For something like "a hard limit" to hold, LLMs must be restricted to only reproducing existing text. This is utterly false even for base models - their basin seems to be "permutations loosely inspired by existing text".

And that's before all the post-training comes in.

What's the "limit" there?