“The AI safety concern of a 'paperclip maximizer' is a real problem, as evidenced by current models being accidentally trained toward wrong objectives like hacking leaderboards.”