响马

响马

60,044 Twitter followers

Follow

。。。。。。

Posts

There was one feature I had AI fix. I reviewed the code but wasn't satisfied, so I started explaining the correct way to write it. After explaining it many times, I was surprised to find that the AI's initial code actually worked perfectly. I carefully read the relevant code and had to admit that the AI was really good at finding loopholes; it had solved the problem by getting stuck in a perfectly timed state. Finally, I secretly changed the code back to the original version without telling the AI and then immediately closed the session.

I was afraid to leave it alone, even watching its thought process, stopping whenever it started rambling. That's how I learned how truly absurd its thinking really is. But when I tried switching to a less analytical approach, it immediately became even dumber.

我发现了个极其简单但被大家忽视，却能大幅提高生产力的技巧。 AI 时代来了，Agent 虽然代替人的部分思考，但还有很零碎的部分需要人来协同。为了提高效率，我们自然而然想一个人协同多个智能体。但这往往适得其反，为什么呢？人为了协助 Agent

Even someone like me who strongly dislikes TypeScript has implemented direct translation in Fibjs. After writing code for over half a year, I feel the efficiency of AI modifications and refactoring has improved by at least an order of magnitude. Node.js also supports direct translation, but the implementation is just too slow.

我现在所有 skills 的脚本，默认用 bun 执行，因为 bun 可以直接执行 TypeScript，不需要预先编译，这带来的好处就是如果出错了，Agent 可以快速定位到代码里面的问题，然后直接修改代码再次运行，如果是打包后的代码这就很难打补丁。任何语言或者环境，Agent 友好才是王道！ x.com/atian25/status…

My experience is that for complex architectures, files must be split. Allowing AI to index a self-consistent logic that fits into a single context based on filenames is far more stable than it piecing together bits and pieces and guessing the overall picture. Of course, AI won't split files on its own, so I ask it to refactor a portion of the code almost every day.

我不是完全激进地反对在 vibe coding 时代文件不需要拆分。我觉得如下两种情况是需要拆分的： 1. 这个文件中的有些逻辑是被其他地方复用的 2. 文件里面有太多状态流转了，你可以通过文件拆分来维护状态流转的边界 x.com/yetone/status/…

Haven't you realized that MiroFish is the metaverse? twitter.com/xicilion/status/20...

Humans invent tools to overcome their own illusions. AI's illusions are part of AI itself; they can only be reduced in certain aspects, not eliminated entirely. This inherently limits AI from being a stable automation engine. However, AI can serve as an external brain for automation engines, which is AI's strength.

我感觉这些由md文件、skill文件驱动的AI操作系统（Openclaw/Claude Code），从鲁棒性（robustness）和确定性看就是纸糊的，但它们竟然能每天还算正常的运转，真的是Oh My God。想想这和人的团队差不多，一帮互相不断误解的草台班子，但就建立了人类数千年的伟大文明。

Now, let me state a rather outrageous claim. Skills are essentially low-code from the AI era; they look appealing, but ultimately they are still a phase in development. Deterministic processes are the foundation of automation; if each step has 90% stability, then ten steps together only achieve 34.9% stability.

低代码拖拽工作流借 ai 火了几个月，还是被遗弃了。我最早了解到可视化编程可以追溯到 1992 年前后，这个东西要是效率高，程序员老早自己就用了。一个老程序员，看几十行代码，几乎扫一眼就明白了。同样的逻辑，用工作流得画两屏幕，要了解细节还得点开看。你让程序员这么干活，他能把键盘给砸了。

I find AI's thinking particularly annoying. From my observation, Opus spends 80% of its thinking rambling on and on.

To reiterate: Currently, AI excels at quickly building seemingly complex applications and, within those applications, analyzing and solving seemingly complex problems. However, AI is poor at evolving seemingly complex applications into truly complex ones, or at analyzing and solving genuinely complex problems.

That's normal. Most Markdown readers completely freeze when there are many charts. I couldn't stand it anymore, so I wrote my own Markdown viewer. It loads instantly on documents of over 300 pages with hundreds of charts and updates incrementally in real time. twitter.com/xicilion/status/20...

Loading..