AskwhoCasts AIJune 5, 202615m

OpenAI Offers A New Policy Blueprint

TL;DR

OpenAI says recursive self-improvement is already showing up: The blueprint calls RSI "potentially the most consequential frontier safety issue of the coming decade" and wants CAISI to make it an urgent monitoring priority.
The document is surprisingly strong by OpenAI standards: Zvi Moshowitz says it "exceeds expectations a lot," praising its support for transparency, democratic governance, state capacity, model evaluations, and compute advantage.
CAISI is the centerpiece, not the NSA: The plan would make CAISI the federal government's main frontier AI safety institution, with mandatory evaluations, independent technical assessments, and public-facing oversight rather than a classified licensing regime.
Preempting state laws is the biggest danger: OpenAI wants a federal framework modeled on California SB 53, New York's RAISE Act, and Illinois SB 315, but Zvi warns that broad preemption could kill stronger future state action while leaving weak or selective federal enforcement.
'Meaningful accountability' has to mean actual power to force compliance: Zvi argues that million-dollar fines, and maybe even billion-dollar fines alone, are not enough for frontier AI risks if regulators cannot actually stop dangerous behavior.
The blueprint is a serious starting point, not a finished answer: Zvi likes the proposal's whole-of-government resilience ideas, from restricting unevaluated frontier models to coordinating internationally on RSI, but keeps asking the hard question: what happens when an evaluation comes back flashing red?

The Breakdown

OpenAI's new frontier AI policy blueprint gets an unusually positive review because it explicitly treats recursive self-improvement as an urgent safety issue and centers federal oversight in CAISI instead of the NSA. The catch is preemption: if Congress blocks state action without creating real enforcement teeth, the whole deal could become a paper shield.