常见问题解答(FAQ)

A: Constitutional AI is an approach by Anthropic that embeds human preferences in a set of written principles, the constitution. The AI model generates responses to prompts, critiques, and revises them based on these principles, and uses this process to further train itself.

A: RLAIF, also known as "scalable oversight," involves using AI models themselves to provide feedback for training, rather than relying on human preferences. This approach is cheaper and potentially more effective.