GPT-5 and the future of mathematical discovery

OpenAI's latest blog post reads less like a product update and more like a lab notebook. UCLA professor Ernest Ryu, an expert in optimization theory, worked alongside GPT-5 to resolve a question that had sat unanswered in his field. The company is framing the result as proof that its models have moved past writing code and summarizing PDFs, into the messier business of pushing human knowledge forward.

Optimization theory is not glamorous. It underpins everything from training neural networks to routing delivery trucks, and its open problems tend to be abstract, technical, and invisible to anyone outside the field. That makes this case interesting precisely because it is small and specific rather than sweeping. Ryu did not ask GPT-5 to hand him a finished proof. He used it as a collaborator, someone to bounce ideas off, check intermediate steps, and suggest directions he might not have tried alone.

That distinction matters more than OpenAI's headline lets on. Mathematicians have spent the last couple of years testing whether large language models can do anything beyond pattern-matching on textbook exercises. Most of that testing has produced mixed results: impressive on contest-style problems, shaky on genuinely novel research. A working mathematician choosing to credit an AI system as a partner in solving a real open question is a different kind of signal, even if it's a single case study picked and published by the company that built the model.

OpenAI clearly wants this story to stand in for a bigger claim: that GPT-5 marks a turning point where AI stops assisting research and starts doing some of it. The honest version of that claim is narrower. One professor, one problem, one model. Whether that generalizes to messier, less well-defined areas of math is still an open question of its own, and probably the more important one.