> But there isn’t any mechanism inherent in large language models (LLMs) that would seem to enable this

Stopped reading here. What is the mechanism in humans that enables intelligence? You don't know? Didn't think so. So how do you know LLMs don't have the required mechanism?