“Releasing AI models while their capabilities are still limited is a helpful strategy for learning how to use them safely and prevent abuse.”