Dwarkesh Patel argues locking in AI safety regimes now is a mistake, because continual learning breaks their core assumption.

His August 7 essay makes a specific structural argument: nearly every current regulation proposal assumes a model is trained once, safety-checked, then deployed frozen. If base models update daily from real work sessions, pre-deployment evaluation becomes a snapshot of something that no longer exists. Miles Brundage picked it up in the context of entity-based rather than model-based frontier regulation. I don't think Patel is arguing for no regulation, and I've seen the piece read that way. He's arguing that the unit of regulation is wrong, which is a harder and more useful claim.