“The first half of Anthropic's Constitutional AI process involves instruction tuning, where a language model critiques and refines instruction-response pairs based on a set of principles.”