Tao's ChatGPT Transcript Reveals How Experts Actually Use AI

The signal: Terence Tao published a ChatGPT conversation showing how he used the model to probe a potential counterexample to the Jacobian Conjecture, and the HackerNews crowd went wild watching a Fields Medalist think out loud with an AI.

Why it matters: This isn’t a story about AI solving open math problems — it’s a story about how the best minds in a field actually use these tools day to day, and that’s a better signal than any benchmark. If you’re building AI products for experts, the workflow Tao demonstrated — fast hypothesis testing, error-catching, notation-wrangling — is the actual product-market fit, not “AI writes a proof.”

Did ChatGPT actually solve part of an unsolved math problem?

No — Tao used ChatGPT as a fast, tireless collaborator to check algebra, explore edge cases, and rule out dead ends, not as an oracle that produced a novel result on its own. The conversation reads like pair programming: Tao steers, proposes structure, and catches the model’s mistakes in real time, while ChatGPT accelerates the grunt work of trying variations. That distinction matters because the discourse keeps swinging between “AI can’t do real math” and “AI just did research-level math,” when the actual value is boringly practical: it’s a fast calculator with a decent model of the literature. Anyone who’s used LLMs for real coding or research work already knows this pattern — the win is iteration speed, not autonomy.

The pattern I’m watching: Every viral “look what AI did with an expert” story follows the same arc — initial breathless takes, then a correction thread pointing out the human did the actual thinking. That correction thread is where the real product insight lives, and it’s the same recalibration behind why “pelicanmaxxing” (labs benchmarking on absurd tasks) is trending today — the industry is quietly redefining what “capability” even means.

What I’d do with this: If you’re building dev tools or research tools on top of LLMs, stop optimizing for “can it get the right answer alone” and start optimizing for loop speed — how fast can an expert propose, test, and discard ideas with the model. Instrument your product for that loop, not for one-shot accuracy, because that’s what Tao’s transcript actually demonstrates as valuable.

Key takeaways

• Terence Tao’s viral ChatGPT transcript shows expert-guided iteration, not autonomous AI discovery, and that distinction is the whole story.
• The real product opportunity in AI-for-experts tools is compressing the propose-test-discard loop, not chasing one-shot correctness.
• Viral “AI did X” stories reliably get walked back once the expert’s steering role becomes clear, and builders should design for that reality now.