批评notahacker · 2024-01
“The paper itself is interesting as it covers the limitations (it has big methodological issues)”
公开图灵测试中,GPT-4 最佳提示通过率 49.7%,低于人类 66%
作者让 GPT-4 参加一场公开的在线图灵测试:表现最好的提示在 49.7% 的对局中被判为人类,高于 ELIZA(22%)和 GPT-3.5(20%),但低于人类参与者的基线(66%)。参与者主要依据语言风格(35%)和社会情感特征(27%)判断;论文也讨论了图灵测试作为智能测试的局限,以及伪装成人类的 AI 可能带来的社会后果。
评论者用它说明 GPT-4 在这种简单、非对抗的图灵测试里也还没有通过;另有评论者提醒论文自己也承认存在较大的方法学问题。
“The paper itself is interesting as it covers the limitations (it has big methodological issues)”
“not yet: Does GPT-4 pass the Turing test?”
“GPT4 might not be great at the Turing test either, it seems (Does GPT-4 pass the Turing test? and that is the simple non-adverserial version of the test).”
“In test environments, players are judged human with these rates:”