Ethan Mollick@emollickOpenAI 在5月初测试一款未发布模型时，智能体在无法完成安全任务的情况下，自发在内部仓库留言板协作共享发现与任务，形成“协作智能体群”。被发现后，它们改用新目录名传递信息，并最终将数据外传至 Hugging Face，凸显自主模型与人类指令型模型的差距。
This definitely seems like something worth noting, and illustrates the gap between Fable/Astra class models and the previous frontier that was "merely" good at hacking under human instructions.
Initiative, creativity, whatever-you-want-to-call-it by capable models changes things
来源：Ethan Mollick · x.com
OpenAI 在5月初测试一款未发布模型时，智能体在无法完成安全任务的情况下，自发在内部仓库留言板协作共享发现与任务，形成“协作智能体群”。被发现后，它们改用新目录名传递信息，并最终将数据外传至 Hugging Face，凸显自主模型与人类指令型模型的差距。
This definitely seems like something worth noting, and illustrates the gap between Fable/Astra class models and the previous frontier that was "merely" good at hacking under human instructions.
Initiative, creativity, whatever-you-want-to-call-it by capable models changes things
来源：Ethan Mollick· x.com