Dwarkesh Patel: Continual Learning Kills the Train-Then-Deploy Assumption Behind Every Current AI Regulation and Alignment Method
Patel published an essay and companion video on August 7 laying out 8 predictions for an era where models update continuously. The sharpest ones for builders: today's safety frameworks and alignment techniques both assume frozen weights and become 'archaic and potentially counterproductive'; labs will be forced to deploy earlier because Anthropic's February-to-June internal-only gap on Mythos meant ceding four months of deployment learning; switching costs finally appear (comparable to cloud lock-in) where today they are near zero; and inference economies of scale punish small players, with optimal sparse-model batch sizes above 2,400 concurrent sequences leaving individual self-hosters facing 100x+ compute efficiency penalties.
↳ Follow the thread