DeepMind's Pushmeet Kohli on AI's Scientific Revolution

So I went to a biology conference and after I gave my talk, a biologist approached me and he said, push me. I have been working on this protein for the last 10 years and I had collected so much lab data to characterize this protein, to figure out its structure, but somehow this had. This has eluded like this has evaded all kind of investigation and we still didn't know the structure, but we had all this data. If we knew the structure, we could sort of validate it very quickly. I ran AlphaFold 2. It gave me the structure. It perfectly fit the answer. I've been working this for 10 years. Wow. What do I do next?

What happens when AI stops just answering questions and starts asking them? In this episode, Pushmeet Kohli discusses DeepMind's Alpha Evolve, a breakthrough evolutionary AI system that discovers entirely new algorithms. Pushmeet reveals how coupling language models with evaluators creates something unprecedented. AI that can tackle decades old math problems and generate human interpretable code that outperforms expert design solutions. PushMeet shares stunning examples of AI uncovering hidden mathematical truths and explains why we're witnessing the emergence of a new scientific method, one where AI doesn't just accelerate discovery, but transforms which problems we can even attempt to solve. Enjoy the show.

Pushmeet, thank you so much for joining us today. We've all been eagerly waiting for the moment that AI is capable of making novel scientific discoveries. Do you think that Alpha Evolve is that watershed moment?

Yeah. So it's certainly a key sort of milestone. What we have shown is that you have AI sort of model, a large language model, when coupled with a harness, is able to discover new algorithms. And not only that, it's basically able to prove new mathematical results which have been studied for many, many years.

You use the words when coupled with a harness. Can you tell us more about that harness?

Yeah. So if you now go back to AI models, the history of AI for science is very long. We have a number of different models that have tried to do scientific discovery. One of the key models in this category is alphafold, which is prototypical example of what can be achieved in by AI in science. We released AlphaFold 2 at the end of 2021 and it won the Nobel Prize last year. So the impact of AI in science is very well understood. Now the question is whether LLMs and foundational models, how can they impact science? In around two years back, we had a agent called Fundsearch in which we took an LLM and we coupled it with an evaluator. And the evaluator allowed the LLM to figure out when it was making new conjectures on sort of or coming up with new ideas to solve problems, whether they were hallucinations or whether they were brilliant insights. So essentially, in this particular case, hallucinations were great because some of those hallucinations were in fact, brilliant new insights that nobody had thought about. So this is where the hardness comes in, that you have this evaluation function and essentially a search protocol associated with the LLM that together is able to come up with completely new discoveries that are really impactful.

You mentioned FundSearch. Could you say a word on the difference between the results you all accomplished with FundSearch versus Alpha Evolve?

Yeah. So FundSearch was our first instantiation of taking a large language model and trying to see if it can discover new algorithms. The models at that time were weaker, right? And the type of search that we were trying to do, we had not sort of explored things much further. So what we asked the LLM to do was essentially try to complete a small function and see if it can do that much better. And surprisingly, it was able to discover completely new algorithms that mathematicians had been trying to study for a long time. But the limitation was that the mathematician or the. Or the researcher had to give a template in which the algorithm should be. Should be found. With Alpha Evolve, we have removed that restriction. Alpha Evolve is not just sort of searching for a few lines. It's basically looking at whole, sort of whole algorithms themselves, right? Very, very large pieces of code and sort of optimizing them over a long period of time. And secondly, FunSearch, our original sort of model, used a lot of function evaluations to discover, to make these new discoveries. Alpha Evolve can work with many fewer function calls, and it can basically, by looking at fewer proposals, it can discover new algorithms much more quickly.

Can you tell us about the role that the evolving Gemini models play in the capabilities of Alpha Evolve? And I think I saw in your blog posts, you have both Gemini, Flash and Pro involved in the harness. What is each responsible for?

Yeah, So I think, see, we have been evaluating or as Gemini improves with various sort of generations, it is becoming much, much better at its understanding of code. Now, if you have a proposal generator which can understand code much more effectively than it gives, it generates proposals which are not only syntactically correct, they are also semantically trying to solve the task. And then you are sampling what are the different ways in which the task can be solved. So as the baseline model, Gemini's abilities to perform coding improve our sample effectiveness in searching for the right Solution on these very hard maths and computational problem becomes much better. So if you want to search in a large space, there are two sort of elements. One is the speed of how you can generate these proposals and then the speed at which you can evaluate those proposals. Right. Like so, first, how quickly can you say, can you give me a new sort of candidate algorithm? And then secondly, how quickly can you evaluate whether the algorithm is any good or not? And both things are really important. And the fact that you have these variants of Gemini, like Gemini Flash, which can do that very efficiently and very quickly, this is really important.

I know AlphaVolve is more of a broad domain model than some of its predecessors. How broad is it? What's in scope, what's out of scope?

Yeah. So Alpha Evolve essentially allows you to search not only in terms, not only in terms of the size of what you can search, search over. You can now discover whole new algorithms, but it also is extremely general in its ability of thinking about algorithms in various different languages. So not only can it sort of search in C, but it can also do it in Python. It can also sort of do in Verilog, which is what the language is for describing chips. Right. In chip design. So the generality of Alpha Evolve is in its ability to search for these large algorithmic spaces, but also in different syntactic and semantic representations. Right. It's not restricted to a particular language like Python, but it can sort of do that search across many different types of languages and many different types of tasks. The only expectation it has is that you have a function evaluator that you can quickly evaluate whatever proposal there is and say how good it is.

It seems like the rough kind of cognitive architecture, so to speak, of generating a bunch of algorithm candidates, evaluating them, and then I guess evolutionarily deciding which ones to keep and then going forward from there. And this, it seems like it roughly mirrors the scientific method. Is that intentional?

Yeah. So I think there are also, like, if you think about it, there is another sort of agent that we released earlier this year, which was called Co Scientist. And in Co Scientist, essentially what you had was Gemini playing the role of the whole scientific academic process. So Gemini playing the role of a hypothesis generator, Gemini playing the role of the critique, Gemini playing the role of ranking different, of reviewing those ideas and ranking those ideas and then editing those ideas. So it was Gemini playing all these roles in a multi agent setup. And these were all sort of Gemini models prompted differently to play different roles and very sort of. Interestingly, this combined multi agent system came up with behavior that went much beyond a single Gemini sort of model's answer. So it was able to give much, much, much better proposals and new ideas compared to a single sort of model.

What's the intuition behind why that works?

Yeah, so I think that it is something that is still being studied, but it is a fascinating sort of thing. The one thing that I actually sort of noticed is that especially with regards to sort of co scientist, you would run sort of co scientist on a particular problem and the very first answer that you might get might not be very different from the baseline Gemini sort of model. But what happens is even as you sort of increase the amount of computation over sort of, you're not talking about just a few minutes or a few hours, but even days as the whole multi agent system sort of looks at the solutions and then refines them and sort of tries to sort of rank them, it just becomes much, much better. So why, why that might be happening? It might be that the proposal that there is deep insights or there is some sort of intuitions that are buried in the tail of the distribution, right. And then somehow Gemini's ability to evaluate which sort of proposal, which idea is better is much better than its capability to come up with sort of a new idea. It's the same sort of thing in computer science, right. Sometimes we are able to find if sometimes we know whether a particular solution is correct or not, but it's very difficult to come up with a solution. Right? So it's the same sort of thing appearing again in this multi agent setup that somehow the agents working together are able to extract many more impactful results.

It seems like the architecture of kind of generators and verifiers, it seems like that paradigm is being echoed across the broad AI space, whether it's very general models or very specific, kind of like AI systems for very specific applications. Is that fair that that's sort of the consensus architecture right now? And do you think that'll be the thing that people continue to push and scale?

Yeah. So I think there is going to be more work in agents. Right. What we are seeing is basically the very start of research on agents. Whether in alpha evolve you had a generator coupled with an evaluator. The generator was a neural network, a foundational model, an LLM and the evaluator was even hand coded. Right. But together with an evolutionary sort of search scheme, you were able to sort of get these much more effective results. In co scientist you didn't have just one agent, you had multiple agents working in a shared memory. Now what is the optimal agent configuration. This is still an open research problem.

Super interesting. Are the results that you're getting, are they different from the ways that humans would derive them? And I'm kind of thinking of the AlphaGo Move 37 stuff. Are the methods different? How do they compare to the ways humans would think about them?

So let's go back to the original motivation for why we started even working on the first iteration of using LLNs for algorithmic discovery, which was FunSearch. So a few years back, like, as you know, DeepMind has spent a lot of has done a lot of work in using AI systems for searching over large spaces. Right. We have done a lot of work in building AI agents which have been trained using reinforcement learning, which can deal with many complex challenges from the game of Go to playing starcraft, which are quite complex challenges. We set ourselves a challenge that can we take the same kinds of models, like the AlphaZero family of models which were extensions of or sort of extensions of what we had done in Go and the development of AlphaGo, can we use the same types of models for discovering new algorithms? And we came up with a new sort of agent called Alpha Tensor, which was particularly sort of focused on finding solutions for the matrix multiplication problem. And we found that this agent was able to improve over the past known results, which had stood for 50 years. But the key question sort of remained, can you do something better? And secondly, can you sort of come up with a solution that is more interpretable at the same time? Like when we were looking at practical problems in Google, like how do you schedule jobs in a data center? Now there has been a lot of work on coming up with new algorithms. And these heuristics have been designed by some of the best researchers and engineers at Google. And because they have a huge amount of impact in terms of compute utilization, if you use a typical reinforcement learning agent on this kind of problem, you might get better results, but it might come at the cost of interpretability. Because now you have a neural network deciding which workloads go to which computers, and if something breaks, then you don't know how do you debug this thing. So what engineers would really prefer is instead of giving them a neural network, you gave them a piece of code that they can interpret and they can run. And this was essentially the motivation, can we now use LLM? So instead of searching in the space of specific algorithms like we had done, metric falsification algorithms, like we had done in Alpha Tensor, or coming up with a neural network policy to directly solve the problem can we come up with an agent which can search in the space of programs and come up with a program that solves this hard problem. And the benefit of course will be interpretability, that you can see the sort of code, you can see what its properties are and so on. And, and that's what is what happened. We found sort of programs that not only were effective, but when the experts actually saw those programs, they could sort of recovers insights. So for instance, one of the maths problems that we had looked at for fun search was called the cap set problem. This is a problem that Terence Tao, one of the famous mathematicians, he is very interested in. And we collaborated with this mathematician, Jordan Ellenberg at NYU and when we looked at the program that FundSearch had produced, he found that there were certain symmetries that were in the problem that had not been recognized before. And somehow the program like OneSearch, the agent had discovered those and was utilizing those to get a better solution.

Can you say a word about, you mentioned working with Terence Tao and other famous mathematicians. Is math considered the gold standard for testing and benchmarking if these models are generating novel scientific results?

Yeah. So math certainly sort of has some properties which are very interesting. Right. The fact that it's very precise. Like you know whether the property that you are looking for, whether you have found it or not. Right. You know the matrix multiplication. For matrix multiplication, how many multiplications you require, like for a four by four matrix, what was known was that you can do it with 49 sort of multiplications. That's by stressor. And we showed that you can do it by 48. So that's a very precise result. Right. There is no sort of arguing about that. So it gives you a very crisp way of evaluating how well you have done. And there is no sort of like RLHF needed in terms of human feedback. Whether this was a nice result or whether this was a nice output or not. And you don't need to rely on sort of an LMSIS score, you just know that you are better.

Yeah. Okay. So then when you go from the beautiful pristine environment that is math to the real world, it seems like you all have found a lot of real world applications in data centers. In the Verilog world. Could you say a little bit about which applications you expect Alpha evolved to be most impactful for?

Yeah. So wherever sort of you can find a good function evaluator. Wherever you can find an evaluator where you can say I really trust this evaluation scheme. If you give me A program. I can tell you very concretely how good it is. If your problem satisfies that setup, then you can use Alpha Evolve. Because alpha, unlike a human sort of programmer who can try ten things or hundred things or a thousand things, Alpha Evolve does not like. It does not sort of. It can go on and on and on and on, right? It can come up with very counterintuitive sort of strategies to find, to solve that problem. Some things that you might not have ever imagined.

Can you have humans be the function evaluators or does that not work?

Humans, Humans can be the function evaluators. It's a question of sort of scale, right? Like how many can you sort of evaluate and whether you, whether you can evaluate the property of the program effectively. So at scale and with the right level of accuracy, how do you do that?

Do you build that into the application itself so that there's a human in the loop evaluating as it goes? Do you do that offline separately, before the application is produced? I guess.

How do you do that?

Or how do you imagine people doing that?

Yeah. So I mean, like, so we haven't used a human in the loop for Alpha Evolve, right? Most of our evaluators were programmatic evaluators. Right? But imagine a hypothetical scenario where Alpha Evolve was told that you have to solve this math problem and come up with a new algorithm to solve this problem. And suppose it came up with many different types of problem, many different kinds of solutions which are all equivalent in performance. Okay, but then which one is the best? It's the, the best is the one which is not only sort of very effective on the, on the problem, but is the most elegant according to a mathematician, or the most simple to understand. Right? And that's a very subjective human thing like simplicity or interpretability. Like we don't have a sort of crisp definition of depends on it is grounded in the human observer.

At what point do you need to pair kind of what's happening in the digital world to any kind of physical world stuff? I think in your blog post you mentioned that you could see Alpha Evolve being useful, for example, for material science. Do you need to be able to connect to a real world laboratory to kind of get any of that feedback? Or do you think all of this can kind of happen in the algorithmic domain?

Yeah, that's a very good question. And I think this goes back to how much do you trust the evaluator? If you, if sort of, if your evaluation was based on a computational sort of method, and the computational method was perfect and you completely trusted it, then you don't have to. Then you think, well, I believe the computational model, the computational model says that the solution that Alpha Evolved came up with is, is satisfies these properties. Job is done. Right. But if you don't believe that the computational model is the perfect characterization of reality, then you want to make sure that you sort of validate that result in the real world. Right. And you see whether that assessment of the evaluator was indeed correct.

As Alpha Evolve becomes more and more successful, as Gemini becomes more and more powerful, what do you think happens to these domains and how will the human scientists and engineers working in them adapt? So, for example, if you take chip design as an example you mentioned, these models are getting very good at creating new, generating verilog, creating new chip designs. Does that mean the role of a chip designer goes away? Changes? How do you think that this changes the world?

Yeah. So I think that's, again, sort of a very interesting question. I'll give you the example of what happened with AlphaFold. So we started working on this problem of protein structure prediction. So for those of you who don't know, like, proteins are the building blocks of life, they are the Lego blocks of life. And for many, many decades, scientists have been trying to figure out what is the shape of proteins. Because if we understand the shape of proteins, we understand how they function, and we can use that to sort of develop new drugs to treating sort of the most sort of challenging diseases on the planet, we can develop better sort of enzymes and so on. Now, in 2021, as I sort of mentioned, we release AlphaFold too. Before that, you used to take or even single protein sometimes 1 to 5 years to find the structure of a single protein, and it might take a million dollars. And there were some proteins which are so notoriously hard that people had been trying to study them for almost one or two decades and had not found the solution. And which is why only 37% recommend roughly 37% of the human proteins, their structure was known. So after we sort of released AlphaFold 2, I went to a biology conference. And because AlphaFold 2, with AlphaFold 2, we could find the structure of all proteins, not just human proteins, all proteins on the planet. And we made the structures available to everyone on the planet. So I went to a biology conference, and after I gave my talk, a biologist approached me and he said, pushmeet, I have been working on this protein for the last 10 years, and I had collected so much lab data to characterize this protein to figure out its structure. But somehow this has eluded like this has evaded all kind of investigation. And we still didn't know the structure. But we had all this data. If we knew the structure, we could sort of validate it very quickly. I ran AlphaFold 2. It gave me the structure. It perfectly fit the answer. I've been working on this for 10 years. What do I do next? So what has happened after AlphaFold 2? What happened is basically suddenly it did three things. It first, advanced structural biology. What was not possible earlier, which would take a synchrotron and six months and a million dollars is now done in a second, it really advanced what was possible. Secondly, it accelerated it and thirdly, it democratized it. Like that particular scientist working in Latin America or South Asia or Africa on some neglected tropical disease had no chance to sort of figure out the structure of their protein. They did not have the funds or have access to instruments that could find them the structure. Now they have access to those things to like the, any sort of parasite that they're working on. So what do they do? They are now working in this new model where structures of proteins are not hard to get. They are everywhere. And so they are working on the next set of things like how do you now use that knowledge to treat diseases and design better drugs? And I think the same thing will happen with the, with Alpha Evolved. Once you have these agents which can go beyond human abilities in solving these problems, then the question becomes which problems do we solve? What are the important characteristics of a chip that we need to improve on? Right. Like we want to make it much more efficient, much more sort of so that it requires less cooling, it requires sort of less expensive construction mechanism, it's more fault tolerant. Many other things you can make the problem more and more sophisticated because now you have more sophisticated systems to sort of optimize them.

While I have you something I've always wondered. The AlphaFold results are phenomenal and the story you shared with us is really impactful. Do you think that it's caused an inflection point in the kind of availability of new drugs? Or have there been other, are there other bottlenecks now that are just. We're faster at one part, but unfortunately everything else is just hard. So we're still slow overall.

No, it's, it's, it's so it has, it has speeded things up. But I think there's a. One has to understand that drug discovery is a long process. Now. What are the, what are the roadblocks for drug discovery? First you have to understand a Target, you have to understand, here's the protein in the body that I need to bind, because this protein is somehow involved in the disease. So if I can somehow bind something to this protein and change its function, it will have an effect that can sort of treat the disease. Like, first you have to come up with that conjecture. Then you have to say, okay, now I have a target protein, how do I develop a drug? How do I develop a small molecule or another protein that binds to it? So for that you needed to understand the structure of the protein, which other proteins that it interacted with, how did it interact with this molecule? This would take a significant amount of time, sometimes two years. Now that process is dramatically sort of accelerated. Now you can do it in sort of a few weeks or a month or a few months that took you multiple years sometimes. But that's not the end of the story after that. You need to now clinically validate it. So you have to go through phase one trials, phase two trials, phase three trials. You have to think about toxicity, all these other sort of things. So what AlphaFold did was take one blocker away, made the overall timeline faster. But there are other sort of blockers which our new generation of AI for biology models are hoping to accelerate and make much faster. So we have taken a big step, but we need to take a few more big steps.

What domains do you think will be most lucrative for this family of models?

I think the question is what is. I mean, the answer to your question is basically what domains do you think are important for society? Because AI is going to accelerate everything. It's going to accelerate health care, it's going to accelerate sort of the ability for us to develop more smart systems from healthcare to sort of material science. Like, if you think about the history of our civilization, we even describe our civilization in the sense of there was first we were sort of cave dwellers and then we went into the Stone Age, and then we sort of went to the Iron Age and then the Bronze Age. And now, depending on who you talk to, you are either in the Silicon Age or in the Plastic Age, whether you're stick or feeling a bit sort of sad. But if you take a step back and you think about what has humanity achieved? What we have achieved compared to any other species is the ability to transform energy, to leverage energy, right? We have been able to leverage energy and do big things with that power. Now if you can come up with, say, a new room temperature superconductor, that completely transforms your ability to handle energy, right? What changes will it bring about in society, they're hard to predict. If you can, if you can deal with energy in that way. Right. If we can unlock fusion. Right. And energy becomes so cheap every. Like, like if you think about geopolitics, if you think about economy, a lot of it is about energy. Right. And suddenly if energy sort of goes down to zero, what will be the impact on, on, on the economics of the whole thing? Similarly, like, if you think about like coding and if you have these agents which can code, what does that mean? If everyone can sort of code like intelligence sort of is completely ubiquitous. Everyone has access to all these different things. So there will be dramatic changes and everything will be impacted. So from, from materials to energy to sort of coding to healthcare.

Really cool. Do you think we're gonna have a fast takeoff moment moment for scientific discoveries? Do you think we're at the ramp of one? You think we're already there?

I think we are living through the middle of it. When you're in the middle, you don't really see it, but I think we are already in that era of AI accelerated scientific discovery.

What do you say is the biggest bottlenecks going forward?

I think sort of two elements. One is validation bringing, bridging the gap between the digital and the real world. Right. How do you validate some of that? So that is one sort of key idea, right. And really sort of capturing what is important for the problem. Right. And, and the second is sort of the other bottleneck is how do you make this technology accessible? You can build the most sophisticated technology. If people don't know how to use it, then then you will not have the, the impact that you want. Right. AlphaFold. Two, the. It was not just impactful and transformative because it had very high accuracy. Because even if like it had, it was quite accurate, but it was not perfect. And suppose it was accurate on 99% of the things that it predicted. It's definitely not at 99%, probably at the 90 or 95% mark. But suppose even it was accurate at 99%. The one person who got unlucky with their prediction and then spent the next sort of one or two years chasing a wrong prediction would then sort of say that I should not use it, I should not sort of use the predictions. So why is, why is everyone using alphafold? They're using Alphafold. Because not only Alphafold, not only is Alphafold good at making these predictions which are accurate, but it's also very good in understanding the limits of its predictions. When it makes mistakes, it basically holds its Hand and says I have made a mistake. So now if it is making you a prediction and saying I'm very confident, like most of the time it's correct and that's great. This is something that the LLMs of today don't have. They don't have calibrated uncertainty.

Should we close out with some rapid fire questions?

Yeah, sure.

Must read paper of the year.

Must read paper of the year. Oh, I would say Alpha evolved or co scientist.

Favorite algorithm nobody talks about.

Oh, the wake sleep algorithm. And very few people know about it, but it's essentially the idea. It's a paper from MIT from Kevin Ellis and Josh Tenenbaum which sort of talks about. It's a way of sort of doing training where you find some exploration and then you somehow build the gist of it, right? So think about sort of library construction and the analogy is library construction, right? You don't just want to write programs, but you want to also create the libraries that have common modules that will make all your future programs much easier to write. Hmm, very cool.

Agree or disagree? Inference time compute will be the next major lag of compute scaling.

Somewhat agree.

Okay, say more.

So I think inference time compute will be very, very important. I think also test time, sort of training time compute will be equally important. Right. We also like, if you look at distillation, how powerful distillation sort of has been, right? So if these models have an ability to sort of understand and conceptualize what these models are able to do and come up with better inherent representations, then they just become much more effective in making predictions. Maybe their sort of, sort of uncertainty improves and so on, they become more efficient, even robotics.

Bullish or bearish?

I'm bullish about everything, so I have to say bullish. I think everything will have an impact, but the question is basically near term or longer term. In the near term it will take some sort of. Getting robotics to work is challenging, but in the medium to long term, I think I'm bullish.

Humanoid robots, bullish or bearish?

We have constructed our world for humans, right? We like the human form. A lot of the non natural world around us is made for humans, has been designed for humans. Like from an architecture perspective. Right now humanoids have the same form as humans, so they will fit in in all these different architectures that we have built. Now whether they are the most optimal thing, that is not clear, but they certainly sort of have an advantage that we designed everything for the human form and now human humanoids have the same form.

Future Nobel Prizes in the sciences Will all of them be won by teams working with AI?

No, I think we would be getting there, but I think humans are still winning Nobel Prizes in the sciences. So I think there will come a point where AI will be indispensable. So it will be sort of humans and AI teams working together to achieving these amazing breakthroughs.

Pushmeet, thank you so much for joining us today. These are really fundamental, really general results that you're pushing forward at DeepMind, and we appreciate you joining us to share more about how you managed to do all this so far and what's ahead. Thank you.

Thank you.

SA.