Agentic AI Outperforms GPU Experts

Agentic optimization (AI agents that run continuous evolutionary search without human-in-the-loop) is now producing kernel and model optimizations that beat most human GPU experts and generalize across related workloads. If robust, this shifts where performance expertise sits — from specialized human engineers to persistent agentic processes running on large compute budgets. — This implies a near-term shift in labor, competition for compute, and who controls performance-critical AI infrastructure, with consequences for jobs, industrial policy, and national security.