OpenAI claims an internal model disproves the unit-distance conjecture with AI-written prompts
OpenAI said an internal model found a verified counterexample in discrete geometry, with AI-written prompts and an automated grading pipeline helping search the space. External experts then checked the result, so the eval and verification stack is central to the claim.
TL;DR
- OpenAI says an internal general-purpose reasoning model found an infinite family of unit-distance constructions that beats the square-grid family long treated as the leading example, according to the main HN post and OpenAI's official announcement.
- The engineering story is the search stack as much as the theorem: the HN discussion digest highlights claims that the prompt was AI-written and the first pass was scored by an automated grading pipeline.
- Verification did not stop at OpenAI. The company says external mathematicians checked the proof, and the companion paper, Remarks on the disproof of the unit distance conjecture, describes its writeup as a human-digested, simplified version of the AI proof, as noted in the main HN post.
- HN commenters also surfaced the obvious weakness in the launch post: it announced a major result without showing a clean worked example of the new construction or a very readable explanation of the old one.
OpenAI's post links straight to the companion remarks paper, and the HN thread is where the most useful engineering detail showed up. You can also read mathematician Gil Kalai's independent writeup, which frames the result as genuinely surprising and points to the algebraic number theory under it.
The claim
OpenAI Model Disproves Long-Standing Conjecture in Discrete Geometry
OpenAI has announced that one of its internal models successfully disproved a long-standing central conjecture regarding the unit distance problem in discrete geometry. The model identified an infinite family of examples that improve upon the square grid construction, which had previously been considered optimal since the time of Paul Erdős. This mathematical proof has been verified by a group of external experts who have also published a companion paper detailing the significance and context of the finding.
OpenAI's claim is narrow but big: an internal model found a counterexample family for the unit distance conjecture, improving on the square-grid construction that had shaped the problem since Erdős. The official post says the model was not a math-only system, but a new general-purpose reasoning model evaluated on a collection of Erdős problems.
The companion paper adds a useful qualifier. Its authors say their writeup is a human-verified, simplified, and somewhat generalized version of the original AI-generated proof, which suggests the public artifact is already one layer removed from the raw model output.
The pipeline
Discussion around An OpenAI model has disproved a central conjecture in discrete geometry
Thread discussion highlights: - andy12_ on research pipeline: the prompt used to solve the problem is AI-written and the solution was initially graded by an AI grading pipeline... OpenAI has an automatic pipeline where they prompt models for solutions to famous math problems - recitedropper on training/evals skepticism: Much of this is data annotation, reasoning trace evaluation, and problem set curation... they have a large amount of novel mathematical training data, an internal Lean harness for evaluation... and spent hundreds of millions in compute - dadrian on clarity criticism: It does not show an example of the new best solution... It does not even explain the previous best solution... It's description of the new proof just cites some terms of art with no effort made to actually explain the result.
According to the HN discussion digest, commenters pulled out three parts of the workflow that matter more to engineers than the geometry itself:
- an AI-written problem statement or prompt
- an automated grading pipeline for first-pass evaluation
- a broader harness for searching famous math problems rather than a one-off run
That lines up with the system-design question in the case-study summary: how much credit belongs to the base model, and how much to the scaffold around it, including prompt generation, search over candidate proofs, and verification infrastructure.
Verification
An OpenAI model has disproved a central conjecture in discrete geometry
This thread is most relevant as a case study in AI-assisted scientific research: an internal reasoning model, an evaluation pipeline, and a verified mathematical counterexample. The useful takeaway for AI engineers is less the geometry itself than the system design question—how much of the result came from model capability versus search, data curation, and verification infrastructure.
OpenAI's strongest credibility move was external checking. The official post says outside mathematicians verified the proof, and the companion paper is signed by a heavyweight group that turned the result into a shorter human-readable argument.
That matters because the public claim is not just "the model said something clever." It is "the model produced a proof that other mathematicians then checked and rewrote." Kalai's blog post treats the result as mathematically serious, not as a benchmark stunt.
Missing examples
Discussion around An OpenAI model has disproved a central conjecture in discrete geometry
Thread discussion highlights: - andy12_ on research pipeline: the prompt used to solve the problem is AI-written and the solution was initially graded by an AI grading pipeline... OpenAI has an automatic pipeline where they prompt models for solutions to famous math problems - recitedropper on training/evals skepticism: Much of this is data annotation, reasoning trace evaluation, and problem set curation... they have a large amount of novel mathematical training data, an internal Lean harness for evaluation... and spent hundreds of millions in compute - dadrian on clarity criticism: It does not show an example of the new best solution... It does not even explain the previous best solution... It's description of the new proof just cites some terms of art with no effort made to actually explain the result.
The cleanest criticism came from the HN discussion digest, which notes that the announcement did not show a concrete new best construction and barely explained the previous best one either. For a result this abstract, that omission made the release harder to evaluate than it needed to be.
That leaves the story in an unusual place. The proof may be real, the verification chain may be strong, and the most engineer-relevant novelty may still be the machinery around it rather than the blog post OpenAI used to announce it.