Every time an AI model cracks a hard math problem, the same story gets told: the machine is getting smarter. Better training. Better reasoning. Maybe even a spark of real mathematical intuition. A widely shared essay on Hacker News, sitting near the top of the front page with 186 points, argues that story is missing something simpler and far less flattering to the models. The AI might not be out-thinking us. It might just be out-remembering us.
That’s the contrarian claim worth sitting with. What we read as superior intelligence may partly be the removal of a limit that has capped human reasoning for as long as humans have done math: our tiny working memory.
The bottleneck we forget we have
Working memory is the mental scratchpad that holds information while you manipulate it. Try multiplying two three-digit numbers in your head. The operations are easy. The hard part is keeping the partial results alive while you do the next step. Write the numbers down and the problem gets simple fast.
As the article puts it, paper doesn’t make you smarter. It expands your effective working memory. Mathematicians lean on notation, diagrams, scratch paper and previously proven lemmas not just to communicate, but to make the thinking possible at all. Experts stretch the limit through “chunking,” recognizing a familiar structure and treating it as one object. But chunking compresses the constraint. It doesn’t erase it.
Why this reframes the AI story
Here’s the move that matters. A language model can hold the entire problem statement, hundreds of intermediate equations, several abandoned approaches, every definition and constraint, and its own earlier reasoning all inside one context window. The essay’s sharpest line: the context window is best understood not as a brain, but as “a gigantic external notebook combined with an imperfect system for searching and using what has been written in it.”
So when a model outperforms a human on a proof, part of that gap may not be reasoning quality. It may be that the human is fighting a biological handicap the machine simply doesn’t have.
The author backs this with cognitive science rather than vibes:
- Alloway and Passolunghi (2011) found working memory made a distinct contribution to math performance beyond verbal ability.
- Alloway and Alloway (2010), a six-year study, found early working memory predicted later numeracy even after controlling for IQ, and predicted it better than the IQ measure did.
- Blankenship and colleagues (2015) reported working memory explained unique variance in math fluency after controlling for IQ and age.
- A Friso-van den Bos meta-analysis (2013) found a consistent working-memory-to-math link across primary school studies.
The honest caveat is included too. Working memory and intelligence overlap heavily, and no statistic cleanly separates them. But the narrow conclusion holds: among people of similar measured intelligence, the ability to hold and juggle information still predicts who does math better.
Why it matters now
This lands right as the industry sells ever-longer context windows as a headline feature and frames benchmark wins as leaps in machine intelligence. If a meaningful slice of that performance is really “bigger notebook, not bigger brain,” it changes how you read the scoreboard.
What stands out to me: this is a testable, deflationary explanation in a field addicted to inflationary ones. It doesn’t say models aren’t useful. It says we may be misattributing where the usefulness comes from.
A few practical takeaways:
- For builders: treat the context window as external memory to engineer, not intelligence to trust. How the model searches and reuses what’s in there is the real bottleneck.
- For buyers: discount benchmark hype that conflates capacity with cognition. Ask what the task actually stresses.
- For everyone: the cheapest performance boost in history is still writing things down.
The full essay, with its citations and its argument about what humans have that models don’t, is worth reading at the original source.