# HA7CH — Full Corpus > HA7CH is an AI-native Builder Lab born at Stanford, the world's first FDE Accelerator. Founded by lawted (https://x.com/lawted2). This file bundles every essay published at https://ha7ch.com/writing for LLM ingestion. Each essay is also available individually at /writing/{slug}/md. # Why You Should Come to Hatch / 你为什么该来 HA7CH > Published 2026-05-21 · By lawted · Canonical: https://ha7ch.com/writing/why-you-should-come-to-hatch ## English Let's just assume you are an AI-native person. Let me ask you a question. What's the coolest thing you have ever built? Maybe it's a final project for some stupid course. Maybe it's a project at a hackathon. Or maybe one weekend you used Claude Code to throw together a small demo, took a screenshot, sent it to LinkedIn, and a lot of people liked it. And then what? A lot of things just stop there. So in the summer or winter, you still go look for an internship. Maybe 40 RMB an hour, doing some dog shit job. You know it's not what you want. You just don't know where else you can go. --- So let me tell you something. There are people out there willing to pay for your ability, and pay you a lot of money. But they are not in the world you are familiar with. They are the bosses you would never get a chance to touch if you stay in school. They run a traditional company, maybe 30 people. ARR maybe a million. Their daily job is this: people are using Excel, PDF, WeChat, some internal system, just copy and paste. They know something is wrong. They want to use AI. But they don't know how to prompt, they don't know how to find these AI-native people, and they don't trust that the consulting or the IT company will give them what they actually want. But you. You can use Claude Code, you know how to build a frontend, you know some shit about backend, and you can build a small system in a week. I gotta say, these two worlds have no intersection. So what HA7CH wants to do is send you there. It's quite hard for you to find this boss. Getting them to trust you is much harder. But this path, you don't need to walk through by yourself. We're going to lead you there. You just need to deliver and ship. That's the most familiar thing for you, right? --- Okay, now let's talk more specifically. A summer. You go to a company like this. Not remote. Literally on site. You follow their employees for a couple of days, maybe a couple of weeks, and see how they do their job. How they match a ticket. How they follow a customer. How they put the same number in four different places. Then you go back. Or you just sit on the ground, on the floor, and give them something they can try in maybe two days. Iterate within a couple of hours. And in the end, the boss pays you by how many people I cut. Not by what tech stack you are using. This ticket, this shit, is much higher than finding an internship at Amazon. --- What's more, you get another thing. You get to know what the real customer is thinking. You get to know the cool feature they're not going to use. You get to know what they are willing to pay for. You get to know after you build it, whether people use it or not. There is not a single thing about this you can learn from class. And if this process is shared across the industry, then it's not just a single business. You use this case to knock the second door, to knock the third door. Maybe it's the start point of a SaaS. Maybe it's the entry point for your startup journey. --- One more thing. We did FDE before, and we know what it looks like. So we have a Hatch House. A place where people live, eat, code, talk shit, and drink together. Even if you did nothing in those two months, you got to know a bunch of people just like you. Which is pretty insane. Pretty interesting, to be frank. --- HA7CH is not a school. We don't teach you how to prompt. We don't need to teach you how to fine-tune. That's easy for you, right? And we are not an outsourcing company. We are not going to pay you by function, by feature, by PRD. HA7CH is just a path. A summer. You go from your course homework into a real industry, finish a real delivery, get a real fucking check. Adding one more line on your resume is so different from running a real business of your own. Those are two different things. You choose it yourself. --- If you think you are this kind of person, just come to ha7ch.com. Our first batch is running right now. See you soon. ## 中文 我们就假设你是一个 AI native 的人。我问你一个问题。 你做过的最酷的东西，是什么？ 可能是某门傻逼课的期末大作业。可能是某次 hackathon 上的一个项目。或者就是某个周末，你用 Claude Code 拼了一个小 demo 出来，截图发了朋友圈，一堆人点赞。 然后呢？ 很多事情就停在那儿了。 寒暑假到了，你还是去找一份实习。一个小时 40 块钱，干一些 dog shit 的活儿。你心里清楚这不是你想要的。但你不知道还能去哪儿。 --- 我跟你说件事。 外面真的有人，愿意为你已经会的能力付钱，付一笔你想象不到的钱。 但他们不在你熟悉的那个世界里。 他们是一群你只要待在学校，就基本不可能接触到的老板。开着一家三十来人的传统公司。一年营收几千万。 他们每天的工作是：员工在 Excel、PDF、微信、几个内部系统之间复制粘贴。他们知道这件事不对劲。他们也想用 AI。但他们不会 prompt，他们不知道去哪儿找你们这种 AI native 的人，他们也不相信外包公司、软件公司能给他们做出他们真正想要的东西。 但你不一样。你会 Claude Code，你能写前端，后端也懂个大概，你一个人一周能跑出一个能用的小系统。 我得说，这两个世界基本没有交集。 HA7CH 要做的，就是把你送过去。 你自己去找这种老板很难。让他们信你，更难。这段路你不用自己走，我们带你过去。你只管交付，只管 ship。这本来就是你最熟悉的事，对吧？ --- OK，再具体说说。 一个暑假。你进到一家这样的公司。不是远程，是真的进现场。你跟着他们的员工看几天，或者一两个礼拜，看他们到底怎么干活。怎么对一张单。怎么跟一个客户。怎么把同一个数字在四个地方填四遍。 你就直接蹲在他们公司里，两天就给他们一个能跑的东西，再过几个小时就迭代一版。 最后，老板按「我省了几个人」付你钱。不是按「你用了什么 tech stack」付你钱。 这一单，比你去大厂找一份实习，钱多得多。 --- 你拿到的还不止是钱。 你会知道真实客户脑子里到底在想什么。你会知道你自以为很酷的那个功能，他根本不用。你会知道他真正愿意为什么东西掏钱。你会知道东西交付以后，到底有没有人在用。 这些没有一件，是你能在课上学到的。 而且如果你做的这个流程，恰好是行业里很多公司都有的，那它就不只是一单生意。你拿着第一家的案例去敲第二家、第三家的门。这可能就是一个 SaaS 的起点。可能就是你创业的起点。 --- 还有一件事。 我们自己做过 FDE，知道前期手头紧是什么感觉。所以我们办了一个 Hatch House。一个让大家住到一起、一起吃饭、一起写代码、一起扯淡、一起喝酒的地方。 就算这两个月你最后什么都没做成，你也认识了一群跟你一样的人。这件事本身就够爽了。说真的，挺有意思。 --- HA7CH 不是学校。我们不教你怎么 prompt，也不用教你怎么 fine-tune。这些对你来说太简单了，对吧？ HA7CH 也不是外包公司。我们不会让你照着 PRD 一条一条按功能报价。 HA7CH 就是一条路。 一个暑假。从课程作业的世界里走出来，走进一个真实的行业，完成一次真实的交付，拿到一笔 real fucking money。 简历上多一行，跟你自己第一次做成一单生意，是完全两件事。你自己选。 --- 如果你觉得你就是这种人，来 ha7ch.com。 我们第一批已经在跑了。 下次再见。 --- # Three Hundred Strangers / 三百个陌生人 > Published 2026-05-20 · By lawted · Canonical: https://ha7ch.com/writing/three-hundred-strangers ## English I was at a supermarket when Lawted texted me: 'bro check out this groupchat on 小红书.' He'd released a prototype of Raily a few hours earlier. By the time I opened the link, three hundred people were already in the chat. Feature requests. Bug reports. Screenshots. Someone asking when the Android version was coming. Nobody in that chat was introduced by anyone's dad. Nobody owed anyone a favor. None of them would have been in the same room a decade ago. The entry fee was that the thing existed and was worth being around. The fact that the room exists means the end of an era most people haven't noticed is ending. --- For a hundred years, 人脉 was the operating system. Your dad knew someone's dad. The dinner was the interview. The red envelope was the contract. You couldn't verify a stranger from outside their network, so the network was the verification. Information was scarce. Capital was scarce. Trust was scarce. All three flowed through the same pipes, and those pipes had names on them. This system built fortunes. It built industries. It built most of the wealth your parents are proud of. It's over. Information isn't scarce anymore. You can verify a stranger in thirty seconds. The work is the signal now, and the signal is public. A twenty-two-year-old shipping from a dorm room is more legible to the people who matter than someone's nephew with a polished résumé and a recommendation letter from a vice president. Capital isn't scarce anymore. Raily cost Lawted a few sleepless nights. That's the whole budget. When building was expensive you needed money, and money moved through networks, which is why the networks mattered. The son of a billionaire used to have a moat ten miles wide. Today he has nothing a hungry working-class kid with a laptop can't match by Friday. And the rooms that matter have new bouncers. The bouncer isn't a person anymore. The bouncer is a filter: can you ship, can you see, are you worth three hours of someone's evening. You can't bribe that filter. You can't get your dad to call it. The work is the only password, and there are no backdoors. Every advantage the old system sold you — access, introductions, the family name, the right school, the right firm — is collapsing in real time. Most of the people holding those advantages haven't realized they're holding nothing but a piece of paper. --- The cliché says you're the average of the five people you surround yourself with. So everyone optimizes for access. How do I get into the room. How do I meet the right people. How do I get the introduction. That's the loser's game now. --- The old system isn't fully dead. There are industries and cities where 关系 still runs everything and will for another generation. Fine. Let the people who inherited that game keep playing it. It's a shrinking board. Every year more verification moves online. Every year more building gets cheaper. Every year more rooms form around the work instead of around the bloodline. The direction is one-way and the slope is steepening. If you're betting on this, you're betting early. Early is the only time the bet pays. Stop trying to climb into rooms. Build something loud enough that the room comes to you. ## 中文 我在超市的时候，Lawted 给我发消息：'哥你看看这个小红书的群聊。' 他几个小时前刚放出了 Raily 的原型。等我点开链接的时候，群里已经三百人了。功能建议。Bug 反馈。截图。有人在问安卓版什么时候出。 这群里没有一个人是被谁的爸爸介绍进来的。没有谁欠谁人情。十年前他们根本不可能聚在同一个房间。入场费就是——这个东西存在，而且值得围观。 这个房间能存在本身，就意味着一个时代的终结。大多数人还没意识到它正在终结。 --- 一百年来，人脉一直是这套系统的底层操作系统。你爸认识谁的爸。饭局就是面试。红包就是合同。你没办法从网络外部去验证一个陌生人，所以网络本身就是验证。信息稀缺。资本稀缺。信任稀缺。这三样东西都通过同一套管道流动，而那些管道上都写着名字。 这套系统造就了财富。造就了行业。造就了你父母引以为傲的大部分家底。 它结束了。 信息不再稀缺。验证一个陌生人只要三十秒。作品就是信号，而且这个信号是公开的。一个在宿舍里发版本的二十二岁年轻人，在真正重要的人眼里，比某某副总裁推荐信里那个简历光鲜的侄子要清晰得多。 资本不再稀缺。Raily 的全部成本就是 Lawted 几个失眠的晚上。仅此而已。建造昂贵的时代，你需要钱；钱在网络里流动，所以网络才那么重要。亿万富翁的儿子曾经有十英里宽的护城河。今天，一个饥渴的、有台笔记本电脑的工薪阶层小孩，周五之前就能追平。 而真正重要的房间，有了新的门卫。门卫不再是人了。门卫是一个过滤器：你能不能交付，你能不能看见，你值不值得别人花三个小时的晚上。你贿赂不了这个过滤器。你爸打电话也没用。作品是唯一的密码，而且没有后门。 旧系统卖给你的所有优势——人脉、引荐、家族名号、对的学校、对的公司——都在实时崩塌。大多数还握着这些优势的人，还没意识到他们手里握的只是一张废纸。 --- 那句老话说，你是你身边五个人的平均值。所以大家都在优化'进入'——怎么进那个房间。怎么认识对的人。怎么搞到那个引荐。 这是输家的游戏了。 --- 旧系统没有完全死。有些行业、有些城市，关系还在主导一切，而且还会再主导一代人的时间。行。让那些继承了那套游戏的人继续玩吧。棋盘在缩小。 每一年，更多的验证搬到线上。每一年，建造的成本继续下降。每一年，更多的房间是围绕作品形成的，而不是围绕血统。方向是单向的，坡度还在变陡。 如果你押注在这个转向上，你是在早期下注。早期是唯一押对的时机。 别再想着挤进哪个房间了。 做点声音够大的东西，让房间自己来找你。 --- # The Ignored Continent / 被忽略的大陆 > Published 2026-05-20 · By lawted · Canonical: https://ha7ch.com/writing/the-ignored-continent ## English Here is what happened. A while ago, my friend Lawted met a boss who runs a traditional manufacturing and transportation business. After they talked through the business, Lawted casually helped him install DouBao on his phone. The boss opened it and tried a few things. Face reading, fortune telling, chatting. Then he said one sentence: “Holy shit, this is fucking insane.” This boss runs a company with annual revenue in the tens of millions and dozens of people under him. But he had never used any AI product before. He did not know what ChatGPT was, did not know what Claude was, did not know what a prompt was. His entire business runs on WeChat groups, Excel handoffs, and people brute-forcing the workflow. What did he use DouBao for? Face reading and fortune telling. And it was not the kind of thing where he tried it once and put it down. He was using it every day. Completely hooked. Almost possessed. It was honestly wild. When I heard this story, I sat there stunned for a while. Not because I thought it was funny. Because I suddenly realized something: those of us soaking in the AI world every day and the bosses actually running businesses on human labor live in two completely different worlds. We are discussing Harness Engineering, multi-agent orchestration, whether Codex or Claude Code is stronger. Codex ships on mobile and some people think it is convenient while others complain it is slow. Our feeds are full of this stuff every day. But outside our field of vision, there is an entire continent we cannot see. On that continent, a boss making tens of millions a year sees AI for the first time through a face-reading app and is stunned speechless. Just picture that scene. This is not a joke. It is a signal. A huge signal that almost everyone has ignored. --- Honestly, I used to think the market for AI was inside internet companies, inside big tech, inside teams that already understand technology. I thought AI deployment meant making better tools, stronger agents, and smoother workflows for people who already knew how to use AI. But ever since we started doing HA7CH and actually began touching traditional industries, my view has been completely changed. The largest market is not there at all. The largest market is on every industrial belt in China that you cannot see. Take the shipping-doc logistics business that boss is in. In Shenzhen alone, there are more than 8,000 companies doing this line of work. More than 8,000. And across these companies, the workflows are almost the same. Orders come in through WeChat, documents are made in Excel, PDFs get passed around, people manually check, manually enter, manually chase payments, manually track progress. A company may have twenty or thirty people whose daily work is basically moving something from one system to another, from one spreadsheet to another, from a sentence in a WeChat group into a cell in Excel. What the boss says is one thing. What the employees do is another. What is written in the system is one thing. How it actually works is another. Many of the key rules are not in any document. They live inside the head of one old employee who has been there for more than a decade. A lot of the information is not structured data. It is a voice message in a WeChat group, a tiny line in a PDF, an abbreviation in the notes column of an Excel sheet. And this is just one industry in one city. Multiply that number across the whole country and across every traditional industry: logistics, education, manufacturing, trade, freight forwarding, building materials, restaurant supply chains. You get a number that makes your scalp go numb. Most of these bosses do not use Jike, do not follow AI news, do not know the difference between ChatGPT and DouBao. The only thing they know is: I spend a huge amount on labor every month, it gives me a headache, I want to cut costs and boost efficiency, but I do not know who to call. This is the real PMF for AI. Not helping people who already know how to use AI use it better. Helping people who have never seen AI see it for the first time and feel their heads explode. --- Some people may wonder: if the market is this large, why are the big companies not doing it? Why has nobody eaten it? To be blunt, it is not that they do not want to. It is that they cannot. When big companies do FDE, companies like Palantir, OpenAI, and Anthropic have mature platforms behind them: Foundry, Claude Enterprise, entire delivery systems. They face Fortune 500 companies and clients with budgets in the millions or tens of millions. They cannot send a dedicated team to deeply customize software for a logistics company with thirty or forty people, a local education business, or a small traditional factory. Domestic companies like MiniMax and Zhipu follow a similar logic. Enterprise customers do not want “I chat with AI for a bit.” Enterprises want models that can enter the intranet, connect to existing systems, and handle concrete business scenarios. Delivery costs are high, so the FDE teams at model companies can only prioritize large clients. VCs do not fund this kind of business either. It is too scattered, too dirty, too non-standardized. It does not tell a clean exponential-growth story. But the problem is, the opportunity is exactly hidden in these places. The more local, messy, and labor-heavy something is, the more room there is for AI to transform it. Traditional outsourcing companies do not do this well either. Communication costs are high, delivery cycles are long, pricing is not cheap, and the final product often does not work well. Outsourcing also charges by feature: you tell me what you want, I build it. But these traditional companies do not have the problem “I need a feature.” Their problem is “my entire workflow is chaotic, and I cannot even explain what I need.” This is the gap. Big companies cannot enter, outsourcing cannot do it well, and the boss cannot figure it out alone. And what HA7CH wants to do is get inside this gap. --- Back to the DouBao story. That boss was blown away by a face-reading app, but what he actually needs is not fortune telling. What he needs is someone to help him rewrite all those repetitive, labor-stacked workflows with AI. What he needs is a young person with a MacBook and a $200 Claude Code plan to walk into his company, sit beside his employees, and watch how they actually work every day. Then, over two or three months, turn the most painful workflows into a system that actually runs. That person is the FDE: Forward Deployed Engineer. But our kind of FDE is different from big-company FDE. Behind a big-company FDE is a massive model platform. Behind our kind of FDE there may just be a MacBook, a $200 Claude Code plan, a few APIs, a WeChat group, and one person brave enough to walk into the company on-site. It sounds very local. But precisely because it is local, it can enter places the big companies cannot enter. And why does this work now? Because AI coding tools have amplified the delivery ability of an individual by too much. In the past, a small enterprise system needed a small team working for two or three months. Now, a sufficiently strong builder using Claude Code, Cursor, Codex, and tools like these may be able to ship an MVP in two weeks. In the past, juggling five or six projects at once was basically impossible. Now, if the method is right, it really can be done. This turns “enterprise solutions” from an organizational capability into a strong-individual capability. --- This is also why we say HA7CH is not an outsourcing company, not a bootcamp, and not a startup community. HA7CH is an FDE accelerator. We send AI-native builders into traditional companies on-site. During the day, they interview operations people and watch how they work, how they fill forms, how they reconcile orders, how they copy and paste, how they move back and forth between WeChat and Excel. At night, they go back and code what they saw into a system. The next day, they bring it back on-site and let the employees try it. If it is wrong, they change it. If it is right, they keep pushing forward. The boss pays based on “how many people did this save,” not based on “what technology did you use.” This logic is very simple and very real. If a company has 10 operators, each making 8,000 RMB a month, that is nearly 1 million RMB in labor cost a year. If your system can turn the work of 10 people into something 3 to 5 people can handle, charging 10% to 20% of the saved cost, 100,000 or 200,000 RMB is not exaggerated at all. And the more important point is: this system is not something you can only sell to one company. In the same industry, workflows are largely similar. After you finish the first company, you take the case study to its peers. You do not need to preach the future of AI again. You just say, this company is already using it. This workflow used to take this many people, now it saves this many people. They used to process this many orders per day, now they can process this many. The first company is the show apartment. The ones after that can become a repeatable industry product. If the first boss is smart enough, he may even become your angel investor. Because he understands very clearly that if this system works, there is no way you will only sell it to him. You will definitely sell it to his competitors. So he will think: can I get a seat first? That is the part of FDE that is genuinely sexy. At the beginning, it may look like you are working for free, like you are doing outsourcing. But if you pick the right industry, the right first customer, and the right repeatable workflow, what comes after is not outsourcing at all. It becomes SaaS. It may even become an industry-level AI system. --- At this point, some people may ask: what kind of people are you actually looking for? I will say it directly. Two traits. You need both. Ability and time. First, ability. You need to be able to build. You do not have to be the strongest engineer, but you must be able to make something from zero. You know how to use Claude Code, Cursor, Codex, and tools like these; how to quickly set up a system; how to connect APIs; how to build a page; how to deploy. If you are already using vibe coding to make small projects in daily life, you probably already have the basic ability. Then, time. This is extremely important. FDE is not remote coding work. You have to actually go on-site, to the company, the office, maybe even the warehouse, and stand beside front-line employees watching how they work. You need to be able to talk to the boss and also to the operations people. You need to understand the very local industry language they use. This takes at least several continuous weeks, and sometimes two or three months. So I genuinely think college students are a very suitable group. Not only college students. If you are a freelancer, a developer who has left a job, or someone on a gap year, as long as you meet the two conditions above, you can do it. But college students have several natural advantages that are hard for other groups to match. They have time. A winter or summer break of two or three months is just enough to go on-site and complete one project. They have momentum. They have not yet been trained by big-company process. They are willing to try, willing to run, willing to throw themselves into an unfamiliar industry. And most importantly, this generation of students already naturally knows how to use AI tools. They know how to use Claude Code, how to use Cursor, how to open issues, how to send PRs, how to quickly fork something and modify it. Think about it: a sophomore spends two months in the summer helping a logistics company build a system. That system later sells to the second and third companies in the same industry. Every month, he can still receive some revenue share from it. That money might cover living expenses, buy equipment, fund the next project. That feeling is completely different from having a job. It lets you feel, for the first time, that something you built can actually make money in the real world. Not salary. Cash flow you created yourself. This experience cannot be taught by school, and big-company internships may not teach it either. Taking ten classes in school is not as good as actually going on-site to a company, watching how a boss makes money, how an operations person works, how a system goes from zero to one and starts running. You learn product, engineering, sales, delivery, business, industry, and human nature all at once. This is the best kind of learning. --- Of course, I do not want to make this sound too beautiful. To be blunt, doing FDE is hard. You do not have a proper desk. You do not have a proper work environment. Everyone in the office may be smoking. The boss may pour tea for you, or he may ask you to drink with him. Your first project will probably only be paid after delivery, maybe with not even one yuan of deposit. You may put your own time into it, go on-site every day, and write code until midnight. And what you face is not standardized requirements. Real business is not written in a PRD. The boss will not write user stories for you. Employees will not tell you the full workflow either. You have to watch, ask, break it down, and judge by yourself. You will run into a mess of fields, historical baggage, and endless cases of “this customer is special.” This is not something everyone can do. But if you can do it, and you dare to do it, the return is also very real. Because in traditional industries, you suddenly become a very scarce person. When you tell them AI can help reconcile orders, organize spreadsheets, automatically generate files, analyze customer data, and take over the daily copy-paste work, they will really think: holy shit, this person has something. Inside a big company, you may just be an ordinary engineer. At Stanford, you may just be an ordinary research assistant. But when you arrive in a traditional industry that truly lacks AI, automation, and technical understanding, you suddenly become a key person. Technology has completely different value depending on where you put it. --- I am not sure where HA7CH will ultimately go. Maybe a few months from now we find out some assumptions were wrong. Maybe Hatch House is a false premise. Maybe the business model needs a major adjustment. All of that is possible. But there is one thing we are certain about. That ignored continent is real. Those bosses who have never seen AI truly need someone to walk in. Those workflows stacked on human labor are truly waiting to be rewritten. And every company can only be AI-ified once. Whoever gets in first owns that workflow. Distilling it later becomes extremely difficult. This window will not stay open forever. Big companies will enter within 12 months, but they will start with large clients. The market of small and mid-sized “local bosses” may still have a 12-month window. So, back to the opening story. A boss making tens of millions a year uses DouBao for face reading and fortune telling, and thinks it is incredible. He does not know what else AI can help him do. He does not know that half of the labor cost he spends every day can be taken over by a system. He does not know that his company is waiting to be distilled. But we know. And what we need to do is find those young builders who have ability, have time, and dare to walk into the field, then send them to the side of these bosses. Let them become the first person to walk into these companies. HA7CH, BUILD IN THE FIELD. HATCH INTO IMPACT. If you feel like you are this kind of person, or you want to learn more, come to ha7ch.com. Our first batch is already running. See you next time. ## 中文 事情是这样的。 我的朋友 Lawted 前段时间见了一个做传统制造运输的老板。聊完业务以后，他随手帮老板在手机上装了一个豆包。 老板打开，试了试。看脸，算命，聊天。 然后老板说了一句，「卧槽，这他妈太屌了。」 这个老板，公司年营收过千万，手底下几十号人。但他从来没用过任何 AI 产品。他不知道什么是 ChatGPT，不知道什么是 Claude，不知道什么是 prompt。他的全部业务靠微信群协调、Excel 流转、人力堆砌。 他用豆包干什么呢？ 看脸算命。 而且不是试一次就放下了。是每天都在用。完全上瘾了，跟着了魔似的，真的很夸张。 我当时听完这个事儿，愣了好一会儿。 不是觉得好笑。是突然意识到一件事，我们这些天天泡在 AI 圈子里的人，和这些真正在用人力堆业务的老板，真的活在两个完全不同的世界里。 我们在讨论 Harness Engineering，在讨论多 Agent 编排，在讨论 Codex 和 Claude Code 到底谁更能打。Codex 上了移动端有人觉得方便有人嫌它慢。我们每天的信息流里全是这些东西。 但在我们视线之外，有一整片我们根本看不见的大陆。 那片大陆上，一个年入千万的老板，第一次见到 AI，是被一个看脸算命的 APP 震撼到说不出话。 你想想这个画面。 这不是一个段子。这是一个信号。一个巨大的、几乎所有人都忽视了的信号。 --- 说真的，我之前一直以为 AI 的市场在互联网公司里，在大厂里，在那些已经很懂技术的团队里。我觉得 AI 落地嘛，就是给已经会用 AI 的人做更好的工具，做更强的 agent，做更丝滑的工作流。 但自从我们开始做 HA7CH，真的去接触传统行业以后，我的想法被彻底改变了。 最大的市场根本不在那里。 最大的市场，在中国每一条你看不见的产业带上。 就拿那个老板做的船单物流来说。光深圳一个城市，做这行的公司就有 8000 多家。 8000 多家。 而且这些公司之间，工作流几乎大同小异。都是微信接单，Excel 做单，PDF 传文件，人工核对，人工录入，人工催款，人工跟踪。一个公司里面可能有二三十个人，每天在做的事情就是从一个系统搬到另一个系统，从一个表格搬到另一个表格，从微信群里的一句话搬到 Excel 里的一个格子。 老板说的是一套，员工做的是另一套。系统里写的是一套，真实操作又是一套。很多关键的规则不在任何文档里，而是在某个干了十几年的老员工脑子里。很多信息不是结构化数据，而是微信群里的一句语音、PDF 里的一行小字、Excel 备注栏里的一个缩写。 这还只是一个行业，一个城市。 你把这个数字乘以全国，乘以所有传统行业。物流、教育、制造、贸易、货代、建材、餐饮供应链。你会得到一个让人头皮发麻的数字。 而这些公司的老板，绝大多数都不刷即刻，不关注 AI 新闻，不知道 ChatGPT 和豆包有什么区别。他们唯一知道的事情是，我每个月在人力上花一大笔钱，我很头疼，我想降本增效，但我不知道该找谁。 这才是 AI 真正的 PMF。 不是让已经会用 AI 的人用得更好。而是让从没见过 AI 的人，第一次见到就炸裂。 --- 可能有小伙伴纳闷，这么大的市场，为什么大厂不做？为什么没有人去吃？ 坦率的讲，不是不想做，是做不了。 大厂做 FDE，比如 Palantir、OpenAI、Anthropic，背后是成熟的平台产品，是 Foundry、是 Claude 企业版、是整套交付体系。他们面对的是世界 500 强，是动不动几百万几千万预算的大客户。他们不可能为了一个三四十人的物流公司、一个本地教育机构、一个传统小工厂，专门派一个团队去做深度定制。 国内的 MiniMax、智谱也是类似的逻辑。企业客户要的不是「我跟 AI 聊两句」，企业要的是模型能进内网、能接现有系统、能处理具体业务场景。这些事情的交付成本很高，大模型公司的 FDE 只能优先服务大客户。 VC 也不投这种生意。太散了，太脏了，太不标准化了，讲不出什么指数增长的故事。 但问题是，机会恰恰就藏在这种地方。 越土、越乱、越靠人堆，才越有 AI 改造的空间。 传统外包公司也不好做这个事儿。外包沟通成本高、交付周期长、报价也不低，最后做出来还经常不好用。而且外包是按功能收费的，你要什么我做什么。但这些传统企业的问题不是「我要一个功能」，而是「我整个流程都是乱的，我自己都说不清我需要什么」。 这就是那个缝隙。大厂进不来，外包做不好，老板自己搞不定。 而 HA7CH 要做的，就是钻进这个缝隙里。 --- 回到那个装豆包的故事。那个老板被一个看脸算命的 APP 震撼到不行，但他真正需要的不是算命。他需要的是有人帮他把那些每天重复的、靠人堆出来的流程，用 AI 重新写一遍。 他需要的是，一个带着 MacBook 和 Claude Code 200 刀套餐的年轻人，走进他的公司，坐在他员工旁边，看他们每天到底怎么干活。然后用两三个月时间，把那些最痛的流程做成一个能跑的系统。 这个人，就是 FDE。Forward Deployed Engineer，前沿部署工程师。 但我们这种 FDE，跟大厂的不一样。 大厂 FDE 背后是一个巨大的模型平台。我们这种 FDE 背后可能就是一个 MacBook、一个 Claude Code 200 刀套餐、几个 API、一个微信群，和一个敢走进企业现场的人。 听起来很土。 但也正因为土，所以它能进入那些大厂进不去的地方。 而且现在为什么这事能成立？因为 AI coding 工具真的把个体的交付能力放大了太多。以前一个企业小系统，需要一个小团队做两三个月。现在一个足够强的 builder，用 Claude Code、Cursor、Codex 这些工具，可能两周就能跑出一个 MVP。以前你要同时 handle 五六个项目基本不可能，现在如果方法对，真的可以做到。 这件事把「企业解决方案」从一个组织能力，变成了一个强个体能力。 --- 这也是为什么我们说 HA7CH 不是外包公司，不是培训班，也不是创业社群。 HA7CH 是一个 FDE 加速器。 我们把 AI-native 的 builder 送到传统企业现场。白天去访谈业务员，看他们怎么工作，怎么填表，怎么对单，怎么复制粘贴，怎么在微信和 Excel 之间来回切。晚上回去 coding，把今天看到的东西写成系统。第二天再拿回现场，让业务员试。 不对就改。对了就继续往下推。 老板按「省了多少人」付钱，不按「用了什么技术」付钱。 这个逻辑非常简单，也非常真实。一个公司如果有 10 个运营，每人月薪 8000，一年就是接近 100 万的人力成本。如果你的系统能帮他把 10 个人的活变成 3 到 5 个人就能做，收他节省成本的 10% 到 20%，收个 10 万、20 万，一点都不夸张。 而且更关键的一点是，这个系统不是只能卖给一家。同一个行业里，工作流大同小异。做完第一家，你拿着案例去找同行，不用再重新讲什么 AI 未来。你直接说，某某公司已经用了，原来多少人做这个流程，现在省了多少人，原来一天处理多少单，现在一天能处理多少单。 第一家是样板间。后面的，是可以复制的行业产品。 如果第一个老板够聪明，他甚至可能直接变成你的天使投资人。因为他很清楚，你这个系统只要好用，不可能只卖给他一家。你一定会卖给他的同行。那他就会想，我能不能先占一个坑。 这才是 FDE 真正性感的地方。 你一开始看起来像是在免费干活，像是在做外包。但如果你选对行业，选对第一个客户，选对那个可以复制的工作流，它后面就完全不是外包。它会变成一个 SaaS。甚至变成一个行业级的 AI 系统。 --- 说到这个，可能有人会问，你们到底想找什么样的人。 我直说了。两个特征，缺一不可。 有能力，有时间。 先说能力。你要会 build。你不一定是最强的工程师，但你一定要能把东西从 0 做出来。你知道怎么用 Claude Code、Cursor、Codex 这些工具，怎么快速搭系统，怎么接 API，怎么做页面，怎么部署。如果你平时就在用 vibe coding 做各种小项目，那你大概率已经具备基础能力了。 再说时间。这一点非常关键。FDE 不是远程写代码的活儿。你要真的去现场，去公司，去办公室，甚至去仓库，站在一线员工旁边看他们怎么干活。你要能跟老板聊，也要能跟业务员聊。你要能听懂他们讲的那些很土的行话。这需要至少连续几周到两三个月的投入。 所以我是真的觉得，大学生是一个非常适合的群体。 不是说只要大学生。如果你是自由职业者、离职程序员、gap year 的人，只要满足上面两个条件，都可以。但大学生有几个天然的优势，是其他群体很难具备的。 有时间。寒暑假两三个月，刚好可以驻场做一单。 有冲劲。还没有被大厂流程驯化，愿意试、愿意跑、愿意把自己扔到一个陌生的行业里。 而且最关键的是，这一代学生已经天然会用 AI 工具了。他们知道怎么用 Claude Code，怎么用 Cursor，怎么提 issue，怎么发 PR，怎么快速 fork 一个东西再改出来。 你想想看，一个大二的学生，暑假两个月，帮一家物流公司做了一套系统。这套系统后来卖给了同行的第二家、第三家公司。他每个月还能从里面拿到一些分成。这些钱可能帮他 cover 生活费，帮他买设备，帮他做下一个项目。 这个感觉和打工是完全不一样的。它会让你第一次感受到，我做出来的东西，真的可以在真实世界里赚钱。 不是工资。是你自己创造出来的现金流。 这种体验，学校教不了，大厂实习也不一定教你。你在学校里学十门课，都不如你真的去一个公司现场，看一个老板怎么赚钱，看一个业务员怎么干活，看一个系统怎么从 0 到 1 跑起来。你会同时学到产品、工程、销售、交付、商业、行业、人性。 这才是最好的学习。 --- 当然，我也不想把这件事说得太美好。坦率的讲，做 FDE 很苦。 你没有工位，没有正经的工作环境，办公室里可能所有人都在抽烟。老板可能给你端茶倒水，也可能要你陪酒。你做的第一单大概率是交付以后才给钱，甚至一分钱定金都没有。你可能贴时间进去，可能天天去现场，可能晚上写代码写到 12 点。 而且你面对的不是标准化的需求。真实业务不是 PRD 里写好的。老板不会给你写 user story。员工也不会告诉你完整流程。你要自己看，自己问，自己拆，自己判断。你会遇到一堆乱七八糟的字段，一堆历史包袱，一堆「这个客户比较特殊」的情况。 这不是每个人都能干的事。 但如果你能干，而且你敢干，回报也是很真实的。 因为在传统行业里，你会突然变成一个很稀缺的人。你跟他们说 AI 可以帮他们对单、整理表格、自动生成文件、分析客户数据、把每天复制粘贴的活儿接走，他们真的会觉得，我操，这个人有点东西。 在大厂里你可能只是一个普通工程师。在 Stanford 你可能只是一个普通研究助理。但你到了一个真正缺 AI、缺自动化、缺技术理解的传统行业里，你突然就会变成一个很关键的人。 技术这东西，放在不同的地方，价值完全不一样。 --- 我自己也不确定 HA7CH 这件事最后能走到哪里。也许几个月后发现某些假设是错的，也许 Hatch House 是个伪命题，也许商业模式需要大调整。这些都有可能。 但有一件事我们是确定的。 那片被忽略的大陆，是真实存在的。 那些从没见过 AI 的老板，是真的需要有人走进去的。 那些靠人力堆出来的工作流，是真的在等着被重写一遍的。 而每一个公司，有且只能被 AI 化一次。先进去的人，就拥有了那个工作流。后面再想蒸馏，就非常困难了。 这个窗口期不会永远存在。大厂 12 个月内会进场，但他们会从大客户做起。中小「土老板」市场，可能还有 12 个月的窗口。 所以回到开头那个故事。 一个年入千万的老板，用豆包看脸算命，觉得太牛逼了。 他不知道 AI 还能帮他做什么。他不知道他每天花在人力上的成本，有一半可以被系统接走。他不知道他的公司正等着被蒸馏。 但我们知道。 而我们要做的，就是找到那些有能力、有时间、敢走进现场的年轻 builder，把他们送到这些老板身边去。 让他们成为走进这些公司的第一个人。 HA7CH，BUILD IN THE FIELD. HATCH INTO IMPACT. 如果你觉得自己就是这种人，或者你想了解更多，来 ha7ch.com 看看。我们第一批 batch 已经在跑了。 我们，下次再见。 --- # Harvard Isn't Harvard, YC Isn't YC / 哈佛不是哈佛，YC 也不是 YC > Published 2026-05-19 · By lawted · Canonical: https://ha7ch.com/writing/harvard-is-not-harvard ## English Lately we've been chewing on one question: what exactly is ha7ch? Not a business-model question. Not a fundraising-story question. Just very plainly: what are we actually building? --- China has a massive number of mid-sized companies. Dozens of people, hundreds, sometimes several hundred. The operations are already too complex for Excel, but they can't afford a traditional software vendor. Before, they had two options: drop several million on a custom platform, or keep grinding it out by hand. So entire industries got stuck in a “semi-digitalized” limbo. Then AI native coding showed up. Claude Code, Cursor, vibe coding... they crushed the cost of writing software to a level no one would have dared to imagine. Suddenly a lot of needs that “weren't worth doing” were worth doing. That was the opportunity we saw first. But later we realized it might only be the surface. --- We suddenly clicked on something: the core of ha7ch might not be software at all. It's filtering people. Today's college students aren't short on tutorials, courses, or Hackathons. What they're short on is the first real entry into the real world. The first time they realize: “Wait, what I built is actually being used.” “Wait, a system I made actually saved a company money.” “Wait, I can make my first real money off my own skill.” After a Hackathon ends, the project never gets opened again. A real company is different. A real company yells at you every day, chases you every day, says there's a bug here every day, says the flow is wrong every day. And precisely because of that, you actually enter the real world. So ha7ch isn't a bootcamp. It's a funnel for AI native builders. We keep filtering: who can actually communicate, who can actually walk into a company, who can actually understand the business, who can actually deliver, who can actually finish. A real builder isn't just someone who can write code. A real builder is someone who can turn the mess of the world into a system. --- A lot of people ask: “Are you guys handing money to top college students?” No. Handing out money has no meaning. What matters is letting them earn money for the first time. That feeling is completely different. The moment someone realizes “holy shit, I actually made money doing this,” their worldview shifts. And the shift is irreversible. A lot of people never get into that state in their whole life. They only live inside the GPA, grad-school, internship, offer, ranking game. The real world has another game. Some people are wired for research, some for enterprise, some for starting things, some for 0-to-1, some for 1-to-100. Jack Ma wasn't Tsinghua's top student. Not everyone has to become the top of the academic ladder. What matters is: have you found your own battlefield? --- Then we figured out one more thing: why do some organizations end up so strong? Not the courses. Not the office. Not the logo. It's the people inside. Why is PayPal Mafia strong? Because that group later scattered across the Valley and built Tesla, LinkedIn, YouTube, Palantir. Why is YC strong? Because it keeps filtering founders, and the alumni network keeps compounding. Why is Harvard, Harvard? Because inside Harvard is that group of people. Without those people, Harvard isn't Harvard. So we increasingly think: the truly valuable thing isn't code. It's who you pulled all-nighters with, who you shipped projects with, who you failed with, who you raced a deadline with in a rented Shenzhen apartment. These relationships stay with you for years and years. --- Honestly, we haven't figured out what ha7ch finally turns into. But one thing is getting clearer: it doesn't necessarily need to be commercial, at least not at the start. The moment you stare at monetization from day one, you start unconsciously doing “things that make money” instead of “the right things.” What we want is to gather people first, let things happen first, let young people enter the real world and get results first. It's more like a hybrid: a community of AI native builders, a filter, a real-world training ground, a resource network for young people. If we have to analogize, it's closer to early YC. Not a commercial product but a mechanism. Its value isn't in a revenue report. It's in the people who walk out of it. “I came out of ha7ch.” We hope one day that sentence carries weight. --- A lot of people like to argue these days about whether AI will replace programmers. But we increasingly think the truly hard-to-replace ability is a different one: can you talk to the boss, can you read a business workflow, can you walk into an unfamiliar industry, can you take the chaos, can you marshal resources, push things forward, land a vague requirement into something real. AI has a hard time replacing these. And this might be the most important ability for the next generation of builders. What ha7ch wants to do is actually simple: use the real world to filter out the next generation of AI native builders. Monetization, later. ## 中文 我们最近一直在聊一件事：ha7ch 到底是什么？ 不是商业模式的问题，也不是融资故事的问题。就是很纯粹地在想：我们到底在做一个什么东西？ --- 中国有大量中型企业。几十人，上百人，甚至几百人。业务已经复杂到 Excel 顶不住了，但又请不起传统软件公司。 以前他们只有两个选择：花几百万上千万搞中台，或者继续人工硬撑。于是大量行业永远卡在“半数字化”的状态里。 然后 AI native coding 出现了。Claude Code、Cursor、vibe coding……把软件开发成本压到了一个以前不敢想的程度。突然之间，很多原本“不值得做”的需求，现在居然值得做了。 这是我们最早看到的机会。但后来发现，这可能只是表层。 --- 后来我们突然意识到：ha7ch 最核心的东西，可能根本不是软件。而是筛人。 现在的大学生不缺教程，不缺课程，不缺 Hackathon。他们缺的是第一次真正进入真实世界。第一次知道： “原来我写的东西真的有人在用。” “原来一个系统真的能帮企业省钱。” “原来我可以靠自己的能力赚到第一桶金。” Hackathon 做完以后，项目一辈子没人打开第二次。但真实企业不是。企业会天天骂你，天天催你，天天说这里有 bug，天天说流程不对。但也正因为这样，你才真正进入了现实世界。 所以 ha7ch 的本质不是培训班，而是一个 AI native builder 的漏斗。我们在不断筛选：谁真的能沟通，谁真的能进企业，谁真的能理解业务，谁真的能交付，谁真的能把事情做完。 真正的 builder，不是只会写代码的人，而是能把混乱世界变成系统的人。 --- 很多人会问：“你们是不是给优秀大学生发钱？” 不是。直接给钱没有意义。真正重要的是，让他第一次赚到钱。那个感觉完全不一样。 一旦一个人发现“卧槽，我靠这个东西真的赚到钱了”，他的世界观会变。而且这种改变是不可逆的。 很多人一辈子都没进入过这种状态。他们只活在 GPA、保研、实习、offer、ranking 的那套游戏里。但真实世界还有另一套游戏。有人适合做 research，有人适合做企业，有人适合创业，有人适合做 0 到 1，有人适合做 1 到 100。 马云也不是清华第一名。不是所有人都要成为学术最顶尖的人。真正重要的是：你有没有找到自己的战场。 --- 后来我们想明白了一件事：为什么有些组织最后会变得那么强？ 不是因为课程。不是因为 office。不是因为 logo。而是因为那群人。 PayPal Mafia 为什么强？因为那群人后来散落到了整个硅谷，创了 Tesla、LinkedIn、YouTube、Palantir。YC 为什么强？因为它持续在筛选创业者，校友网络越滚越大。哈佛为什么是哈佛？因为哈佛里面是那群人。 如果没有那群人，哈佛也不是哈佛。 所以我们越来越觉得，未来真正值钱的东西不是代码，而是：你和谁一起熬过夜，你和谁一起做过项目，你和谁一起失败过，你和谁一起在深圳的出租屋里赶过 deadline。这些关系会跟着你很多很多年。 --- 说实话，我们现在也没想清楚 ha7ch 最终会长成什么样。但有一件事越来越确定：它不一定需要商业化，至少前期不需要。 如果一开始就盯着变现，你会不自觉地去做“能赚钱的事”，而不是“对的事”。我们现在更想做的是，先把人聚起来，先让事情发生，先让年轻人进入真实世界拿到结果。 它更像一个混合体：一个 AI native builder 的社区，一个筛选系统，一个现实世界训练场，一个年轻人的资源网络。 如果非要类比的话，可能更接近早期的 YC。不是一个商业产品，而是一种机制。它的价值不在营收报表里，而在于从里面走出来的那群人。 “我是从 ha7ch 出来的。”我们希望有一天，这句话是有分量的。 --- 现在很多人喜欢讨论 AI 会不会替代程序员。 但我们越来越觉得，真正难替代的是另一种能力：你能不能和老板聊天，能不能看懂业务流程，能不能进入一个陌生行业，能不能扛住混乱，能不能组织资源、推动事情、把模糊需求真正落地。 这些东西，AI 很难替代。而这也可能是下一代 builder 最重要的能力。 ha7ch 想做的事情其实很简单：用真实世界，筛出下一代 AI native builder。 商业化的事，以后再说。 --- # Stop Saying "Jiushi" / 不要再说“就是” > Published 2026-05-17 · By lawted · Canonical: https://ha7ch.com/writing/stop-saying-jiushi ## English This is not exactly a "practical tips" essay. It is more like a small language alarm, and also a bit of life thinking from the AI era. Lately I have felt more and more strongly that there is one word we should probably say less. Ideally, we should consciously try to quit it. That word is "jiushi." Of course, "jiushi" in Chinese is not some original sin. It has many normal uses. Sometimes it is just an ordinary copula. Sometimes it is just a spoken connector. Sometimes it is even just a filler sound people reach for while thinking. The problem is not the word itself. The problem is that we often use it to skip ahead. A lot of the time, once a "jiushi" comes out, the thinking that should follow has already been sealed off in advance. The most common scene is not explanation, but rhetorical questioning. For example: "Isn't that what you mean?" "Isn't this just avoidance?" "Doesn't that just show you do not really want it?" On the surface, this sounds like discussion. In reality, the conclusion has already been stuffed into the question. It is not opening understanding. It is setting the default. It is not inviting the other person to think together. It is speaking the other person's meaning to death. A lot of people feel oppressive in conversation not because their tone is especially fierce, but because this posture of "I have already summarized you" arrives too quickly. After spending a long time with AI, this actually becomes easier to see. A reasonably tuned model usually will not rush into a rhetorical question like that, and it usually will not slap a sentence like "aren't you just..." onto someone's head right away. What it does more often is first try to understand the context, first identify ambiguity, first offer several possible interpretations, and then slowly narrow them down. A lot of the time, it is even clumsy in how much it confirms: am I understanding this correctly? Is this what you mean? That kind of caution can feel verbose, but at least it shows one thing: real understanding should not be built on defaults. It should be built on space. The most dangerous thing about "jiushi" is not that it is rude, and not that it is too colloquial. It is that it can so easily create the illusion of "I have already figured this out." Especially in Chinese, the word is too convenient. So convenient that a lot of the time, before the mind has really turned the corner, the mouth has already defined things on its behalf. It looks fluent. In reality, it is skipping steps. A question that was still worth thinking through one more layer, a place where one more question could still be asked, a relationship that had not yet been truly clarified: once a "jiushi" covers it over, what follows often stops unfolding. That is also why I have recently become especially sensitive to this word. Model thinking takes time. AI today is already faster and faster, and better and better at simulating the feeling of "I get it." But any reasoning that is halfway decent still needs context, still needs disambiguation, still needs to put several possibilities next to each other and compare them. People are actually the same. But in real life, many people open their mouths with "jiushi," as if within one second they have already completed understanding, judgment, summarization, and classification. But how could it be that fast? Many so-called "jiushi" moments are not expressions after thinking has finished. They are shortcuts before thinking has begun. So lately I have been somewhat serious about quitting this word. Not because it is low-class, and not because it lacks elegance, but because once you say it a little less, you realize that many times you actually had not thought that far. The place that "jiushi" wanted to jump over is exactly the place most worth pausing. Why do I understand it this way? Is there another possibility? Am I stating a vague problem too fully? Am I sealing off something that could still be questioned further? In a sense, quitting "jiushi" is not training diction. It is training a more honest way of thinking. It forces a person to admit: I may not understand this yet. I may still need to think about this. I cannot reach a conclusion that quickly. This sense of pause feels more and more important in this era. AI is already very clear in its contextual logic. It is good at organizing information, sorting out structure, and laying out several possibilities. If we still keep some advantage that is more decent than the machine's, it may not be speed, and it may not be being "smarter." It may be understanding. The understanding I mean here is not just "knowing" something. It is really entering into it, admitting that it may be more complicated than your first reaction, admitting that you may not have grasped the point immediately, and being willing to leave some room in discussion between people. Understanding is not some lofty posture of empathy, either. It is a very plain ability: knowing that you have limits, being willing to ask further, being willing to let a question stay with you for a little longer instead of rushing to answer first. I originally wanted to connect this to Andrej Karpathy, but the more accurate version is that in his public discussions in recent years, Karpathy has repeatedly pushed human taste, judgment, and understanding to the front, rather than simply offering the slogan "the only moat humans have is understanding." A steadier way to put it is: the stronger AI becomes, the more important human judgment, taste, and understanding become. I agree with that direction. Because AI can already "think for a minute before answering." It can already simulate caution, simulate reasoning, and simulate reflection. But it cannot always truly notice where it does not understand. Humans at least still have one ability: in a certain moment, to honestly admit that I may have misunderstood this, I have not thought this through, I need to ask one more question, I need to learn a little more. That action itself is already powerful. So in the end, what I want to say is not language purism, and it is not that "jiushi" should be deleted from Chinese entirely. I just increasingly feel that the way a person uses "jiushi" reveals many things. It reveals whether they are too eager to judge, too eager to summarize other people, too eager to simplify something complex into a ready-made default. It also reveals whether they leave room for understanding, whether they leave time for thinking, and whether they realize that they may not have arrived there yet. If the AI era still leaves humans with any decent homework, I suspect one piece of it is this: do not live yourself into a machine that only knows how to buzz in first. A little less "isn't this just," a little more "let me think again"; fewer defaults, more real understanding; less rushing to define, more willingness to ask. Start by saying one less "jiushi." Maybe that is not a bad exercise. Not because the word is guilty, but because a lot of the time, it arrives too fast. And understanding was never supposed to be that fast. ## 中文 这不是一篇特别“干货”的文章。更像是一篇语言上的小警报，也是一点 AI 时代的生活感想。最近越来越强烈地觉得，有一个词，真的应该少说，最好能有意识地戒掉。这个词就是“就是”。 中文里的“就是”当然不是原罪。它有很多正常用法，有时候只是一个普通的系词，有时候只是口语里的连接词，有时候甚至只是人在思考时顺手垫一下的语气。问题不在这个词本身，而在于它常常被我们用来偷跑。很多时候，一个“就是”出来，后面的思考其实就已经被它提前封口了。 最常见的场景，不是解释，而是反问。比如，“你不就是这个意思吗？”“这不就是在逃避吗？”“那不就是说明你根本不想吗？”这种说法表面上像在讨论，实际上已经把结论塞进了问题里。它不是在打开理解，而是在设定默认值。它不是在邀请对方一起想，而是在替对方把话说死。很多人说话之所以让人觉得有压迫感，不是因为语气有多凶，而是因为这种“我已经替你总结完了”的姿态来得太快。 这一点，和 AI 相处久了之后，反而会看得更明显。一个正常被调教过的模型，通常不会那么急着反问，也不会上来就把一句“你不就是……”扣在人头上。它更常做的事情，是先试图理解上下文，先辨认歧义，先给出几种可能的解释，然后再慢慢收束。很多时候，它甚至笨拙得有点过头，会反复确认：我理解得对不对？是不是这个意思？虽然这种谨慎有时让人觉得啰嗦，但它至少说明了一件事：真正的理解，不该建立在默认值上，而该建立在空间上。 “就是”最危险的地方，不在于粗鲁，也不在于口语化，而在于它特别容易制造一种“我已经想清楚了”的幻觉。尤其是在中文里，这个词太顺手了。顺手到很多时候，脑子还没真正转过去，嘴已经先替自己下了定义。看上去像表达流畅，实际上是在跳步。原本还值得再想一层的问题，原本还可以再问一句的地方，原本还没有真正厘清的关系，一旦被一个“就是”盖过去，后面往往就不会再继续展开了。 这也是为什么我最近开始对这个词特别敏感。因为模型思考是需要时间的。今天的 AI 虽然已经越来越快，越来越会模拟那种“我懂了”的感觉，但真正像样一点的推理依然需要上下文，需要辨义，需要把几种可能性放在一起比一比。人其实也一样。可现实里，很多人一开口就是“就是”，好像一秒钟之内就已经完成了理解、判断、归纳和定性。可哪里有那么快。很多所谓的“就是”，不是思考完成后的表达，而是思考还没开始时的捷径。 所以我最近有一点想认真戒掉这个词。不是因为它低级，也不是因为它不够优雅，而是因为一旦少说一点，才会发现，很多时候自己其实并没有想到那个地步。那个“就是”原本想跳过去的地方，恰恰是最值得停一停的地方。为什么会这样理解？有没有别的可能？是不是把一个模糊的问题讲得太满了？是不是把一个本来还可以继续追问的东西，提前封口了？ 某种意义上，戒掉“就是”，不是在训练措辞，而是在训练一种更诚实的思考方式。它逼着人承认：这里我可能还没懂，这里我还需要再想一下，这里不能那么快地下结论。这种停顿感，在现在这个时代反而显得越来越重要。因为 AI 的上下文逻辑已经很清楚了，它很擅长整理信息、梳理结构、摊开几种可能性。我们如果还保留一点比机器更像样的优势，未必是在速度上，也未必是在“更聪明”上，而更可能是在理解上。 这里说的理解，不只是“知道”一个东西，而是真实地进入它，承认它可能比自己第一反应更复杂，承认自己可能没有一下子抓到重点，也愿意给人与人之间的讨论留一点空间。理解也不是一种高高在上的共情姿态，而是一种相当朴素的能力：知道自己有局限，愿意追问，愿意让一个问题在自己这里多停留一会儿，而不是急着抢答。 我原本想把这件事和 Andrej Karpathy 联系起来，但更准确的说法应该是，Karpathy 在近年的公开讨论里，反复会把人的 taste、judgment、understanding 往前推，而不是简单给出一句“人类唯一的护城河就是理解”的口号。更稳妥地说，AI 越强，人的判断、品味和理解力就越重要。我很认同这个方向。因为 AI 当然已经能“思考一分钟再回答”，也已经能模拟谨慎、模拟推理、模拟反思，但它不总能真的意识到自己哪里没懂。人至少还有一个能力，是可以在某个瞬间认真承认：这里我可能理解错了，这里我还没想透，我得再问一句，我得再学一点。这个动作本身，就已经很厉害了。 所以到最后，我想说的其实不是语言洁癖，也不是要把“就是”从中文里彻底删掉。我只是越来越觉得，一个人怎么用“就是”，其实会暴露很多东西。暴露他是不是太急着下判断，太急着替别人总结，太急着把复杂的东西简化成一个现成的默认值。也暴露他有没有给理解留空间，有没有给思考留时间，有没有意识到自己可能还没到那个地步。 如果说 AI 时代还给人留了什么像样的功课，我猜其中一个就是这个：不要把自己活成一台只会抢答的机器。少一点“这不就是”，多一点“我再想想”；少一点默认值，多一点真的理解；少一点急着定性，多一点愿意追问。 先从少说一句“就是”开始，也许是个不坏的练习。不是因为这个词有罪，而是因为很多时候，它来得太快了。而理解这件事，本来就不该那么快。 --- # Baseball and the Blame Game / 棒球与职场甩锅 > Published 2026-05-14 · By lawted · Canonical: https://ha7ch.com/writing/baseball-and-the-blame-game ## English I got into baseball last year and realized: baseball and the office are basically the same thing. Same rules, same playbook, no one really watching. --- Nine of them against one of you. When it's your turn, you're alone against all of them. The office is no different — you think you have teammates, but you don't. There's exactly one situation where a coworker actually cares about you: they've already taken a base, they need you to not strike out, they need you to move them forward. The moment your interests line up, they care. The rest of the time, nobody is catching the ball for you. --- Every pitch is someone trying to pin something on you. The ball flies in, you don't know if it's meant for you. If it isn't in your zone, that's a ball — don't move. The blame doesn't land, and the person who threw it just exposed themselves. The second you swing, it's yours. The most important skill is reading whether it's coming into your zone. The mistake juniors make is panicking and swinging. --- If it's in your zone, you have to swing. The point isn't to hit the ball back — it's to throw the blame somewhere else. But you can't put it directly into someone's hands. That's a caught fly ball. Everyone saw it. Everyone knows it came from you. Either knock it out of the park — home run — and the blame vanishes. Or hit it where nobody can catch it cleanly, then run like hell and get yourself on base. Getting on base is grabbing onto a boss's leg. The blame is still floating around, but you're standing on something solid. --- Standing on base isn't winning, though. People keep coming for you. Three strikes and you're out. The round is over. How does the strikeout happen so easily? Because your coworkers all know you're going to throw the blame somewhere. They've already taken up every position on the field — wherever you want to throw it, they're already standing there. The angle and force of your swing? They've predicted it. Real home runs — the ones that actually clear the park — are rare. Most balls land inside their range. They reach out and catch. --- The hard part of baseball isn't swinging. It's knowing when not to. But there's a harder call than that — whether you're on the field today, or up in the stands. It's not a difference in job. It's a difference in posture. In the same company, some people are out there grinding for every at-bat, while others sit in the stands with a beer and watch the whole thing play out. --- But honestly, neither of those is right. Baseball is baseball. The office shouldn't be baseball. Nine guys on the field grinding through a game nobody's watching — that's their job. The office isn't supposed to be like that. The office is supposed to be a place where you produce value. What you ship runs or it doesn't. It saved a headcount or it didn't. A customer paid for it or they didn't. There's no blame to pass, because there's nothing to blame anyone for. Over the next few years, there's going to be a new player on the field. More specifically: an AI player. The person who brings him onto the field is what's now being called an FDE — a Forward Deployed Engineer. This player can pitch, catch, run, and deflect — all at once. He knows your strike zone. He knows where you want to throw the blame. He can stand at all nine positions at the same time. Hit it clean and he reaches out and catches. Even a home run — he gets to the wall first. In the office, he's the handoff that used to take three people a week — now it takes ten minutes. Reading this, you might be thinking: isn't this exactly the work that's going to replace mine? Yes. Partly. But more precisely — he's not here to replace you. He's here to replace the game itself. This game nobody was watching, he can play it alone. The only real question left is whether you want to keep stepping up to bat, or stop playing this game. You might still be thinking: even if I stop, the coach isn't going to let me have any of that saved time off. He won't — if you're still on his team. Here's another angle. The AI player can do everything, but he doesn't know what to do. Which process is broken. Which Excel everyone hates. Which handoff is the worst — none of that lives in the documentation. None of it lives in the data. It lives in the break room complaints. In the 5:47 PM message someone fires into the group chat and deletes a minute later. In the 'don't tell the boss' that a coworker drops over a cigarette right before telling you anyway. People only say this kind of thing to other people. That's the part AI can't take. You've been on this field for years. So you know. Take that — the things only humans tell other humans — into another company, bring the AI player with you, and you're the FDE. Freelancing, building your own product, working solo with a craft — these count too. None of them put you on this field. Of course, not every one of those paths will work for everyone. That's OK — this game isn't going to wrap up in a day, and you don't have to leave in one either. Knowing what AI can do, and what it still can't, is enough. Where you go doesn't matter. Just stop playing. People who stop playing this game come home tired and can still go watch a real one. A beer, some peanuts, friends. ## 中文 我从去年开始喜欢看棒球，意识到棒球和职场其实是一回事。一样的规则、玩法，没人在看。 --- 球场上九个人对你一个。轮到你打球的时候，你一个人面对所有人。职场也是——你以为有同事，其实没有。 只有一种情况他们会真正关心你：他已经站上了某个位置，需要你别砸，需要你把他往前推。利益捆在一起的那一刻，他才在乎。剩下的时候，没人替你挡球。 --- 每一次投球，是一次甩锅。 锅飞过来，你不知道它是不是冲着你来的。没飞进你的区域，那是坏球——你不动，扔锅的人自己心虚。但你只要挥棒，这口锅就是你的。 所以最重要的是看清楚有没有飞向你的区域。新人最容易犯的错，就是慌着挥。 --- 飞向你了，你必须挥。挥不是为了把球打回去，是把锅甩出去。 但不能甩到别人手里——那叫接杀，所有人都看见是你甩的。 要嘛打出场外，全垒打，锅找都找不着。要嘛打到没人接得住的地方，自己冲上去安全上垒。 安全上垒，就是抱住了领导的大腿。锅还在场上，但你站稳了。 --- 但是站稳不等于赢。 人会一直来找你茬。三振出局，这一轮就完了。 三振出局是怎么发生的？你的同事都知道你会甩锅。他们已经在场上每一个位置都站好了——你想往哪甩，他们就在哪等着。你挥棒的方向、力度，他们早就预判过。 真正能打出场外的全垒打，很少。大多数球，都落进他们的射程里，伸手就接住。 --- 棒球的难，不在挥，在判断什么时候不挥。 但还有一个更难的判断——你今天到底是在场上，还是在看台上。 这不是工种的区别，是心态的区别。同一家公司里，有人在场上为这一棒拼命，有人在看台上嗑着瓜子把整场看完。 --- 但说实话，这两种都不对。 棒球就是棒球。职场不应该是棒球。 球场上九个人玩这场没人看的硬仗，那是他们的本职工作。职场不该是。职场该是一个生产价值的地方——你交付的东西要么跑要么不跑，要么省了人要么没省，要么客户买单要么没买单。这种地方没有锅可甩，因为根本没有锅。 未来几年球场上会多一个人。准确说，是一个 AI 球员。带他上场的人，现在流行叫 FDE（前线部署工程师）。 这个球员同时会投会接、会跑会甩。你的好球带他知道，你想把锅甩到哪他也知道。他可以同时站在场上的九个位置——你打得再准他也伸手就接，全垒打也接得下来。 具体到办公室里，他就是那个每天三个人扯一星期的对接，现在十分钟自己跑完了。 看到这儿你可能想：这不就是来取代我的工作的吗？ 对，部分是。但更准确地说——他不是来取代你。他是来取代这场球本身的。这场没人看的比赛，他一个人就能玩。剩下的问题只有一个：你想继续上场挥棒，还是不打这场球。 你可能还会想：就算我不打了，省下来的时间教练也不会让我休息。 对，他不会——如果你还在他的球队里。 换个角度想——AI 球员什么都会做，但他不知道该做哪件。哪个流程最烂、哪个 Excel 大家最恨、哪个对接最折磨人——这些事不在文档里，也不在系统数据里。 它们在茶水间的抱怨里，在群里 5 点 47 分发出来又秒删的那条吐槽里，在同事跟你抽烟时说的「你别跟领导说啊」后面那半句话里。 人只跟人说这种话。这是 AI 拿不走的。 你打了这么多年球。所以你知道。 把这些「只有人才会告诉人」的事，带到另一家公司，再带上 AI 球员——你就是 FDE。 自己接活、做产品、靠手艺单干，也都行。这些人都不在这个球场里。 当然，这些路不一定每条都走得通。也没关系——这场球不是一天散的，你也不必一天就走。看清楚 AI 能干什么、还干不了什么，已经够了。 去哪不要紧。不打就行。 不打的人，累完回家还能去看一场真的。一杯啤酒，一把瓜子，一群朋友。 --- # Claude Code for Everything / Claude Code，用于一切 > Published 2026-05-14 · By lawted · Canonical: https://ha7ch.com/writing/claude-code-for-everything ## English Not long ago I wanted to ask another developer: did you write this feature? I stopped myself. Wait — why am I asking him? When did I start doing that? I went and asked Claude Code instead. The answer came back faster, broader, and more complete than anything I would have gotten from the person who wrote it. I didn't need to wait. I didn't need to bother anyone. I didn't need to set up any context at all. --- I'm not saying asking people doesn't work. I'm saying Claude Code has a deeper understanding of the entire codebase than any single person has of their own part of it. When you ask a person, you first check if they're free. Then you set the context. Then you wait. Then you accept that they might remember it wrong. Every single stage has friction. Ask Claude Code, and all of that is gone. --- And now I use Claude Code for everything. Articles, ideas, content production, all of it. This article itself: I dictate in Chinese using WeChat voice input on Mac, drop it into Claude Code, and get a Chinese draft with an English version alongside it. The English always sounds too AI, so I don't use it directly. I take both versions, run them through Claude one more time, and that produces the final thing you're reading. --- This is also why we're building zero-token products. Zero-token doesn't mean zero AI. It means the reasoning happens inside your own workspace, not inside some chatbot widget embedded in a product. Claude Code is the workspace. Not the tool. Everything flows through it: coding, writing, thinking, executing. That's the default architecture for everything we build now. --- So my conclusion is simple: everyone needs their own Claude Code or Codex. Not a recommendation. A must have. This is the infrastructure of 2026. ## 中文 前不久，我想问我们另一个开发一个问题：这个功能是不是你写的？ 然后我突然停住了。喂，为什么我要问他呢？ 我直接去问了 Claude Code。 Claude Code 的回答比本人更快、更全、更准。不需要打扰别人，不需要等待，也不需要给他铺垫上下文。 --- 不是另一个开发不行。 而是 Claude Code 对整个 codebase 的理解，比任何一个人对他自己的那块代码的理解都要深。因为我们所有人都是 Vibe Coding 的。 当你想问一个人的时候，你需要先去问他有没有空，然后要把背景说清楚，然后等他想起来，他有可能还记错了。中间全部都是摩擦点。 而问 Claude Code，这些摩擦点会全部消失。 --- 我已经把 Claude Code 用于一切。改文章、整理思路、内容生产，都在这里。 就连这篇文章，工作流是这样的：用 Mac 微信的语音输入说中文，丢进 Claude Code，它先出一个中文初稿。然后我对着这个中文初稿，口说英文翻译，再把英文丢给 Claude Code，让它将两个版本结合起来。因为这样的话，中文初稿里那些 AI 腔的语言，我是不会翻译成英文的。这样一结合，就更容易出一个人类易阅读的版本。 --- 这就是为什么我们一直在做零 Token 产品。关于零 Token 产品，可以看我们之前的文章。 Claude Code 是一个 workspace，不是一个工具。所有的工作都以它为中心展开：编码、写作、思考、执行。我们所有产品的默认架构，都是这样的。 --- 最后我的结论很简单：每个人都应该有自己的 Claude Code 或者 Codex。 这不是一个推荐，而是一个 must have。 因为他妈的，2026 年的基础设施，就是这个。 --- # Question Every Instinct / 质疑你的每一个直觉 > Published 2026-05-13 · By lawted · Canonical: https://ha7ch.com/writing/question-every-instinct ## English Since shipping Raily, we get a ton of suggestions every day. They all sound reasonable. They all sound like the obvious next step. But I've noticed that most of them, if you just follow your gut and do them, are wrong. I mean it. All wrong. --- Example one. We have a RedNote group of about 1000 people dropping feedback all day. A bunch of people told me: you should build a RedNote agent. Have the agent read the group messages, turn them into a requirements doc, then you can build the requested features faster and you don't have to keep providing emotional value in the group. I told them: you're dead wrong. Completely wrong. Absurdly wrong. First, the reason we built this group is to learn how to do ops. We've been writing fucking code for five years. Why would we want to learn more code? What we need to learn right now is how to give people emotional value. That's our current bottleneck. Not code. Second, building a RedNote scraper agent walks straight into anti-bot hell. It's a bottomless pit. Burns time, burns brain, returns close to zero. What's the right move? You provide the emotional value yourself, in the group, and while you're at it you distill the requirements yourself. Once the requirements are clean, you hand them to Claude Code or Codex and let it write the code. That's the right way. And the technical difficulty of that second half is zero. That's literally today's workflow. Right? Following the gut, you were going to use the agent for the part that's most worth doing by hand, and spend your own time on the part the agent is best at. Completely flipped. --- Example two. We started doing writing, and people said: you should add email subscriptions. Nope. Email is a thing on the way out. If my read is right, we're heading into a zero-token era. Everyone will have their own AI agent. So what should we build? A CLI, or an MCP, so people can plug their agents in. Whenever their agent wants to read our writing, it can just ask the AI. The AI summarizes, extracts, translates into whatever language. That's how the consumption is going to happen. Or, if they want it automated, their agent crawls our site on a schedule and pulls in the new content. So the move is not to embed a chatbot widget on the page, and not to slap on an email subscription button. The move is to make our site more agent-friendly. Easier to read, easier to parse, easier to consume. Following the gut, you were trying to keep people on your webpage. But the future is that people aren't going to come to your webpage at all. --- So now, every time I hear a suggestion or feel an instinct, I stop. I ask: which era does this instinct come from? Is it a paradigm from the previous era? Is it my old comfort zone? Most of the time, the answer is yes. ## 中文 做了 Raily 以后，每天会有非常多的人给我们提建议。 听起来都很有道理，听起来都像是顺理成章的下一步。但是我发现，大部分的建议如果你顺着直觉去做，全是错的。 我说真的，全是错的。 --- 举个例子。我们小红书群里大概 1000 个人，每天嘎嘎地提需求。听了好多人跟我说：你可以做一个小红书的 agent，让 agent 去读群里面的消息，整理成需求文档，然后你就可以更方便地去做，不需要在群里面提供情绪价值了。 我说大错特错，大错特错，错到离谱。 第一，我们建这个群是因为我们要学习的是如何去运营。我们写代码他妈的已经写了五年了，我们还学什么代码？我们要学的就是如何去给别人提供情绪价值，这是我们当下的瓶颈，不是写代码。 第二，你做小红书的 agent，反爬非常严重。你掉进去就是个无底洞，烧时间、烧脑子、收益接近于零。 那正确的做法是什么？你他妈应该自己在群里一边提供情绪价值，一边把需求整理出来。把需求整完了以后，丢给 Claude Code 或者 Codex 自己去写。这才是正道。 而且这后面的技术难度是零，就是现在的工作流，对不对？ 也就是说，你顺着直觉是想用 agent 去做你最值得人工做的部分，然后自己花时间去做 agent 最容易做的部分。完全反了。 --- 再举个例子。我们这边开始做了 writing，然后有人说：你们应该给 writing 加上邮件订阅。 我说不对呀。 邮件是正在退场的东西。如果我的判断是对的，后面大家会进入零 token 时代，所有人都会有自己的 AI agent。 那这种情况下我们应该做什么？应该做一个 CLI，或者一个 MCP，让大家可以把他们的 agent 接进来。 他们的 agent 在任何时候，如果他想起来要读一下我们的文章，可以直接问 AI。AI 也可以帮他总结、帮他提炼、帮他翻译成各种各样的语言。 或者，如果他想自动化，应该是他的 agent 每天定时来爬一下我们的网站，把新的内容拉走。 所以我们要做的不是在网页里嵌一个聊天机器人，也不是加一个邮件订阅按钮。我们要做的是让我们的网站对 agent 更友好，更容易被读、更容易被理解、更容易被消费。 你顺着直觉想做的是把人留在你的网页上。但未来的事情是，人根本就不会来你的网页。 --- 所以我现在每次听到一个建议、产生一个直觉，先停下来。 问自己：这个直觉到底来自哪个时代？是来自上一个时代的范式吗？是来自我以前的舒适区吗？ 大部分的时候，答案是是的。 --- # The Frog in the Well / 井底之蛙 > Published 2026-05-13 · By lawted · Canonical: https://ha7ch.com/writing/the-frog-in-the-well ## English I failed Computer Science in high school. I'm studying accounting in college. I am, by every reasonable measure, not the person who should be writing this. But three internships in three industries taught me something most software engineers will never see. --- I didn't learn what FDE meant from a blog post. I learned it from three internships where I watched smart people do stupid things, and nobody around them knew it was stupid. Private equity firm (2023): Hundreds of interns. Powerful CRM, genuinely good leads, real money on the table. The job: copy-paste the same boilerplate outreach message to every lead. Hundreds of interns, all doing the exact same thing, by hand, every day. I built a pipeline that replaced the work of 200 interns and personalized every message at the click of a button. They didn't want to pay me for the tool. Cited my work contract. Hell nah, I left. Fuel company, $5B in revenue (2024). I was a tax accounting intern. I watched two CPAs (people who passed one of the most rigorous certifications in the country) manually copy-paste customer addresses, one by one, to look up tax rates. This was 50 percent of their job. This is what the company pays them to do. Mindless. Their specialization is needed elsewhere but their time is soaked up here. One week. I find my own shapefiles, build a Python pipeline (a few hundred lines of code), done. Whole company's fuel tax calculations, automated. Civil engineering firm (2025). Engineers eyeballing 50 years of time series data, trying to pattern-match temporal sequences visually. With their eyes. Took them 10 hours per week. I built a parsing engine with dynamic time warping that surfaced mathematically high-tier matches in seconds. --- Three internships. Three industries that have nothing to do with each other. Same story every time. I wasn't in any of these companies as a software engineer. I entered each of these companies as a business intern. A SWE intern would never have been in those rooms. The reason I saw these workflows is that I was sitting next to the people doing them. These weren't dumb companies. PE firms aren't dumb. CPAs aren't dumb. These civil engineers were incredibly smart. And yet all of these workflows were structurally insane. Nobody on the inside saw it. Not because they were stupid. Because they were the frog at the bottom of the well. 井底之蛙. You don't know what you don't know. You can't see the sky from down there. The water you swim in is the only water there is. --- This is what's sitting in every traditional company in the world right now. Not a few. Every single one. Some workflow that absorbs the brainpower of three people, or thirty, or three hundred, that one person with a laptop and the right instinct could collapse in a few weeks. The bottleneck is that the people who can see the workflow can't see the AI, and the people who can see the AI never walk into the room. YC and SF are bright and flashy right now. Founders pitching wrappers that will be murdered in the next Claude release, bragging on Twitter how they're going to scam VCs out of a seed round, building the same stupid dating app for the seventh time. The bubble is real. Many of those companies will not exist for long. Meanwhile, there's a logistics company outside Seattle doing $80M a year, running dispatch out of a spreadsheet a guy named Dave built in 2011. Dave retired in 2019. Nobody knows how half the formulas work. They hired two people last year just to babysit the file. The owner has heard of ChatGPT because his daughter showed him. He has real revenue, real margin, and a real problem that AI can solve this afternoon. That's where the value is. Not glamorous. Not on Twitter. A back office in a strip mall, fluorescent lights, a printer that jams, a whiteboard with last quarter's numbers still on it. Just a workflow that's been running on human brainpower for thirty years, waiting for one person to walk in and see it. --- The people who can do this are not normal software engineers. A traditional SWE wants a ticket, a spec, a code review, a staging environment. None of that exists here. You walk in with nothing. The workflow lives in someone's head. Half the time the boss can't even tell you what his employees do. You have to sit next to them and watch. You will not understand anything going in. You have to learn fast or you're cooked. You have to be able to talk to people who are nothing like you, who think AI is magic or a scam or both. You have to have tact when the boss is wrong, and you have to know when to push anyway. You have to ship something on Tuesday that you didn't understand existed on Monday. The things you build will have no documentation. There is no way to paste it into Claude Code and have it write itself. The schema is in someone's notebook. The business logic is in an employee's head. The edge cases live in a Teams chat from 2020. You have to dig it out, structure it, and ship. --- This is not for everyone. But if you can do it (and maybe that person is you) there is a window right now that we've never seen before. Every traditional company gets distilled exactly once. Whoever does it first owns that workflow. And there are millions of these companies. I learned this by accident. Starting from an accounting degree, three internships, and a habit of figuring shit out and building the thing when I see something too stupid to be done manually. Most of the people who'll thrive at HA7CH probably learned it the same way- from somewhere they weren't supposed to be. --- The frog at the bottom of the well doesn't know the well is a well. Let's go build some fucking ladders ## 中文 我高中计算机课挂了科。我大学读的是会计。怎么看，我都不该是写这篇东西的人。 但三段实习，三个完全不同的行业，让我看到了大多数软件工程师一辈子都不会看到的东西。 --- 我不是从哪篇博客上学会什么叫 FDE 的。我是从三段实习里学会的——在那些地方，我亲眼看着一群聪明人做着蠢事，而他们身边没有一个人意识到那有多蠢。 私募股权公司（2023年）：几百个实习生。强大的 CRM、真正优质的线索、桌面上摆着真金白银的钱。工作内容是什么？把同一段模板化的开发信复制粘贴给每一个线索。几百个实习生，每天手动重复着完全一样的动作。 我搭了一条流水线，一键替代了 200 个实习生的工作，而且每封信都做了个性化处理。他们不愿意为这个工具付我钱，搬出我的劳动合同来压我。去他妈的，我走人。 燃油公司，年营收 50 亿美元（2024年）。我是税务会计实习生。我看着两个 CPA（全美最难考的资格证之一的持证人）一个一个手动复制粘贴客户地址，去查税率。这占了他们工作量的 50%。这就是公司花钱请他们干的事。完全是脑死亡的活。他们的专业能力本该用在别的地方，时间却全耗在了这里。一个星期。我自己找到了 shapefile 数据，搭了一条 Python 流水线（几百行代码），搞定。整个公司的燃油税计算，全自动化了。 土木工程公司（2025年）。工程师们盯着 50 年的时间序列数据，试图用眼睛去识别时间模式。用他们的肉眼。每周要花 10 小时。我写了一个解析引擎，用动态时间规整（DTW）算法，几秒钟就把数学意义上的高匹配项给捞出来了。 --- 三段实习。三个毫不相干的行业。每次都是同一个故事。我进这些公司都不是以软件工程师的身份。每一次我都是以商科实习生的身份进去的。SWE 实习生根本进不了那些房间。我能看到这些工作流，是因为我就坐在干这些活的人旁边。 这些都不是傻公司。私募股权公司不傻。CPA 不傻。那些土木工程师聪明得很。但所有这些工作流在结构上都是疯狂的。公司里没有一个人看得出来。不是因为他们蠢，而是因为他们是井底之蛙。 井底之蛙。你不知道自己不知道什么。从井底，你看不到天空。你游的那点水，就是你认识的全部的水。 --- 这就是现在全世界每一家传统企业里都摆着的东西。不是少数几家。是每一家。某个吞噬着三个、三十个、甚至三百个人脑力的工作流，而一个带着笔记本电脑和正确直觉的人，几个星期就能把它压成一行命令。瓶颈在于：能看见工作流的人看不见 AI，能看见 AI 的人从来不会走进那个房间。 YC 和旧金山现在又亮又花。创始人们推销着下一个 Claude 版本一发布就会被秒杀的套壳产品，在 Twitter 上炫耀自己怎么把 VC 的种子轮骗到手，第七次做同一个傻逼约会软件。泡沫是真的。这些公司里很多撑不了多久。 与此同时，西雅图郊外有家物流公司，年营收 8000 万美元，整个调度系统跑在一个叫 Dave 的人 2011 年搭的 Excel 表上。Dave 2019 年退休了。没人知道那些公式里有一半是怎么算出来的。去年他们专门请了两个人来伺候这张表。老板听说过 ChatGPT，因为他女儿给他看过。他有真实的营收、真实的利润，以及一个 AI 今天下午就能解决的真实问题。 价值在那儿。不光鲜。不在 Twitter 上。在某个购物广场后面的办公室里，荧光灯，会卡纸的打印机，白板上还写着上个季度的数字。一个跑了三十年、全靠人脑撑着的工作流，等着一个人走进来看见它。 --- 能干这活的人，不是普通的软件工程师。传统 SWE 想要的是工单、需求文档、code review、staging 环境。这里什么都没有。你两手空空地走进去。工作流住在某个人的脑子里。一半时间老板自己都说不清他的员工到底在干什么。你必须坐在他们旁边，看着。 刚进去的时候你什么都看不懂。你必须学得很快，不然就完蛋。你必须能跟那些跟你完全不是一类人的人说上话——那些觉得 AI 是魔法、是骗局、或者两者皆是的人。老板说错的时候你得有分寸，但该顶回去的时候你还得顶回去。你必须在星期二交付一个你星期一根本不知道它存在的东西。 你写出来的东西不会有任何文档。没法直接丢给 Claude Code 让它自己写出来。数据库结构在某人的笔记本里。业务逻辑在某个员工的脑子里。边角情况躺在 2020 年的某个 Teams 聊天记录里。你必须把它们挖出来，理清楚，然后交付。 --- 这不是每个人都能干的活。 但如果你能干（也许那个人就是你），现在有一扇我们从来没见过的窗口。每一家传统公司只会被蒸馏一次。谁先把它做了，谁就拿下那个工作流。这样的公司有几百万家。 我是误打误撞学会这件事的。一个会计学位、三段实习，加上一种看见太蠢的人工活就忍不住想搞清楚、把东西造出来的习惯。能在 HA7CH 干得好的人，大概也都是这么学会的——从一个他们本不该出现的地方学会的。 --- 井底之蛙不知道那口井是一口井。 我们去造他妈的梯子吧 --- # Walk on Two Legs / 两条腿走路 > Published 2026-05-12 · By lawted · Canonical: https://ha7ch.com/writing/walk-on-two-legs ## English At dinner today a friend brought up this question: why does the US political system have an auto-repair mechanism? It's pretty simple. Every government can publicly say the previous one was wrong. That sounds normal, but think about it. It means the system can go left, can go right, can go left-right-left-right, and it just keeps moving. Working with two legs. Most people can't do this. Because we're always scared. Scared that when we publish our ideas, someone will say you're fucking stupid. Scared that something we told a friend last week gets screenshotted and thrown back at us: look how much you've changed. So we either say nothing, or we hold a position we already know is wrong. --- But you can treat every period of yourself as a different person. Next week, you can say: what I wrote last week was wrong. I wasn't clever enough then. After that I saw something, I experienced something, so my judgment changed. That's a good update. Everything you write today, you don't need to protect it with your whole life. It's just today's version. Look at the date on this article. May 12. This is a snapshot of May 12. It might change tomorrow, next week, next month. Or it might be right forever. Both are fine. --- Your mind only has before and after. There is no high and low. Because it's dynamic. Yesterday is right and today may be wrong. And even worse, you don't know you're wrong today. But that's very possible. Every day is different. Who you meet, what you read, what you talk about at dinner — it all has a huge impact on your judgment. That's pretty normal. So don't chase a straight upward mind curve. That curve doesn't exist. If you pursue it, you're just going to end up saying nothing. Walk on two legs. Left right, left right. That's how we work. ## 中文 今天和朋友吃饭，聊到一个观点：为什么美国的政权有一种自动修复机制？ 很简单。每一届政府都可以公开说，上一届是错的。这件事听起来很普通，但你仔细想，这其实非常牛逼。它意味着整个系统可以左，可以右，可以左右左右地走，但它一直在走。两条腿走路。 我们很多人做不到这件事。不是因为认知不够，而是因为怕。怕发出来以后别人觉得你蠢，怕上周说了一件事这周被打脸，怕被人截图说你前后矛盾。所以大多数人就干脆不说，少做少错，不做不错。 --- 但其实你完全可以把每一个时期的自己当成不同的人。 下周的你，完全可以站出来说：上周那些话是错的。我那时候认知不够。后来我看到了什么，经历了什么，所以我今天的判断变了。这不是打脸，这是更新。 你今天写下来的东西，不是一份需要你用余生去捍卫的声明。它只是你今天这个版本的快照。 --- 认知只有先后，没有高低。 而且它是动态的。昨天是对的，今天可能是错的。更麻烦的是，你今天意识不到自己今天是错的。这也完全有可能。 每天都会不一样。你今天遇见了什么人，读了什么，吃饭的时候聊了什么，都会影响你的判断。这不是软弱，这是正常的。 所以，不要去追求一条永远笔直向上的认知曲线。那种曲线不存在，追求它只会让你越来越不敢开口。 两条腿走路，左右左右，一直往前走。这才是真的在走路。 --- # HA7CH Is a FDE Accelerator / HA7CH 是一个 FDE 加速器 > Published 2026-05-11 · By lawted · Canonical: https://ha7ch.com/writing/ha7ch-is-a-fde-accelerator ## English Over the last couple of days a lot of people have been asking us what HA7CH actually does. The answer is simple: HA7CH is a FDE accelerator. --- YC helps founders become companies. HA7CH helps people become founders. YC takes people who already stand out and gives them resources. HA7CH finds people before they stand out, on campuses and in build in public, and puts them in front of real problems, real users, and real deliveries. YC asks: "Who is worth investing in?" HA7CH asks: "Who can be ha7ch'd?" In a world where products are easy to copy, the moat is no longer the product itself. The moat is the people you can gather, ignite, and move forward together. --- So what are we actually doing right now? On one side, we're running a RedNote group of around 1000 people who really love Raily, which, to be fair, has turned out pretty well. They drop feedback in there all day long. On the other side, we're running a WeChat builder group, made up of people who like vibe coding, with their heads on right. People who want to build things together, and who align with some of what HA7CH thinks. --- Let's start with Raily. A lot of people want us to charge. We don't want to. We'll ship it, list it on the store, and at peak popularity we'll just open source the whole thing. Raily didn't cost us much. A few sleepless nights and some dev. But what did we get? A user base. Feedback. Affection. These matter way more than the few bucks of subscription money we could have collected. What we're after is momentum, pulling more users and more attention onto our products. There's a second use too: when we want to break out and push the next product, if someone in the builder group has an idea, they don't have to cold start. We drop what they build straight into the user group and let real users react. If they like it, we keep going. If they don't, we drop it and start the next one. So Raily is really our first branding move. Lawted's branding, HA7CH's branding. Which is fucking important. This is how we attract a certain kind of builder: people with serious AI and coding skill, with belief, with ideas, with execution. What are these people good for? They're good for FDE. --- FDE is Forward Deployed Engineer. You walk straight into a traditional-industry company and rewrite their workflow with AI. Why Shenzhen? Because Shenzhen is crawling with bosses of old-school industries, and every one of them sits on a workflow built up over decades of human labor. Employees who've been there for decades have all the implicit knowledge in their heads: how to talk to clients, how to price, how the flow goes, who to call when shit breaks, what each field means, what each piece of jargon decodes to. None of it is documented. None of it lives in a system. It's all in heads, or in Excel. Their systems are mostly Excel, PDF, and Word. AI can obviously replace this. We all believe a lot of people in these companies will be laid off in two years. So why do those people get laid off? Because the boss suddenly learned to vibe code and built a system himself? I don't think so. It's because one fucking person walked into the company, came in every day with a Mac, and over two or three months distilled them, sorted them out, raised efficiency, and replaced a chunk of headcount. This is how we let this group of people earn their first real fucking money. A first bucket of gold. Not a salary. Something they made themselves. --- So how do our people get to Shenzhen to do FDE? We want to set up a Hatch House in Shenzhen. You'd be surprised how much office space is sitting around. Sponsors, borrowed rooms, or just renting somewhere for parties, any of it works. Pull these builders into one place, then connect them to companies. Every day they head out from the house to wherever they're embedded. A summer might be enough. These builders were probably students, or programmers at big tech. I believe that if they can come through and survive our 2C filter, they can do this. If you haven't been through that filter, I don't think you can be a real FDE. The FDE working environment can be brutal: the boss might pour you tea, might want you out drinking with him, no workstation, no proper workspace, the office might be full of smoke. But that's where you have to find your first bucket of gold. --- What these builders are actually doing is distilling the first workflow out of a company. And we have a bold claim: this is going to be a massive wave. Because every traditional company can be distilled exactly once. After that, all the knowledge lives inside that AI system. Even if the industry keeps evolving, all the thinking, all the business knowledge ends up inside the same system. Trying to re-distill later becomes very hard. And the bosses won't want to switch vendors. On the labor cost question, what's cheaper than a builder walking in barefoot? Once you've done the first company, you can almost always keep going in that industry. Workflows in a given industry are roughly the same. So HA7CH helps these bosses sell the same AI system to competitor #2 and #3. With those workflows in hand, the boss can turn his company into an AI company. Costs drop hard, competitors can't survive. It's the same play as Raily: we can be free because we used AI to write the code, far more powerful than the older apps, so we can just give it away. --- We're aiming for the first FDE delivery in May–June, then start pushing people through Hatch House, then horizontal replication and operations and so on. We might pull it off. We might not. Maybe in June I realize the path doesn't work. Maybe Hatch House is a false premise. All of it might be wrong. But what HA7CH is, we know. HA7CH is a FDE accelerator. --- P.S. The term 'old-school bosses' in this piece doesn't refer to any specific person and isn't pejorative. Huge thanks for the opportunities Reform and Opening Up created. ## 中文 这两天好多人来问我们 HA7CH 到底是干什么的。答案很简单：HA7CH 是一个 FDE 加速器。 --- YC 帮助创始人成为公司。HA7CH 帮助人成为创始人。 YC 挑选的是已经冒尖的人，然后给他们资源。HA7CH 找的是还没冒尖的人，在校园里找，在 build in public 里找，让他们去解决真实的问题，了解真实的用户，完成真实的交付。 YC 会问：谁值得投资？HA7CH 会问：谁可以嗨起（ha7ch）？ 在产品极易被复制的世界里，护城河不再是产品本身。护城河是你能聚集、ignite、一起向前走的人。 --- 所以我们现在在做什么？ 我们一边在运营小红书的群，群里有大概 1000 个人，非常喜欢我们 Raily 这个产品，确实做得也不错，他们每天在那里嘎嘎地提反馈。 另一方面，我们在运营一个微信的 builder 群，群里有一些喜欢 vibe coding、认知比较高的人。大家想一起做一些事情，也认同 HA7CH 的某些观点。 --- 先说 Raily。很多人说希望我们收费，但我们其实并不希望收费。我们决定很快就会上线上架，然后在它最火爆的时候直接开源。 因为 Raily 其实没有多少成本，成本就是熬一些夜、一些开发。但是我们得到了什么？得到了一个用户群，得到了反馈，得到了大家的喜爱。这些东西都很重要，比那几十块钱的订阅费重要太多。 我们希望的是造势，拉更多的用户、更多的注意力到我们的产品上。这件事还有另外一个好处：如果我们想出圈、想推下一个产品，builder 群里有人有 idea，他就不用冷启动，我们可以直接把它丢到用户群里去，让真实的用户做反馈。如果喜欢，我们继续做；如果不喜欢，就丢掉，做下一个。 所以 Raily 其实就是我们的第一次 branding，Lawted 的 branding，HA7CH 的 branding。Which is fucking important. 就这样，我们希望吸引到一些志同道合的 builder。这些 builder 会有非常高超的 AI、coding 能力，他们有信念、有想法、有执行力。 这样的人适合做什么呢？适合来做 FDE。 --- FDE 其实就是 Forward Deployed Engineer，直接驻进一家传统行业公司，用 AI 把他们的工作流重写一遍。 为什么是深圳？因为深圳一抓一大把传统行业的土老板，每个老板手里都有一套跑了几十年、用人肉堆出来的工作流。员工在这里干了几十年，脑子里全是隐性的知识：客户怎么谈、价格怎么算、流程怎么走、出问题找谁、每个字段该怎么填、这是什么东西、那是什么黑话。这些东西没文档、没系统，全部都在人的脑子里，或者在 Excel 里。他们的系统大部分都是 Excel、PDF、Word 组成的。 AI 当然可以代替这些东西。我们现在所有人都相信，两年后这些公司将会裁掉很多人。 那为什么这些人会被裁掉？是因为这些土老板突然学会了 vibe coding，自己做了一个系统吗？我并不这么觉得。 我觉得是他妈的有一个人走进这家公司，每天带着一个 Mac，两三个月，帮他们蒸馏、帮他们梳理，最后提高效率，裁掉一波人。 就这样，我们可以让这批人赚到他们自己的第一笔 real fucking money。这是第一桶金，不是工资，是他自己的创造。 --- 那我们的人怎么来到深圳做 FDE 呢？ 我们希望在深圳搞一个 Hatch House。你知道深圳的办公室多得一批，赞助、借场地都行，我们也可以直接租一个开 party。把这些 builder 全部聚集在这里，然后帮他们去链接公司。他们每天从 house 出发，去那些公司驻场做事。 我觉得一个暑假可能就够了。 这群 builder 可能之前是学生，可能是大厂里的程序员。我相信，如果他们能进入并通过我们 2C 这一关的验证和考验，他们应该能做到。 如果你没有经过这个考验，我觉得你很难成为一个真正的 FDE。因为 FDE 的工作环境可能非常恶劣：老板可能给你倒茶，也可能要你陪酒，没有工位，没有正经的工作环境，甚至办公室里大家都在抽烟。但你就要在这里找到你的第一桶金。 --- 这些 builder 真正做的事，就是把第一个工作流蒸馏出来。 我们有一个大胆的断言：这件事会是一个巨大的风口。因为每一个传统公司有且只能被蒸馏一次。再往后，所有的知识都会进入这个 AI 系统。即使这个传统行业还在迭代，所有的思考、所有的业务知识也都会进入这个 AI 系统。这样即使别人想蒸馏，也会变得非常困难。 而且土老板们也不会想切换方案。关于人力成本，还有什么比一个光脚走进来的 builder 更便宜的东西呢？ 做完第一家以后，你肯定还能继续做这个行业。因为工作流在一个行业里其实大同小异。 所以 HA7CH 会帮他们把这些 AI 系统卖给同行业的第二家、第三家。有了这些工作流，土老板可以把自己的公司变成 AI 公司，成本大大降低，让竞品无法生存。 就像我们做 Raily 一样：我们能免费，是因为我们用 AI 去写代码，比之前其他的 APP 强大太多，所以我们就可以直接做到免费。 --- 我们希望在 5-6 月跑出第一笔 FDE 交付，然后开始往 Hatch House 输送人，接着是后面的横向复制、运营等等。 这件事我们可能跑得出来，也可能跑不出来。也许 6 月我就发现走不通，也许 Hatch House 是个伪命题。可能都是错的。 但是 HA7CH 是什么，我们知道了。HA7CH 是 FDE 加速器。 --- P.S. 本文中「土老板」并不针对任何具体的人，也并非贬义。感谢改革开放创造的巨大机会。 --- # FDE Is The Future / FDE 才是未来 > Published 2026-05-11 · By lawted · Canonical: https://ha7ch.com/writing/fde-is-the-future ## English A lot of people have already noticed something happening abroad. OpenAI, Anthropic, these companies aren't just selling models or APIs anymore. They started sending people into enterprises, sitting right next to the client, asking how they work each day, how their systems run, what their industry jargon means, and then using AI to rebuild those workflows from the ground up. That's FDE: Forward Deployed Engineer. Inside China, I think there's a better name for the role: AI BP, AI Business Partner. They're not traditional consultants. They don't hand you a PPT or a proposal. They sure as hell don't hand you some shitty SaaS account. It's one person with a laptop, on-site at the company, working through the workflow line by line with the boss and the employees, and then turning it into an AI-based solution that actually runs. --- We live in a bubble most of the time. We assume everyone is already using AI. We assume everyone uses Claude Code, vibe codes, has AI write code, build spreadsheets, organize documents, run automations. Walk through traditional industries in Shenzhen for a day and you'll see it's nothing like that. A lot of bosses haven't even used DouBao. A while back I set up DouBao for a boss and he was overjoyed. To us this is a completely ordinary thing. To him it was like a door to a new world opened up, and he wanted to take us out for Moutai. The skills that are completely mundane inside the AI bubble are still rare and valuable in traditional industries. Another time I was talking to a boss and showing him a product I'd built. He said, you can use AI to write code? I said of course, obviously, everyone uses AI to write code now. But it suddenly hit me: in his entire circle, there might not be a single person who knows how to use AI to write code. He immediately grabbed my hand and said, you have to come help us cut costs and boost efficiency. --- Over the next two years, everyone knows a huge chunk of jobs will be replaced by AI. But how does that replacement actually happen? Is it the bosses who haven't even touched DouBao suddenly learning to vibe code, then organizing their entire company's workflow themselves, setting up their own agents, plugging in their own APIs, deploying themselves? No way. What actually happens is this: a young person walks into a traditional company with a MacBook Air, sits in the office, maybe sits on the floor, and one by one asks the employees. What's the first thing you do every day? Who does this Excel get sent to? What does this field mean? What does that industry term mean? Why do you copy-paste this every time? Which step of this process is the most tedious? Which is the most error-prone? Which is the hardest to handle? They write it all down and turn it into tools using AI. Maybe it's an internal system, maybe a Chrome extension, maybe a small utility, maybe an agent, maybe an I-don't-know-what-the-fuck-is-this. But here's the result. AI stops being a cool thing. The boss realizes that what three people used to do, one person can do now. What used to take a day takes ten minutes. What used to require a senior employee mentoring you to ramp up, a new hire can just do with AI. Stuff that used to need someone constantly nagging, AI now proactively messages on its own. That's it. So that's what FDE actually is. --- I think this is a huge opportunity for ordinary people. Because what traditional industries are short on isn't a stronger foundation model. A lot of people building startups are fighting against the model itself. Every week they worry: is Claude Code going to ship some update next week that kills my whole product? But think about it. When you're doing FDE, the stronger the model gets, the stronger you get. The cheaper the model gets, the more powerful you become. --- FDE asks a lot of you. You have to be willing to meet these bosses. You have to be willing to go on-site. You have to understand what they're saying. You can't mind the dirt, the mess, or the annoyance. Every day starts with Excel files, WeChat screenshots, emails, handwritten receipts, and you turn all of it into an AI workflow. And you have to be fast. You just had tea with the boss this morning, you need to ship a prototype this afternoon. You just understood something today, this weekend you go heads-down and turn it into system logic. You don't have to come from the industry. You can know absolutely nothing about it. But you have to be proactive enough to dive in and turn AI into productivity. That's FDE. That's AI BP. --- Over the next two years, the last mile of AI deployment is going to depend on this kind of person. Carry your laptop into the room. Let companies actually spend less, make fewer mistakes, hire fewer people. That's it. ## 中文 很多人已经看到，在国外，OpenAI、Anthropic 这些公司已经不再是单纯卖模型、卖 API。他们开始派一批人到企业里面，直接坐在客户旁边，问他每天怎么工作，问他们系统怎么跑，问他们的行业黑话是什么，然后用 AI 把这些流程重新搭一遍、重构一遍，把 AI 完全地植入进去。这就是 FDE，Forward Deployed Engineer。 如果放到中国，我觉得它有一个更好的名字：AI BP，AI Business Partner。 他不是那种传统的咨询，他不给你 PPT，不给你方案。他更不是把一些什么傻逼的 SaaS 账号给你，而就是一个人带着电脑，直接到企业现场，帮老板和员工一条一条地把工作流拆出来，然后用 AI 做成能够真正跑起来的 solution。 --- 其实我们经常活在一个错觉里面。我们以为所有人都已经在用 AI 了，以为所有人都会用 Claude Code、都会 vibe coding、都会让 AI 写代码、做表格、整理文档、跑自动化。 但是你只要去深圳的传统行业转一圈，你就会发现完全不是这样。 很多老板连豆包都没有用过。我前段时间给一个老板装了一个豆包，他都高兴得不行。对我们来说这可能是一个再普通不过的东西，但对他来说，直接就像打开了新世界的大门，高兴得要请我们喝茅台。所以在 AI 圈里非常稀疏平常的东西，在传统行业里依然是非常稀缺的技能。 还有一次我跟一个老板谈话，给他介绍我写的产品。他说，你会用 AI 写代码？我说当然啦，废话，现在所有人都在用 AI 写代码。但是突然我也意识到一件事：在他身边，也许连一个会用 AI 写代码的人都没有。于是他立马就握着我的手说，你一定要过来帮我们降本增效。 --- 未来两年，所有人都知道有大量工作会被 AI 替代。但是这些岗位到底是怎么被替代的？ 是那些连豆包都不会用的老板突然开始学习 vibe coding，然后自己把公司的工作流全部整理出来，自己搭 Agent、自己接 API、自己部署吗？不可能。 真正发生的事情一定是这样：一个年轻人背着 MacBook Air 走进一家传统公司，他坐在办公室里，可能就坐在地上，一个一个地问员工：你每天第一件事干什么？Excel 发给谁？这个字段什么意思？行业术语代表了什么？为什么你每次都要复制粘贴？这些流程里哪一个最繁杂、哪个最容易出错、哪一个最难搞？ 然后他把这些东西记下来，用 AI 做成工具。可能是个内部系统，可能是一个 Chrome 插件，可能是一个小工具，可能是一个 Agent，可能是一个 I don't know what the fuck is this。 但是结果是什么？结果就是 AI 不再是一个很酷的东西。老板会发现原来三个人做的事情，现在一个人就可以做。以前一天做不完的，现在 10 分钟就可以做完。以前需要老员工手把手带着才能上手做项目的，现在新员工直接用 AI 开始做。原来需要人一直去盯着、去催的东西，现在 AI 都可以主动去发消息。That's it。 所以这就是 FDE 的本质。 --- 我觉得对于普通人来说，这里有非常大的机会。因为传统行业缺的并不是一个更强的大模型。 很多人创业都在跟模型对着干，每天都要担心：下一周 Claude Code 是不是又更新一个东西把我干掉了？但是你仔细想，我们做 FDE，就是模型越强我越强，模型越便宜我越牛逼。 --- 当然 FDE 的要求也很高。你要愿意认识这些老板，你要愿意走到现场，你要听得懂他们的话，你要不嫌脏、不嫌乱、不嫌烦，每天从 Excel、微信截图、邮件、手写的单据开始，然后把这些东西全部变成 AI workflow。 而且你要做得非常的快。你上午刚刚跟老板喝完茶，下午就必须要跑出一个 prototype。你今天刚刚理解了一个东西，周末就要开始狂干，然后把它写进系统的逻辑里。 你可以不是这个行业出身，甚至你可以对这个行业一无所知。但是你必须要足够主动，能够钻进去，把 AI 变成生产力。这就是 FDE，这就是 AI BP。 --- 未来两年，AI 落地的最后一公里，靠的就是这种人。背着电脑走进现场，让企业真的少花钱、少犯错、少用人。That's it。 --- # Code Agent and Token Cost / Code Agent 和 Token Cost > Published 2026-05-11 · By lawted · Canonical: https://ha7ch.com/writing/code-agent-and-token-efficiency ## English This one runs a bit long — about ten minutes. If you're a heavy VibeCoding user or curious about how LLM billing actually works under the hood, it's worth finishing. VibeCoding is becoming infrastructure for a lot of engineers. When you hit a rate limit or context ceiling in the middle of a long task, the quickest fix is obvious: throw $200 at a GPT Pro subscription, throw another $200 at Claude Max — problem gone. But if we don't just buy our way around this, and instead ask the real question — where the hell do all the tokens go — aren't you curious? I sure am. --- Let's start with billing. Most mainstream LLM APIs, including Claude and OpenAI, charge by token count, with separate rates for input and output. Input is cheaper; output costs more. In the specific context of Code Agents, input tokens are the overwhelming majority — often over 80% of total consumption. Some providers offer a KV Cache Discount: when the server detects that the input prefix of a new request heavily overlaps with a previous one, the overlapping portion hits the cache and gets a significant discount. The mechanism makes physical sense — it avoids redundant attention computation. This will become a plot point. We'll come back to it. --- Most Code Agents today, including Claude Code, still run on the classic ReAct framework. The original ReAct paper is over three years old. Its core loop: the model produces a Tool Call, receives an Observation, reasons through a CoT, then produces the next Tool Call. It was an elegant design when it was proposed. But over the past three years, the agent research community has produced a lot of optimization paradigms — Plan Before Act, Hierarchical Planning, Task Decomposition with Memory… Code Agents have adopted basically none of them. The arrogance isn't entirely unjustified — engineering stability will always outrank academic novelty at the product level. But that doesn't mean we can't look at the cost. The cost lands on token consumption, and ultimately on the user. The worst offender is context management. The only way to describe it is: A Piece of SHIT. Each loop iteration, the Agent appends the full user Query, the current Tool Call, the Observation, and the model's own CoT into the Context — then feeds the entire thing back unchanged on the next round. Under this model, Context grows quadratically, and most of it becomes historical noise that's nearly useless for whatever the model is actually trying to do right now. When Context approaches the model's limit, Claude Code triggers Auto Compact — an interesting mechanism in itself. It's not a semantic summarization pass; it's a rule-based structural pruning at the linguistic level. Codex takes a blunter approach and just terminates the session. Claude's context window is around 200K tokens. That sounds large until you've done a few dozen tool calls on any reasonably-sized codebase — then it's gone fast. --- Solutions to this problem fall into two camps: Harness Level and Model Level. Harness Level means engineering the Agent's runtime framework without touching the model itself. Two main approaches: First, fine-grained Context management — actively filtering and compressing historical information, keeping only what's genuinely relevant to the current task. Compressing Observations is the primary lever. Second, introducing a Plan Before Code paradigm — having the Agent complete an explicit planning pass before execution, reducing aimless exploratory Tool Calls and cutting token consumption by reducing loop iterations. Papers along these lines have been coming out for about a year. The arrogant CC has shown zero interest. From an engineering standpoint, adding structural complexity always introduces potential side effects. The standard academic benchmark for validating these methods is SWE-Bench. Hit good numbers there and you can publish. But SWE-Bench is fundamentally a closed evaluation with deterministic answers. Most real Code Agent usage is open-ended exploration — unfamiliar codebases, undefined requirements — far beyond what SWE-Bench covers. Academic proof doesn't translate cleanly to actual user experience. Model Level is an entirely different angle: keep the Agent code unchanged, but use a better model. If it can solve your problem in 3 loops instead of 15, token consumption drops on its own. The most striking news on this front came from DeepSeek. DeepSeek V4's API pricing is near-disruptive — after two rounds of discounts, it lands at roughly one-tenth of the baseline price. At that level, almost no token compression technique can match the savings, because you can't algorithmically optimize your way to a 1000% efficiency gain. What's even more counterintuitive: because of KV Cache Discounts, some optimization approaches that reduce raw token count actually end up costing more — because restructuring the input breaks the cache hit pattern. Counterintuitive, but completely logical once you understand the billing mechanics. --- Worth noting: some Claude Code developers are pretty dismissive of Harness Level approaches. They don't want to introduce complex context management at the harness layer. Their position is that whatever CC can't solve today, the next model will handle. Maybe that's principled engineering conservatism. Maybe it's passing the buck. There's another take: on a long enough timeline, obsessing over token counts is just a phase. A mentor of mine compared tokens to mobile data — we might be in the 3G era right now. When 5G arrives, nobody cares how much data a single request burns. That analogy has some weight. Compute costs will keep falling. Context windows will keep growing — a 1M context window isn't unthinkable. What feels like a bottleneck today might genuinely be a historical footnote in the transition period. --- My own take: I'm open on this field, but I lean Model Level right now. Partly hindsight — Harness Level approaches have been around for a year and none of them have made it into production at scale, which tells you something. Partly distribution — if Model Level solves the problem, it'll spread like DeepSeek did. People running low on tokens will find the better API on their own. And from a vendor's perspective: token cost is temporary. Time and compute are forever. If you have a take or a solution, reach out — happy to talk. ## 中文 这篇文章略长，大概需要十分钟。如果你是 VibeCoding 的重度用户，或者对 LLM 的计费机制感兴趣，值得读完。 VibeCoding 正在成为很多工程师日常开发的基础设施。当你在某次长任务中撞上了 rate limit 或者 context 上限的时候，最直接的解决办法当然是给 GPT 充一个 200 刀的 Pro、给 Claude 充一个 200 刀 Max——问题立刻消失。 但如果我们不从财力上绕开这个问题，而是从原理上真正问一句「token 到底去哪了」，你难道不好奇吗？反正我是很好奇。 --- 先说计费机制。目前主流的 LLM API，包括 Claude 和 OpenAI，都按 token 数量计费，区分输入和输出。输入会便宜一点，输出会贵一些。在 Code Agent 这个具体的情景之下，Input 的 token 消耗占绝对大头，甚至达到 80% 以上。 部分厂家的 API 会提供一种 KV Cache Discount：当 Server 端检测到本次请求的输入前缀与历史请求高度重叠时，重叠部分的计算可以命中缓存，因此给出相当幅度的折扣。这个机制设计得很合理，背后的物理含义是避免了重复的注意力计算。 这个机制会成为伏笔，我们接下来会讲到。 --- 当前绝大多数 Code Agent，包括 Claude Code，依旧运行在经典的 ReAct 框架上。ReAct 最早的论文距今已有三年多，其核心循环是：模型产生一个 Tool Call，收到 Observation，结合 CoT 思考下一步，再产生下一个 Tool Call。 这个框架在它提出的年代是相当优雅的设计。但三年以来，Agent 领域涌现出了大量优化范式——Plan Before Act、Hierarchical Planning、Task Decomposition with Memory……Code Agent 几乎一个都没有采用。 傲慢并非没有理由，毕竟工程稳定性的优先级在产品层面永远高于学术新颖性，但这不妨碍我们审视其代价。 代价落在 token 消耗上，最终落在消费者身上。Token 消耗的重灾区是上下文管理，这里只能用「A Piece of SHIT」来形容。 每一轮循环，Agent 会把当前的用户 Query、本轮的 Tool Call、Observation、以及模型自身的 CoT 全量追加进 Context，然后在下一轮把整个 Context 原封不动地喂回给模型。这种模式下，Context 以二次方速度膨胀，且大量内容是对模型当前任务几乎不再有用的历史噪声。 当 Context 逼近模型的上限时，Claude Code 会触发 Auto Compact——这个机制本身也颇有意思，它并非调用一次模型做语义层面的摘要压缩，而是从语言学结构角度做规则性的删减；Codex 则更为简单粗暴，直接终止本次对话。 Claude 的 Context 窗口大约在 200K token 量级，这个数字听起来很大，但在一个稍有规模的代码库上执行几十轮工具调用之后，很快就会见底。 --- 目前针对这个问题的解决思路大致分两派：Harness Level 和 Model Level。 Harness Level 的核心是在不修改模型本身的前提下，对 Agent 的运行框架做工程改造。核心思路有两条： 一是精细化管理 Context，主动过滤和压缩历史信息，只保留对当前任务真正有价值的内容，以压缩 Observation 的思路为主力军； 二是引入 Plan Before Code 的范式，让 Agent 在实际执行之前先完成一次显式的任务规划，从而减少无效的探索性 Tool Call，从减少循环次数来减少 token。 这类论文已经陆续发表了约一年，傲慢的 CC 依旧没有任何反应。因为从工程实践角度来说，让结构变复杂必然带来潜在的 Side Effect。 学界验证这类方法的标准工具是 SWE-Bench，跑下来指标漂亮的话足以发表，但 SWE-Bench 本质上是一个有确定性答案的封闭评测。大多数人使用 Code Agent 的场景是开放性探索——面对一个陌生的代码库、一个未定义的需求——这类场景的复杂度远超 SWE-Bench 所能覆盖的边界，学界的证明因此很难直接转化为对用户实际体验的保证。 Model Level 则是一个视角完全不同的方向：构建 Agent 的代码保持不变，调用的模型更牛逼了，3 轮循环就能把你问题解决了，token 消耗自然下来了。 这个方向最牛逼的新闻来源于 DeepSeek。DeepSeek V4 的 API 定价策略是近乎颠覆性的——两轮打折后折扣力度达到了基准价格的一折。在这个价格体系下，几乎没有任何一种 token 压缩方法能够产生与之匹配的效益，因为你无法通过算法优化做到 1000% 的效率提升。 更吊诡的是，由于 KV Cache Discount 的存在，某些优化方案在减少了 token 消耗的同时，却因为改变了输入结构导致缓存命中率下降，实际扣费不减反增。这是一个反直觉的结果，但从计费机制的逻辑来看完全合理。 --- 值得一提的是，Claude Code 的部分开发者对 Harness Level 的改造方案持相对消极的态度，不太倾向于在 Harness Level 引入复杂的上下文管理逻辑。他们的观点认为，现在 CC 解决不了的问题，等新模型出来之后就能解决了。 这或许是出于工程保守主义的考量，但也可能是在为自己的工作甩锅。 还有一种观点认为，从更长的时间轴来看，现在对 token 斤斤计较这件事本身就是阶段性的。我的一位导师曾把 token 比作流量：我们现在可能处于 3G 时代，当 5G 到来的时候，无人在意一次请求消耗了多少流量。 这个比喻有它的说服力。计算成本的下降是可以预期的，模型的 Context Window 也在持续提升（比如可能会实现的 1M 上下文），今天被视为瓶颈的问题，未来或许真的只是一个过渡期的历史注脚。 --- 本人对这个 Field 持开放态度，目前来说站 Model Level。一个是马后炮唯结果论的原因，Harness Level 的方法现在工业界一个都没有用上，那总有它的原因在。还有一个是推广度上的看法，如果从 Model Level 解决了问题那么就会像这次 DeepSeek 一样，无需过多推广，缺 token 的人自然会使用你的 API。 更何况从厂商的角度来说，token 的减耗是暂时的，时间和算力的减耗是永远的。 任何观点和方法都可以联系我们进行探讨。 --- # Attention is All You Need > Published 2026-05-11 · By lawted · Canonical: https://ha7ch.com/writing/attention-is-all-you-need ## English I think the era of programming we-media has arrived. In the past, when we wanted to do a project, it would take a lot of time. Design, develop, test, publish — the whole process needed a few months, or even a fucking year. But right now, you can do it in a real quick vibe-coding session, and ship an MVP within a couple hours. It's like writing a book. In the past, when you wanted to publish an idea, you had to choose the topic, do the editing, proofread, print, distribute. It took a long time. But right now you can publish in a real quick. Just post your idea on Twitter. A couple seconds. And at this moment, the most important thing changes. Just like I said — we are entering the we-media era. Attention is all you need. Imagine you make a great movie. You go through all the processes, you spend a lot of time, and finally people can see it in the cinema. But there is fucking no one. Meanwhile some random livestream is on, and a bunch of people are watching, dropping comments, even sending gifts. But a lot of people, when they're starting a business, are still thinking: I don't need to find users first. Promotion is not that important. I just need to build the stuff, build the workflow, and use that to find a VC, get the money. Then I'll use that pile of money to do promotion afterwards. Honestly, I was trapped in this exact idea last year. But I figured out it's fucking wrong. Because this era has already passed. You could do this two years ago. You can't do it now. Why? Because two years ago everybody was caring about the concept. AI was pretty new. So you'd say, "Okay, we'll do an AI browser. We'll do AI plus fucking medicine, AI plus fucking health." AI plus this, AI plus that — you'd quickly raise a seed round. Then the seed-stage guy would tell you how to wrap this fucking idea to help you raise a Series A, so he could exit and make money. That was the old logic. Right now this logic has been turned upside down. Investors right now are very grounded. If you don't have ARR, no traction, they won't give you any fucking money. What you need to do now — if you're going toC, it has to be something genuinely useful. Otherwise, you might as well come back to toB. By the way, I really like the FDE role. In China it's called AIBP, in the US it's called FDE. I'm doing FDE in Shenzhen right now. I'll talk about this with everyone in the next post. Now back to normal people. I've talked a lot in previous essays about how you should just build an MVP and throw it on social media to test if anyone is interested. If you get a user, keep going. Don't build a fucking heavy code shit on day one. For normal people — code is pretty cheap right now. You can just vibe-code something. Don't go build some workflow or some complicated system on day one. First, it might not be what people want. Second, it's gonna make you start really fucking slow. You'll spend a lot of time thinking about how the workflow should be built. You might even be afraid of building something that big. For normal people, the most important thing is how fast you can start. Can you ship it this afternoon? Can you ship it before you go to sleep tonight? That's what determines how fast your product can land in front of users, and how fast it can catch their attention. ## 中文 我觉得编程自媒体时代已经到来了。 原来我们做一个项目，需要花很长时间。设计、开发、测试、再发布，整个流程跑下来动辄几个月，甚至几年。但是现在，你可以做得非常快。Vibe coding 一下，几个小时就是一个 MVP。 这有点像写书。原来你写一本书，需要发版、需要选题、约稿、编辑、校对、印刷、发行，一两年是一个常态。但是现在你可以发布得非常的快，公众号、博客、推特，写完就发，几秒钟就完事了。 在这种时候，最重要的事情就变了。它就像自媒体一样，最重要的是要找到你的用户。 Attention is all you need. 就像你拍了一个很好的电影，走过了所有流程，花了很长时间，终于他妈的大家可以在电影院里面看到了。结果你发现，没有人看你的电影。反而小杨哥的直播间非常的土，但却有一群人在看。 但是很多人创业的时候还是会想：我先不需要找用户，推广不重要。我先把东西搭出来，先把工作流搭出来，然后再通过这套东西去找投资人，找完投资人拿到钱，拿了钱以后再去做推广。 我去年其实也是这个想法。后来发现，这个想法大错特错。因为这个时代已经过去了。 两年前你可以这么做，但是现在你不能。 为什么？因为前两年大家都在炒概念，AI 很新。你可以先说，「哎，我要做一个 AI browser，我要做一个 AI 加他妈的医疗，AI 加什么健康」，AI 加什么、AI 加什么，你都能很快融到种子轮。然后种子轮的人会告诉你如何继续包装这个概念，帮他融到 A 轮，让种子轮的人能从中赚钱退出。 这是原来的逻辑。但是现在，逻辑已经发生了翻天覆地的变化。 现在的投资人已经非常务实了。你没有 ARR、没有 traction，他基本上不会给你一分钱。 现在大家需要做的，如果是 toC，那就要做一些真正有用的东西。不然的话，不如回到 toB 上面来。 By the way，我现在一直很看好 FDE 这个角色。在中国叫 AIBP，在美国叫 FDE。我自己在深圳就在做 FDE。下一篇文章有机会再和大家讲一讲这个内容。 那回到普通人。我前面几篇文章其实已经说过了，你应该先把东西做一个 MVP，然后丢到社交媒体上，看有没有人感兴趣。有用户用了，你再继续做。而不是一上来就背一堆代码债。 对于普通人而言，现在 code 这么便宜，你自己 vibe coding 写一些东西就好，不要一上来就搭什么工作流、什么复杂的系统。 第一，它不一定是大家想要的。 第二，这样会让你启动得非常慢。你会花一堆时间去想这个工作流应该怎么搭，甚至你会害怕做一个那么大的东西。 对一个普通人而言，最重要的就是你能多快启动。今天下午能不能 ship 出去？睡觉之前能不能 ship 出去？这件事情决定了你的产品能多快到用户面前，多快抓住他们的注意力。 --- # Two Pairs of Eyes / 两双眼睛 > Published 2026-05-10 · By lawted · Canonical: https://ha7ch.com/writing/poetry-and-the-plaza ## English I've been thinking a lot about San Francisco and Silicon Valley lately. More specifically, I've been thinking about why these two places, only forty minutes apart by car, feel like two different countries. Silicon Valley is not a city. It's a suburb. Plaza after plaza, neighborhood after neighborhood, corporate park after corporate park. Every town has a so-called downtown, but compared to SF, those downtowns are basically just one block. But geography is just the surface. The real difference is the gaze. When you live in Silicon Valley, you can always feel someone watching you. But honestly, those people aren't actually watching you. The truth of American life is that you spend more time alone than you think. The gaze you can feel — almost all the time — is yourself watching yourself. It's just that the eyes you're watching yourself with were installed by the place you live. Once they're installed, you can't take them off. Even in an empty room, on an empty highway, at 2am with nobody around, those eyes are still open. Silicon Valley installs a pair of eyes with very clear markings on them. They keep asking you: did you get the promotion? What's your company's valuation? Which school district did your kid get into? These questions all have clear answers. You can answer them at any moment, and at any moment you know exactly where you stand. These markings are double-edged. On one hand, they slowly grind you into the shape they measure. You end up becoming a person mostly defined by these numbers. On the other hand, a person without any markings also has a hard time pushing anything to its limit. I know this take isn't popular, but I think it's correct. Caring about whether those eyes are scoring you and caring about whether you've actually done a good job are the same internal structure. A person who genuinely doesn't care what anyone thinks usually doesn't write code with tests, doesn't ship a product polished enough for a thousand strangers to use, doesn't sustain a thirty-year career arc. Patience, rigor, finishing — these are qualities that come more naturally to people who have an audience. Even if the audience is only imagined. Silicon Valley is Silicon Valley because those eyes have, over decades, produced the world's best engineers, the most ruthless product standards, the longest patience. They kill creativity. But they also feed excellence. It's the same thing turning into different things at different stages of your life. Invention needs those eyes closed. Building needs them open. I think this is similar to what happens in Asia. Asia installs the same kind of eyes — invisible, internalized, always there. They eat freedom, but they also build the kind of order that lets generations live stable lives. It's not a question of good or bad. It's a question of whether you need them right now. San Francisco has its own gaze too. I don't want to write SF as a city without a gaze, because it isn't. But the eyes SF installs are different. The Silicon Valley eyes ask you: do they see you? The SF eyes ask you: do you see yourself? The first question has an answer. The second one doesn't. The answer to the first is on your paycheck, your LinkedIn, your kid's acceptance letter. The answer to the second is something only you can faintly feel at 3am, alone, and very often you can't feel it at all. So SF isn't actually a lighter city. It just changes the direction of the work. It swaps the work of being seen for the work of seeing yourself. The first has KPIs. The second doesn't. The first you can perform. The second you can't. When you do the first one right, people clap. When you do the second one right, nobody knows. A lot of people think moving to SF lets them escape the gaze. It doesn't. They just move from one gaze to another. And many of them quietly bring the SV ruler with them — they live an SF life but score themselves with an SV scorecard, and end up paying both costs. That said, the SF gaze is genuinely thinner, because from the very beginning this city was a refuge for weird people. Gold rushers, missionaries, drifters, poets, coders, people who didn't want to get married, people who didn't want kids, people who didn't want to wear normal clothes, people who didn't want to do normal jobs. They all found their corner here. The architecture says the same thing. It's not telling you the city is beautiful. It's telling you: you don't need to be normal to live well. Silicon Valley's architecture speaks a different language. Low, flat, clean, white, beige, lawns always trimmed, Teslas always parked. These buildings don't tell you anything. They just wait for you to fit them. That kind of fitting isn't romantic, but it has an underrated kind of good. A house that doesn't demand anything of you is, for someone who has been demanded of all day, its own kind of rest. It doesn't ask you to be interesting. It doesn't ask you to be edgy. It doesn't ask you to prove anything. You can be tired. You can be bored. You can stare into space. You can spend three hours being nobody at all. The beige and the lawn don't ask you whether you spent the day living meaningfully enough. AI brought the young people back to SoMa, to Hayes Valley, to the Mission. Twenty-somethings, not married, no kids, no Tesla, renting a studio, or sleeping on a friend's air mattress, writing code in a cafe until 2am. They're building things that might change the world, or might pivot in two weeks. Silicon Valley is mostly houses. SF is mostly apartments, condos, townhouses. I think there's something interesting hiding behind that. When you live in a house, your connections point inward. You have a family, a backyard, a dinner time. The relationships that matter most to you are already inside these walls. You don't need to go looking. And your thoughts stop changing that easily. When you live in a studio, your connections point outward. The apartment is too small to want to stay inside, so you go out — to cafes, to coworking, to random meetups, to meet people you've never met. You don't know where your next idea will come from, but you know it won't come from inside your 120 square feet. Both are connection, just in different directions. One tends to what's already there. The other keeps reaching outward. That's probably also why people in SF change their minds more easily. And eventually I figured it out: these two places really represent two states a person can be in at different stages of life. Silicon Valley is for people who already know where they're going. Their lives are meant to be optimized. SF is for people who haven't figured out where they're going. Their lives are not meant to be optimized. They're meant to be invented. That's the difference between poetry and the grind. But the grind isn't bad. The grind is taking the ruler somebody else wrote and being as good as possible on that ruler. Poetry is throwing that ruler away and trying to find one nobody has made yet. Most people will probably need both at some point. First use someone else's ruler to lay a foundation, then throw it away and look for your own. Or the other way around — wander around with your own ruler for a while, then accept the more mainstream one and grind it into your own. The question isn't which one is higher. The question is which stage you're in right now, and whether you're being honest with yourself about it. Honestly, every time I drive north out of Silicon Valley, past Daly City, and the SF skyline emerges from the fog, I feel myself exhale. ## 中文 我最近一直在想旧金山和硅谷的事。 更准确地说，我在想为什么这两个地方开车只要四十分钟，但它们给人的感觉差得像两个国家。 硅谷其实根本不是一座城市。它是一片郊区。一个又一个 plaza，一个又一个住宅区，一个又一个公司园区。每个城市都有一个所谓的 downtown，但是放到旧金山来看，那个 downtown 也就只能算一个街区。 但地理只是表面。真正不一样的是凝视。 你住在硅谷的时候，你总是能感到有人在看你。 但说实话，那些人其实并没有真的在看你。美国生活的真相是，你独处的时间比你以为的多。你能感觉到的那种凝视，绝大多数时候是你自己在凝视你自己。 只是这双眼睛是你住的这个地方装上去的。它装好以后，就再也卸不下来了。哪怕在没有人的房间里、没有人的高速上、没有人的凌晨两点，那双眼睛也还是开着的。 硅谷给你装的，是一双有非常清晰刻度的眼睛。它会一直问你：你 title 升了吗？你公司估值多少？你小孩进了哪个学区？这些问题都有清楚的答案。你随时可以回答，也随时知道自己在哪一档。 这种刻度是双刃的。一方面，它会慢慢把一个人磨成它所测量的形状。你最后变成了一个主要由这些指标定义的人。 但另一方面，一个永远没有刻度的人，也很难把一件事做到极致。我知道这个观点不讨喜，但我越来越觉得它是对的。 在意自己有没有被这双眼睛打分，和在意自己有没有把事情做好，其实是同一种内在结构。一个完全不在意外界目光的人，往往也写不出有 test 的代码、做不出能给一千个陌生人用的产品、坚持不了一个三十年的事业曲线。耐心、严谨、收尾——这些都是有「观众」的人才容易具备的品质。哪怕那个观众只是想象中的。 硅谷之所以是硅谷，正是因为这双眼睛在过去几十年里逼出了世界上最厉害的一批工程师、最严格的产品标准、最长期的耐心。它杀创造力，但它也喂养卓越。这是同一个东西在你不同阶段会变成不同的东西。发明需要这双眼睛闭上，建造需要这双眼睛睁开。 我觉得这其实和亚洲很像。亚洲装上的也是这样一双眼睛，无形的、内化的、永远在那里。它消耗自由，但它也建造了让一代一代人能稳定生活的秩序。这不是一个好坏的问题，是一个你现在需不需要它的问题。 旧金山也有自己的凝视。我不想把旧金山写成一座没有凝视的城市，因为它不是。 但旧金山给你装的眼睛不一样。 硅谷的眼睛问的是：他们看见你了吗？ 旧金山的眼睛问的是：你看见你自己了吗？ 第一个问题有答案。第二个问题没有。 第一个问题的答案在你的工资条上、你的 LinkedIn 上、你小孩的录取通知书上。第二个问题的答案，只有你一个人在凌晨三点的时候能隐约感觉到，而且经常感觉不到。 所以旧金山其实并不是一座更轻松的城市，它只是把工作的方向换了。它把「被世界看见」的工作，换成了「看见自己」的工作。前者有 KPI，后者没有。前者可以表演，后者表演不了。前者你做对了别人会鼓掌，后者你做对了没人知道。 很多人以为搬来旧金山就能从凝视里逃出来。其实不能。他们只是从一种凝视搬到了另一种凝视。而且很多人最后会偷偷把硅谷那把尺子带过来——他们用旧金山的方式过日子，但用硅谷的标准给自己打分，结果是两边的代价都付了。 但即便如此，旧金山的凝视确实更稀薄一些，因为这座城市从一开始就是各种奇怪的人的避难所。淘金的、传教的、流浪的、写诗的、写代码的、不想结婚的、不想要小孩的、不想穿正常衣服的、不想做正常工作的人，都在这里找到了他们的角落。 旧金山的建筑也是这样。它不是在告诉你这座城市多漂亮，它是在告诉你：你不需要正常才能过得好。 而硅谷的建筑说的是另一种语言。低矮的、平的、整洁的、白的、米色的、永远在修剪的草坪，永远停着的 Tesla。这些建筑不会告诉你任何事情，它们只是在等你去配合它们。这种配合不浪漫，但它有一种被低估的好。一栋不要求你的房子，对一个白天一直被要求的人来说，本身就是一种喘息。它不要求你 interesting，不要求你 edgy，不要求你证明什么。你可以疲惫，可以无聊，可以发呆，可以一连好几个小时什么都不是。米色和草坪不会问你今天有没有更有意义地活着。 AI 又把年轻人带回了 SoMa，带回了 Hayes Valley，带回了 Mission。他们二十几岁，没结婚，没小孩，不开 Tesla，租一个 studio，或者睡在朋友的 air mattress 上，每天在 cafe 里写代码到凌晨。他们在做一些可能会改变世界，也可能两个礼拜就 pivot 掉的产品。 硅谷大部分是 house，旧金山大部分是 apartment、condo、townhouse。我觉得这背后藏着一件挺有意思的事。 住在 house 里，你的连接是向内的。你有一个家庭，一个院子，一个晚饭时间。你最重要的关系都已经在这堵墙里面了，你不太需要往外找。你的想法也就不那么容易被改变。 住在 studio 里，你的连接是向外的。你的房子小到你不愿意一直待在里面，所以你会出门，去 cafe，去 coworking，去随便一个 meetup，去认识完全不认识的人。你不知道你下一个想法会从谁那里来，但你知道它一定不会从你这一百二十平方英尺的墙里冒出来。 两种都是连接，但方向不一样。一种是在养护已经有的关系，一种是在不断地往外伸触手。这大概也是为什么旧金山的人想法更容易变。 我后来想明白了，这两个地方其实代表了一个人在不同阶段的两种状态。 硅谷是已经知道自己要去哪儿的人住的。他们的人生是要被持续优化的。 旧金山是还没想好自己要去哪儿的人住的。他们的人生不是要被优化的，是要被发明的。 诗和苟且其实就这一点区别。但苟且不是不好。苟且是你拿着别人写好的尺子，然后在那把尺子上把日子过到最好。诗是你扔掉那把尺子，自己去找一把还没人造出来的。 一个人这一辈子，大概率两种都需要。先用别人的尺子把基础打牢，再扔掉它去找自己的。或者反过来，先用自己的尺子瞎走一段，最后接受那把更主流的尺子，把它打磨成自己的。问题不是哪个更高级，问题是你现在在哪个阶段，以及你有没有诚实地知道自己在哪个阶段。 说到底，每次开车从硅谷北上，过了 Daly City，旧金山的天际线在雾里浮出来的那一瞬间，我都会松一口气。 --- # MVP as Research > Published 2026-05-09 · By lawted · Canonical: https://ha7ch.com/writing/mvp-as-research ## English In the AI era, how should we ship a product? Traditionally, when we want to build something, we start with user research. Then we write the requirements, draft the idea, do the design, develop it, test it, and finally release it. That was how big tech worked. And it made sense at that time, because the cost of development was high. If a product direction was wrong, the cost was serious. You couldn't really afford to be wrong. But now I increasingly feel that for normal people, especially people who can use Claude Code and do vibe coding, this process is just too slow. At least fifty times too slow. So what is the better way? I think it is: MVP as research. Don't think too much at the beginning. Just build the thing you want to build. Build the ugliest version. Or build a version where you can at least look at it and say, "Okay, I understand what this thing is." Because if you want to build it, that already means it is attractive to you. You want to use it, or you believe it might be useful. That is enough. Just build it. Some of my friends always say, "What if I build this thing and someone has already done it?" Or, "What if I build this and someone comes after me? What if I get into legal trouble?" And I usually say: after you build it, if someone really comes after you, then hire the best lawyer and fight the case. But if you don't even have the money to hire the best lawyer, why would they come after you in the first place? A lot of the time, we are not blocked by reality. We are blocked by tiny-probability events inside our own head. For example, I built Raily Friend in maybe two hours. I built CV.PRO in three or four hours. Raily took a little longer because it is an app, maybe a few days. But even so, compared with the traditional product development process, this speed is fucking insane. I saw one data point before: Flighty spent about two years in beta, from 2019 to 2021. And I basically replicated the core feeling of it in around twenty days. This is AI coding. This is the AI era. So I think the new process should not be: research first, then development. It should be: develop first, then use the launch itself as research. That is MVP as research. You build the thing first, then post it on WeChat Moments. Because people in your Moments know you. They have some relationship with you. Many of them live in a similar environment, share similar interests, or have a similar mindset. So when you publish it, someone will jump out. Some people will say, "This is wrong." Some people will say, "This is pretty good." Some people will give you suggestions. Some people will directly say, "Fuck it, I want to use this." Anyway, all of these are signals. And the most important thing is: if you do user research first, you don't even know who you should research with. You don't know which motherfucker in your Moments is actually interested in your thing. But when you drop a real product out there, the interested people naturally come out. At that moment, you can pull them in. Let them become your early users. Let them give you advice. Let them analyze it. Let them talk shit about it. Let them tell you why it sucks. That is much better than sitting there and imagining user needs by yourself. And another thing I think is very important: don't make the product too perfect at the beginning. I know a lot of people talk about this, but I mean it in a very practical way. You can build the product to 40 points, then package it like it is 60 points. If someone uses it, if someone complains about it, then you can hire another person, or hire an agent, or invest more time and money to push it to 80 points. I once built a tool that could import my school timetable into iCloud Calendar. At the beginning, it was extremely simple. You had to run it with Python. That means if you didn't know Python, you basically couldn't use it. It was just a CLI. But somehow, a few people were actually using it. And even more surprisingly, someone submitted a PR. That was the signal. So I kept going. I reverse-engineered the school login system so students could log in directly with their student account and password. Then I posted it on the school forum. After that, hundreds, even thousands of students started using it. So my understanding is simple: you should first drop the food on the ground and see if anyone eats it. If someone is willing to eat it even when it is on the ground, then you can give them a plate. ## 中文 在 AI 的时代，我们到底应该怎么去 ship 一个 product？ 传统来说，我们做一个产品，可能会先做用户调研，然后写需求，做设计，开发，测试，最后再发布。以前大厂里面都是这么干的。因为当时开发成本很高，一个产品如果做错了，代价会很严重。 但是我现在越来越觉得，对于普通人来说，尤其是对于会用 Claude Code、会 vibe coding 的人来说，这套流程实在是太慢了。至少慢了五十倍。 现在更好的方式是什么？我觉得是：MVP as research。 你不要一开始想太多。你直接先把你想做的东西做出来。做一个最简陋的版本，或者说，做一个你自己觉得「OK，我能看懂这是什么」的版本。 因为只要你想做这个东西，就至少说明它对你自己是有吸引力的。你想用，或者你觉得它可能有用。那就够了。先把它做出来。 我有些朋友经常会说，哎，那如果我做了这个东西，别人是不是已经做过了？如果我做了这个东西，会不会有什么影响？会不会有人过来找我麻烦？ 我一般就说，你等你做出来以后，他真的过来找你麻烦了，你再请最好的律师去跟他打官司不就完了吗？那如果你连请最好的律师的钱都没有，人家为什么要过来找你麻烦呢？ 很多时候，我们根本不是被现实拦住了，而是被脑子里那些极小概率的事情拦住了。 比如我之前做 Raily Friend，可能就花了两个小时。做 CV.PRO，也就三四个小时。Raily 时间稍微长一点，因为它是一个 App，可能花了几天。但即使这样，这个速度跟传统产品开发比起来，还是快得离谱。 我之前还看到一个数据，Flighty 的 beta 阶段做了两年，从 2019 年做到 2021 年。 所以我现在觉得，新的流程不应该是：先调研，再开发。而应该是：先开发，再用发布本身来做调研。也就是 MVP as research。 你先把东西做出来，然后发朋友圈。朋友圈里面的人是认识你的，跟你有关系，很多人的生活环境、兴趣、认知结构也跟你比较接近。这个时候你发出来，就会有人跳出来。 有的人会说，你这个做得不对。有的人会说，你这个挺好。有的人会给你提建议。有的人会直接说，我想用。Anyway，这些都是信号。 而且很关键的一点是，如果你一开始就去做用户调研，你根本不知道应该调研谁。你也不知道你朋友圈里面到底谁会对这个东西感兴趣。但是当你把一个真实的产品丢出去以后，感兴趣的人自然会冒出来。 这个时候你就可以把他们拉进来，让他们成为你的早期用户，让他们帮你提意见，帮你分析，帮你骂你。这比你自己坐在那里幻想用户需求啥都不做要好。 还有一点我觉得挺重要的：一开始千万不要把产品做得太完美。 有的时候，你需要先把产品做到 40 分，然后把它包装成 60 分。如果真的有人用，有人骂，有人给你提 PR，那你再去招一个人，或者拉一个 agent 进来，把它做到 80 分。 比如我之前做过一个学校课表导入 iCloud 日历的工具。一开始那个东西非常简陋，甚至必须要用 Python 跑。也就是说，如果你不懂 Python，你根本用不了，因为它就是一个 CLI。 但这玩意儿还真的有几个人在用，而且竟然还有人给我提 PR。 所以我就开始继续往下做。我去逆向了学校的登录系统，让大家可以直接用账号密码登录，然后再把它发到学校论坛里面去宣传。后来，几百个人，甚至上千个人都开始用这个东西。 所以我现在的理解是：你应该先把食物丢在地上，看有没有人吃。如果丢在地上都有人吃，那你再给他安排一个盘子。 --- # Powerball Effect > Published 2026-05-08 · By lawted · Canonical: https://ha7ch.com/writing/powerball-effect ## English Recently, I found this new thing. Maybe just tonight. I'm planning to call it the Powerball Effect. Yeah, fuck it. I named it. It's not that Powerball really makes you a fucking rich person. It means that before you get the result, you briefly live in a totally different world. I first felt this last Christmas, when I was on the way to Sequoia Park with my friend Jin. That day, the Powerball jackpot was something like $1.7 billion. So when we passed by a gas station, we just bought two tickets. And after buying those tickets, the world was fucking different. We started seriously talking about what if we actually won. At that time, I was visiting the U.S. with my B1/B2 visa, so I started thinking, if I really won this thing, how am I gonna take the money? Would the IRS tax me? Do I need a lawyer? Do I need to set up a trust? Should I not go back to China first? And my friend Jin was even more insane. He said, if you really win, maybe you won't be able to walk out of the hotel alive. I said, what? And he said, because if the winning ticket is yours, he might just take the ticket, drive the car away, and I will never see him again in my whole life. That was fucking crazy. We were laughing so hard in the car. Then later, when we got back to the hotel, I said I was going to take a shower. Jin said, don't you need to wait until 9? Because the result was coming out at 9. What he meant was, if the result came out while I was showering, and the winning ticket was mine, then maybe by the time I came back from the bathroom, the ticket would already be gone. And maybe he would be gone too. That was fucking stupid, but we were really high. And the craziest thing is, in the end, my ticket actually got three numbers right. Of course, I didn't win the $1.7 billion. But during those few hours, we had already lived a $1.7 billion life. That's what I mean by the Powerball Effect. It's not about actually winning. It's about holding a ticket that might change your life, and before the result comes out, you start seriously imagining another version of your life. And I found that startups are kind of the same thing. Recently, I've been building Raily, an app that feels like Flighty for rail. I posted something on RedNote and did some small promotions. At first, I didn't really think it would become anything. But then tonight, around 1 or 2 a.m., I suddenly found that some people actually liked this thing. Maybe just four or five people. But my feeling was: Oh my God, that's fucking a lot. Finally, someone saw me. So I posted another video. A couple hours later, about 70 people had joined the group. They wanted to try the app. The video had around 2,000 views, which means every 10 or 20 people, someone wanted to join the group and try this thing. And at that moment, I entered the Powerball Effect. I started thinking: What if this really works? What if Raily really becomes the Flighty for rail? Flighty is for airplanes, and it can do millions in ARR. What about rail? There are so fucking many people taking trains every day. China, Europe, Japan, the UK, Amtrak, commuters, students, business travelers, rail fans. So if Flighty can do $6M ARR, why can't Raily do $20M? Fuck that. Maybe even bigger. Anyway, this is the Powerball Effect. ## 中文 我最近发现一个东西，我暂时叫它 Powerball Effect。不是 Powerball 真的会让你发财，而是，在开奖之前，你会短暂地活在一个完全不同的世界里。 我第一次很强烈地感受到这个东西，是去年圣诞节，和我朋友金哥一起去 Sequoia Park 的路上。那天 Powerball 的奖池好像有 1.7 billion，我们两个在路上就顺手买了两张票。然后从买完票开始，整个世界就变了。 我们开始认真讨论，如果真的中了怎么办。我当时还是旅游签去美国，我就在想，如果我真的中了这个钱，我要怎么把它拿走？IRS 会不会扣很多税？我是不是要请律师？是不是要成立 trust？我要不要先别回国？ 金哥更离谱。他说，如果你真的中了，你可能不一定能从旅馆里活着走出来。我说什么意思？他说，因为如果开奖之后发现是你的票，他可能直接拿着票跑了，这辈子我就再也见不到他了。然后我们两个人在车上笑疯了。 那天晚上好像九点开奖。到了酒店之后，我说我要先去洗澡。金哥说，你不等九点开奖以后再洗吗？意思就是，如果我洗澡的时候开奖了，然后真的中了，那我出来以后可能票已经没了，人也没了。很傻逼，但真的很嗨。 最离谱的是，我那张票最后还真的中了三个数字。当然，没有中 1.7 billion。但那几个小时里，我们已经把 1.7 billion 的人生过了一遍。 这就是我说的 Powerball Effect。它不是中奖本身，它是你手里拿着一张可能改变命运的彩票，然后在开奖之前，你开始认真幻想另一个版本的人生。 我发现，创业也是这样的。这几天我在做 Raily，一个有点像高铁版 Flighty 的 app，然后我发了一些小红书，做了一些推广。一开始也没觉得会怎么样。 结果今天凌晨一两点的时候，我突然发现，有几个人开始真的喜欢这个东西。可能也就四五个人。但我当时的感觉是：Oh my God, that's fucking a lot. 终于有人看到我了。 于是我又发了一个视频。几个小时之后，有七十多个人加了群，视频两千多 views，大概十几二十个人里就有一个人愿意进群。我当时整个人就开始进入 Powerball Effect。 我开始想，what if this really works? What if Raily really becomes the Flighty for rail? Flighty 做飞机，可以做到几百万 ARR，那铁路呢？中国高铁每年那么多人坐，欧洲铁路也很多，日本、英国、Amtrak、跨城通勤、留学生、商务差旅、铁路爱好者，全世界坐铁路的人并不比坐飞机的人少，甚至很多地方铁路比飞机更高频。那如果 Flighty 可以做 $6M ARR，Raily 为什么不能做 $20M ARR？Fuck that, maybe even bigger. Anyway, 这就是 Powerball Effect. --- # Zero Token Design / 零 Token 设计 > Published 2026-05-07 · By lawted · Canonical: https://ha7ch.com/writing/zero-token-design ## English I'm not saying the product doesn't need AI. What I mean is: an AI product doesn't always need to burn its own tokens at runtime. Think about how we built AI SaaS products in the past. The logic was simple: every time a user clicks, the product calls the model. Another click, another request. Another request, another token burned. That made sense in 2023, because back then, most users didn't really have their own AI workspace. But now it's different. Codex, Claude Code, OpenCode, Cursor, and other local agent workspaces are getting more and more mature. They are no longer just chatboxes. They can read your directory, read your documents, run bash commands, edit code, run npx or uvx, and handle a whole workflow inside the user's own environment. So I think the next generation of AI products can be designed differently. Not every product needs to put a chatbot inside the webpage and pay for all the reasoning costs by itself. A better way might be: let users finish the reasoning inside their own agent workspace, and then push the result back into your product. And this is different from Bring Your Own Key. BYOK still asks users to apply for an API key, configure the key, understand billing, and put that key into your product. Honestly, that's just transferring the bill to the customer, but the UX is still heavy. Zero-token design is not about asking users to configure a key inside my product. It is about letting my product work with the tools they already use, like Codex, Claude Code, OpenCode, or Cursor. CV.pro is a zero-token AI product in my understanding. It is definitely an AI product, because resume parsing, JD tailoring, and content rewriting all need AI. But these things don't have to happen on CV.pro's server. The more natural way is: the user copies a quick-start prompt, and maybe there is an npx command inside that prompt. Then they paste it into their own agent workspace. The agent runs the command, parses the file, handles the errors, generates the structured schema, and pushes that schema back to CV.pro. So CV.pro is responsible for the schema, database, URL views, versions, rendering, and distribution. Basically: Let the user's agent do the work. Let the product handle the result. And I think every founder building AI products should think about this carefully. Should your product do all the reasoning by itself? Or should it become a system that can be operated by agents, capture the result, and distribute it beautifully? That's the difference. ## 中文 它不是说产品不用 AI，而是说：AI 产品不一定要在自己的 runtime 里烧 token。 以前做 AI 产品，默认逻辑是用户点一下，产品调用一次模型；用户再点一下，产品再烧一次 token。这个在 2023 年可能合理，因为用户没有自己的 AI 工作台。但现在不一样了。 Codex、Claude Code、OpenCode、Cursor 这种本地 agent workspace 已经越来越成熟。它们不只是聊天框，而是可以在用户自己的项目目录里读文件、跑命令、改代码、执行脚手架的工作台。 所以我觉得，下一代 AI 产品的默认架构应该变了。 不是每个产品都自己包一个 chatbot，然后自己承担推理成本。更好的方式是：让用户在自己的 agent workspace 里完成推理，把结果写回产品。 这和 Bring Your Own Key 不一样。BYOK 还是让用户去申请 API key、配置 key、理解 billing，本质上只是把账单转嫁给用户，体验很重。零 Token 设计不是让用户把 key 塞进我的产品，而是让我的产品进入用户已经在用的 Codex / Claude Code / OpenCode 工作台。 CV.pro 就是我理解里的零 Token AI 产品。它当然是 AI 产品，因为简历解析、JD tailoring、内容重写都需要 AI。但这些动作不应该都发生在 CV.pro 的服务器里。更自然的方式是，用户复制一段 quick start prompt，丢进自己的 agent，agent 自己去跑 npx、解析文件、处理错误、生成结构化数据，然后写回 CV.pro。 CV.pro 负责的是 schema、数据库、URL view、版本、展示和分发。 说白了就是，让 agent 去干活，让产品把结果接住。 这可能也是我现在对下一代 AI 产品的一个判断：不要急着在网页里塞一个聊天框。先想清楚，你的产品到底应该自己推理，还是应该变成一个能被 agent 操作、能沉淀结果的系统。 --- # So WTF is HA7CH > Published 2026-04-30 · By lawted · Canonical: https://ha7ch.com/writing/so-wtf-is-ha7ch ## English 3 a.m. on my side. I was chatting with my American friend on WeChat, morning for him. Mid-conversation, he tossed out an idea: what if we build Raily Friends, a high-speed rail travel buddy? I said, let's fucking go. Two hours later, my frontend was done. I went to sleep. He picked up the relay on the U.S. side, building out the chat backend. By noon China time, the thing was actually usable. We posted it on WeChat Moments right on schedule. It got a wave of attention and feedback the moment it went live. That was the moment we realized something. We really can ship in 48 hours. So, WTF is HA7CH? It's a tiny little club. Anyone here can speak up with an idea, we build it fast, or build it together with you. The moment it ships, we throw it on WeChat Moments and RedNote for people to use. If people use it, we keep maintaining it. If nobody does, we drop it and move to the next one. The reason this works is because vibe coding is fucking fast right now. As long as you have a decent idea, you don't need to wait for requirement docs, you don't need to wait for design mocks, you don't need to wait for anything. Just build it. And because we ship fast, we don't get too attached to any single product. We don't cling. Maybe this week it's this product, next week it's the next one. Before ByteDance hit it big, Zhang Yiming had built 5 products. He failed 5 times. Back then there was no AI, and shipping anything took forever. One or two years per product was the norm. In the AI era, failing 500 times before you hit it is fine. After all, our shipping speed is 100 times faster. ## 中文 凌晨三点，我跟我美国的朋友在微信上聊天。他那边是早上。聊着聊着，他突然冒出一个想法：要不我们做一个 Raily Friends，高铁搭子？ 我说，let's fucking go。 两个小时后，我这边前端就做完了。做完我就去睡了。他在美国那边接力，把后端的聊天功能都补齐。等中国中午十二点的时候，东西已经能用了。准时发朋友圈。 一上线就收获了一波关注和反馈。 那一刻，我们意识到一件事。We really can ship in 48 hours. So, WTF is HA7CH? 它就是一个小小的 club。 在这里，谁有 idea 就直接说出来，我们快速把它做出来，或者拉上大家一起做。做完立马丢到朋友圈、小红书，给大家用。 有人用，我们就继续维护下去。没人用，立马放掉，去做下一个。 之所以能这样，是因为 vibe coding 现在快到这种程度。你只要有一个不错的想法，不需要等需求文档、不需要等设计稿、不需要等任何东西，直接做就行。 而且因为做得快，我们对单个产品的期待反而没那么高了。我们不会死守。也许这一周做这个产品，下一周就开始做下一个了。 字节跳动成功之前，张一鸣做了 5 个产品，失败了 5 次。那个时候没有 AI，做东西很慢，一个产品做一两年是常态。 在 AI 时代，我们成功之前失败 500 次也没什么。毕竟，我们的开发速度提升了 100 倍。