OpenAI's Model Just Disproved Erdős's Most Stubborn Geometry Conjecture

Key Takeaways

- Timothy Gowers, Fields Medal winner, reviewed the result and called it "a milestone in AI mathematics," recommending it for the Annals of Mathematics without hesitation. - The unit distance conjecture, first posed by Paul Erdős in 1946, was disproved by an OpenAI reasoning model on May 21, 2026. Marking the first time a prominent open problem was autonomously resolved by an AI system. - The paper (arXiv:2605.20695v1) was independently verified by nine external mathematicians including Will Sawin; no retraction came. - Forty-eight hours later, Google DeepMind's AlphaProof Nexus tackled nine additional open Erdős problems, all formally verified in the Lean proof assistant. - The real bottleneck isn't generating proofs anymore. It's checking them. That shift matters for anyone running AI-powered systems.

---

On May 21, 2026, an OpenAI reasoning model published a paper claiming to disprove the unit distance conjecture. The paper appeared in arXiv (arXiv:2605.20695v1) and showed the full reasoning trace. Nine external mathematicians. Including Will Sawin. Verified the proof independently over the following days. No retraction came. Fields Medalist Timothy Gowers reviewed the result and called it significant.

That's the short version.

Here's why it actually matters.

What the unit distance conjecture asks

Erdős phrased it like a puzzle: place n points on a plane.

How many pairs can sit exactly one unit apart?

Simple question. Brutally hard math.

The intuition says: not too many. Space is big. Angles are weird. You can't pack unit distances efficiently because geometric constraints fight back. Erdős guessed the maximum was around n^1+o(1). Basically linear, with fudge factors that vanish as n grows.

For 80 years, mathematicians believed him. Couldn't prove it.

Nothing contradicted it either.

Then the AI found a construction that blows past that bound.

The approach didn't require new math

The model used algebraic number theory. Specifically: number fields with many symmetries, algebraic numbers of magnitude 1 and bounded denominator, mapped into the plane to generate unit-distance pairs.

It showed infinitely many configurations hitting at least n^1.014 unit-distance pairs.

That's asymptotically better. Not marginal. Not a rounding error. A meaningful gap in the best human results.

Here's the thing.

None of these techniques were new. Human mathematicians knew about number fields. Knew about algebraic conjugates. Knew about the geometry of discrete point sets.

The AI just combined them in a direction nobody had fully explored.

Gowers — Fields Medal winner, serious mathematician. Said he spent an evening convinced the AI had broken something fundamental in math. Turned out: no paradox. Just a construction Erdős hadn't anticipated.

A blind spot in 80 years of human intuition.

That's not mathematicians being dumb.

That's how intuition works. We optimize for what's been productive. We miss the corners nobody's looked at.

The AI had no such loyalty. It just searched.

Nine more problems fell 48 hours later

After OpenAI's announcement, Google DeepMind published that AlphaProof Nexus had tackled nine additional open Erdős problems in a single stretch. Two of those had been unsolved for over 50 years. All nine were formally verified in Lean. A proof assistant that checks the logic mechanically.

Nine problems. Two days. One announcement.

The Erdős problems website lists around 20 claimed solutions sitting in a backlog nobody's audited.

Terence Tao flagged this first: AI generates proofs faster than humans can check them.

The math community is dealing with a verification bottleneck. Your business probably is too.

If you're running AI tools. And if you're reading this, you are. You're sitting on systems capable of research-grade work. The model that cracked Erdős wasn't built for discrete geometry. It was built to reason well. It found the problem, solved it, and handed the result to humans who could check it.

That's exactly what good automation does.

It finds the work. Does the work. Defers to you for the judgment calls.

What you should actually do with this

Three things worth sitting with.

First: general-purpose reasoning models are doing math research now. The same category of model that drafts your reports and answers support tickets just disproved a central conjecture in discrete geometry. You don't need a specialized academic system. You need a capable model and a well-formed question.

Second: the AI didn't invent new mathematics.

It combined existing ideas in a non-obvious way. That means the scarce skill is knowing which existing ideas to combine. And being able to verify the result when the model produces something unexpected.

Third: the verification bottleneck is real and getting worse.

If you're building automation on top of AI outputs. And you are — you need to think about how you know the output is correct. Not "probably correct." Actually correct.

Anyone can run a model. Knowing what to ask it. What to do with the answer. That's the work that stays human.

The conjecture took 80 years to solve. Question: how fast you can figure out whether your AI is right about the next one.

Sources

- OpenAI: Model Disproves Discrete Geometry Conjecture - UnderstandingAI: OpenAI's Milestone Math Breakthrough - Gil Kalai: Amazing Erdos Unit Distance Problem Was Disproved by AI - Arxiv: Remarks on the Disproof of the Unit Distance Conjecture - OpenAI Chain-of-Thought Document (PDF)