AlphaGoAlphaGo

In March 2016, DeepMind's AlphaGo faced Lee Sedol, one of the greatest Go players of his generation, in a five-game match in Seoul watched by an estimated 200 million people. Go had long been considered the grand challenge of game AI: its branching factor is so vast that the brute-force search that conquered chess was hopeless.

In game two, on move 37, AlphaGo placed a stone on the fifth line in a 'shoulder hit', a play that professional commentators initially assumed was a mistake. AlphaGo's own policy network estimated that a human professional would choose that move roughly one time in ten thousand. Lee Sedol left the room to compose himself; the move went on to shape the whole game, which AlphaGo won.

The system combined deep neural networks with Monte Carlo tree search. A policy network, trained first on human expert games and then refined by self-play reinforcement learning, proposed promising moves, while a value network judged positions without playing them out. The search tied both together, exploring only the sliver of the game tree the networks considered worth exploring.

AlphaGo had already beaten European champion Fan Hui 5-0 in October 2015, a result published in Nature in January 2016 by David Silver, Aja Huang, and colleagues. But Lee Sedol was expected to expose it. Instead AlphaGo won 4-1, with Lee's single victory in game four, sparked by his brilliant move 78, celebrated as a triumph of its own.

The 2017 successor, AlphaGo Zero, discarded human data entirely and learned from scratch through pure self-play, surpassing the version that beat Lee within days of training. The message was uncomfortable and profound: in some domains, human knowledge is not just unnecessary for machines, it can be a ceiling.

Move 37 became shorthand for machine creativity, the moment a learned system produced something valuable that lay outside accumulated human practice. Lee Sedol retired from professional play in 2019, saying that AI had shown it could not be surpassed.