Actually, the story is pretty complicated in the sense that the leading labs at the time — so us included — and actually both DeepMind and the Google lab, Google Brain, had large language models. In fact, the Google Brain side of it had a product, a very early one, when it was called LaMDA in those days, and there had been a conscious decision not to release that. And it was obviously, in retrospect, probably the wrong decision, but you can sort of see — and I wasn't involved in that because we were at the time just supposed to be working on research, right? Because DeepMind, prior to that, prior to Google DeepMind being formed, we were just sort of running independently, mostly as a research organization. So, but I think that's an interesting thing of where you can be sometimes too close to your own technology. So at building that technology, those early language models, they're very interesting, but they were very flawed. Even today's ones, very impressive as they are, have things that we all know when we use the chatbots, they can't do very well. They make mistakes, they hallucinate, and those early ones were doing that a lot. So you can imagine — and I can understand that I wasn't involved in the decision-making — but you can sort of think about, you know, Google's known for accurate information. That's what Google is for. Google Search is for organizing the world's information. That's its bread and butter. That's what all of us come to Google Search for. So I think it's sort of doubly tricky to think about putting out a bot that, you know, 20% of the time gives you the wrong information. And so that's — there's, I think, there's a clear issue there for a company like Google, whereas obviously a startup, especially someone with someone like Sam leading it, you know, it's like, go for it, you know, you only live once, YOLO it, and it's fine, right? And then — and the thing is, to be fair to us as well, from a research perspective, no one knew how big that was going to get, ChatGPT, including OpenAI, because — and it's obvious, I mean, Sam's admitted that, but it's also obvious from the name. ChatGPT is a programmer's name. You can — you just like, you know, they would have named it something catchier if they knew how big it was going to get, right? It's not a catchy name. Although, you know, it's obviously synonymous with AI now. So you can see, and I think it's — they've talked about they just thought it was going to be a research experiment, and obviously they could take more risks with, given that they were a startup with reputation and other things. If it makes mistakes, so what? They don't have any existing products that rely on reliability. So the thing that caught everyone by surprise, I would say, including them, was actually people finding use for even these early language models, that even though they had all of these flaws, you know, people still found — obviously very creatively, once millions of people use it, oh, you can summarize things, can help with homework. So there were these sort of low-hanging fruit that it didn't matter — do some brainstorming. It didn't matter if it made some errors, right, that weren't to do with information seeking, and that's what — these were new use cases that hadn't existed before. And obviously there was this — turned out there was this huge demand for even those kind of early use cases, and that's what caught everyone by surprise in terms of the technology being mature enough to come out of the lab and be put into the general public's hands. And one of the upshots was that for Google, Google Brain and DeepMind merged. You now run essentially all AI at Google, and obviously you've not just closed the gap, but on some metrics are leading in terms of some of these models that you are putting out with Gemini. That in itself — I mean, we probably have time to go into the intricacies of that merger, but that itself, Sebastian, you were saying that merger of Google Brain and DeepMind stands out to you as, in terms of just in terms of a corporate merger, as something of a standout moment.
Yeah. I mean, the conventional wisdom if you ask a business school professor about mergers is that, you know, they're very difficult because there's always cultural differences. It's painful to smush two different organizations together and so forth. And in the case of, you know, Google Brain, which was doing AI research in Mountain View, and then DeepMind, which was doing AI research in London, you're putting together two teams which had eight hours of time difference between them, which doesn't make it easier. And they had a little bit of a record, which, you know, I hope you don't mind me saying this, there was a bit of a record of competition between the two, because to some extent, if DeepMind got compute resources to, you know, run some tests and build a model, the people in Mountain View might see that as at the expense of their experiment they were trying to do. So there was, at least on the level of kind of engineering gossip, you know, quite a lot of back and forth about, you know, this bad blood. So here you are, you've got to do a merger, you've got eight time zones of difference, there's a bit of a record of competition, and you're in the middle of this knockdown, drag-out capitalist fight, right, with OpenAI racing ahead and Google's got to catch up. What are the odds of this succeeding? I think most business school professors would have said, 'Nuh-uh, not going to work.' It worked. Like two and a half years later, at the end of 2025, DeepMind has an AI model, Gemini 3.0, which beats the ChatGPT rival on the leaderboards on these independent metrics of power and accuracy and so forth. So, in two and a half years, it was a total turnaround. It's like, I think this will go down not merely as a story about a frontier technology, but as a sort of case study in business mergers.
And this is helping us, this merger and being under Demis and the ambitions around AGI, is helping you to get to that long-term ambition that's driven you since you were a student, which is AGI and superintelligent machines. So I want to get to the title of the book belatedly, maybe the infinity — what is it, Demis? What is the infinity machine as you understand it?
Well, I mean, I think it's a great title for the book, and I think actually Sebastian mentioned it in the opening. The way we look at problems, and it started really with AlphaGo and then onto AlphaFold, is I think we've found — and this is what AGI should be — a kind of general solution that can be applied to many, many problems. And the way I look at both Go and protein folding, which ostensibly look like really different types of problems, is you're effectively navigating — there's a needle-in-a-haystack solution, but the haystack is infathomably large, right? So, for example, in Go, there are more possible board positions than there are atoms in the universe. So, you know, that means you can't brute force and enumerate every option. There's just not enough time in the universe. And the same with proteins is even more complicated. An average-sized protein, you know, has 10 to the 300 possible conformations, possible shapes. So you can't find the right shape by looking at every single shape. You're going to have to do something much smarter. And so what we generically have built with our AI systems, at least applied to science, is using deep learning, build neural networks, give them the data that you have on that particular domain — so Go games or, you know, protein structures that are known — and it's got to create a model of how that domain works. And then once you have that model, you can apply a search process, but in a smart way, instead of having to search everything. So in a Go game, instead of, you know, you have a hundred possible moves in a position, you don't search all of them, you just search — the model tells you which ones are the most likely, the most fruitful ones to look at, so you can spend your compute budget on very useful alternatives and only look at the ones that are likely to be fruitful. And so that's the way you can actually cut down these enormous search spaces into something that's tractable in potentially a very fast amount of time. And so this is a very — I sort of realize this is a very generic solution to many problems if you can couch it in this way: this massive combinatorial space, you're trying to find the needle in the haystack. And the final thing you need is an objective function. You need to have a clear goal you can hill climb against that the search is going to do and the model is going to help you find. And in games — and why we used games at the beginning, apart from the fact that I love games, is that it's very convenient to train your AI systems on. In fact, before Go, we used Atari games as the earliest measures of how our systems are doing because they have clear objectives: win the game or maximize the score. But in science, there are also clear objectives, you know, minimize the energy in the system, for example, is a very clear one that often you find in biology and chemistry and physics. So as long as you can couch the problem in that way, and most of the time — in a surprisingly large amount of times — you can, then these types of methods that we are using can be very, very effective.
