Looking Deeper at Deconfusion — LessWrong