Full Breakdown
OpenAI’s AI Model Disproves Erdos’s Planar Unit Distance Conjecture
5/22/2026, 3:23:30 AM
Background: An Eight-Decade-Old Geometry Puzzle
In 1946 Hungarian mathematician Paul Erdos posed the planar unit distance problem: given N points on a plane, how many pairs can be exactly one unit apart? Erdos conjectured that the maximum number of such pairs grows only slightly faster than N, and for decades the “square-grid” arrangement was regarded as essentially optimal.
The Breakthrough Model and Its Proof
OpenAI announced that a new general-purpose reasoning model produced an original proof that the square-grid bound is too low. The model identified a previously unknown family of point configurations that yields a polynomial improvement over the traditional arrangement, thereby disproving Erdos’s long-standing conjecture. The proof was generated autonomously, without a system specifically trained for mathematics, and was subsequently validated by several independent mathematicians.
Key Figures and Supporting Mathematicians
- Thomas Bloom – Maintainer of the Erdos Problems website, co-author of the companion paper.
- Tim Gowers – Fields Medalist, co-author of the companion paper.
- Arul Shankar – University of Toronto mathematician who reviewed the work.
- Noga Alon and Melanie Wood – Provided supporting remarks.
- Andrew Rogoyski – Institute for People-Centred AI, University of Surrey.
Data & Technical Insight
- Problem age: ? 80 years.
- The AI’s construction offers a “polynomial improvement,” meaning the advantage grows significantly as the number of points increases.
- The proof does not yet determine the exact asymptotic growth rate of unit-distance pairs; it only shows that the previously conjectured bound is insufficient.
Official Statements & Responses
OpenAI emphasized that the result demonstrates the capacity of general-purpose reasoning models to maintain long, complex chains of inference across mathematical domains. Thomas Bloom highlighted the broader implication for exploring “the cathedral of mathematics.” Andrew Rogoyski described the development as evidence that AI is becoming a “fundamental tool of future scientific research.” Tim Gowers called the achievement “a milestone in AI mathematics.”
Criticism, Prior Overclaims, and Skepticism
Seven months earlier, former OpenAI VP Kevin Weil claimed GPT-5 solved multiple Erdos problems, a claim later retracted after it was shown the model had merely rediscovered existing literature. Critics such as Yann LeCun and DeepMind CEO Demis Hassabis questioned the earlier announcement. The current proof, however, has been independently verified by the mathematicians listed above, addressing prior concerns about overstatement.
Conflicting Reports & Gaps
Sources agree the conjectured bound is disproved, but they differ on the problem’s remaining status: the exact rate at which unit-distance pairs can increase remains unresolved. No source provides a definitive solution to the full asymptotic question.
Verbatim Quotes
- “For nearly 80 years, mathematicians believed the best possible solutions looked roughly like square grids,” — OpenAI, X post
- “An OpenAI model has now disproved that belief, discovering an entirely new family of constructions that performs better.” — OpenAI, X post
- “It’s becoming clear that AI is impacting the world of creative thought and will become a fundamental tool of future scientific research,” — Andrew Rogoyski, Institute for People-Centred AI
- “a milestone in AI Mathematics.” — Tim Gowers, Fields Medalist
- “In my opinion, this paper demonstrates that current AI models go beyond just helpers to human mathematicians – they are capable of having original, ingenious ideas, and then carrying them out to fruition,” — Arul Shankar, University of Toronto
What’s Next
OpenAI suggests the model’s reasoning abilities could be applied to complex problems in biology, physics, engineering, and medicine. Researchers plan to explore further AI-generated constructions in other open mathematical questions, anticipating that similar “unexpected connections” may emerge as AI continues to engage with deep-seated scientific challenges.
