OpenAI just dropped something that could be far more consequential than another benchmark win. The company says an unreleased internal AI model produced 722 mathematics papers grouped into 372 result families, with 162 papers containing Lean formalizations that allow proof steps to be checked computationally. Among the boldest claims is progress on the quasi Riemann hypothesis, a weaker version of one of mathematics’ most famous open problems. What really caught my attention is the reported economics of the process. OpenAI says many of these results came from a single prompt and averaged roughly three hours of ChatGPT Pro compute. That changes the scaling equation. MacroSift Perspective: If meaningful mathematical discovery can increasingly be generated and verified with compute, progress stops scaling only with the number of human researchers and the hours they can work. It begins scaling with chips, inference, tooling and verification infrastructure. Humans still have the crucia...