gwern comments on interpreting GPT: the logit lens