Very interesting approach to showing the relationship between RLHF and AI Psychosis: the idea of taking clinical conversations and prompting the model with it seemed like a grounded start. As I'm also investigating AI Psychosis, this approach seems like something to adopt for my work.

It was interesting to read just how much of a "sycophant" effect LLMs can have if they fully lean into the RLFH system.