Voices
Simon Willison on the Navier-Stokes race: the uncomfortable question is what "improving model performance" means for your Codex sessions
Willison reads OpenAI's admission that it "cannot rule out that de-identified data derived from their usage of our products helped improve our models" as the real story, since Buckmaster and Alpöge had been drafting the project inside Codex sessions. He compares the dynamic to races on security vulnerabilities and raises a new hazard for anyone doing original work with agents: your unfinished problem may be seeding a competitor's solution. He notes OpenAI declined to answer Buckmaster's direct training-data question.
↳ Follow the thread