Sometime near the end of last year, I came across a blog post by Scott Alexander giving an overview of Anthropic’s recent work on language model interpretability.