Inside The Infinity Machine

Sebastian, last time we talked, this idea came up that a lot of these figures in AI, let's say, have superpowers, right? Like Sam Altman seems to be really good at raising capital. Elon Musk seems to be quite good at that too, but. Demis, I think, has a more unique superpower. So I thought maybe to start off, you could talk a little bit about what sets Demis apart from a lot of the other folks in this space.

Yeah, so look, when I was thinking about writing a book about artificial intelligence, I had no doubt that Demis Asabis would be the person I want to put at the center of the story. And the reason is that he combines both the scientific, know, outstanding contribution, Nobel Prize scientist and so forth, and also the entrepreneurial leadership. You know, you've got Jeff Hinton. or Jan LeCun or, you know, Yoshua Benjo, lots of people who contributed to the science in an amazing way. And then you've got entrepreneurs like Sam Altman who've pushed the story forward, but Demis is both. And also he was super early on this, right? He started DeepMind in 2010 before AI could even recognize the picture of a cat. Nothing was working. So that combination of the early-ness, the entrepreneurial.

contribution and the scientific excellence is what really sets him apart.

Yeah, and in 2010, it was, mean, you know, in those days, it wasn't even kind of considered probabilistic. We just deterministic like everything's a rule, you know, how do we, how do we make a bunch of rules? How do we break down intelligence into a set of rules and hierarchy of rules? What is the high level hierarchy of rules? That would be general intelligence. Is there some algorithm? And I think that obviously eventually leads to an algorithm that is probabilistic, that surprises a lot of folks instead of this collection of deterministic rules that we perceive as intelligence. I wonder where he was in 2010. Was he thinking, you know, every sentence could be broken into some sort of pre-written rule? Or if he, he always understood this as, more of a, probabilistic situation.

Yeah, well, it's a great question because actually the answer goes right back to when he was an undergraduate at Cambridge. And, you know, he was at Cambridge 1994 to 1997 doing computer science. And most of the teaching was, as you were saying, know, deterministic, the search for algorithms, symbolic logic and all that stuff that is basically deductive thinking. And what he realized, even when he was sort of 19, is that

There's a lot of stuff we do which is inductive. We look at lot of examples and we induce a conclusion. We look at categories of objects and we create categories which are not logically describable, but we sort of know. And he once said to me, when I was at Cambridge, I realized that first order logic couldn't be the full answer to building intelligence because look, I speak on the grammatically all the time, he said, and yet you understand me. We, you know, just using your example of sentences, we often say stuff where we miss words out or we kind of change what we're trying to say halfway through a sentence. And yet we understand each other. So, yeah, he was super early in terms of giving up on the idea of first order logic and all that stuff. wanted to, he believed in much fuzzier thinking, much more inductive thinking.

It's actually why I call the book, Infinity Machine, because when you do induction, you have to have a lot of examples, right? Because otherwise you're gonna reach a wrong conclusion. If I look at 10 New Yorkers and I see what they do every morning, I'm gonna conclude that human beings always have coffee in the morning. But if I look at a million New Yorkers or even a million people, right? I would realize, no, there are exceptions. Not everybody has coffee in the morning. So you need a lot of examples to get it right. And in fact, you need almost

Right. Hmm. Right.

ideally, infinity examples. And so an infinity machine is a machine that can do induction and get away from logic.

Yeah. Yeah, I mean, especially if you're going for some sort of general intelligence concept, you need examples of every edge case and then every edge case of the edge cases and then every edge case of the edge cases. So it's interesting. think, you know, a good way to look at it that a lot of folks ascribe to is the idea that with this way of thinking,

Hmm. Right, right, right.

It's programming machines through examples versus through instructions. It's a new way of imagining how we program machines, but it does help think about, okay, how many examples now? Examples of what? How diverse are those examples? This all becomes the new way of programming.

Yeah, and the essence of the modern scaling laws is that the more data you put in the system, the more examples there are, the better the machine will be able to make judgments. One other amusing thing on the sort of earlyness of Demis's understanding of all this stuff is that he read the book Gödel Escherbach when he was still a teenager. This was a Pulitzer-pricing book. I think it was probably published in the early 70s.

No, sorry, I'm 80s maybe. Anyway, he read it when he was a teenager, something like, you know, 1991, 1992. And the point about the book in its invocation of Kurt Gödel in the title is that Gödel was famous for the incompleteness theorem, which showed that it was simply impossible for logicians to provide a complete description of how thinking works just through logic. No mathematical system. could encompass all of the possible truths in mathematics. It was a statement about the limits of logic. And so he was absorbing this stuff before he even went to college, which is just, this illustrates why Demis is so special and unusual in the field of lots of brilliant people. But I don't think we could say the same even about Elon Musk or someone like that.

No, well, it's interesting to how much of it comes through games too, right? Like very early on, he's programming his own games and then he ends up at, I think it was Bullfrog, which is a game manufacturer in the UK, correct? Yeah, and then, you know, they want to build this amusement park game that's going to be very complex and I love how you write about it. Like everyone there was sort of like, no thanks, that sounds like too much work and.

Hmm. Mm That's right.

Demis was like, bring it on. And then he sets about to work on this game, including like just an insane amount of variables, right? Like you can decide how much salt you want to be putting in the food at the concession stands. And like, if you make rides too fast, people are getting sick everywhere. And then they bring this game, right, to an AI conference where everything they're seeing is not very impressive to them. And they kind of start blowing minds with what they've been able to do through this game. And it's interesting how much of AI comes through gaming. You have what Demis is doing and then even the hardware, like GPUs, end up becoming the key ingredient to a lot of this.

Yeah, totally. mean, that's completely right. you know, reinforcement learning is the particular strand of AI that was advanced by gaming because you had these creatures in games that were being produced kind of in the later nineties, but still super early in terms of AI development, where if the player rewarded one of the avatars with a slap, I mean, punished it with a slap or rewarded it with a stroke or something, know, the character would learn from that and would do more of whatever you reward it for. And this is just an early implementation of the idea of reinforcement learning, which became central to breakthroughs like AlphaGo and even the reasoning systems, the reasoning large language models that we've seen more recently.

Yeah, I was thinking we've talked about like the next sort of big leap that we're going to see. And our belief is it's around simulation and testing of systems, essentially making them safe through, generating simulated situations, running them through those before we actually put them in the real world. that to me, you know, has led me to this, you know, this sort of thought. I was listening to how kids talk before they learn to read or write. And I was realizing that it's just one long word. You know, when they want something, and it just dawned on me, like, I don't know why it just dawned on me. that while I'm looking at them, like you don't know what letters and words are. All you're doing is mimicking sounds in one long stream. And that these sounds, they just trigger different actions and different sounds back, right? They have no concept, which then made me think like maybe the internet and all these words are just a simulation. of our intelligence, you know, that's all that sits out there. And then we've consumed this simulation of our intelligence. And because we have words, letters, and essentially tokens, we're now able to reproduce back, you know, what is these sounds, essentially that add up to certain actions. it just led me to think like he is so still

So well, even maybe better positioned at this point as we go into like this simulated component of these machines, his whole background has like perfectly led him to lead in this area.

Yeah, yeah, mean, gaming has played an important role in DeepMind's success under Demis' leadership, partly because, you the early demonstrations that AI could do stuff was in games. So you had chess before he started in 1997, but then you had Go in 2016, which was AlphaGo, the, you know, Demis' creation. And you also had before that Atari, the system which was built to play lots of different Atari games like Breakout and Pong and so forth. And so these were the proving grounds for the mid 2010s AI systems. And then you had StarCraft II was a big deal for DeepMind. They built a multiplayer system that could compete in a multiplayer real-time game. where if you think about those games, there's no like, it's my turn, it's your turn. There's just a constant flow of decisions and you have imperfect information because you can't see the whole map of the territory on which the game is taking place. So it's kind of like real life. You've got imperfect information, you have to these continuous decisions on different levels. Some is about the strategy.

of long-term plans to build up your civilization so you're stronger. And others is like immediate tactical responses to somebody attacking you. It's very kind of real world like. And that kind of then seemed to be the wrong path at one point because OpenAI did large language models and made this amazing progress and released chat GPT. But now it's coming back and I agree with you, know, that simulations, particularly if we want to build robots, it's too expensive to have actual physical robots, you know, try something, fall over, break their arms, have to fix them. No, you're going to do that in simulation. So if one of the next challenges for AI is to do really good robotics, it does play into something that DeepMind and DemisysAb is led on for much of their history.

Yeah, yeah. And speaking of GPUs, think that's Nvidia's play too, right? Is simulation and specifically for robotics because you don't want robots, I guess, breaking real dishes while they're learning to do the dishes. That's the thinking.

Yeah. the rules for them, know, physics rules somehow applying in the simulated world so that we can break robot arms in data.

Yeah, and here quietly games have just been that in the background and now they'll move to the foreground, right? Like when you look at a lot of like really even like Grand Theft Auto or something, right? It's just a huge simulation. you can just keep building more into that, go deeper on those simulations.

There was actually one really cool moment around 2017 when DeepMind had been doing these various simulations, which were kind of games. And then they realized that all of the games. kind of, know, everything is very regular. And even if you think about the urban environment, the built environment, you know, I've got a flat floor under my feet, I've got a flat ceiling above my head, there are right angles. If you look at a chair in my kitchen, some designer designed the chair and then, you know, the factory made, you know, 15,000 of them or whatever number. There's a lot of regularity and kind of easy classifiability.

in the built environment. But if you look at nature, you know, have these massive tall trees, you have these tiny blades of grass, each leaf or something is going to be a little bit different, bushes are completely irregular. And so DeepMind had this idea that human intelligence is born of interacting with the natural world or least earlier in evolution it was. And maybe artificial intelligence needs to interact with something like the natural world, way more irregular. And so they built a simulation of a natural environment. They called it Gaia. They had this vision when they were building their new offices in London that there would be an entire wall covered by a screen showing what was going on in Gaia. At all times you'd have these

reinforcement learning agents, you tell them to go and pick the apples or you tell them to go and, you know, find the path through the woods to something or other, right? And the problem was in 2017, the compute wasn't there yet. And so if you think about the way that reinforcement learning works, you're supposed to, and the agent does trial and error. And then if it gets a reward signal because it chose the right option, then it

that reinforces that choice and it will do more of that choice until finally it lands on the right set of behaviors. But if it's just too complicated, there's no, you you let it run for kind of 10 million attempts. And if it still hasn't found how to pick the apple, then it's intractable. There isn't enough reward signal for the reinforcement learning to work. And so what's exciting now, I think, is that they're going back to simulations.

and nine years on from that time when they had a setback, maybe it's going to work.

Yeah, that just the it was it was about getting the right amounts of each ingredient and There was this ingredient missing. This is core ingredient salt. Let's call it salt just because that seems appropriate for the book It was like this needs more salt and that's all we needed Yeah, I I'm fascinated by this area I I think this is the this is the part

Hmm. Ha

people miss today as we predict where AI might be going at least at a high level, the details, who knows, can anyone predict that? But at a high level understanding that if machines are smart enough to take over the world, then they're smart enough to predict that they will take over the world. And they're smart enough to predict who will do what when.

So this idea that it can simulate and predict and trying to factor that into the equation, very difficult for us to do. But it really seems like that's where our attention is going now is to say, yeah, like we need to understand what the effects on our environment will be by these machines before they before they actually affect the environment. And so they need to be able to do that. And And yeah, all of what he did is essentially that and language is, it's an interesting way to think about it, or it's either really obvious way to think about it, that if you're going to create a simulation and you're going to try to create a language around what the results of that simulation are, is it just language or is there a different language? for communicating a simulation in the future. As they were doing AlphaGo, there was some algorithmic state machine language that they kind of came back to. And it is interesting to think that that's what language is. One of its primary things is language is for communicating a future simulation. to someone, here's what's going to happen. And then we get, then machines get to re-consume those, or is there some more mathematical version of that simulation, like general intelligence as a general simulation algorithm or model?

Yeah, think listening, you one of the things to me which is fascinating about natural language is the way that it turns out to encompass so much of what one needs to know to be at least a simulation intelligence, right? So one of the reasons why Demis and Zabis did not go big on, you know, competing with the early versions of GPT, right? So in.

2018, the first open AI GPT is released and then 2019, the second one comes. And then this is like, you know, whatever. And the reason he thought that is because he didn't think that, you know, just by ingesting all of the language on the internet, the machine could really learn how to be intelligent because he thought, look, You got to be grounded in the real world. We were talking about robotics and simulations and world models and the physics of the world before. And he viewed all that stuff as super important. You couldn't just wish it away. just, if you're learning from a bunch of symbols, like words, and then you map those symbols onto images, like cats, that's a symbol onto a symbol. It's not really connected to anything real. So he was.

He was bothered by this as a sort of philosophical, philosophically, can that get you to AGI? And he thought not. And so he thought, well, that's kind of a dead end. And then he reversed his position in 2020 when GPT-3 came out, because suddenly, you couldn't say, well, this is just regurgitation. This was actually looking like proper intelligent responses to prompts. And so what he then.

thought to himself was, hey, this is like telling us that the range of human experience is smaller than I thought. It's actually not bigger than all of the tokens on the internet. So if you download the contents of the internet, you train on that. You're not only getting a bunch of symbols and language, you're actually getting a description of the physical world because those words were written by humans who are in the.

physical world. And so by studying everything on the internet, you kind of do know what gravity is. You do know what it feels like to fall over maybe. And then also, of course, there was reinforcement learning from human feedback on these systems. so grounded humans who were embodied in the real world were giving additional instruction to the machines. Anyway, so the point being that sometimes we think of

different representations of intelligence or knowledge, like some is like linguistic, some is images, some might be mathematical symbols and logic. And part of the art of building AI is to figure out which to bet on. is the most fruitful training data hidden?

Yeah, we've been talking a lot about the importance of knowledge management, more for like businesses that are trying to leverage AI, giving these powerful models a source of truth about, about an organization, what its policies are, how it wants to interact with customers. There's, there's so many different facets to that. And I think, you know, eventually that starts to become. simulation too, right? Like if you have all of your knowledge organized canonically within your organization and it's dynamic and people are always updating it and checking in on it, you know, then you can start to use that information to run simulations. And so then, how many decisions within an organization suddenly become simulated first, right? Like we're going to run a bunch of simulations. I think that we like to think that could lead to a state

we call it a OWAGI organizational AGI, but almost where like suddenly an organization now does know everything that it knows to a pretty high degree. And it has self-awareness at least about what it's supposed to be doing. And that might actually be sort of a precursor to a broader AGI.

Yeah, I mean. There's a similar debate about medicine, right? If you could build a simulation of a cell or something, you could run tests on, would this medicine work or not? Rather than trying it out on people, it'd be super fast in terms of we could iterate and get your way to the right compound much better. I think, yeah, trying to create, you know, virtual replicas of all kinds of things, whether it's your organization, whether it's your product that you're making. whether it's the human body, this is a pretty fruitful way of accelerating how fast we learn.

Yeah, I think it's easy just to get lost in interacting with these things in natural language to forget that their number one value is being prediction machines. We're often valued as humans for our prediction capabilities, especially in business. Did you do the right thing? Did you make the right product that people like? Did you make the right predictions? And if that's true, then simulation becomes like a massive advantage for machines to make better predictions. Now it just needs that data. But within organizations, they don't need to have AGI. They just need their organizational knowledge that's relevant. If you're a parking lot company, you don't need to understand all of the rules of the universe. to figure out that you have five parking spaces available right now. And one guy didn't pay.

Yeah, mean, you know, this reminds me of something that Demis was quite frustrated with when I was speaking to him for hours and hours and hours while I was researching my book is that, you know, prediction, as you say, is kind of the core to what makes humans good at lots of things, including business. But then people turn around and say, this large language model is just next word prediction, as if that's, you know, to diminish it. But of course, if the AI could predict the next word, it can predict the word after the next word, and pretty soon it can predict the whole book. And so there is enormous power. That shouldn't be regarded as a put down. That is a compliment if you say, it's predicting the next thing.

Yeah. Right. The put downs on ourselves that we can't think in compounding ways that we fall down when it compounds, but sorry, keep going.

Yeah, totally. totally. No, yeah. So I mean, I also talked from my book to Elia Satskeva, who was, of course, the chief scientist at OpenAI and a big Jeff Hinton star undergraduate student. At one point, he said to me, you he was very irritated with people who kept on saying it's just a stochastic parrot and all these ways of putting down the models. And he said, you know, it's not just statistics. And then he said,

Hahaha

He said, no, it is statistics, the statistics represent all these kinds of human thought, all the richness of human experience and how we express ourselves. And that's what's brilliant about it. The reason why computers can operate so effectively is that they translate all of these symbols that we use, language, images, into numbers, into matrices. And then you can.

create these multi-dimensional maps where everything is sort of situated, you know, with umpteen dimensions. In one sense, a computer might be right next to a mouse. In another sense, a computer is next to the word Apple. In another sense, you know, human beings have different contexts and

And the same way words have different contexts, but you can map all that if you turn it into matrices And so the fact is that it's just statistics. No, no, no statistics is what's good about it. That's why machines are so good

yeah, I go back to the McLuhanism where, you know, every piece of art is half done and then the observer completes the art themselves and it's always co-creation, right? And so if these things are outputting art, if words are art, I don't think we can, you know, that's not an argument at least that I'm aware of out there. I think we all agree that words become art and If the observer completes the art, then when someone says, that's not art, you're insulting their own work in their minds, right? They're like, wait a second, that is art. I made it art by looking at it and interpreting the way I interpret it. And so to each person, even if half of the art was done by a machine, they completed it. That's therefore human co-creation. I'm getting to this point here where it. A long time ago, I looked at words as sort of a packaging or almost a train that carries a payload and the payload or ideas and that next word prediction is just predicting the next word. But over time, it's also an idea predictor because within words, they're packages for ideas and ideas start emerging on top of the words. And yes, the machine is not packaging ideas. It's packaging words, but just like looking at the core elements of a plant and understanding that eventually it can grow an orange, right? Or a lemon or whatever. These ideas do sort of carry a level of intelligence that we ascribe to the system and say, well, you just happened to say something smart, right? Doesn't count because you weren't consciously planning to say something smart. But since these things are not planning out, you know, what they're going to do next, and they're not planning the thousandth word yet from the beginning, we decide it's less intelligent just because it sort of becomes emergent when potentially that's how we do it. Maybe we're not planning out our intelligence, it's emergent.

Yeah, know, Rob, I mean, listening to you, you it's a reminder of how this field of computer science, building intelligent machines becomes a sort of discourse on what is knowledge, epistemology, philosophy. And I think that's why, again, know, Demis Asabis is a fantastic character for a story about artificial intelligence because he has that range. where he did computer science as an undergraduate, but he also was a huge reader of philosophy, loves quoting Immanuel Kant and Spinoza and so forth. He also then did a PhD not in computer science or AI, he did it in neuroscience because he wanted to understand how human intelligence works.

and the different components of our intelligence and how those components interact and imagination and memory and the ability to plan and all these different things. so for him, think neuroscience was a way of just grappling with what the heck we mean when we say intelligence. And indeed his co-founder, Shane Legg had written a PhD all about defining what intelligence in a machine What is it? How should we think about it?

And then when they created DeepMind at the beginning, they hired people from neuroscience, they hired people who were like into probabilistic math. They were, you know, physicists, they were computer scientists, there was just a range of people, engineers. you know, AI hadn't been a field, I don't think, and it was kind of being created as a field by bringing together these different people from different backgrounds and putting them together and having them share ideas and see what they came up with. And then, you know, by 2015 when OpenAI got started and certainly by, you know, five years later or so when Anthropic started, these other labs sort of were born into a very different climate. You didn't have to invent the field anymore because that had been done. Now you sort of had to engineer it, implement it, produce a product. Very different challenge. so, but you're in, in the way you talk about it, you're taking me back to that early phase of, know, Demisysabas founding DeepMind and putting all these crazy ideas into a kind of blue sky exploration exercise. It's heady stuff.

Yeah, it is. And I guess when you're doing it, you can draw lines, you realize you're just solving a problem. You're obsessing about a problem of saying, like, how do we how do we make intelligence? And it just always leads you up the chain. I think that's where McLuhanism I think it's like a big missing piece is understanding how we like what is art. were so logical and we think we're so practical, but yet so much of how we interact with each other in technology is more artistic. And one of the things I really connect with McClune is the fact that 24 frames per second in film, if we actually saw each frame, it would stutter, it means we like, we add frames, right? It's like drawing in a book and flipping the book. Our minds fill in frames, which means like every time we watch a movie, half of it we invented in our heads and half of it was given to us, which means no art is complete and that art is always co-created. And so yeah, when someone's saying, the Mona Lisa is not just a picture, right?

They're saying that their completion of it, what they see, what she's feeling and thinking when they look at it is their own version of it. And so does it matter if the painter was a human or a machine if half of the painting is completed by you? And if it's abstract art, then more is completed by you, right? It's like two frames per second or one frame per second. And saying.

Hahaha

saying it's not art, saying it's not anything is to almost diminish the fact that the art's not complete until someone sees it.

Yeah, I don't think it's just art also. mean, so when Demis was doing his PhD in neuroscience, one of the things that really fascinated him was whether memory is something you sort of recall out of your brain like a film with a precise reconstruction of the event. Or is it sort of just a little bit of, you know, dry powder and you sort of add water to that and then it expands into a memory. But

but you just have a few kind of little dry husks of powder stored away in your brain and then you go to them when you need it. In other words, is memory reconstructive or is it sort of, yeah, yeah, yeah, yeah. And he concluded, he believed strongly that it was reconstructive, that only a fraction of the actual experience is stored in the mind. And then we reconstruct it because

Right, like two frames per second or one frame per minute.

Reimagining, yeah.

And that's why in all these experiments with police questions, know, the police can kind of suggest to the witness what they think they saw by sort of leading them a bit, because our memories are very susceptible to that because it's not actually a film or a photograph that you pull out of your head. It's just the sort of fragments. And then that led him to think, well, maybe memory is connected to imagination.

And indeed, through experimentation, he showed that the same part of the brain, the hippocampus, is crucial to both. And people who have damaged hippocampus suffer loss not only of memory, but also of imagination. And so these were the kinds of things that preoccupied Demis before he founded DeepMind. And I think that's why he comes at the construction of machine intelligence in this incredibly broad way, where he's thinking on multiple dimensions. He's not just trying to. know, scale a neural network. He's way more, he's way bigger than that, which again is why he's so fascinated.

Yeah. Yeah, yeah, and well positioned to hit this next chapter, right, I do think there's big missing pieces still left. I think this was a major, sort of missing piece, this, this idea that you mentioned, he thought, I'm not sure the fidelity is high enough.

in language and that we have enough data to complete a picture of intelligence with just language. And then he's like, maybe it is, but it's still incomplete. what completes it, right? Well, those memories and that ability to figure out how do we store memories in systems? How do we keep them learning? And then how do we reconstruct them on the way out? What fidelity do we save them at? And how much making up is okay? How much making up is not okay in between these ideas? so where is he today? Is he still as excited? Yeah.

About the whole field, yes, I think he is, absolutely. He's interesting because he combines this entrepreneurial side and he ran that game company called Elixir after he got out of college and before he did his neuroscience PhD. And so he's shipped product and he likes to win. He's the most competitive individual I've ever come across. I crazy, crazy competitive. And so when he got into the large language model race after ChatyBT came out,

They did this merger between Google Brain in Mountain View and DeepMind in London. And if you'd asked a business school professor, you're merging these two companies, it's in the middle of this crazy race against the rival company, OpenAI. And by the way, these two companies were competitors before, didn't really like each other. And what's more, they are eight time zones away from each other, right? California to London. What are the chances of this merger actually working and producing good products? I think the business school professor would have said, no, it's not gonna work. And they made it work. What's more, they made it work fast. So by two and a half years later in late 2025, Gemini 3 , the Google system was beating OpenAI's Chat GPT, whichever it was at the time, five or something, on the leaderboards. So. So they went from being behind to overtaking in two and a half years despite the merger. And I think that tells you that Demis has actually a zeal for leading a product organization because he has that competitive gene and because he has the experience of gaming. But then also he's excited as a scientist and that's a different part of his personality. The blue sky research, he often says,

know, whenever you kind of line up these AI leaders on the stage at Davos or something, you say, when is AGI coming? It's kind of a dumb question because nobody's agreed on the definition of AGI, but anyway, people ask the question. And Demis tends to be pushing out the expected time point more than some of the other people. And the reason is that he can't stand the idea that, you know, the excitement of some major scientific breakthrough.

isn't going to happen. He wants there to be more challenge on the, know, so yeah, know, the transformer paper is cool, but you know, we want to do more than that. We want three more transformer papers. Yeah, yeah, yeah. Because he just loves the scientific inquiry side as well as he loves the product competition side.

he needs another level.

Yes, I want to get to level three.

Yeah, I get that. It's motivating and you need that. Even if you invent the enemy.

You know, when I was talking to him, we would have these long conversations. would be, you know, there was a little pub in London near his house in North London. And he knew this secret staircase, which went up to the kind of room upstairs. Nobody else went there. And we would sit there for a couple of hours. So we had a lot of time to get into things. And he'd often kind of start saying, you know, so much noise at the moment in this AI race, I can't stand it. I want to, you know, retire to Princeton, to the Institute of Advanced Studies where Einstein.

you know, spent time and where Oppenheimer went after the Manhattan Project, you know, that's what I want to do. I love that place. And he actually went and gave a lecture there after he got the Nobel Prize. And it really that that place, you know, it's quiet and it's where you think. And that's what he loves that idea. But then you go like, Demis, are you really going to, you know, step away from this crazy capitalist race? It's like, no, no, no,

No. Yes, it's like playing tennis and not keeping score, know, like, no, it gets boring fast.

Yeah, there's a duality there. Yeah, I feel like it's strange because as a writer, find the experience of interacting with AI to be somewhat deflating because it's like, my God, it's getting faster and better at this every day. But then on the other side of it, I feel like it says something remarkable. Yeah, on one hand, am I that predictable? Yeah, I guess maybe I am. But on the other hand,

there's such richness in the way we communicate. There's so much still to be uncovered that even though these models have eaten all of the internet, they still need more, right? And they're still more locked away in language. And so it feels like it points to knowledge being or intelligence being a shared attribute, not something that's held inside your head, but also one that benefits from interaction. Like it continues to grow as long as people are continuing to communicate. And so, It seems like AI, and especially the way Demis looks at it, it's tapped into that. It's intertwined.

Yeah, it's so fascinating, Josh. mean, the range of reactions that I got from people I interviewed about the AlphaGo match when, you Man was beaten by Machine in the ancient Chinese game of Go. You know, there were some people who literally just gave up playing. And they, as you say, they were so deflated that if the Machine was going to win, they just gave up. And then there were others who sort of said, ooh.

There's a new beauty, a new depth to this that we didn't understand before and it's cool. Look at the additional beauty we've discovered thanks to the machine. of course, playing by humans has continued because human audiences want to watch humans. We identify with other humans, we want to watch humans. We don't really want to watch two machines play against each other. That would be boring. But the humans that we watch these days, and this is true of chess, hugely as well as Go.

they train with the machine. The machine teaches them to be better. And so it actually, in that sense, it's inflating human capacity, expanding it, right? And this is the paradox.

Yeah, and it's a treadmill for someone who's going to sprint in the Olympics, right? No one's like, you cheated. You used the treadmill to train. No, it just comes naturally to us. Of course we're going to use a treadmill. And I think we'll start to see it this way as like mental exercise. It's just a treadmill.

Yeah. So. there can be a world where it grows alongside us. It's not necessarily about replacement. It's more augmentation and parallel growth almost, Because AI needs us too, right? To give it ideas and push it.

Yeah, yeah, yeah, yeah. know, Demis has all these, obviously he understands all these dualities and he can feel both sides of this question. And he said, after the defeat of the Korean Go champion by his machine, he said, you know, I couldn't celebrate because he had been a competitive chess player as a kid, playing in tournaments all over the place. And he was the captain of the junior UK chess team and all that. And so he was, you know, He grew up on games competition and so he knew what it was to be defeated. And yet, of course, it was his machine, it was his scientific breakthrough. Of course, he was also delighted. And, you know, I puzzled about this, you know, is he really anti-machines and praver machines? How do we describe this? And then I realized what we describe this as is that he's a macrocosm, like an enlarged version of what all of us feel. We all feel. excited by technology and yet also frightened. And yet we go ahead with it, right? Because otherwise we would as human beings still be living in caves. We wouldn't have made any progress, right? So, so yeah, you know, you, so I think, I think, you know, Demis is sort of an enlarged version of, of me and most humans.

Yeah, yeah, it's the dark forest. Yeah, I wonder what he would say and maybe you can kind of take a guess for me, but whenever I'm trying to solve these puzzles, I always try to step outside, you know, and look at us through like an alien's eye, right? And I think of the like obscure movie reference. but it's this concept that aliens have consumed our media and then use that to kind of come back to earth and then communicate with us, right? Galaxy Quest, yeah. And in a way, like that's what an LLM is. tapped into our internet, right? And it consumed all of our language.

Galaxy Quest, yeah.

and all the words we said, and now it's coming back and it's echoing it to us and communicating after gathering all that language. And it just watches us like whales making these sounds at each other in different sequences, right? All these different frequencies as we make a bunch of sounds. It's almost like singing to each other. Like we're sitting there watching two birds in a tree. I wonder what they're talking about.

And it's kind of bizarre, right? It's sort of interesting how much we just go point our faces at different people's faces and make weird sounds and they make weird sounds back. if two people face each other and don't make any sounds, we call that awkward. You're not making any sound, you know? You're just staring at me, what's going on? Except when you play like a game like AlphaGo.

where they stare at each other and make no sound, but they're communicating in an entirely different way. And it's a conversation that's happening through movement, whether it's like dancing in a way, you know? And that it could go on for hours with no sound and that's okay, that's not awkward. And so there's like this other language that's completely nonverbal that is in these gaming situations is okay. And better than silence.

Well, I'll give you a thought that occurred to me and then I'll try and channel Demis if I can. But I think, you know, as you've just described, we have ways of communicating which might be linguistic or sometimes it could be moving pieces on a go board. The AI has now arrived and both understood our language and understood those movements on the go board very deeply. So your other example was what about Birdsong?

Surely we could now have an AI which, you know, watches the birds and listens to their song and really understands deeply what they're saying to each other if they're saying something. Like, what is the meaning being exchanged? And maybe human ornithologists have made decent progress with this, but I bet an AI could process more data and get deeper on the question. All kinds of animal communication would be fascinating to do that. scientific work.

if this is a different language that's not necessarily it's not reproduced in words. It's now reproduced in math So so go is a language of itself expressed in mathematics and and that's why it doesn't necessarily Transfer to to the language side right because because go is is another language And so we've got all these different dialects

Mm. Hmm. Yeah, I think for Demis and DeepMind, the difference between Go and language is not so much that one is mathematics and one isn't, because as we've said already, computers translate language into math, matrices, they translate Go moves into math. To the computer, it's all numbers. So all of these forms of communication are in some sense linked, and machines are in.

instantiation of that linkage because they just reduce everything to numbers. What's different though between Go and language though is that a large language model, least in the kind of early versions, goes to the internet, ingests the entirety of the internet and finds the patterns and can start to do next word prediction, which is different to Go in the sense that there isn't enough data on Go games between good players that have been written down with each move tabulated. There isn't enough data to train on for the system to become really good. And so therefore the way that Go systems are built is that they play against themselves, the computers. So they're simulating, you know, a human Go game. And by having tons of self play by the machine, you generate fresh data. And then eventually there's almost an infinity of data, which as I said, is what an infinity machine needs to really understand a data set deeply. So, and of course what the machine was doing is called reinforcement learning. It's playing and generating new data through experience as opposed to just sucking it up from the internet, pre-existing data. And one of the funny things about, you know, the sort of history of

modern AI is that this data on the internet is kind of a weird coincidence that it was there, ready for AI to use it, because the data wasn't put on the internet to train AI. The data was put on the internet because we were writing emails, we were doing e-commerce, we were writing blogs and, you know, publishing our scientific papers or whatever we were doing. We had a bunch of other reasons for populating the internet with all this language.

And it's a bit, mean, Demis is the one who pointed this out to me. And he said, it's kind of like, you know, the industrial revolution. And I go, what do mean? He says, well, you know, I mean, what would it have been like if things like electricity and steam were invented, but there was no coal in the ground. I mean, coal is a bunch of dead dinosaurs that happened to turn into coal very conveniently and then oil. And, you know, we could extract these hydrocarbons, but they weren't

Hmm. Right.

put there with a view to powering machines in the 20th or 19th century, right? It's like a divine coincidence that something, know, dinosaurs died, they got underground, they get, know, and in that sense, the internet is like dinosaurs underground. It's like a bunch of text that wasn't there for the purpose of making AI work, but it happens to make it work.

Right. Right. Yeah. Yeah, I like what you're saying. And this is kind of helps me more synthesize my my question, is really the goal of intelligence is the next level, right? Like level three that that we want to tackle. If we look at what happened with LLMs, it was it was people humans boiling down language to mathematics, right? If you look at AlphaGo, it was people boiling down the game to mathematics. Is our ability to invent an algorithm or machine that can boil down things in the world to mathematics, is that the intelligence that we're all talking about that's missing? once we have machines that can do that, that can look at the world and convert it to mathematics so that it can learn it, right? Is that the holy grail?

Yeah, I think that is the Holy Grail. And I would just add that you need a lot of it. You need a lot of mathematics. You need a lot of examples. Because as I said earlier, the challenge with learning from data, from induction, is that you need a lot of examples, otherwise you're going to come to that wrong conclusion that all human beings drink coffee every morning. And so the infinity machine that is the Holy Grail is a machine that gets an almost infinity of

data which is crystallized in mathematics as you're saying.

Yeah, I was watching. Yeah.

listening to birds is probably a good start then, right? Because if you're studying the bird song and you want to know what it's saying, ultimately it's going to start touching other things in the environment. Like, it's singing that, it's signaling that way because there's this seed over here that it likes. And then it all starts to thread.

Yeah. And by the way, that's a really good point because that points to what you were saying earlier, Josh, that it's sort of super exciting when you start thinking about, it's not just the bird song. relates to the seeds and how the pollen is going to be distributed by the bees. it's like, there is so much to understand. with AI, we have a shot at understanding it. I think this is why I grapple with this question.

Why do scientists create technologies that could be super dangerous and even challenge human existence? And all of the creators, including Demis, maybe especially Demis, were early in understanding that this is what they're doing. They're making something that is properly dangerous. So why do they do it? What is it? know, Jeff Hinton has this line which he borrowed from Robert Oppenheimer where he goes, you know, technology is sweet. as a scientist, can't resist inventing something if you have the opportunity to do that. And then you sort of figure out the consequences later. And that's, think, very powerful. I use that, the sweetness, I use that in my introduction. But I think that, you know, why are things sweet? And I think it's, I know that with demos, it's that flash of overwhelming excitement about how much there is to learn about the universe, about nature. And.

the desire to understand it. And, you know, I have a really good friend who's also a former Jeff Hinton PhD graduate student who went into a totally different field of finance. And he actually read my book, an early copy just recently, and kind of sent me this huge email saying how much he identified with Demis's quasi-spiritual urge to build AI because It is about understanding nature. It's like, when you see that, you just can't resist it. You've got to know.

I like that you chose him. I definitely feel like you picked the right person when history is being made it's hard to really feel like history is being made and and when you're documenting it as it's happening this is the fidelity that people will use to look back on and try to understand what was happening so it's so important that is being documented as it's happening or as close as it's happening as as can be because as you point out our memories are not very good. And the further we get away from the events, the worse that those interpretations will be. So I really love the book, love the person you chose to look at this through, the eyes and the perspective of how he sees this and how this is so key to why we're all here. Fantastic book.

Definitely. Yeah. And thanks. Thanks for taking the time with us in your afternoon, Sebastian. Yeah.

It's been a super fun conversation. Yeah. Thank you. Thank you, Rob. Thank you, Josh.