OpenAI Did Real Math This Time, And Got It Vetted First
OpenAI made this claim once before and got burned. Seven months ago they said GPT-5 had cracked ten open Erdos problems. It turned out the model had just rediscovered solutions that already existed in the literature, the math community called it out, and Kevin Weil quietly deleted his victory post. So when OpenAI says it again, the first instinct is to roll your eyes.
Except this time looks different. A general-purpose reasoning model, not a math-specific system, disproved a conjecture Paul Erdos posed back in 1946 about the planar unit distance problem. For roughly eighty years mathematicians assumed the best constructions looked like tidy square grids. The model found an entirely new family of constructions that beats them, and it got there by pulling in algebraic number theory, a corner of math that has nothing obvious to do with the original question.
What makes it credible is who signed off and when. The same mathematicians who exposed the earlier embarrassment, Noga Alon, Melanie Wood, and Thomas Bloom who runs the Erdos Problems site, vetted this result before the announcement went out. Verification first, press release second. That ordering is the whole story.
The bigger deal is the word autonomously. This was not a solver built for one problem, it was a general reasoning model that found something humans had missed for eight decades. The line between an AI that helps you do research and an AI that does the research itself just moved, and this is the cleanest evidence yet that it moved for real. Details at https://openai.com/index/model-disproves-discrete-geometry-conjecture/
← Back to all articles
Except this time looks different. A general-purpose reasoning model, not a math-specific system, disproved a conjecture Paul Erdos posed back in 1946 about the planar unit distance problem. For roughly eighty years mathematicians assumed the best constructions looked like tidy square grids. The model found an entirely new family of constructions that beats them, and it got there by pulling in algebraic number theory, a corner of math that has nothing obvious to do with the original question.
What makes it credible is who signed off and when. The same mathematicians who exposed the earlier embarrassment, Noga Alon, Melanie Wood, and Thomas Bloom who runs the Erdos Problems site, vetted this result before the announcement went out. Verification first, press release second. That ordering is the whole story.
The bigger deal is the word autonomously. This was not a solver built for one problem, it was a general reasoning model that found something humans had missed for eight decades. The line between an AI that helps you do research and an AI that does the research itself just moved, and this is the cleanest evidence yet that it moved for real. Details at https://openai.com/index/model-disproves-discrete-geometry-conjecture/
Comments