Move 37
In one sentence
Move 37 is a decision or intervention that conventional expertise would classify as an error, but that an agent operating beyond human-learned heuristics identifies as correct — and that, once observed, permanently reshapes what practitioners in that field consider possible.
The origin
March 10, 2016. Game 2 of the AlphaGo–Lee Sedol match. AlphaGo places a stone on the fifth line of the board, early in the game. Go masters watching the live feed call it a mistake. Fan Hui, the European Go champion acting as commentator, leaves the room to think about it. The move violates a principle beginners are taught and professionals carry without examining: stones on the fifth line, early, are exposed and weak. Except this stone was neither. One hundred moves later it was in exactly the right place to decide the game. AlphaGo won.
Demis Hassabis, who had been doing arithmetic on his phone during the game, described it as “the moment I’d been waiting for” — not because AlphaGo won, but because of how it won. The move was not found by searching through human Go games. AlphaGo had discovered something no human had considered, and then been right about it. That was new.
What it names now
Move 37 has left the Go board. It has entered the vocabulary of AI researchers, practitioners, and — increasingly — managers and strategists as a shorthand for any moment when an AI system produces a solution that:
- contradicts expert intuition
- turns out to be demonstrably correct
- permanently shifts the boundary of what practitioners believe is possible
The protein folding moment — when AlphaFold predicted the 3D structure of a protein with near-experimental accuracy and the team realised they could fold every known protein in a year — had the same shape. The move was obvious after, absurd before. So did AlphaTensor’s discovery of a faster algorithm for matrix multiplication: five percent faster, which at scale represents billions of dollars of compute savings, found by treating the problem as a game and exploring moves no human had tried.
The term matters because it names a failure mode in human judgment: we call things mistakes because they violate our heuristics, and our heuristics were built from human data. An agent not trained on human data has no reason to inherit those heuristics. The fifth line is fine if you are AlphaGo. The question is whether, in any given domain, your intuition is tracking something real or something merely traditional.
Why it matters in a management context
Move 37 moments are arriving outside Go. They are arriving in drug discovery, chip design, materials science, logistics, and — less visibly — in the kinds of strategic decisions that management consultants and analysts have spent careers learning to make. The practitioner who cannot recognise a Move 37 when it appears will dismiss it. The practitioner who can will have a material edge.
This is not a call to defer to AI systems uncritically. Lee Sedol won Game 4 of that same match with a move — Move 78 — that AlphaGo could not handle, a counterintuitive probe that broke the machine’s evaluation function. The lesson is symmetric: humans and machines have different blind spots, and the gap between them is the productive space. A practitioner who understands where their own heuristics are load-bearing and where they are merely inherited is better placed to recognise which kind of move they are looking at.
See also
Root Node Problems · Capability Overhang · Single-Arrow Fallacy
Proposed May 9, 2026. Source: Demis Hassabis interview, Huge Conversations / Cleo Abram, May 2026; AlphaGo vs. Lee Sedol, Game 2, March 10, 2016.