Take some comfort, if you like, in the notion that LLMs are limited to emitting convolutions of their training, plus whatever is stuffed into context.