Terence Tao — arguably the greatest living mathematician — sat down with ChatGPT and walked it through one of the most stubborn open problems in algebraic geometry: the Jacobian Conjecture. The conversation, shared publicly, is not just a curiosity. It is a stress test of frontier AI reasoning conducted by someone qualified to grade the answers. The results are instructive for anyone building or evaluating AI-assisted tools.

What the Jacobian Conjecture Actually Is

The Jacobian Conjecture, unsolved since 1939, asks a deceptively simple question: if a polynomial map from ℂⁿ → ℂⁿ has a Jacobian determinant that is a non-zero constant everywhere, must the map be invertible? In two dimensions this feels intuitive — a locally invertible smooth map ought to be globally invertible. Polynomial maps, however, are rigid enough that the conjecture seems like it should be provable, yet flexible enough that every proposed proof has eventually collapsed.

Tao has long been interested in whether claimed counterexamples hold up. His ChatGPT session explores whether a particular line of reasoning around a potential counterexample survives scrutiny.

What the Conversation Reveals About AI Mathematical Reasoning

Reading the exchange carefully, a few patterns emerge.

Strengths the model showed:

  • Fluency with standard algebraic geometry vocabulary. It could discuss polynomial maps, the Jacobi criterion, and degree arguments without prompting.
  • Useful as a sounding board. Tao explicitly used it to externalize and probe his own reasoning — essentially talking to a well-read collaborator who never gets tired.
  • Capable of catching surface-level inconsistencies when the logical chain was short enough to hold in a single context window.

Where it struggled:

  • Multi-step deductive chains over three or four logical hops became unreliable. The model would occasionally assert a step as valid when it was not, and only retract the claim after Tao explicitly challenged it.
  • It could not independently verify whether a proposed map genuinely satisfied the Jacobian condition — it reasoned symbolically but could not perform the kind of careful case analysis a graduate student would write out by hand.
  • Confidence was poorly calibrated. Correct and incorrect claims were delivered with similar assertiveness, placing the full burden of verification back on the human expert.

This last point matters enormously. The model does not know what it does not know — at least not reliably enough to be trusted on novel mathematics.

The "Expert-in-the-Loop" Model, Not the "AI Replaces Expert" Model

Tao himself has spoken publicly about viewing AI as a tool that can accelerate mathematical exploration without replacing mathematical judgment. This session is a live demonstration of exactly that dynamic. He used ChatGPT the way a skilled engineer uses a fast prototype: to surface ideas quickly, then apply real expertise to validate them.

For software teams, this is the operative framing. The most productive uses of LLMs in technical work look like this:

  • Drafting — the AI produces a first pass; a senior engineer refines it.
  • Exploration — the AI maps the solution space; the team decides which path to take.
  • Verification assistance — the AI checks for obvious errors; a human checks the AI's checking.

The failure mode is inversion: letting the AI's confident output substitute for expert review on high-stakes decisions.

What This Means for AI-Assisted Development Tools

If a model struggles with multi-hop logical verification in mathematics — a domain with absolute ground truth — it will struggle with analogous tasks in software: proving the correctness of a concurrent algorithm, verifying security properties of an authentication flow, or reasoning about subtle state transitions in a distributed system.

This is not a reason to avoid AI tooling. It is a reason to design workflows that respect the model's actual capability profile:

High-confidence AI tasks          Lower-confidence AI tasks
─────────────────────────────     ──────────────────────────────
Boilerplate generation            Correctness proofs
API surface suggestions           Security boundary analysis
Test case scaffolding             Novel algorithm design
Documentation drafts              Complex refactor planning

Teams that treat the left column as AI territory and the right column as human territory — with AI as a fast first-draft partner — consistently outperform teams that either avoid AI entirely or hand it unchecked authority.

The Broader Signal for SaaS and ML Builders

The Tao session is also a data point in an ongoing calibration exercise the industry is running: how capable are frontier models at genuine reasoning, versus sophisticated pattern-matching over their training distribution?

The honest answer, visible in this exchange, is that the line is blurry and context-dependent. The model performed well on patterns well-represented in mathematical literature. It degraded on genuinely novel inference. This gradient — strong on familiar, weaker on novel — is exactly what you would expect from a system that has learned to predict rather than to reason from first principles.

For ML practitioners building products on top of these models, that gradient is a design constraint, not a dealbreaker. Retrieval-augmented generation, structured output validation, and human-in-the-loop checkpoints are engineering responses to a known limitation. The limitation is not going away in the next model generation; it is shrinking, but slowly.

Why This Matters for Your Project

If you are building a product that exposes LLM reasoning to end users — a legal drafting tool, a code review assistant, a financial analysis platform — the Tao session is a useful calibration reference. The model's failure modes on expert-level mathematics are structurally similar to its failure modes in your domain. Design your human-review checkpoints around those failure modes, not around the model's best-case demos. The difference between a trustworthy AI-assisted product and a liability is almost always in that architecture decision.

Source: Terence Tao's public ChatGPT conversation on the Jacobian Conjecture, shared via Hacker News — https://chatgpt.com/share/6a5fdc7a-d6f8-83e8-bbea-8deb42cfed56