HN Debrief

Claude Fable produced a counterexample to the Jacobian Conjecture

  • AI
  • Mathematics
  • Research
  • Developer Tools

The post points to a tweet from Levent Alpöge claiming the Jacobian Conjecture is false, giving an explicit polynomial map from C^3 to C^3 with constant nonzero Jacobian determinant and three distinct inputs that all map to the same output. That is enough to kill the conjecture. You do not need a deep proof to check it. You just compute the Jacobian determinant and plug in the points. That made the conversation unusually concrete for a story this big. People were not arguing over a 200-page paper. They were staring at a short certificate that any competent user of Sage, SymPy, Wolfram, or multivariable calculus could verify.

Treat this as another sign that frontier models are becoming strong at narrow but real mathematical discovery, especially search over structured spaces with fast verification. For teams doing research, the practical move is to build workflows around generation plus rigorous checking, not to wait for a clean story about whether the model or the human "really" did it.

Discussion mood

Amazed and slightly unsettled. Most commenters accepted that the counterexample itself is straightforward to verify and took the result as strong evidence that frontier models are becoming genuinely useful for mathematical discovery, while a smaller but persistent group pressed for prompt traces and clearer accounting of the human contribution before drawing big conclusions about AI capability.

Key insights

  1. 01

    Verification is trivial, discovery is not

    The important distinction is that checking this object is easy once you have it. Computing the Jacobian determinant and confirming that multiple inputs hit the same output is an undergraduate exercise. That lowers the burden of trust in the claim, but it does not make the original search cheap. Confusing those two steps leads people to badly underestimate what was achieved.

    When evaluating AI research claims, separate "can I verify the output fast" from "could I have generated this candidate myself." Fast verification makes these workflows much more credible and much more scalable.

      Attribution:
    • nxtfari #1
    • mswphd #1
    • jibal #1
  2. 02

    It may be a repaired near-counterexample

    One plausible reconstruction is that the model did not conjure the map from nowhere. A shared Claude analysis suggested it may have started from a known false counterexample that only worked away from a pole, then modified that structure to remove the singularity while keeping the bad global behavior. Comments from algebraic geometers found the resulting geometry believable, especially after noticing discriminant structure and the map's degree-3 cover viewpoint.

    The near-term research pattern to watch is not pure one-shot genius. It is models mining the borderland of failed attempts, almost-examples, and geometric reformulations, then turning them into exact witnesses.

      Attribution:
    • pja #1
    • jlev1 #1 #2
  3. 03

    Earlier search intuitions pointed the wrong way

    People with background in the area remembered searches and lower-bound heuristics that seemed to push possible counterexamples toward much higher degree or different setups. The surprise here is not just that the conjecture fell. It is that it fell in dimension 3 with degree 7. That means human priors about where to look were likely steering effort away from the actual weak spot.

    If your research process depends heavily on human taste to narrow search, add a parallel track that ignores your favorite heuristics. Models may pay off most where expert intuition has become overconfident.

      Attribution:
    • koito17 #1 #2
    • NooneAtAll3 #1
  4. 04

    One counterexample collapses related conjectures

    Because the Jacobian, Poisson, and Dixmier conjectures are linked, this is not an isolated hit. A valid counterexample here also knocks out the Dixmier conjecture in the third Weyl algebra and, by comments in the thread, the Poisson side as well. The object is small, but the blast radius is large.

    Watch for cascade effects when AI lands a result in a tightly connected theory stack. A single verified object can obsolete assumptions across several adjacent literatures at once.

      Attribution:
    • pedrocr #1
    • mdurana #1
    • civilchaos #1
  5. 05

    Formal proof is for scaling, not for this case

    Several comments used Lean and other formal tools as the obvious next step, but mathematically sophisticated replies made a sharper point. Formalization adds little confidence for this specific counterexample because the certificate is already easy to check by hand or CAS. The real role of Lean is elsewhere. It gives AI a feedback loop during discovery and will become more important as proofs get longer and less human-readable.

    Use formal methods where they change the economics of search or verification. Do not confuse them with prestige markers. In easy-certificate cases, they are workflow infrastructure, not epistemic rescue.

      Attribution:
    • ibarrajo #1
    • iib #1
    • olmo23 #1

Against the grain

  1. 01

    The tweet does not say Claude did it alone

    The wording of the original post matters. It credits a collaboration with a human who asked about the problem and Fable who worked on it. That is weaker than the headline framing that Claude "produced" the counterexample by itself. It does not undercut the result, but it does undercut stronger capability claims people are smuggling in.

    Be precise about agency in AI-assisted discoveries. If you care about benchmarking capability, insist on the exact workflow before upgrading your beliefs too far.

      Attribution:
    • YeGoblynQueenne #1
  2. 02

    This may involve unreleased internal models

    A speculative but credible concern is that employees at Anthropic or OpenAI may be using internal next-generation models or harnesses, which makes public replication claims muddy. Follow-up comments softened the certainty, but the core point remained. Without model identity, prompt, and tooling, outsiders cannot tell whether this result reflects current public capabilities or private frontier stacks.

    Do not map internal lab demos directly onto what your team can do today. Separate "AI can do this somewhere" from "our current production model can do this reliably."

      Attribution:
    • tristanj #1 #2
    • sailingparrot #1

In plain english

C^3 ↩
Three-dimensional complex space, meaning triples of complex numbers.
CAS ↩
Computer algebra system, software that manipulates symbolic mathematical expressions exactly.
counterexample ↩
A specific example that shows a general claim is false.
degree 7 ↩
For a polynomial map, this means the highest total exponent appearing in any term is 7.
discriminant ↩
A quantity derived from a polynomial that detects when its roots collide or become repeated.
Dixmier conjecture ↩
A major conjecture in noncommutative algebra that is known to be closely connected to the Jacobian Conjecture.
Jacobian Conjecture ↩
A long-standing math conjecture claiming that a polynomial map from n-dimensional space to itself with constant nonzero Jacobian determinant must have a polynomial inverse.
Jacobian determinant ↩
A single number computed from the matrix of partial derivatives of a multivariable function, used to capture local invertibility behavior.
Lean ↩
A proof assistant used to write formal machine-checked mathematical proofs.
pole ↩
A type of singularity where an expression blows up because of division by zero.
reasoning trace ↩
The intermediate chain of thought or summarized internal steps a model shows while working on a problem.
Sage ↩
An open source mathematics software system used for symbolic and numerical computation.
SymPy ↩
A Python library for symbolic mathematics.
Weyl algebra ↩
An algebraic structure important in representation theory and differential operators.

Reference links

Background on the conjecture

Candidate mechanism and mathematical interpretation

Verification and formalization

Related commentary and surveys

Historical and adjacent references