Move 37
In one sentence
Move 37 is the Dictionary’s name for a machine-generated decision that violates expert intuition, survives validation, and expands what practitioners consider possible.
The origin
On 10 March 2016, during Game 2 of AlphaGo’s match with Lee Sedol, AlphaGo played its thirty-seventh move on the fifth line. Commentators initially struggled to explain it. DeepMind later reported that AlphaGo assigned only a one-in-ten-thousand probability that a human would choose the move. It became pivotal to AlphaGo’s victory and a durable symbol of machine-generated surprise.
The surprise did not come from a system untouched by human play. The 2016 AlphaGo first learned from expert human games and then improved through reinforcement learning from self-play. Move 37 is therefore not evidence that freedom from human data automatically produces originality. It shows that a system can begin with human examples, search beyond familiar heuristics, and return a move that experts can test on the board.
The test
Not every odd output is a Move 37. The term should be reserved for a result that:
- conflicts with established expert intuition;
- can be evaluated against an external standard;
- proves useful or correct; and
- changes subsequent expert practice or understanding.
The validation requirement is the load-bearing part. Without it, “Move 37” becomes a flattering name for a hallucination.
The analogy may apply beyond Go—to a protein structure, an algorithm, a design, or an operating decision—but each field needs its own evidence. AlphaFold and AlphaTensor are related examples of machine-assisted discovery; neither justifies a generic claim that surprising AI outputs are correct or worth billions.
Game 4 supplies the necessary symmetry. Lee Sedol’s Move 78 surprised AlphaGo and contributed to the human’s only victory in the five-game match. Human and machine judgment can each expose the other’s blind spots.
Sources
- DeepMind, AlphaGo.
- Silver et al., Mastering the game of Go with deep neural networks and tree search, Nature, 2016.
See also
Lee Sedol · Root Node Problems · Capability Overhang · Single-Arrow Fallacy