What's this about?
The weekend's most-read piece argues that when an AI system solves a hard mathematical problem, the usual explanations — it absorbed millions of examples, reinforcement learning taught it better strategies, it's developing intuition — may all be partly true but overlook something simpler. The AI has access to a vastly larger working memory. The image the piece leaves you with: not an electronic Einstein, but a machine-amplified von Neumann — immense speed, breadth, and symbolic memory.
What exactly is "working memory"?
Working memory is what you can actively hold and manipulate at once, as distinct from what you know. The classic estimate is around seven items. Humans get around it by chunking — grouping details into single units, the way a chess master sees a board position rather than sixteen separate pieces. But as the article notes, chunking compresses the information without removing the ceiling. A system that can hold hundreds of thousands of tokens in active attention isn't doing the same thing faster; it's doing something humans structurally cannot, which may look like insight from outside.
What's the community saying?
- "It's also out-brute-forcing them." The most valuable addition: a mathematician who spends a week on a research direction that fails gets discouraged and needs a break. The machine is never tired, never demoralized, and simply proceeds to the next direction.
- The comprehension worry. A widely-shared and genuinely unsettling thread: if AI produces arguments requiring a working memory of hundreds of items, humans may lose the ability to verify them. One commenter went further — "the age of humans comprehending things is coming to an end."
- It generalizes past mathematics. Several noted the same applies anywhere: you can have a model read an entire field's literature and cross-reference all of it at once, which no researcher can.
- Substantial pushback on the framing. "'It's not X, it's Y' hot take" and "the goalposts must move once again" both got traction — the objection being that this is a redescription rather than a deflation, since out-remembering at sufficient scale may just be what out-thinking looks like.
- A meta-complaint about AI prose. Multiple commenters flagged the article itself as AI-polished and asked for a way to filter such writing — the same irritation that surfaced about security write-ups last week.