HUGE* Conversations: w/ Google DeepMind CEO Demis Hassabis (Transcript)

So who couldn’t be interested in that?

But also, when I first heard it, I thought the kind of problem it was would be suitable for AI one day, even though, of course, this is in the late ’90s, we didn’t have any kind of AI that would be possible to work on this. But I thought one day that would be possible.

And then the final thing was just the impact it would make if you cracked it, because it would open up all these downstream possibilities for research, and especially in things like drug discovery and understanding disease. Which I think is the most important thing to apply AI to — improving human health.

CLEO ABRAM: And the reason that this would be huge for human health is that up until now, in order to develop new medicines, we’d have to spend hundreds of thousands of dollars and years of human effort to find out the structure of a single protein by shooting X-rays at it. So we had figured out some protein structures, but it was slow and expensive.

So I’m skipping over an enormous amount of hard work here by you and your team. But I think by the way that I’m asking the questions, it is very obvious to people that you solved it.

DEMIS HASSABIS: Yeah.

The Breakthrough Moment: Folding Every Protein in Existence

CLEO ABRAM: So there’s this moment where you realize that it is genuinely useful and you have solved what had been called one of the most important unsolved problems in modern medicine. And it’s 2021. You’re in a meeting. I am so glad that there was a camera in this meeting.

DEMIS HASSABIS: Yes, randomly.

CLEO ABRAM: Because it is one of the most incredible moments I have ever seen.

DEMIS HASSABIS: “Can we use AlphaFold to solve it?”

CLEO ABRAM: I think you’re talking with your team about setting up a system where scientists could send in a request for a specific protein, like a website, and then get the protein folded. And then someone else has a very different idea.

DEMIS HASSABIS: Yes.

CLEO ABRAM: Can you walk me through what happens in that meeting? And then your reaction is incredible. And I really want to know what you were thinking.

DEMIS HASSABIS: Yeah, sure. Well, look, it was funny that the cameras were there, happened to be in that particular meeting. It was crazy that it was that day. They very rarely followed us, but it was for that meeting.

And normally what happens for these sorts of prediction models is you— the traditional thing is you kind of set up a server, and then other scientists send you their protein sequences. And they say, “Oh, I’m interested in this protein, can you send me back the predictive structure?” So that’s how it’d been done in the whole field for the last 40-plus years. And the reason is that most of the prediction algorithms are quite slow. So maybe it would take a few days, and then you’d email back the structure, and then you’d ask for the next one.

But once I realized, sort of in that meeting actually, not only how accurately we could fold the proteins, but how quickly — in a matter of seconds — I was just sort of doing the back-of-envelope calculation. Like, how many proteins are there known to science, known in nature? 200 million. And then how many computers do we have? And how many would we need? And if we folded one every 10 seconds, I sort of realized, in the middle of that meeting, fiddling on my phone, that it would be possible in a year.

And so why go to all the effort of building the servers and the databases and the email client and all of that, when we could just actually fold everything ourselves — everything anyone could ever request and ever want — and then put it on a database somewhere for free for all the scientists in the world to use. So it just suddenly hit me, we should just do that.

“Why don’t we just do that? There’s this— we should just— that’s a great idea. We should just run every protein in existence. And then release that.”

Suddenly all these things must have been going on in the back of my mind, and I suddenly realized that that would be the obvious thing to do, and it would actually probably be less effort than standing up the server. So it would actually save us time.

A Free Database for the World’s Scientists

CLEO ABRAM: And you, in that meeting, your reaction is something like, “Why don’t we just do that? That would be way better. We should clearly do that.” And then you do. All of a sudden, this crucial process that had been so hard is suddenly fast and easy, and it’s being used by scientists all over the world. This huge unsolved problem now solved. Is it correct to say that we have now predicted the structure of almost all proteins known to science?

DEMIS HASSABIS: Yes. And we keep updating it. So every time somebody scoops a pail out of the ocean somewhere and there are loads of different types of organisms in that bucket of water, seawater, and then they sequence them all — the sequencing technology has obviously improved many orders of magnitude since the human genome was sequenced. So now the problem was that structural biology, the finding of these 3D structures, was far lagging behind the genetic sequencing.

So now with these computational resources like AlphaFold 2, we can actually keep up. “Oh, here’s a new million genetic sequences from some new strange organisms we found. Oh, here are the structures.” And so we have a small team at the European Bioinformatics Institute that keeps updating every year all the new sequences that have been found that year. So we’re now always at the cutting edge — we know what all of these different protein structures mostly look like.

CLEO ABRAM: That’s so awesome.

DEMIS HASSABIS: It is pretty amazing. It’s especially amazing, actually, for the people — researchers that work on slightly more obscure organisms or animals and things like, for example, wheat. A lot of plants have way more genomic data than mammals and humans, which is very strange. They seem to have multiple copies of their genome and things. It’s a kind of strange and bizarre world, I think, the plant world.

My plant scientist friends — they don’t have the resources that, you know, the human genome has had a lot of work done on it. But some of these more obscure organisms that are still really important for humanity, like crops and things like that, now we’re able to immediately jump to the science around what they want to do with the proteins. Maybe help them be more resilient to climate change, things like that. And they can jump straight to the problem they’re actually interested in, rather than getting bogged down with trying to crystallize the proteins that they’re interested in.

Another boon is for researchers who work on neglected diseases that affect primarily the more developing parts of the world — things like malaria, or Chagas disease, or leishmaniasis. These affect hundreds of millions of people around the world. But there’s not a lot of money in that if Big Pharma tried to research that and find cures, because they’re in the more poorer parts of the world. So they tend to be neglected, the research that goes into that.

So there are these amazing nonprofit organizations that do the research on that. But they don’t have a lot of money or resources. So giving them the structures of the proteins that are involved in, say, malaria is a huge boon for them too, because they can go straight to the drug discovery phase.

From Protein Structures to Real-World Drugs

CLEO ABRAM: That was one of the hardest things to figure out as I was doing this research, because there’s this moment where scientists all around the world have access to AlphaFold. You can see, like, the map lights up. You can see that people are using it.

DEMIS HASSABIS: Yeah.

CLEO ABRAM: But I wasn’t easily able to figure out what a great example would be to talk to you about — of a scientist using AlphaFold and then that speeding up a drug process that results in a drug that I could now take.

DEMIS HASSABIS: Yeah.

AlphaFold and the Future of Drug Discovery

CLEO ABRAM: What is your favorite example of a scientist using AlphaFold for something the audience might understand or have seen?

DEMIS HASSABIS: So over 3 million scientists are now using AlphaFold. We think it’s pretty much every biologist in the world at this point. And one scientist at a pharma company said to me that almost every drug developed from now on will have probably used AlphaFold in its process, which is sort of mind-blowing, really, and amazing that that’s the impact it’s having.

So it still takes time with drug discovery. We’re still mostly in the fundamental biology stage of understanding the disease. What is the protein we’re targeting? Is that the right biological mechanism? And then, as I understand it, some of these drugs are now in the clinical trials phase. And then hopefully, we’ll see in a few years’ time whole dozens of drugs that were partially helped by at least AlphaFold.

In terms of my favorite breakthrough so far that has happened with the help of AlphaFold, there’s this protein called the nuclear pore complex. It’s one of the biggest proteins in the body — it’s huge for a protein. What it does is a very important job: it’s basically the gateway that opens up and closes to let nutrients come in and out of the cell nucleus. So it’s basically the gate, it’s like a big doughnut ring that opens and closes. But we didn’t know until very recently what the structure of this was, because it’s so big and complicated. It’s pretty hard to crystallise and actually see.

Almost, I think it was pretty much 6 months or a year after we put AlphaFold out, some teams used it, along with experimental data, to finally work out what this beautiful shape was of this gateway protein. And that was amazing to me, as one of the biggest proteins in the body — and AlphaFold was very useful in helping determine that structure.

CLEO ABRAM: And so perhaps we can design drugs or treatments that better — that use that somehow, that better access the —

DEMIS HASSABIS: Yes, potentially. I think that was more for fundamental biology understanding. But obviously, I mean, we ourselves have spun out a new company, Isomorphic Labs, that actually tries to build on AlphaFold and uses it to, indeed, as one of the pieces of the puzzle, to massively speed up drug discovery.

So on average, it takes like 10 years to develop a drug. It’s a crazy long time, an unbelievable amount of hard work, very expensive, and huge failure rates. Only about 10% of drugs actually get through all the clinical stages. So we need to vastly improve that if we want to improve human health, I think. And I think the way to do that is by using in silico methods, AlphaFold being one of those components.

But knowing the structure of a protein is only one small part of the drug discovery process. You need a lot of chemistry — like what compound should you design to bind to it, all of these things. So we’re trying to, at Isomorphic, build these — you can think of them as adjacent systems that work with AlphaFold, more advanced AlphaFold, AlphaFold 3, AlphaFold 4, you could call it. And then end-to-end, basically create these drugs that have very minimal side effects, and are incredibly effective at addressing the type of disease we’re trying to help with.

We’re working on, I think at this point, like 18, 19 different drug programs across the gamut of things, from cardiovascular heart disease, to cancer, to immunology. So I think eventually, these types of technologies should be able to help across almost every therapeutic area.

CLEO ABRAM: In prep for this, I did a background interview with your fellow Nobel Prize winner, John Jumper. He really stressed that it’s one part of a larger problem of drug discovery. And so that brings us to the cutting edge today. I’ve taken some of the examples that I want to talk about. What is the cutting edge now?

The Cutting Edge: AI-Driven Drug Discovery

DEMIS HASSABIS: Sure. So we’re building many different components that can kind of go together. AlphaFold is one of the linchpins — that’s the structure of the protein. But if you think about it, let’s say you understand what the shape of the protein is. Now you know which bit of the protein is the important part that does its function. So now, if you think about drug discovery, say you want to block the effect of that protein or enhance it in some way, you need to know which part of the protein surface you have to bind to. So now you have to discover a chemical compound that will attach to the right place on the protein. And you want to know how strong will it attach — and then on top of that, even more important is not just will it attach to the thing you’re interested in, but make sure it doesn’t attach to other things. Because if it does, that would be toxic. So you don’t want it to have these side effects.

But because now we have all of these amazing algorithmic tools, we can do a virtual screen of, oh, here’s a compound one of our AI systems has designed. This is our prediction of how strong it binds to the protein surface. And then we can check that very quickly, like in a matter of hours — that particular compound, how does it attach to any of the other 20,000 proteins in the human body? We can just do it within a few minutes, and then keep modifying the compound so that it has less and less side effects, ideally none, on any of the other proteins, but an increasingly strong effect on the one that you want.

So you can see I’ve just outlined a self-improvement process, or self-modification process. This is extremely fast and efficient if you can do it in silico, on computers. And then only at the final stage do you check it in the wet lab. You still have to validate it. You make all these predictions, you do all your search in silicon, but then at the final stage, you check your final proposed compounds in the wet lab, and check it really does what the predictions say.

But you can imagine that would save — you can search thousands of times more compounds, or maybe even millions at some point, more quickly and efficiently that way. And then just at the end, check that they’re correct. That’s so much more efficient than doing the search in the wet lab, which is what effectively is done today.

AlphaGenome and the Future of Gene Editing

CLEO ABRAM: One of my favorites also is AlphaGenome.

DEMIS HASSABIS: Yes.

CLEO ABRAM: So I reached out to yet another Nobel Prize winner — good grief — Dr. Jennifer Doudna, who I’ve had on the show.

DEMIS HASSABIS: Oh, yes, fantastic.

CLEO ABRAM: And she sent a question for you. So I’m going to read this question from Dr. Doudna.

DEMIS HASSABIS: Okay.

CLEO ABRAM: She says: CRISPR, the gene editing technology that she pioneered, can now target nearly any DNA sequence. But for most genetic diseases, we still don’t fully understand which changes in the DNA are actually driving the problem, especially in the 98% of the genome that doesn’t code for proteins. With tools like AlphaGenome starting to decode that 98%, how close do you think we are to the moment where AI can reliably point to the exact genetic change causing a patient’s disease so that technologies like CRISPR can fix it?

DEMIS HASSABIS: Yeah, what an awesome question. I’ve discussed this with her actually in the past, and it is really exciting. I think with AlphaGenome, which is exactly that kind of technology, it takes the big, long genetic sequences, and then it tries to predict, if you had made a mutation to this particular single letter, single position in the genetic sequence, will that be a harmful mutation that might cause disease, or is it benign and it won’t do anything?

AlphaGenome, which we just released, is the best system in the world for predicting that. So that’s exactly what you then want — it’s still not probably good enough yet, but you can imagine a future version of AlphaGenome that is accurate enough to really know, oh, that particular mutation in combination with this other one. That’s the hard part — what if they’re multigenic diseases where there are cascades of mutations that cause the problem? Those are even harder to detect, but actually perfect for AI to try and help with.

Then you could go in with something like CRISPR maybe one day and fix that mutation, and fix the problem. So a kind of combination of things like AlphaGenome and CRISPR could be incredibly powerful. And hopefully one day we’ll be collaborating with the likes of Jennifer on that.

From DeepMind to Google DeepMind: What Was Gained and Lost

CLEO ABRAM: Last year, you said something to The Guardian that I found really interesting. You said that if you’d had your way, you would have left AI in the lab for longer. And the quote is: “Done more things like AlphaFold, maybe cured cancer or something like that.” From the outside, it looks like the story goes: you founded DeepMind with the mission to solve intelligence and use it to solve everything else.

DEMIS HASSABIS: Yes.

CLEO ABRAM: And then you sell to Google specifically because they will allow the freedom to explore science in this way.

DEMIS HASSABIS: Yes.

CLEO ABRAM: And for a long time, that’s your exclusive focus.

DEMIS HASSABIS: Yeah.

CLEO ABRAM: And then ChatGPT comes out, Google goes code red, and you become the head of all Google AI, including the consumer products that you weren’t spending as much time on before. And it feels to me like watching that from afar, it mirrors somewhat the larger experience of AI, which is just this incredible change.

DEMIS HASSABIS: Yes.

CLEO ABRAM: In the last couple of years.

DEMIS HASSABIS: Yeah.

CLEO ABRAM: What was gained and what was lost in that change?

The Race to AGI: Benefits, Pressures, and the Path Not Taken

DEMIS HASSABIS: Yeah, I think that’s exactly right. What you describe is sort of what, how it felt from the inside too. And for me, as I mentioned earlier, was that the AI, the best use case of AI was to improve human health and accelerate scientific discovery.

In fact, for me, I got into AI in the first place because I was interested in all the big questions in the world, the nature of reality, nature of consciousness, these kinds of things. And I felt we needed a tool to help us, even the best scientists, to help us make sense of the amount of data and information out there and find insights in that. And that’s happening, which is amazing. And obviously AlphaFold was our first and, so far best expression of that, let’s say.

And I always had that on my mind and many other problems like that. So it would have been great, I think, to— and given how important AGI is and how transformative a technology is, maybe the most transformative one in human history, then I thought it would be best to approach these kinds of the sort of latter stages of building it, which we’re in now, using the scientific method very carefully, very precisely, very thoughtfully, and rigorously with all the best scientists kind of in my ideal world collaborating in a CERN-like way effort on making sure we understood each step as we got to the final goal of building AGI. Seems to me like that would make the most sense with a technology like this.

And then, of course, you don’t have to wait. So that might take a lot longer, maybe a decade, even two decades longer. But I think that would make sense given the enormity of what we’re dealing with. And then my other idea was, but we don’t have to wait till AGI arrives to start getting the benefits of AI. We could use more specialized systems that maybe make use of the general technologies, the general algorithms we’re developing for AGI, but are not in themselves general intelligences. They’re narrow AIs, if you want to call them, like AlphaFold, which does a specific purpose and only that purpose.

And we could still, I’m still doing this, create many types of AlphaFolds and Isomorphics while we’re building AGI in this careful scientific way, and then humanity could benefit from the proceeds of that, like cures for cancer, or maybe new energy sources or new materials.

So I felt that that would be, maybe looking at this from 20, 30 years ago when I started out on all of this, that would have been the ideal way for it to play out, in my opinion.

Now, it didn’t happen like that because technology is unpredictable. And in fact, it turns out that things like language were a lot easier than we were all expecting. Even those of us who were obviously optimists about the whole technology and eventually we’ll crack language, but it seems funny to think of it now, but language and concepts and abstractions, things that the current models, foundation models like Gemini do incredibly well. We thought that maybe there would be 1 or 2 or 3 more breakthroughs needed before we could get there. But it turned out Transformers, which my Google colleagues invented, and some reinforcement learning as well on top, was enough to crack things like language.

And we were sort of playing around with that. So were the other leading labs. But of course, with ChatGPT, and fair play to OpenAI, they scaled it and then they put it out there. And I think even they say it was kind of a research experiment. They didn’t realize it would go so viral. And I think none of us did.

And we had sort of fairly equivalent systems at the time, because I think when you’re building that technology, you are so close to it, you’re very aware of the things it can’t do, the flaws it has. And you don’t realize that actually people out there would find use, even though it was hallucinating and doing other things that we’re obviously all still trying to improve on now, still not completely fixed. But there are still interesting use cases like summarizing things or brainstorming, things like that, that people use — everyone uses chatbots for today.

The Double-Edged Sword of Rapid AI Progress

Now, the downside of it is that we’re in this sort of ferocious commercial pressure race that everyone’s sort of locked into currently. And then on top of that, there’s geopolitical issues like the US-China race and so on. So there’s sort of multiple levels of pressure to sort of move fast.

The benefit of that, of course, you get faster progress, obviously. The progress is just at lightning speed these days. So that’s good for all the good use cases. The second benefit is that everybody, all of the viewers out there, everyone — you’re all getting to use the most cutting-edge AI technology, perhaps only 3 to 6 months behind what is actually in the labs. So that’s kind of mind-blowing.

It’s also great because I think it gives everyone a feeling for — it’s democratizing AI. It’s giving everyone a feeling for what it’s like to interact with cutting-edge AI and what it can do and what it can’t do. And I think that’s good for society to start getting — normalizing itself to what is going to be an enormous change with this technology coming. So it’s probably better that we get to sample that in incremental steps rather than it just being a shock to the system. There’s no AGI, and then here’s AGI one day. Probably that’s not good, although I think there could be many ways it could have rolled out.

And then the final thing that’s actually on the benefit side is that you can’t really fully understand your systems until they’re stress tested by millions of people. So it doesn’t matter how good your testing is, your in-house testing — obviously millions of smart people trying out things and then you seeing what bubbles to the top or the feedback you get is really important for building more robust systems and better systems.

So I think there are positives and negatives about the way it’s gone. It’s not the way I dreamed about years ago, where we would be sort of contemplating this philosophically and carefully considering each next step. We’re not in that world. And although I’m a scientist first and foremost, I’m also a pragmatic engineer. So we have to deal with the world as we find it and make the best of that. And we try to do that by advancing the frontier, but also trying to be as responsible as we can as we deploy these very powerful technologies like Gemini and AlphaFold.

Move 37: The Moment AI Creativity Changed Everything

CLEO ABRAM: There’s another story happening at the same time as this, and I want to get back to your concerns and how you weight those concerns and the cost. In order to understand that, I think we need to tell a story about AI being very creative, unexpectedly creative. And that story begins — let me find my Jenga block — that story begins here. So let’s go back to March 10, 2016.

DEMIS HASSABIS: Yeah.

CLEO ABRAM: There’s a very famous Go player that sits down to play against a system that you designed. And at this point, computers have beat humans at all kinds of games. But Go is really interesting because there are more potential moves in Go than atoms in the universe. They go back and forth, they’re playing. And then your system makes a move that is so surprising because it is incredibly unlikely that a human would figure out a move like that. Move 37. And you see Lee Sedol sitting there, he’s just got this shock on his face. He’s got his head in his hands like this.

And it really was this moment where I think people like yourself saw ahead to the creativity that we would find in AI systems that are very different than the systems that we’ve talked about so far. So there’s a category where you’re giving a huge amount of data and you’re asking to make new predictions. And I understand this is much more complicated than this oversimplification, but then there’s a category where you’re not giving data, you’re giving rules, like with math or physics or games like Go. And it has this incredible opportunity for creativity.

DEMIS HASSABIS: Yeah.

CLEO ABRAM: Where were you when that moment happened? And what future did you see ahead?

The Dawn of Modern AI: AlphaGo and Move 37

DEMIS HASSABIS: Yeah, it was an incredible moment that you’re describing, and it’s actually almost exactly 10 years ago now, which feels like a century ago, actually. But I think in many ways it was the dawn of the modern AI era, because until that point, there were many AI programs that could beat world champions at games, things like chess, but they were done with what’s called expert systems.

So they were systems where a team of smart programmers with a team of smart, in that case, chess grandmasters came together, tried to distill the knowledge the chess grandmasters have into a set of rules and heuristics. And then the programmers would build a system, kind of a brute force system, that would use a lot of compute, like on a supercomputer, like IBM did with Deep Blue to beat Garry Kasparov. And they would sort of encapsulate the rules they were given by the chess experts. And then the system would sort of dumbly execute those rules and heuristics and do millions and millions of searching of moves and then try and work out against those heuristics, which is the best one to do.

Now, the thing with that is, for me, that was not satisfactory when I saw that in the ’90s. I was doing my undergrad at the time. I didn’t feel like that was proper AI because that system, let’s take Deep Blue, okay, it’s world champion level at chess, at chess, but it can’t do anything else. Not only can’t it do language and robotics or any of those kinds of things, it can’t even play a strictly simpler game like tic-tac-toe, right?

So something’s obviously not quite right about the definition of intelligence, right? In the sense that no human — you could imagine a human grandmaster not being able to learn how to play tic-tac-toe. It would make no sense because it’s strictly simpler. So there’s something sort of wrong about its generalization capability and the fact that it didn’t learn. It was just given the answer, right?

So, if you could ask, for something like Deep Blue, where did the intelligence reside in the system? Well, it wasn’t in the system. It was in the minds of the chess grandmasters and the programmers. They solved the problem of chess and then implemented the solution. The program just dumbly executed the solution.

Why Go Was the Final Frontier

Now, Go, as you mentioned, is the sort of final frontier for games. It’s the most complex game humans have ever invented. It’s also the oldest game, so it’s just amazing in many ways. And it’s also very beautiful. So in Asia, where they play in China and Japan, Korea, it takes the place of chess basically and occupies that intellectual echelon. But it’s a much more intuitive game, sort of artistic game almost. So you play patterns that look beautiful and they turn out to be really strong, which is why the game has a little bit of a mystical element.

The top Go players would say to you it encapsulates the mysteries of the universe in the game. I think that’s how the ancient Chinese thought about it. And also just its raw complexity, as you mentioned, has more possible ball positions — 10 to the power 170 — than there are atoms in the universe. So what that means is there’s no way you can brute force it in the way that we did with chess.

Furthermore, because the game’s so intuitive and so esoteric, there aren’t really these rules that you can encapsulate easily for a machine to follow. So when you talk to a Go master, unlike a chess master, they’ll tell you things like, “Oh, why did you play there?” They’ll say, “It felt right.” Whereas a chess player will never say that. They would say, “I did it because I’m calculating this, this,” and then they’ll tell you the calculation. So that intuitive feeling is obviously very hard to encapsulate in a system. You can’t really program that directly.

AlphaGo and the Birth of Deep Reinforcement Learning

So it’s the perfect proving ground, I would say, for these new techniques that we were pioneering in the early days of DeepMind — deep reinforcement learning. Can you build systems that learn from themselves directly from experience?

So in the case of AlphaGo, AlphaGo started by looking at all the games on the internet that humans have played and learning the types of moves humans would do. But then we overlaid it with a Monte Carlo tree search that allowed it to sort of discover new branches of the tree of knowledge, if you like, in Go, starting with what humans knew and then going beyond that. And that’s what we hoped was going to happen.

So the amazing thing about that match, which ended up being watched by 200 million people around the world, was that not only did we win the match 4-1, that was the main objective, but in game 2 specifically, it played this famous Move 37 that you talk about — this creative move. It was on the 5th line of the board and early in the game, and it’s sort of a big no-no to do that in Go, right? Like, if you were being taught by a Go master, they would slap your wrist for playing on that because it’s just regarded as a bad move.

But not only was that a great move, it ended up winning the game for AlphaGo. Like, 100 moves, 200 moves later, it was in the right place, as if it sort of presciently put the stone there. So it was the critical move — not only was it a surprising move, it was the critical move for later, for it to be exactly in the right place to decide the game.

So obviously it’s changed the way all Go players play Go, but for me, it was the moment I’d been waiting for in terms of building a system — we’d already spent 6 years by then building these types of learning systems — that could achieve something no other system could. This sort of Mount Everest of games AI, the final frontier, if you like: can you beat the Go world champion? But also, not only did it win the match, but it was how it won, with these creative new ideas like Move 37. And that for me was the signal that we were ready to turn it to scientific problems like AlphaFold.

CLEO ABRAM: To say this back to you, the reason why it’s important that this audience that wants to understand the future understand what happened with Move 37 and Go is because the implication is if DeepMind can build a system that can do that, it can also perhaps build a system that can play any game. It can also perhaps build systems that can figure out in real-world problems what is the best solution — in quantum computing or in nuclear fusion or in matrix multiplication or chip design.

DEMIS HASSABIS: So many projects.

CLEO ABRAM: Could you tell me about the cutting edge here? Pick one of these systems. What is the Move 37 of these — the surprising creative element going on?

AlphaZero: Learning From Scratch

DEMIS HASSABIS: Yeah, I think AlphaZero is very interesting to talk about, which was the evolution of AlphaGo. So after we won, got to the pinnacle of Go and showed that it could come up with new ideas — at least in Go, Move 37, and actually many other ideas that it came up with, which has revolutionized how professionals play Go now — we then generalized it further to a system called AlphaZero, which I think is going to turn out to be a very important system today as well.

With AlphaGo, we started with all the human games that we could find on the internet. And also, there were a few other things that were specific about Go that were built into the AlphaGo system, like the symmetry of the board and things like that. So we wanted to get rid of all of those assumptions completely and actually start from scratch, as if the program and the algorithm didn’t know anything about what it was trying to do to start off with. And that’s what the “zero” refers to in AlphaZero — it’s sort of like AlphaGo, but now removing any human-crafted knowledge, both in the data and in any of the heuristics that we’d given the system.

So AlphaZero starts almost Tabula Rasa. Obviously it has a learning system, it’s got a neural network, we set up the parameters, but we didn’t give it any domain-specific knowledge about Go or any other game.

What we tested AlphaZero on was, first of all, could it learn Go from scratch and then beat AlphaGo? And we managed to do that. It takes 17 evolutions of the program. So you can imagine what happens: AlphaZero starts off random to begin with. It only has the rules of the game, plays randomly. Obviously it’s terrible at playing. It creates its own dataset by playing 100,000 games against itself. And then it can see which moves won or lost. And even though it’s playing more or less randomly to begin with, there’ll be some moves that are slightly better than other moves.

So now it takes those 100,000 games, and we train a new version — version 2 of AlphaZero — with that new data. That version 2 is slightly better than version 1. So now it’s not random anymore, but it’s not great, it’s not good, but it’s playing okay moves. And then those okay moves end up being better. And so then version 2 gets trained, version 3, version 4. Each time that new system gets played against the old system to see if it’s significantly better or not.

It turns out that at least in Go and chess and things like that, around 16 or 17 generations of that is enough to go from random to better than world champion. At least in the case of chess, which I actually once watched live happen — because I was fascinated, obviously playing chess myself — it starts in the morning random, then by lunchtime I could still just about compete with it myself. And then by tea time, it’s better than all grandmasters. And then by dinner time, it’s better than the world champion. And you’ve just seen the entire evolution of that from scratch.

Also, it’s playing interesting new chess that even chess computers like Stockfish — the more expert system brute force ones — haven’t discovered. So AlphaZero was the full generalization of the AlphaGo ideas.

Bringing AlphaZero Ideas Back to Foundation Models

And interestingly, I think we need these types of ideas back in now with our foundation models — the new Gemini and these kinds of things, which you can think of as generalized models of everything: language, the world around us, not just a game like Go. But we still need this ability to search and think and reason on top of those models. Sometimes we call those world models. And I think that still hasn’t fully been cracked yet — how to do that, bringing back some of these AlphaGo ideas, but now instead of just a narrow game, applying it to the whole world, and maybe interestingly, parts of science too, like material design and chip design and quantum computers.

All of these cool projects — when I see all these areas, I can’t believe we’re actually working on all these things, but it’s true. This is sort of the dream: I love every branch of science, and I get to indulge myself in all these different areas of science because AI is such a general tool. It can really make a huge difference to all these areas.

So maybe one example I give is just designing new materials. If we want a material with a special type of property, can we go beyond what is currently known in material science? And I think AlphaGo-like processes could be very useful there.

CLEO ABRAM: And the equivalent of a Move 37 would be like AlphaTensor finding a new algorithm that makes matrix multiplication better.

AlphaTensor, AlphaChip, and the Road Ahead

DEMIS HASSABIS: Exactly. Exactly. So you can apply it in algorithmic space, which is quite exciting because then the algorithm itself gets faster. So there’s some circular sort of improvement there. And yes, AlphaTensor, just making the matrix multiplication — which is the basis of all neural networks, it turns out everything’s matrix multiplication — if you just make that 5% faster, that’s a huge cost saving, given tens of billions being spent on training.

And so these are good examples of ideas. Things like the design of chips on a die, making it as efficient as possible, the routing — it’s a kind of NP-hard problem, like the traveling salesman: what’s the shortest distance you can wire up all of these things? AlphaChip and programs like that are really good, better in some cases than human chip designers at dealing with that.

So I think we’re just scratching the surface of what’s going to be possible in the next few years with today’s more general systems combined with these types of ideas from AlphaGo and AlphaZero. I think those ideas are going to come back.

CLEO ABRAM: These two categories — the story that starts with AlphaFold, the story that starts with AlphaGo — these are the kinds of AI that make me feel really optimistic. I also think that being really optimistic, and you do this a lot in public, which I appreciate, is fully thinking through the ways in which something can go wrong and what we can do to prevent that. So I want to insert one other in here.

DEMIS HASSABIS: Sure.

AI, Governments, and the Bigger Picture

CLEO ABRAM: This one.

DEMIS HASSABIS: Yes.

CLEO ABRAM: And the reason why I bring up this game is this is a real-time war game.

DEMIS HASSABIS: Yeah.

CLEO ABRAM: And in the videos where this system is absolutely crushing humans, you can see the engineers cheering for the victory of their system. But of course, as someone who didn’t build the system, I’m thinking to myself, what if that’s real?

And we’re speaking right now during a time when the debate about militaries and governments using AI is a huge topic of conversation. I want this conversation to last for 10 years. I want it to be useful for that long. So I don’t want to talk about specific companies, specific terms of service.

I also think people are in some way missing the forest for the trees here because bigger picture, governments are going to use AI. And so what I want to know from you as someone building these systems is, if you could wave your magic wand, what would you hope that they use it for?

DEMIS HASSABIS: Well, look, I think governments should be using AI, and we want to support all sorts of democratically elected governments. And I think the things I would love to see them use it for, and what we’re trying to build our systems to be good for, is things like improving public health, education. I mean, all of these things need to be rethought. The efficiency gains and the amount of good governments could do with it for their citizens could be incredible. And I think some countries are doing it — like Singapore and UAE, I think, are leaning into these types of use cases.

I would love to see it being used for things like energy, like optimizing energy grids. We did that with our data centers and saved 30% of the energy used for the cooling systems. I think there’s enormous societal gain from applying AI at scale to these types of areas. So that’s what I’ve always thought about and hope that governments will pick up and use, and we want to support all of that.

Of course, geopolitics of the world is very complicated right now. And these are dual-purpose technologies. I worry about a couple of use cases, things that can go wrong with AI. In the bigger picture, as you say, I think sometimes people get bogged down in the details, but actually there are two big-picture things to worry about.

One is bad actors — whether that’s individuals or all the way up to nation states — repurposing these technologies that we’re trying to build for good, like curing diseases and advancing material science and energy and so on, for harmful ends. Whether that’s inadvertently or intentionally.

And then the second branch of things I worry about is the AI itself going rogue or going off the rails. As these systems get more powerful — that’s not today’s systems, but maybe in the next 2, 3, 4 years — especially as we go towards more of the agentic era, which we’re entering now. And by agents, I mean systems that are capable of completing entire tasks on their own. We want those because they’ll be very useful, like as an assistant or something like that. But also that means they’ll be increasingly capable and autonomous.

And so how do we make sure — as one of the frontier labs, and the frontier labs all have to think about this — that the guardrails are put in place, that we can ensure that they do exactly what they’ve been told to do, or the goals they’ve been given, and they’ve been specified clearly enough, and there’s no way of them circumventing that or accidentally breaching those guardrails? And that’s an incredibly hard technical challenge if you think about how powerful and how smart and capable these systems eventually are going to get.

So I tend to worry about those — you could call them the medium term now, even though 3, 4 years is not really medium term. But those are the things I think people are perhaps not paying enough attention to at the moment. And I think they will be the biggest issues that we’re going to have to contend with if we’re going to get through the AGI moment in a way that’s beneficial for humanity.

What to Worry About — and What Not To

CLEO ABRAM: Yeah, one of the biggest questions I came in for you with — if I get an hour with you in my life — was: next time I read a headline, how do I weight the concerns that we’re all going to have over the next 30 years? Like, what are the things that people are worrying too much about? And what are the things that they are not worrying enough about?

DEMIS HASSABIS: So I think the two things I just mentioned are the things that maybe the average person is not worrying enough about. But even some of the experts and the scientists in the field — I feel like those are the key things that are more societally affecting.

There are other things that we need to worry about too, like deepfakes, and we try to help with those. Those are immediate-term worries — misinformation, deepfakes, these kinds of things. And we work on this system called SynthID, which is a watermarking system, actually an AI watermark — probably somewhere in one of these bricks. Yeah, one of them. And it uses AI to actually digitally watermark any generated image. So all the Google technologies — Veo and everything else — they all have this watermarking technology. We can detect and flag to the user or government or whoever that these are fake.

And I think, actually, I would advocate that all companies working on generative AI should build in something, some kind of technology like that. So at least it can be detected, or they can detect which things have been built with their technologies. And I think that’s going to be increasingly important.

But I think that still pales, as a small issue, compared to some of these bigger issues around AGI itself becoming very capable and how do we make sure that guardrails are put in place, that we understand what those types of systems are capable of as we get towards AGI. I think a lot more research and a lot more effort needs to go into that from everyone. And actually, I would love to see international cooperation amongst the leading labs around these safety issues — including with places like the AI safety institutes and also academia — to help work out how we navigate that next step, because it’s unprecedented to create technology like that.

The Limits of AI — and What Makes Us Human

CLEO ABRAM: If we play this out, what’s the limit here? What are the things that you think AI cannot do that humans can do? You’ve called this the central question of your life.

DEMIS HASSABIS: Yes, it is. And it’s very related to some scientific thinking of some of my all-time heroes like Alan Turing. He described Turing machines, which were these theoretical constructs that all modern computers are basically — Turing machines that are able to compute anything that’s computable. So anything that can be described as an algorithm, this type of machine can compute. And I think that the systems we’re building are approximate Turing machines. And potentially, a lot of neuroscientists, including me, think that maybe the brain — a good model for the brain — is an approximate Turing machine.

So the question is — and there are others like friends of mine like Roger Penrose, who believes there might be some quantum effect in the brain. And I’m sure you’ve probably done videos about that. We’ve had some very good-natured debates about this, but so far neuroscience hasn’t found any quantum effects in the brain. Doesn’t mean they won’t be found, but so far people have looked quite carefully and we haven’t found any. So it looks like most of what’s going on in the brain is kind of classical computation.

And so therefore, it’s not clear what the limit would be in terms of eventually what an AI system could do and could mimic. But I think that’s an empirical question. I think that’s one of the questions around consciousness. I mean, I don’t think it’s very well defined what it is, but we all intuit what it is. And I think this journey we’re on of building an intelligent artifact — I think we’ll have almost like a controlled study comparison to the human mind. And then I think we’ll see in this journey what the differences are and what’s unique about the mind.

And I’m very open-minded about that. I think there could be unique things, and certainly unique connections between humans that will never be replicated by these AI systems. But I think a lot of things that we currently think are not in reach — like long-term planning and reasoning and maybe some forms of creativity — I think eventually AI systems will be able to—

CLEO ABRAM: I want to be honest about what’s happening in my mind right now, and it is that I am doing exactly the thing that humans have done throughout history. I am trying to find the reason why we are special.

DEMIS HASSABIS: Yes.

CLEO ABRAM: It is that we have to be at the center of the universe. Oh wait, we’re not. We have to be the ones that are emotionally attuned. Oh wait, elephants have funerals. Oh, we must be the ones that can be creative and create art. Oh wait, Gemini can do that. Like, what? Oh, we must be special. Do you find yourself doing that as well? That’s my reaction as you’re describing this future of AI.

DEMIS HASSABIS: Yeah, no, I think we are special. And I think there are a lot of deep mysteries about how the universe works, including a lot of the things that are in our minds, but also things out there in physics.

I think that’s why I decided from a very young age to do AI — because I was obsessed, when I was a kid at school, with the big questions. Physics was my favorite subject at school because that is the subject you’re supposed to study when you’re interested in all the big questions. But the thing was, I realized — I guess as a young teenager reading all these science books and biographies on the best scientists — Richard Feynman is one of my all-time heroes — that although we’ve discovered a lot and we know a lot about the world, there’s so much we don’t know. Like, it’s just incredible.

Like, we don’t know what time is. I mean, this is insane to me. We can’t even describe something as fundamental as that — we’re just swimming in it. But what is it? Of course it’s entropy and things like that, but there’s nothing satisfactory about what it really is. And we don’t understand a lot of quantum effects and gravity properly, and consciousness —

CLEO ABRAM: All —

DEMIS HASSABIS: — actually, most of the things we care about. And we just sort of — I feel most people just distract themselves all day with TV shows and games and things and don’t worry too much about it. But I’ve never been like that. These deep mysteries kind of play on my mind all the time.

And I think I’m quite open-minded about what the answers might be eventually about what’s going on here, the nature of reality. I think that’s ultimately what I’m after. And I want to use AI as a tool to help us understand the nature of reality. And I’m quite sanguine about whatever the answer might be. I guess I’m a true scientist in that sense — I don’t actually have any pre-described notion of what the answer should be. I just want to know the answer.

The Sci-Fi Movie in Demis’s Head

CLEO ABRAM: Me too. One way to describe what you’re trying to do is effectively this — to create a system that wouldn’t be especially good at one thing or another thing, but rather to create, as you’ve been saying, AGI, artificial general intelligence, that would be good at it all.

DEMIS HASSABIS: Yes.

CLEO ABRAM: I know you’re a fan of sci-fi. I am too.

DEMIS HASSABIS: Yes.

CLEO ABRAM: Could you play out for me the plot of the sci-fi movie in your head that is the future where you actually do this?

A Vision for the Future: Post-AGI World and Human Flourishing

DEMIS HASSABIS: Yeah, I can. I think I love sci-fi too, and probably I read too much of it when I was a kid, may explain a few things. But one of my favorite series was the Culture series by Iain Banks. I think it just paints a really interesting, actually post-AGI world. He didn’t call it AGI, but that’s what he was describing, like 1,000 years in the future.

But I think even 50 years, some of this could happen where we’ve got through the AGI moment safely. It’s built, it’s helpful for society, and it’s here, and maybe we will have it in our Markets even. And we’ve used it to crack some of these, what I call root node problems in science. AlphaFold was one of those, right? So these are problems, if you think of the tree of all knowledge, these are kind of root node problems, which if you cracked it, it would unlock a whole branch of new research or new applications.

And I think there are other things like fusion we briefly mentioned, or better, maybe room temperature superconductors at atmospheric pressure. That you could then combine with optimal batteries and things like that. I think though there will be a solution to the energy problem. So free, pretty much free, renewable, clean energy one way or another, fusion or better solar.

And then that will unlock us to really travel the stars because the main cost of — Elon does amazing work with SpaceX and those things, but the main cost is still the rocket fuel, right? The energy cost. So if that’s sort of zero somehow, because we can just make infinite rocket fuel out of seawater because we’ve cracked fusion, so we can have catalyst plants and desalination everywhere, then that really unlocks space. And then we’ll be able to get a lot more resources because we can mine asteroids.

All of these things that are the purview of science fiction become, I think, very plausible. In the next 50 years. Dyson spheres around the sun, Mercury’s sort of conveniently in the right place actually, and made of the right material, which is kind of amazing if you think about what’s going on in the universe.

And then that should hopefully lead to maximum human flourishing, and we help cure all these terrible diseases, so we live much longer, healthier lives, and traveling to the stars, bringing consciousness to the rest of the galaxy. That would be, I think, an amazing outcome, and I think could happen within the next 50 years.

CLEO ABRAM: I believe you. You’re saying these things, and when you’re saying them, I believe you.

DEMIS HASSABIS: That’s what I’m trying to do, at least.

Final Reflections: Legacy and Purpose

CLEO ABRAM: Yeah. So this is my last question. If I were a fly on the wall at my own funeral, after they said she loved her husband and her family friends. I would hope that they would say that she spent her life trying to help people see optimistic futures so that they can be part of making them happen, that they can make them happen more quickly or better for more people or whatever it is that people decide to do with the vision that they see. And so my last question for you is, what do you hope that they say about you?

DEMIS HASSABIS: I would hope that they would say that my life was of benefit and service to humanity. That’s, I think, what I’m trying to do. So that maybe would be the best thing.

CLEO ABRAM: Thank you so much for your time.

DEMIS HASSABIS: Thank you.

CLEO ABRAM: Really appreciate it.

DEMIS HASSABIS: It was really fun.

CLEO ABRAM: Thanks.

DEMIS HASSABIS: Awesome.

The Jenga Game: DeepMind Projects on Bricks

CLEO ABRAM: If you want to play Jenga anytime, we have a simplified version of Jenga.

DEMIS HASSABIS: You did that very well. So yeah, this is actually awesome. I can’t believe how many projects we’ve done. It’s really crazy when I saw the bricks. So they all got our — yeah, they have all got our projects on them. Did you memorize where everything was? Yeah.

CLEO ABRAM: Okay.

DEMIS HASSABIS: Okay, of course.

CLEO ABRAM: So the game is you pull it out, and we were playing this. It’s unfair to play with you, but it would be — you have to say what that project was, and you don’t get the point if you get it wrong.

DEMIS HASSABIS: Oh my God.

CLEO ABRAM: So for example, it would be, Noam, this is material science.

DEMIS HASSABIS: Yeah, it’s a little bit unfair on you. I mean, I would hope I would win this game, but — although you’re probably way better at Jenga than me. What is this one? There you go. Okay. AlphaCode.

CLEO ABRAM: Yeah, that one.

DEMIS HASSABIS: That one’s clearer, right? Codeforces. Yeah.

CLEO ABRAM: That makes sense. This is genetics, but the 2% that codes for proteins?

DEMIS HASSABIS: Yes. We have to do this now. I’ve got time. I can push back my next one.

CLEO ABRAM: Great. Wait, I have one more question.

DEMIS HASSABIS: Do you know AlphaEvolve?

CLEO ABRAM: AlphaEvolve is coding.

DEMIS HASSABIS: Yeah, it can be used for coding.

CLEO ABRAM: Programming?

DEMIS HASSABIS: It’s combining genetic algorithms with Gemini. So this is our — this is one attempt at doing like AlphaGo stuff beyond what is known.

CLEO ABRAM: So I wouldn’t get the point for that one?

DEMIS HASSABIS: No, half a point, half a point.

CLEO ABRAM: Okay, one more question for you then while I have — I’m just going to keep going while I have you because why not?

DEMIS HASSABIS: Sure, we’re still rolling.

What We Didn’t Cover: Simulations and the Opportunity Space

CLEO ABRAM: Obviously. Okay, what did I not ask you that you think is important for people to know?

DEMIS HASSABIS: What did they not ask me? I think we covered a lot actually.

CLEO ABRAM: GenCast. This is weather prediction.

DEMIS HASSABIS: Yes. Oh yeah, we didn’t cover that. Navier-Stokes. I completely forgot about solving that whole bunch of things.

CLEO ABRAM: I completely forgot about that thing I did solving that.

DEMIS HASSABIS: So that was one interesting thing is simulations. We didn’t talk much about that or Genie, which is the role of simulations to DQN, of course, started it all off, the Atari stuff, simulations to help you to understand some area of science, or even social science, like economics, that you can’t — are very hard to run, either expensive to run experiments, or you can’t run controlled experiments in. So I’ve always loved simulation.

Oh yeah, ISO, there you go. We’re both very competitive, I think. So this is going to be quite serious. Actually, in Jenga, is the rules if you touch it, you have to move it? We are playing a loose — the easier version.

CLEO ABRAM: Also, because we were doing a creative thing where you’re allowed to push them together, you can use two hands also.

DEMIS HASSABIS: Oh, okay. You’re not allowed to normally do that, right?

CLEO ABRAM: Right.

DEMIS HASSABIS: Okay. I’ll just take this one. I’m going to cheat with AlphaCode again.

Advice for Those Who Want to Be Part of the Future

CLEO ABRAM: One of the questions I think people will have for you is if they’re watching this and they are very optimistic — Gemini, everybody.

DEMIS HASSABIS: Yes.

CLEO ABRAM: They’re very optimistic about the futures that you’ve described. They have all of your same concerns. They generally have gotten to the end of this conversation and they’re thinking, “I believe in this future and I want to be part of it.” How would you — CodeMender, I think that finds bugs in code.

DEMIS HASSABIS: Yes, very good.

CLEO ABRAM: How would you advise them to participate if this is all about helping people participate in the future?

DEMIS HASSABIS: When I do talks at universities and schools, I would say you’ve got to just go with the flow of the direction. I would immerse myself in every tool available and just become almost like superpowered with those tools and those capabilities, because I think my impression is even at the Frontier Labs, there’s so much work that has to go into just making the next versions of these Frontier models and then all the adjacent models.

So for us, like VEO and Nanobanana and Gemini, even we can only explore a fraction of what the applied things you could do with it, the applications you could make with it. And I think that gap’s getting bigger and bigger in terms of the overhang of the capabilities, all the cool stuff from the latest models. And the release schedules are getting faster and faster on that.

So I think the opportunity space is getting huge for people who are really expert at using those tools and then apply it to some new domain. I think a kid these days could probably start a multi-billion-dollar business in some ways using these tools in some new way that no one had thought about. And I think things like OpenClaw is a good example of that.

CLEO ABRAM: Yeah.

DEMIS HASSABIS: Yeah. Maybe we should call it a draw because I don’t think either of us could bear to lose that, right? It’s your move. It’s your move. Is it my move?

CLEO ABRAM: Yeah, it’s your move.

DEMIS HASSABIS: We can end on your move if you want. I will try my end of my move. Go on then. You’re going to make me make a move.

The Sticky Note Board: What’s Next?

CLEO ABRAM: I knew it was going to get me another move. In 2016, you had a sticky note on your board that said, “Solve protein folding, smiley face.” Yeah.

DEMIS HASSABIS: What is it like?

CLEO ABRAM: Yes.

DEMIS HASSABIS: Okay.

CLEO ABRAM: What is on the board now in your proverbial sticky notes?

DEMIS HASSABIS: Oh my gosh. I’ve got a pile of about 100 sticky notes on my desk.

CLEO ABRAM: What’s on it? Alpha chip, this is chip.

DEMIS HASSABIS: What’s on it? What’s in it? I can’t actually remember. It will be a list of about 30 things that need to be done by like this evening. So I better probably get to them. But look, great. Should we — do you want to actually —

CLEO ABRAM: I’m going to keep going until you stop.

DEMIS HASSABIS: So you can stop whenever you want. Yeah, let’s — what time is it? Okay. I’ll do one more move. But now we’re kind of cheating. We’re using the pieces that are already —

CLEO ABRAM: I’m going to go ambitious in our last — Oh, come on.

DEMIS HASSABIS: Come on. Come on. Come on.

CLEO ABRAM: If I get this one, I get another question.

DEMIS HASSABIS: Yeah. Okay. That seems fair.

CLEO ABRAM: Oh, God.

DEMIS HASSABIS: How is that going to balance? Surely not.

CLEO ABRAM: No.

DEMIS HASSABIS: No! Yes! All right. Thank you so much. That was awesome. Thanks. That was a great, great idea to have that.

Related Posts

• WWDC26: Platforms State of the Union – June 8, 2026 (Transcript)
• Transcript: Apple WWDC 2026 June 8 Keynote: Siri AI and More
• Elon Musk’s Interview in 2007 with PBS Wired Science (Transcript)
• TRIGGERnometry: w/ Dwarkesh Patel – People Have No Idea What Is About To Happen (Transcript)
• Generative AI Tools for Software Testing: How QA Is Getting Smarter in 2026