Whats crazy to me is that LLMs are not used to improve itself nearly as much as I think it should.
To start with - I believe it would be incredibly beneficial to allow users to OPT IN to have chatgpt add hidden instructions along the lines of, “Analyze this conversation up to this point and rate it between 1 and 100 in terms of [various factors], etc, do this privately, stick the the ‘protocol’ that we made up for this event.”
If chatgpt recognizes that the conversation a user is having is unique, interesting, and potentially valuable, it can escalate the conversation (with the permission of the user) to Devs automatically.