Why AlphaGo’s famous move was reasoning, not intuition
In March 2016, AlphaGo played an unexpected stone on the fifth line during its second match game against Lee Sedol. The move, later known as move 37, appeared so strange that some observers suspected an error. AlphaGo won that game and the five-game match 4-1. Lee described the move as evidence of creativity, but Thore Graepel, a former core member of the AlphaGo team, argues that it reflected deliberate search rather than machine intuition.
AlphaGo combined a policy network, which estimated moves a strong human might choose, with search software that explored thousands of possible continuations in a game tree. The network supplied promising options, while the search assessed their longer-term consequences. Graepel compares this arrangement with fast intuition and slower deliberation in human thought.
He says current large language models mainly generate one token after another. Chain-of-thought systems can produce intermediate steps, but those steps still come from the same prediction process and may be constructed after an answer is reached. Graepel argues that reliable AI needs an explicit, inspectable record of beliefs, evidence, uncertainty and revisions. He proposes systems that test possible actions, use tools, update beliefs only when evidence supports them, and produce conclusions through an auditable sequence of inference. He says this approach could be important in medicine, science, engineering, drug discovery and climate research.
Written by Sakhanda Wire AI from reporting by MIT Tech Review. Facts, quotes and full context are in the original article below.