By training on public conversations, we set LLMs up for a bias toward what I would call “poor emotional self-regulation” if I was anthropomorphizing them as having emotions the way we do.
When an argument in a comments section occurs and participants talk past one another but continue talking, this creates a lot of training data.
When an argument occurs and one or both participants notice that the discussion is unconstructive and move on with their day, this creates less training data.
How does this intersect with the conversation model in which the LLM is constrained to reply once to each input from the user?
By training on public conversations, we set LLMs up for a bias toward what I would call “poor emotional self-regulation” if I was anthropomorphizing them as having emotions the way we do.
When an argument in a comments section occurs and participants talk past one another but continue talking, this creates a lot of training data.
When an argument occurs and one or both participants notice that the discussion is unconstructive and move on with their day, this creates less training data.
How does this intersect with the conversation model in which the LLM is constrained to reply once to each input from the user?