Anthropic said some of the essays the model writes may be informed by "very minimal prompting" or past entries, and has predicted everything from essays on AI safety to "occasional poetry." The company also admitted that the concept might be seen as "whimsical," but is a reflection of its intention to "take model preferences seriously."
另一层更致命的是责任漂移。模型输出参与决策、代理系统参与执行，过失主体更容易在供应链里移动，从部署方漂到集成商，再漂到平台与模型提供者。巴伦指出什么算AI、什么算AI使用在司法与理赔中仍存在解释空间，这会拉长争议、抬高准备金不确定性，也迫使承保条件更前置。
。heLLoword翻译官方下载对此有专业解读
On today’s pod: late drama and penalty shootout chaos in the FA Cup as Tottenham edge London City Lionesses in a 17-penalty epic to set up a quarter-final with holders Chelsea. The panel discusses Lize Kop’s heroics, Spurs’ resilience and what the result means for both clubs.
9点1氪丨语音误关大灯致车祸，领克道歉；OpenAI获1100亿美元融资；米哈游内部通报员工意外离世