Move 37
In the [YOU] on AI Field Guide
Move 37 functions in the cycle as the empirical hinge between the two easy stories about machine creativity. The first easy story is that machines only do what they are told and therefore cannot be creative in any meaningful sense. The second is that machines have become general creative agents and human creativity is obsolete. Move 37 disproves the first story without establishing the second. The engineers at DeepMind did not tell the system to play the fifth line on move thirty-seven. The system played it because its self-taught evaluation of the position reached a conclusion no human intuition had reached. That is a real and remarkable thing, and the cycle’s insistence on honest accounting requires acknowledging it.
The episode also carries a precise lesson about the right kind of surprise. The move was not surprising because it was random but because it was deep—it lay beyond the horizon of human understanding while remaining inside the logic of the game. Du Sautoy uses this distinction to define the kind of surprise that earns the name creative, as distinct from the cheap surprise of noise or error. A system that played the fifth line because of a glitch would also have surprised everyone, and we would have called it broken. AlphaGo surprised everyone because it saw further, and that difference is the whole difference.
The limit du Sautoy draws around the episode is equally important for the cycle. Move 37 is exploratory creativity within a closed and well-defined game. The harder forms of human creativity unfold in open domains where the rules are negotiable and the goal is not given. Whether a machine can be creative there—in mathematics, in art, in the construction of new frameworks of thought—is a question Move 37 does not settle. It only proves that the first door has opened. The second door, the door to transformation, remains closed and may require something the machines do not yet possess.
Origin
AlphaGo was developed by DeepMind and trained using a combination of supervised learning from human expert games and reinforcement learning through self-play. The self-play phase produced the evaluation function that generated Move 37: the system played against itself millions of times, updating its assessments of board positions until it developed intuitions that no human trainer had explicitly encoded. The resulting system defeated Lee Sedol four games to one in a five-game match watched by an estimated two hundred million people.
The match was widely covered as a milestone in artificial intelligence, but du Sautoy’s contribution was to identify precisely what milestone it represented. It was not evidence that machines are generally intelligent or conscious. It was evidence that a machine, under the right conditions, can perform an act of genuine exploratory creativity within a well-defined domain. That is a smaller claim than the popular coverage suggested and a more important one: it is precise, verifiable, and grounded in an actual move that actual experts were unable to anticipate.
Key Ideas
Passes the Lovelace test at the exploratory level. The move was new (no record of a human master playing it in that situation exists), surprising (professional commentators could not account for it), and of value (it contributed to a historic win). And it was not merely an expression of any programmer’s intent. By the Lovelace test’s four conditions, Move 37 holds up as well as any machine-generated act in the historical record.
Exploratory, not transformational. Du Sautoy is careful not to overclaim. The move does not change the rules of Go; it finds a brilliant position within them. The space of legal Go positions is larger than the number of atoms in the observable universe, and the machine’s achievement was to search that space with an intuition that exceeded the human one. This is a staggering feat of exploration. It is not a redrawing of what the game is. The taxonomy of creativities keeps this distinction visible.
The right kind of surprise. Move 37 surprised because it was deep, not because it was random. This distinction between depth-surprise and noise-surprise is the difference between creativity and malfunction, and du Sautoy uses the episode to define it precisely. A system trained to maximize performance within the rules of Go can achieve depth-surprise through the quality of its evaluation function; no amount of random output can.
Debates & Critiques
The move provoked lasting debate about what it demonstrated and what it did not. Optimists argued that a system capable of Move 37 had crossed into genuine intelligence; du Sautoy resisted this, insisting that excellence in a closed game is evidence of excellence in that game and no more. The case of Go is unusually favorable for machine exploration: the rules are perfectly specified, the evaluation function can be defined unambiguously (win or lose), and the training signal is clear. Open-ended creative domains have none of these features. A second debate concerns whether the self-play process that generated Move 37 is in principle extensible to open domains—whether a system that plays mathematics or music or fiction against itself could develop analogous intuitions that exceed any human training data. Du Sautoy does not rule this out but notes that the evaluation problem in open domains is the hard problem: knowing which move is brilliant requires a sense of value that, in open domains, is not given by the rules.
Further Reading
- Marcus du Sautoy, The Creativity Code: Art and Innovation in the Age of AI (Fourth Estate, 2019), Chapter 4
- David Silver et al., “Mastering the Game of Go with Deep Neural Networks and Tree Search,” Nature 529 (2016)
- David Silver et al., “Mastering the Game of Go without Human Knowledge,” Nature 550 (2017)
- Cade Metz, Genius Makers: The Mavericks Who Brought AI to Google, Facebook, and the World (Dutton, 2021)