How Anthropic Designed Itself to Avoid OpenAI's Mistakes