I think he argues more that LLMs are a form of supervision because you need a whole batch of text about specifically what you want the LLM to learn for it to be useful at that one topic. That they predict on human supplied data instead of learning purely from existing in the world and from observations.