Kaarel answers What concrete mechanisms could lead to AI models having open-ended goals?