DeepMind's AlphaEvolve Delivers Its Own 'Move 37' Moment

A decade after AlphaGo's now-famous Move 37 rattled the world's best Go players, Google DeepMind CEO Demis Hassabis is arguing the same jarring, 'how did a machine come up with that?' moment is showing up in serious math. That is the thread of a recent WSJ column by Ben Cohen, which frames a novel matrix multiplication result from DeepMind's coding agent AlphaEvolve as, in Hassabis's words, its own 'Move 37 moment.'

The original Move 37 was the second-round placement AlphaGo made against Lee Sedol in Seoul in March 2016 that DeepMind calculated a human player would produce one time in 10,000. The move was so unconventional and counterintuitive that commentators initially read it as a blunder, then as a masterstroke. Lee, who lost the match 4-1, later said it made him think about Go in a new light. Cohen uses that as a reference point: the striking thing was not that a computer beat a champion, it was that the computer produced an idea no champion would have.

Why this matters beyond Go: DeepMind has long used games as a training ground because they hand back fast, measurable feedback. The claim in the column is that the same pattern is now showing up in domains with real economic weight. When AlphaEvolve, a Gemini-powered coding agent, produced a new matrix multiplication method (the arithmetic operation underlying virtually all modern neural networks), it was doing what AlphaGo did, in a corner of computing where a small improvement compounds across every AI system in production.

The honest caveats: this is an opinion column, not a research write-up, so it does not benchmark AlphaEvolve's math against independent verifiers, name specific mathematicians who inspected the result, or quantify the speedup. It is Hassabis's framing, retold by a columnist, and the 'Move 37' label is doing a lot of rhetorical work. Take the specifics as reported, not settled.

The reason to pay attention anyway is the incentive structure. If frontier labs really are getting genuinely novel outputs from their coding agents, the buyers who benefit first are the ones running deep, structured search: algorithm design teams, chip and compiler shops, and cloud providers trying to squeeze inference cost. Everyone else should at least know the argument is being made.