Because if you have an explicit world model you can optimize against it.

It sounds like you are imagining tacking a world model onto an LLM. That's one approach but not what LeCun advocates for.