Consider Representative Data Sets — LessWrong