AlphaFold, it's already superhuman, right? No human can figure out just by looking at the sequence cannot tell you the structure of the protein.
Let's hype up everyone. Today, it's an honor to meet none other than Dr. Pushmeet Kohli, the name behind one of the biggest scientific discoveries, or you can say one of the most useful thing AI has ever created. His name is right next to Demis Hassabis on the research papers that won Nobel Prize in chemistry for AlphaFold. He has led the science team at Google DeepMind and cut down work of biologists from discovering protein structures, which took 10 years to now down to few minutes. So, hi Pushmeet. Can you please introduce yourself and what breakthrough are you most proud of?
I'm Pushmeet Kohli. I am vice president for science and strategic initiatives at Google DeepMind. I'm a computer scientist by background. And I lead the science team here at Google DeepMind. And in terms of what I'm proud of, I think with AlphaFold, not only was it a scientific breakthrough, but we made this tool available to the wider scientific community. And the fact that more than 3 million people have used it. If you had told me that there are 3 million people interested in protein structure prediction, I would not have believed you. But the fact that 3 million people have used it is like a big achievement. And I think I'm very proud of that fact.
And in fact, as you have mentioned in your podcast, like the time it used to predict the protein structures was almost 10 years. Cutting it down to few minutes, it's one of the biggest feat we have achieved. It's one of the most powerful use case of AI we have ever seen, right?
Yeah, absolutely. I mean, essentially, proteins are the building blocks of life. They are the LEGO blocks that make you and me. And the fact that we did not know the shape of many of these building blocks is, it shows you how much, how little, in some sense, humanity, although we have achieved a lot, how little we still know about the universe and our own selves. And the fact that AlphaFold was able to make these predictions, which are extremely accurate about the structure of these key constituents of life, was a very remarkable achievement.
And in fact, I remember 9 years ago you were giving a session when I actually joined my journey as a computer scientist. Your sessions were played in my university. And you were mentioning that Bayesian theorem is not going to take us to, you know, AGI or is going to take us to the biggest breakthroughs we're going to see in AI. And can you talk about how many breakthroughs were needed to reach this feat? Because your name is behind not just one breakthrough in science, multiple breakthroughs. And now we are talking about how with AlphaEvolve, like, you know, basically, AI can take over your job and actually discover, you know, scientific breakthroughs. How does it feel and how and why we reach right now here?
I think first thing to sort of mention is that with machine learning and AI and with many of these large breakthroughs, it's like a team sport. A number of people have contributed key ideas into the project. Let's take the example of AlphaFold. AlphaFold, although it leverages transformers, which is another sort of key breakthrough that came out of Google and has powered the modern LLM sort of revolution, it needs to be sort of refined to the protein structure prediction task. So, we came up with a different form of transformer called EvoFormer that was the basis of AlphaFold.
Now, there were many other technical innovations that went into making AlphaFold a reality. And there was a large collection of machine learning researchers, but also engineers and structural biologists that came together tried to integrate some of the things that we know about biology, about chemistry, about physics right into the structure of the AI model and how you train it to get it to accomplish this amazing feat.
Now, coming to the second part of your question, if we think about AlphaEvolve, the development of AlphaEvolve was motivated by the fact that if you are going after artificial general intelligence and AI which can sort of achieve many different things, then it should be able to also do tasks in computer science. And many of these sort of tasks and problems have stood for many years and sometimes many decades. To give you an example, modern AI and machine learning is all sort of at the heart of it is a very simple operator of matrix multiplication. So, modern neural networks are essentially multiplying matrices together.
Yet, even this very simple operation, if you ask someone, what is the computational complexity of matrix multiplication? Now, we were taught in university and in school that it's cubic. Like if you have an n by n matrix and you have another n by n matrix that to multiply it, you can do it by n to the power three sort of number of multiplications. But Strassen in 1969, while trying to prove that this was the case, showed that you could do it much faster. And still, we do not know the lower limit of how efficient you can do it in terms of the number of multiplications.
So, we asked some of our most advanced AI agents, first AlphaTensor and then now AlphaEvolve, to take on this problem. And what they revealed is that they can come up with these amazing sophisticated algorithms that can go beyond what Strassen was able to find in 1969. Right? So, it's amazing that we are living in a time where you have these amazing tools at our disposal that can come up with concepts that the human mind has not been able to discover so far. So, this is the most exciting time to be a scientist. Rather than replace scientist, I think AI will augment scientist to discover amazing new things and will, like explorers, so earlier we were only able to go to the next sort of like in ancient times, you would go to a new country and that would be a major thing, right? And as our technology sort of advanced, we were able to sort of travel to different continents and travel to the moon and maybe hopefully to the planets. So, it's the same thing, right? With AI.
With AI, you can take on many more scientific problems. And it's a great tool for scientist rather than it will augment science rather than sort of replace scientist.
Absolutely. You talked about Strassen problem. There's also couple of other problems written in the AlphaEvolve paper, such as the orange kissing problem, but exact is kissing problem, I believe, where you discover that the minimum amount of oranges that can touch the single orange in the center is 589, 588, which was thought before. This is mind-blowing how much we can discover decades after which was proven. And while discovering, while, you know, seeing so many breakthroughs and discoveries through AI, what are the difference you see humans discovering sciences versus AI or versus AI doing automated science?
I think science is, the discipline of science is to sort of solve, to understand the world, right? To understand the world, make predictions about it, and so on. Now, there is a, when you frame that as a problem, when you frame scientific tasks as prediction problems, what these AI models try to do, like AlphaFold or even sort of AlphaEvolve as an agent, they try to figure out the solution to that problem. And figure out how can you make these predictions? And they are much better at it than any human.
But then whether the concepts that they come up with, the calculations that are involved in making that predictions are human understandable, that's a different matter altogether. Right? So, as sort of the human scientists come up with theories which are also predictive, but are also understandable by other humans. Right? But AI models can come up with these prediction mechanisms that a human mind might not be able to comprehend because of the complexity of the calculations involved, but still, it's predictive of what actually happens.
So, the biggest bottleneck is not just making AI do discoveries, but also make it human discoverable. That's the point you're making, right?
I think the key point is basically that these models will sort of come up with predictions about the world. But then there will also be a challenge in figuring out how do they make those predictions, right? And sort of making, sort of, and interpreting how these models work will be a sort of discipline in its own right.
And will that be specific niche oriented? For example, right now we are building models specific for each niche like AlphaFold for protein structures, AlphaEvolve for specially, you know, creating algorithms with evaluators. And similarly, for each niche they are created. There are sort of lot of startups coming up. Some are partnering with DeepMind for each niche. Do you think the final goal is one system that can do discoveries in all area or it's going to be each co-scientist for specific niche areas?
I think there will be a balance of both. In certain types of problems they are so deep and involved that you will need to refine general purpose models for that specific problem at hand. So, I gave you the example of the transformer. The transformer architecture is pretty general, right? And it's powering the whole LLM revolution, but in order for us to use it for the task of protein structure prediction, we had to change it, modify it, refine it for solving protein folding, and so we had to come up with a variant of transformer called EvoFormer that was used in AlphaFold, right?
So, as what you will see is that as general models become more expressive and more powerful, they will be able to do many scientific tasks, right? And we are already seeing that with our LLMs, they can even tackle IMO level maths problems, the International Math Olympiad level sort of maths problems. But can they make progress on research level mathematics? Yes, maybe future versions of them may be able to contribute to research mathematics. But they will be a sort of point where a model which is fine-tuned for that particular task or has been designed in a way which is specific to the domain might still have some more advantages, right? In that domain.
Excellent. And now I want to get into your education journey and advice for the youngsters out there. You did your bachelor's in computer science from India, then you did PhD and also post-doc in UK and Oxford Brookes University. I want to ask you, let's say you're hiring for researchers who are going to shape the decade of AI, what is one single trait that separates the top researcher that you want to hire versus the 99th percentile of those researchers that you will consider?
I think the people who we generally hire are people who are really driven. In some sense it's not sort of even a job for them. It's sort of a life mission, right? They're really sort of focused on the type of, they're focused on the problems that are meaningful, that are going to have impact the world, and they care deeply about those sort of problems. There are different types of people, right? Some people are sort of really focused on the problem. Some people are very focused on the technique. But more generally, the basic sort of skills that they have is this innate ability to learn.
To learn more about different disciplines, not only being a master in their own discipline, but also learn new disciplines. And I think that's where a lot of scientific and research breakthroughs happen. When multidisciplinary research teams work together and you have ideas from computer science or biology and physics and chemistry coming together. So, I think the key advice to youngsters is yes, sort of really focus on the ability to learn, right? And think about concepts across multiple disciplines and don't think that your skill set is focused on one particular niche. You should always be constantly adding to your sort of bag of intuitions and bag of skills and bag of sort of competencies that you can exhibit for tackling a research problem.
