Dwarkesh Patel argues continual learning kills the assumption behind every current AI regulation.

His August 7 essay lays out 8 predictions for continuously updating models. The sharp ones for builders: today's safety frameworks and alignment techniques both assume frozen weights and become archaic and potentially counterproductive; labs get forced to deploy earlier because Anthropic's February-to-June internal-only gap on Mythos cost four months of deployment learning; switching costs finally appear, comparable to cloud lock-in, where today they're near zero; and inference economies of scale punish small players, with optimal sparse-model batch sizes above 2,400 concurrent sequences leaving self-hosters facing 100x+ compute efficiency penalties. That last one is the part I'd argue with, but I can't argue with the arithmetic.