Will it take a ‘Chernobyl-scale disaster’ for us to regulate AI? | Stuart Russell
TL;DR
Stuart Russell treats Anthropic's recent turmoil as a warning sign: reports of recursive self-improvement, new frontier models and autonomous cyberattack capability make AI risk operational, not theoretical. According to the Guardian piece, the White House imposed export controls on Fable 5 and Mythos 5; Anthropic then took the models offline because the rule also hit foreign staff.
Nauti's Take
Russell lands the uncomfortable point: AI companies want the prestige of aviation, medicine and energy, but often resist the pre-release discipline those sectors live with. The Chornobyl comparison is deliberately heavy and can slide into doom rhetoric.
The cleaner standard is practical: models that can accelerate cyber operations or their own improvement should need licensing, third-party tests and shutdown rules before release. Otherwise safety stays a press line while the public carries the downside.
Briefingshow
The issue is less one model than the feedback loop between coding automation, security gaps and geopolitical control. If AI systems accelerate their own improvement while scaling cyberattacks, after-the-fact liability is too slow. Pre-release licensing becomes infrastructure policy, not tech-sector paperwork.
Sources