Last updated: August 17, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
AI isn’t outthinking mathematicians when it blazes through a proof. It's out-remembering them. That's the argument independent researcher Davide Piffer laid out in an August 2026 essay that spread fast among mathematicians and AI researchers. His claim: the sharpest gains in AI math performance may not come from deeper reasoning. They come from a working-memory edge no human brain has. A person doing hard math can hold only a handful of unfamiliar symbols and half-finished sub-proofs in mind before something falls out. A large language model keeps the whole problem, every intermediate equation, every dead-end attempt, and every earlier conclusion inside its context window at once, and can search back through all of it on demand. That structural difference, not superior insight, may explain most of what looks like sudden mathematical talent.
Short answer: AI's apparent skill at hard math may come less from better reasoning than from raw working memory. A human can hold only a few unfamiliar ideas at once; an AI model keeps the whole problem, every failed attempt, and every earlier step inside its context window, searchable on demand — a structural edge, not necessarily a smarter one.
I pulled Piffer's original essay instead of working from the Hacker News thread quoting it. I also cross-checked the numbers below against each AI vendor's own documentation in August 2026, not a roundup site. The goal isn't to settle whether AI "really" understands math. It's to give you a working theory of what these tools are good at, so you stop asking them to do the one thing they're still weak at.
What you'll need
Nothing exotic. You don't need to be a working mathematician to use this. The value here is a way of thinking about what AI is good at, not a math tutorial. Open an AI chat with a real context window: ChatGPT, Claude, or Gemini all qualify, since the memory edge Piffer describes shows up in all three. Bring one real problem with actual structure. Several constraints, a multi-step derivation, or a proof with more than two or three moves works well. Skip the one-line question a search engine could answer just as fast. And bring patience for verification. Memory isn't correctness, so you'll need to check the output either way.
Step-by-step: put the memory advantage to work
1. Load the whole problem before you ask anything
Paste the full problem, every constraint, every definition, and any partial work you've already done into one message. Don't compress it into a summary first. Piffer's argument only pays off if the model actually has everything in view at once. Summarizing away detail throws away the exact advantage you're trying to use.
2. Let it carry the failed attempts, not you
Keep exploring inside the same conversation instead of opening a fresh chat for every new angle. A huge context window can hold your abandoned approaches alongside the current one. That means the model can reference the thing that didn't work three tries ago without you re-explaining it.
3. Do the reframing yourself first
Piffer is specific about where humans still hold the edge: finding an entirely new way to represent a problem. If the sticking point is genuinely conceptual, not more cases to check but a different way of seeing the problem, that's still your job. Spend your own effort there before you hand anything to the model.
4. Ask it to run out every consequence of your framing
Once you've picked a representation, the working-memory edge does the most work. It can track every branch, every case, every step of a long derivation without losing the thread. Ask explicitly for full case coverage, not just a final answer.
5. Verify every step, don't trust the confidence
A long, tidy derivation isn't the same as a correct one. Ask the model to show each step on its own so you can check it. Independently verify any number, formula, or citation before you rely on it. Terence Tao’s own framework for this is the fastest filter I've found: could you defend the answer to a room full of people, without further AI help?
Example prompts you can copy
These are built around the working-memory framing, not generic solve-this requests. Paste your own details into the brackets:
- Here is the full problem, every constraint, and every definition I'm working with: [paste everything]. Keep all of this in mind and don't drop any constraint as we go.
- I tried these three approaches and abandoned each one for [reasons]. Before trying anything new, tell me exactly what each attempt already ruled out.
- I've decided to represent the problem this way: [your framing]. Now work out every consequence of that choice, including edge cases I might have missed.
- List every step of your derivation on its own line, including the ones that feel obvious, so I can check each one against my own reasoning.
- Where in this derivation are you filling a gap from pattern-matching, rather than actually deriving the next step from what's already on the page?
Common mistakes to avoid
In my testing, the most common failure is treating a long, confident derivation as correct because it's long. Piffer's whole point is that memory volume and correctness sit on different axes. A model can hold a hundred steps in view and still get one wrong. A bigger notebook doesn't audit itself.
Second: restarting the chat mid-problem. Each fresh session throws away the working-memory edge that's supposedly doing the heavy lifting. You end up paying for AI's weaknesses without keeping its main strength, a live, searchable record of everything already tried.
Third: asking for the conceptual leap instead of making it yourself. When I tested this on problems that needed one genuinely new representation, not more casework, the answers got noticeably weaker and more circular. That's exactly the boundary Piffer's theory predicts.
Fourth: skipping verification because the notation looks rigorous. Symbolic precision reads as authority. It isn't the same thing as being right.
The context-window numbers, verified
Piffer's argument is about working memory in general. In current chatbots, that maps most directly onto context window size: how much of a problem a model can hold before it starts losing track of earlier parts of the conversation. I checked each vendor's own documentation directly rather than trust a secondhand table.
| Assistant | Current flagship | Context window | Source |
|---|---|---|---|
| ChatGPT | GPT-5.6 (Sol/Terra/Luna) | 1.05M tokens | OpenAI model docs |
| Claude | Opus 5 / Sonnet 5 | 1M tokens | Anthropic model docs |
| Gemini | 3.1 Pro | 1M tokens | Google DeepMind model card |
For scale, a million tokens is roughly 750,000 words. That's more than a mathematician could realistically hold across a week of scratch paper, let alone in working memory at one time. Compare that to FrontierMath, a benchmark of research-level problems introduced in November 2024. At launch, the paper reports that current state-of-the-art models solved under 2% of its problems. Today's scores come from models running with roughly 500 times more working room than the ones that failed almost everything on day one. That's itself evidence for Piffer's theory, not against it.
Tools that make this easier
If you don't already have accounts open, my guides to ChatGPT, Claude, and Gemini cover setup on each. Free AI tools is the place to start if the cost of testing all three is the blocker. Want to see how differently the major models respond to identical input? My one-prompt, 11-model comparison is a more direct test than anything in this piece. Pair Piffer's framing with Terence Tao's own framework for trusting AI math output. Tao's verification test is the other half of what this article covers. Students specifically should look at the best AI tool for study before settling on a default assistant for coursework.
My take
AI isn’t outthinking mathematicians, it’s out-remembering them, and once you accept that framing, what matters isn't the mechanism. It's the prediction it makes. If the theory holds, AI should keep improving fastest on math that rewards holding more in view at once: long case analysis, multi-constraint optimization, bookkeeping-heavy proofs. It should keep lagging on problems that need one genuinely new idea, no matter how big the context window gets. That's a testable claim, not a vibe. It gives you a concrete way to decide when to trust a long AI derivation, and when the real work is still yours to do.
Frequently Asked Questions
Is the working-memory theory of AI math skill peer-reviewed?
No. Piffer's essay is an independent analysis published on his Substack in August 2026. It draws on established working-memory research in human cognition, not new experiments on AI models. Treat it as a well-argued hypothesis, not a settled result.
Does a bigger context window guarantee better math answers?
No. A larger context window means a model can hold more of a problem without losing track of it. Holding more information doesn't make each individual step correct. You still have to verify the derivation yourself.
Which AI tool has the biggest context window right now?
As of August 2026, ChatGPT's GPT-5.6 models lead at roughly 1.05 million tokens. Claude's and Gemini's current flagship models follow close behind at 1 million tokens each, per each vendor's own documentation.
Is this free to try?
Yes. ChatGPT, Claude, and Gemini all offer free tiers that are enough to test the prompting approach in this guide. Free tiers typically cap the context window well below the flagship figures above, though.
What's the easiest way to apply this without reading Piffer's full essay?
Start with step 1. Paste your entire problem, with every constraint and definition, into a single AI conversation instead of feeding it in pieces. That alone captures most of the practical difference this article describes.