Self-supervised training doesn’t involve human labelers, and that’s what’s responsible for most of the recent gains in LLM performance.