GPT-5 as a proof-exploration tool: what Ernest Ryu's NAG result actually shows

The 40-year-old question about NAG's stability

The problem Ryu tackled is narrow and old. The Nesterov Accelerated Gradient method, introduced by Yurii Nesterov in 1983, speeds up convergence by calculating a function's gradient at a "look-ahead" point rather than the current one. Researchers observed for decades that this added momentum made algorithms dramatically faster without destabilizing them — but no one had produced a proof explaining why.

Ryu, who has spent 15 years in applied mathematics and optimization theory, had tried this before with colleagues and failed. His intuition told him the answer might be simple; humans just hadn't found the path. That framing matters: this was a problem he already understood deeply, not one the model discovered or defined.

The workflow: generate, discard, verify in a new chat

The mechanics OpenAI describes are the most transferable part. Over about twelve hours across three days, Ryu ran nearly a dozen approaches. GPT-5 proposed directions — many wrong, some unexpected — and he assessed each quickly, pivoting on dead ends and pursuing anything promising.

Two habits stand out. First, GPT-5 "often produced arguments that looked plausible but did not hold up on inspection," so Ryu checked every calculation step himself. Second, when asking the model to check its own work, he found more success starting a fresh chat rather than continuing in the same one — feeding results into clean conversations to keep errors from accumulating.

The turning point came when GPT-5 suggested restructuring the equations governing NAG. The suggestion was not correct as written, but Ryu recognized a meaningful structural feature, developed it rigorously himself, and used targeted questions to test viability. That line became the backbone of the proof he wrote out.

What the model was good at, and what it was not

OpenAI is unusually candid about the limits. GPT-5 "was not inventing new mathematical tools and principles"; it was wielding existing ones and pulling equations and ideas from papers slightly outside Ryu's specialty. It could not assemble a complete proof on its own, even though, in Ryu's words, "several of the key steps that ultimately mattered were suggested by GPT-5."

The described strength is exhaustive search: rapidly proposing and discarding variations, and borrowing tools from adjacent subfields — cognitively exhausting work for a human. The described weakness is reliability. The model's value was speed of exploration, not correctness, and speed only paid off because a domain expert was ruling out bad paths in real time.

The overlooked variable: persistence, not just proof

Ryu estimates the work could have taken weeks at a focused pace — but he says that was never going to happen. "After three days of trying hard, I would have given up," he said. Mathematicians constantly decide whether to abandon a problem and move on.

But with GPT-5, he had the sense of making rapid progress, which kept him going. In his view, this created a psychological shift: the steady flow of new ideas made the problem feel within reach longer than it might have otherwise.Montana Labs

This is a subtler claim than "AI solved a math problem." The tool's contribution was partly emotional — a constant supply of fresh directions that changed Ryu's calculus about when to quit. For teams evaluating AI's research value, that effect on human persistence is real but hard to measure and easy to overstate.

The authorship decision sets the precedent this case is really about

The specific implication here is how Ryu chose to credit the model. He deliberately did not list GPT-5 as a co-author, concluding it was used as a tool. But he named it in the paper's title and abstract and explained its contributions throughout, then wrote the final proof and narrative in traditional style himself.

He framed this as a case study for classical mathematicians, and the framing is the point: to persuade skeptics, he conformed to existing scholarly convention rather than inventing new credit categories. The preprint is public and faces 12-18 months of peer review, so the field's verdict is still pending.

For applied teams, the durable lesson is the division of labor Ryu enforced — model for breadth and speed, human for judgment, verification, and accountability — not the headline that a decades-old problem fell. His own summary was blunt: "If you want the model to fail, it will fail." The result depended on someone who could tell the difference.