AlphaGo beats Lee Sedol

Go has about 10¹⁷⁰ legal positions and no evaluation function anyone could write down, and after Deep Blue it was the game that stood for what search could not do. Experts in 2015 guessed a decade before a machine would beat a professional. On 9 March 2016, in Seoul, DeepMind's AlphaGo won the first of five games against Lee Sedol, holder of eighteen international titles, and went on to win the match four to one. Two hundred million people watched.

AlphaGo combined three things. A policy network, trained first on thirty million human moves and then by playing itself, suggested moves. A value network, trained on the outcomes of self-play games, judged positions. Monte Carlo tree search used both to look ahead, sampling lines rather than enumerating them. The learning was TD-Gammon's, the networks were AlexNet's descendants, and the compute was a cluster of Google's.

In game two AlphaGo played its 37th move on the fifth line, a shoulder hit that no strong human would have considered and that commentators called a mistake until it won the game. Lee's own move 78 in game four, the one win, was called divine. The match was the moment the public understood that machines could be creative at something, and the moment Go players started learning from them.