OpenAI Model Cracks 80-Year-Old Erdős Conjecture
3 min readAn internal reasoning model from OpenAI has autonomously disproved a central conjecture in discrete geometry that the legendary mathematician Paul Erdős first posed in 1946. The result, announced on May 20, lands as the clearest sign yet that frontier AI systems can now produce original mathematics worthy of a top journal.
The 80-Year-Old Erdős Conjecture
The puzzle is famously easy to state. Place a large number of points on a flat sheet of paper. How many pairs of those points can be exactly one unit apart? Erdős believed the answer grew very slowly, essentially like a square grid arrangement, and the so-called planar unit distance conjecture has frustrated combinatorial geometers for decades.
According to OpenAI’s announcement, the system attacked the problem by reaching into algebraic number theory, a branch of math that on its face has nothing to do with dots on a page. Using machinery from Golod-Shafarevich theory and infinite class field towers, the model constructed an entire family of point arrangements with strictly more unit-distance pairs than any grid.
What Happened
The proof, roughly 125 pages long, shows that the number of unit distances among n points can grow as n^(1+δ) for some fixed δ greater than zero. Princeton’s Will Sawin has since refined that exponent to about 0.014, giving the broader mathematics community a concrete figure to inspect. As Scientific American reports, mathematicians including Noga Alon, Melanie Wood, and Thomas Bloom independently checked the argument, and a companion paper has been released to explain it to the wider field.
Fields medalist Tim Gowers called the result “a milestone in AI mathematics,” and several researchers said it would deserve publication in a leading journal even if it had been produced by humans alone.
Why It Matters
Until now, AI models have mostly recited or rearranged known mathematics. The OpenAI Erdős breakthrough looks different because the system stitched together two distant fields on its own and produced a result that working mathematicians did not anticipate. That is the kind of cross-domain reasoning many researchers thought was years away.
The proof still needs full peer review, and a single problem does not make a science. But the pattern is becoming hard to ignore. As reasoning models move from passing benchmarks to publishing potential journal papers, expect the conversation about AI in research, education, and even hiring to shift quickly. Watch for how the formal verification community responds and whether OpenAI releases the underlying model behind this work.