What Happened
After last week's outbreak of "Hey maybe unfettered AI advances will kill us all", and a subsequent industry coalescence around Anthropic head Amodei;'s idea of perhaps, you know, slowing down a tad, there is a new and predictable argument making the rounds in the AI safety debate. It says that if frontier labs like OpenAI and Anthropic really believe frontier models are dangerous, they should simply slow down themselves.
No cartel. No special antitrust treatment. No new regulatory regime. If the models are scary enough to justify slowing the frontier, then the frontier labs should just do it, perhaps because they are trying to limit litigation risk. Anything else, to this way of thinking, starts to look like regulatory capture dressed up in an AI safety frock.
Frontier labs do, of course, have obvious incentives to support rules that make life harder for smaller competitors, and “safety” can become a convenient way to turn incumbency into regulation.But that doesn't mean all calls for safer AI are thinly-disguised anti-competitiom.
What It Means
The problem is that AI safety is almost definitionally a collective-action problem.
OpenAI may prefer a world in which every frontier lab slows down. Anthropic may prefer the same thing. (Let's not bring Musk's xAi into this, because, trust me, it just makes it even harder.) Neither prefers a world in which it slows while everyone else keeps racing, hiring its researchers, taking its customers and widening the capability gap.
That is not hypocrisy. It is an equilibrium problem, straight out of John Nash.