tl;dr: Frontier LLMs' headline wins (Navier-Stokes, CVE discovery) come from rare domains with rigorous, machine-checkable specifications, but most knowledge work lacks such specs and writing them is prohibitively expensive—leaving models dependent on human oversight that doesn't scale. As a result, autonomous LLM deployment will be viable mainly for three narrow classes: firms that tolerate cheap failure, narrow guardrailed tasks, and domains already paying for rigorous validation (chip design, drug discovery). Most of these use cases are price-sensitive or favor swarm width over reasoning depth, undermining the frontier labs' valuations premised on drop-in knowledge worker replacement.