Anthropic's training method for creating helpful, harmless, and honest AI systems through self-supervision.