Failing safely is the anomaly — LessWrong