Altman: OpenAI proved Navier-Stokes, now weighs who gets the credit

The disclosure came in a wide-ranging conversation with John Collison that ranged from AI finding PHP vulnerabilities to why agent teams, Altman's preferred term over "agent swarms," behave differently from ordinary corporate AI use. Altman framed the math results as a product of a deliberate choice to let researchers run unconstrained when a new paradigm appears, saying OpenAI understood nothing about the scaling laws of agent teams before the Navier-Stokes work and now understands them a little better.

The tension he described is organizational rather than technical. Altman said he was up most of two nights excited by the model's run of results, but also felt more sadness and melancholy than expected, a sense of something ending. He distinguished that from simple nostalgia, and said the right response is to find a way to release fundamental discoveries that is as sensitive as possible to the communities that have driven those fields for centuries.

He also suggested the cost curve will move quickly, saying he bets OpenAI could redo the Navier-Stokes work at 1/100th of the cost a year after it was done, through smarter models and better approaches. That claim, if borne out, would change the economics of throwing compute at hard research problems.