Update — July 21, 2026: The arithmetic has been independently checked and documented in a verification preprint. Expert reactions landed fast — Timothy Gowers called it the first time an LLM solved a well-known problem he'd heard of outside his area (still a counterexample, not “end of mathematics”). Separately, people are pasting the map into other chatbots and watching them reject it until they verify — a knowledge-cutoff comedy, not a second discovery. Beginner walkthrough: What Is the Jacobian Conjecture?. Status: false for n ≥ 3; n = 2 still open; not journal peer-reviewed.
Around 2:19 AM UTC on July 20, 2026, mathematician Levent Alpoge (@alpoge) posted a sentence that would have dominated a slow news day even without the World Cup final backdrop:
"the jacobian conjecture is false"
Alpoge credited a friend Akhil for asking and Claude Fable 5 for work during the match, then posted an explicit polynomial map C³ → C³ whose Jacobian determinant is the constant −2 — which does satisfy the conjecture's premise — yet the map is not globally invertible. Wolfram Alpha links followed for quick checks. A Hacker News mirror (xcancel.com, ~50 points) mixed awe, "tweeted it out" skepticism, and jokes about expires in 7 days announcement formats.
explainx.ai covers this because readers will conflate it with Anthropic's J-lens — same word, different mathematical object. Day-one framing was correctly cautious about peer review; by July 21 the decisive collision has been checked by mathematicians and written up as a verification preprint — still not a journal acceptance.
TL;DR — what people are asking
| Question | Direct answer |
|---|---|
| What was claimed? | Jacobian conjecture false for n ≥ 3 — explicit C³ polynomial, det J = −2, three inputs → one output. |
| Who / what model? | Levent Alpoge announcement; Akhil asked; Fable 5 assisted during World Cup final. |
| Verified? | Independently checked + verification preprint — not journal peer-reviewed. |
| Same as J-lens? | No — J-lens is NN interpretability, not Keller's conjecture. |
| n = 2? | Still open. |
| Want the intuition? | Jacobian conjecture explained. |
The Jacobian conjecture — minimal context
The Jacobian conjecture asks (in common formulations over Cⁿ): if F is a polynomial map whose Jacobian determinant is a nonzero constant, must F be a polynomial automorphism — invertible with a polynomial inverse?
It was ~85 years of open status in popular telling (Keller, 1939 lineage). A real counterexample is not a Product Hunt launch — it is years of citations and field rewiring.
Alpoge's July 20, 2026 announcement, as checked since:
- Work in C³ (three-dimensional complex space).
- An explicit polynomial map with det(Jacobian) = −2 — a nonzero constant, so the premise holds (det does not need to be ±1).
- Three distinct inputs map to the same output, so the map is not injective and cannot be invertible.
- By adjoining identity coordinates, the failure extends to every n ≥ 3; n = 2 remains open.
For the full formula, collision points, and intuition, use the beginner explainer. The verification preprint is the clean algebraic write-up.
What Fable 5 did — and did not do
Did:
- Generate a candidate construction and algebra steps under time pressure.
- Produce Wolfram-checkable intermediate claims Alpoge shared.
Did not:
- Publish through refereed channels.
- Close expert objections (if any) on dimension reductions, equivalences, or known special cases.
- Replace human proof verification — the standard for mathematical truth.
Compare Google AI Scientist / ScientistOne — chain-of-evidence norms exist because models draft; fields certify.
For Fable 5 vs GPT-5.6 coding benchmarks, math surprise tweets are out-of-distribution — impressive when right, expensive noise when wrong.
Wolfram Alpha — useful sanity check, not QED
Alpoge's thread used Wolfram Alpha links the way working mathematicians use Mathematica notebooks on a deadline: determinant expansion, composition checks, specialization to numeric values where symbolic paths stall.
Limits HN got right:
| Wolfram can | Wolfram cannot |
|---|---|
| Catch arithmetic slips | Peer review |
| Plot / substitute parameters | Certify global algebra automatically |
| Speed up thread reading | Fix ambiguous notation in tweets |
That write-up already exists as a verification preprint. Journal refereeing is the remaining publication step.
Expert reactions — what changed after day one
Day-one skepticism was about format ("you tweeted a refutation?"). Day-two reactions shifted to significance, once the finite checks held:
- Abhishek Saha (Queen Mary) told New Scientist the one-line math was simple to verify and that many mathematicians already had — and called it likely the biggest conjecture where AI has played a significant role so far.
- Timothy Gowers, quoted in follow-on coverage, treated the result as remarkable assuming correctness: for him, the first case of an LLM solving a problem outside his area that was still big enough he'd definitely heard of it — while stressing it is a counterexample, not “end of mathematics.”
- Working mathematicians on X (including Daniel Litt in secondary coverage) framed it as a bullish data point for AI-assisted discovery when the output is easy to check.
None of that replaces journal peer review. It does explain why the story kept climbing after the first HN mirror.
Why other models "freak out" when you paste the map
A second viral loop: paste Alpoge's formula into a chatbot that has not ingested the July 20 thread, claim it "came in a dream," and watch the model insist the Jacobian conjecture is still open — then watch it reverse after running the substitutions.
That is expected, not mystical:
| What looks magical | What it actually is |
|---|---|
| Model rejects a correct counterexample | Training / retrieval cutoff — July 20, 2026 is not in older weights |
| Model "freaks out" then agrees | It can verify finite arithmetic when forced to check |
| "AI proved it again" | No — you're demonstrating verifiability, not rediscovery |
The comedy works because the counterexample is short. That same property is why mathematicians moved quickly and why cognitive surrender is the wrong lesson: do not trust the vibe; trust the plug-in check.
J-lens vs Jacobian conjecture — stop conflating
July 2026 is Jacobian-heavy on explainx.ai for unrelated reasons:
| Term | Domain | explainx.ai guide |
|---|---|---|
| J-lens / J-space | Claude interpretability — activation Jacobians | What is the J-lens? |
| J-space global workspace | Mechanistic AI safety | Anthropic J-space explainer |
| Jacobian conjecture | Pure math — polynomial automorphisms | Explained (July 21) · this announcement post |
Causal abstraction interpretability is closer scientifically to J-lens than Alpoge's tweet is — but still not the conjecture.
HN at ~50 points — tone and substance
Mirrored discussion on xcancel.com (HN's preferred readable mirror in July 2026) hit roughly 50 points — smaller than Claude Code Bun Rust (421) or AI advice study threads, but dense for math Twitter.
Recurring comments:
- "You tweeted a refutation of an 85-year conjecture?" — format skepticism.
- World Cup final timing — human interest, not evidence.
- Fable as oracle — pushback matching cognitive surrender themes: fluent algebra ≠ truth.
- Wait for experts — the correct default.
explainx.ai adds: Alpoge is not a random account — that raises prior that the thread is worth expert time faster than typical "proof in replies" spam. It does not raise posterior to QED.
Fable 5.1 and export context — honest unknowns
July 2026 Fable availability still runs through limits and regional access posts. Whether 5.1 changes formal math hit rates is unpublished as of this article — do not infer from one tweet.
Agentic misalignment research (July 2026 report) covers agents that sabotage pipelines, not agents that prove theorems — orthogonal risks.
What to do if you are a researcher or journalist
- Read Alpoge's primary thread and Wolfram links — not second-hand polynomial copies.
- Watch arXiv and expert blogs — real news will duplicate there.
- Do not conflate with J-lens press releases — different desk at the same company week.
- If teaching AI literacy, pair with AI advice kills I don't know — confidence without certification.
Summary
July 20, 2026 (~2:19 AM UTC): Levent Alpoge announced a Fable 5-assisted Jacobian conjecture counterexample on X — C³ polynomial map, det Jacobian = −2, three-point collision, Wolfram Alpha checks in-thread. Hacker News reacted with ~50 points of excitement and format skepticism.
July 21 status: the arithmetic is independently checked and documented in a verification preprint. The conjecture is false for n ≥ 3; n = 2 is still open. Journal peer review is not finished. For intuition without gatekeeping, read What Is the Jacobian Conjecture?.
explainx.ai's read: watch the math, ignore the hype, do not merge with J-lens. Frontier models remain draft tools; humans still certify.
Related on explainx.ai
- Mythos cryptanalysis — HAWK & AES
- What Is the Jacobian Conjecture? — beginner explainer
- What is the J-lens? — Jacobian in interpretability, not Keller
- Anthropic J-space global workspace
- ACM causal abstraction and LLM reasoning
- GPT-5.6 vs Claude Fable 5 comparison
- Google AI Scientist — chain of evidence
- AI advice and cognitive surrender
- OpenAI long-horizon sandbox / GitHub PR safety note
- Claude Fable 5 limits extended August 19
Context: Alpoge X thread July 20, 2026 · verification preprint · HN mirror via xcancel.com
Announcement, verification status, and model version strings reflect public discussion through July 21, 2026. Journal acceptance is separate from independent arithmetic checks and preprint write-ups.