>Just to clarify, when I mentioned "safely used as a supplement", I was thinking from a holistic emotional standpoint. I came across the post on "Chatbot Psychosis" (https://news.ycombinator.com/item?id=46045674) and it got me pondering...
What I think is happening there. A model might only reasonably have 65,000 context. You quickly use it all up and it doesnt tell you it's truncating context, but it defaults in the center.
Eventually your chat is silently removing context and the message you think you're sending the LLM is X, Y, Z and what it's processing is X, Z. Which gives a different answer than the missing Y context.
My thought, use a new chat regularly.