An AI Math Milestone: What OpenAI Disproving Erdős' Unit Distance Conjecture Means

Fri, 22 May 2026 22:21:46 +0800

On May 20, 2026, OpenAI announced an unusual research result: an internal general reasoning model had found a new construction for the planar unit distance problem, overturning an upper-bound conjecture that mathematicians had long believed to be true.

This was not a casual answer from a chatbot. It was a proof produced by OpenAI’s internal general reasoning model during a set of Erdős problem evaluations. The proof has been checked by external mathematicians, and OpenAI also released the proof text, companion remarks, and an edited summary of the model’s reasoning.

What is the problem

The planar unit distance problem was posed by Paul Erdős in 1946. The problem is easy to state: if you place n points in the plane, what is the maximum possible number of pairs of points whose distance is exactly 1?

Mathematicians usually denote this maximum by u(n). If the points are arranged on a line, one can get about n - 1 unit-distance pairs. If they are arranged in a square grid, each point forms unit distances with its vertical and horizontal neighbors, giving roughly 2n such pairs. Erdős also gave a more refined scaled square-grid construction that reaches the order of n^(1+C/log log n) unit-distance pairs.

For a long time, mathematicians broadly believed that these grid-like constructions were close to optimal. The corresponding conjecture can be written roughly as: u(n) should not exceed n^(1+o(1)). Here o(1) tends to 0 as n grows, meaning the number of unit-distance pairs may grow slightly faster than linearly, but should not enjoy a fixed exponent advantage.

OpenAI’s model broke that intuition. It constructed an infinite family of examples: for infinitely many values of n, one can obtain at least n^(1+δ) unit-distance pairs, where δ is a fixed positive constant. OpenAI’s article notes that the original AI proof did not give an explicit value of δ, but Will Sawin later improved the result to allow δ = 0.014.

Why the proof process is special

The most interesting part of this breakthrough is not only the conclusion, but the route of the proof.

Erdős’ early construction can be understood through Gaussian integers. Gaussian integers have the form a+bi; they extend ordinary integers into the complex plane while preserving a property similar to unique factorization. This number-theoretic structure helps explain why certain scaled grids can produce many unit distances.

OpenAI’s model did not keep following ordinary geometric intuition. Instead, it moved the problem into more sophisticated algebraic number theory. According to OpenAI’s explanation, the new proof uses more general algebraic number fields, exploiting their richer symmetry structures to create many differences of unit length and thus produce more point pairs at distance exactly 1 in the plane.

More technically, the proof involves infinite class field towers and Golod-Shafarevich theory. These tools are familiar to researchers in algebraic number theory, but their sudden appearance in a combinatorial geometry problem in the Euclidean plane is what external experts found so illuminating.

The process can be roughly broken into four steps:

Start from the traditional grid construction for the unit distance problem, and translate “the difference between two points has length 1” into a problem about norms and differences in an algebraic structure.
Replace Gaussian integers with more complex algebraic number fields, increasing the number of available unit-length differences.
Use infinite class field towers and Golod-Shafarevich theory to prove that the required number fields exist.
Map the algebraic construction back into planar point sets, obtaining more than n^(1+o(1)) unit-distance pairs for infinitely many n.

In other words, the AI was not simply searching through known proofs. It connected combinatorial geometry with algebraic number theory and proposed a construction outside the dominant human intuition around the problem.

Expert reactions

OpenAI’s article included comments from several mathematicians. Their overall response was strongly positive, though they emphasized different points.

Combinatorialist Noga Alon noted that this was one of Erdős’ favorite problems and that almost every researcher in combinatorial geometry had thought about it. What surprised him was that the correct answer did not fit the long-believed n^(1+o(1)) picture, and that the new construction used advanced algebraic number theory in an elegant way.

Fields Medalist Tim Gowers called the result a milestone for AI mathematics. His judgment was weighty: if the paper had been written by humans and submitted to a top mathematics journal, he would have had no hesitation recommending acceptance. That assessment highlights the quality of the proof, not merely the fact that AI was involved.

Number theorist Arul Shankar focused on model capability. In his view, the paper shows that current AI models are no longer just assistants to mathematicians; they can also propose original and clever ideas and carry them through to complete proofs.

In the companion remarks, Thomas Bloom offered a more cautious standard: the key question in evaluating an AI-generated proof is whether it helps humans understand the problem better. For him, this result gives a careful yes. It suggests that number-theoretic constructions may have a deeper impact on discrete geometry than previously imagined.

These reactions point to the same conclusion: the mathematical community is not accepting the result because “AI did it.” It is accepting it because the proof can be checked, the route explains the problem, and the conclusion genuinely changes the prior understanding.

Does this mean AI is replacing mathematicians

Not yet.

In this case, AI proposed the key construction and proof route, but turning the result into serious mathematics still depended on external mathematicians checking, explaining, and supplementing it. The companion paper also matters: it places the AI proof back into mathematical context, explains why the construction is important, how it relates to existing work, and which problems it may influence next.

A more reasonable conclusion is that AI is beginning to enter the upstream part of mathematical research, but it has not pushed human experts out of the process.

In recent years, AI’s role in mathematics has mostly involved solving contest problems, generating proof drafts, assisting formalization, retrieving references, or rewriting arguments. In those tasks, humans typically still specify the direction. What is different about the unit distance result is that the model faced a long-standing open problem, proposed a new construction, and advanced the argument to a checkable state.

This may change the division of labor in mathematical research. Models may be better at trying many long-chain routes, connecting distant bodies of knowledge, and exploring directions researchers might not prioritize first. The value of human mathematicians will concentrate even more on higher-level questions:

Choosing which problems are worth studying.
Judging whether AI-generated results are trustworthy.
Explaining where a result sits within the field.
Deciding which routes deserve further investment.

Implications for future research

The significance of this event for the AI industry may be even larger than its significance for a single mathematical conjecture.

Mathematics is an ideal setting for testing reasoning ability. Problems are clearly defined, proofs can be checked step by step, and a long argument collapses if a link in the middle fails. If a model can maintain coherence through complex mathematical reasoning and connect tools across fields, similar capabilities may transfer to other areas of research.

OpenAI’s article also extends the implications to biology, physics, materials science, engineering, and medicine. This should not be simplified into “AI will soon make scientific discoveries automatically.” A more realistic change is that AI may first become a route generator and hypothesis amplifier in research: it proposes many possible paths, human experts filter, verify, and explain them, and then push a few valuable paths forward.

This brings three kinds of change.

First, research speed may increase. Many open problems are not unsolved because nobody can understand them, but because there are too many possible routes and the cost of crossing disciplines is high. If AI can continuously propose checkable constructions, it will expand researchers’ search radius.

Second, cross-disciplinary connections will become more common. The unit distance problem belongs to combinatorial geometry, yet the new proof draws on algebraic number theory. Similar “long-distance knowledge transfer” may become a key value of AI research tools.

Third, expert review will become more important. The more routes AI generates, the more reliable verification mechanisms are needed. Mathematics can filter errors through proof checking; other experimental sciences also need experiments, data, reproduction, and safety evaluation. The more AI resembles a researcher, the less human judgment can be skipped.

How this differs from IMO and contest problem solving

In recent years, AI mathematical ability has often been demonstrated through contest problems, such as IMO-level tasks, university mathematics problems, or formal proof benchmarks. These tests are important, but they are not the same kind of event as this unit distance breakthrough.

Contest problems usually have a clear statement, a definite answer, and a relatively bounded solution space. The model’s job is to find a verifiable solution within limited time. Even when the problem is difficult, it remains a “designed problem” and usually has a human problem setter’s expected path behind it.

Open mathematical problems are different. They have no standard answer and no guarantee that existing methods can solve them. Researchers must judge which directions are worth trying, which tools might transfer across fields, and which constructions could be counterintuitive yet viable. This is where OpenAI’s result matters: the model did not merely solve a known problem; it proposed a new construction in a long-standing open problem and changed the original conjectural picture.

So this breakthrough is closer to mathematical research than to a mathematics exam.

Why mathematics is a good test of AI reasoning

Mathematics is a high-pressure environment for testing AI reasoning because fluent expression is not enough to get by.

A mathematical proof must hold at every layer. Experts can inspect whether definitions are accurate, lemmas are applicable, derivations skip steps, and conclusions truly cover the target proposition. If one step in the middle fails, the entire proof fails.

That makes mathematics a better reasoning test than many open-ended writing tasks. A model must not only give an answer that looks plausible; it must produce an answer that survives review. The unit distance problem is especially representative: the conclusion matters, and the proof route can be checked and explained by external mathematicians.

Of course, mathematics is not the only standard. Real-world scientific research also involves experimental error, data quality, equipment constraints, and engineering limitations. But mathematics offers a clear window: if a model can produce a new proof here, it at least shows that its long-chain reasoning and cross-domain connection abilities deserve serious attention.

Why AI proofs still need human mathematicians

An AI-generated proof does not mean human mathematicians can leave the room.

First, proofs need verification. AI-generated arguments may contain gaps, hidden assumptions, or symbolic misuse, and experts must check them. Second, proofs need explanation. Why a result matters, how it relates to existing theory, and what new questions it opens are not automatically settled once a formal proof exists.

Third, proofs need improvement. OpenAI’s original proof did not give an explicit δ; Will Sawin later improved it to allow δ = 0.014. This shows that human experts still compress, clarify, and strengthen the result.

More importantly, mathematical research is not only about “having a proof.” Researchers also judge which routes are valuable, which problems are worth pursuing, and which constructions might transfer elsewhere. AI can expand the search space, but scholarly judgment still requires humans.

What this means for OpenAI’s model direction

From a product perspective, this event suggests that OpenAI’s model direction is shifting from “chat assistants that answer questions” toward “reasoning systems that can participate in complex tasks.”

Chat assistants emphasize dialogue, summarization, writing, and tool use. Scientific reasoning systems must maintain goals over long horizons, combine knowledge from multiple fields, generate verifiable intermediate steps, and organize exploration results in a form experts can review. The unit distance result shows part of that second category.

This also explains why OpenAI published the proof, companion remarks, and model reasoning summary. For research tasks, the final answer is not enough; the process must also be inspectable. Future models for science, engineering, and professional knowledge work are likely to place more emphasis on traceable reasoning, reviewable outputs, and interfaces for expert collaboration.

In other words, models are not merely becoming better conversationalists. They are becoming systems that can share part of the work of research exploration.

How general readers should view this result

This result should neither be mythologized nor dismissed.

It should not be mythologized because AI has not become an independent scientist. This result still needs human mathematicians to check, explain, and improve it, and it still needs to be examined over time by the mathematical community. One breakthrough does not imply that all scientific problems are about to be solved automatically by AI.

It should not be dismissed because it crosses an important threshold. The model did more than repeat knowledge or solve a similar problem from training. It produced a new construction in an open problem, and experts judged that it had mathematical value.

The steadier interpretation is that AI is becoming a powerful collaborator for researchers. It may first change exploration speed, cross-disciplinary connection, and proof drafting, rather than replacing the academic community overnight. For general readers, the key question is not “will AI replace mathematicians?” but “how can humans use AI to expand the range of problems we can study?”

Conclusion

The importance of OpenAI’s result is not only that it overturned a conjecture nearly 80 years old. It also demonstrates a form in which general reasoning models can participate in frontier research: proposing constructions, connecting tools across fields, and producing proofs that experts can review.

It is not the endpoint of an “independent AI scientist,” but it is no longer just a simple problem-solving assistant. In the next few years, mathematics may remain one of the clearest windows for observing AI’s research capabilities: which problems models can advance, which proofs humans need to complete, and which cross-disciplinary connections will be rediscovered are all worth watching.

References:

OpenAI, “An OpenAI model has disproved a central conjecture in discrete geometry”: https://openai.com/index/model-disproves-discrete-geometry-conjecture/
OpenAI proof PDF: https://cdn.openai.com/pdf/74c24085-19b0-4534-9c90-465b8e29ad73/unit-distance-proof.pdf
OpenAI companion remarks: https://cdn.openai.com/pdf/74c24085-19b0-4534-9c90-465b8e29ad73/unit-distance-remarks.pdf
OpenAI model reasoning summary: https://cdn.openai.com/pdf/1625eff6-5ac1-40d8-b1db-5d5cf925de8b/unit-distance-cot.pdf

Combinatorial Geometry on KnightLi Blog