AlphaGo beat Lee Sedol. Four games to one. The Go world’s confidence lasted exactly zero games.
But the score isn’t the story. The story is Move 37, game two: AlphaGo played a shoulder hit on the fifth line that every professional commentator initially called a mistake, a move with a 1-in-10,000 chance of being played by a human, per AlphaGo’s own model. It wasn’t a mistake. It was better than what humans knew, and it reorganized the whole board fifty moves later. Lee Sedol left the room. Three thousand years of accumulated human theory, and the machine found a gap in it on live television.
And then, this matters just as much, game four: Lee played move 78, the “divine move,” a wedge AlphaGo’s model rated equally improbable, and the machine collapsed into confusion for thirty moves. The human found the gap in the machine. One each.
I watched at 2am with the Dota group chat, all of us aware we were watching our own professional futures negotiate terms. My take, for the record, 2016: the tools that beat us at intuition become the tools; the players who study with AlphaGo will be better than any player before them. Centaurs over engines. Ask me again in ten years.
TIL: policy networks vs value networks, one suggests moves, one judges positions. Proposal and review. Even the machine pair-programs.