This is Sir Demis Hassabis. He's the co-founder and CEO of Google DeepMind and Isomorphic Labs. For over two decades, Demis has been all in on artificial intelligence. He believes that within the next decade, AI could help us cure every disease on earth: cancer, Alzheimer's, Parkinson's, and even diseases we haven't discovered yet. So when his team at Google DeepMind invited me to Mountain View the day after Demis closed the I/O keynote, I had exactly one question I needed to ask him: what has this mission cost you that success can't repay?
It's a great question, I think.
Before we get to that question, let me tell you what I was actually doing in that room, because I didn't fly out to Mountain View just to ask Demis what AGI will look like in the next few years. I flew out because for the last 12 months, Demis has been telling anybody who will listen that the most important thing AI is going to do isn't making movies, replacing developers, or running a search engine. He believes that the single most important thing AI will do for humanity is to end disease for good. I wanted to know how he thinks that could actually happen. So that's where I started our conversation. You said something which I agree with. You said the number one application of AI is to improve human health.
And so I'd love to dig into Gemini for science. As I noticed yesterday, the most vigorous applause was for the scientific research.
Yes. What exactly does Gemini for science unlock for scientists? It's one of the things I'm most excited about. It sort of pulls together a few different projects we've been working on, probably most notably Code Scientist, which you can sort of think of as a fine-tuned version of Gemini that has extra additional tools and harnesses, citations, looking up literature, reading graphs, the kinds of things that scientists need to do.
Almost like it's this great researcher system that you have. I've had the chance to sit with a lot of tech executives over the last few years, and almost none of them get visibly excited about a product that isn't going to show up in the next quarter's earnings call. But Demis is different. This is the version of AI he cares about most, the version that turns scientific breakthroughs once thought to be impossible into things that suddenly aren't. To understand why he believes what he believes, you need a bit more context on his origin story. In 2020, his team released a system called AlphaFold. It does one core thing: it takes the amino acid sequence of a protein and predicts what shape that protein will fold into. Now, that might not sound groundbreaking, but trust me, it is. Every single thing happening in your body right now is mediated by protein: your muscles contracting, your neurons firing, your immune system recognizing a virus and fighting it off, all of it. And until AlphaFold, scientists had only mapped the structures of about 1% of all known proteins. It took them 60 years to get there. AlphaFold mapped the other 99%, 200 million structures, in about a year. By every standard that existed before it, that was supposed to be impossible. Then Demis and his DeepMind colleagues gave the entire database away for free to the public. That's the work that Demis was awarded the Nobel Prize for in October of 2024, alongside John Jumper and David Baker. But the AlphaFold story isn't a one-off. Demis didn't get lucky once. He's been at the forefront of AI for over two decades. Back in 2010, he founded DeepMind in London with two friends at a whiteboard. At the time, most of the world had never even heard the phrase 'artificial general intelligence' outside of science fiction. Google bought DeepMind four years later in 2014 for around $650 million. It was the biggest European acquisition Google had ever made, and Demis insisted on one non-negotiable condition: DeepMind could continue doing fundamental science research to push humanity forward. Which brings me to the people Demis is still working alongside on Gemini today, people you've probably heard of. I've heard that Larry and Sergey are actually back and reinvigorated over the mission.
Yes, of course. Yeah. They're actually coding away in the weeds of Gemini. And Larry is always around at the board meetings. And strategically, he's very brilliant with sort of far future planning. So it's super fun talking to them all the time.
Larry Page is in his early 50s, a billionaire many times over, and he's still in the codebase. The kind of person who doesn't actually retire, whose hands need to be on the thing, pushing society forward. Which is reassuring because there's a problem with the 'cure all disease' mission that nobody has fully addressed yet. If an AI eventually discovers a cure for something humans don't fully understand, if the math is too complex for a human brain to follow, are we supposed to just blindly follow it? That's the kind of question I wasn't sure I was smart enough to put into words, but I asked anyway. A question that I actually wrote down. I was like, you know what, I'm not smart enough to understand this. Let me ask Demis, a guy who is. And I said, if AI discovers a really complex cure, and it works, but the math and the logic behind it the human brain cannot actually understand, are we then at that point just having to blindly trust the AI?
Actually, this is a great question. The good thing about the practical sciences is we can test the answer empirically. So with a drug, for example, obviously you wouldn't just trust what the model says. You would need to test it in clinical trials and test it in the laboratory. The thing that takes all the time, and a lot of the cost, is the searching to find the needle in the haystack. But you still need to validate it at the final step. That would be kind of empirical proof that it's safe, but you don't already have to understand how it works. Although maybe if you pair an AlphaFold with a Gemini, a future version of that could explain what AlphaFold did in a way that a human scientist can understand. But even if not, when it comes to solving diseases, we really just care about the efficacy of the final outcome. What they do need to know, though, is any uncertainty around which bits the model thinks it's very confident about. And AlphaFold does that. So it outputs its confidence on the different parts of the structure.
That answer reassured me, because empirical validation actually works for drugs. You give them to test subjects, you measure what happens, and you find out if they worked. That's been the standard model for 100 years. But the part Demis kind of glossed over is that empirical validation takes time. I mean, we're talking over a decade and more than $1 billion to bring a single drug to market. Today, Demis has said publicly that AI can compress the discovery part, the needle in a haystack part, from years down to weeks. And I believe him. I mean, he's already done it once with AlphaFold. But if AI cuts the search to weeks and validation still takes ten years, that's still ten years to a cure. Which made me wonder: if the prize is curing every disease, why are most of the GPUs at Google DeepMind going somewhere else entirely right now? And I figured the only person who could actually answer that was the man sitting across from me. You've been really vocal that you think it's a possibility that we can cure all disease in the next ten years. But you've also been vocal that you believe the commercial success of generative AI is actually a hamper in that mission. Do you still believe that?
No, I don't think it's a hamper. I would like to see more work going on in the scientific fields, in the medical fields. A lot of the technologies that we're developing for things like chatbots and assistants and so on, not only do they increase productivity and create the commercial flywheel, which then allows us to fund our science work and in some cases give it away for free like AlphaFold, but it also develops some of the underlying technologies which can be used for the scientific domain. I would like to see more work being done in those kinds of medical fields.
I read the book. It's phenomenal. You are a humanitarian scientist. You also have to play the role of CEO of a massively commercial entity. Tell me about that tension when it comes to things like allocating computational resources, open-sourcing things versus proprietary things.
