Thore Graepel, chair of machine learning at University College London and formerly a core member of DeepMind's AlphaGo team, argued in MIT Technology Review on 2 October 2026 that today's language models do not reason. His reference point is AlphaGo's Move 37 — played out of explicit search, not intuition.
Key takeaways
- Published 2 October 2026 in MIT Technology Review
- Graepel left Google DeepMind to pursue a different route to machine reasoning
- Three charges: no persistent epistemic state, knowledge fused with reasoning in the weights, post-hoc rationalisation
- Move 37, in game two against Lee Sedol, had roughly a 1 in 10,000 chance of being played by a human expert
- The proposal: reasoning as operations that change an explicit epistemic state
Move 37 was not a hunch
The split into two modes is the crux. AlphaGo's policy network?Policy network: A model that, for a given position, proposes a distribution over plausible moves. In AlphaGo it generated candidates for further analysis. supplied candidate moves — fast, intuitive pattern matching, System 1?System 1: Fast, automatic thinking in Daniel Kahneman's split. System 2 is slow, deliberative thinking. in Daniel Kahneman's terms. Deliberation came from tree search, testing those hunches against the moves and countermoves to follow. Chain-of-thought stretches the first mode without adding the second.
Three charges against chain-of-thought
| Charge | What it means | Consequence |
|---|---|---|
| No explicit state | The model keeps no open ledger of hypotheses, confidence, evidence and open questions | There is no way to check what it treats as settled |
| Knowledge fused with reasoning | Both sit interwoven in the network's weights | Facts cannot be swapped without disturbing inference |
| Post-hoc rationalisation | The justification is built after the answer is settled | The reasoning trace stops being evidence |
That last point has hard backing. An arXiv study of fourteen reasoning models trained on outcome-based rewards found most chains of thought illegible, and using only the comprehensible portions cut accuracy by 53%.
What replaces more scaling
Instead of extending System 1, Graepel proposes systems holding an explicit epistemic state?Epistemic state: A maintained record of what the system treats as settled, doubtful, ruled out and still open. Closer to an investigator's notebook than to a run of sentences..
An explicit epistemic state is four buckets the system updates as it reasons:
Reasoning becomes a sequence of moves reducing uncertainty, with tool access through APIs and an auditable trace. The case rests on medicine and science, where the stakes rule out unverifiable justifications.
Why it matters
The position is strong because it rests on architecture rather than benchmarks. If reasoning requires a separate, explicit state, no volume of data or parameters closes the gap, and the current path is an expensive detour. The claim still has to be settled empirically, but it moves the question from how large the model is to how the system is structured.
What's next?
- Graepel left Google DeepMind to develop an alternative approach to machine reasoning
- If chains of thought stay illegible, oversight by humans and by other systems loses its footing, as the fourteen-model study indicates
Sources
- MIT Technology Review — Don't be fooled—LLMs don't reason
- arXiv — Reasoning Models Sometimes Output Illegible Chains of Thought
