SAEs are highly dataset dependent: a case study on the refusal direction — LessWrong