in this paper, it's clear that the authors don't think modern LLM-based systems are just stochastic parrots.