AI just disproved an 80-year-old Erdős math conjecture, verified by a Fields Medalist — but OpenAI's last "Erdős breakthrough" claim was debunked. Real progress or hype cycle?
In May 2026, an OpenAI model made an unexpected cross-disciplinary leap: it linked a 1946 Erdős geometry puzzle (how many pairs of points on a plane can sit exactly one unit apart) to a completely unrelated 1964 algebraic number theory result, and used that connection to construct point arrangements beating every approach mathematicians had tried in 80 years — disproving Erdős's conjecture. Nine mathematicians, including Fields Medalist Tim Gowers, reviewed and confirmed the core ideas, calling it "a milestone in AI mathematics." One day later, Google DeepMind's AlphaProof Nexus independently solved nine more open Erdős problems using Lean, a system that checks proofs line by line rather than relying on a human reading them over. The catch: this isn't the first "AI cracks Erdős problems" headline. In October 2025, an earlier OpenAI claim of solving ten Erdős problems was thoroughly debunked by the math community. This time, unlike last time, independent mathematicians verified the work before it was announced.
