Terence Tao's ChatGPT Math Session Redefines AI's Role

The signal: Terence Tao posted a ChatGPT conversation where he used the model to help probe a potential counterexample to the Jacobian Conjecture, and it topped Hacker News today.

Why it matters: This isn’t a chatbot generating a proof from scratch — it’s one of the sharpest mathematicians alive using an LLM as a fast, tireless collaborator to check algebra, explore cases, and rule out dead ends. For builders, that’s the actual pattern worth copying: expert-in-the-loop reasoning, not autonomous discovery. It also quietly resets the “AI can’t do real math” narrative that’s dominated HN threads for two years.

Does this mean AI can now solve unsolved math problems?

No — Tao was using ChatGPT as a research assistant to accelerate his own reasoning, not letting it independently crack a decades-old conjecture. The conversation shows him steering the model through specific algebraic manipulations, catching its errors, and using it to rapidly test hypotheses he’d already framed. That’s the same workflow good engineers use with Copilot or Cursor: the human holds the problem structure, the model handles grunt-work exploration at high speed. The real value isn’t “AI is smart enough to do novel research alone” — it’s “AI collapses the cost of testing an idea from hours to minutes.”

The pattern I’m watching: Every viral AI story this year follows the same arc — skepticism, then a credible expert quietly demonstrates a narrow, well-scoped use case, then the skepticism shifts to “okay but where’s the ceiling.” Tao doing this with a famously hard, unsolved conjecture is the highest-credibility version of that arc we’ve seen. Expect more domain experts — not AI researchers — to start publishing raw prompt transcripts as a credibility signal of their own.

What I’d do with this: If you’re building dev tools, stop marketing “AI writes your code” and start marketing “AI checks your work at a speed no human can match” — that’s the framing that just got endorsed by a Fields Medalist. If you’re an engineer, start treating your own LLM sessions as artifacts worth sharing; Tao just showed that transparent work, mistakes included, builds more trust than a polished writeup.

Key takeaways

• Terence Tao used ChatGPT as a research collaborator to explore a Jacobian Conjecture counterexample, not to autonomously solve it.
• The winning AI pattern right now is expert-in-the-loop acceleration, where humans hold problem structure and models handle rapid exploration.
• The “AI can’t do real math” narrative is weakening not because of benchmark scores, but because credible experts are demonstrating narrow, well-scoped wins in public.