We used sparse autoencoders to explain LLM moderation flags of violent threatsvariance.co6 points·karinemellata··0 commentsOpen articleSaveView on HN