x
Future-as-Label: Scalable Supervision from Real-World Outcomes — LessWrong