Davide Piffer Says GPT Isn't Smarter, Just Less Forgetful

Everyone wants to believe the models got smarter. The truth might be dumber and more interesting: they just stopped forgetting.

On August 4, 2026, Davide Piffer published a piece with a title that should sting anyone who's spent the last two years marveling at AI math benchmarks: "AI Isn't Outthinking Mathematicians. It's Out-Remembering Them." It landed on Hacker News with 486 points and 415 comments — the kind of engagement that usually means someone said the quiet part out loud.

Here's the quiet part: what looks like reasoning might just be storage.

The Argument, Stripped Down

Piffer's claim isn't that AI is fake or that GPT-whatever can't do math. It's that the mechanism behind the math might not be cognition in any meaningful sense — it's capacity. Human mathematicians juggle maybe four or five things in working memory before they need paper. AI models hold thousands of tokens of problem statements, assumptions, equations, and half-finished proofs in an "enormous external symbolic workspace" simultaneously.

That's the whole thesis in one sentence, and it's uncomfortably plausible.

Think about how mathematicians actually work. They don't hold entire proofs in their heads — they use paper, whiteboards, notation systems, entire external scaffolding built specifically to compensate for the brain's tiny working memory. Piffer's point is that context windows are just a much, much bigger whiteboard. Not a better brain.

The Elephant in the Room

Here's what nobody wants to say directly: if this thesis is even partially true, a huge chunk of the AI reasoning benchmark industry is measuring the wrong thing.

Benchmarks love to frame results as evidence of "reasoning ability." But if a model's edge is really about holding 50 intermediate results in view without losing track of any of them, then we're not testing intelligence — we're testing RAM. That's a category error dressed up as a scientific finding, and it's been driving billions in valuation narratives.

Noah Smith apparently picked up the thread too, suggesting on social media that true superintelligence might be more about working memory scale than reasoning depth. That's a notable admission from someone who writes seriously about AI economics — it's not just contrarian bloggers making this case anymore.

But let's not pretend this settles anything. The counterargument matters: mathematical competence also requires learned abstractions, pattern recognition, and something that at least resembles algorithmic judgment. Memory alone doesn't write a valid proof. A bigger whiteboard doesn't help you if you don't know what to write on it.

What Actually Changes If This Is True

For developers building on top of these models, the implications are concrete, not philosophical:

1. Test context degradation directly. Don't assume a model "reasoned through" a problem — shrink the context window and see what breaks.

2. Treat memory as an allocated resource, not a side effect. Scratchpads, retrieval, structured intermediate state — these become architecture decisions, not implementation details.

3. Reconsider vendor selection criteria. Context robustness and state persistence may matter more than raw benchmark scores.

The market implication is sharper: if the real differentiator is context handling rather than algorithmic breakthroughs, then the orchestration layer — memory stores, agent frameworks, retrieval pipelines — becomes more valuable, not less. The model itself starts to look like a commodity sitting on top of infrastructure that actually does the remembering.

Nobody's proven Piffer's thesis definitively. It's an interpretation, not a settled result. But it's the kind of interpretation that should make every benchmark chart in every pitch deck a little less trustworthy than it looked last week.