GDM’s Pushmeet Kohli on solving science's biggest challenges with AI

0:00Today we're joined by Pushmi Kohli, who is the head of our science and strategic initiatives team. We are not looking for something which is an incremental improvement. We are really sort of looking at goals which are going to have a transformative impact. One of my general takeaways is like the pace of science advancement inside DeepMind is crazy. We will have more and more powerful models which will be more general, which will be able to do many more interesting things. Yes, we will have this amazingly powerful general intelligence. The question for our team is, what will we use it for? I love that framework.

0:35As we get intelligence, how do you actually wield intelligence to solve these problems? Our job is to leverage all that progress, add to it, to solve the next impossible thing for the benefit of humanity. Hey, everyone. Welcome back to Release Notes. My name is Logan Kilpatrick. Today, we're joined by Pushmi Kohli, who is the head of our science and strategic initiatives team. Pushme, thanks for being here. Excited to chat all things science. Yeah, thanks for having me, Logan. Let's dive in. I feel like we have, one of my general takeaways is like the pace of science advancement inside DeepMind is crazy.

1:12Like we have all these new alpha programs always coming out. There's tons of stuff. And hopefully we'll be able to deep dive into a bunch of these today. But do you want to, for folks who have maybe seen some of the launches, do you want to give us an overview of perhaps maybe like the three most recent alpha science launches that we've done? Yeah, so very happy to. So I think even in the last three months, we have been having a number of launches come up. The first one was AlphaEvolve, our super optimization, code optimization agent, which really showed a lot of impact in solving a number of optimization problems from optimizing data centers to speeding up Gemini training itself.

1:54Then came Alpha Genome, which was our model for deciphering the human genome, right? And that was very exciting. And then very recently, we had Alpha Earth, which was another sort of geospatial model for sort of giving, shedding light on what's happening on the planetary scale. I love that. I know you'll get a bunch of ignorant questions for me because I have a much better world view of the Gemini ecosystem than I do of some of our science work. But maybe to start, like for Alpha Earth, can you sort of double click as far as like what if people miss the headlines? Like, what is it? Why should people be excited?

2:34What problem is it solving? All that stuff. Yeah. So if you think about Google Earth, what Google Earth was able to do was able to integrate all this amazing information about our planet and making it accessible to everyone. What Alpha Earth does, it basically again sort of combines a lot of information that has been captured by remote sensing satellites and so on into one single representation. And what this semantic representation is then able to do is you can use it to propagate information that you know about a specific sort of attribute of a location. You might say, well, I'm looking at habitats of certain species, and I have found this species in this particular position on the map, right?

3:25Now propagate that and try to reason about where else will I find the same thing, right? So it essentially captures the semantics of what's happening at the geospatial sort of level at the planetary scale. I love that. How do you think about, as you described, just even those most recent three, there's like such a there's a big range of that work that's happening, I think, for your team, but also just generally for DeepMind. Like, how do you think about what what are those problems are we tackling versus what are they not? Because they feel like very like it's almost surprising to me that we'd be able to make progress in such different domains.

4:06So I'm curious to double click on how you how you think about that framework. Yeah, so I think if you think about the whole spectrum of intelligence, you can categorize it into maybe three broad categories. The first category are intelligence competencies that everyone sort of has, most humans have, right? Like all of us will be able to look at an image and say what is in that image. We will be able to recognize the characters in the image. We will be able to read handwriting and so forth. Then there is a second category of intelligence, which is more focused on expertise. If you see someone's symptoms, can you figure out what's happening to them?

4:54Can you make an accurate diagnosis? So that's expert level intelligence, right? Similarly, if you are given a specification, can you write a program that implements that specification? Again, that requires a lot of expertise. So that's expert level intelligence. Now, then there comes a point where you have certain tasks which maybe no human can solve. If you give a sequence of amino acids, which is a protein, and if you give that protein sequence to any human, like one of the smartest human beings, and say, can you figure out what is its 3D structure? We'll not be able to do that just by reasoning about it, right?

5:42It takes, before we released AlphaFold, it used to take almost a million dollars and sometimes multiple years to figure out the structure of a single protein. And so with AlphaFold, we were able to do that in a matter of a second for the cost of a few cents. And so those are the kinds of competencies that our science program tries to aim for, things that are not possible today, and really sort of building intelligent systems that can solve really the unknown. We were talking before off camera about what the sort of the different levels of impact from all these different initiatives. And I feel like actually the three that we released most recently are great examples of like the variance and the level of impact across like different categories in the ecosystem.

6:38So I'm curious to also double click on this like framework of how we think about impact from all this different science work. Yeah. So I think essentially the science and strategic initiatives per unit, we really focus on transformative impact. The important piece is transformative impact that will benefit humanity. And what is the impact type? The impact could be scientific, it could be commercial, or it could be social impact. I can give you sort of examples of all these three different categories. So the exemplar for scientific impact is AlphaFold. I mean, everyone sort of knows about it. It was essentially something that we sort of pioneered, one of the great success stories of modern AI, where this amazing fundamental root node problem in science, which has so many different applications from drug discovery to understanding human health to understanding, frankly, life.

7:40itself sort of was open for many, many decades. And AlphaFold was this remarkable system that could, from a protein sequence, figure out what is its 3D structure. And it made that big breakthrough at the height of the pandemic in 2020. And even in the last four years, it has been used in so many different places. And the scientific impact has been amazing. It was recognized as one of the most cited scientific papers and won the Nobel Prize last year for Demis and John. And so you can imagine in such a short span of time the amazing impact that AlphaFold had in the scientific community. Now, moving on from AlphaFold, I mean, of course, it had commercial impact as well because we spun off isomorphic labs with the specific mission of really revolutionizing drug discovery.

8:43And I think they are doing a very good job at that. Now, double-flicking on the commercial sort of impact point, one of the agents that we released earlier this year is AlphaEvolve. Now, AlphaEvolve is a Gemini-powered coding agent. It's not sort of giving you a solution for something that you do every day. it's essentially trying to do things which no computer scientist has been able to do. So imagine over the last two decades people have tried to make sure that how we optimize our data centers is fine-tuned. And some of the smartest people at Google have tried to find out what are the right algorithms so that how do we distribute jobs to different compute clusters.

9:38And that is an incredibly important thing because there's a lot of compute and monetary impact is to be had if you could refine this. And by running AlphaEvolve, not only did we make an improvement, but we made a very significant improvement. almost 0.7 % of the whole compute fleet we were able to save. And that's a huge sort of saving when you actually multiply it with the amount of compute that we use. Similarly, it was able to speed up Gemini training itself, which is a very important element. And now, again, although it was focused on commercial impact, it had amazing scientific impact as well.

10:25We ran it on a bunch of open maths problems. and for 75 % of them, it found the state-of-the-art solution, which the best mathematicians had been able to find. And for 20%, it went beyond them. So these solutions were not known. So that was sort of commercial impact. Now, coming to social impact. And a great example of social impact is SynthID. So SynthID is essentially our state-of-the-art watermarking system. Now, we have been one of the pioneers of Generative AI. And if you think about both the Vio launches as well as the Genie launch, which happened recently, you could see the impact this technology is going to have.

11:13But there are sort of also risks associated with that, right? That in the information ecosystem, the quality of these models and the quality of these results are so much that it's difficult for users to even sort of see what is synthetically generated versus what is natural, right? And this was one of the key sort of elements. Like for Google, it's really important that users understand what was synthetically generated and what was sort of physically represented the physical sort of world. And so SynthID was born out of that problem motivation. And I'm very sort of happy and proud to say that we are one of the first sort of companies that every Gen.AI content across all modalities at Google, whether it's text, whether it's images, whether it's videos, is watermarked.

12:04With this imperceptible signal, which is robust to different transformations of those signals and can be detected. Yeah, I don't think people appreciate how difficult technically of a problem that it's to solve synthetic. So happy to double click on that. But just in general, I feel like part of the challenge I imagine for your team, and I'm curious to get your reaction to this, is just like you have on some ends of the extreme, you know, alpha fold and like Nobel prize science impact. On the other hand, you're probably saving Google billions of dollars by, you know, the 0.7 % improvement in fleet efficiency and Gemini training in general.

12:43How do you like, from your team's perspective, is it just like, there's all of these huge opportunities, just like low hanging fruit sitting everywhere. And it's like, we can just go tackle a bunch of these. Or is it like, Are you actually really having to turn over a bunch of stones and do a whole lot of exploration to find these places where AI can have a huge impact and Google is well-positioned to solve? I'm curious how that initial process works. Yeah, so I think that's a very good question. And as you can see, there are a lot of problems that we could solve, but we sort of think about it in a very specific way.

13:25I think in terms of the what we try to solve, there is a very simple sort of algorithm that we follow. The first point of that algorithm is to make sure that the problem is going to have transformative impact and is going to be feasible, right? So we are not looking for something which is an incremental improvement. We are really sort of looking at goals which are going to have a transformative impact, whether it's scientific, whether it's commercial, whether it's social. It needs to be sort of transformative. And it needs to be feasible in the sense that everyone in the community believes that it's going to be transformative and it's going to be achieved at some point of time.

14:05We don't want to work on time travel. We want to work on things which are actually going to happen. So that's the first requirement. The second requirement is that there should be sort of almost consensus that no one will be able to achieve it in the next five to ten years. So it needs to be hard. If someone is working on it and thinking that it will be solved in the next sort of six months, then it's not a problem for us. Right. Like our team is particularly focused on these things which are incredibly hard and that require multidisciplinary, which require the best AI research coupled with the best engineering, coupled with the best compute, the best data to come together to solve this incredibly hard problem.

14:47And then the third requirement is we think that the consensus in the community that this is going to take five to 10 years is wrong. We can do it in half or one third of the time. If those three conditions are met, we take it on. Yeah, that makes complete sense. I'm actually curious through this framework, how does AGI fit into that? Like, is there a world where like, and actually as you solve, as you think about these, like you're solving very, it seems like domain specific problems. And like, is there a world where this like intelligence sort of generalization makes it easier to solve these problems?

15:25Or like you still think even in a world where we have really advanced general purpose technology from an AI perspective, you'd still need to go and like build a bunch of bespoke stuff or leverage those systems in a very specific way. And like the Gemini Alpha Evolve one is a great example of this, maybe, where it's like the model is really good and then you can supercharge it in a bunch of ways to help solve some domains of a specific problem. Exactly. It's exactly. The idea is that progress will happen in AI and we will have more and more powerful models, which will be more general, which will be able to do many more interesting things.

16:01like when we were thinking about reinforcement learning and being able to solve search problems, whether it was AlphaGo, whether it was AlphaChess and so on, there is development that is happening which is making things possible that was not possible earlier. And our job is to leverage all that progress, add to it to solve the next impossible thing for the benefit of humanity. Right. And the key question is, yes, we will have these this amazingly powerful general intelligence. The question for our team is, what will we use it for? The specification, the problem that we will sort of that we will apply this amazing technology that will eventually sort of be created.

16:49Yeah. I love that framework of like as we get intelligence, how do you actually wield intelligence to solve these problems? And this is my plug for the DeepMind team in general, which is I think so many of the labs and I think this Demis said something like this, like so much of the narrative about AI right now is that in the future it will be used for scientific impact. And I feel like DeepMind historically is the only place where like it is actually today being used for scientific impact across the ecosystem. I feel like a bunch of the examples you gave are like the proof is in the pudding that DeepMind is doing this work, which is awesome to see.

17:25I'd love to, again, through my worldview of like, I spend much more time thinking about Gemini stuff than I do our science initiatives, just because it's closer to my world. How are those two worlds connected? Like the science unit and the work that's happening across Gen.AI and across the Gemini world, like how do you collaborate with those teams? Like how does the technology transfer happen? You talked about Alpha Evolve where some of the foundational components are the same. I'm curious to hear that story. Yeah, so I think this is the other thing about DeepMind, I think, and Google DeepMind, which is the culture, which is extremely collaborative.

18:02We collaborate very, very closely with Gemini. I can sort of give you many, many examples, whether it's sort of collaborations on the base architecture as to what is the next sort of generations of architectures that we should be developing that would make our model, which will make Gemini be able to sort of understand and perform much better on scientific tasks. So this is one sort of area that we sort of collaborate on. We collaborate on evaluation. what are the next types of things that we should evaluate the system on to really see how that's making progress so and then we also collaborate on data right what kind of data should we train gemini on so that it has a really good understanding of these amazing specific domains that we want to go after whether it's a sort of biology whether it's chemistry materials sort of code cyber security, like we really want to make sure that Gemini has a good handle of all these different types of domains.

19:06And then of course, there are projects that we work on together. For instance, the IMO project, like has been a collaborative project from Gemini and the science team. And as part of it, we first worked on alpha proof and alpha geometry. And then this year, we basically learned a lot and then tried to integrate all of that into the Gemini DeepThink model. And we have got this amazing sort of result, which really sort of shows that both the generality as well as the pioneering aspect of sort of being able to go for that very ambitious challenge can work for both kinds of models. Yeah, let's actually double click on this example, because I think it's perhaps like one of the best versions of the story, which is I think the original, so we got like the IMO, I think it was silver metal.

20:00We were like one point away from gold last year. That was a domain-specific model. That like wasn't the base Gemini model that was available to the world. And that actually was, I don't remember, was that like a custom alpha? Was it an alpha version of the model? Yeah. Or do we brand it in that way? Yeah, so it actually was a collection of two models. One was called alpha proof and the other one was called alpha geometry. So alpha geometry, by its name, you can sort of guess that it was specifically focused on solving geometry problems. And then there was alpha proof. And alpha proof had this very interesting sort of aspect is that there was an underlying LLM in alpha proof.

20:39But what it was doing is essentially searching in the space of all proofs. Right. So it will formulate the problem. It will take your math problem, convert it into a formalized problem specification in a domain-specific language called Lean. And what this sort of language is able to do is it's able to give you proofs. If you solve the problem, you will get a formal proof so that you know that it's actually correct. That if you find the solution, it is actually correct. And just to clarify, for folks who aren't familiar with, and I know pretty much nothing about formal math proofs, I'll put myself in this camp, the nice thing is that it has this verifiable component.

21:21Exactly. It's not like a question of if it runs, then essentially you know that this thing is a true version of it's accurate. Yeah, it's given a problem statement, and it's told to prove that problem statement. and if it finds the sort of a proof and says oh i've proved this then it's actually correct and is it always just a double clock in this anytime you have a proof like that is it always like mathematically uh like i'm thinking about code as another proxy yeah i could write a bunch of like horrible programs that are like garbage code to solve some problem and like it's not like i wouldn't uh it can still be very like a bad solution is that the same in this like proof world or like if it solves the problem it doesn't matter if it's like not elegant or something it's like it's a everyone's happy with it solving the problem that way yeah so i think mathematicians would tell you that math is about elegance so of course basically the proof can be elegant or the proof could be like really really long and imo actually from the actual competition does like grade based on the elegance to a certain extent right it sort of grades on both aspects right Like it's not as if you just say, oh, here's the answer.

22:34You will not get full marks. You have to actually show how did you arrive at that answer. Right. So in that sense, basically all our proofs were given full marks because they showed how you arrived at that answer in a very specific way. Now, whether those proofs are understandable by a mathematician or how difficult were they to be understood by a mathematician? That's another question. Yeah. Yeah. So we have those two models last year. Can you tell any part of the story of how do we get from those two domain-specific models, AlphaGeometry, AlphaProof, to now DeepThink, built on top of Gemini 2.5 Pro as a base Gemini model?

23:14Is there a bunch of synthetic data generation that then feeds into the mainline model training? Or what's the technology transfer in that direction to mainline Gemini? Yeah. So I think the first sort of element is that Gemini itself has improved a lot, regardless of what we incorporated, right? The sort of the deep think, the thinking versions are generally much more advanced in terms of their reasoning. Now, in terms of what we learned from AlphaProof is AlphaProof had this amazing ability to find valid proofs. And as we sort of said that when it finds a proof, it is valid. We know it's correct.

23:52So now you have this amazing ability that you can generate hundreds of thousands or even millions of problems and try to solve them using alpha proof. And wherever it's successful, you know, here's the problem and here's a valid solution. So now you can generate training data. So this is training data that can be used for training Gemini saying, see, alpha proof found it. You should also be able to find it. So it's an amazing data generation strategy for training the next generation of Gemini. Something that I've, to delve into this a little bit more, something I've struggled with is how that math skill transfers to general domains.

24:36I'm curious if you have a point of view on this, but like I don't intuitively when I hear as like a somebody who wants to build like I'm building a customer support app or I'm building, you know, even like a code generation app or like some vibe coding app or whatever it is, you know, models being good at math does not intuitively in my mind translate towards like I should care about this as someone who's building with the technology. So I'm curious, like, is it is there some underlying principle of math that, like, makes it so that it's interesting? Or is it there are math use cases and we should make the model better at math if we can, but we don't actually expect that to generalize across other domains?

25:13Yeah. So I think that that question itself is a research question in the sense that there's no sort of conclusive answer. It's very much an empirical thing at the moment. Right. It could have a huge impact. It could not have a it might not have an impact, but it's all analytical and empirical. We have to really see if we add this data in, of course, it should have impact in maths, but which other abilities it impacts? That is something that we look at very systematically through ablations, saying, okay, if we add this, what is the effect? If we delete it, what's the effect? On various different evaluations.

25:52And then that's how you get the real picture. too. Yeah, I feel like instruction following is maybe like as you're describing what alpha proof and alpha geometry do in this like explainable how to solve a problem. I feel like that's like that sort of generally tracks, I feel like to a lot of language model capabilities where you want it to. Here's an example of me doing something in practice, follow a bunch of examples like this, generalize to other random variables that might show up. So I do feel like there's, I could, I could grok how that might be beneficial generally. Just to double click on this, I think part of the story of this IMO achievement is that we're actually not using a domain-specific math model anymore.

26:34The capability is now generalized into the DeepThink model. Can you double-click on this? Because I think that's been one of the cool things about the DeepThink moment. Yeah, exactly. I think what happened this year is that Gemini made a number of significant advances in one go. So not only did we move from silver to gold, but we did that using natural language specifications. So the problems are not now translated in a specific mathematical language like Lean. They are just normal, your English language, anyone can sort of specify them. And Gemini could take those problem specifications and solve the problem.

27:17The second thing that has happened is that this is almost a generally available sort of model, not a very custom alpha-proof model that was only available sort of within DeepMind and required a significant amount of search. It is almost a model that can be used by anyone on the planet. So I think these are amazing sort of transformations that have happened even in just the last one year. Yeah, I want to double click on this, how we make the technology available to a much broader audience. But I just as another point, like we, we announced Genie 3. And this was my immediate reaction. I think the world's reaction is like, we want to use this stuff.

27:59And I feel like it's so cool that for deep think and for this IMO gold model, we're actually we have it in a way that people can access it and like experience the model themselves and play around with it. But tying back to some of the other initiatives, and I think like, on the IMO side, like good example, deployed through the through the Gemini app and customers can access it there. For some of the other initiatives that we're doing across the science spectrum, how are those actually showing up in the world? How are people accessing them? What's the deployment story for those things? Yeah, so I think this is a very good question.

28:31And from the very start, the whole initiative and the whole unit was focused on the idea that we are not just making these breakthroughs. Part of it was to make this available to the world to have that impact in the world. So to give you an example with AlphaFold, it's not that we just created this AlphaFold system and kept it with us, right? We actually made it accessible through an API. And what we did with AlphaFold was we, in fact, predicted the structure of almost every known protein on the whole planet. and then we sort of placed it on a database called Alpha Pole Database and now anyone on the planet could sort of use it.

29:15So that researcher in Brazil or in Africa who was working on some neglected tropical disease who had no sort of way to get a structural characterization of the protein target that they were interested in now can just put their protein onto a web page, click a button, get the structure. And that sort of completely democratizes sort of AI. And we are not just doing it for AlphaFold. We have done it for almost all the models. Like Alpha Genome, we built a custom sort of UI for allowing researchers to figure out what is the effect of mutations or variation in the human genome. And so again, sort of all these models have required innovation to bring them to the broader scientific community so that they can eventually use it.