DeepMind Researcher Revisits AlphaGo's Move 37 to Argue LLMs Still Don't Reason

A researcher who worked on the AlphaGo program has pushed back on claims that large language models reason, using the system's most famous move as the counterexample. In an account published by MIT Technology Review, the author recalls sitting in Seoul in March 2016 as the program placed a stone on the fifth line of the Go board during game two of its five-match series against Lee Sedol. Commentators initially read Move 37 as a blunder, then watched it turn the game.

The detail that matters is not the move itself but how it was produced. AlphaGo arrived at that placement through Monte Carlo tree search, evaluating positions against a value network trained on millions of games and refined through self-play. The program explored branches, scored outcomes and selected the line with the best expected result. Its output looked strange because the underlying calculation was deep, not because the system was improvising.

That distinction is the core of the argument against treating today's chatbots as reasoning engines. A large language model generates a token by predicting what plausibly follows from the tokens before it, drawing on statistical patterns in its training corpus. It can produce a chain of steps that resembles deliberation, but the steps are not verified against a world model or a search tree before the next one is emitted. Fluency and correctness come from the same mechanism, which is why confident errors survive intact.

The timing of the essay is pointed. Labs are shipping models that write proofs, debug code and plan multi-step tasks, and vendors increasingly describe these systems as reasoning models. Benchmarks that reward final answers can be passed by pattern matching on familiar problem shapes, so a high score does not prove the model built and tested an internal argument. Researchers have shown that changing names, numbers or surface wording in a problem can flip a model's answer, a fragility that a search-based system would not show.

None of this means language models are useless. They compress enormous amounts of text into a tool that can draft, translate, summarize and retrieve, and they do it at a scale no hand-written rule set could match. The disagreement is about the label and the expectations attached to it. If reasoning means searching over possibilities and checking results, then current systems mostly approximate the output of reasoning without running the process.

What to watch is whether the next generation of systems adds explicit search, tool use and verification loops on top of language generation, or whether scaling alone closes the gap. AlphaGo needed both a policy network to propose moves and a search procedure to judge them. The open question is whether language models can reach the same reliability without that second component, and how the industry will describe them in the meantime.