The math that OpenAI keeps for itself

OpenAI published over 700 new mathematical results proved by its latest model, a release that reshapes how people think about machine-led mathematics. A closer look shows clear gaps: cryptography and large areas like statistics, machine-learning theory, optimization, and numerical analysis are conspicuously underrepresented. Those fields account for roughly 20% of arXiv theorems but under 1% of the OpenAI results, a disparity that remains statistically significant even when restricting to papers that solve open problems. Public concerns about withheld breakthroughs in cryptography surfaced immediately, and earlier demonstrations (for example, GPT-5-Pro improving a convex optimization bound in 2025) show models can attack applied-theory problems; yet the published corpus skews toward other domains. The author used programmatic analysis (Semantic Scholar) to quantify these patterns and notes only one result that would have mattered professionally to his previous research focus: a Unique Games-based inapproximability for Max-Cut.

The argument is that mathematics is uniquely easy for AI to scale - synthetic data, cheap verification, and straightforward curricula make compute-to-knowledge conversion efficient - so rapid progress in math should be expected and will feed back into model training. Labs may withhold publishable results to retain competitive advantage, which makes certain high-value fields less visible. The recommendation is urgent: researchers should stop thinking small, learn deep learning and its theory, and prepare ambitious research agendas now so the right questions and safeguards exist when AI-driven mathematics reaches capabilities that can be wielded far faster than human understanding.