Human Mathematicians Beat ChatGPT: Three-Person Proof of Talagrand's Convexity Conjecture

In 1995, mathematician Michel Talagrand posed a question: in any dimension, can convexity be achieved through fixed-degree Minkowski sums (the geometric operation of pairing and adding every point of two sets to form a new set)? Talagrand put up a $2,000 reward, saying candidly: "I made that bold conjecture with absolutely no evidence—it was just a wild guess. When you say something like that, you feel it can't possibly be right."

Thirty-one years later, Caltech's Minghua Dong and Anthony Song, together with Princeton's Stefan Tudose, posted the proof on arXiv.

The mathematical community reacted on two levels: disbelief that Talagrand would actually pay out the bounty, and surprise that the proof route ran through probability theory rather than convex geometry itself.

Timeline and Key Facts

• 1995: Talagrand proposes the convexity conjecture with a $2,000 reward
• 1985/1998: Joel Spencer tightened the lower bound to log N; Wojciech Banaszczyk tightened it to √log N
• 2025: Another mathematician proved the stronger version—"convex operations replacing Minkowski sums"—does not hold
• August 2026: Dong, Song, and Tudose release the final proof as an arXiv preprint
• Core equivalent statement: any 1-degree Gaussian random vector in n-dimensional space can be decomposed into the sum of three standard Gaussian random vectors
• Three-step process: Song and Dong first rewrote the conjecture as a probability-theory statement → tried ChatGPT for assistance → Tudose joined with a more general proof

The ChatGPT Episode

The most striking detail: Song and Dong initially tried ChatGPT.

According to their account, the LLM answered some questions and brought them closer to a solution, but it was Tudose who delivered the decisive proof. The team ultimately did not adopt the results produced with ChatGPT. In their paper, they wrote that Tudose's proof was "more general and more theoretical."

This is worth unpacking:

1. AI is not useless in conjecture-solving, but its role is "advancing the problem one step." It can provide constructions, lemmas, and accelerate exploring paths—that's what it's good at. But when the final blow requiring "more generality and more theory" was needed, the AI's version was replaced by a human mathematician. This contrasts with OpenAI Astra solving 10 open math problems (sorry=0, fully machine-verifiable in Lean 4) on August 1: there, it was "AI-led + Lean verification"; here, "human-led + AI assistance + peer review."

2. The proof channel itself is highly unusual. Convexity is a geometric problem; the standard play is constructing convex sets. But the trio rewrote the statement as a probability-theory proposition—any 1-degree Gaussian random vector in n-dimensional space decomposes into the sum of three standard Gaussian random vectors. Once this equivalent statement holds, convexity follows. It's a battlefield-switching, dimension-reducing strike.

3. The bounty payout's irony. Talagrand initially believed no one could solve it and didn't really intend to pay. When he transferred the $2,000 to Tudose's team in 2026, his public comments amounted to admitting he'd been proven wrong—a settlement of the math bounty, and a lightweight retraction of "I felt it couldn't possibly be right."

August 2026: Math × AI Heats Up

• May: OpenAI solves the Erdős unit distance problem, manually reviewed by 9 top mathematicians
• July: Axiom Math (founded by Lکتong “Leyton” Hong) completes Lean 4 formal verification of 246 theorems
• August 1: OpenAI Astra solves 10 open math problems overnight at a total cost of $2,000, with independently verifiable Lean certificates
• August 7: OpenAI locks down Astra over "critical-level" safety concerns
• August 17: Axiom Math valued at $1.6 billion; Ken Ono serves as founding mathematician
• August 18: Talagrand's convexity conjecture resolved; the trio announces the proof

Three clear paths emerge:

• AI-led path (Astra, AxiomProver): AI produces complete Lean 4 proofs, independently machine-verifiable. Advantages: extremely low cost, reproducible. Disadvantage: the open problem must have a "verifiable" structure.
• Human-AI collaboration path (the Talagrand trio; AI-assisted Erdős modeling): AI pushes in the middle, but the final blow is human. Advantages: can handle non-machine-verifiable propositions. Disadvantage: depends on top experts' attention.
• Formalization backfill path (Axiom's 246 theorems, theorem-library rewrites): already-published theorems rewritten in Lean 4. Advantages: reproducible results. Disadvantage: not new mathematical discovery.

All three paths matured in parallel in August 2026. AI is not replacing mathematicians—it's occupying different positions across three workflows: machine-verifiable propositions go AI-led; creative final blows go to humans; formalization rewrites go to toolchains. This is the counterpoint to the "AGI will soon replace mathematicians" narrative—AI is getting stronger, but the layers of mathematical workflow are getting thicker, too.

Back to Talagrand's 1995 remark: "I made that bold conjecture with absolutely no evidence." Thirty-one years later, he might agree there's a second half: "But my bounty amount was just enough to buy three mathematicians three nights of coffee."

One-line Takeaway

Human mathematicians weren't replaced by AI, but AI fired the first shot—conjecture rewriting, construction experiments, and review processes all accelerated; the final, "more general and more theoretical" blow remains human.