Move 37

Move 37 refers to a specific play made by DeepMind's AlphaGo during the second game of its historic five-game match against world Go champion Lee Sedol in March 2016. On the 37th turn, AlphaGo placed a stone on the fifth row near the board's edge — a position so unusual that professional commentators initially assumed it was a mistake. Human Go tradition, built over thousands of years, strongly disfavored such placements at that stage of the game. Yet the move proved to be a stroke of profound strategic genius, ultimately contributing to AlphaGo's victory in that game and the match overall.

What made Move 37 technically remarkable was that it emerged from AlphaGo's combination of deep convolutional neural networks and Monte Carlo tree search, trained through both supervised learning on human expert games and self-play reinforcement learning. AlphaGo's policy network estimated the move had roughly a one-in-ten-thousand probability of being chosen by a human player — meaning the system had genuinely departed from the distribution of known human strategies. Rather than interpolating between human moves, AlphaGo had discovered a novel strategic concept through its own simulated experience, one that humans had simply never converged on.

The broader significance of Move 37 extends well beyond the game of Go. It served as a vivid, public demonstration that modern AI systems trained with deep reinforcement learning are not merely pattern-matching engines that mimic human behavior — they can explore strategy spaces in ways that transcend their training data and surface genuinely new knowledge. This challenged a common assumption that AI would always be bounded by the ceiling of human expertise it was trained on.

Move 37 has since become a cultural and technical touchstone in AI discourse, frequently cited when discussing creativity, superhuman performance, and the limits of human intuition in complex domains. It accelerated interest in applying similar reinforcement learning techniques to other high-dimensional decision-making problems, from protein folding to chip design, cementing its place as one of the most consequential single actions in AI history.