I like this analysis: OpenAI's approach to alignment is about model obedience. Anthropic's approach is about model character and integrity. Anthropic's approach seems more likely to succeed. https://meaningalignment.substack.com/p/model-integrity-and-character