An OpenAI model has disproved a central conjecture in discrete geometry

## An OpenAI Model Disproves an 80-Year-Old Conjecture in Discrete Geometry

### 1. What Actually Happened

An OpenAI reasoning model produced a proof disproving the central conjecture of the **unit distance problem**, an open problem dating to the 1940s. The conjecture, associated with Paul Erdős, concerns the maximum number of times a unit distance can appear among *n* points in the Euclidean plane. The conjectured upper bound was O(n^(1+ε)) for any ε > 0 — the model produced a counterexample or construction that invalidates this bound, with technical details still under verification by the mathematical community.

This is not an assisted result where a human guides the model step by step: OpenAI describes autonomous resolution, with the model exploring the space of combinatorial and geometric constructions to identify the invalidating structure.

### 2. Why This Specific Problem Matters

The unit distance problem is not an isolated academic exercise. It is structurally connected to several open conjectures in geometric combinatorics, including the Hadwiger-Nelson problem (chromatic number of the plane) and Szemerédi-Trotter bounds on point-line incidences. A disproof here disrupts an entire network of working assumptions used in computational geometry, lattice-based cryptography, and certain discrete optimization problems.

Before this announcement, the best known result was the Spencer-Szemerédi-Trotter bound (1984): O(n^(4/3)) unit distance occurrences. Decades of work had neither pushed below this bound nor proven it optimal. The model appears to have settled the question in an unexpected direction.

### 3. What This Reveals About Current Reasoning Model Capabilities

The signal here is not "AI does math" — it is the **nature of the problem solved**. Discrete geometry problems require a rare combination: constructive intuition (finding the right structure), formal verification (proving the structure has the desired properties), and navigation through an exponential search space with no clear gradient to follow.

Standard benchmarks like MATH or GSM8K have been saturating for several quarters — GPT-4o already reached ~95% on MATH. This result operates in a different register: an open research problem, not present in training data as a solved problem, requiring genuine generalization rather than pattern memorization.

This places the result above olympiad performance (IMO 2024, where models earned gold medals): olympiad problems have known solutions and recognizable structures. Here, there was no solution to memorize.

### 4. Losers and Open Questions

**Direct losers:** Pure mathematics research teams who had spent years on this problem see their comparative advantage reduced on a specific type of task — exhaustive search guided by intuition in combinatorial spaces. PhD students and postdocs whose theses relied on incremental approaches to this problem are in a difficult position.

**Indirect losers:** Mathematics journals and conferences will need to adapt peer review processes. If a model produces a 50-page proof, who verifies it? With what authority? Human verification remains the bottleneck — and it is unclear the community has the tools to absorb an accelerated flow of results of this type.

**Critical open questions:** Has the proof been formally verified (Lean, Coq, Isabelle) or only by human mathematicians? OpenAI had not published the full paper at the time of this announcement. Reproducibility — can the model be re-run to produce the same proof, or a different but equally valid one? — remains to be established.

Finally, the generalization question: is this result the product of a general mathematical reasoning capability, or a specific confluence between training data and the particular structure of this problem? The answer determines whether we are looking at a systematically usable mathematical research tool, or a remarkable but isolated result.