AlphaGo defeats Lee Sedol 4–1 in Seoul

Background

Go is a board game for two players who take turns placing black and white stones on a grid and try to claim more territory than the opponent. For a long time it was treated as one of the harder targets for artificial intelligence, because the number of possible positions is enormous and strong play is difficult to reduce to a short list of rules.

AlphaGo was built by DeepMind. Before the Seoul match it had played the European champion Fan Hui and won 5–0 in October 2015, and the system was first described in public in a Nature paper in January 2016 ^[2].

How it worked

AlphaGo combined deep neural networks with tree search ^[2]. In outline, a neural network is a layered model that takes a position and produces a judgement about it, or a set of plausible moves; tree search treats those moves as branches and looks ahead through the continuations that follow them. Working together, the network narrows down what has to be considered and the search works out which of the candidate moves leads somewhere useful.

The match

The opponent was Lee Sedol, a 9-dan professional with 18 world titles ^[3]. The five games were played on 9, 10, 12, 13 and 15 March 2016 at the Four Seasons Hotel in Seoul ^[4]. AlphaGo won games 1, 2, 3 and 5, and Lee won game 4, for a final score of 4–1 ^[3][4]. Nature reported the result on 15 March 2016 ^[3].

Two moves

In game 2 AlphaGo played a move that became known as "Move 37". The system's own data estimated that a human player would choose that move with a probability of 1 in 10,000 ^[1]. In game 4 it was Lee's turn to surprise the machine: his move 78 was also estimated at 1 in 10,000, and it led to his win ^[1] — his only victory in the five games.

Audience

Interest was enormous. According to Demis Hassabis, tens of millions of people in China watched live streams of the match, and the match hashtag reached 200 million pageviews on Sina Weibo ^[1]. DeepMind says the match was watched by over 200 million people worldwide ^[2].

Prize and rank

The prize was $1 million. DeepMind said it would donate the money to organisations supporting STEM education and Go, and to UNICEF ^[4]. After the match, AlphaGo received an honorary 9 dan from the Korea Baduk Association — the first time a computer Go player had received the highest possible certification ^[1][4].

Aftermath

DeepMind later built successors including AlphaZero and MuZero on AlphaGo's approach ^[2]. Lee Sedol announced his retirement from professional Go on 19 November 2019, citing the dominance of AI ^[2][4].

Why it mattered

Board games have long served as a testing ground for AI, because the rules are fixed and the outcome is unambiguous. The Seoul match was a public demonstration of how far the approach behind AlphaGo could go: deep neural networks combined with tree search ^[2]. That combination was not specific to Go, and the systems DeepMind built afterwards reused it ^[2]. The match also reached well beyond research circles: by DeepMind's count, more than 200 million people worldwide watched it ^[2].