← Back to Lessons So Far

系统改版 · Redesigning My System

2026/10/03

从 2024 年 12 月 5 号到今天,这套系统换过很多次。
Since December 5, 2024, this system has changed many times.

从记事本,到 TickTick,到 Notion,再到现在的 Obsidian 加 Notion。
It went from a paper notebook, to TickTick, to Notion, and now to Obsidian plus Notion.

从极简,到繁琐,再到平衡。
From minimal, to heavy, and then to balance.

花这么多力气整理,不是每个人都需要。对我来说,它是值得的。
Not everyone needs to spend this much effort on a system. For me, it has been worth it.

有时我会读历史,看大事件背后的团队和一次次决策。
Sometimes I read history: the teams behind big events, and the decisions they made one by one.

那些不太出名、却起了重要作用的人,常常让我印象很深。
The people who were not famous, but still played an important part, often stay with me.

世界一直都是群星闪耀的,今天也一样。
The world has always been full of shining stars. It still is today.

这次改版,来自一个问题:什么样的错误,最容易被平静地修正?
This redesign started with one question: which kind of mistake is the easiest to fix calmly?

做计算机和 AI 的人,大多对代码出错很坦然。
People in computer science and AI are usually calm about bugs in their code.

训练模型时,loss 不对、梯度出了问题,我们会直接去改。
When the loss looks wrong or the gradient breaks during training, we simply go and fix it.

很少有人会因为一个 bug 难过很久。修好它,就继续往前走。
Few people feel bad about a bug for long. You fix it, and you move on.

计算机领域对错误很宽容,也愿意把错误公开出来,一起改。
The field is kind to mistakes. People even share them in the open and fix them together.

于是我想:日常的事情,能不能也当成一个训练过程来看?
So I wondered: could daily work also be seen as a training process?

Obsidian 就是日志。遇到的问题,写成一个个可以观察的指标。
Obsidian is the log. Each problem becomes a signal that I can watch.

不同的事,节奏不一样:有些像 LSTM,比如会议;有些像 Transformer,比如知识卡片的更新。
Different things move at different speeds. Some are like an LSTM, such as meetings. Some are like a Transformer, such as how my knowledge cards update.

我知道,这样的结构会有不确定性。
I know this kind of structure comes with uncertainty.

但就像写递归:相信当前这一层返回的是对的,最后的答案也会是对的。
But it is like writing recursion: trust that the current call returns the right result, and the final answer will be right too.

建这套系统,是为了让判断少受情绪的影响。
I built this system so that my judgment depends less on my mood.

我还给文件建了数据库表,有清楚的主键和外键。
I also built database tables for my files, with clear primary keys and foreign keys.

结构清楚以后,很多事情变得更容易看清。
Once the structure was clear, many things became easier to see.

一个问题,往往只是某个指标需要调整。
A problem is often just one signal that needs a small adjustment.

检查,调整,继续,过程顺畅了很多。
Check, adjust, continue. The process became much smoother.

也许有经验的人很快就能做到这些。我需要一个脚手架,帮我慢慢练。
Experienced people may do this quickly. I need a scaffold to help me practice, step by step.

这个脚手架还能记录我需要多久才能平静下来。看着这个数字慢慢变小,挺有意思的。
The scaffold can even record how long it takes me to calm down. It is fun to watch that number slowly go down.

在这次整理中,我也学到了一些和 AI 合作的经验。
While doing this, I also learned a few things about working with AI.

效率提高了,但也要留意其中的陷阱。
Speed goes up, but there are traps to watch for.

AI 给出的往往不是一个确定的答案,而是一些可能性。
AI often gives you possibilities, not one fixed answer.

有人喜欢把它限定得很严。我更倾向于给它一定范围内的自由。
Some people prefer tight control. I lean toward giving it some freedom within clear limits.

整理时我还发现,和 AI 的对话只保留了我的部分,AI 的回复没有保存下来。
I also found that my chats with AI only kept my side. The AI replies were not saved.

一共 466 个对话,完整保留 269 个,大约 880 万字,4000 轮。
There were 466 conversations. 269 were fully kept: about 8.8 million characters and 4,000 rounds.

一开始觉得可惜。后来想,只留下自己的话,也许更能看出这一路的变化。
At first, it felt like a loss. Later I thought: my own words alone may show the changes along the way even more clearly.

我想记住这种状态:认真看问题,得到反馈,再慢慢调整。
I want to remember this way of working: look at the problem carefully, get feedback, and adjust slowly.

进度像水流,顺着事情发展就好。
Progress is like water. Let it follow the shape of things.