Ten years ago in Seoul, a DeepMind system placed its 19th stone on an unusual point along the fifth line of a Go board during Game 2 of a five-game match against world champion Lee Sedol. Professional commentators assumed it was a malfunction. It was not. About 100 moves later, that stone sat in exactly the right position for AlphaGo to win the game.
Move 37 was not a lucky guess. AlphaGo combined deep neural networks with Monte Carlo tree search, trained first on recorded games from human experts, then through hundreds of thousands of self-play rounds where the strongest winning strategies were reinforced. The policy network assigned probabilities to candidate moves; the value network estimated win probability from any given board position; the search algorithm then explored only the most promising branches of a decision tree covering 10 to the power of 170 possible board states—more positions than atoms in the observable universe.
The move's statistical rarity was the point. Reinforcement learning across that search space elevated a strategy humans had never played because human experts never had the compute time to discover it. Go professionals described the stone as creative relative to any prior high-level game. More than 200 million people watched the match live. The psychological blow to elite human judgment was as significant as the technical achievement: a machine had generated an idea that surpassed expert intuition, not merely automated it.
That claim has since been tested. The same reinforcement-learning architecture that found Move 37 was later aimed at protein folding, a problem biologists had worked on for 50 years. In 2020, AlphaFold 2 predicted the three-dimensional structure of proteins with enough accuracy to be declared a solution to the field's grand challenge. DeepMind then folded the structures of all 200 million proteins known to science and released the results in a free database. Today more than three million researchers use it, working on projects from malaria vaccines to plastic-degrading enzymes. In 2024, Demis Hassabis and John Jumper were awarded the Nobel Prize in Chemistry for leading the project.
The progression from AlphaGo to AlphaFold followed a clear internal logic. AlphaGo Zero, built after the original system, learned Go from random play alone—no human expert games, no prior knowledge—and became the strongest player in history. AlphaZero generalized further, teaching itself chess, Go and shogi from scratch. It mastered chess in hours and beat Stockfish, the dominant specialized chess engine, while producing opening and middlegame strategies that decades of computer-assisted analysis had never surfaced.
DeepMind's latest system in this lineage is AlphaEvolve, a coding agent. Hassabis said it had its own Move 37 moment: AlphaEvolve discovered a novel method for matrix multiplication, a mathematical operation that underlies nearly every modern neural network. Matrix multiplication sits at the core of transformer inference, which means any efficiency gain there compounds across every AI workload running in production today. DeepMind has tested AlphaEvolve on data center optimization alongside the mathematical work.
AlphaProof, the most direct architectural descendant of AlphaGo, learned to prove formal mathematical statements through a combination of reinforcement learning and symbolic reasoning. The system tackles problems that have historically required human mathematical creativity, not pattern matching. Where AlphaGo found Move 37 by searching a game tree, AlphaProof searches a proof tree.
The commercial stakes behind these research milestones are real. Nvidia's GPU business, which supplies the hardware running virtually all frontier AI training and inference, trades at $223.96 today, up 2.3 percent on the session. The broader Nasdaq is up 1.3 percent. The infrastructure spending that AlphaGo's 2016 win intellectually justified now runs to tens of billions of dollars annually across Alphabet, Microsoft, Meta and Amazon. Each of those companies is betting that the AlphaGo lesson—that AI trained at scale will find strategies humans would never discover—applies to their core products.
The business model question the anniversary raises is whether Move 37-style discoveries translate into durable competitive moats. AlphaFold's database is open source, which means the research breakthrough generated scientific value but not proprietary revenue for DeepMind directly. AlphaEvolve's matrix multiplication finding, if it materially reduces compute cost per inference token, is a different kind of asset—one that sits inside Google's infrastructure and does not get released to competitors.
The distinction matters at the P&L level. Alphabet earns roughly 60 percent of its revenue from advertising that runs on AI-ranked feeds and search results. Every efficiency gain in transformer inference reduces the marginal cost of serving those results. A novel matrix multiplication method that cuts compute by even a few percent across Alphabet's data centers is worth more to the company's operating margin than most product launches. Alphabet trades at $354.30 today, the only major tech name in the red, down one percent on the session.
Lee Sedol retired from professional Go in 2019, citing AlphaGo as a reason: there was no longer a meaningful human benchmark at the top of the game. The same dynamic is appearing in mathematical research, where AI systems are beginning to propose proof strategies that researchers then verify rather than originate. The direction of labor—human checks machine, rather than machine assists human—is the real consequence of what Move 37 started.
