algon33 comments on interpreting GPT: the logit lens