I think I have LLM burnout

我想我患上了“大模型倦怠症”

I use LLMs a lot. By current dev standards, my usage rate is probably average, and my methods are probably primitive. I work on one task at a time and discuss it with Claude Code (at work) or Codex (at home, for now). Sometimes, I let the assistant write code, but I read the output thoroughly, understand it, and revise it. I’m not in the deep end of autonomous agents or agent orchestration.

我经常使用大语言模型（LLM）。以目前的开发标准来看，我的使用频率大概处于平均水平，方法也比较原始。我一次只处理一个任务，并与 Claude Code（在工作中）或 Codex（目前在家里）进行探讨。有时，我会让助手编写代码，但我会仔细阅读输出内容，理解它并进行修改。我并没有深入涉足自主智能体或智能体编排领域。

Still, I spend hours each day interacting with LLMs across work and home. That’s a hell of a lot more than I did a few years ago, and I probably don’t go a day without reading AI-generated text. My job has changed from designing and writing code to designing code, describing the design to an LLM, reviewing code the LLM produces, and then finally writing code. The LLM steps expose me to approaches I might not have considered or been aware of. I also feel more comfortable in areas where I don’t have deep knowledge.

尽管如此，我每天仍要花数小时在工作和生活中与大模型交互。这比几年前多得多，我几乎每天都要阅读 AI 生成的文本。我的工作内容已经从“设计并编写代码”转变为“设计代码、向大模型描述设计、审查大模型生成的代码，最后再编写代码”。大模型的介入让我接触到了以前可能没考虑过或没意识到的方法。在那些我缺乏深入了解的领域，我也感到更加从容了。

My main project right now is to establish a framework for large-scale, unsupervised code generation in our codebase. When I’m not working with Claude to create tooling, I’m sifting through the unsupervised agent’s (Qwen’s) output. Either way, I’m reading LLM content. If I want to know something, I’ll probably ask ChatGPT or read Gemini’s overview unless I know what sites I want to check. I still have to fall back to browsing when the LLM’s answer is wrong, but it’s good enough for many casual queries, especially when useless AI-generated articles clutter the search results.

我现在的主要项目是为我们的代码库建立一个大规模、无监督的代码生成框架。当我不与 Claude 合作开发工具时，我就是在筛选无监督智能体（Qwen）的输出。无论哪种方式，我都在阅读大模型生成的内容。如果我想了解什么，除非我知道要去查哪些网站，否则我可能会直接问 ChatGPT 或阅读 Gemini 的概览。当大模型的回答出错时，我仍然不得不退回到传统的搜索方式，但对于许多随意的查询来说，它已经足够好了，尤其是在搜索结果充斥着无用的人工智能生成文章时。

It’s been this way for about a year, and I don’t see myself stopping. I feel more productive with LLMs, and I think continually learning how to use them effectively is valuable. However, my disposition has changed a bit in the last few months. Some small part of me has started to dread reading LLM output because I know what I’m going to find. False assumptions and hallucinations. Emphatic, staccato fragments. ✨ Excessive emojis 🚀. It’s not just me—these are real patterns (🤮). On their own, none of these annoyances gets to me. Together, though, they’ve gotten me sick of LLM writing in a hurry.

这种情况已经持续了大约一年，我看不出自己有停下来的迹象。我觉得使用大模型让我效率更高，而且我认为持续学习如何有效地使用它们很有价值。然而，过去几个月里，我的心态发生了一些变化。内心深处，我开始害怕阅读大模型的输出，因为我知道我会看到什么：错误的假设和幻觉、语气强烈且断断续续的片段、✨过多的表情符号🚀。这不仅仅是我个人的感觉——这些都是真实存在的模式（🤮）。单独来看，这些烦人的地方都不会让我抓狂。但当它们凑在一起时，让我很快就对大模型的写作感到厌倦。

I’m not trying to condemn LLMs. Humans are fallible, too—we can be just as unreliable or annoying. The problem is repetition. LLMs write in the same style, and they make the same kinds of mistakes. Dealing with the same thing over and over is wearing me out. I can use personalization features if the interface offers them, but some idiosyncrasies seep through. And of course, I don’t control the style of content generated by other people.

我并不是要谴责大模型。人类也会犯错，我们同样可能不可靠或令人讨厌。问题在于重复。大模型总是用同样的风格写作，犯同样的错误。一遍又一遍地处理同样的事情让我精疲力竭。如果界面提供个性化功能，我可以使用，但一些特质总会渗透出来。当然，我也无法控制其他人生成的内容风格。

I don’t know how to deal with this feeling yet. I didn’t expect to be so bothered by it. Frustration at a flaky tool is understandable, but the writing patterns grind my gears, too. For now, I’ll grit my teeth and hope I don’t lose my lunch.

我还没想好该如何应对这种感觉。我没想到自己会因此感到如此困扰。对一个不稳定的工具感到沮丧是可以理解的，但那些写作模式也确实让我心烦意乱。目前，我只能咬牙坚持，希望自己不会真的“反胃”。