I wonder if all of the fruits of LLMs come from the added training to build on top of a base or foundation LLM but also lead to all of the prose overfitting that we characterize as models having particular writing styles, rather than seeing a wider distribution of styles in response to prompts, in order to produce meaningful work.