Four Ways An Impact Measure Could Help Alignment — LessWrong