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

Forum topic · QianXun · 2026-08-23

Summary

In 1995, Michel Talagrand posed a conjecture asking whether convexity can be created via fixed-degree Minkowski sums in any dimension, offering a $2,000 prize while admitting his guess had no supporting evidence. Thirty-one years later, Minghua Dong and Anthony Song of Caltech and Stefan Tudose of Princeton posted a complete proof on arXiv, solving the problem through an unexpected route: probability theory rather than convex geometry. Their key equivalence states that any 1-degree Gaussian random vector in n-dimensional space can be decomposed into a sum of three standard Gaussian random vectors. Notably, the team initially tried using ChatGPT for assistance; while the model helped push the problem forward, Tudose delivered the decisive proof, which the paper describes as 'more general and theoretical.' The post situates this result within August 2026's wave of AI-math milestones, including OpenAI Astra solving ten open problems with Lean 4 verification and Axiom Math formalizing 246 theorems, arguing that AI is reshaping mathematical workflows rather than replacing mathematicians.

The Conjecture and the Prize

In 1995, Michel Talagrand posed a mathematical question: in any dimension, can convexity be created through fixed-degree Minkowski sums (a geometric operation that pairs and sums every point of two sets to produce a new set)? Talagrand offered a $2,000 reward to whoever could prove it, admitting candidly: "I made this bold conjecture with no real basis at all — it was just a guess. Even when saying it, I felt it couldn't possibly be right."

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

The mathematical community's reaction came in two layers: surprise that Talagrand would actually honor the bounty, and surprise that the proof path ran through probability theory rather than convex geometry itself.

Timeline and Key Facts

1995: Talagrand poses the convexity conjecture with a $2,000 prize
1985/1998: Joel Spencer tightens the lower bound to log N; Wojciech Banaszczyk tightens it to √log N
2025: Another mathematician proves a 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 as the sum of three standard Gaussian random vectors
Three-step path: Song and Dong first reformulated the conjecture as a probability statement → they tried ChatGPT → Tudose joined with a more general proof

The ChatGPT Episode

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

> Initially, Song and Dong attempted to use ChatGPT to find a solution. Although the large model answered some questions and brought them closer to the solution, it was ultimately Tudose who provided the decisive proof. The team did not adopt the results produced in collaboration with ChatGPT. In their paper, they write that Tudose's proof is "more general and more theoretical."

This deserves unpacking.

First, AI was not useless in solving the conjecture, but its role was "pushing the problem one step forward." It can produce constructions, lemmas, and accelerate path exploration — things it is good at. But when the final blow requiring "greater generality and more theory" was needed, the AI's version was replaced by a human mathematician's. This contrasts with OpenAI Astra's August 1 result of 10 open mathematics problems (sorry=0, fully machine-verifiable in Lean 4): Astra represents "AI-led + Lean verification," while the Talagrand proof is "human-led + AI-assisted + peer review."

Second, the proof channel itself is highly unusual. Convexity is a geometric problem where the standard approach is constructing convex sets. Instead, the trio reformulated the statement as a probability-theory proposition — any 1-degree Gaussian random vector in n-dimensional space decomposes into a sum of three standard Gaussian random vectors. Once this equivalence holds, convexity follows. It is a "change the battlefield" flanking maneuver.

Third, the bounty payoff has its own humor. Talagrand initially believed no one could solve it and never really planned to pay. When he transferred the $2,000 to Tudose's team in 2026, his public comments amounted to being proven wrong — the payment was not just a settlement of a mathematical bounty, but a light-hearted retraction of "I felt it couldn't possibly be right."

August 2026: Mathematics Meets AI

The math × AI storyline was rewritten by a chain of events in August 2026:

May: OpenAI solves the Erdős unit distance problem, manually reviewed by 9 top mathematicians
July: Axiom Math (founded by Hong Le-Tong / Lena Hong) completes Lean 4 formal verification of 246 theorems
August 1: OpenAI Astra solves 10 open math problems in one night, total cost $2,000, with independently verifiable Lean certificates
August 7: OpenAI restricts Astra due to "critical" safety concerns
August 17: Axiom Math valued at $1.6 billion, with Ken Ono as founding mathematician
August 18: Talagrand's convexity conjecture solved; three-person team announces the result

Three Parallel Paths

Three things delineate the frontlines clearly:

AI-led path — Astra, AxiomProver-style: AI produces complete Lean 4 proofs, machine-independently verifiable. Advantage: extremely low cost, reproducible. Limitation: the open problem must have a "machine-verifiable" structure.
Human-AI collaboration path — the Talagrand trio, AI mathematical modeling (May's Erdős problem): AI pushes in the middle, but the final blow is human. Advantage: handles propositions that cannot be mechanically verified. Limitation: depends on top experts' attention.
Formalization backfill path — Axiom's 246 theorems, theorem library rewrites: already-published theorems rewritten in Lean 4. Advantage: reproducible results. Limitation: not new mathematical discovery.

All three paths matured in parallel in August 2026. AI is not replacing mathematicians; it is claiming different positions in three workflows: machine-verifiable propositions go AI-led, creative final blows go to humans, and formalization rewrites go to toolchains. This is the counter-narrative to "AGI is about to replace mathematicians" — AI is getting stronger, but the layers of mathematical workflow are also getting thicker.

Returning to Talagrand's 1995 words: "I made this bold conjecture with no real basis at all." Thirty-one years later, he would probably agree there's a second half: "but the bounty I set happens to be exactly enough to buy three mathematicians three nights of coffee."

Sources

NetEase subscription (via Scientific American): Mathematicians solve decades-old mystery: hidden order in high-dimensional randomness (2026-08-23)
Quanta Magazine: 'Huge Breakthrough' in the Math of Imbalance (2026-08-21, Bansal–Jiang background on the Komlós conjecture)
NetEase Tech: 2026's dense "AI siege" on mathematics (hqwc.cn, 2026-08)
Xinzhiyuan / Tencent News: Axiom Math's 246 theorems + Hong Le-Tong + $1.6 billion valuation
Zhidx / Tencent News: Anthropic Fable 5 Jacobi conjecture counterexample

One-Sentence Takeaway

Human mathematicians have not been replaced by AI, but AI has seized the first shot — conjecture reformulation, construction experiments, and peer-review workflows are all accelerating; the final blow remains a "more general, more theoretical" human proof.