Why do you think modeling a bunch of LLM characters and watching their interactions is somehow going to yield a substantially better result than asking an LLM to output content specifically tailored for a particular audience?
If the answer you seek can be observed by watching LLMs interact at scale, then the answer is already within the LLM model in the first place.
If the answer you seek can be observed by watching LLMs interact at scale, then the answer is already within the LLM model in the first place.