AI Isn't Outthinking Mathematicians — It's Out-Remembering Them

When a language model solves a hard math problem, we tend to say it has “gotten smarter”: it absorbed millions of examples, learned better reasoning strategies, acquired something resembling intuition. Davide Piffer proposes a more prosaic explanation: AI's key advantage isn't superior thinking, but a colossal external symbolic “scratchpad” that serves as working memory.

Mathematics Is Constrained by Memory

A person can hold only a few unfamiliar items in mind at once. Try multiplying two three-digit numbers in your head: the operations are simple, but the difficulty lies in retaining the intermediate results. Paper doesn't make us smarter — it expands our effective working memory. The same goes for notation, drafts, and previously proven lemmas: experts compress information into “chunks,” but they don't remove the biological ceiling.

The author cites research (Alloway, Passolunghi, and others) in which working memory predicted children's mathematical achievement even after controlling for IQ. The conclusion is cautious but important: at equal intelligence, differences in the ability to hold and update information still affect the outcome.

The Context Window as a Giant Scratchpad

A model holds in context the problem statement, the definitions, dozens of intermediate equations, discarded approaches, and its own earlier conclusions. Its reasoning is externalized: having written “x=6” and later “x+3=9,” it can turn back to those lines. A person struggles to keep five conditions active; an AI keeps hundreds. This isn't an exact analogue of human working memory (it's more accurate to speak of “augmented symbolic memory”), and not everything in a long context is retrieved perfectly, but the difference in potential capacity is enormous.

Mathematics is perfectly suited to such an architecture: almost everything can be written down explicitly and unambiguously — premises, constraints, cases already ruled out. A person may understand the strategy but divide by x, forgetting that x≠0. That's not an error of intellect but of bookkeeping. An AI keeps a “ledger” of the problem's state and therefore copes with long chains A→B→C→D, where the difficulty is determined not by any single step but by the number of steps that must be kept consistent. What looks like deeper thought sometimes turns out to be a broader search across a large scratchpad.

Where the Advantage Disappears

In informal domains — “why did Maria stop answering?”, in politics, history, psychology — what matters isn't the volume of memory but a correct causal model applied to incomplete and ambiguous data. No amount of memory can reconstruct facts that were never observed, and words like “fair” or “successful” lack the precision of variables. Mathematics, by contrast, rewards systems capable of generating, storing, and checking explicit intermediate states — and here code and formal verifiers such as Lean help.

Piffer recalls Wigner's observation: von Neumann amazed people with the speed and breadth of his mind, but Einstein understood more deeply and could reconceive the very problem. Today's AI resembles a machine-amplified von Neumann with a limitless scratchpad more than it does an Einstein. The real threshold will be crossed when AI learns to recognize that a problem has been posed incorrectly and to invent a fundamentally new way of understanding it.

Source: Hacker News