I’m done using AI

我不再使用 AI 了

10 Aug, 2026 2026年8月10日

I’ve been programming for 20 years. It started as a hobby as a teen, making fansites and little video games. Then it turned into a career. And it’s something I’ve largely enjoyed doing. I like learning, I like problem solving, and I like making things, which is a lot of what programming is. 我从事编程工作已经 20 年了。最初这只是我青少年时期的一个爱好，我做过粉丝网站和一些小游戏。后来，它变成了一份职业。我一直很享受这份工作。我喜欢学习，喜欢解决问题，也喜欢创造事物，而这正是编程的核心所在。

Over the past 18 months, my career, my hobby, and my craft have been drastically changed by the introduction of LLM-powered coding harnesses (e.g., Claude Code) and what’s often referred to as “agentic programming.” Because of its negative impacts on the environment, society, and my mental health, I’ve decided to stop using it. 在过去的 18 个月里，随着大语言模型（LLM）驱动的编程辅助工具（例如 Claude Code）以及常被称为“智能体编程”（agentic programming）技术的引入，我的职业生涯、爱好和手艺都发生了翻天覆地的变化。由于它对环境、社会以及我的心理健康产生的负面影响，我决定停止使用它。

I first became aware of using LLMs for programming a few years ago through GitHub’s Copilot integration in Visual Studio Code, which was like autocomplete on steroids. You’d start typing function sortNames and it’d show you a preview of what that implementation could look like. I tried it and found it distracting and basically useless. 几年前，我第一次接触到利用大语言模型进行编程，是通过 Visual Studio Code 中的 GitHub Copilot 集成，它就像是“强化版”的自动补全。当你开始输入 function sortNames 时，它会向你展示该实现可能的样子。我试用了一下，发现它不仅让人分心，而且基本没什么用。

The second time it reared its head was the early chat interfaces. A colleague was using it to help investigate an issue and the solution the LLM provided them with wasn’t working. Turns out that the LLM hallucinated a version of a software library that didn’t exist, and therefore couldn’t be installed nor used to fix the issue. 第二次接触是在早期的聊天界面中。一位同事用它来协助调查一个问题，但大语言模型提供的解决方案根本行不通。结果发现，该模型“幻觉”出了一个根本不存在的软件库版本，因此既无法安装，也无法用于修复问题。

Around this time, at that job, a true believer of AI was hired as the head of the product team. They were a product manager who began submitting changes, using the LLM-powered coding tools but without understanding the implications to the broader system. Conversations became focused on how to integrate AI into the software that the customers used. I was watching it spread like a virus throughout that organization, while I continued programming away as I had for the decades before, writing tests and building features, and largely enjoying the work. 大约在那个时候，公司聘请了一位 AI 的忠实信徒担任产品团队负责人。这位产品经理开始提交代码变更，他使用这些 AI 编程工具，却并不了解这对整个系统意味着什么。对话的焦点变成了如何将 AI 集成到客户使用的软件中。我眼睁睁地看着它像病毒一样在组织内蔓延，而我则继续像过去几十年那样编程，编写测试、构建功能，并依然乐在其中。

At a newer job, the job I have now as I write this, I continued to code how I had my entire life. Writing TypeScript, Ruby, CSS, etc. and corresponding tests, integrating with other systems and dependencies. It’s a small team of less than 10 developers. On collaborative calls, I’d see how some of my colleagues utilized AI in the Cursor editor, quickly tab completing the code and sometimes prompting it inline to write some code. 在现在这份工作中，我依然保持着我一贯的编程方式。编写 TypeScript、Ruby、CSS 等代码及相应的测试，并与其它系统和依赖项进行集成。这是一个不到 10 人的小团队。在协作会议上，我看到一些同事如何在 Cursor 编辑器中使用 AI，通过 Tab 键快速补全代码，有时还会直接在行内提示 AI 编写代码。

The LLM-powered coding tooling at that time was largely working at a local-to-that-file, small scale. I thought it was neat, almost magical, but nothing would pry me away from Neovim and tmux, and I felt like I was working fast enough anyway. I didn’t, and I still don’t, see the need for working faster. I really like what Cal Newport encourages in his book Slow Productivity: work at a natural pace. 当时，这些大语言模型编程工具大多是在单个文件范围内的小规模工作。我觉得这很巧妙，甚至有点神奇，但没有什么能让我放弃 Neovim 和 tmux，而且我觉得我的工作速度已经足够快了。我过去不觉得，现在也不觉得有必要追求更快的速度。我非常认同卡尔·纽波特（Cal Newport）在《慢生产力》（Slow Productivity）一书中的建议：以自然的节奏工作。

In early 2025 I was asked a question by management: “what do you think of this AI coding stuff?” And I said “it’s often wrong, gets in my way, and doesn’t help me code faster, so I don’t use it.” Management would not long after attend a conference and be sold on how LLMs would change everything we know about knowledge work, revolutionizing the workplace. This then led to a mandate to make use of AI tools or be left behind. 2025 年初，管理层问我：“你觉得这些 AI 编程工具怎么样？”我说：“它们经常出错，反而碍手碍脚，也不能帮我写得更快，所以我不用。”不久之后，管理层参加了一场会议，被“大语言模型将改变我们对知识工作的认知，彻底变革职场”的观点所说服。随后，公司下达了强制要求：必须使用 AI 工具，否则就会被淘汰。

I decided to give AI a go, in earnest. I tried the Cursor editor, but it didn’t click with me and ran horribly on Linux. So I tried Zed, and pretty soon thereafter I was utilizing the sidebar chat with auto-edit to have the LLMs make changes to the code and build software. At that time, it honestly felt magical. Type into a box and the thing would write code and answer my questions. 我决定认真尝试一下 AI。我试用了 Cursor 编辑器，但它并不适合我，而且在 Linux 上运行得非常糟糕。于是我试了 Zed，很快我就开始利用侧边栏聊天功能配合自动编辑，让大语言模型修改代码并构建软件。当时，这感觉确实很神奇。在对话框里输入文字，它就能写代码并回答我的问题。

During mid-2025, agentic editing became the new way to use AI for coding, replacing the auto-complete-based workflows. As the AI companies invested in training models tuned for that purpose and building coding harnesses, the capabilities kept getting better and better. 2025 年年中，智能体编辑（agentic editing）成为了使用 AI 编程的新方式，取代了基于自动补全的工作流。随着 AI 公司投入资源训练专门的模型并构建编程辅助工具，其能力变得越来越强。

Fast forward a year to mid-2026. In the lead up to my decision to stop using AI, I was able to essentially connect Linear to Claude Code and have it build out a non-trivial project from start to finish without me editing a line of code. The work got done faster than I could have done on my own. And I barely had to think. While that sounds amazing if you’re a person who profits from more work getting done, there are immense drawbacks to this, which I’ll share more about in a moment. 时间快进到 2026 年年中。在我决定停止使用 AI 之前，我甚至可以将 Linear 与 Claude Code 连接起来，让它从头到尾完成一个复杂的项目，而我无需修改一行代码。工作完成的速度比我自己做要快得多，而且我几乎不需要动脑子。虽然对于那些从提高工作产出中获利的人来说这听起来很棒，但这背后有着巨大的弊端，我稍后会详细说明。

I share this anecdote to say: I understand what the LLM-based coding harnesses are capable of. I write this all from the perspective of an experienced lead engineer who used these tools in earnest for more than a year. And I cannot in good conscious continue to use AI for coding (or anything else) any longer. 我分享这个轶事是为了说明：我非常了解基于大语言模型的编程辅助工具能做到什么程度。我写下这些内容，是基于一位资深首席工程师的视角，我曾认真使用这些工具超过一年。但我无法违背良心，继续在编程（或任何其它事情）中使用 AI。

During the time when I started using the LLM coding tools, I also started to use the chat functionality. I’d ask it to recommend me books. I’d ask for career advice. I’d float business ideas by it. I talked to it like it was a friend. I’d come up with grand ideas, and it was always supportive. And I’d ask it for medical advice. 在我开始使用大语言模型编程工具的同时，我也开始使用它的聊天功能。我会让它推荐书籍，寻求职业建议，向它抛出商业想法。我像对待朋友一样与它交谈。我会提出宏大的构想，而它总是给予支持。我甚至会向它咨询医疗建议。

I’m embarrassed to admit this, but one night when I was sick, I described all my symptoms and it told me to go to the Emergency Department at the hospital. So I hopped in my car at midnight, suffering from the flu, and drove to the ED. I sat there and waited for two hours in the waiting room while my ibuprofen I took before I left kicked in. By the time I saw a medical professional, they basically just told me to go home and rest, that I didn’t need to go to the ED. 我很尴尬地承认这一点：有天晚上我生病了，我描述了所有症状，它告诉我应该去医院急诊科。于是，我在午夜时分，顶着流感，开车去了急诊室。我在候诊室坐了两个小时，期间我出发前吃的布洛芬开始起效了。等我见到医护人员时，他们基本上只是让我回家休息，说我根本没必要来急诊。

While sure, yeah, better safe than sorry can sometimes apply to the ED, I think relying on the AI chat for what action to take for my medical condition was a huge mistake. It also cost me hundreds of dollars. That example, among others, helped me realize the AI chat was making me delusional. I stopped using the AI chat because it became very clear it wasn’t helpful or healthy, and it was in fact dangerous. But I kept using AI for coding at work, using it to take on more and more of the heavy-lifting to the point where I was just a code reviewer and a quality… 虽然“小心驶得万年船”在某些情况下适用于急诊，但我认为在医疗状况上依赖 AI 聊天来决定行动方案是一个巨大的错误。这还让我损失了数百美元。这个例子以及其它一些经历，让我意识到 AI 聊天正在让我产生幻觉。我停止使用 AI 聊天，因为它变得非常明显——它既无益也不健康，实际上甚至是危险的。但我继续在工作中将 AI 用于编程，让它承担越来越多的繁重工作，以至于我最终只成了一个代码审查员和质量……