Gradient-descent step-size bounds and novelty

The GPT-5 convex-bound announcement

Model / AI: GPT-5

announced on X as GPT-5 proving new mathematics by improving a gradient-descent step-size condition to 1.5/L; the source paper had already published the sharp 1.75/L months earlier, and the eventual writeup concedes both that gap and that internal models derived the optimal bound from scratch.

Source review pending. Imported from the original notes; linked claims and artifacts have not been re-audited in this restructuring.

• Paper — The GPT-5 convex-bound announcement
• Paper — the source paper