Dev doubles down: OpenAI could easily target training runs at frontier research ability

ctjlewis · x · 2026-09-10

Continuing the OpenAI math-model data debate, ctjlewis argues there are plenty of ways to customize training runs or feed research agents: give all agents access to frontier research from OpenAI's and Anthropic's people for reference, or test models in training on reproducing recent cutting-edge papers and use that as a boost. "Many many ways, very lazy examples" — the point being there's no technical excuse for not optimizing this, in an all-or-nothing competition OpenAI has now flipped.

Related event: Debate: Feed Frontier Papers to Training or Reproduce from Scratch(2 posts)→

Original post →

More from Companies & People

Companies & People channel →