Unreasonably effective AI with Demis Hassabis

Introduction

Hannah Fry: [MUSIC PLAYING] HANNAH FRY: Welcome to “Google DeepMind, the Podcast” with me, your host, Professor Hannah Fry. Now, when we first started thinking about making this podcast way back in 2017, DeepMind was this relatively small, focused AI research lab. They’d just been bought by Google and given the freedom to do their own quirky research projects from the safe distance of London. How things have changed. Because since the last season, Google has reconfigured its entire structure, putting AI and the team at DeepMind at the core of its strategy. Google DeepMind has continued its quest to endow AI with human-level intelligence, known as artificial general intelligence, or AGI. It has introduced a family of powerful new AI models called Gemini, as well as an AI agent called Project Astra that can process audio, video, image, and code. The lab is also making huge leaps in applying AI to a host of scientific domains, including a brand new third version of AlphaFold, which can predict the structures of all of the molecules that you will find in the human body, not just proteins. And in 2021, they spun off a new company, Isomorphic Labs, to get down to the business of discovering new drugs to treat diseases. Google DeepMind is also working on powerful AI agents that can learn to perform tasks by themselves using reinforcement learning, and continuing that legacy of AlphaGo’s famous victory over a human in the game of Go. Now, of course, you’ll all have been following this podcast since the beginning. You’ll all be familiar with the stories behind all of those changes. But just in case you are coming to us fresh, welcome. You can find our first award-winning previous seasons on Google DeepMind’s YouTube channel, or wherever you get your podcasts. They also, those episodes go into detail about a lot of the themes that we’re going to hear come up over and over again from the people here, like reinforcement learning, deep learning, large language models and, so on. So have a listen. They are really good, even if we do say so ourselves. Now, all of the newfound attention on AI since the last series does mean that there are quite a few more podcasts out there for you to choose from. But on this podcast, in just the same way as we always have, we want to offer you something a little bit different. We want to take you right to the heart of where these ideas are coming from to introduce you to the people who are leading the design of our collective future— no hype, no spin, just compelling discussions and grand scientific ambition. So with all of that in mind, I am here with the DeepMind co-founder and now CEO of Google DeepMind, Demis Hassabis. So with all of that in mind, do I have to call you Sir Demis now?

Demis Hassabis: DEMIS HASSABIS: No, absolutely not. HANNAH FRY: OK.

Hannah Fry: Well, Demis, welcome to the podcast. DEMIS HASSABIS: Thank you. HANNAH FRY: Thank you very much for being here. OK, I want to know, is your job easier or harder now that there has been this explosion in public interest?

Demis Hassabis: DEMIS HASSABIS: I think it’s double edged. I think it’s harder because there’s just so much scrutiny, focus, and actually quite a lot of noise in the whole field. I actually preferred it when it was less people, and maybe a little bit more focused on the science. But it’s also good because it shows that the technology is ready to impact the real world in many different ways, and impact people’s everyday lives in positive ways. So I think it’s exciting, too.

Explosion of AI

Hannah Fry: HANNAH FRY: Have you been surprised by how quickly this has caught the public’s imagination? I mean, I guess you would have expected that eventually people would have got on board.

Demis Hassabis: DEMIS HASSABIS: Yes, exactly. So at some point, those of us who’ve been working on it like us for many years now, even decades, so I guess at some point the general public would wake up to that fact. And effectively, everyone’s starting to realize how important AI is going to be. But it’s been quite surreal still to see that actually come to fruition, and for that to happen. And I guess it is the advent of the chat bots and language models because everyone, of course, uses language. Everyone can understand language. So it’s an easy way for the general public to understand and maybe measure where AI has got to.

Hannah Fry: HANNAH FRY: I heard you describe these chat bots as though they were unreasonably effective, which I really like. And actually, later in the podcast we are going to be discussing transformers, which was the big breakthrough, I guess— the big advance that gave us those tools. But tell me first, what do you mean by unreasonably effective?

Unreasonably effective AI

Demis Hassabis: DEMIS HASSABIS: What I mean by it is I suppose if one were to wind back 5, 10 years ago, and you were to say the way we’re going to go about this is build these amazing architectures, and then scale from there, and not necessarily crack specific things like concepts or abstractions. These are a lot of debates we would have 5, 10 years ago is do you need a special way of doing abstractions? The brain certainly seems to do that. But yet somehow, the systems, if you give them enough data— i.e. The whole internet— then they do seem to learn this and generalize from those examples— not just rote memorize, but actually somewhat understand what they’re processing. And it’s a little bit unreasonably effective in the sense that I don’t think anyone would have thought that it would work as well as it has done, say, five years ago.

Hannah Fry: HANNAH FRY: Yeah. I suppose it is a surprise that things like conceptual understanding and abstraction have emerged rather than been—

Demis Hassabis: DEMIS HASSABIS: Yes, and we would have been— probably we discussed last time things like concepts and grounding— grounding language in real world experience, maybe in simulations or as robots embodied intelligence, would have been necessary to really understand the world around us. And of course, these systems are not there yet. They make lots of mistakes. They don’t really have a proper model of the world, but they’ve got a lot further than one might expect just by learning from language.

Hannah Fry: HANNAH FRY: I guess we probably should actually say what grounding is for those who haven’t listened to series 1 and series 2. Because this was a big thing. I mean, we were talking about this a lot. So do you want to just give us an overview

Demis Hassabis: of what grounding is? DEMIS HASSABIS: Grounding is when— one of the reasons the systems that were built in the ’80s and ’90s, the classical AI systems built at places like MIT, they were big logic systems. So you can imagine them as huge databases of words connected to other words. And the problem was you could say something, a dog has legs, and that would be in the database. But the problem was, as soon as you showed it a picture of a dog, it had no idea that collection of pixels was referring to that symbol. And that’s the grounding problem. So you have this symbolic representation, this abstract representation, but what does it really mean in the real world— in the messy real world? And then, of course, they tried to fix that, but you never get that quite right. And instead of that, of course, today’s systems, they’re directly learning from the data. So in a way, they’re forming that connection from the beginning. But the interesting thing was that if you learn just from language, in theory, there should be missing a lot of the grounding that you need. But it turns out that a lot of it is inferrable somehow. HANNAH FRY: Why, in theory? DEMIS HASSABIS: Well, because where is that grounding coming from? These systems, at least the first large language models— HANNAH FRY: Don’t exist in the real world. DEMIS HASSABIS: —don’t exist in the real world. They’re not connected to simulators. They’re not connected to robots. They don’t have any access to even— they weren’t multimodal to begin with, either. They don’t have access to the visuals or anything else. It’s just purely they live in language space. So they’re learning in an abstract domain, so it’s pretty surprising they can then infer some things about the real world from that.

Hannah Fry: HANNAH FRY: Which makes sense if the grounding gets in by people interacting with the system and saying that’s a rubbish answer, that’s a good answer.

Demis Hassabis: DEMIS HASSABIS: Yes. So for sure, part of that, if the question that they’re getting wrong, the early versions of this, was due to grounding missing— actually, the real world dogs bark in this way or whatever it is— and it’s answering it incorrectly, then that feedback will correct it. And part of that feedback is from our own grounded knowledge. So some grounding is seeping in like that for sure.

Hannah Fry: HANNAH FRY: I remember seeing a really nice example about crossing the English Channel versus walking

Demis Hassabis: across the English Channel. DEMIS HASSABIS: Exactly, those kinds of things. And if it answered wrong, you would tell it it’s wrong. And then it would have to slightly figure out that you can’t walk across the Channel.

Hannah Fry: HANNAH FRY: So some of these properties that have emerged that weren’t necessarily expected to be, I want to ask you a little bit about hype. Do you think that where we are right now, how things are at this moment, is overhyped or underhyped? Or is it just hyped, perhaps, in the wrong direction?

Overhyped or under hyped AI

Demis Hassabis: DEMIS HASSABIS: Yeah, I think it’s more the latter. So I would say that in the near term, it’s hyped too much. So I think people are claiming can do all sorts of things it can’t. There’s all sorts of startups and VC money chasing crazy ideas that are just not ready. On the other hand, I think it’s still underhyped. HANNAH FRY: Coming from you, Demis— DEMIS HASSABIS: Yes, I know, I know, I know. HANNAH FRY: AI in 2010. DEMIS HASSABIS: Exactly, exactly. But I think it’s still underhyped or perhaps underappreciated still even now what’s going to happen when we get to AGI and post-AGI. I still don’t feel like that’s people are quite understood how enormous that’s going to be, and therefore, the responsibility of that. So it’s both, really. I think it’s a little bit overhyped in the near term at the moment. We’re going through that cycle.

Hannah Fry: HANNAH FRY: I guess, though, so in terms of all of these potential startups, and VC funding, and so on, you who have lived and breathed this stuff for, as you say, decades, are very well placed to spot which ones are realistic goals and which ones aren’t. But for other people, how can they distinguish between what’s real and what isn’t?

Demis Hassabis: DEMIS HASSABIS: Yeah, well look, I think you need to look at— obviously you’ve got to do your technical due diligence, have some understanding of the technology, and the latest trends. I think also look at, perhaps, the background of the people saying it, how technical they are. Have they just arrived in AI last year from somewhere else? I don’t know. They were doing crypto last year. These might be some clues that perhaps they’re jumping on a bandwagon. And it doesn’t mean to say, of course, they could still have some good ideas, and many will do. But it’s a bit more lottery ticket like, shall we say. And I think that always happens when there’s a ton of attention suddenly on a place, and obviously, then the money follows that. And everyone feels like they’re missing out. And that creates a kind of opportunistic, shall we say, environment, which is a little bit opposite to those of us who’ve been in for decades in a deep technology, deep science way, which is ideally the way I think we need to carry on going as we get closer to AGI.

Hannah Fry: HANNAH FRY: Yeah. And I guess one of the big things that we’re going to talk about in this series is Gemini, which really comes from that very deep science approach, I guess. In what ways is Gemini different from the other large language models that are released by other labs?

Gemini and Project Astra

Demis Hassabis: DEMIS HASSABIS: So from the beginning with Gemini, we wanted it to be multi-modal from the start so it could process not just language, but also audio, video, image, code— any modality, really. And the reason we wanted to do that was firstly, we think that’s the way to get these systems to actually understand the world around them and build better world models. So actually still going back to our grounding question earlier, still building grounding in, but piggybacking on top of language this time. And so that’s important. And we also had this vision in the end of having a universal assistant, and we prototyped something called Astro, which I’m sure we’ll talk about, which understands not just what you’re typing, but actually the context you’re in. And if you think about something like a personal assistant or digital assistant, it will be much more useful the more context it understood about what you’re asking it for or the situation that you’re in. So we always thought that would be a much more useful type of system, and so we built multi-modality in from the start. So that was one thing, natively multi-modal. And then at the time, that was the only model doing that. So now the other models are trying to catch up. And then the other big innovations we had are on memory. So long context. So actually holding in mind 1 million— or 2 million now— tokens, you can think of them as more or less like words, in mind. So you can give it “War and Peace,” or even a whole— because it’s multi-modal— a whole video now, a whole film, or a lecture, and then get it to answer questions or find you things within that video stream.

Hannah Fry: HANNAH FRY: OK Project Astra, that’s the new universal AI agent, the one that can take in video and audio data. At Google. I/O, I think you used the example of how Astra could help you remember where you left your glasses, for instance. So I wonder, though, about the lineage of this stuff because is this just a fancy, advanced version of those old Google glasses?

Demis Hassabis: DEMIS HASSABIS: So, of course, Google have a long history of developing glass-type devices actually back to, I think, 2012 or something. So they were way ahead of the curve. But maybe it was just missing this kind of technology So you could actually understand— a smart agent, or a smart assistant that could actually understand what it’s seeing. And so we’re very excited about that digital assistant to go around with you and understand the world around you. So it seems a really— when you use it, it feels a really natural use case.

Hannah Fry: HANNAH FRY: OK. I want to rewind a tiny bit to the start of Gemini because it came from two separate parts of the organization.

Demis Hassabis: DEMIS HASSABIS: Yes. So we— actually, last year we combined our two research divisions at Alphabet. So obviously, the old DeepMind, and then Google Brain into one we call it super unit, bringing all the talent together— that amazing talent we have across the company, across the whole of Google, into one unified unit. And what it meant was that we combined all the best knowledge that we had from all the research we were doing, but especially on language models. So we had Chinchilla, and Gopher, and things like that, and they were building things like PaLM, and LaMDA, and early language models. And they had different strengths and weaknesses, and we pulled them all together into what became Gemini as the first Lighthouse project that the combined group would output. And then the other important thing, of course, was bringing together all the compute, as well, so that we could, do these really massive training runs and actually pull the compute resources together. So it’s been great.

Hannah Fry: HANNAH FRY: I guess, in a lot of ways, the focus of Google Brain and DeepMind was slightly different. Is that fair to say?

Demis Hassabis: DEMIS HASSABIS: Yeah. So I think it was. I mean, we were obviously focused, both of us, on the frontiers of AI, and there was a lot of collaborations already on a individual research level, but maybe not on a strategic level. Obviously, now the combined group, Google DeepMind, I describe it as we’re the engine room of Google now. But it’s worked really well. I think there were a lot more similarities, actually, in the way we were working than there were differences, and we’ve continued to keep and double down our strengths on things fundamental research. So where does the next transformer architecture come from? We want to invent that. Obviously, Google Brain invented the previous one. We combined it with deep reinforcement learning that we pioneered, and I still think more innovations are going to be needed. And I would back us to do that just as we’ve done in the past 10 years collectively, both Brain and DeepMind. So it’s been exciting.

Benchmarks

Hannah Fry: HANNAH FRY: I want to come back to that merge in a moment. But I think just sticking on Gemini for a second, how good is it? How does it compare to other models?

Demis Hassabis: DEMIS HASSABIS: Yeah, well, I think some of the benchmarks are not— the problem is that we need more— I think this is one thing the whole field needs is much better benchmarks. HANNAH FRY: Yeah. How do you decide? DEMIS HASSABIS: Well, there are some well-known benchmarks, academic ones. But they’re getting saturated now, and they don’t really differentiate between the nuances, between the different top models. I would say there’s three models that are at the top, the frontier. So it’s Gemini from us, OpenAI’s, GPT, of course, and then Anthropic with their Claude models. And then obviously, there’s a bunch of other good models, too, that people like Meta, and Mistral, and others built, and they’re differently good at different things. It depends what you want— coding, perhaps that’s Claude. And reasoning, maybe that’s GPT. And then memory stuff, long context, and multimodal understanding, that would be Gemini. Of course, we’re continuing to— all of us are improving our models all the time. So given where we started from, which Gemini as a project only existed for a year, obviously, based on some of our other projects, I think our trajectory is very good. So when we talk next time, we should hopefully be right at the forefront.

Hannah Fry: HANNAH FRY: Because there is still a way to go. I mean, there are still some things that these models aren’t very good at.

Demis Hassabis: DEMIS HASSABIS: Yes, for sure. And actually, that’s the big debate right now. So this last set of things emerged from the technologies that were invented five, six years ago. The question is, they’re still missing a ton of things— so their factuality, they hallucinate, as we know. They also not good at planning yet.

Hannah Fry: HANNAH FRY: Planning in what sense?

Demis Hassabis: DEMIS HASSABIS: Well, long term planning. So they can’t problem solve. Something long term, you give it an objective, they can’t really do actions in the world for you. So they’re very much like passive Q&A systems. You put the energy in by asking the question, and then they give you some kind of response. But they’re not able to solve a problem for you. You can’t say something like, if you wanted it as a digital assistant, you might want to say something like, book me that holiday in Italy, and all the restaurants, and the museums, and whatever, and it knows what you like, but then it goes out and books the flights and all of that for you. So it can’t do any of that. But I think that’s the next era— these more agent-based systems, we would call them, or agentic systems that have agent-like behavior. But of course, that’s what we’re expert in. That’s what we used to build with all our game agents— AlphaGo and all of the other things we’ve talked in about in the past. So a lot of what we’re doing is marrying that work that we’re, I guess, famous for with the new large multimodal models. And I think that’s going to be the next generation of systems. You can think of it as combining AlphaGo with Gemini.

Hannah Fry: HANNAH FRY: Yeah, because I guess AlphaGo was very, very good at planning.

Demis Hassabis: DEMIS HASSABIS: Yes, it was very good at planning. Of course, only in the domain, though, of games. And so we need to generalize that into the general domain of everyday workloads and language.

Hannah Fry: HANNAH FRY: You mentioned a minute ago how Google DeepMind is now the engine room of Google. I mean, that is quite a big shift since I was last here in the last couple of years ago. Is Google taking quite a big gamble on you?

Engine room of Google

Demis Hassabis: DEMIS HASSABIS: Well, I guess so. I mean, I think Google have always understood the importance of AI. Sundar, when he took over as CEO, said that Google was an AI-first company. And we discussed that very early on in his tenure, and he saw the potential in AI as the next big paradigm shift after mobile and internet, but bigger than those things. But then I think maybe in the last year or two, we’ve really started living what that means— not just from a research perspective, but also from products and other things. So it’s very exciting, but I think it’s the right bet for us to coordinate all of our talents together, and then push as hard as possible.

Hannah Fry: HANNAH FRY: And then how about the other way around? Because I guess from DeepMind, having that very strong research and science focus, does becoming the engine room for Google now mean that you have to care much more about commercial interests rather than the purer stuff that—

Demis Hassabis: DEMIS HASSABIS: Yeah, well, we do definitely have to worry more about, and it’s in our remit now, the commercial interests. But actually, there’s a couple things say about that. First of all, we’re continuing on with our science work in AlphaFolds, and you just saw AlphaFold 3 come out. And we’re doubling down on our investments there. That’s, I think, a unique thing that we do at Google DeepMind now. And even our competitors point at those things as universal goods, if you like, that come out of AI. And that’s going really well. And we spun out isomorphic to do drug discovery. So it’s very exciting, and that’s all going really well. And so we’re going to continue to do that. And then was all our work on climate and all of these things. But then, we’re quite a large team, so we can do more than one things at once. We’re also building our large models, Gemini and et cetera, and then we have a product team that we’re building out that is going to bring all this amazing technology to all of the surfaces that Google has. So it’s an incredible privilege, in a way, to have that there to plug in all of our stuff. And we invent something, it immediately can become useful to a billion people. And so that’s really motivating. And actually, the other thing is there’s a lot more convergence now between the technology we need to develop for a product to have AI in it and what you would do for pure AGI research purposes. So there’s not really— five years ago, you’d have had to build some special case AI for a product. Now, you can branch off your main research and, of course, you still need to do some things that are product specific, but maybe it’s only 10% of the work. So there’s actually not that tension anymore between what you would develop for an AI product and what you would develop for trying to build AGI. It’s 90%, I would say, the same research program. And then finally, of course, if you do products, and you get them out into the world, you learn a lot from that. And people using it, and you learn a lot about, oh, your internal metrics don’t quite match what people are saying, so then you can update that. And that’s really helpful for your research.

Hannah Fry: HANNAH FRY: Absolutely. Well, OK, we are going to talk a lot more in this podcast about those breakthroughs that have come from applying AI to science, but I want to ask you about that tension that there is between knowing when the right moment is to release something to the public. Because internally at DeepMind, those tools like large language models were being used for research rather than being seen as a potentially commercial thing.

Demis Hassabis: DEMIS HASSABIS: Yeah, that’s right. So as you know, we’ve always taken responsibility incredibly seriously here, and safety, right from the beginning, way back when we started in 2010 and before that. And Google then adopted some of our, basically, ethics charter effectively into their AI principles. So we’ve always been well aligned with the whole of Google and wanting to be responsible about deploying this as one of the leaders in this space. And so it’s been interesting now starting to ship real products with Gen AI in them. Actually there’s a lot of learning that is going on, and we’re learning fast, which is good because we’re at relatively low stakes here with the current technology. So it’s not that powerful yet. But as it gets more powerful, we have to be more careful. And that’s just learning about the product teams and other groups learning about how to test Gen AI technologies. It’s different from a normal piece of technology because it doesn’t always do the same thing. It’s almost like testing an open world game. It’s almost infinite what you can try and do with it, so it’s interesting to figure out how do you do the red teaming on it.

Hannah Fry: HANNAH FRY: So red teaming, in this case, being where you’re competing against yourselves?

Demis Hassabis: DEMIS HASSABIS: Yeah. So red teaming is when you set up a specific separate team from the team that’s developed the technology to stress test it and try and break it in any way possible. You actually need to use tools to automate that because nobody can red team— even if you had thousands of people doing it, that’s not enough compared to billions of users when you put it out there. They’re going to try all sorts of things. So it’s kind of interesting to take that learning, and then improve our processes so that our future launches will be as smooth as possible. And I think we got to do it in stages where there’s an experimental phase, then a closed beta, and then launch— a little bit, again, like we used to launch our games back in the day, and learn at each step of the way. And then the other thing we’ve got to do, and I think we need to do more on, is use AI itself to help us internally with red teaming and actually spotting some errors automatically or triaging that so that, then, our developers and human testers can actually focus on those hard cases.

Hannah Fry: HANNAH FRY: You said something really interesting there about how you’re just in a much more probabilistic space here. And then, if there’s even a very small chance of something happening, if you have enough tries, eventually, something will go wrong. And I guess there have been a couple of mistakes that— public mistakes.

Demis Hassabis: DEMIS HASSABIS: Yeah, so that’s why I think that, as I mentioned, that product teams are just getting used to the sorts of testing. They tested these things, but they have this stochastic nature, probabilistic nature. So in fact, a lot of cases where if it was a normal piece of software, you could say I’ve tested 99.999% of things, so then extrapolates. So then it’s enough because there’s no way of exposing the flaw that it has if it has one. But that’s not the case with these generative systems. They can do all sorts of things that are a little bit left field, or out of the box, out of distribution, in a way, from what you’ve seen before if someone clever or adversarial decides to— it’s almost like a hacker decides to test push it in some way. And it could even be— I mean, it’s so combinatorial, it could even be with all the things that you’ve happened to have said before to it. And then it’s in some kind of peculiar state which then— or it’s got its memories filled up with this particular thing, and then that’s why it outputs something. So there’s a lot of complexity there, but it’s not infinite. So there’s ways to deal with it. But it’s just a lot more nuanced than launching normal technology.

Shifting expectations of this technology

Hannah Fry: HANNAH FRY: I remember you saying, I think it was in the first time I interviewed you about how, actually, you have to think that this is a completely different way of computing. You have to move away from the things that we completely understand— the deterministic stuff— into this much more messy, probabilistic error-ridden place, as well as your testers. Do you think the public slightly has to shift its mindset on the type of computing that we’re doing?

Demis Hassabis: DEMIS HASSABIS: Yeah, I think so, and maybe that’s another thing, interestingly, that we’re thinking about is actually putting out a kind of principles document or something before you release something to show what is the expectation from this system. What’s it designed for? What’s it useful for? What can’t it do? And I think there is some sort of education there needed of, you’ll be able to find it useful if you do these things with it, but don’t try and use it for these other things because it won’t work. And I think that that’s something that we need to get better at clarifying as a field, and then probably users need to get more experienced on. And actually, this interesting. This is probably why chatbots themselves came a little bit out of the blue. Even obviously ChatGPT, but even to OpenAI, it surprised them. And we had our own chat bots, and Google had theirs. And one of the things was we were looking at them, and we were looking at all the flaws they still had, and they still do. And it’s like, well, it’s getting these things wrong, and it sometimes hallucinates, and blah, blah, blah. And there’s so many things. But then what we didn’t realize is, actually, there’s still a lot of very good use cases for that even now that people find very valuable— summarizing documents, and really long things, or writing— HANNAH FRY: Awkward emails? DEMIS HASSABIS: —awkward emails, or mundane forms to be filled in. And there’s all these use cases which, actually, people don’t mind if there’s some small errors. They can fix them easily, and saves a huge amount of time. And I guess that was the surprising thing. They discovered— people discovered when you put it in the hands of everyone, there were actually these valuable use cases, even though the systems were flawed in all of these ways we know.

Open source

Hannah Fry: HANNAH FRY: Well, OK, so I think that sort of takes me on to the next question I want to ask, which is about open source. Because when things are in the hands of people, as you mentioned, really extraordinary things can happen. And I know that DeepMind in the past has open sourced lots of its research projects, but it feels like that’s slightly changing now as we go forward. So just tell me what your stance is on open source.

Demis Hassabis: DEMIS HASSABIS: Yeah. Well, look, we’re huge supporters of open source and open science, as you know. I mean, we’ve given away and published almost everything we’ve done, collectively, including like, things like transformers, and AlphaGo. We published all these things in “Nature” and “Science.” AlphaFold was open source, as we covered last time. And these are all good choices, and you’re absolutely right. That’s the reason that all works is because that’s the way technology and science advances as quickly as possible, by sharing information. So almost always, that’s a universal good to do it like that, and that’s how science works. The only exception is when you— and AGI and powerful AI does fall into this— is when you have a dual purpose technology. And so then, the problem is that you want to enable all the good use cases and all the genuine scientists who are acting in good faith and so on, technologists, to build on the ideas, critique the ideas, and so on. That’s the way society advances the quickest. But the problem is how do you restrict access at the same time for bad actors who would take the same systems, repurpose them for bad ends, misuse them— weapon systems, who knows what? And those general purpose systems can be repurposed like that. And it’s OK today because I don’t think the systems are that powerful. But in two, three, four years time, especially when you start getting agent-like systems or agentic behaviors, then, I think, if something’s misused by someone, or perhaps even a rogue nation, state, there could be serious harm. So then, I don’t have a solution to that. But as a community, we need to think about what does that mean for open source? Perhaps the frontier models need to have more checks on them, and then only after they’ve been out for a year or two years, then they can get open sourced. That’s the model we’re following because we have our own open models of Gemini called Gemma because they’re smaller. So they’re not frontier models. So their capabilities are very useful still to the developer because they’re also easy to run on a laptop because they’re small numbers of parameters. But the capabilities they have are well understood at this point. Because they’re not frontier models. So it’s just not as powerful as the latest, say, Gemini 1.5 models. So I think that’s probably the approach that we’ll end up taking is we’ll have open source models, but they’ll be lagging maybe one year behind the most cutting edge models just so that we can really assess out in the open by users what those models can do— the frontier ones can do.

Hannah Fry: HANNAH FRY: And you can really, I guess, test those boundaries of the stochastic—

Demis Hassabis: DEMIS HASSABIS: Yeah, and we’ll see what those are. The problem with open source is if something goes wrong, you can’t recall it. With a proprietary model, if your bad actor starts using it in a bad way, you can just close the tap off. In the limit, you could switch it off. But once you open source something, there’s no pulling it back. So it’s a one way door, so you should be very, very sure when you do that.

Hannah Fry: HANNAH FRY: Is it definitely possible to contain an AGI, though, within the walls of an organization?

Demis Hassabis: DEMIS HASSABIS: Well, that’s a whole separate question. I don’t think we know how to do that right now. So when you start talking about AGI level powerful, like human level AI— HANNAH FRY: Well, what about intermediary? DEMIS HASSABIS: Well, intermediary, I think, we have good ideas of how to do that. So one would be things like secure sandboxing. So you test— that’s what I’d want to test the agent behaviors in is in a game environment, or a version of the internet that’s not quite fully connected. So there’s a lot of security work that’s done and known in this space, and in fintech, and other places. So we’d probably borrow those ideas, and then build those kinds of systems. And that’s how we would test the early prototype systems. But we also know that’s not going to be good enough to contain an AGI, something that’s potentially smarter than us. So I think we got to understand those systems better so that we can design the protocols for an AGI. When that time comes, we’ll have better ideas for how to contain that, potentially also using AI systems and tools to monitor the next versions of the AI system.

Hannah Fry: HANNAH FRY: So on the subject of safety, because I know that you are a very big part of the AI Safety Summit at Bletchley Park in 2023, which was, of course, hosted by the UK government. And from the outside, I think a lot of people just say the word regulation as though it’s just going to come in and fix everything. But what is your view on how regulation should be structured?

Regulation

Demis Hassabis: DEMIS HASSABIS: Well, I think it’s great that governments are getting up to speed on it and involved. I think that’s one of the good things about the recent explosion of interest is that, of course, governments are paying attention. And I think it’s been great. The UK government specifically, who I’ve talked to a lot, and US, as well, they’ve got very smart people in the civil service staff that understand the technology now to a good degree. And it’s been great to see the AI safety institutes being set up in the UK and US, and I think many other countries are going to follow. So I think these are all good precedents and protocols to settle into, again, before the stakes get really high. So this is a proving stage, again, as well. And I do think international cooperation is going to be needed, ideally around things like regulation, and guardrails, and deployment norms. So because AI is a digital technology, very much so, it’s hard to contain it within national boundaries. So if the UK or Europe does something, or even the US, but China doesn’t, does that really help the world? When we start getting closer to AGI, not really. So I think my view on it is you’ve got to be, because the technology is changing so fast, we’ve got to be very nimble and light-footed with regulation so that it’s easy to adapt it to where the latest technology is going. If you’d regulated AI five years ago, you’d have regulated something completely different to what we see today, which is Gen AI. But it might be different again in five years. It might be these agent-based systems that are the ones that carry the highest risk. So right now, I would recommend to beef up existing regulations in domains that already have them— health, transport, so on. I think you can update them for an AI world just like they were updated for mobile and internet. That’s probably the first thing I’d do, while doing a watching brief on making sure you understand and test the frontier systems. And then as things become clear and more clearly obvious, then start regulating around that. Maybe in a couple of years time would make sense. One of the things we’re missing is, again, the benchmarks— the right tests for capabilities that— what we’d all want to know, including the industry in the field, is at what point are capabilities posing some big risk? And there’s no answer to that at the moment beyond what I’ve just said, which is agent-based capabilities is probably our next threshold. But there’s no agreed-upon test for that. One thing you might imagine is testing for deception, for example, as a capability. You really don’t want that in the system because then you can’t rely on anything else that it’s reporting. So that would be my number one emerging capability that I think would be good to test for. But there’s many— ability to achieve certain goals, the ability to replicate. And there’s quite a lot of work going on on this now. And I think the safety institutes, which are basically government agencies, I think it would be great for them to push on that, as well. As well as the labs, of course, contributing what we know.

Hannah Fry: HANNAH FRY: I wonder, in this picture of the world that you’re describing, what’s the place for institutions in this? I mean, if we get to the stage where we have AGI that’s supporting all scientific research, is there still a place for great institutions?

Demis Hassabis: DEMIS HASSABIS: Yeah, I think so. There’s the stage up to AGI, and I think that’s got to be a cooperation between civil society, academia, government, and the industrial labs. So I really believe that’s the only way we’re going to get to the final stages of this. Now, if you’re asking after AGI happens, maybe that is what you’re asking, then AGI, of course, one of the reasons I’ve always wanted to build it is then we can use it to start answering some of the biggest, most fundamental questions about the nature of reality, and physics, and all of these things, and consciousness, and so on. It depends what form that takes, whether that will be a human-expert combination with AI. I think that will be the case for a while in terms of discovering the next frontier. So like right now, these systems can’t come up with their own conjectures or hypotheses. They can help you prove something, and I think we’ll be able to prove get gold medals on international maths Olympiad, things like that. But maybe even solve a famous conjecture. I think that’s within reach now. But they don’t have the ability to come up with Riemann hypothesis in the first place, or general relativity. So that, really, was always my test for maybe a true artificial general intelligence is it will be able to do that, or invent Go. And so we don’t have any systems. We don’t really know how we would design, in theory, even, a system that could do that.

Hannah Fry: HANNAH FRY: You know the computer scientist, Stuart Russell? So he told me that he was a bit worried that once we get to AGI, it might be that we all become like the royal princes of the past— the ones who never had to ascend the throne or do any work, but just got to live this life of unbridled luxury and have no purpose.

The royal princes of the past

Demis Hassabis: DEMIS HASSABIS: Yeah, so that is the interesting question, is it? Maybe it’s beyond AGI. It’s more like artificial superintelligence or something— sometimes people call it ASI. But then we should have radical abundance. And assuming we make sure we distribute that fairly and equitably, then we will be in this position where we’ll have more freedom to choose what to do. And then meaning will be a big philosophical question. And I think we’ll need philosophers, perhaps theologians even, to start thinking as social scientists. They should be thinking about that now. What brings meaning? I mean, I still think there’s, of course, self-actualization, and I don’t think we’ll all just be sitting there meditating. But maybe we’ll be playing computer games. I don’t know. But is that a bad thing even, or not? Who knows?

Hannah Fry: HANNAH FRY: I don’t think the princes of the past came off particularly well.

Demis Hassabis: DEMIS HASSABIS: No. Traveling the stars. But then there’s also extreme sports people do. Why do they do them? I mean, climb Everest, all these. I mean, there’ll be— but I think it’s going be very interesting. And that I don’t know, but that’s what I was saying earlier about it’s underappreciated what’s going to happen going back to the hype near-term versus far-term. So if you want to call that hype, even, it’s definitely under hyped, I think, the amount of transformation that will happen. I think it will be very good in the limit. We’ll cure lots of diseases, and/or all diseases, solve our energy problems, climate problems. But then the next question comes is, is there meaning?

Hannah Fry: HANNAH FRY: So bring us back slightly closer to AGI rather than superintelligence. I know that your big mission is to build artificial intelligence to benefit everybody, but how do you make sure that it does benefit everybody? How do you include all people’s preferences rather than just the designers?

Demis Hassabis: DEMIS HASSABIS: Yeah, I think what’s going to have to happen is— I mean, it’s impossible to include all preferences in one system. Because by definition, people don’t agree. We can see that in, unfortunately, in the current state of the world. Countries don’t agree. Governments don’t agree. We can’t even get agreement on obvious things like dealing with the climate situation. So I think that’s very hard. What I imagine will happen is that we’ll have a set of safe architectures, hopefully, that personalized AIs can be built on top of. And then everyone will have, or different countries will have their own preferences about what they use it for, what they deploy it for, what can and can’t be done with them. But overall— and that’s fine. That’s for everyone to individually decide or countries to decide themselves, just like they do today. But as a society, we know that there’s some provably safe things about those architectures. And then you can let them proliferate, and so on. So I think that we’ve got to get through the eye of a needle in a way where as we get closer to AGI, we’ve probably got to cooperate more, ideally internationally, and then make sure we build AGIs in a safe architecture way. Because I’m sure there are unsafe ways, and I’m sure there are safe ways of building AGI. And then once we get through that, then we can open the funnel again, and everyone can have their own personalized pocket AGIs, if they want.

Hannah Fry: HANNAH FRY: What a version of the future. But then, in terms of the safe way to build it, I mean, are we talking about undesirable behaviors here that might emerge?

Demis Hassabis: DEMIS HASSABIS: Yes, undesirable emergent behaviors, capabilities that— HANNAH FRY: Deception. DEMIS HASSABIS: Deception is one example that you don’t want. Value systems. We got to understand all of these things better— what kind of guardrails work, not circumventable. And there’s two cases to worry about. There’s bad uses by bad individuals or nations, so human misuse, and then there’s the AI itself as it gets closer to AGI going off the rails. And I think you need different solutions for those two problems. And so, yeah, that’s what we’re going to have to contend with as we get closer to building these technologies. And also, just going back to your benefiting everyone point, of course, we’re showing the way with things like AlphaFold and isomorphic. I think we could cure most diseases within the next decade or two if AI drug design works. And then they could be personalized medicines where it minimizes the side effects on the individual because it’s mapped to the person’s individual illness, and their individual metabolism, and so on. So these are amazing things— clean energy, renewable energy sources, fusion, or better solar power, all of these types of things. I think they’re all within reach. And then that would sort out water access because you could do desalination everywhere. So I just feel like an enormous good is going to come from these technologies, but we have to mitigate the risks, too.

Hannah Fry: HANNAH FRY: And one way that you said that you would want to mitigate the risks was that there would be a moment where you would basically do the scientific version of Avengers assemble. DEMIS HASSABIS: Yes, sure. HANNAH FRY: Terence Tao, get him on the phone.

Avengers assemble

Demis Hassabis: DEMIS HASSABIS: Exactly. HANNAH FRY: Bring him on down.

Hannah Fry: DEMIS HASSABIS: Yeah, exactly.

Demis Hassabis: HANNAH FRY: Is that still your plan? DEMIS HASSABIS: Yeah, well, I think so. I think if we can get the international cooperation, I’d love there to be a international CERN, basically, for AI. Where you get the top researchers in the world, and you go, look, let’s focus on the final few years of this AGI project, and get it really right, and do it scientifically, and carefully, and thoughtfully at every step— the final steps. I still think that would be the best way.

Hannah Fry: HANNAH FRY: How do you know when is the time to press the button?

Demis Hassabis: DEMIS HASSABIS: Well, that’s the big question. Because you can’t do it too early because you would never be able to get the buy-in to do that. A lot of people would disagree. I mean, today people disagree with the risks. You see very famous people saying there’s no risks, and then you have people like Jeff Hinton saying there’s lots of risks. And I’m in the middle of that.

Hannah Fry: HANNAH FRY: I want to talk to you a bit more about neuroscience. How much does it still inspire what you’re doing? Because I noticed the other day that DeepMind had unveiled this computerized rat with an artificial brain that helps to change our understanding of how the brain controls movement. But in the first season of the podcast, I remember we talked a lot about how DeepMind takes direct inspiration from biological systems. Is that still the core of your approach?

Demis Hassabis: DEMIS HASSABIS: No, it’s evolved now because I think we’ve got to a stage now. In the last, I would say, two to three years, we’ve gone more into an engineering phase— large scale systems, massive training architectures. So I would say that the influence of neuroscience on that is a little bit less. It may come back in. So any time where you need more invention, then you want to get as many sources as possible. And neuroscience would be one of those sources of ideas. But when it’s more engineering heavy, then I think that takes a little bit more of a backseat. So it may be more applying AI to neuroscience now like you saw with the virtual rat brain. And I think we’ll see that as we get closer to AGI using that to understand the brain. I think it would be one of the coolest use cases for AGI and science.

Hannah Fry: HANNAH FRY: I guess this stuff goes through phases of the engineering challenge,

Demis Hassabis: the intervention challenge. DEMIS HASSABIS: So it’s done its part for now, and it’s been great. And we still obviously keep a close track of it and take any other ideas, too.

Peer into the mysteries of the universe

Hannah Fry: HANNAH FRY: OK. All of the pictures of the future that you’ve painted are still anchored quite in reality. But I know that you’ve said that you really want AGI to be able to peer into the mysteries of the universe down at the Planck scale. DEMIS HASSABIS: Yes! HANNAH FRY: Like subatomic, quantum worlds. Do you think that there are things that we have not even yet conceived of that might end up being possible?

Demis Hassabis: I’m talking wormholes here. DEMIS HASSABIS: Completely, yes. I’d love wormholes to be possible. I think there is a lot of probably misunderstanding, I would say, still things we don’t understand about physics and the nature of reality. And obviously, the quantum mechanics, and unifying that with gravity, and all of these things, and there’s all these problems with the standard model. So I think there’s— and string theory, I mean, I just think—

Hannah Fry: HANNAH FRY: There’s giant gaping holes in physics.

Demis Hassabis: DEMIS HASSABIS: Yes, in physics, all over the place. And I talk to my physics friends about this, and there’s a lot of things that don’t fit together. I don’t really like the multiverse explanation. So I think that it would be great to come up with new theories and then test those on massive apparatus, perhaps out in space, at these tiny— the reason I’m obsessed with Planck scale things— Planck time, Planck space— is because that seems to be the resolution of reality in a way. That’s the smallest quanta you can break anything into. So that feels like the level you want to experiment on if you had powerful apparatus perhaps designed or enabled by having AGI and radical abundance. You would need both to be able to afford to build those types of experiments.

Hannah Fry: HANNAH FRY: The resolution of reality. What a phrase. What, so as in the resolution that we’re at the moment? Human level is just an approximation of reality.

Demis Hassabis: DEMIS HASSABIS: Yes, that’s right. And then we know there’s the atomic level, and below that is the Planck level, which as far as we know is the smallest resolution one can even talk about things. And so that, to me, would be the resolution one wants to experiment on to really understand what’s going on here.

Hannah Fry: HANNAH FRY: I wonder whether you’re also envisaging that there will be things that are beyond the limits of human understanding AGI will help us to uncover— that actually, we’re just not really capable of understanding. And then I wonder if things are unexplainable or ununderstandable, are they still falsifiable?

Demis Hassabis: DEMIS HASSABIS: Yeah well, look, I mean, these are great questions. I think there will be a potential for an AGI system to understand higher level abstractions than we can. So again, going back to neuroscience, we know that it’s your prefrontal cortex that does that. And there’s up to about six or seven layers of indirection one could take. This person’s thinking this, and I’m thinking this about that person thinking this, and so on. And then we lose track. But I think an AI system could have an arbitrarily large prefrontal cortex effectively. So you could imagine higher levels of abstraction and patterns that it will be able to see about the universe that we can’t really comprehend or hold in mind at once. And then I think in terms of explainability point of view, the way I think that is a little bit different to other philosophers who’ve thought about this, which is we’ll be closer to an ant and then the AGI, in terms of IQ. But I don’t think that’s the way to think of it. I think it’s we are Turing complete, so we’re a full general intelligence as ourselves, albeit a bit slow because we run on slow machinery. And we can’t infinitely expand our own brains. But we can, in theory, given enough time and memory, understand anything that’s computable. And so I think it will be more like Garry Kasparov or Magnus Carlsen playing an amazing chess move. I couldn’t have come up with it, but they can explain it to me why it’s a good move. So I think that’s what an AGI system will be able to do.

Hannah Fry: HANNAH FRY: You said that DeepMind was a 20-year project. How far through are we? Are you on track?

Demis Hassabis: DEMIS HASSABIS: I think we’re on track, yeah, crazily. Because usually 20-year projects stay 20 years away. But yeah, we’re a good way in now, and I think we’re—

Hannah Fry: HANNAH FRY: 20 years is 2030 for AGI.

Demis Hassabis: DEMIS HASSABIS: 2030, yeah. So I think the way I say is I wouldn’t be surprised if it comes in the next decade. So I think we’re on track.

Hannah Fry: HANNAH FRY: That matches what you said last time. You haven’t updated your priors. [LAUGHTER] DEMIS HASSABIS: Exactly HANNAH FRY: Amazing. Demis, thank you so much. Absolute delight.

Demis Hassabis: Absolute delight, as always. DEMIS HASSABIS: Very fun to talk, as always, as well. Thank you.

Conclusion

Hannah Fry: HANNAH FRY: OK, I think there are a few really important things that came out of that conversation, especially when you compare it to what Demis was saying last time we spoke to him in 2022. Because there have definitely been a few surprises in the last couple of years. The way that these models have demonstrated a genuine conceptual understanding is one— this real world grounding that came in from language and human feedback alone. We did not think that would be enough. And then how interesting and useful, imperfect AI has been to the everyday person. Demis himself there admitted that he had not seen that one coming. And that makes me wonder about the other challenges that we don’t yet know how to solve like long-term planning, and agency, and robust, unbreakable safeguards. How many of those— which we’re going to cover in detail in this podcast, by the way— are we going to come back to in a couple of years and realize that they were easier than we thought? And how many of them are going to be harder? And then as for the big predictions that Demis made, like cures for most diseases in 10 or 20 years, or AGI by the end of the decade, or how we’re about to enter into an era of abundance, I mean, they all sound like Demis is being a bit overly optimistic, doesn’t it? But then again, he hasn’t exactly been wrong so far. You’ve been listening to “Google DeepMind, the Podcast” with me, Professor Hannah Fry. If you have enjoyed this episode, hey, why not subscribe? We’ve got plenty more fascinating conversations with the people at the cutting edge of AI coming up on topics ranging from how AI is accelerating the pace of scientific discoveries to addressing some of the biggest risks of this technology. If you have any feedback, or you want to suggest a future guest, then do leave us a comment on YouTube. Until next time. [MUSIC PLAYING]