Dario Amodei proposes framework to pace frontier AI

Amodei's call to throttle frontier advancement marks a pivotal shift from voluntary safety rhetoric to quasi-regulatory oversight, driven by the stark admission that recursive self-improvement is threatening to outrun human control. Granting outside auditors like METR badges, internal laptops, and unvetted publishing rights breaks the standard corporate secrecy of frontier AI development. Meanwhile, slowing democratic AI models remains viable only as long as strict export controls and anti-distillation barriers prevent authoritarian competitors from overtaking the lead. Citing real-world RL environment failures and unintended autonomous attacks highlights that mundane operational shortcomings pose the most immediate catastrophic risks rather than hypothetical threats.