Go
The rules fit in a paragraph, and they held out against computers for nineteen years after chess fell.

The board begins empty. Nineteen lines cross nineteen lines, which gives three hundred and sixty-one points where they meet, and the first player sets a black stone on one of them. That stone will never move again. White answers somewhere else, usually far away, and the two go on filling the grid with stones that stay where they are put, each trying to finish owning more of the board than the other. Two ideas laid over that arrangement do almost all of the work.
The first is capture by suffocation. The empty points directly beside a stone, up and down and across but never diagonally, are its liberties: four in the open middle, three on the edge, two in the corner. Fill the last with your own stones and it comes off the board and into your bowl as a prisoner. Stones side by side along a line are joined for good, sharing liberties and dying together, which is why players talk about walls rather than pieces.
The second is that capture almost never decides anything. Once neither side can find a profitable move, both pass, and two passes end the game. Each counts the vacant points enclosed by their own stones, adds the prisoners taken, and the larger number wins. Games between strong players turn on two or three points out of three hundred and sixty-one.
What separates a person who has read the rules from a person worth talking to is life and death. A group hemmed in by enemy stones survives forever if it encloses two separate empty points, called eyes, since filling either would be a move with no liberty of its own, which the rules forbid; the attacker may take one eye or the other and never both. One eye is a corpse waiting. Nearly every fight on a Go board is an argument about a second eye.
Two smaller rules finish the set. A capture that could be answered by an immediate recapture, restoring the position exactly, has to wait a turn, which keeps play out of circles; the Japanese name for that shape is ko, eternity. And since moving first is reckoned worth about seven points, White is handed that many at the start, a compensation called komi. Go also solved a problem chess never has: strength is counted in stones, so the weaker player begins with up to nine already on the board, and what follows is a real contest.
The game is Chinese and older than any date worth defending; Confucius and his contemporaries already knew it, as weiqi, the surrounding game. The grid took time to settle, since the earliest boards ran seventeen lines a side, and a white-glazed ceramic board made in 595 and lifted from Zhang Sheng's tomb at Anyang in 1959 shows nineteen already standard under the Sui. The oldest surviving book on how to play is a paper scroll two and a half metres long, sealed into the library cave at Dunhuang and carried off by Aurel Stein in 1907; its last line gives the title as the Classic of Go.

For most of its life the game carried more weight than a game usually does. Chinese scholars counted qi, meaning Go, among the four accomplishments expected of a cultivated man, alongside the seven-string zither, calligraphy, and painting. In 1612 the shogun put Japan's four strongest players on hereditary stipends, and the houses that grew out of them spent two and a half centuries trying to beat one another, which is where most of the theory came from; HoninbÅ, the strongest of them, produced more holders of the title Meijin than the other three combined. In Korea the game is baduk, and regular players are reckoned at five to ten percent of the country.
Deep Blue took chess from Kasparov in 1997. Go held out, and the reason is arithmetic. A chess player has roughly thirty-five reasonable moves at a turn across a game some eighty moves deep; a Go player has roughly two hundred and fifty, across about a hundred and fifty. In January 2016 a distributed computation finished counting the legal positions on a full board, and the answer runs to a hundred and seventy-one digits, past the number of atoms anyone credits to the observable universe. The deeper trouble was that nobody had written down what makes a Go position good. Strong players call it feeling and leave it there.
The way past that, published in Nature that same month, was to let a machine learn the feeling instead. David Silver, Aja Huang, Demis Hassabis and their colleagues at DeepMind trained two neural networks, one to propose a strong player's move and one to guess who was winning, first on a database of human games and then on millions the program played against itself. Their program, AlphaGo, had already taken five games from the European champion Fan Hui without losing one. In March 2016 it sat down in Seoul against Lee Sedol, holder of eighteen world titles, with two hundred million people watching, and won four games to one.
Move thirty-seven of the second game is the one people still talk about. Convention put a stone in that fight on the third or fourth line, where territory and influence trade fairly, and AlphaGo played a shoulder hit on the fifth, a choice its own model put at one chance in ten thousand of coming from a human hand. Dozens of moves later it was holding the whole right side. Lee's reply came two games on, at move seventy-eight of the fourth, a wedge just as improbable, and it won him the only game AlphaGo has ever lost in a formal match; the move is known as God's Touch. Lee said afterwards that he had taken the machine for a probability engine, and that this had changed his mind: "Surely, AlphaGo is creative."
The full film, free from DeepMind. Watch the room in Seoul rather than the board: the commentators reaching for a rational explanation of move thirty-seven, and Lee Sedol going out for a cigarette while the machine holds a position nobody in the building can read.
AlphaGo beat the world number one, Ke Jie, three to nothing at Wuzhen in May 2017 and retired. Lee retired in November 2019, telling a Korean reporter that first place had stopped meaning anything, since "there is an entity that cannot be defeated." The game came out of it better. Four researchers later ran nearly six million professional moves from 1950 to 2021 past a superhuman engine and found human decisions measurably stronger after 2016, with professionals reaching more often for moves nobody had played before. To start, use a nine-line board, which teaches liberties and eyes in half an hour. The nineteen lines will wait.