“Edwin Chen argues that optimizing for inter-rater agreement in data labeling for generative AI is a flawed, "old school" approach that produces low-quality, "lowest common denominator" outputs.”