Sebastien Bubeck is a researcher at OpenAI. He recently shared an experiment where he posed an open problem from a convex optimization paper to GPT-5-Pro, asking it to improve a known bound on step sizes in gradient descent.

The issue was determining when the curve of function values remains convex. The original paper showed it holds below 1/L but fails above 1.75/L, with a gap in between. GPT-5-Pro produced a proof tightening the bound to 1.5/L in just 17 minutes, which Bubeck verified as correct and novel.

Humans later closed the gap fully at 1.75/L, but the AI's independent contribution showed it wasn't relying on memorized data. This raises questions about AI's role in generating original math.

His takeaway is clear: we're entering an era where machines can advance research frontiers, not just apply existing knowledge. Build AI to collaborate on problems, enhancing human discovery without overstepping into creation myths.