“The primary goal of AI alignment should be to make models good enough to prevent catastrophic outcomes and enable continuous improvement, rather than achieving a perfect theoretical solution.”