Demis Hassabis on AlphaFold, AGI, and Why Biology Is the Hardest Problem

Demis Hassabis has been thinking about intelligence — human, artificial, and otherwise — since he was a teenager programming games in his bedroom in North London. He co-founded DeepMind in 2010, sold it to Google in 2014, won the Nobel Prize in Chemistry in 2024 for AlphaFold, and is now leading Google DeepMind’s push toward what he has publicly described as artificial general intelligence.

We spoke for ninety minutes at Google DeepMind’s London offices. The conversation ranged from protein folding to consciousness to what Hassabis thinks AlphaFold’s success tells us about the nature of scientific discovery. What follows is an edited transcript.

AlphaFold solved a problem that biologists had worked on for fifty years. In retrospect, why was it solvable by a machine learning system when it resisted analytical approaches?

The protein folding problem is, at its core, a pattern recognition problem at a scale and dimensionality that human intuition cannot navigate. Analytical approaches tried to derive folding rules from first principles — from physics, from chemistry. That approach kept running into the combinatorial explosion of the conformational space. There are more possible configurations for a mid-sized protein than atoms in the observable universe.

What machine learning systems are good at is finding structure in high-dimensional spaces without needing to enumerate the space. AlphaFold didn’t derive folding rules. It learned the statistical regularities that govern which configurations proteins actually adopt, from the training data of experimentally determined structures. The answer was always in the data. We needed the right approach to find it.

Does that suggest that there are other fifty-year problems waiting for the same approach?

Many. We’re working on several of them. Drug-target interaction prediction. Materials discovery — finding novel materials with desired properties without having to synthesise and test them experimentally. Climate modelling at resolution and timescales that current approaches can’t reach. The common thread is that they’re all domains where the answer is, in principle, latent in existing data, but the data is too high-dimensional for human analysis.

Where does that leave human scientists?

Asking better questions. That’s not a consolation prize — it’s the most valuable thing scientists do, and it’s the thing that AI is furthest from being able to do autonomously. AlphaFold predicted protein structures. It didn’t decide that protein structure prediction was important, or design the experiments that generated the training data, or understand why the results mattered for drug discovery. Human scientists did all of those things.

The scientists who will thrive in the next twenty years are the ones who develop fluency with AI tools and use them to ask questions that couldn’t be asked before, rather than continuing to ask the same questions in the same ways.

On AGI — you’ve been more specific than most about what you mean by it. What do you mean?

I mean a system that can learn any cognitive task that a human can learn, with human-like or better efficiency. Not just perform tasks it was trained to perform. Not just generalise within a domain. Genuinely learn new domains from limited data, form new hypotheses, conduct experiments to test them, update its beliefs based on results. The full cycle of scientific reasoning, applied to arbitrary domains.

I think that’s achievable. I think it’s probably closer than most people outside the field believe, and I think the timeline uncertainty is still substantial enough that anyone who gives you a specific year should be treated with scepticism. Including me.