An AI settles a problem mathematicians chased for decades
OpenAI shared what it bills as a genuine milestone: an internal model autonomously resolved a famous open question in combinatorial geometry - the planar unit distance problem first posed by Paul Erdos in 1946, which asks how many pairs among n points in the plane can be exactly distance 1 apart.
What was actually proven
For decades the prevailing belief was that rescaled "square grid" constructions were essentially optimal, and Erdos conjectured an upper bound just barely above linear growth. The model disproved that conjecture, constructing an infinite family of configurations that do measurably better - on the order of n^(1+delta) for a fixed positive exponent. The original proof didn't pin down the exponent, but a follow-up refinement from a Princeton mathematician showed you can take delta = 0.014. External mathematicians checked the work and wrote a companion paper laying out the argument and its significance.
Why the math community is paying attention
Two things make this land harder than a typical result:
- It's described as the first time a prominent open problem central to a subfield has been solved autonomously by AI - and the proof came from a general-purpose reasoning model, not one trained specifically for math, scaffolded to search proof strategies, or aimed at this particular problem.
- The method was a genuine surprise: it pulls deep tools from algebraic number theory - generalizing the Gaussian integers to richer number fields, using machinery like infinite class field towers - to attack an elementary geometric question nobody expected them to touch.
The endorsements are notable. Fields medalist Tim Gowers called it a milestone in AI mathematics and said he'd have recommended a human-authored version for a top journal without hesitation; number theorist Arul Shankar argued it shows models going beyond helpers to having original ideas and carrying them through to completion.
The bigger takeaway
OpenAI is candid that the point is bigger than this one problem. The same abilities - holding a long argument together, connecting distant areas of knowledge, surfacing approaches experts deprioritized, and producing work that survives scrutiny - transfer to biology, physics, materials science, and ultimately AI research itself. The company frames it as evidence of progress toward more automated research, while stressing that human judgment still chooses the problems and interprets the results - one reason, it argues, that expertise becomes more valuable, not less.