AlphaGo

Go was the last board-game holdout: the search space is far too large for brute force. The solution was two neural networks — a policy network proposing moves and a value network judging positions — steering a Monte Carlo tree search. First trained on human games, then improved through self-play.

Move 37 in game two was so unconventional that commentators assumed it was a mistake; it turned out to decide the game. Successor AlphaGo Zero (2017) skipped human games entirely and was stronger anyway.