One interesting bit from Anthropic's emotion concepts work is that there are separate present speaker/other speaker emotional representations, but that these aren't specific to the assistant/human. The way the model represents the assistant, human, and story character emotions when each are active..
Post
is very similar. This is not what I would have expected! I would have thought there'd be a representation of assistant emotional state that was also active on the user turn, or at least during a character dialogue during the assistant turn.
transformer-circuits.pubtransformer-circuits.pub
0 reposts1 like
No replies yet.