“The primary goal for AI alignment at Anthropic is not to achieve a perfect theoretical solution, but to make models good enough to prevent terrible outcomes and allow for continued iteration and improvement.”