AlphaGo Veteran Questions LLM Reasoning, Sparking Researcher Debate
Thore Graepel, a former Google DeepMind researcher and AlphaGo team member, published an op-ed in MIT Technology Review on the 10th anniversary of AlphaGo's 4:1 victory over Lee Sedol, questioning whether today's LLMs can truly reason. Researchers including David Duvenaud (Toronto) and Max Tegmark (MIT) responded publicly, sparking a debate over LLM reasoning.
Confirmed
- Graepel's core argument: Move 37 was made possible by explicit search; what modern LLMs lack is not longer chains of thought but auditable reasoning—generating longer chains does not equal genuine reasoning.
- His listed deficiencies: no explicit, persistent, inspectable cognitive state; knowledge not separated from how it is operated on; post-hoc fabricated explanations.
- In follow-up discussion with Duvenaud, Graepel clarified: humans alone are weak reasoners, improve markedly with pen and paper, and become excellent reasoners when strictly following structured processes such as the scientific method—the key is structured process, not any fixed form.
- Duvenaud countered that these objections apply equally to humans, who also lack inspectable cognitive states, so one cannot conclude LLMs cannot reason.
- Tegmark responded to Duvenaud, arguing that the bare human brain likewise does not reason in that sense.
Why it matters
- Timed at the AlphaGo anniversary, the debate grounds the question of LLM reasoning in concrete technical criteria: auditability, separation of knowledge and operations, and whether search is necessary.
- With prominent researchers from DeepMind, Toronto and MIT involved, the exchange reflects a fundamental schism over the long-chain-of-thought reasoning paradigm, directly relevant to interpretability and AI safety discussions.
2026-10-02 ~ 2026-10-03 · 6 related posts
Primary sources
- DeepMind vet Graepel: ten years after AlphaGo, LLMs still don't reason — nordicinst · 2026-10-02
- [source] Ex-DeepMind researcher argues LLMs still lack auditable reasoning, calls for AlphaGo-style rethink — ThoreG · 2026-10-02
- [source] David Duvenaud rebuts "LLMs don't reason": all the stated reasons apply to humans too — DavidDuvenaud · 2026-10-03
- [source] Max Tegmark to Duvenaud: Humans Alone Don't Reason Well Either — ThoreG · 2026-10-03
- Duvenaud and Thore debate: structured processes, not raw minds, make true reasoners — DavidDuvenaud · 2026-10-03
1 near-duplicate retellings: ThoreG