Nice explainer. One nuance I’d be careful with: “the brain doesn’t think in words” is broader than what the study seems to establish. The stronger result is that the natural-language network is not necessary for the kinds of inductive and deductive reasoning tested here.

The AI implication is fascinating though. We tend to build systems where the language model is simultaneously the language interface, memory, planner, reasoner and controller. Human cognition may be pointing in the opposite architectural direction: separate persistent state, reasoning/control mechanisms and language, then let language be one tool among several.

That distinction seems more interesting to me than simply asking whether LLMs “reason in language.”