Yes—an OpenAI model produced a proof that disproves a prominent conjecture in combinatorial geometry. The result concerns the maximum number of pairs exactly one unit apart among points in the plane. External mathematicians then checked the argument and prepared a human-digested exposition. It is a significant research-mathematics achievement, not evidence that AI can solve mathematics broadly or reliably.
Which mathematical problem did the AI solve?
The result is about the planar unit-distance problem: given n points in the Euclidean plane, what is the largest possible number of pairs whose distance is exactly 1? The points need not form a regular grid, and the question is about maximizing the count over possible arrangements.
As an Amazon Associate I earn from qualifying purchases.
Paul Erdős conjectured that this maximum grows no faster than n1+o(1). In plain terms, that predicts growth close to linear: for every fixed positive amount added to the exponent, the maximum should eventually grow more slowly than n raised to 1 plus that amount. The conjecture was a central, accessible-to-state question in discrete geometry—not a Millennium Prize problem or a claim about mathematics as a whole. OpenAI’s announcement describes its place in the field.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What did the new proof establish?
The construction shows that for infinitely many values of n, there are planar point sets with at least n1+δ unit-distance pairs for a fixed positive δ. That polynomial improvement over linear growth contradicts Erdős’s proposed n1+o(1) behavior.
#1 Best Overall
- Paul Hoffman, The Man Who Loved Only Numbers: The Story of Paul Erdos and the Search for Mathematical Truth, paperback
| Result | What it says |
|---|---|
| Erdős’s conjecture | The maximum count grows as n1+o(1), meaning near-linear growth in the exponent. |
| Earlier lower-bound constructions | Rescaled square grids achieved growth on the order of n1+C/log log n, for a constant C. |
| Previously cited upper bound | O(n4/3), from work by Spencer, Szemerédi, and Trotter in 1984. |
| AI-generated construction | At least n1+δ pairs for infinitely many n, with fixed δ > 0. OpenAI says the original proof did not specify δ; a later refinement by Princeton mathematician Will Sawin obtained δ = 0.014. |
The new result is a lower bound: it shows that some point sets have many unit-distance pairs. It does not determine the exact maximum for every n, nor does it close the gap to the cited upper bound. OpenAI’s explanation and the mathematicians’ companion note give the mathematical context and construction: OpenAI’s announcement and “Remarks on the Disproof of the Unit Distance Conjecture”.
How does the counterexample work?
The surprising ingredient is algebraic number theory. Earlier constructions use rescaled square grids, which can be understood through Gaussian integers. The new construction draws on richer number-field structures: it builds many algebraic numbers of magnitude one and uses them as differences between points. The exposition invokes increasingly large layers of suitable number fields, including infinite class field towers of Golod–Shafarevich type. Those structures make it possible to produce point sets with many distinct differences of length one.
This is a conceptual outline, not a substitute for the proof. The important shift is that the advance is not simply a denser version of the old grid construction; it uses number fields and their symmetries to obtain a stronger lower bound.
Recommended Free Tools
What does “AI solved it” mean in this case?
OpenAI says the proof came from an internal general-purpose reasoning model, rather than a system trained specifically for mathematics or tailored to this conjecture. The companion exposition, by Noga Alon, Thomas F. Bloom, W. T. Gowers, Daniel Litt, Will Sawin, Arul Shankar, Jacob Tsimerman, Victor Wang, and Melanie Matchett Wood, describes its proof as a “human-digested, somewhat simplified, and somewhat generalized version” of the AI proof.
Rank #3
That distinction matters. The model generated the core mathematical argument; external mathematicians scrutinized it and produced an explanation suitable for human readers. The account also says the argument was generated in one shot and its exposition later refined through human interactions with Codex. The cited materials establish expert checking and a human-verified exposition. They do not establish publication in a peer-reviewed journal or formal verification in Lean.
Mathematicians involved in the discussion offered strong but attributed assessments. Tim Gowers, a mathematician and Fields Medalist, said he would have recommended acceptance if asked for a quick opinion on a human-written submission. Thomas F. Bloom’s more qualified assessment was that the result taught mathematicians something new, but only “a moderated yes.” Those are individual judgments, not evidence of unanimity or journal acceptance. OpenAI’s announcement includes their comments.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How does this compare with AI’s math-contest results?
An open research conjecture and a timed contest are different tests. An Olympiad has a fixed set of problems, defined conditions and a scoring rubric; a research problem can require finding a new direction, developing it, and making the argument withstand scrutiny.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →| Example | Reported result | How to interpret it |
|---|---|---|
| 2024 International Mathematical Olympiad | Google DeepMind reported that AlphaProof and AlphaGeometry 2 together solved four of six problems and scored 28 of 42 points—the equivalent of a silver medal. The solutions were scored under IMO rules by prominent mathematicians. | A strong performance on a fixed contest; it is not the same as resolving an open research conjecture. |
| FrontierMath Erdős benchmark | A September 2026 preprint by Adamczewski and Bloom reports that a pre-release GPT-6 Astra resolved two of 68 selected open Erdős problems. Four other evaluated models resolved none. The setup allowed a $300 budget per problem and 72 hours of working time. | A useful, bounded check on claims of broad capability. The benchmark is a preprint and does not directly evaluate the unit-distance result. |
The benchmark authors caution that celebrated examples do not yet provide a systematic account of AI’s mathematical capabilities. They identify concerns including reporting bias, compute disclosure, human scaffolding and possible contamination. The figures describe that benchmark’s selected problems and evaluation setup, not a universal measure of mathematical ability. Sources: Google DeepMind’s IMO account, the 2025 Nature paper on Olympiad-level formal reasoning, and the FrontierMath Erdős preprint.
Best Value
What the result does—and does not—tell us about AI
This is more than an AI system checking arithmetic or assisting with a proof written by a person: the reported contribution was an original construction that overturned a long-standing conjecture, and mathematicians found it worthy of serious scrutiny. It also shows why a striking result should be described precisely. One successful proof does not establish that current models can consistently choose important open problems, solve them, or produce correct arguments without expert review.
The best-supported conclusion is specific: an OpenAI model generated a counterexample construction disproving Erdős’s near-linear conjecture for planar unit distances, and external mathematicians checked and explained the argument. The result is a milestone for AI-assisted mathematics, while the broader reliability question remains open.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →
