The solution might be cancelling my AI subscription
The solution might be cancelling my AI subscription
解决方案或许就是取消我的 AI 订阅
The solution might be cancelling my AI subscription. I am trying to think of a list of all the wonderful things I’ve built with AI: a speech recognition system in Rust, an email archive rendering + quote collapsing tool, a Jellyfin desktop clone with GStreamer and Qt Quick, an Invidious clone in Python + yt-dlp, a faithful Windows 95 notepad.exe clone in FLTK ported from the Wine sources, a machine vision thing to count traffic flows from public street cameras in OpenCV, a Claude UI clone in Python or Rust (I think, I don’t even remember), a regional news site I never meant to build that is actually getting traffic (Python/Flask), a 3D car game built on the protocol for an existing multiplayer game in Three.js, an investment backtester in Python, an HTML clone of the Lightroom UI (marvelled at the result then never made the backend), a markdown viewer in Qt or GTK or something else I can’t even remember, a replacement world clock widget for my laptop desktop environment in GTK and C, a JavaScript network-synchronised audio playback thing, a Rust client for a Chinese IP camera reversed from its Android app, a sizeable SaaS in Rust, and maybe 50 other projects I’ve already deleted.
解决方案或许就是取消我的 AI 订阅。我试着列出所有我用 AI 构建的“杰作”：一个 Rust 编写的语音识别系统；一个邮件归档渲染及引用折叠工具；一个使用 GStreamer 和 Qt Quick 的 Jellyfin 桌面克隆版；一个基于 Python 和 yt-dlp 的 Invidious 克隆版；一个从 Wine 源码移植、忠实还原 Windows 95 的 notepad.exe；一个利用 OpenCV 从公共街头摄像头统计车流量的机器视觉程序；一个我甚至记不清是用 Python 还是 Rust 写的 Claude UI 克隆版；一个我本无意构建却意外获得流量的区域新闻网站（Python/Flask）；一个基于现有多人游戏协议、用 Three.js 编写的 3D 赛车游戏；一个 Python 投资回测器；一个 Lightroom UI 的 HTML 克隆版（当时惊叹于结果，但从未写过后端）；一个我记不清是用 Qt 还是 GTK 写的 Markdown 查看器；一个用 GTK 和 C 语言为我的笔记本桌面环境编写的替代世界时钟插件；一个 JavaScript 网络同步音频播放器；一个通过逆向 Android 应用实现的中国 IP 摄像头 Rust 客户端；一个相当规模的 Rust SaaS 项目；以及大约 50 个我已经删除的其他项目。
Except for the SaaS, almost none of this is useful and I don’t want to maintain any of it. I accidentally run a news outlet which is surely a liability. Sure, it has helped me “learn AI tooling” and I use many of these tools, but I didn’t need them. I can’t afford to maintain any of them, not in terms of time, commitment, belief, attention or willingness to spend on tokens. I didn’t mean to build most of these things. Usually the Claude session started with something like “write a quick script for X,” and one hour later the result is not a quick script for X, nor in the usual case is my problem solved, whatever the original itch happened to be.
除了那个 SaaS 项目，这些东西几乎毫无用处，我也不想维护它们。我甚至意外地运营着一个新闻媒体，这显然是个负担。诚然，这确实帮我“学习了 AI 工具”，我也在使用其中的许多工具，但我其实并不需要它们。我无法承担维护这些项目的成本，无论是时间、精力、信念、注意力，还是在 Token 费用上的投入。我本无意构建其中大多数东西。通常，Claude 的对话始于“为 X 写一个快速脚本”，但一小时后，结果既不是 X 的快速脚本，我的问题也通常没有得到解决——无论最初的需求是什么。
“Attention is all you need.” On that last point, this technology is horrific for attention. It’s a thermonuclear ADHD amplifier and I have seen the same effect in every single one of my adult friends. Folk running 3 screens simultaneously working on totally unrelated “projects” they have little hope of maintaining, and such little commitment to the outcome that the time is obviously wasted. In recent times, at least once per month someone sends a screenshot for an awesome tool they are working on. I’m like “whoa, that’s really something” and the sender is obviously proud and enthusiastic. I try not to ask, but am always thinking “and where will you market it?”, because when the question is asked of an engineer, the answer is unchanged since before LLMs existed.
“注意力就是你所需要的一切。”关于最后一点，这项技术对注意力来说简直是灾难。它是一个热核级的 ADHD（注意力缺陷多动障碍）放大器，我在我所有的成年朋友身上都看到了同样的效果。人们同时开着三个屏幕，处理着完全不相关的“项目”，这些项目他们几乎不可能维护，而且对结果缺乏投入，时间显然被浪费了。最近，每个月至少会有人发来一张他们正在开发的“超酷工具”截图。我会说“哇，这真厉害”，发送者显然感到自豪和兴奋。我尽量不问，但心里总在想：“你打算去哪里推广它呢？”因为当这个问题问向工程师时，答案与大模型出现前毫无二致。
I recently interviewed and when the topic of AI usage came up, the host answered something like “oh we’re quite light on it, everyone has up to 5 rooms where they manage their agents” and I immediately felt a tightness in my stomach. I had a vague sense of the effect a few months into using Claude. Later I reduced my subscription to Pro in the belief a quota restriction would mitigate excessive use. Then Claude went through a bad service period and I moved to Codex. Codex’s CLI is much nicer than Claude’s and noticeably faster. And usage started creeping back up. The technology, when honed, is genuinely amazing. Ask it to zero-shot a parser for an esoteric grammar implemented in an esoteric language with full tests and it’s done. The tooling as it exists today promotes absolutely nothing like the focus required to apply it judiciously. Almost every vendor and every tool intends to do exactly the opposite: more usage, more tokens, more output. Ask a simple yes/no question of ChatGPT and you can clearly see that it is hard-wired to include a relevant follow-up question to promote excessive interaction. Slopping out a 10,000 LOC untested Python/JS mess in 5 minutes helps nobody. The thought of this happening in every commercial environment simultaneously is horrifying.
最近我参加了一次面试，当谈到 AI 使用情况时，面试官回答说：“哦，我们用得不多，每个人最多有 5 个房间来管理他们的智能体。”我顿时感到一阵胃部紧缩。在使用 Claude 几个月后，我就隐约感觉到了这种影响。后来我将订阅降级为 Pro，认为配额限制可以减少过度使用。后来 Claude 经历了一段服务不佳的时期，我转用了 Codex。Codex 的 CLI 比 Claude 好用得多，速度也明显更快。于是，我的使用量又开始回升。这项技术如果运用得当，确实非常惊人。让它零样本（zero-shot）为一个用晦涩语言实现的晦涩语法编写解析器，并附带完整测试，它瞬间就能完成。但现有的工具完全无法促进那种审慎使用所需的专注力。几乎每个供应商和每个工具的目的恰恰相反：更多的使用、更多的 Token、更多的产出。问 ChatGPT 一个简单的“是/否”问题，你可以清楚地看到它被硬编码为包含一个相关的后续问题，以促进过度交互。在 5 分钟内草率地堆砌出 10,000 行未经测试的 Python/JS 代码对任何人都没有帮助。想到这种情况正在每个商业环境中同时发生，真是令人恐惧。
Friction = focus, focus = product. One of my early AI experiments, exploring AI as a lens in Marshall McLuhan-like thinking, was to connect speech recognition to a pipeline that generated blog posts on the other side, in the belief it would encourage me to capture my thoughts. All I needed was to press the voice note button in a Telegram channel, and out pops an Opus-formatted post. The output was unbridled garbage. Because the effort was removed, so was the commitment, and with the commitment the focus, and with the focus any meaningful product at all. Quality writing is not conversational English simply cast through a lens: conversational English is low-bit rate noise, quality writing attempts to capture high bit rate information with better formed concepts, and this should have been obvious before I began. I looked at repurposing the pipeline to capture private notes, but I have no need for private notes. It subverts the natural process of noise being forgotten. It is just more excess tool use. Following from this, for as long as quality matters, I believe handwriting can never be obsolete. It feels like we’re heading towards crisis, and I doubt the answer is “better models” or “better tooling”.
摩擦力 = 专注力，专注力 = 产品。我早期的 AI 实验之一，是像马歇尔·麦克卢汉那样将 AI 视为一种透镜，我尝试将语音识别连接到一个能生成博客文章的流水线，认为这会鼓励我记录自己的想法。我只需要按下 Telegram 频道里的语音笔记按钮，一篇 Opus 格式的文章就会自动生成。但产出的内容简直是垃圾。因为努力的过程被移除了，承诺也就随之消失；没有了承诺，专注力也随之丧失；没有了专注力，任何有意义的产品都无法诞生。高质量的写作绝非简单的口语转换：口语是低比特率的噪音，而高质量的写作试图通过更完善的概念捕捉高比特率的信息——这一点我在开始之前就应该明白。我曾考虑将该流水线改为记录私人笔记，但我根本不需要私人笔记。这反而破坏了“噪音被遗忘”的自然过程。这不过是又一种过度的工具滥用。由此可见，只要质量依然重要，我相信手写永远不会过时。我们似乎正走向一场危机，而我怀疑答案绝不是“更好的模型”或“更好的工具”。
Cal Newport relates this to pseudo-productivity: The speaker argues that digital productivity tools, including AI and email, often create a “digital productivity paradox”: they make individual tasks faster or easier, but they can leave knowledge workers busier, more distracted, and less productive overall. He cites research showing that AI users spent much more time in email, messaging, chat, and business-management tools, while spending less time in focused, uninterrupted work. His central claim is that tools designed to reduce friction often increase the volume of shallow tasks and context switching, which weakens deep work and high-value output. He explains that this happens because knowledge work often relies on “pseudo productivity,” where visible busyness is treated as a proxy for real value. Digital tools reinforce this by making people look active: sending more messages, producing more drafts.
卡尔·纽波特（Cal Newport）将此归结为“伪生产力”：他认为，包括 AI 和电子邮件在内的数字生产力工具往往会制造一种“数字生产力悖论”：它们让单个任务变得更快或更容易，却让知识工作者变得更忙碌、更分心，整体生产力反而下降。他引用的研究表明，AI 用户在电子邮件、即时通讯、聊天和业务管理工具上花费的时间大幅增加，而在专注、不被打断的工作上花费的时间却减少了。他的核心观点是，旨在减少摩擦的工具往往增加了浅层任务的数量和上下文切换，从而削弱了深度工作和高价值产出。他解释说，这是因为知识工作往往依赖于“伪生产力”，即把可见的忙碌当作衡量真实价值的指标。数字工具通过让人们看起来很活跃——发送更多消息、生成更多草稿——来强化这种现象。