AI Isn’t Outthinking Mathematicians. It’s Out-Remembering Them.

gist

Davide Piffer proposes that some of AI’s mathematical advantage comes from augmented symbolic working memory rather than deeper reasoning. A long context, written scratch work, tools, and verification let a model preserve constraints and explore long formal chains beyond unaided human capacity. Mathematics converts this architectural advantage into performance because its state is explicit and checkable. The hypothesis predicts smaller gains on problems whose main difficulty is a short, original reframing rather than extensive bookkeeping.

ideas

• Writing is part of the computation. A model’s generated tokens act as a persistent notebook of assumptions, intermediate results, and abandoned branches.
• Formal domains amplify external memory. Mathematical symbols keep definitions stable and make constraints easier to retrieve, manipulate, and verify than ambiguous real-world evidence.
• Long chains create a coordination burden. A person may understand every local inference yet fail to generate the whole proof because maintaining its global state exceeds working-memory capacity.
• Context is not perfect human memory. Long windows are passive, retrieval is fallible, and models remain weak at maintaining a small private state that can be continuously updated.
• The hypothesis is testable. Restricting scratch work or usable context should hurt long, constraint-heavy tasks more than problems dominated by one conceptual leap.