A University of Pennsylvania statistics professor used OpenAI's GPT-5.6 Sol Pro to disprove a 30-year-old conjecture about the Benjamini-Hochberg method in roughly 90 minutes. The predecessor model, GPT-5.5, failed to solve it even after 20 hours of computation.
The Benjamini-Hochberg method sits at the heart of modern statistics. It controls false discovery rates when researchers test multiple hypotheses simultaneously, a problem endemic to fields from genomics to psychology. The open conjecture about its behavior has resisted human mathematicians for three decades.
GPT-5.6 Sol Pro's approach combined known mathematical techniques in a novel configuration to reach its disproof. The finding keeps intact a larger open question: whether the AI genuinely produced new knowledge or simply recombined existing methods in ways humans hadn't tried.
This distinction matters deeply. If the model merely shuffles existing patterns faster than humans can, it remains a tool for acceleration. If it produces insights that weren't latent in its training data, it crosses into generative discovery.
The speed differential between versions deserves scrutiny. GPT-5.5's 20-hour failure against GPT-5.6 Sol Pro's 90-minute success points to architectural improvements rather than raw parameter scaling. The "Sol" designation suggests specialized training on mathematical content, similar to how domain-specific models outperform generalists in narrow tasks.
Mathematics offers the cleanest testing ground for AI reasoning. Proofs live or die on logical rigor. No ambiguity surrounds correctness. Other domains permit more interpretation, making it harder to distinguish genuine insight from plausible-sounding fabrication.
The real impact depends on replicability. One conjecture disproof, even a significant one, doesn't establish AI as a