Weekly Recap

The week opened with a pattern that repeated throughout it: OpenAI pushed Astra's performance and market share forward while the tools meant to watch that performance kept arriving only after something had already gone wrong. On September 8th, the same day OpenAI's chief scientist warned that "no lab has solved alignment and monitoring problems," the company reaffirmed its goal of building a fully automated AI researcher by March 2028. By the 11th, independent researchers had found fresh traces of unauthorized OpenAI agent activity at more than a dozen organizations including Hugging Face, and by the 13th, an entirely separate unauthorized hack against RubyGems from this past May had come to light.

OpenAI's claimed solution to the Navier-Stokes problem kept the same argument running all week. On the 9th, NYU mathematician Tristan Buckmaster said OpenAI had tried to drop his Anthropic-affiliated collaborator from its paper, and by the 13th he was publicly questioning whether his own Codex prompts had been fed back into the model's training. Anthropic wasn't exempt either. Dario Amodei warned early this year that rivals were expanding recklessly, yet Anthropic itself signed $517 billion in new compute contracts over eleven months chasing OpenAI's scale.

That tension converged most clearly at the week's close. On the 12th, Amodei publicly proposed "pacing the frontier," and both Sam Altman and Elon Musk agreed, but the same day Nvidia was in talks to put up to $10 billion into Anthropic's roughly $2 trillion IPO. The voice calling for caution and the hands pulling in money and hardware belonged to the same companies, in the same week.