First of all, we've got to start with congratulations on your amazing win for them. In case you haven't heard, John was awarded the 2024 Nobel Prize in Chemistry for his work on Alpha Fold. It uses A.I. to predict the 3D structure of a protein from its chain of amino acids. Yeah, give it up. Okay. So tell us about this. When that what you just heard was a little bit of the call in which they told him they told him that he won. And it was it was funny because if you go on to say you thought you had a 10% chance of winning. Tell us what you meant by that and also how you felt when you found out about this. So. You know, for a long time or a long time ago when I was an undergrad or in grad school, I remember I had read this wonderful speech by Richard Hemmings, an incredible computer scientist.

You know, you and your research. And he says that you should, you know, scientists should try and do, you know, research that's worthy of a Nobel Prize. Write something. Something that kind of really changes how we think about the science that really enables us to do new things. And, you know, the prize itself is only given to kind of one discovery per year. And you never really know. And there's so many incredible pieces of science that come out kind of. You know, we had known for some time that this mattered to scientists. We had really seen it. We had seen something like 30,000 papers that had cited our work and used it and really used it to do their science right. We build tools that enable other people to do science, but it was kind of hard to imagine, really, that we'd get this kind of call from Sweden. And I remember thinking kind of, you

know, maybe one in ten rated people talked about it. It was starting to be kind of discussed as something that could get a Nobel. But I thought, not really. And in fact, I knew the call came about an hour before the press conference. And so it had gotten to 30 minutes before the press conference. And I remember I turned to my wife and I said, well, I guess not this year and not 30 seconds later had this phone call with a Swedish area code. And I was really glad it wasn't the meanest prank ever. And and just absolutely kind of unbelievable and extraordinary that they they chose to recognize us this year and and such a prestigious prize and recognized what I was doing for scientists. So some people were surprised to hear that a Nobel Prize for chemistry and also one for physics was awarded for work that's very focused on air. What do you think it indicates Might there do we need a new category for Nobel Prizes?

Does it show how much A.I. has been integrated into all kinds of research? I think and there's really the two different Nobel Prizes are for two different aspects. The Nobel Prize in physics is a really wonderful prize that recognizes kind of scientists that studied how learning can arise from physical systems, how it's kind of within the laws of what's called statistical mechanics or physics. I used to be a physicist. I'll try not to be too technical, but really, like, how does learning arise? And kind of the early moments that I remember a lot of that work was actually very influential to me when I was younger. Scientist And then the Nobel Prize in chemistry is about kind of computational science being able to solve problems we need to solve. And it's really, you know, about how A.I. and really the program we worked on Alpha Fold is solving problems that we don't otherwise know how to solve.

And I think that's it's it's about that. It matters to scientists who don't necessarily have an interest in AI or in tech, but they, you know, have questions about, you know, how does this piece of the body work? And then they turn to our tools to get answers to those questions. And so it's really about the meaning of this for scientists. Let's talk about Alpha fold in progress. What kind of progress have we seen with Alpha four fold thus far? So what we do this problem that we work on, so we really predict the results of a scientific experiment. So we're trying to use a computer to say, this is what you're likely to get if you were to do this experiment. And the experiment that we predict is called, for example, X-ray crystallography. It's taking pictures of proteins and proteins. Are these nanomachines of your cell, if you hear that DNA is the blueprint of

life? Well, it's the blueprint basically for how to build proteins. And these are a few thousand atoms, big, smaller than the wavelength of light. And when they're made, they're made. It's basically a string that folds up itself just due to the laws of physics into this intricate 3D shape. And that's what functions. For example, there are proteins in your muscle cells, and when your muscles contract, it's basically pulling via these proteins. And so scientists have long tried to get the shape of these proteins and understand it just from what you can read out of the DNA. And it still takes scientists a year or more to get a single answer right? It's an incredibly difficult experiment, if you want to put it in economic terms. It's about $100,000 an investment on average to get one of these structures. And but most importantly, it's often getting scientists want to study

Alzheimer's or they want to study. You know, there's a lovely example on fertilization or many others. They want to study these processes, and then they have to take a year to do this experiment. And we've built an AI system that gives very accurate predictions as well as a strong indication of how confident we are in those predictions and does it in about five or 10 minutes. So, it's really taking a year to 5 minutes, so it's shortening that timeline. But then, what happens after that? Are we seeing any meaningful drug discovery as of yet or anything else that outperformed, has led to that you can point to? So, we're seeing some really wonderful applications across both structural biology and drug discovery. And what we really see is that scientists are using this tool to discover new hypotheses that they then go test in the lab. One really wonderful example, I guess, is from Fong Jones Lab at I Believe, in MIT.

And they were studying this particular system found in possibly bacteria, I can't remember, which injects one protein into a cell in a very targeted manner. And they used Alpha fold to look at the structure of this and find the piece that targets this what's called contractile ejection system. And they swap it out for a new design. And they showed that in mice they can inject a particular protein into a specific type of cell in the mouse brain. So they're developing new types of drug delivery systems as the ultimate goal. We're seeing also with the latest versions of Alpha fold that we can predict how drugs will attach to proteins. And people are starting to say, okay, well now I can do rational design of drugs more quickly. And we're seeing really, I think kind of still at the early stages, drug design still takes something like ten years, but we're seeing a lot in which people are integrating these predictions and

skipping experimental steps to get further along drug development. When you think about the future, what kind of timetable do you think there might be for getting, let's say, some drugs, a drug on the market that would be used initially, alcohol abuse initially to help produce that? So I think we're already seeing the influence on early stage drug design. Now, of course, you still have to do clinical trials. There's nothing that skips clinical trials. We need to do this. So you're still, I think, seeing, you know, five, seven plus years to get the first drugs to market. But what I think you'll also see is really an acceleration. And we'll see this, we'll see many other tools. I also think Alpha Fold is probably a first tool, but we'll have many others that together will help us compress early stage drug discovery quite dramatically. But I think really the first we'll see

is probably five, seven years just because of clinical trial timelines. Okay. In a recent essay that he published, Dario Ahmadi of Anthropic praised our fold and wrote about a feature in which you use A.I. to perform, directs and improve upon nearly everything biologists do, which is quite broad, and that the use of A.I. tools such as Alpha Fold and others could shrink the next 50 to 100 years of biological progress into 5 to 10 years. That seems like a lot of shrinking. How likely does that sound to you? I think it I think in some parts of it I think are quite reasonable. I think the timeframe in which you'll start to see that happen, I'm not sure. I'm not sure it will be within five years. It will depend, of course, on progress in I, I think what we will get definitely almost certainly is very powerful tools in specific areas.

And we've actually already seen this in biology where we have, you know, very powerful tools to read DNA. And then we build very clever experiments which end in reading DNA because we're so very good at it. So I think we'll see quite a lot of that. And then I think it will depend really on how much further do we get in general reasoning systems. And it is a very interesting point that Alpha Fold is really showing that for certain kind of narrow problems in science, just like we've seen previously in say no problems in games, air systems can be vastly, vastly more effective than anything else we know how to build. And then humans at the same task, right? There's no humans that are good at predicting the structure of proteins. So I think if we are, you know, fortunate will really see this acceleration within I and I won't I can't guess 100 years in the future, but

I believe it will be dramatic. You mentioned with Alpha Fold, it's working on something that people could do, but it's extremely time-consuming. It's very expensive. It takes a really long time. It's not something that one person alone could do or could do very quickly. A lot of the systems that we see now are aimed at automating things that humans can do in a relatively short time span. A lot of task-related work. Do you feel like we should be? Ah, we may be concentrating too much on that sort of short-term work. Should we be concentrating more on these longer-term bets? I think there's a healthy mix. I think it is really the case that some of the places where we're going to see maybe some of the most dramatic within a single task improvements are probably going to be more narrow and scientific. We've really got this incredible general, incredibly general new technology that lets us learn from data, and for some problems like emulating what humans can do, we're making

progress. We're going out toward really general learning systems, but we're already seeing the dividends more narrowly. And I think that's going to be very exciting. And we also will benefit from all of the investment into GPUs and TPUs, into ideas and algorithms for how to learn. Those are all going to pay off enormously in the scientific domains and other domains. And I think we'll see some really exciting work that isn't just making us faster at human tasks, but otherwise. And I think it's hard to say we're underinvested in anything in AI, but I think it will be a very exciting time. So when you were doing the work on AlphaFold, it was, I mean, at this point, it was a number of years ago. So how was the experience of working on Bleeding Edge via Google changed over the past five years? I mean, a lot has changed in AI. Do you feel like a team like yours could

do work such as this that you could do five years ago? Could you do this work now? I think so. I think it's, you know, in a lot of ways gotten easier, in part because it's been proved out like the notion that AI is really within DeepMind and now Google DeepMind, of course, you know, there was always believed and always believed in the applications of science. But I think this notion that it will be absolutely transformative in fields has now been well-established in that way. It's easier. Also, you know, computers increase. There are lots of lots of things that have actually improved. If anything, it's gotten harder. It's it's that there are so many people now working in it that you have to find the right niche, although even I think still DeepMind has this advantage, or at least we've always worked with really great teams, we've always been really focused.

And I think one of the just really exciting things as a person working in AI is how much influence a small group of people can have working on the right problem in the right way, that it's actually very small teams that build a lot of really great things. Yeah, you mentioned when we spoke the other day that it was quite a small team behind Alpha Fold initially. How many people was it? It was about 15 that worked on it full time. Okay. So some you know, support from the wider Google DeepMind org, but the core team is around 15 and in fact not even very few of us with much biology experience. I was, I guess, biophysics for my Ph.D. but you know, maybe the worst biologist in my group but but the second best on the Alpha fold team. So I think it really is a small group of people that can find the right problem and change the direction you rent. You mentioned Richard Hemming, you in

your research, you mentioned it here. You also mentioned in the call when they told you that you had one and the idea that some people who win the little prizes can then feel like they have to focus on big problems to solve at the expense of not nurturing smaller things that can turn into big things. So how are you thinking about avoiding that? What is one of the small things that you're working on that you hope will turn into something big? So I won't go too much into I do try and find the right problems. I think for me, you know, I have it's weird. I feel like the crossroads of two moments. One is kind of, you know, as you mentioned in Dario's essay, this notion that I really, really transform the sciences. And I do believe that. So then it's how we should work on grand things, because the sciences are about to be transformed.

But then you have to make incremental progress. And I think one of the lessons of AlphaFold for me is, you know, we had this huge breakthrough, but on, say, our leaderboard, we moved 1% at a time and no idea was worth more than, you know, one point on this 100-point scale. And so, you know, by inches, we got to where we are. And I think, you know, some of the questions I'm very interested in how do we go beyond this kind of narrow world of proteins and atoms and think just a little bit wider, think about the cell, think about how proteins interact with each other, how we start to look at the cell and data around it, and how do we connect these predictions made at the kind of nanoscale, epic scale out to the microscopic scale. But we'll see. And one of the things, you know, the other reason I don't like to talk about what I'm going to do next is I want the

freedom to change it. And I think it's more important, especially as fast as technology is, to be responsive to the right thing to do tomorrow, which might not have been the right thing to do yesterday. I'm going to start saying that to my editor. I don't want to tell you what I'm working on next because I want the freedom to change it. Yeah, well, once you say everyone, I'm going to do this, then they'll ask you in a year or so, do we have Exactly? Yeah. Where is it? What's the timeline? Trust is one of the themes of this conference. I feel like it comes up more and more and more as consumers are increasingly accustomed to using AI in their daily lives. I mean, we're using machine learning in many ways previously, but it wasn't super, super obvious, I think, to a lot of people, except for until a few

years from now. What have you learned from your work that we can apply more generally as far as helping people know when they can trust information that is generated with A.I.? So this was one of the things we worried a lot about as we were planning to release Alpha Fold. And we were thinking like, how do we do this responsibly? And so, you know, and the thing that worries me is I worked in a mixed wet and dry lab with experimentalists and computational into my PhD, and I didn't want someone to look at our prediction and go do the wrong thing for the next six months. And so what? And as we were coming out with this, we said we need some way of knowing if it's right. And we actually ended up with a very kind of direct method, surprisingly direct as we teach the algorithm to produce an answer and a prediction of what the error would be if you knew the right answer of the answer it gave.

Kind of like if you take a test and someone asks you, What grade do you think you got? And that worked extremely well. In fact, it gave very what's called calibrated, but it gave a very clear answer. So it'll say, you know, I have to give some answer. But this part I believe, this part I don't believe, I maybe don't believe how these come together. And I've seen that really integrated into the work of scientists. And, you know, scientists are used to dealing with uncertain information. Scientists don't even necessarily trust experiments at other labs. And they've done, I think, a really good job at integrating the confidence that these A.I. systems produce. And it's an advantage that we have over large language models at the moment and integrating it into their practice in a very effective way. And I think it's going to be essential for scientific A.I. systems. Do you think maybe we could use a little

bit more of that in consumer-oriented systems? Like if you get an answer from a chatbot, perhaps it should include a confidence score. I hope people will develop things like this. It becomes more complex if we get. One of the nice things about science is that you can agree on the right answer, at least after the fact. Yeah, right. You know, the measurement was performed. We resolve disputes by actually measuring, but I think it is something that is going to have to be integrated. So we're going to have to kind of acknowledge that sometimes it's going to be confident, sometimes it's not going to be confident, much like sources on the Internet that we already do. And we have to figure out how to come up with this kind of view of how sure a machine is and the answer. Thank you so, so much. It was really wonderful to talk to you, and congratulations again. Thank you.

Going Inside Google DeepMind

Welcome to the Bloomberg Technology YouTube channel. Subscribe for original series and mini-docs on game-changing technologies, profiles on innovators, live conversations with experts, and breaking technology news & analysis from Bloomberg TV.

Going Inside Google DeepMind

Going Inside Google DeepMind

Scripts