna be hard for these types of systems, but at least empirical, uh, uh, bounds around what these systems can do. Um, and that's why I think about things like deception and, as being quite root know traits that you don't want, because if, if you're confident that your system is, is, is, is, is tell- is, is sort of exposing what it actually thinks, then you could potentially... Or that opens up possibilities of using the system itself to explain aspects of itself to you.
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
Um, the way I think about that actually is like, um, if I was to play a game of chess against Garry Kasparov, right, which, which I played in the past-
- DPDwarkesh Patel
(laughs)
- DHDemis Hassabis
... or Magnus Carlsen, you know, the amazing chess players of all time.
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
You, you, I wouldn't be able to come up with a move that they could, but, but they could explain to me, um, why they came up with that move, and I could understand it, uh, uh, uh, post hoc, right? And, and that's the sort of thing one could imagine, uh, uh, uh, one of the, uh, uh, uh, um, capabilities that we could make use of these systems, is for them to explain it, uh, it to us then even maybe the proofs behind why they're thinking something, certainly in a mathematical se- uh, any mathematical problem.
- DPDwarkesh Patel
Got it. Um, do, do you have a sense of what the, the converse answer would be? So, what would have to be true where tomorrow morning you're like, "Oh man, I, I didn't anticipate this." You see some specific observation tomorrow morning where like, "We, we got to stop Gemini 2 training." Like what, what is, is there, what would specifically be the opposite?
- DHDemis Hassabis
Yeah. I could imagine that, like, um... And this is where, uh, you know, things like the sandbox simulations.
- DPDwarkesh Patel
Yeah, yeah.
- DHDemis Hassabis
I, I would hope we'd, we're, we're experimenting in a, in a safe, secure, uh, environment. And then, you know, something happens in it where, um, very unexpected happens, a new unexpected capability or something that we didn't want, you know, explicitly told the system we didn't want, that it did but then lied about. You know, these are the kinds of things where one would want to then dig in-
- DPDwarkesh Patel
Yeah, yeah.
- DHDemis Hassabis
... carefully, um, you know, now with the systems that are around today, which are not dangerous in my opinion d- today, but in a few years they might be, uh, have, have s- potential. Um, and then you would sort of ideally kind of pause, and then really get to the bottom of, um, uh, why it was doing those things before one continued.
- 28:42 – 35:30
Gemini training
- DHDemis Hassabis
- DPDwarkesh Patel
Mm. Yeah. G- going back to Gemini, I'm curious, uh, what the bottlenecks were in the development. Um, like why not make it immediately one order of magnitude bigger, uh, if, if like scaling works?
- DHDemis Hassabis
Well, look, first of all, there are practical limits, how much compute-
- DPDwarkesh Patel
Yeah, yeah.
- DHDemis Hassabis
... that can you actually fit in one data center.
- DPDwarkesh Patel
Sure, sure.
- DHDemis Hassabis
And actually we, you know, you're, you're up, you're bumping up against very interesting, um, uh, uh, dist- you know, distributor computing kind of challenges, right? Where fortunately we have some of the best people in the world on, on those challenges, and, and you know, cross data center training, all these kinds of things. Very interesting challenges, hardware challenges, and we have our TPUs and so on that we're building and designing all the time, uh, as well as using GPUs. And so, um, there's all of that. And then you also have to... The scaling laws, you know, you, they don't, they don't just work by magic. You sort of, you still need to scale up the hyper parameters, 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, uh, and that's a bit of an art form in a way, and you have to sort of almost get new data points. If you try and extend your predictions, extrapolate them, say, several orders of magnitude out, sometimes they don't hold anymore, right? Because, um, new capabilities, there can be step functions-
- DPDwarkesh Patel
Right.
- DHDemis Hassabis
... in, in terms of new capabilities, and, and, and, and some things just, some things hold and other things don't. So, often you, you do need those intermediate data points actually to, to correct, uh, uh, some of your hyper parameter optimization and other things, so the, the, the scaling law continues to be true. So, um, so there's sort of various practical limitations on to, on to that. Um, so, you know-... kind of one order of magnitude is about probably the maximum that you want to, that you want to carry on, uh, you want to sort of do between each, uh, each era.
- DPDwarkesh Patel
Oh, that's so fascinating. Uh, you know, in the GPT-4 technical report, they say that they were able to predict the, the training loss, um, of, you know, me- tens of thousands of times less compute than GPT-4, that they could see the curve. But the point you're making is that the actual capabilities that loss implies, um, it may not be so clear.
- DHDemis Hassabis
Yeah, the downstream capabilities-
- DPDwarkesh Patel
Yeah, yeah.
- DHDemis Hassabis
... sometimes don't follow from the... You can often predict the-
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
... the core metrics like training loss or, or something like that.
- DPDwarkesh Patel
Yeah, yeah.
- DHDemis Hassabis
But then, um, it doesn't actually translate into MMLU or-
- DPDwarkesh Patel
Sure, sure.
- DHDemis Hassabis
... math or some other actual, uh, capability that you care about.
- DPDwarkesh Patel
Right.
- DHDemis Hassabis
It's... They're not, they're not necessarily linear all the time.
- DPDwarkesh Patel
Yeah, yeah.
- DHDemis Hassabis
So there's sort of nonlinear effects there.
- DPDwarkesh Patel
What was the biggest surprise to you during the development of Gemini of s- some- some- something like this happening?
- DHDemis Hassabis
Um, well, I, I mean, the... I don't... I wouldn't say there was one big surprise, but it's, it was very interesting, you know, trying to train things at that, uh, that size, and, and, and learning about, um, uh, all sorts of things, from organizational, how to babysit such a system, and, and to track it. And, and I think things like getting a better, uh, understanding of, of the, the metrics you're optimizing versus the, the final capabilities that you want. Um, I would say that's still not a perfectly understood s- uh, uh, uh, uh, mapping. But, but it's an interesting one that we're getting better and better at.
- DPDwarkesh Patel
Yeah, yeah. There, there's a perception that maybe other labs are more compute efficient, uh, than, uh, DeepMind has been with Gemini. I, I don't know what you make of that perception.
- DHDemis Hassabis
Uh, I don't think that's the case. I mean, you know, it's, uh, uh, uh, I, I think that, that actually Gemini 1.0 used roughly the same amount of compute, maybe slightly more than, than what was rumored for GPT-4. I don't know exactly what was, was used. So, um, I think it's, was in the same ballpark. Um, I think we're very efficient with our compute, and we use our compute for many things. One is not just the scaling, but going back to earlier to these n- more innovation and, and, ah, ideas. You've got to... You know, it's only useful, a new innovation, a new invention if it also can scale. So, so in a way, um, you also need quite a lot of compute to do new invention, uh, because you've got to test many things at at least some reasonable scale and make sure that they work at that scale. And also, some new ideas may not work at a toy scale but do work at a larger scale. And in fact, those are the more valuable ones. So you actually, if you think about that exploration process, you need quite a lot of compute to be able to do that. Um, I mean, the good news is, is I think, you know, we... That you... We're pretty lucky here at Google that we... I think we... This year certainly we're gonna have the most compute by far of, of any sort of research lab. And, you know, we hope to make very efficient and good use of that in terms of both scaling, uh, a- and the capability of our systems and also new inventions.
- DPDwarkesh Patel
Yeah. What's been the biggest surprise to you, uh, if you go back to, uh, yourself in 2010 when you were starting DeepMind in terms of what AI progresses looked like? Did you anticipate back then that it would in some large sense amount to spend as... Uh, you know, dumping billions of dollars into these models? Or did you have a different sense of what it would look like?
- DHDemis Hassabis
We thought that, and actually, you know, a few... I know you've interviewed my, my colleague Shane, and, and, and he, he always thought that in, in terms of, like, um, compute curves and, and then maybe comparing roughly to, like, the brain and how many neurons and synapses there are very loosely. But we're actually interestingly in that kind of regime now, roughly in the right order of magnitude of, you know, number of synapses in the brain and, and, and the sort of compute that we have. But I think more fundamentally, you know, we, we always thought that, um, we bet on generality and learning, right? So tho- those were always at the core of the... Any technique we would use. That's why we triangulated on reinforcement learning and search and, and, and, and deep learning, right? As three types of algorithms that, that would scale and, um, and, and would be very general, and, and not require a lot of handcrafted human priors, which we thought was the sort of failure mode really of, of the efforts to build AI, uh, in the '90s, right? Places like MIT where, where there were very, you know, logic-based systems, expert systems. You know, masses of hand-coded, handcrafted human information going into it that turned out to be wrong or, or too rigid. So we wanted to move away from that, and I think we spotted that trend early and, uh, became... You know, and obviously we, we used games as our proving ground, and we did very well with that. And I think all of that was very successful, and I think ins- maybe inspired others, uh, to... You know, things like AlphaGo I think was a big moment for inspiring many others to think, "Oh, actually these systems are ready to scale." And then, of course, with the advent of transformers invented by our colleagues at Google Research and Brain, that was the then, you know, the, the type of deep learning that allowed us to ingest masses of amounts of information. And that, uh, of course has really turbocharged where we are today. So I think that's all part of the same lineage. Um, I... You know, we couldn't have predicted every twist and turn there, but I think the general direction we were going in, um, uh, was the right one.
- DPDwarkesh Patel
Yeah. A- a- and in fact, it's, it's, like, fascinating because actually if you, like, read your old papers or Shane's old papers, uh, Shane's thesis I think in 2009 he said like, "Well, you know, the way we test for AI is if it... Can you compress Wikipedia?" And that's, like, literally the loss function of LLMs.
- 35:30 – 40:42
Governance of superhuman AIs
- DHDemis Hassabis
of that.
- DPDwarkesh Patel
Yeah, yeah, yeah. Um, when, when you... When you extrapolate all this out forward and you think about, uh, superhuman intelligence or as, um... Uh, w- uh, f- like, what does that landscape look like to you? Is it, is it, like, con- still controlled by a private company? Like, what should the governance of that look like, uh, concretely?
- DHDemis Hassabis
Yeah, look, I, I would love... Um, you know, I think that this has to be, uh, uh... This is so consequential, this technology. I think it's much bigger than any one company or, or, or, or even industry in general. I think it has to be a big collaboration with many stakeholders from civil society, academia, government. And the good news is, I think with the popularity of the recent chatbot systems and so on, I think that has woken up, uh, uh, many of these other parts of society that this is coming and what it will be like to interact with these systems. And that's great-... rate. So it's opened up lots of doors for very good conversations. I mean, an example of that was the Safety Summit at, in the UK hosted a few months ago, which I thought was a big success to start getting this international dialogue going. And, and, and, you know, I think it, the whole of society needs to be involved in deciding, what do we want to deploy these models for? How do we want to use them? What do we not want to use them for? You know, I think we've got to try and get some international consensus around that, uh, and then also making sure that the benefits of these systems, uh, uh, benefit everyone, you know, for the good of everyone and society in general. And that's why I push so hard things like AI for science, and, and I hope that, you know, with things like our spin-out, Isomorphic, we're gonna start curing diseases, you know, terrible diseases with AI and accelerate drug discovery. Amazing things, climate change and other things, I think big challenges that face us, uh, and face humanity. Um, massive challenges actually, which I'm optimistic we can solve, uh, because we've got this incredibly powerful tool coming along down the line of AI, uh, that we can apply and I think help us in, uh, solve many of these problems. So, you know, ideally, we would have a big, uh, uh, uh, consensus around that and, and, and a big discussion, you know, sort of almost like the UN level if possible.
- DPDwarkesh Patel
Mm. You, you, you know, o- one interesting thing is if you look at these systems, they... L- you chat with them and they're, they're immensely powerful and, uh, intelligent, um, but i- it's interesting to the extent of which they haven't, like, automated large sections of the economy yet, um, whereas if five years ago I showed you, uh, Gemini, you'd be like, "Wow, this is like di- you know, totally coming for a lot of things." So h- how do you account for that? Like, what's going on where it hasn't, uh, had, had the broader impact yet?
- DHDemis Hassabis
Yeah. I think it's... We're still... I think that just shows we're still at the beginning of, of, of this new era.
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
Um, and I think that for these systems, I think there are some interesting use cases, you know, um, you know, where you can use things to sum... You know, these, these, these chatbot systems to summarize, uh, stuff for you and, and maybe do some simple writing and, uh, uh, maybe more kind of boilerplate-type writing. But that's only a small part of wha- you know, we, we all do every day, so I think for more general use cases, um, I think we n- still need new capabilities, uh, things like, um, planning and search but also maybe things like personalization and, uh, memory, episodic memory. So not just long context windows, but actually remembering what I s- what we'd spoke about 100 conversations ago. Um, and I think once those start coming in, I mean, I'm really looking forward to things like recommendation systems that, that help me find better, more enriching material, whether that's books or films or music and so on. You know, I would use that type of system every day, so I think we're just scratching the surface of, uh, uh, what these AI, say, assistants could actually do, uh, for us in our general everyday lives, and also in our work contexts as well. I think they're not reliable yet enough to do things like science with them, but I think one day, you know, once we fix factuality and grounding and other things, um, I think they could end up becoming, like, you know, the world's best research assistant for, for you as a, as a scientist or as a, as a, as a clinician.
- DPDwarkesh Patel
Hmm. Uh, I wanna ask about memory, by the way. Um, you had this fascinating paper in 2007 where you, uh, talk about the links between memory and imagination-
- DHDemis Hassabis
Yeah.
- DPDwarkesh Patel
... and how they, in some sense, are very similar. Um, uh, y- people often claim that these models are just memorizing. How do you think about that claim that people make? Um, is, is memorization all you need? Because in some d- some deep sense, that's compression or... But, you know, what's your intuition here?
- DHDemis Hassabis
Yeah. I mean, sort of at the limit one, one maybe could try and memorize everything, but it wouldn't generalize out of, out of your distribution, and I think these systems are clearly... I think the early, the early, uh, um, criticisms of these early systems, uh, were that they were just regurgitating and memorizing. But I think clearly, the new era, the Gemini GPT-4 type era, they are def- definitely generalizing to new constructs. Um, so... But actually, m- you know, in my thesis in tha- in that paper particularly, uh, that started that area of imagination in neuroscience was showing that, you know, first of all, memory, certainly at least human memory, is a reconstructive process. It's not a videotape, right? We sort of put it together back from components that seems familiar to us, the, the ensemble, and that's what made me think that imagination might be the same thing except, in this case, you're using the same semantic components, but now you're putting it together into a way that your brain thinks is novel, right?
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
For a particular purpose like planning. And, um, and so I do think that, uh, that kind of idea is still probably missing from our current systems, this sort of pulling together different, um, parts of your world model to simulate something new that then helps with your planning, uh, which is what I would call
- 40:42 – 47:00
Safety, open source, and security of weights
- DHDemis Hassabis
imagination.
- DPDwarkesh Patel
Yeah, for sure. So yeah. Th- now, now you guys have the best models in the world, um, y- you know, with the Gemini models. Uh, uh, do you ha- do you have s- uh, do you plan on putting out some sort of framework like the other two major AI labs have of, you know, once we see these specific capabilities, unless we have these specific safeguards, we're not gonna continue development or we're not gonna ship the product out, uh
- NANarrator
(...)
- DHDemis Hassabis
Yes, we, we have... Actually, we... I mean, we have already lots of internal checks and balances-
- DPDwarkesh Patel
Sure.
- DHDemis Hassabis
... but we're gonna start publishing, actually.
- DPDwarkesh Patel
Oh, great.
- DHDemis Hassabis
You know, sort of watch this space as we're working on a whole bunch of, um, blog posts and technical papers that, uh, we'll be putting out in the next few months that, um, you know, along the similar lines of things like responsible scaling laws and so on. We have those, uh, uh, implicitly internally and various, uh, safety councils and so on, people like Shane, Chiara, and so on. Um, but, but, uh, it's time for us to talk about that more publicly I think, so we'll be doing that throughout the course of the year.
- DPDwarkesh Patel
Oh. Th- that's great to hear. Um, and I, I... Another thing I'm curious about is, um... So it's not only the risk of, like, uh, you know, the, the deployed model being something that people can use to do bad things, but also, uh, rogue actors, be- foreign agents, so forth, being able to steal the weights and then fine-tune them to do crazy things. Um, uh, how do you think about securing the weights to make sure something like this doesn't happen? Uh, making sure a, a very, like, key group of people have access to them and so forth?
- DHDemis Hassabis
Yeah, it's interesting. So first of all, there's sort of two parts to this. One is security, one is open source. Maybe we can discuss. But the security, I think, is super key, like ma- just the sort of, um, normal cybersecurity type things. And I think we're lucky at Google DeepMind, we're kind of behind Google's firewall and, and cloud protection which is, you know, I think best in, you know, best in class in the world.... corporately. So we already have that protection, and then behind that, we have specific, uh, uh, DeepMind, uh, uh, uh, protections within our code base. So it's sort of a double layer of protection. So I feel pretty good about that, that that's... I mean, we, you know, you can never be complacent on that, but I feel it's, it's, it's already sort of best in the world in terms of cyber, uh, uh, defenses. Um, but we got to carry on improving that. And again, things like the hardened sandboxes could be a way of doing that, uh, as well. And, and maybe even there are, um, you know, uh, specifically secure data centers or hardware solutions to this too that we're thinking about. I think that maybe in the next three, four, five years, we would also want, um, air gaps and various other things that are known in the security community. So I think that's key, and I think all frontier labs should be doing that because otherwise, you know, nation states and other things, rogue, rogue nation, you know, states and other, other dangerous actors, um, that, that there'd be obviously a lot of incentive for them to, to steal things like the weights. Um, and then, you know, of course, open source is another interesting question, which is, we're huge proponents of open source and open science. I mean, almost every, you know, we've published thousands of papers and, and things like AlphaFold and Transformers, of course, and AlphaGo, all of these things we put out there into the world, uh, uh, published and, and open sourced many of them. Uh, GraphCast most recently, our weather prediction system. But when it comes to, uh, uh, you know, the core technology, the foundational technology and very general purpose, I think the question I would have is, um, if you, you know, uh, uh, for sort of open source proponents is that, how does one, uh, uh, stop bad actors, um, individuals or rogue, you know, up to rogue states, um, taking those, uh, same open source systems and repurposing them, because they're general purpose, for harmful ends, right? So we have to answer that question.
- DPDwarkesh Patel
Yeah, yeah.
- DHDemis Hassabis
Uh, uh, and, and I haven't heard a compelling... I mean, I don't know what the answer is to that, but I haven't heard a compelling, clear answer to that from, uh, uh, uh, proponents of just sort of open sourcing everything.
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
So I think there has to be some balance there, but, um, you know, obviously it's a complex question of, of to what that is.
- DPDwarkesh Patel
Yeah, yeah. I, I feel like tech doesn't get the credit it deserves for, like, funding, you know, hundreds of billions of dollars' worth of R&D.
- DHDemis Hassabis
Yeah.
- DPDwarkesh Patel
Um, and, you know, obviously we have DeepMind with systems like AlphaFold and so on.
- DHDemis Hassabis
Yeah.
- DPDwarkesh Patel
Um, but when we talk about securing, uh, the weights, uh, you know, as we said, like maybe right now, it's not, uh, something that, like, is gonna cause the end of the world or anything. But as these systems get better and better, the worry that, uh, yeah, some, a foreign agent or something gets access to them. Presumably right now, there's like dozens to hundreds of researchers who have access to the weights. How do you... Uh, what, what's the plan for, like, getting into, like, the situation, or getting the weights in a situation room, so if you're like, well, if you need to access to them, you, you, it's like, you know, some extremely strenuous process. You, nobody, no individual can really take them out.
- DHDemis Hassabis
Yeah. Yeah, I mean, one has to balance that with, with, with allowing for collaboration-
- DPDwarkesh Patel
Sure.
- DHDemis Hassabis
... and speed of progress. Actually, another interesting thing is you, of course you want, uh, uh, you know, brilliant independent researchers from academia or, or things like the UK AI Safety Institute, and US1, um, to be able to, uh, uh, uh, uh, kind of red team these systems. So, so one has to expose them to a certain extent, um, although that's not necessarily the weights. Um, and then, you know, we have a lot of processes in place about, uh, making sure that, um, you know, only if you need them, that, that you have access to, you know, those people who need access have access. Um, and right now, I think we're still in the early days of those kinds of systems being at risk. And as that, as these systems become more powerful and more general and more capable, um, I think one has to look at the, the access question.
- DPDwarkesh Patel
Mm-hmm. Uh, so some of these other labs have specialized in different things, uh, relative to safety, like, uh, Anthropic, for example, with interpretability.
- DHDemis Hassabis
Yes.
- DPDwarkesh Patel
And, um, uh, do you have some sense of where, uh, you guys might have a edge? Where, as so that, uh, you know, now that you have the frontier model, you're gonna keep, scale up safety, where you guys are gonna be able to put out the b- the best frontier research on safety?
- DHDemis Hassabis
Yeah. I think, you know, well, we, we helped pioneer RLHF and other things like that-
- DPDwarkesh Patel
Sure, yeah.
- DHDemis Hassabis
... which can also be obviously used for performance-
- DPDwarkesh Patel
Right.
- DHDemis Hassabis
... but also for safety. Um, I think that, uh, um, you know, a lot of the self-play ideas and these kinds of things could also be used potentially to, to auto-test, uh, a lot of the, the, the, the boundary conditions that you have with the new systems. I mean, part of the issue is that, um, with these sort of very general systems, uh, there's so much surface area to, to cover, like, about how these systems behave. So I think we are going to need some automated, uh, testing. And, and again, with things like simulations, I think, and games environment, very realistic environments, uh, virtual environments, I think we have a long history in, in that and, and using those kinds of systems and making use of them for, for, for building AI algorithms. So I think we can leverage all of that, uh, history. Um, and then, you know, around at Google, we're very lucky. We have some of the world's best cyber security experts, hardware designers. So I think we can bring that to bear and, and, you know, for security and safety
- 47:00 – 54:18
Multimodal and further progress
- DHDemis Hassabis
as well.
- DPDwarkesh Patel
Great. Great. Um, let's talk about Gemini.
- DHDemis Hassabis
Yeah.
- DPDwarkesh Patel
Um, so, you know, now, you know you guys have the best model in the world. Um, I'm, and so, uh, uh, I'm curious, you know, the default way to interact with these systems has been through chat, uh, 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?
- DHDemis Hassabis
Yeah, I think we're just at the beginning of actually understanding what a full multimodal model system, uh, how exciting that might be to interact with. And, and, and, uh, it'll be quite different to, I think, what we're used to today with the chatbots. I think, um, uh, uh, the next versions of this over the next year, 18 months, you know, maybe we'll have some contextual understanding around the environment around you through a camera or whatever it is, a phone. Um, you know, I could imagine that as the next awesome glasses or the next step. Um, and then I think that, that we'll start becoming more fluid in understanding, oh, let's, let's, let's, let's sample from a video. Let's, let's use voice. Um, uh, um, maybe even eventually things like touch and, and, you know, if you think about robotics and other things, uh, you know, sensors, other types of sensors. So I think, uh, uh, 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 multi-modality means.
- DPDwarkesh Patel
Mm-hmm. Um, uh, on the robotics subject, uh, 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. Um, uh, I mean, you guys have put out different things like robot transformer and other things.
- DHDemis Hassabis
Yes.
- DPDwarkesh Patel
How, what... Do you think that's still a bottleneck for robotics progress or will we see, uh, progress in the world of atoms as well as robots?
- DHDemis Hassabis
Yeah. Well, we're very excited about our progress with things like GATO and, and, and, and, and RT-2, you know, robotic transformer. And, uh, and we actually think... Um, so we've always liked robotics and we've, we've had, you know, amazing research and now we still have that going now, because we like the fact that it's a data-... poor regime, 'cause that pushes us on s- on very interesting research directions that we think are gonna be useful anyway, like sampling efficiency and data efficiency in general and transfer learning, uh, learning from simulation, transferring that to reality, all of these very, you know, sim-to-real, all of these very interesting, uh, actually general challenges that we would like to solve.
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
Um, so the control problem. So, um, we've always pushed on that. And actually, I think, uh, uh, uh, so, so Ilya's right, that, that is more challenging because of the data problem. Um, but it's also, I think we're starting to see the beginnings of, um, these large models being transferrable, uh, to the robotics regime, learning in the general domain, language domain, and other things. And then just treating tokens like GATO as any type of token. You know, the token could be an action, it could be a word, it could be, uh, part of an image, a pixel, or whatever it is. And that's what I think true multimodality is, and to begin with, it's harder to train a system like that than a straightforward, uh, uh, text language system. Um, but, uh, actually, you know, going back to our early con- conversation of, of transfer learning, you start seeing that a true multimodal system, the other modalities benefit, uh, some, some different modalities. So, you get better at language because you, you now understand a little bit about video. So, um, I do think it, uh, it's harder to get going, but actually ultimately, um, we'll have a more general, more capable system like that.
- DPDwarkesh Patel
Uh, what- whatever happened to GATO? Like that was super fascinating that you could have, like played games and-
- DHDemis Hassabis
Yeah.
- DPDwarkesh Patel
... also do, like video and also do text.
- DHDemis Hassabis
Yeah, we're still, we're still working on those kinds of systems, but you can imagine we're just trying to, uh, those ideas we're trying to build into our, our, our future generations of Gemini-
- DPDwarkesh Patel
Oh, great.
- DHDemis Hassabis
... you know, to be able to do all of those things and, and, and robotics transformers and, you know, things like that are, are kind of, you can think of them as sort of, uh, follow-ups to that.
- DPDwarkesh Patel
Mm-hmm. Well, we see asymmetric progress towards the domains in which the self-play kinds of things you're talking about will be especially powerful, so math and code. You know, obviously recently you have these papers out about this, um, where, where yeah, you can, you can use these things to do, um, uh, really cool novel things. Uh, will, will they just be like superhuman coders, but like in other ways they might be still worse than humans? Or how do you think about that sort of-
- DHDemis Hassabis
Yeah, so look, I, I think that, that, that, um, you know, we're making great progress with math and, and, and, and things like theorem proving and, and coding. Um, but, uh, it's still interesting, you know, if one looks at, uh, I mean, creativity in general and scientific endeavor in general, I think we're getting to the stage where our systems could help the best human scientists make their breakthroughs quicker, like almost triage the search space in some ways, uh, or perhaps find a solution like AlphaFold does with a protein structure. Um, but it can't, it's, they're not at the s- at the level where they can create a hypothesis themselves or, or ask the right question. And any, as any top scientist will tell you, that, that's the hardest part of science is actually asking the right question. Uh, boiling down that space to like what's the critical question we should go after, the critical problem? And then formulating that problem in the right way to attack it. And that's not, um, something our systems or we have really have any idea how our systems could do. Um, but they can, uh, they are suitable for searching, uh, large combinatorial spaces if one can specify, uh, the problem in that way with a clear objective function. So that's very useful for already, uh, many of the problems we deal with today, but not the, the most high level creative problems.
- DPDwarkesh Patel
Mm-hmm. Um, w- when you, uh, so DeepMind obviously has, uh, published all kinds of interesting stuff in the, you know, speeding up science in different areas. Um, how do you think about that in the context of if you think AGI is gonna happen in the next 10, 20 years? Uh, why not just wait for the AGI to do it for you? Uh, why build these domain-specific solutions?
- DHDemis Hassabis
Yeah. Well, I think, um, we don't know how long AGI is going to be.
- DPDwarkesh Patel
Uh-huh.
- DHDemis Hassabis
And, and we always used to say, uh, you know, back even when we started DeepMind that, that, uh, uh, uh, we don't have to wait for AGI in order for, to bring incredible benefits to the world. Um, and, uh, es- especially, you know, my personal passion has been AI for science and, and, and health, and, and you can see that with things like AlphaFold and all of our various nature papers of different domains and material science work and so on. I think there's lots of exciting directions. Uh, and also impacting the world through products too I think is very exciting, uh, and a huge opportunity, unique opportunity we have as, as part of Google of, of, of, of the, you know, they, they, you know, they got dozens of, of, of billion user products, right-
- DPDwarkesh Patel
Yeah.
- DHDemis Hassabis
... that we can immediately ship our advances into and then, uh, billions of people, uh, can, can, you know, uh, improve their daily lives, right, and enriches their daily lives and enhances their daily lives. So, I think it's, it's a fantastic opportunity for impact on all those fronts. And I think the other reason from a point of view of, of AGI specifically is that it, it battle tests your ideas, right? So you don't want to be in a sort of, uh, research bunker where you just, you know, theoretically are pushing things, some things forward but then actually your internal metrics start deviating from, uh, uh, uh, real world things that would c- people would care about, right, or real world impact. Um, so you get a lot of feedback, uh, direct feedback from these real world applications that then tells you whether your systems really are scaling or, or actually is, you know, do we need to be more data efficient or sample efficient? Because most real world, uh, uh, challenges, uh, require that, right? And so it kind of keeps you honest and, um, pushes you, you know, keep sort of-... 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, society benefits from that, on the way, many, many, maybe many, many years before AGI arrives.
- 54:18 – 1:01:33
Inside Google DeepMind
- DHDemis Hassabis
- DPDwarkesh Patel
Yeah. Um, well, the, uh, the development of Gemini is super interesting because it comes right at the heels of merging these, uh, different organizations, uh, Brain and DeepMind. Um, well, yeah, I'm curious, what, what have been the challenges there? What have been the synergies? Uh, ha- ha- ha- and it's, it's been successful in the sense that you have the best model in the world now. What, what's that been like?
- DHDemis Hassabis
Yeah, it's been... Well, look, it's, it's, it's been fantastic actually over the last year. Of course, it's been challenging to do that, like any, any big integration coming together. Um, but you're talking about two, you know, world-class organizations, um, long storied histories of inventing many, many important things, um, you know, from deep reinforcement learning to transformers. And so it's very exciting actually, um, pooling all of that together and, and collaborating much more closely. We always used to be collaborating, but more on a, on a, on a, you know, sort of project by project basis versus a, a, a much deeper, broader collaboration like we have now. And Gemini is the first fruit of, of that, uh, uh, collaboration, uh, including the name Gemini, actually-
- DPDwarkesh Patel
(laughs)
- DHDemis Hassabis
... you know, implying twins. And, uh, and of course, a lot of other things are made more efficient, like pooling compute resources together and ideas and engineering, which, um, I think 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, um, I think it makes sense to, to coordinate that more closely.
- DPDwarkesh Patel
Yeah. So I mean, you, you and Shane started DeepMind, um, partly because you were concerned about safety. Um, you saw AGI coming as like a live possibility. Do you, uh, do you think the people who were formerly part of Brain, the half of Google DeepMind now, do they, do you think they approach it in the same way? Have there been cultural differences there in terms of that question?
- DHDemis Hassabis
Yeah, no, I think overall, and this is why, you know, I, I, I think one of the reasons we joined forces with Google back in 2014 was I think, um, the entirety of Google and Alphabet, not just Brain and DeepMind, take these questions very seriously of, of responsibility. And, um, you know, our kind of mantra is to try and be bold and responsible with these systems. So, you know, I would, I would, I would class it as I'm obviously a huge techno-optimist, but I, I want us to be cautious with that given the transformative power of what we're bringing, bringing into the world, you know, collectively. And, um, I think it's important. Uh, uh, you know, I think it's gonna be one of the most important technologies humanity will ever invent. So we, we've got to put, you know, all our efforts into getting this right and be thoughtful and sort of also humble about what we know and don't know about, uh, uh, uh, the systems that are coming and the uncertainties around that. And in my view, the only, the only sensible approach when you have huge uncertainty is to be sort of 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 of that before it happens. You know, you don't want to be live A/B testing out in the world-
- DPDwarkesh Patel
Sure.
- DHDemis Hassabis
... with these very consequential systems because unintended consequences may be, may be quite severe. So, um, you know, I, I want us to move away as a, a, a, as a field from a sort of move fast and break things attitude, which has, you know, maybe served the valley very well in the past and obviously created, uh, uh, important innovations. Um, but, but I think in this case, you know, we want to be, uh, uh, bold with the, with the positive things that it can do and make sure we realize things like medicine and science and advancing all of those things whilst being, um, you know, responsible and thoughtful with, with, uh, as far as possible with, with mitigating the risks.
- DPDwarkesh Patel
Yeah, yeah. And that, that's why it seems like the, the responsible scaling policies are something that, that is a very good empirical way to pre-commit to these kinds of things.
- DHDemis Hassabis
Yes, exactly.
- DPDwarkesh Patel
Um, yeah, and I'm curious if you have a sense of like, for example, when you're doing these evaluations, if it turns out your next model, um, could h- help a layperson build a pandemic-class bioweapon or something, uh, how you would think, first of all, of sec- making sure those rates are secure so that that doesn't get out, and second, uh, what would have to be true for you to be comfortable deploying that system? How comfortable, like how, how would you make sure that that, that latent capability isn't exposed?
- DHDemis Hassabis
Yeah. Well, first, I mean, you know, the, the, the secure model part, I think we've covered with the cybersecurity-
- DPDwarkesh Patel
Yeah, yeah.
- DHDemis Hassabis
... and, and make sure that's well classed and you're monitoring all those things. I think, um, if a capability was, was, was discovered like that through red teaming or, or external testing by, you know, uh, uh, uh, government institutes and, or academia or whatever-
- DPDwarkesh Patel
Mm-hmm.
- DHDemis Hassabis
... independent testers, um, then we would have to fix that, that loophole, well, depending what it was, right? Um, if that required more, um, uh, uh, a different kind of perhaps constitution or, or, or different guardrails or more RLHF to, to avoid that or removing some training data, um, there could... I mean, depending on what the problem is, I think there could be a number of, of, of mitigations. And, uh, so the first part is making sure you detect it ahead of time. So that's about the right evaluations, uh, and right benchmarking and right, and right testing. And then, um, the question is how one would fix that before, you know, you deployed it, so.
- DPDwarkesh Patel
Sure, sure.
- DHDemis Hassabis
But I think it would need to be fixed before it was deployed generally, for sure. If, if, if that was an ex- uh, exposure service.
- DPDwarkesh Patel
Right, right. Um, final question. Um, uh, you know, y- you've been thinking in terms of like the end goal of AGI at a time when other people thought it was, uh, ridiculous in 2010. Now that we're seeing, um, this like slow takeoff where we're actually seeing these like generalization and intelligence, um, h- h- what has like the psychologically seeing this, what has that been like? Has this just like sort of priced into your world model so you like, it's not new news for you? Or is it like actually just seeing it live, you're like, "Wow," like, uh, "This is something's like really changed," or how, what does it feel like?
- DHDemis Hassabis
Yeah, well, for me, n- um, yes, it's, it's already priced into my world model of how things were gonna go, at least from the technology side. But, um, obviously, I didn't, we didn't necessarily opi- anticipate, um, the general public would be that interested this early in the sequence, right, of, of things. Like maybe one could think of if we were to produce more, if, if say, like, uh, uh, ChatGPT and, and chatbots hadn't got the, kind of got the interest they'd ended up getting, which I think was quite surprising to everyone, that people were ready to use these things, uh, even though they, they were lacking in certain directions, right? Impressive though they are. Um, then we would have produced more specialized systems, I think, built off of the main track, like AlphaFolds and AlphaGo's and, uh, and so on in our scientific work. And then, um, I think the, the, the, the, the general public maybe, um, would have only paid attention later down the road, where in a few years time where we have more generally useful assistant-type systems. So that's been interesting. So that's created a different type of, uh, environment that we're now all operating in as a, as a, as a, as a field. So, um, and it's a little bit more chaotic because there's so many more things going on and there's so much VC money going into it and e- everyone's sort of almost losing their minds over it, I think.
- DPDwarkesh Patel
(laughs)
- DHDemis Hassabis
And I, and I, and I, and what I just, the thing, thing I worry about is I want to make sure that as a field, we act responsibly and thoughtfully and, and scientifically about this and use the scientific method to approach this in a, in a, as I said, an optimistic but careful way. And I think that's the... I've always believed that's the right approach for, for, for something like AI. And, um, I just hope that doesn't get lost in this huge rush.
- DPDwarkesh Patel
Sure, sure. Well, I think that's a great place to close. Demis, so much... Thanks dude, thank you so much for your time and for coming on the podcast.
- DHDemis Hassabis
Thanks. It's been a real pleasure.
- DPDwarkesh 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 in Twitter, your group chats, et cetera. It just blitz the world. Appreciate you listening. I'll see you next time. Cheers. (instrumental music)
Episode duration: 1:01:33
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Transcript of episode qTogNUV3CAI
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome