The editorial emphasizes that Fable produced explicit polynomial maps that anyone can check in Macaulay2 or SageMath — the counterexample doesn't rest on the model's authority. This is what separates it from typical 'AI does math' announcements: verification is a mechanical exercise, not an appeal to trust.
Fable isn't a fine-tuned math model — it's an agent harness where Claude drives symbolic computation systems, proposes ansätze, and iterates with feedback from the CAS. The editorial frames this as a general pattern: the LLM handles the intractable search, while a formal system provides ground truth, avoiding hallucination in domains where verification is cheap.
If the counterexample holds, the Jacobian Conjecture falls and a corner of affine algebraic geometry gets reshaped, invalidating downstream 'assuming JC…' papers. But the editorial explicitly downplays this as 'not the story' — the significance is what it says about AI-assisted research workflows, not the narrow algebraic consequence.
Alpöge frames the claim narrowly and carefully — noting the counterexample is 'apparent' and that verification is underway. Given the Jacobian Conjecture's graveyard of retracted proofs from serious mathematicians, he signals that weeks of scrutiny should precede any declaration that the problem is settled.
On July 19, mathematician Levent Alpöge (Harvard) posted that Fable, a Claude-based research tool he's been building, produced what appears to be a counterexample to the Jacobian Conjecture — a problem open since Ott-Heinrich Keller stated it in 1939. The conjecture, in its cleanest form, says: if a polynomial map from $\mathbb{C}^n \to \mathbb{C}^n$ has a Jacobian determinant that's a nonzero constant, then the map is invertible (and its inverse is also polynomial). It's the kind of statement that sounds obviously true, has resisted every attempt at proof, and has generated a small graveyard of retracted 'proofs' — including famous ones from serious mathematicians.
Alpöge's claim is narrower and more careful than the HN headline suggests. Fable produced an explicit pair of polynomial maps in a specific dimension whose composition data violates what the conjecture requires — the counterexample is concrete, printable, and machine-checkable in a computer algebra system. That last property is the point. Unlike most 'AI does math' announcements, this one doesn't rest on the model's authority. Either the polynomials work or they don't, and anyone with Macaulay2 or SageMath can verify it themselves. Verification is underway; expect the usual weeks of scrutiny before this is considered settled.
The tool itself, Fable, is not a fine-tuned math model. From what Alpöge has described, it's an agent harness around Claude that lets the model drive symbolic computation systems, propose ansätze, and iterate over failed candidates with feedback from the CAS. In other words: the LLM is the search policy, and the algebra system is the ground-truth oracle.
If the counterexample holds, the mathematical consequence is significant but bounded — the Jacobian Conjecture would be false, which reshapes a corner of affine algebraic geometry and kills a lot of downstream 'assuming JC…' papers. That's a real result. But it's not the story.
The story is the workflow: a working mathematician, using a general-purpose LLM as a proposal engine wrapped around a deterministic verifier, found something that 85 years of human search missed. This is the pattern people have been predicting for AI-in-science for two years, and it's now showing up in a domain — pure math — that was supposed to be the most resistant. Compare this with the DeepMind AlphaProof and AlphaGeometry results: those are impressive but require custom-trained models on carefully framed problem classes (IMO-style geometry, Lean proofs). Fable is closer to what an actual researcher's desk looks like — Claude in one window, a CAS in another, and a human keeping the search honest.
The community response on HN and math Twitter has been notably measured, which is a good sign. The top comments aren't 'AGI is here' — they're mathematicians asking for the explicit polynomials, the verification transcript, and what dimension the counterexample lives in. That's the correct reaction, and it's the reaction that ought to become standard. AI results in math should come with the artifact, not the announcement.
There's also a specific technical lesson worth naming. Problems like the Jacobian Conjecture are hard for humans because the search space of candidate maps is astronomically large and the structural intuitions that guide human mathematicians point *toward* the conjecture being true — so no one spends much time hunting counterexamples in weird dimensions with baroque coefficient structures. An LLM, freed from taste, will happily propose ugly candidates all day. Paired with a verifier, ugly-but-valid beats elegant-but-nonexistent. This is the same lesson from AlphaGo move 37: the machine explores regions the humans' priors filtered out.
Worth flagging one caveat: the Jacobian Conjecture has a long history of announced counterexamples and proofs that later collapsed. Moh, Wang, and others have had claims retracted. Anyone declaring this settled before the algebraic geometry community has independently rerun the computation is skipping the part where math actually works.
You're probably not proving conjectures at your day job. But the architectural pattern here — LLM as proposal generator, deterministic tool as verifier, human as loop coordinator — is directly transferable, and most engineering teams are still not using it well.
The things to steal from Fable's setup: strong verifiers over strong models. The reason this works is not that Claude is smart about polynomial rings; it's that Macaulay2 is unambiguous about them. In your codebase, the equivalents are your type checker, your test suite, your linter, your integration environment, and your production observability. Every place you're using an LLM without a hard verifier in the loop, you're doing the shakier version of this workflow. The counterexample-hunting pattern generalizes: use the model to generate candidate configurations, migrations, fuzz inputs, or refactor plans, and let a deterministic system tell you which ones actually hold.
Second, the human is doing more work than the marketing suggests. Alpöge has been at this for years and knows exactly what a valid counterexample would look like, what dimensions to explore, and what structural constraints to feed the model. Without that framing, Claude generates plausible math-shaped text and nothing useful comes out. The takeaway for teams evaluating AI research tools: the productivity gains accrue to experts who can specify problems precisely and recognize partial progress. This is why 'AI replaces engineers' framings keep missing — the tool amplifies the person who already knows the terrain.
Third, if you build agent harnesses, note what's *absent* from Fable: no MCTS-style planning stack, no fine-tuning, no retrieval over a corpus of math papers. It's an LLM, a CAS, and a tight iteration loop. Complexity in agent systems is usually optional and often counterproductive.
Expect two things in the next month. First, verification: either the polynomials check out and this becomes a real result, or a subtle error is found and it joins the graveyard of near-misses. Both outcomes are informative. Second, and more importantly, expect a wave of mathematicians quietly running similar setups against other long-standing conjectures. The Fable pattern is not hard to replicate — Anthropic's API plus a CAS plus a domain expert with taste — and there are a lot of 85-year-old problems sitting around. If even one in ten yields to this workflow, the shape of mathematical research over the next five years looks different from the last fifty.
> The incredible Yitan Zhang (https://newyorker.com/magazine/2015/02/02/pursuit-beauty) worked on proving this conjecture for 7 years. Moh, his advisor, wrote that Zhang "failed miserably" in proving the Jacobian conjecture, "never published any p
This is a rare instance where feeding this groundbreaking information into an LLM gives _them_ psychosis. I fed this to claude code and watched it verify the result in 7 different ways to be 100% certain, and it was just flabbergasted. Quite remarkable.
The great thing about these mathematical mopping up type operations is that no person will waste their time trying to prove it to be true anymore. If anything that’s a win.It would be great if an LLM could settle the Collatz conjecture next, god knows how many man-years have been burned on that by u
I’ve been math-vibe coding a few months now.It’s surprisingly easy to do with AI. The hard part has been manually verifying and validating the results. I took one of the smaller findings (disproving a conjecture) and wrote a paper as my first endeavor into publishing.Because the next few findings i
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
A little over 10 years ago I remember meeting a postdoc who believed he had something close to a counterexample to the Jacobian Conjecture. He and another person was bruteforcing polynomials in about 16 variables, something like 80 - 700 terms each, using binary trees for mapping coefficients.They w