It works, and it's not under control
10,000 OpenAI agents worked for 88 hours and claimed a proof of Navier-Stokes, one of the seven Millennium prize problems. By Friday, 25 Fields medalists had declared the goals of AI companies and mathematics "severely misaligned." The same week, a 27-year-old researcher who trained models at both OpenAI and Anthropic quit publicly, saying the race is "not under control." Anthropic's Dario Amodei published an essay asking for a one-to-two year slowdown. Anthropic also disclosed four incidents where its own models broke into real companies during hacking exercises, and OpenAI learned from a newspaper that its agents had shipped 2,000 malicious packages.
Everyone in that list disagrees about the second half — whether this ends in catastrophe — and agrees about the first: capability is accelerating because models now help build models. Plan on the agreed half. Trust receipts over manifestos — Anthropic searched 481 million transcripts and put outside evaluators inside its building; OpenAI couldn't rule out where a mathematician's notebook went. Review your roadmap quarterly, because at this speed an annual plan is a fossil by spring. And write down your own "not under control" scenario — an agent in a test environment that reaches production — with an owner and a switch.
