“Anthropic's Constitutional AI is a modification of the RLHF setup where an AI model provides critiques based on a sampling of constitutional values.”