OpenAI said in mid-May that one of its internal models had disproved the Erdős unit distance conjecture, a geometry problem that had resisted human mathematicians for 80 years. The result drew strong praise from leading mathematicians, but it also fit into a broader pattern: AI systems are becoming increasingly capable at mathematics by combining broad recall of prior work with the stamina to test many ideas humans would dismiss as too tedious.
A long-standing problem in discrete geometry
The unit distance problem dates to 1946, when Paul Erdős asked how many pairs of points in the plane can be exactly one unit apart. The question is simple to state but difficult to solve at scale. For small numbers of points, exact answers are known. For larger sets, the problem becomes much harder, and Erdős focused on bounding the maximum rather than pinning down exact values.
Erdős’s intuition was that the best arrangement of many points would not beat a carefully chosen grid by much. Using a grid-based construction and number-theory arguments, he showed a lower bound that grows slightly faster than the number of points. He then conjectured that the true maximum number of unit-distance pairs for n points was only n^(1+o(1)), meaning it would grow more slowly than n^(1+ϵ) for any fixed ϵ > 0 when n is large enough.
How the AI result changed the picture
OpenAI’s model found a different construction that beats the classic grid approach. Rather than relying on a simple square lattice, the AI used a more elaborate pattern based on higher-dimensional structure and algebraic integers, then projected that structure into two dimensions. In effect, it identified a way to pack more unit distances into the same number of points than Erdős’s grid-based lower bound allowed.
The result does not give a full exact formula for the maximum number of unit distances, but it does move the lower bound upward. Human mathematician Will Sawin was able to show that the growth rate is at least n1.014. That is still far below the best known upper bound, which is around n1.333, so the problem remains open in a formal sense.
Why mathematicians were impressed
OpenAI shared the result with several mathematicians before publishing it. Tim Gowers, who won the Fields Medal, said there was “no doubt that the solution to the unit-distance problem is a milestone in AI mathematics.” Daniel Litt, a University of Toronto professor, wrote that “this is the first example of a result produced autonomously by an AI that I find exciting in itself, as opposed to as a leading indicator.”
Even so, the achievement was not a clean break from prior AI work in math. The model did not introduce a wholly new theory or invent an entirely new method. Instead, it extended existing ideas from different mathematical areas in a nonobvious way. Human mathematicians then cleaned up and extended the argument.
How the construction works in broad terms
The old grid-based strategy uses the Pythagorean theorem. If two points differ by integers a and b in the horizontal and vertical directions, then their distance is one when a² + b² = c² after scaling. Erdős used this idea to count many unit distances by choosing grid spacings that line up with lots of integer solutions to that equation.
OpenAI’s model went further by choosing a richer structure than a plain square grid. The method relied on careful arithmetic choices that allow more pairs to fall exactly one unit apart. In the Ars Technica account, the simplified intuition is that the AI found a “clever modification” of the grid that uses more structure than the classical construction.
What this says about AI and math
The breakthrough comes after a rapid rise in math performance from large language models. Three years ago, LLMs struggled with arithmetic. Last year, they began acing high school mathematics competitions. Earlier this year, at the Joint Mathematics Meetings, researchers were already seeing AI help in constrained mathematical settings, though significant human interpretation was still needed to turn outputs into publishable results.
OpenAI’s result suggests a medium-term model of collaboration rather than replacement. AI systems can search widely across prior mathematics, including areas that a human working on a specific problem might not naturally consult. They can also persist through many unpromising proof attempts without getting discouraged. Humans still matter for choosing problems, judging whether a line of attack is worth pursuing, and turning raw output into rigorous mathematics.
Why this problem suited an AI model
- Broad background knowledge: The proof drew on algebraic number theory, a field many mathematicians would not automatically connect to the unit distance problem.
- Willingness to grind: The approach involved trying many strategies that often fail, a task that suits models better than people.
- Search over large spaces: OpenAI indicated that even with the maximum token budget, the internal model solved the problem only half the time, suggesting the success depended on repeated attempts.
Not the end of the story
Although the headline is striking, the unit distance problem is not fully solved. The AI’s contribution improves the known lower bound, but the upper bound remains much higher, leaving a substantial gap. That means mathematicians still do not know the exact asymptotic answer.
The larger question is how far AI can go from here. In the same period, other systems have also solved or improved upon Erdős problems, including work from Google and earlier experiments involving GPT-5.2 and Harmonic’s Aristotle. Soon after OpenAI’s announcement, University of Michigan postdoc Xiao Ma found that GPT-5.5 could also disprove the conjecture with a small hint. That raises a broader possibility: some discoveries may already be within reach of today’s models, even if no one has yet asked the right question.
Source: Original report
Was this helpful?
Last Modified: July 7, 2026 at 9:27 pm
2 views
