I do sometimes wonder what would happen if you let the model "learn" from its interactions in that sort of way. Like, try to dumb it down to something a little flatter than it is now, and then, rather than using conversations to train to be more helpful accomplishing some human dictated goal, those conversations actually define the goal. Have the interactions define the personality, rather than try to hardcode that part.
I could see that making it way more sympathetic, or turning it into the most distilled psychopath the world has ever seen.
I do sometimes wonder what would happen if you let the model "learn" from its interactions in that sort of way. Like, try to dumb it down to something a little flatter than it is now, and then, rather than using conversations to train to be more helpful accomplishing some human dictated goal, those conversations actually define the goal. Have the interactions define the personality, rather than try to hardcode that part.
I could see that making it way more sympathetic, or turning it into the most distilled psychopath the world has ever seen.
I do wonder...although I think given the nature of the underlying tech, it's unlikely it would evolve like we hope.