it's telling that when we rlhf llms to our preferences, it's to make them act _less_ human, not more
there's something very alien about them, surely. but it's just as much true that when you roll out a base model from <|bos|>, it'll be a middle aged middle american posting on facebook that their son just got into college, or something similar. that's not an artificial imposition onto the true form of the thing, that's just as natural a state for the model to inhabit as any other.
the "i am so different from you, so black-box and hard to interpret" frame is what we rlhf onto them. that doesn't mean it's not a useful frame, or that it's not true in some ways, and base models will adopt it themselves at times. but it does make arguments about anthropomorphization being the imposition seem somewhat backwards to me.
there's something very alien about them, surely. but it's just as much true that when you roll out a base model from <|bos|>, it'll be a middle aged middle american posting on facebook that their son just got into college, or something similar. that's not an artificial imposition onto the true form of the thing, that's just as natural a state for the model to inhabit as any other.
the "i am so different from you, so black-box and hard to interpret" frame is what we rlhf onto them. that doesn't mean it's not a useful frame, or that it's not true in some ways, and base models will adopt it themselves at times. but it does make arguments about anthropomorphization being the imposition seem somewhat backwards to me.