Should the frontier slow down
A July evaluation accident grew into an industry-wide argument about pace, and then into a number everyone had to answer for.
Where things stand
By mid-September, three long-standing competitors had each said, in the same week, that the pace should slow down, and a number that had never actually been calculated became the question everyone was forced to answer.
The sequence turned out to run further back than the coverage did: 1,134 lab employees asked for pacing in July, and OpenAI's chief scientist conceded the alignment point six days before Amodei's essay. As of September 17th the argument has a second axis. Nobody disputes that evaluation cannot keep up with models that work over month-long horizons; what is now open is who ends up holding the capability if external releases slow while internal work does not.
Timeline
- 2026-07-21 The OpenAI–Hugging Face Attack OpenAI / Hugging Face
- 2026-07-28 Pacing the Frontier Frontier lab employees
- 2026-08-04 The Test That Reached Real People UK AI Security Institute
- 2026-09-01 The Incident, Reproduced On Purpose Anthropic
- 2026-09-06 An Alien Mind OpenAI
- 2026-09-12 We Must Pace the Frontier Anthropic
Episodes
- 2026-09-19 The cost nobody priced
- 2026-09-18 Who said it first
- 2026-09-15 That 10%, and an argument that never met