Beyond the Grid

What Erdős Actually Asked

The setup is simple enough to explain at a kitchen table. You put some number of points on a flat surface. Between every pair of points, there's a distance. Some pairs are exactly 1 unit apart. The question Erdős asked: if you have n points, what's the maximum number of pairs that can be exactly 1 unit apart?

The answer turns out to be somewhere around n to the power of some exponent. Getting the exponent right is what the problem is about. And for almost 80 years, the consensus was that the square grid, points arranged in a regular lattice, was essentially optimal. The square grid gives you roughly n pairs at unit distance. If you believed the grid was optimal, you believed the answer was n raised to the first power, maybe with some small logarithmic corrections, but basically linear.

Erdős conjectured the answer was n to the power of 1 plus some small positive delta. He was asking whether there existed any arrangement of points that beats the grid. For 80 years, no one could prove it either way.

How a General-Purpose Reasoner Cracked It

OpenAI's announcement is careful about what the model was. It was not a system trained specifically for mathematics. It was not scaffolded with specialized tools for proof search. It was a general-purpose reasoning model, the kind of thing being used simultaneously for code debugging, legal drafting, and customer service. It was pointed at the Erdős problem and asked to reason about it.

What it produced draws on a branch of mathematics that would not have been obvious to most people working on the unit distance problem. The construction relies on ideas from algebraic number theory. Specifically, the model's argument connects the unit distance problem to results due to Ellenberg-Venkatesh on sum-product phenomena, to Golod-Shafarevich on class field towers, and to work by Hajir, Maire, and Ramakrishna on infinite extensions of number fields.

This is not a simplification. That's the point. A human mathematician working on the unit distance problem might reasonably spend years in combinatorial geometry without ever thinking to reach for class field towers. The model connected domains that had not been meaningfully connected before.

Will Sawin, a professor of mathematics at Princeton, published a companion paper the same day, making the exponent explicit. Where the model had shown that some positive delta exists, Sawin pinned down delta ≥ 0.014. The lower bound on the number of unit-distance pairs for n points is now provably n to the 1.014 or higher. The grid is not optimal. Erdős was right.

What External Verification Means Here

Mathematical proof is different from other kinds of scientific evidence. It doesn't require statistical significance thresholds or replication studies. It either goes through or it doesn't. The mathematical community has spent decades developing adversarial peer review precisely because flawed proofs are common, even from eminent mathematicians. Proofs are wrong more often than the public understands.

OpenAI released the full proof text, companion remarks, and a summary of the model's reasoning steps. The proof has been checked by external mathematicians who found it correct. At 125 pages, it is not a trivial thing to verify, but it's not an unusually long proof by modern standards. What's unusual is that a machine wrote it without a human walking it through the relevant literature.

This matters because it speaks to what kind of mathematical capability has actually arrived. There's a meaningful difference between a system that can verify proofs (which AI has been able to do for years), a system that can fill in gaps in partially specified proofs (which is also well-established), and a system that can independently originate a non-obvious solution to an open problem in a domain it wasn't specifically trained for. The third category is what happened here.

One important qualifier: the model was not operating in a vacuum. OpenAI doesn't disclose exactly what was in the model's training data. It's possible, and not particularly surprising, that the model had encountered relevant algebraic number theory literature during pretraining. But working from a corpus of mathematical knowledge to an original proof connecting several prior results in a new way is still, by any reasonable definition, mathematical reasoning. That's what mathematicians do.

The Question the Proof Opens

The Erdős unit distance problem is now resolved in one direction. The square grid is not optimal. Something beats it. What remains unknown is how much better than n^1.014 you can get. The upper bound on the number of unit-distance pairs, the absolute ceiling on how many you can ever have, sits at n^(4/3). The gap between 1.014 and 1.333 is wide. What lives in there is an entirely open question.

The broader context is the list of Erdős problems. Erdős was a famously itinerant mathematician who spent his career distributing problem sets and small cash prizes to anyone who could solve them. Many remain open. The unit distance problem was one of the better known. Its resolution by a machine, on a Tuesday afternoon in May 2026, signals something that mathematicians are still processing.

It's not that AI has become a better mathematician than the best humans. The delta of 0.014 is a lower bound on a lower bound. The field has a long way to go. What's changed is that the threshold for useful autonomous mathematical reasoning has been crossed in a visible, verifiable way. Erdős offered $500 for this problem. He didn't specify what form the solver had to take.