OpenAI math model generates 722 papers and Lean checked proofs

If even a fraction of these 722 AI written math papers hold up to expert review, this is a meaningful shift in how we think about machine contributions to frontier science. OpenAI is claiming that an internal model, running on what it describes as modest compute, can generate hundreds of families of new or advanced results, some of which are formalized in Lean and therefore mechanically checked. ([therundown.ai](https://www.therundown.ai/news/openai-722-math-papers)) That moves us beyond toy arithmetic and exam problems into an environment where AI systems are proposing and verifying arguments in parts of math that historically took small teams years.

For the race to AGI, this looks like an early version of “AI as junior collaborator” in theoretical research. The compute budget per result matters because it tells us what happens if you multiply these workflows by an order of magnitude in hardware and model scale. If relatively small models can produce this volume of candidate work, then the bottleneck shifts to validation, interpretation and integration into human research programs. That raises new questions about how grants, credit and academic careers adapt when draft conjectures and proofs can be spun up in bulk by proprietary models sitting behind an API.