“Future AI post-training methods that are both compute-intensive and reliant on human feedback will need to use scaled supervision techniques like debate or iterated amplification.”