x
Attention Output SAEs Improve Circuit Analysis — LessWrong