💭 given a well used llm with enough user data I think you could fine tune on a per user basis by mapping user rating and behavior to some sort of user taste embedding and include that as part of the input structure. Then when tuning the model the examples would also have such embeddings and it'd learn what different people want.

Idk if that's valuable. What I want is an llm that produces ideas I think are cool rather than leaning generic and vague. I believe llms have that potential, it's just lost in training.