gwern comments on interpreting GPT: the logit lens

[gwern]

Related layer visualizations: “Looking for Grammar in All The Right Places”.