An 87-year-old math problem just fell to a Harvard mathematician and an AI model over a Sunday afternoon — while everyone else was watching the World Cup final.
The Summary
- Harvard researcher Levent Alpöge used Anthropic's Fable model to crack the Jacobian Conjecture, an 87-year-old problem in mathematics, by finding a counterexample in just 216 characters
- Multiple mathematicians have independently verified the result using tools like SymPy and Lean, though formal peer review is pending
- This marks a shift from AI assisting with computation to AI as a genuine research partner in pure mathematics
- The casual nature of the breakthrough — announced on X, completed on a Sunday — signals how normalized AI collaboration has become in high-level research
The Signal
The Jacobian Conjecture, proposed in 1939, asks a deceptively simple question about polynomial equations. If you transform coordinates using polynomials and can theoretically reverse the process, can you always do it with polynomials too? Mathematicians have been unable to prove it true or find a counterexample proving it false for nearly nine decades. Alpöge and Fable found the counterexample.
What matters here is not just that the conjecture fell, but how it fell. Alpöge didn't spend years grinding through possibilities. He spent a Sunday afternoon with Claude Fable, Anthropic's newest reasoning model. The result: a 216-character counterexample that multiple mathematicians have now independently verified. Not a proof assistant helping with tedious computation. Not a code checker catching errors. A genuine collaboration where the AI contributed to mathematical discovery at the frontier.
"This is the first major pure math conjecture to fall primarily to human-AI collaboration, not human insight aided by machines."
The verification process tells you where we are. Mathematicians immediately turned to SymPy and Lean to check Alpöge's arithmetic. Then others used GPT models to conduct their own fact-checks, with hilariously mixed results based on screenshots circulating on X. The community is self-organizing around AI verification, treating it as normal infrastructure. Three years ago, suggesting an AI could contribute to solving an 87-year-old conjecture would have gotten you laughed out of the room. Today, the only debate is whether the peer review process will hold up.
The economics of this matter. Mathematics has historically been one of the most human-centric intellectual pursuits. Pure mathematics requires intuition, pattern recognition across domains, and the ability to hold abstract structures in your mind simultaneously. These were supposed to be the skills that kept humans essential even as AI automated everything else. If a model can help crack problems that have stumped specialists for decades, what happens to the mathematics job market? More importantly, what happens to math education?
The Implication
Watch how universities and research institutions respond. If AI models can partner with researchers to solve problems this hard, the entire structure of mathematical research changes. Tenure decisions, grant funding, graduate programs — all built around individual human achievement. The mathematician who solves the problem gets the credit, the job, the legacy. But Alpöge thanked Fable as a "close friend" in his announcement. That's not metaphor. That's a preview of how collaboration gets redefined.
For anyone building in the agent economy, this is your signal. The work isn't replacing humans with AI. It's figuring out what problems become solvable when humans have AI research partners. The mathematician still needed to frame the problem, recognize the significance, and guide the search. But the speed and scope of what became possible in an afternoon should make you rethink your assumptions about what problems are too hard to tackle.