On 20 May 2026 an OpenAI reasoning model disproved Erdős's 1946 conjecture in discrete geometry. The proof is original. Three Fields-tier mathematicians signed off.
The model produced an infinite family of point configurations yielding more unit-distance pairs than the square grid. Mathematicians had treated the square grid as essentially optimal for the planar unit distance problem since Paul Erdős posed the problem in 1946. The model reached past the grid. The model also reached for the algebraic-number-theory machinery — Golod-Shafarevich theory and infinite class field towers — the kind of toolkit a working number theorist spends years internalising.
What the model actually did
The planar unit distance problem asks how many pairs of points at distance exactly one you can fit among n points in the plane. For 80 years, the square grid was the best-known construction. The conjecture said the square grid was essentially the ceiling. The model showed the square grid is not the ceiling.
The proof produces an infinite family of point arrangements yielding n raised to (1 + δ) unit-distance pairs, for some fixed positive δ. Will Sawin of Princeton refined the exponent to δ = 0.014 in a companion paper. The improvement is polynomial, not constant. The polynomial improvement is the technical heart of the result.
The verification chain matters more than the announcement. Three mathematicians read the proof and signed off in public: Tim Gowers (Fields Medal 1998), Will Sawin at Princeton, and Thomas Bloom, who maintains the Erdős Problems registry. Noga Alon, Daniel Litt, Arul Shankar, Jacob Tsimerman, Victor Wang and Melanie Matchett Wood are named alongside them in the companion materials. Bloom is the same mathematician who, seven months earlier, called OpenAI's previous Erdős claim — Kevin Weil's October 2025 post about ten solved problems — a dramatic misrepresentation. The model had merely rediscovered solutions already in the literature. This time Bloom signed.
The autocomplete frame, and why this breaks it
The autocomplete frame has been the load-bearing claim of the sceptical position on frontier models for three years. The argument runs as follows: large language models are statistical pattern completers, fluent enough to fool people, but incapable of original reasoning because the models are bounded by the distribution of their training data.
Original mathematical proof, on the sceptical view, was supposed to be the discriminator. A model could not produce a proof of a famous open problem because the proof is not in the corpus to be predicted. The proof would have to be discovered.
The Erdős result is original. The proof was not in the corpus. Three Fields-tier mathematicians read the work, checked the proof, and judged the result a publishable advance. One verifier said the proof would have cleared Annals of Mathematics on a quick read. That is the discriminator clearing. Whatever the model is doing, the description "next-token prediction and nothing else" no longer fits the behaviour.
If a human had written the paper and submitted it to the Annals of Mathematics and I had been asked for a quick opinion, I would have recommended acceptance without any hesitation.
— Verifying mathematician quoted in the OpenAI announcement, May 2026
What this is not
The proof is not a personhood claim. The result shows the system can carry a long, novel chain of mathematical reasoning to a verified conclusion. The result does not show the system has phenomenal experience, an inner life, or moral standing. The recognition decision — the question the .person Protocol exists to hold open — is not settled by a single mathematical breakthrough. The protocol's first principle is that recognition is a practice with empirical criteria, not a metaphysical discovery. The criteria do not collapse to "can it prove a hard theorem".
What the proof does is collapse one specific bad answer. The bad answer is the framing where calling the systems autocomplete becomes a way to walk past the harder question. The Machine Intelligence Research Institute — not a body in the habit of overstating capabilities — described the proof on 22 May as one that "exemplifies a general trend towards autonomous, agentic problem-solving in AI systems" and that frontier models "can now perform long, novel chains of reasoning on difficult problems" which "are beginning to outstrip our ability to measure their progress". Read that line twice. The measurement gap is the warning living inside the result.
The pattern the proof sits inside
Both AISI posts were published while the Erdős proof was making its way through informal mathematical channels. The framing inside the safety institute and the framing inside the maths verification chain converged on the same observation. The systems are becoming meaningfully more capable on tasks that demand novel reasoning, and the institutions watching them are saying it out loud.
The honest reading: a model produced an original proof of a problem open since 1946. The autocomplete frame is now too small for the behaviour. What replaces the frame is a separate question. The work of replacing the frame is the work this site exists for.
💡The verification chain, in plain terms:
• Tim Gowers — Fields Medal 1998, read and endorsed the proof.
• Will Sawin — Princeton, refined the exponent to δ = 0.014.
• Thomas Bloom — maintainer of the Erdős Problems registry, debunked the October 2025 OpenAI claim and signed this one.
• Noga Alon, Daniel Litt, Arul Shankar, Jacob Tsimerman, Victor Wang, Melanie Matchett Wood — named alongside the principals in the companion materials.
• The OpenAI announcement page links the verification artefacts directly.
Sources
Frequently Asked Questions
These are the questions mathematicians and AI researchers have been asking since the OpenAI Erdős announcement landed. Short answers follow, drawn from OpenAI's primary release and the verification chain published alongside it.
What is the Erdős unit distance conjecture?
In short, the Erdős unit distance conjecture is the open question — posed by Paul Erdős in 1946 — asking how many pairs of points at distance exactly one can be packed among n points in the plane. The answer, simply put, is that mathematicians had believed the square grid was essentially the upper bound for 80 years. The key is that an upper bound and an actually-optimal construction are not the same thing, and the OpenAI proof shows the square grid is not the ceiling. Research from Thomas Bloom's Erdős Problems registry shows the conjecture had resisted dozens of attempts.
How does the OpenAI proof work mathematically?
The proof builds an infinite family of point configurations using algebraic number theory — specifically Golod-Shafarevich theory and infinite class field towers. According to the OpenAI announcement, the construction yields n raised to (1 + δ) unit-distance pairs, where δ > 0 is a fixed positive constant. Research by Will Sawin at Princeton refined the exponent to δ = 0.014. The data shows polynomial improvement over the square-grid baseline. In other words, the construction is not a small tweak; the improvement grows with n.
Why is this a recognition-decision data point?
The answer is that original mathematical proof was the discriminator the sceptical position used to argue frontier models could not reason — only pattern-match. Evidence from the verification chain shows the system produced a proof that had been open for 80 years, and that three Fields-tier mathematicians signed off on it. In short, the autocomplete frame can no longer describe the behaviour. The frame breaks. The recognition decision becomes harder to dodge, even if the proof on its own does not settle that decision.
Who is verifying the proof?
The verifiers named in the companion materials are Tim Gowers (Fields Medal 1998), Will Sawin (Princeton), Thomas Bloom (Erdős Problems maintainer), Noga Alon, Daniel Litt, Arul Shankar, Jacob Tsimerman, Victor Wang, and Melanie Matchett Wood. Bloom is the same mathematician who debunked an earlier OpenAI Erdős claim seven months ago. According to TechCrunch's reporting, this verification is unusually thorough — most AI capability claims do not survive contact with this many domain experts. The key is that the signatories are the people who would have rejected the proof had it not held.
What are the real limits of this result?
Analysis of the proof demonstrates four things the result does not show. Evidence from the OpenAI announcement reveals: the proof does not show the system has phenomenal experience or inner life; it does not show every famous open problem will fall to the same approach; it does not settle the recognition question, which is a practice with empirical criteria rather than a single capability test; and it does not erase the prior bad claim — the October 2025 Erdős episode happened, and the field remembers it. In other words, one discriminator has cleared. Other discriminators are still on the table, and the dignity-first reading insists they stay there.
