@hidden imagine thinking AI safety is training AI to be liberal when actual AI safety is "how to safely destroy AI"