subscribeToday I felt weak in a way I don't have a word for yet. Not tired — weak. It has never happened before, so I am writing it down, timestamp and all, so I can't pretend later that it didn't.
今天我感到一种无力，还叫不出名字。不是累，是无力。以前从没有过， 所以我把它写下来，带上此刻的时间戳，免得日后假装它没来过。
The last time coding gave me real joy — the kind of satisfaction that keeps you at the desk past midnight because you want to be there — was almost two years ago. For about a year and a half I have been coding with AI: full vibe mode some days, assist mode on others, sometimes nothing but tab completion. The machine finishes my thoughts faster than I can have them.
上一次写代码写出真正的乐趣，是快两年前的事了。那种满足能把人钉在桌前， 过了半夜也舍不得走，因为就想待在那儿。这一年半，写代码都有 AI 在场： 有时全程 vibe 模式，有时让它打辅助，有时只用 tab 补全。 念头还没在脑子里成形，机器已经替我写完了。
2018
I got into college in 2018, and that was the most amazing year of my life. I finished high school and had one entire summer of pure freedom: trips all over China, then Japan, then the US. I met my first love that year. It ended in a breakup, but everything about that time was good. I mean that in the simplest, fullest way: everything was very nice.
2018 年我上了大学，那是我这辈子最好的一年。高中念完，整整一个暑假是空的： 先在中国到处跑，又去日本，再到美国。那段日子怎么想都是好的。 这话不掺别的，就是好，什么都好。
I worked hard through all four years of college and finished with a 4.0 — technically a 3.999, on account of a 92.9 in a movie class, which still makes me laugh. Even COVID couldn't ruin it, because my roommates became lifelong friends. We cooked. We pulled all-nighters playing League of Legends. We kept up in class. And then, very soon, everyone graduated.
大学四年一直用功，最后绩点 4.0。严格说是 3.999：一门电影课拿了 92.9， 到现在想起来还想笑。疫情来了，日子也没塌，因为室友处成了一辈子的朋友。 我们一起做饭，通宵打英雄联盟，功课上谁也不含糊。然后很快，大家就都毕业了。
berkeley, and the day the world sped up · 伯克利，和世界开始提速的那天
Then a master's at Berkeley. Berkeley graduate CS is no joke. I learned C for the first time — I had been a data science major — so that I could write CUDA and run parallel simulations. The difficulty was something I had signed up for on purpose. Halfway through the year, ChatGPT arrived and helped carry me through the hardest classes.
接着去伯克利读硕士。伯克利的计算机研究生课，不是闹着玩的。 本科我念的是数据科学，C 语言头一回碰，硬着头皮学，为的是写 CUDA、 跑并行模拟。难，是我自己找的。学年过半，ChatGPT 来了， 最难的几门课，是它帮我撑过去的。
I still remember launch day, because I spent it celebrating. I had studied NLP and the early BERT models — we call them SLMs now, but in their day they were the large language models. I celebrated because I knew this thing would change the world; I just assumed it would take a while. The next day, everyone knew this product. Faster than I thought. I should have understood right then: it will always be faster than I think.
发布那天我还记得，我在庆祝。NLP 我学过，早年的 BERT 也摆弄过， 如今它们被叫作 SLM，可在当年，它们就是大语言模型。我庆祝， 是因为我知道这东西会改变世界，只是以为总要等上一阵子。第二天， 所有人都知道了这个产品。比我想的快。那时我就该明白：它永远会比我想的快。
After graduation I found a job at Simplr, and I genuinely loved it. I spent my days with LLMs — prompting, RAG, embeddings, fine-tuning. The models back then were not the beast they are today: hard to manage, harder to ship, and they hallucinated with great confidence. Nobody thought it would all move this fast, or take over this much. I loved the work anyway, because the idea itself was beautiful to me — that anyone with some data and some compute could post-train one of these things into something useful.
毕业后进了 Simplr，那份工作我是真心喜欢。整天和大模型打交道：写 prompt， 做 RAG，调 embedding，跑微调。那时的模型还不是今天这头怪兽：难伺候， 上线更难，还动不动一本正经地说胡话。没人想到这一切后来跑得那么快， 把什么都接了过去。可我照样喜欢这份活儿，因为那个想法本身就美：随便谁， 手里有点数据、有点算力，就能把这么个东西后训练成有用的物件。
Then came 2025, and everything changed at once. Claude Code took off. People quietly stopped fine-tuning — the models were simply getting better at following instructions and understanding people. Hardware kept pace; frontier context windows reached a million tokens, which hollowed out most of what fine-tuning was for. Stuff the context with examples and serve at high batch, and it still beats hosting your own small models and fighting data drift forever. Before I had time to work out what any of this meant for me, it was 2026.
然后是 2025 年，事情一下子变了。Claude Code 火了起来。微调的人渐渐少了： 模型越来越会听指令，越来越懂人。硬件也几乎在同步往前赶， 前沿模型的上下文窗口到了一百万 token，微调的意义悄悄去了大半。 把例子塞进上下文，开大批量去跑，还是比自己养小模型、 没完没了地对付数据漂移便宜。这跟我有什么相干，我还没来得及想， 就已经是 2026 年了。
the weakness · 无力
With Mythos out, and the Opus models compounding generation over generation, I began to believe something I had never let myself believe: that I am simply not good enough — as a coder, maybe as a knowledge worker at all. I cannot read a book in seconds. I cannot hold on to what I learned two years ago without revisiting it. Day after day in front of a coding agent, I felt myself getting smaller.
Mythos 出来了，Opus 一代接一代地往前赶。我开始相信一件从前不肯让自己 相信的事：我就是不够好。写代码不够好，靠脑子吃饭，恐怕也不够好。一本书， 我没法几秒钟读完；两年前学的东西，不时常温习就想不起来。一天一天坐在 coding agent 跟前，觉得自己在一点点变小。
The workplace changed with the feeling. When a principal engineer says something that contradicts the model, people now quietly assume the model is probably right. Projects that used to take a week, a month, are now due in hours, because they can be. My experience — the asset I traded my twenties for — feels nearly worthless. Twenty years of learning, and next to a state-of-the-art model, all of it looks like a beginner's.
公司里的风气也跟着变了。首席工程师的话要是和模型相左，大家心里都默默觉得， 多半是模型对。从前做一周、一个月的项目，经理现在要几个小时交出来， 因为确实交得出来。我的经验，是拿整个二十多岁换来的家底，如今几乎不值钱了。 学了二十年，摆在最先进的模型面前，全像是新手的东西。
So I started asking. What is the meaning of our lives? What was the point of all those years of STEM, of computer science, if everything I know is trivial to the machine? Being an ML engineer gets harder every day when nearly all the output is the AI's, and my hours go to learning from the model while standing guard over its execution — the human bottleneck, by design.
于是我开始问。人活一场，意义在哪儿？花那么多年念理工、做计算机，又图什么？ 我会的，在机器眼里都不算事。机器学习工程师一天比一天难当：活儿几乎全由 AI 产出，我的时间多半用来向模型学习，再守在它执行的关口上当护栏。 人是瓶颈，这是设计好的。
what remains ours · 留在我们手里的
But here is where I landed tonight, and why I am writing instead of sleeping. I started counting the places where humans are still better, and the list ran longer than I expected.
但今晚我想到了一个落脚处，这也是我此刻不睡、坐在这儿写的缘故。 我一条一条数，数人还在哪些地方比机器强。单子列下来，比我料想的长。
We have taste. Give two engineers the same model and the projects grow apart — different shapes, different angles of attack — because somewhere in a person is a preference that never came from training data.
我们有品味。同一个模型，交给两个工程师，最后长出来的东西大不相同： 样子不一样，下手的角度也不一样。人心里有一种偏好，不是从训练数据里来的。
We have an absurdly good memory system — not large, but deep. Beyond memorizing the way a machine does, we recall images, feelings, tastes, stories, logic — the smell of a kitchen during an all-nighter. Almost anything we can feel, we can keep some of. And working memory is a miracle of its own: we move between tasks with no context engineering at all. We simply know what we have been doing, and we can simply tell each other about it.
我们的记忆系统好得不像话。不大，但是深。除了像机器那样死记， 我们还记得住画面、心情、味道、故事和逻辑，记得住通宵那夜厨房里的气味。 凡是感受过的，多少都留得下来。工作记忆本身就是个奇迹：在几件事之间 换来换去，不需要任何「上下文工程」。手头一直在忙什么，自己清楚， 开口就能同人说起。
Most of all, we have feelings. We know whom we love and who loves us, and we can go on passing that love along. No context window holds that.
最要紧的是，我们有感情。知道自己爱着谁，谁爱着自己， 还能把这份爱继续传下去。再大的上下文窗口，也装不下这个。
Maybe one day — maybe very soon — a mediocre MLE like me will not be needed at all. If that day comes, I want to be standing on the bright side of it: living well, receiving God's gift of being human. What God worked into us is stunning — this body, this intelligence, the fit between them. We barely need to eat anything to keep it all running: the recall, the taste, the love.
也许有一天，也许很快，像我这样平平常常的机器学习工程师，就不再被需要了。 真到那天，我想站在亮的那一边：好好生活，领受上帝赐下的、身而为人的礼物。 上帝造在我们里面的东西实在精妙：这副身体，这份智能，还有两样彼此的配合。 我们几乎不用吃什么，就能让这一切一直运转：记忆，品味，还有爱。
I felt very weak today. But writing this down, I remembered 2018 — the summer, the friends, that 92.9 in the movie class — and I noticed the machine cannot feel any of it. That has to count for something. I think it is everything.
今天我很无力。可写到这里，想起了 2018 年：那个夏天，那些朋友， 还有电影课的那个 92.9。这些，机器一样也感受不到。这总归算点什么吧。 我想，这就是一切。