Terence Tao and ChatGPT: Deconstructing the Jacobian Conjecture Counterexample

Expert-Led AI Collaboration in Advanced Mathematics

A shared transcript of a conversation between Fields Medalist Terence Tao and ChatGPT reveals a high-bandwidth collaboration where the AI acts not as a simple chatbot, but as a symbolic reasoning colleague. By providing specific mathematical constraints and directing the AI toward structural simplifications, Tao uses the model to verify a counterexample to the Jacobian Conjecture and map it to a more intuitive coordinate system.

Symbolic Verification of the Jacobian Counterexample

The core of the exchange focuses on the verification of a polynomial map with a constant nonzero Jacobian, a central element in the study of the Jacobian Conjecture. The AI successfully performs several complex symbolic tasks:

Proving Polynomial Isomorphism

The AI verifies that a specific set $X$ in $\mathbb{C}^5$ is polynomially isomorphic to the affine space $\mathbb{A}^3$. It does this by defining two maps, $\Phi$ and $\Psi$, and symbolically demonstrating that $\Psi \circ \Phi = (id)_{\mathbb{A}^3}$ and $\Phi \circ \Psi = (id)_X$. This confirms there are no hidden restrictions or exceptional loci at the boundary conditions (e.g., $a = 0$).

Constructing the Keller Map

By composing the isomorphism $\Phi$ with a specific map $F(a,b,c,d,e) = (ac, ad + bc, ae + bd, be)$ and removing the constant term $ad + bc = 1$, the AI derives a polynomial map $G: \mathbb{C}^3 \rightarrow \mathbb{C}^3$. The AI then performs direct symbolic differentiation to prove that the determinant of the Jacobian matrix is constant:

$$\boxed{det \frac{\partial (G_1, G_2, G_3)}{\partial (a, y, z)} = -1}$$

Mapping to the Original Counterexample

Finally, the AI demonstrates that this constructed map $G$ is mathematically equivalent to a known counterexample $F_{orig}$ through a series of linear transformations. By setting $a = z_1, y = z_2, z = -\frac{1}{2}z_3$, the AI proves that $F_{orig}$ is simply a composition of $G$ with linear maps $A$ and $B$ ($F_{orig} = B \circ G \circ A$), thereby recovering the original counterexample exactly.

Analysis of Expert Prompting Patterns

Observers of the transcript note that the quality of the AI's output is directly tied to the sophistication of Tao's inputs. Several key patterns emerge from this interaction:

• Avoidance of Leading Questions: Tao avoids yes/no questions, instead using "what" and "why" to force open-ended, analytical responses, which reduces the likelihood of AI sycophancy or hallucinations.
• High-Density Jargon: The prompts use precise mathematical machinery and terminology. This signals the model to operate at a professional level, effectively "meeting the user at their level."
• Iterative Steering: Rather than asking for a solution in one go, Tao guides the AI through a progression of sub-results, using the AI to brute-force symbolic expansions while he provides the strategic direction.

Community Insights on AI as a Research Colleague

The Hacker News community highlighted several theoretical and practical implications of this interaction:

"The LLM is not acting as a tool here... but very much like a colleague... Tao is guiding it to where he wants to go. But also, Tao is actively learning from it and relying on its explaining, analysis, and inference abilities."

The "Unsolved Problem" Paradox

One contributor noted a specific cognitive challenge for LLMs regarding the Jacobian Conjecture: the models possess the mathematical skill to verify a counterexample symbolically, but their training data tells them the conjecture is an unsolved problem. This creates a tension where the model may be hesitant to claim a solution unless explicitly pushed by a user providing the proof structure.

Breaking Disciplinary Silos

There is a suggestion that AI may help mathematicians overcome the extreme specialization of the field. By acting as a bridge between isolated sub-disciplines, LLMs can help experts connect concepts and break down boundaries that would otherwise require years of study in a separate specialization.