postgrok

thore-graepel

Properties1
tagsperson, researcher, alphago, deepmind, reasoning

2 min read

Thore Graepel

Machine learning researcher, alpha-go team member, and author of graepel-2026-llms-dont-reason. His position in that article, that today’s LLMs reason no better than an unassisted intuition, is stated from inside the tradition that built the counterexample.

Two facts anchor the position. He watched move 37 land in Seoul in March 2016 as a member of the team that built the program, and he says he left Google DeepMind over the disagreement with the current frontier industry.

Roles

  • Chair of Machine Learning, University College London, Department of Computer Science. The UCL profile page reads “Chair of Machine Learning.” Wikipedia calls the role part-time. The article’s byline says chair without qualification.
  • DeepMind 2015 to 2021, on alpha-go, AlphaGo Zero, AlphaZero, and MuZero. His personal site names David Silver as the leader of that work.
  • Google DeepMind 2025 to 2026, Post-AGI team, per his personal site. Wikipedia describes the role as research group lead working on AGI futures.
  • Microsoft Research 2003 to 2015. Co-developer of the TrueSkill ranking system and the AdPredictor ad-ranking work, per Wikipedia.
  • Altos Labs 2021 to 2025, machine learning for cellular rejuvenation, per his personal site.
  • Physics to a machine learning PhD, Berlin, per his personal site. Wikipedia calls him an Anglo-German physicist and computer scientist.

After DeepMind

The article says the departure happened recently. His personal site states the plan behind it: “building machines that reason and act,” a new project bringing AlphaGo-style reasoning, a policy that suggests, a value that weighs, and a tree that looks ahead, to frontier AI. The article names no company or lab. So the program he argues for in the essay is the program he went to build. That is context for the claim, not a refutation of it.

Sources

Graph View

Backlinks