Demis Hassabis on what AlphaGo's Move 37 set in motion, ten years on

Ten years after AlphaGo beat Lee Sedol, Demis Hassabis revisits why that match still matters, and the argument is more specific than nostalgia. The point of Move 37 was not that a machine won Go but that it found a strategy no strong human would have played, evidence that a system could discover approaches outside human expertise rather than just imitate them.

He draws a straight line from there to concrete science. AlphaFold 2 solved protein structure prediction in 2020, the database is now used by more than 3 million researchers on work from malaria vaccines to plastic-eating enzymes, and it earned a Nobel Prize in Chemistry. AlphaProof and AlphaGeometry 2 became the first system to reach medal standard, a silver, at the International Mathematical Olympiad. AlphaEvolve found a new way to multiply matrices, the operation underneath nearly every neural network, which he frames as a second Move 37. The thread is consistent: the same search-plus-learning recipe, pointed at problems where the answer can be checked.

Why it matters

If you plan research bets, the useful read is the pattern, not the trophies: these wins clustered in domains with a clear verifier, which is a practical filter for where today's models are likely to produce real discoveries rather than plausible guesses.