It is getting to the heart of the problem when the claim made is that "no matter how advanced the model" they can't be 'much more than just "really good autocomplete."'.

Given that they are Turing complete when you put a loop around them, that claim is objectively false.