GPT-5 不给力?

1作者: anupshinde9 个月前
总结 - 我一直在努力让 GPT-5 稳定地“合拍”。 我不知道 GPT-5(在 ChatGPT.com 上)出了什么问题。表面上看,一切都很好,但当你深入了解细节时,就会发现问题。 有时候它太聪明,有时候又很愚蠢。这一切都发生在同一个聊天会话中。 有时候它会犯错,有时候又不会。 有时候它会记住上下文,有时候它会忘记,需要多次提醒。 对于代码 - 它会编写更简单的代码或更复杂的代码 - 有缺陷的代码或可运行的代码 - 并且很难预测它何时会做什么。 问题通常有多种解决方案。而 GPT-5 似乎也遵循这一原则。 即使给它同一个问题两次,它也会给出不同的解决方案。但并不能每次都可靠地做到这一点。 与 Claude 模型或 GPT-4o 相比,使用它(或与它“合拍”)变得相当困难。 GPT-4o 模型并没有更好 - 但至少它很可靠。它的行为有一种微妙的模式。 我知道它什么时候会起作用,会犯什么错误,以及如何让它工作,以及什么时候它不会起作用。 而且在写作时,GPT-4o 通常听起来像一位哲学家诗人。 首先,GPT-5 在炒作发布后令人失望。现在又这样。 这种体验主要基于 ChatGPT,以及 GitHub Copilot 的一点点。对于 Copilot,我坚持使用 Claude Sonnet 4。 Canvas 感觉也很 buggy。当你让它编辑文本的一部分时,它有时也会删除其他部分。尽管我不太使用 Canvas。 我以为这可能是由于自动路由造成的。并尝试通过选择 GPT-5 类型来修复它,但没有成功。 还有其他人遇到过同样的问题吗?你们是如何解决的?
查看原文
TL;DR - I have been struggling with GPT-5 to get it to &#x27;vibe&#x27; reliably.<p>I don&#x27;t know what&#x27;s wrong with GPT-5 (on ChatGPT.com). On the surface, everything looks hunky-dory until you get into the details.<p>Sometimes it is too smart, and sometimes it is stupid. All this within the same chat-session. Sometimes it makes mistakes, other times it does not. Sometimes it remembers the context, sometimes it forgets and has to be reminded multiple times. With code - it writes easier code or complex code - buggy code or working code - and it is hard to predict when it would do what.<p>Problems often have multiple solutions. And GPT-5 seems to adhere to that philosophy. Given it the same problem twice, it will come up with different solutions. But not reliably do so every time.<p>It has become quite difficult to use it (or &#x27;vibe&#x27; with it) - compared to Claude models or GPT-4o.<p>The GPT-4o model was not better - but at least it was reliable. There was a subtle pattern to its behavior. I knew when it would work, what mistake it would make, and how to get it to work, and when it would not work. And when writing, GPT-4o would usually sound like a philosophical poet.<p>First, GPT-5 disappointed after the hyped launch. And now this.<p>This experience is mostly based on ChatGPT, and a bit on GitHub Copilot. With Copilot, I stick to Claude Sonnet 4.<p>The Canvas also feels buggy. When you ask it to edit a part of the text, it sometimes deletes other sections as well. Albeit, I do not use Canvas as much.<p>I thought it may be due to auto-routing. And tried to fix it by selecting the GPT-5 type, no luck.<p>Has anybody else faced the same problem? And how did you resolve it?