AI's math edge may be memory, not reasoning

As AI systems post results on hard math problems, Davide Piffer offers a deflating explanation: the models may not be outthinking mathematicians so much as out-remembering them. His claim is that most of the gain comes from working memory, the ability to hold a problem statement, definitions, intermediate steps, and earlier results all in view at once, rather than from any deeper reasoning. A large context window, in this reading, is a gigantic external notebook, and the model is a very fast von Neumann rather than an Einstein having a conceptual breakthrough.

He grounds it in cognitive science. Piffer points to work by Alloway and colleagues showing that working memory predicts math performance on its own, sometimes more strongly than IQ. Mathematics is unusually friendly to this kind of advantage because almost everything relevant can be written down explicitly, and formal proofs give strong feedback about whether a step is right.

The useful part is that the hypothesis makes predictions. AI should do best on problems with many interacting constraints, long calculations, and heavy case analysis, and worst on problems that hinge on a single conceptual leap or a fresh way of framing the question. That is a test, not just a vibe.

Why it matters

If you use AI for technical work, this tells you where to trust it and where to stay skeptical. Lean on it for long, bookkeeping-heavy problems, and check it carefully when success depends on one clever insight rather than careful tracking of many parts.