Paul Erdős tossed out a geometry puzzle in 1946 and offered $500 to anyone who could settle it. For nearly eight decades the problem stood untouched at its core. Then an unreleased OpenAI reasoning model stepped in. It delivered a disproof. The math world took notice. Yet the story did not end with the announcement.
Will Sawin, a number theorist at Princeton, received an email from OpenAI researchers late on a Friday in May 2026. The message contained the model’s output on the unit-distance problem. Sawin read it. He could not stop thinking about it. By Monday he had written his own paper sharpening the result. The episode reveals how human mathematicians still steer the conversation even when machines produce the first breakthrough.
The question itself sounds simple. Place n points in the plane. How many pairs can sit exactly distance one apart? Erdős guessed the answer hovered near n to the power of one plus a vanishingly small term. Square grids seemed optimal. Decades of work produced an upper bound of roughly n to the 4/3 but no one closed the gap from below. OpenAI’s announcement on May 20, 2026 changed that picture.
The model constructed an infinite family of point sets. For infinitely many n these configurations deliver at least n to the power of 1 plus delta unit-distance pairs. Delta sits above zero, a fixed positive constant. That polynomial improvement kills the conjecture. The proof draws on algebraic number theory, specifically infinite class field towers and ideas linked to the Golod-Shafarevich theorem. It generates number fields of growing degree. Those fields supply richer symmetries than the Gaussian integers that power traditional grid constructions.
OpenAI stressed the model was general-purpose. It had not been trained on this problem or fine-tuned with specialized scaffolding. A single prompt set it loose. The chain of thought ran long. Estimates later placed the internal reasoning at more than 75,000 words. The company released a 19-page companion document filled with remarks from outside experts. Those comments carry weight.
Timothy Gowers, a Fields Medalist, called the work “a milestone in AI mathematics.” He added that if a human had submitted the paper he would have recommended acceptance. Noga Alon of Princeton said the AI achieved what many excellent human researchers had tried and failed to do. Daniel Litt at the University of Toronto found it the first AI-generated result that genuinely excited him on its own terms. Thomas Bloom, who maintains the Erdős Problems website, had sharply criticized an earlier OpenAI math claim in 2025 as a “dramatic misrepresentation.” This time he offered measured support.
Sawin’s reaction proved more personal. He spent the weekend turning the argument over in his mind. The AI’s core idea, allowing the degree of the number field to grow with n, struck him as something humans had simply overlooked. He optimized each step. His arXiv paper delivered an explicit exponent near 1.01. Later refinements pushed it to 1.03. The original model had left the constant implicit. Sawin made it concrete.
He also wrestled with deeper questions. What counts as a proof? The AI output an informal argument. Mathematicians could in principle formalize every line. Sawin judged it sound. Still he wondered how much compute had run in the background. The company has not disclosed full details. That opacity fuels quiet skepticism even among supporters.
Yet Sawin does not dismiss the accomplishment. “AI is a tool,” he told Gizmodo in an article published June 1, 2026. It searches literature efficiently. It checks steps. It persists where humans grow tired. But it does not yet generate the surprising conceptual leaps that define great mathematics. At least not in ways he recognizes as fully novel.
The distinction matters. Previous AI claims in mathematics sometimes repackaged known proofs or relied on massive search rather than insight. OpenAI’s 2025 attempt at another problem collapsed under scrutiny. This time the verification held. Multiple independent mathematicians examined the construction. They found no gaps. The result stands.
Reactions across the community mixed astonishment with caution. Some posted on X that the news felt like science fiction. Others noted the upper bound remains untouched. The problem is only half solved. And the AI did not invent a wholly new branch of mathematics. It recombined existing tools from distant fields. Algebraic number theory and discrete geometry had rarely spoken to each other so directly before.
But recombination at this level still counts. The model spotted a path humans had not walked. It refused to accept the square grid as the final word. That stubbornness echoes Erdős himself, who loved to provoke mathematicians into unexpected directions. Now a machine has returned the favor.
Sawin’s follow-up work illustrates the new division of labor. The AI supplies a seed. Humans cultivate it. They tighten constants. They explore neighboring questions. They debate what the result implies for other Erdős problems that still sit open. One mathematician described the process as AI helping explore the “cathedral” of mathematics and wondering what other unseen wonders wait in the corners.
The broader implications stretch beyond pure math. If general-purpose reasoning models can crack central problems in combinatorial geometry, similar techniques may apply to biology, materials science or optimization challenges in engineering. OpenAI itself highlighted those possibilities. The company positioned the event as evidence that AI can act as a genuine research partner rather than a search engine or calculator.
Of course the partnership remains asymmetric. Humans set the problems. Humans judge the outputs. Humans write the papers that enter the permanent record. Sawin’s quick intervention turned an intriguing preprint into a sharper, citable advance. Without his paper the community might have taken longer to absorb the exact strength of the new lower bound.
Look closer and the episode exposes limits. The model required significant resources to reach its conclusion. Exact figures stay private but the effort exceeded casual experimentation. Future systems may need less. Or they may demand more as problems grow harder. Either way the economics of discovery could shift.
Mathematicians have grown accustomed to computers verifying enormous proofs. The four-color theorem, Kepler conjecture and others passed through massive computation before winning acceptance. This feels different. The AI proposed the argument. It did not merely check someone else’s work.
Gowers captured the unease. “It will become very difficult for humans to compete with AI at solving mathematical problems.” He said it without panic. The tone sounded more like recognition of a new colleague than fear of replacement. Other voices struck similar notes. They welcomed the help. They insisted on verification. They planned the next steps.
Sawin returned to his own research after posting his paper. He views the AI as somewhere between solving the impossible and doing nothing at all. The nuance feels right. The disproof does not rewrite every textbook overnight. It does force the field to redraw one important boundary. And it leaves a trail of new questions that human minds will chase for years.
Additional coverage in recent days has reinforced the human role. The Wall Street Journal reported three days ago on the math community’s mixture of excitement and demand for rigor. It quoted Litt and Gowers at length and emphasized how the AI’s cross-domain synthesis surprised researchers who had grown used to incremental progress. TechCrunch noted on May 20 that this success follows a prior embarrassing claim, underscoring the value of independent verification by figures like Melanie Wood and Bloom.
Nature weighed in on May 22. It described the chatbot producing the disproof after a single prompt and quoted experts astonished that algebraic techniques from number theory provided the key. The coverage across outlets paints a consistent picture. The machine found the counterexample. The mathematicians made it rigorous, explicit and useful.
So the conjecture falls. The bounty goes unclaimed because no one proved Erdős right. Instead a new lower bound enters the literature with an asterisk noting its unusual origin. Future students will learn that an AI model first spotted the construction. They will also read the human papers that followed within days.
Sawin could have left the initial proof alone. He chose otherwise. That choice may define the next chapter of mathematical discovery more than any single bound. Tools get sharper. Questions multiply. Humans remain the ones who decide which ones matter.
WebProNews is an iEntry Publication