I don’t think LLMs contradict the Pinker description - it just turns out that the output stream embodies a lot of information, and you can construct a model (of some reasonable fidelity) of the original sources from sufficient examples of the output stream.