The mathematical world is processing a remarkable claim: that Claude Fable, Anthropic's frontier model, produced a counterexample to the Jacobian Conjecture — an open problem in algebraic geometry that has resisted proof or disproof for over 85 years. To get a handle on what the result actually says, Fields Medalist Terence Tao did what many serious researchers now do: he opened a chat window and started asking questions.
Tao shared his full ChatGPT conversation publicly, walking through the structure of the alleged counterexample step by step. The exchange is notable not as a novelty but as a working method — a world-class mathematician using a large language model as an interactive sounding board to rapidly digest and interrogate a complex technical claim. The conversation covers the core construction, where the argument depends on, and what would need to hold for the counterexample to be valid.
The Jacobian Conjecture, stated simply, asks whether a polynomial map from n-dimensional complex space to itself with a nowhere-vanishing Jacobian determinant must be invertible. It sounds approachable; it has defeated generations of algebraists. A genuine counterexample would not just close the problem — it would overturn intuitions baked into large areas of commutative algebra and complex analysis.
What makes this episode practically significant for builders is the workflow it demonstrates. Tao is not using the AI to do the mathematics for him — he is using it to compress the time needed to orient himself in unfamiliar technical territory, surface the load-bearing assumptions, and formulate sharper questions. That is exactly the use case where current LLMs add the most reliable value: structured exploration of a known body of material, not autonomous proof generation.
The Hacker News threads linked to both the original Claude Fable result (508 comments) and Tao's digestion session (133 comments and 740 points for the ChatGPT share) show the research community actively cross-checking the claim. Independent verification is still ongoing. If the counterexample holds up under scrutiny, it will stand as the most consequential mathematical result attributed — even partially — to an AI system to date.
For technical readers: watch the verification threads closely rather than treating the claim as settled. The more immediate lesson is methodological — using LLMs to rapidly build working familiarity with dense mathematical or technical material is a legitimate and high-leverage practice, as Tao's public session makes concrete.