“Making AI models more useful for products requires solving safety problems, which aligns the incentives for capability and safety research.”