HN 提问:Codex 配合 GPT 5.5 Extra High 是否被“降智”了?
2 分•作者: setnone•3 个月前
各位 HN 的朋友们,我只是想发泄一下,看看有没有人能感同身受。
产品和我几个月前注册时已经不一样了,我和 Claude Code 在 Opus 4.6-4.7 版本上也经历了同样的变化。
最能描述这种差异的方式是:你雇佣了一位可靠的“聪明”的技术主管,他“很在乎”,但过了一段时间,你最终得到的是一个过度自信的初级开发人员,他像一个破坏性的代币燃烧器。
在过程中没有思考,甚至没有真正遵循指示。
这几乎是一个二元性的转变,根据我的经验,无法通过提示来解决。
我们无法真正了解模型内部发生了什么,但从第一个回复就能注意到这种差异。
我使用的是最高级别的 GPT 5.5,始终注意上下文窗口,使用计划模式等。
是模型在“变笨”,还是出于某种原因被悄悄地路由到另一个模型?
是框架的问题吗?
还是我自己的想象?
很好奇你们最近的经历。
附注:前沿模型应该在“思考:最大化”、“努力:最大化”之外,再加上“在乎:最大化”,并让它们真正起作用。
查看原文
Hi HN, just want to rant and see if anybody can relate.<p>The product is not the same as i signed up for few month ago and the same shift i've experienced with Claude Code on Opus 4.6-4.7<p>The best way to describe the difference is you hire a reliable 'intelligent' tech lead who 'gives a shit' and in some time you eventually get over-confident junior dev that acts as destructive token burner.<p>No thinking during the process, not even really following instructions.<p>It's pretty much a binary shift that happens and in my experience can't be cured with prompting.<p>There is no way to actually tell what's happening under the hood with models but the difference is noticeable from the first reply.<p>I'm on max, GPT 5.5 extra high, always mindful of context window, use plan mode etc.<p>Is it the model: dumbing down or quietly route to another model for some reason?
Is it the harness?
or is it my imagination?<p>Curious what's your recent experience.<p>PS: frontier models should add "give-a-shit: max" along with 'thinking: max', 'effort: max' and make them actually work.