“AI training methods that rely on human feedback at large compute scales will need to use scaled supervision techniques like debate or iterated amplification.”