AlphaGo's Move 37 was search, not intuition, says Thore Graepel

The popular telling of AlphaGo's 2016 match against Lee Sedol often reduces Move 37 to a moment of pure machine intuition — a flash of insight so strange that some commentators initially suspected a bug. Graepel, who worked on the program, draws a different line: AlphaGo combined a policy network trained to guess human-like moves with a search mechanism that built a game tree of thousands of branches, letting it compare how proposed moves would play out. That pairing mirrors the distinction Daniel Kahneman popularized between fast, intuitive "System 1" thinking and slow, deliberative "System 2" thinking — with the search providing the deliberation.

The contrast Graepel draws is with today's LLMs, which keep predicting the next token rather than exploring alternatives. That difference matters most where the stakes are high. In medicine and science, he argues, it is not enough for a system to land on an answer; it must be possible to check how it got there. Whether that kind of verifiable reasoning can be built at scale is likely the central question for applying AI in those domains, and Graepel suggests recent progress may be making it more feasible.