创建

登录 ReadmeX

或

了解我们如何处理你的信息: 隐私政策

ReadmeX

陪你读懂 AI 动态

今天,想先了解什么?

选一份动态总结,或直接聊聊你的问题。

社区和关注动态总结只属于你自己。

资讯

Reddit 用户用三个笑话测试 Astra 幽默感

AI 总结

一名 Reddit 用户询问 LLM 为何仍不擅长讲笑话,并分享了 Astra Light、Medium 和 High 三个档位生成的三个笑话。作者认为这些例子反映了更广泛的能力短板,但未将其作为受控基准测试,也未独立验证 Astra 消耗 Codex 额度的说法。

深度见解

该帖以社区实验形式考察 Astra 的幽默生成能力,并引发对 LLM 喜剧能力落后的讨论。

I remembered that this was one of my evaluation approaches from 2 years ago to test a model's general capability. I remember wondering in the near future how bad things could get if models could begin to just spit out funny engaging content and set humanity well off on the path to Wall-E future if a cheap slop stream could be just as interesting as all of independent cultural output, and failing that, if slop could be scaled up with no limits on engagement and "perceived quality" that media companies could get people more addicted to their output than ever.

But it seemed like this was always and continues to remain well beyond reach. It's interesting to see that LLMs were able to make such a dent in an edifice so prominent as the field of mathematics and yet basically make no progress on this.

I'm trying to be cheap with my tokens... somehow utilizing Astra even from Chat is using up my codex limits. All I have to offer as examples are 3 examples. They illustrate the point well though.

#GPT Astra#Codex

原文约 3 分钟读完阅读原文
群聊

还没有评论,来说说你的看法。