AI isn't outthinking mathematicians: it's out-remembering them

An essay posted on davidepiffer.com and widely discussed on Hacker News argues that AI's apparent skill at hard mathematics owes less to superior reasoning than to a far larger working memory than the human brain has. When an AI system solves a difficult problem, the usual explanations are that it has absorbed millions of worked examples, that reinforcement learning taught it better strategies, or that it is developing something like genuine mathematical intuition. The essay does not dismiss these, but argues they overlook a simpler factor: AI systems have access to an enormous external symbolic workspace, the context window, that performs many of the functions working memory performs in humans. Some of what looks like superior reasoning, the essay argues, may instead be the removal of one of the biggest biological limits on human thought: how little unfamiliar information a person can hold in mind at once.

Human working memory is sharply limited. Solving an equation means tracking what each variable represents, which operations are already done, and the current goal; a proof means tracking assumptions, intermediate lemmas, exceptions and multiple cases at once. The essay illustrates the limit with mental arithmetic: multiplying two three-digit numbers in your head is hard not because the individual operations are difficult, but because you must preserve partial results while computing further ones. Writing the numbers down does not make a person smarter; it expands their effective working memory. The essay argues professional mathematicians do the same with notation, scratch paper and diagrams, not just to communicate a solution but to make the reasoning cognitively possible. Experts also compensate through chunking: a novice sees a long sequence of symbols, while an expert recognizes a familiar structure and treats it as a single conceptual object, letting far more information fit inside the same biological limit. But chunking does not eliminate that limit. It only compresses the information.

The essay backs this with four cited studies of human subjects. Alloway and Passolunghi (2011) examined working memory, verbal ability and mathematical skills in children and found working memory made a distinct contribution to math performance beyond general verbal ability. Alloway and Alloway (2010) measured children at age five, then examined their academic achievement six years later in a six-year longitudinal study; early working memory predicted later literacy and numeracy even after controlling for IQ, and was reported as a stronger predictor of those outcomes than the IQ measure the study used. Blankenship and colleagues (2015) found working memory explained unique variation in mathematical fluency and calculation after statistically controlling for IQ and age. A meta-analysis by Friso-van den Bos and colleagues (2013) found a consistent relationship between working memory and mathematics across primary-school studies, though the strength varied by the type of working-memory and math task measured. The essay cautions against over-reading these results: working memory and intelligence overlap substantially; statistical control cannot perfectly separate them as independent mechanisms; and none of it implies that commercial working-memory training would produce large gains in intelligence or math ability. The narrower conclusion it draws is that among children with similar measured intelligence, differences in the ability to hold, update and manipulate information still predict differences in math performance. It treats this as a clue for understanding AI: if human mathematical performance is partly capped by a working-memory bottleneck, giving a machine an enormous symbolic workspace changes the nature of the contest, and a model may appear more mathematically intelligent partly because it is far less constrained by the cognitive limit that suppresses human performance.

The essay argues the context window is not identical to human working memory: it is better understood as a gigantic external notebook combined with an imperfect system for searching and using what has been written in it. Unlike a human, who can silently hold and transform a private mental state without writing anything down, a standard language model's most stable memory is the sequence of tokens it has already generated. Once a model writes 'x=6' and later 'x+3=9,' both statements stay inside the context and can be attended to again. Its reasoning is externalized: the generated text is not just a report of a finished thought; it is part of the mechanism by which the reasoning happens. The scale dwarfs a human's: the essay estimates an unaided person may struggle to keep five unfamiliar conditions active at once, while an AI can preserve dozens or hundreds of them explicitly, alongside an entire problem statement, hundreds of intermediate equations, abandoned approaches, definitions, constraints and earlier conclusions. The essay adds a caveat here too: models do not retrieve every item in a long context perfectly; they can overlook information, get distracted or lose track of details. Advertised context length, it notes, is not the same as perfectly usable memory.

The essay argues this advantage is unusually well matched to mathematics because a math problem can almost entirely be written down: assumptions, definitions, known equations, the current objective, results already proved, eliminated cases and the conditions under which each step stays valid. Once written, mathematical facts stay stable; if a variable is defined as an integer at the start of a proof, it remains one unless the proof explicitly redefines it. As an example, the essay describes a solver who must remember that a number is odd, another is prime, and a variable is nonzero, with one branch already producing a contradiction. A human may understand the right strategy and still divide by that variable before confirming it is nonzero, an error the essay attributes not necessarily to a lack of intelligence but to a failure of bookkeeping, the kind an AI can avoid by restating its active assumptions at each step. Because proofs are compositional, built as chains where each step depends on the last, a larger working space lets a model sustain a longer chain before losing the thread. The essay notes a person can be capable of following every local step of a 100-step argument while still being unable to generate the whole argument unaided, because the length exceeds their ability to hold the global structure in mind. This, it suggests, is why giving a model more 'thinking time' often improves its performance: the extra computation may amount to a broader search inside a much larger notebook rather than a deeper thought process.

The essay is explicit that this advantage does not carry over to informal reasoning. Its example: a much larger context window could let an AI scan years of correspondence for changes in tone, timing and vocabulary to guess why someone named Maria has stopped replying to messages. But the decisive fact, whether she is angry, busy, ill, has lost her phone, or is avoiding something else entirely, may simply never have been observed, and no amount of memory can retrieve a fact that was never recorded. Informal reasoning also leans on concepts with unstable meanings, such as 'fair,' 'successful,' 'responsible,' 'harmful' or 'intelligent,' whose sense depends on culture, goals and context in a way mathematical variables do not; a model can remember every sentence of a discussion and still misread what the participants mean. The essay extends the same limit to political analysis, historical interpretation, business strategy and psychological judgment, arguing the central difficulty there is usually not memory capacity but identifying the right causal model when the evidence is incomplete and ambiguous, a problem a bigger notebook alone does not solve.

Key facts

The essay argues AI's edge on hard math problems may come less from superior reasoning than from a vastly larger symbolic working memory, the model's context window.
It backs the claim with four cited studies linking working memory to math performance beyond IQ: Alloway and Passolunghi (2011), the six-year longitudinal study by Alloway and Alloway (2010) that began measuring children at age five, Blankenship and colleagues (2015), and a meta-analysis by Friso-van den Bos and colleagues (2013).
As an illustration, the essay says an unaided human may struggle to hold five unfamiliar conditions active at once, while an AI can preserve dozens or hundreds of them explicitly.
It argues a person can follow every local step of a 100-step argument yet be unable to generate the whole argument unaided, because the length exceeds their ability to hold the global structure in mind.
The essay limits its own claim to formal, symbolic reasoning: it says the advantage does not extend to informal reasoning about ambiguous concepts or hidden, unobserved causes, such as guessing why someone has stopped replying to messages.

Why it matters

If the essay's framing holds, part of what reads as AI becoming more intelligent at mathematics may really be a side effect of giving it a far bigger place to keep its work, the same effect a person gets from switching from mental arithmetic to pen and paper, scaled up by orders of magnitude. A model's success on a hard, multi-step problem may then reflect that it held hundreds of intermediate steps and constraints without losing any of them, not that its underlying reasoning ability improved. The essay also uses this to explain why giving a model more 'thinking time' tends to help: the extra computation may be a wider search of a large notebook rather than a qualitatively deeper thought process, which changes how much credit belongs to 'reasoning' as opposed to memory.

Who it affects

This is aimed at anyone reading AI math or reasoning results and asking how much credit belongs to the model getting smarter versus the model simply forgetting less: researchers, engineers and anyone comparing systems on multi-step symbolic tasks such as proofs, algebra or formal derivations, where the essay's mechanism, an explicit, stable, searchable record of the problem state, predicts a real advantage. Its caution matters just as much to anyone tempted to carry the same reasoning into informal domains: judging an AI's handling of ambiguous, real-world questions is a different test than judging its handling of symbolic ones, and the essay says the memory advantage does not transfer.

How to use it

The essay works better as a diagnostic than a tool: when an AI's performance on a hard problem looks impressive, it suggests asking whether the win comes from tracking more of the problem at once rather than reasoning about it more cleverly, since the two can look identical from the outside but carry different implications for how far the capability generalizes. The same logic runs the other way for people. The essay's own point about why mathematicians lean on notation and scratch paper, that writing does not make a person smarter but expands their effective working memory, doubles as a reminder that writing a hard problem down, rather than holding it in your head, closes part of the gap with a system that has a far bigger notebook to work with. The essay is explicit that none of this transfers to informal, ambiguous reasoning: it is not a lens for judging an AI's handling of social or interpretive questions, only its handling of formal, symbolic ones.

How solid is it

The essay's central claim, that context-window size explains part of AI's math performance, rests on analogy rather than a new experiment: it draws on four existing studies of human working memory and math ability, Alloway and Passolunghi (2011), the six-year longitudinal study by Alloway and Alloway (2010) that began measuring children at age five, Blankenship and colleagues (2015), and a meta-analysis by Friso-van den Bos and colleagues (2013), all of them studies of human subjects rather than AI systems. No specific AI model, product, lab or benchmark is named anywhere in the text, and no numeric figure is given for the actual size of an AI context window or of human working-memory capacity; the human-versus-AI contrast, five items against 'dozens or hundreds,' is stated as an illustration, not a measurement. The essay hedges consistently, using 'may,' 'some of it' and 'partly' throughout, and does not claim the context-window explanation fully accounts for AI's mathematical performance, only that it may account for part of it. No author name or publication date appears anywhere in the text; only the four cited studies carry years.

Risks and caveats

The essay warns against over-reading its own source studies: working memory and general intelligence overlap substantially; statistical control cannot cleanly separate them as independent mechanisms; and the findings do not imply that commercial working-memory training would meaningfully raise intelligence or math skill. On the AI side, it cautions that advertised context length is not the same as perfectly usable memory, since models can overlook information inside a long context, get distracted, or lose track of details. The essay limits its own claim to formal, symbolic reasoning above all: it explicitly does not extend to informal reasoning about ambiguous concepts or about hidden causes that were never observed, such as guessing why a person stopped replying to messages, or to political analysis, historical interpretation, business strategy and psychological judgment, where the essay says the harder problem is identifying the right causal model under incomplete evidence, not holding more of it in memory.

“The context window is not identical to human working memory. It is better understood as a gigantic external notebook combined with an imperfect system for searching and using what has been written in it.”

— the author