An internal OpenAI model has disproved the Erdős unit distance conjecture, an 80-year-old problem in discrete geometry that human mathematicians failed to crack since Paul Erdős first posed it in 1946. OpenAI announced the result in mid-May and gave several mathematicians early access before publishing their reactions. It is arguably the first time an AI system has produced a full proof resolving a major open conjecture in mathematics.
Tim Gowers, who won the Fields Medal — the most prestigious prize in mathematics — wrote that the solution is a milestone in AI mathematics. Daniel Litt, a mathematics professor at the University of Toronto, said it was the first autonomous AI result he found exciting on its own terms rather than as a leading indicator of future capability. The proof has since been cleaned up and extended by human mathematicians.
The conjecture concerns how many pairs of points can sit exactly one unit apart in a 2D plane. With 5 points and 10 possible pairs, the best arrangement yields 7 unit-distance pairs. Optimal arrangements have been worked out for 6, 7, 8, 9 points and on up through 21 points by Boris Alexeev, Dustin G. Mixon, and Hans Parshall — but the exact answer remains open beyond that.
“there is no doubt that the solution to the unit-distance problem is a milestone in AI mathematics.”— Tim Gowers, Fields Medalist
Key facts
- 01OpenAI announced in mid-May that an internal model disproved the Erdős unit distance conjecture, a problem open since 1946.
- 02Fields Medalist Tim Gowers called the solution a milestone in AI mathematics after reviewing early access to the result.
- 03The proof used a grid with spacing 1/√65, where each point sits one unit from 16 neighbors via 1²+8²=65 and 4²+7²=65.
- 04Erdős, author of more than 1,500 papers, had conjectured the maximum number of unit distances among n points grows as n^(1+o(1)).
- 05Human mathematicians have since cleaned up and extended the AI-generated proof, which combined existing techniques rather than inventing new ones.
Erdős instead asked for upper and lower bounds on the unit-distance count as the number of points n grows large. For a lower bound, he laid points on a grid and tuned the spacing using the Pythagorean theorem. At grid spacing of 1, each interior point has 4 unit-distance neighbors. Shrink the spacing to 1/5 and each point has 12 neighbors, because 0² + 5² = 25 and 3² + 4² = 25 both give integer diagonals.
The OpenAI proof leans on the same construction at larger scale. Choosing c² = 65 — satisfied by 1² + 8² = 65 and 4² + 7² = 65 — and scaling the grid to 1/√65 produces a configuration where each interior point of a 13×13 grid has 16 unit-distance neighbors. Number-theoretic results including Jacobi's two-square theorem govern how dense these solutions can get as c² grows.
“this is the first example of a result produced autonomously by an AI that I find exciting in itself, as opposed to as a leading indicator.”— Daniel Litt, University of Toronto mathematics professor
Erdős conjectured the maximum number of unit distances grows as n^(1+o(1)) — meaning for large enough n, fewer than n^(1+𝜖) for any 𝜖 > 0. His own lower-bound construction landed at n^(1 + C/(log log n)) for some constant C. The open question for 80 years was whether the truth sits near the lower bound, as Erdős predicted, or higher. The OpenAI model showed it sits higher, disproving the conjecture.
Notably, the AI did not invent new mathematics. It applied existing techniques drawn from graph theory, number theory, and discrete geometry to assemble a full proof — combining subfields in a way no human had pinned down. That distinction matters: three years ago LLMs struggled with arithmetic, and only last year did they begin acing high school math competitions. At January's Joint Mathematics Meetings, AI contributions to research still required heavy human interpretation to become publishable theorems.
The result is not a clean break from prior trends, despite the headline framing. It sits on a fast-moving curve. Paul Erdős wrote more than 1,500 papers in his lifetime, the most in history, and built his reputation on simply-stated problems with deep structure — the kind of target where an AI with broad recall and high tolerance for tedious case analysis has a real edge over a human chasing intuition.
Some context tempers the milestone. The AI's proof needed human cleanup and extension before publication, and the techniques it deployed were known. Mathematicians remain better at framing the deeper questions and choosing which problems are worth attacking. A medium-term future where AIs grind through proof strategies while humans set direction is the plausible shape of the field for now.
For OpenAI, the result is a research-credibility win at a moment when the company is under pressure to show that its frontier systems do more than ship consumer features. Mathematics is a domain where outputs are verifiable, and a named open problem closed by an AI is the kind of artifact that travels in academic circles. Whether the same model class can climb from Erdős to the Riemann Hypothesis is a different question — but the curve from arithmetic in 2023 to a Fields Medalist calling something a milestone in 2026 is the data point that matters for anyone modeling where this goes next.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
