The Real Math Edge for AI Isn't Smarter Reasoning — It's a Bigger Notebook

When an AI cracks a hard math problem, the reflex is to credit better reasoning, more training data, or emerging intuition. This piece argues for a plainer explanation: the model simply has far more working memory. Human mathematicians can only juggle a handful of unfamiliar elements at once, which is why scratch paper, notation, and lemmas exist — not to boost intelligence, but to offload the memory a proof demands. A language model, by contrast, can hold the full problem statement, hundreds of intermediate steps, abandoned approaches, and constraints inside its context window at the same time.

The author leans on cognitive research showing working memory predicts math performance even after controlling for IQ — across children of similar measured intelligence, those who can hold and manipulate more information do better. If a biological memory bottleneck caps human math ability, then handing a machine a vast symbolic workspace doesn’t just make it faster; it removes the constraint that suppresses human performance in the first place. The apparent intelligence gap may partly be a memory gap.

The context window isn’t a perfect analogue to human working memory, though. It behaves more like an enormous external notebook paired with an imperfect retrieval system: models can overlook, get distracted by, or lose track of information they technically still hold, so advertised context length overstates usable memory. The advantage is also uneven — it pays off most in mathematics precisely because math translates cleanly into stable, explicit symbols that stay put once written down, unlike domains that resist full symbolic capture.