Dream HubDream Hub

时间，给了我们梦想成真的勇气 Time Gives Us the Courage to Make Dreams Come True

十五年前，一个小女孩写下她的"玻璃房子"。十五年后，世界模型开始把这种想象搬进真实的物理仿真。中间走过的，不是技术，而是时间。 Fifteen years ago, a little girl wrote about her "glass house." Fifteen years later, world models are beginning to move that kind of imagination into real physical simulation. What stretches between the two is not technology — it is time.

I.李飞飞所说的"世界模型"What Fei-Fei Li Means by "World Model"

李飞飞，定义"空间智能"与"世界模型"方向的领军科学家，最近反复强调一个判断：世界不是由文字构成的，语言无法承载物理世界的全部规则与因果。仅靠大语言模型，无法抵达 AGI。她把通用智能的根基重新拉回到三件事——三维空间感知、物理世界仿真、具身行动。

Fei-Fei Li — a leading scientist defining the direction of "spatial intelligence" and "world models" — has lately been consistent on one judgment: the world is not made of words. Language cannot carry the full set of rules and causality of the physical world. Large language models alone will not reach AGI. She pulls the foundation of general intelligence back to three things — three-dimensional spatial perception, physical-world simulation, and embodied action.

她进一步指出，世界模型分三层：渲染器（Renderer）、规划器（Planner）、仿真器（Simulator）。行业真正长期攻坚的方向，是物理仿真级别的世界模型，而非表层的视频生成或对话 Agent。

She further frames the world model as three layers: Renderer, Planner, Simulator. The real long-horizon problem the industry must solve is the physics-grade simulator — not surface-level video generation, not chat agents.

II.十五年前，一个小女孩的"玻璃房子"Fifteen Years Ago, a Little Girl's "Glass House"

那是 2011 年前后，我的女儿还是个懵懂的小学生。她把对未来房间的全部想象一笔一笔地写了下来。我把原文照录如下——一个孩子用最朴素的语言，描述了一个"可编程、多模态、温度自适应、按语音指令调度物体"的房间。

It was around 2011. My daughter was still a young primary-schooler. She wrote down every detail of her imagined future room. I am reproducing her text verbatim below — a child describing, in the plainest language, a programmable, multi-modal, temperature-adaptive, voice-commanded room.

A child's letter to her future room · circa 2011

我的玻璃房子

要进入我的房间，不需要任何卡或钥匙，只需要说出我的秘密咒语："巴利巴利——碰！"房门便会自动打开。

进入了我的房间，你第一眼便会看到一个大大的、像果冻一样的东西。是什么东西呢？告诉你吧，这是我的床，这种床是用粉红色的软玻璃做的，非常柔软舒适，而且它还配有一个遥控器。我一按"切换模式"键，床就会变成沙发，沙发还会变成床。而且，它还能变换颜色呢！一按"蓝色"键，我就会来到海绵宝宝居住的海底世界，不仅床变成了蓝色，墙面也会变成蓝色的，我置身在海底世界里，周围有海绵宝宝以及许多他的朋友陪伴着我。一按"绿色"键，我就来到了喜羊羊居住的青青草原，床和墙面都变成了绿色的，还有喜羊羊、美羊羊和我做伴哦！再按"还原"键，床就又变成粉红色的了。我的床还有一个功能，那就是温度自动变换。夏天，我躺上去的时候，它就会变得非常凉快；冬天我躺上去的时候，它就会变得非常温暖。

我的书桌也是玻璃做成的，它会随着床颜色的变换而变换。书桌也配有遥控器。我一按"笔"键，就会出现一个画着许多种笔的三维立体收纳单出现在我的眼前，只要盯住一支笔，并说声"这个"，这支笔就会以 0.001 秒的速度出现在我的手中。

我的衣橱还是用全透明的玻璃做成的。只要盯紧里面的一件衣服，说一声"这个"再指一指我身上的衣服，说"进去"我盯住的那件衣服就会自动穿在我的身上，而我原来穿的衣服就会重新挂在衣橱里。

天天过这么舒适的生活，会不会发胖呢？爱漂亮的女生一定会问。不用担心，只要设置一个具体时间，房间里的闹钟就会发出"该运动啦！该运动啦！"的声音，我只要说出我想尝试什么体育项目，房间里运动器材区会用 1 秒的时间把我想要的器材准备好，到了一定的时间，又自动收回。

拥有这么舒适的房间，你一定很羡慕我吧！告诉你，我未来的房间 2022 年就会开始动工啦！

语音解锁、模态切换、3D 物体调度、姿态/手势交互、温度感知反馈、行为提醒系统——一个 2011 年的小学生，用作文的形式，提前描述了一个 物理世界的小型仿真器。她还没听说过"世界模型"，却已经在脑子里跑了一遍渲染器、规划器和仿真器。

Voice unlock, modal switching, 3D object dispatch, gaze/gesture interaction, temperature feedback, behavioral reminder loops — an elementary-schooler in 2011, through a primary-school composition, had already described a small simulator of a physical world. She had never heard the phrase "world model" — yet she had already run a renderer, a planner, and a simulator inside her head.

III.今天，Seedance 2.0 Mini 把它"渲染"了出来Today, Seedance 2.0 Mini Rendered It Out

今天，我用 Seedance 2.0 Mini 模型，把女儿当年那篇"玻璃房子"作文交给了 AI——让它把她写下的画面、色彩、家具、空间，第一次以视频的形式跑了出来。它当然还不是李飞飞口中的"物理仿真级世界模型"，但它已经是渲染器层级真实可用的一步。这一帧帧画面，对我来说不是 demo，而是一个母亲见证一个孩子十五年前的句子，在十五年后真正"生成"出来的瞬间。

Today, I gave that "glass house" composition to Seedance 2.0 Mini and asked it to render — for the first time — the scenes, the colors, the furniture, the spaces my daughter had written down. It is of course not yet the "physics-grade world model" Fei-Fei Li speaks of. But it is a real, usable step at the renderer layer. To me, these frames are not a demo. They are a mother watching a sentence her child wrote fifteen years ago finally "generate" itself.

视频 · 由 Seedance 2.0 Mini 根据 2011 年的作文文本生成 Video · Generated by Seedance 2.0 Mini from a 2011 composition

IV.为什么这件事让我相信"时间"Why This Makes Me Believe in "Time"

2011 年我女儿写下这篇作文时，没有任何渠道能让她相信这间房子会以任何形式被造出来。她在作文最后乐观地写："我未来的房间 2022 年就会开始动工啦！"——结果 2022 年没有动工，但 2026 年，AI 替她跑了第一帧画面。

When my daughter wrote this in 2011, there was no channel by which she could believe such a room would ever take any form. She optimistically ended her essay with "My future room will start construction in 2022!" — 2022 came, construction did not. But in 2026, AI rendered her first frame for her.

这十五年发生了什么？只是时间过去了。但时间不是空的——它是 GPU 算力翻了几个数量级、是数百万小时的视频被结构化、是世界模型从一个论文术语变成产业方向、是李飞飞这样的人坚持把"空间智能"四个字喊到被听见。

What happened in those fifteen years? Time simply passed. But time was not empty — GPU compute jumped orders of magnitude, millions of hours of video were structured, "world model" went from a paper term to an industry direction, and scientists like Fei-Fei Li kept pushing "spatial intelligence" until the phrase was heard.

所以当我今天作为一个 70 后外贸人坐下来学习 AI、学习世界模型、学习一个 12 岁孩子曾经凭直觉描述过的渲染器与仿真器时，我并不焦虑。我看的不是"我能不能赶上"，而是"时间会不会站在愿意走进去的人这一边"。

So when I — a 1970s foreign-trade practitioner — sit down today to learn AI, world models, the renderer and simulator a twelve-year-old once intuitively described, I am not anxious. I am not asking "can I catch up." I am asking "does time stand with those who choose to step inside."

那个写下"玻璃房子"的小女孩，今天在医学院读博。
她当年的句子，今天被 AI 第一次还了一帧。

原来时间从不催促，它只是悄悄替我们，把想象变成可能。 The little girl who once wrote about a "glass house" is now a PhD candidate in medicine.
The sentences she wrote back then have, today, been rendered one frame at a time by AI.

Time, it turns out, never hurries us. It quietly turns our imagination into the possible.

AI 时代，我为什么反而更看好 OPC（一人公司） Why I'm More Bullish on the One-Person Company in the AI Era

说起来有点感慨。二十多年前我刚成立公司的时候，几乎没人提"一人公司"这个概念。那时候做外贸的人，更习惯叫 SOHO。名字变了很多次，但本质一直没变：一个人，对自己的业务负责。 It's a strange thing to look back on. Twenty-some years ago, when I first started my company, almost no one used the phrase "one-person company." People doing foreign trade back then called it SOHO. The name has changed many times, but the essence never did: one person, accountable for their own business.

我的公司也经历过很多阶段。最开始有跟单、发货、对账，也请过员工。后来随着业务调整，一个岗位一个岗位慢慢减少，到最后，反而稳定成了我一个人。

My company went through many phases. In the beginning there were people doing order follow-up, shipping, and reconciliation. I hired employees. Later, as the business shifted, one role after another quietly disappeared — and in the end it settled into just me.

现在回头看，很难说这是主动选择，还是一路摸索后的自然结果。只是有一天，我忽然发现，公司变小了，自己反而睡得更踏实了。

Looking back now, I can't quite say whether it was a deliberate choice or just a natural outcome after years of feeling my way forward. I only know that one day I noticed: the company had gotten smaller, and I was sleeping more soundly.

二十多年外贸，我见过行业繁荣，也经历过低谷；见过有人迅速做大，也见过不少公司悄悄消失。最近几年，AI 来了。我看到很多人重新燃起创业的热情，也看到不少人开始认真思考：未来，一个人还能不能做好一家公司？

Twenty-some years in foreign trade — I've seen the industry boom and I've been through the troughs. I've seen people scale up fast, and I've seen no shortage of companies quietly disappear. In recent years, AI arrived. I've watched many people rekindle the urge to start something, and I've watched many more begin to seriously ask: in the future, can one person still run a company well?

这篇文章，没有标准答案，只是想分享一些这些年让我越来越确信的事情。

This essay has no definitive answer. I just want to share a few things that, over the years, I've become more and more sure of.

I.一人公司，到底靠什么活下来？What Actually Keeps a One-Person Company Alive?

我的业务模式其实很简单。接海外订单，对接国内工厂，帮助双方完成合作。没有自己的工厂，没有库存，也没有流水线。很多时候，我更像是连接两端的人。这样的模式，我做了将近三十年。

My business model is actually quite simple. Take overseas orders, connect them with domestic factories, help both sides complete the deal. No factory of my own, no inventory, no production line. Most of the time, I'm more of a bridge between two sides. I've been doing this for nearly thirty years.

后来我慢慢发现，比起商业模式本身，更重要的是两件事情。

Along the way, I gradually realized: more important than the business model itself are two other things.

第一，你要拥有别人不容易复制的价值。这个价值，不一定是什么高深技术。它可能是多年积累下来的客户信任。可能是对某个行业足够深入的理解。也可能是别人一句话听不出来，而你能立刻判断客户真正需求的经验。

First, you have to own value that's hard for others to replicate. That value isn't necessarily some advanced technology. It might be customer trust accumulated over many years. It might be a deep enough understanding of one particular industry. Or it might be the experience to hear what a customer really needs from a sentence that others would miss entirely.

这些东西，都不是今天注册公司、明天买套课程就能获得的。很多公司，团队越来越大，但真正支撑业务的，其实始终是创始人的判断能力。而一人公司的优势，也恰恰来自这里。不是因为人少，而是因为最重要的价值，本来就在自己身上。

None of this can be acquired by registering a company today and buying a course tomorrow. Plenty of companies keep growing their teams, but the thing actually holding the business up remains the founder's judgment. And that's exactly where the one-person company's edge comes from — not because there are fewer people, but because the most important value already lives inside you.

第二，你解决了别人没有解决好的问题。任何行业都会有一些"空白地带"。我当年切入外贸，就是因为工厂不会开发海外客户，而海外客户也很难找到真正合适的供应商。我没有生产产品。我只是把两边连接起来。

Second, you're solving a problem others haven't solved well. Every industry has its blank spaces. I moved into foreign trade because factories didn't know how to develop overseas customers, and overseas customers had a hard time finding truly suitable suppliers. I didn't manufacture anything. I simply connected the two sides.

后来越来越觉得，一人公司未必要什么都会。真正重要的是，你是否找到了那个属于自己的位置。那个位置，不需要很大，但最好别人暂时替代不了。

Over time I came to feel that a one-person company doesn't have to be good at everything. What actually matters is whether you've found the position that belongs to you. That position doesn't need to be big — but ideally, no one else can readily take it.

II.AI 来了，它到底改变了什么？AI Arrived — What Has It Actually Changed?

这是最近大家最关心的话题。我的感受可能没有那么激进。AI 当然很重要。但我越来越觉得，它更像一个"能力放大器"，而不是"能力制造器"。

This is the question everyone has been asking. My own take is probably less radical. AI is of course important. But I've come to see it more as an amplifier of ability than as a maker of ability.

如果你原本就在某个领域积累了经验，AI 会让这些经验发挥得更快、更高效。如果还没有建立自己的专业能力，AI 也许能帮助完成很多工作，但真正需要判断的时候，依然要靠自己。

If you've already built up experience in a domain, AI will let that experience move faster and hit harder. If you haven't yet built professional capability of your own, AI can certainly help you complete a lot of work — but the moment real judgment is needed, you're still on your own.

举个很简单的例子。以前写一封复杂的外贸邮件，我可能要花半小时。现在借助 AI，五分钟就能完成初稿。省下来的时间，我可以去沟通客户、了解供应链、思考新的机会。效率确实提高了。

A simple example. Writing a complex foreign-trade email used to take me half an hour. With AI, I can produce a solid draft in five minutes. The time I save goes into talking to customers, understanding the supply chain, and thinking about new opportunities. Efficiency really has improved.

但邮件为什么这样写、客户真正担心什么、这单值不值得接，这些决定依然来自过去很多年的积累。至少目前，在我的工作里，这部分仍然很难完全交给工具。

But why the email is written that way, what the customer is actually worried about, whether the order is even worth taking — those decisions still come from many years of accumulation. At least for now, in my work, that part is very hard to hand over to a tool.

所以现在我的工作方式越来越简单：真正需要我亲自完成的，是判断、沟通、决策和风险控制。其他大量重复性的工作，则尽量交给 AI 和各种工具。这样，我能把更多时间留给真正重要的事情。

So my way of working has become simpler over time: what I need to do myself is judgment, communication, decision-making, and risk control. The large volume of repetitive work goes to AI and other tools wherever possible. That way, I can spend more of my time on the things that actually matter.

III.这些年，我越来越相信的两件事Two Things I've Come to Believe More Deeply

第一，执行越来越容易，判断越来越珍贵。过去，一个业务员只要足够勤奋，就有机会跑出不错的成绩。今天，很多重复性的执行工作，AI 已经可以完成得很好。

First: execution keeps getting cheaper — judgment keeps getting more precious. In the past, a diligent sales rep could produce respectable results on effort alone. Today, much of that repetitive execution is already something AI does well.

这并不意味着努力不重要。只是努力的方向，正在发生变化。真正越来越有价值的，是判断力。知道什么时候该坚持，什么时候该放弃。知道客户真正关心什么。知道供应链出现问题时，应该先联系谁。

This doesn't mean effort no longer matters — only that where to put the effort is shifting. What genuinely grows in value is judgment. Knowing when to hold the line and when to walk away. Knowing what the customer actually cares about. Knowing who to call first when the supply chain breaks.

这些经验，很难一夜之间获得。它们往往来自一次次真实的合作，也来自很多并不轻松的经历。

This kind of experience can't be picked up overnight. It comes from real deals, one at a time — and from more than a few experiences that were anything but easy.

第二，一人公司的自由，也意味着更多责任。很多人向往一人公司的自由。我也确实享受这种工作方式。但如果说没有压力，那肯定不是事实。

Second: the freedom of a one-person company also means more responsibility. Many people are drawn to the freedom of it. I do enjoy this way of working. But it would be dishonest to say there's no pressure.

没人安排工作，也没人替你承担风险。订单出了问题，需要自己解决。市场变化了，需要自己调整。很多决定，都只能自己做。有时候，这种自由其实也伴随着孤独。

No one hands you the work, and no one shares the risk. When an order goes sideways, you fix it. When the market shifts, you adjust. Many decisions have to be made alone. And sometimes, that freedom carries loneliness with it.

所以，如果有人问我："一人公司是不是最好的选择？"我的回答一直都是：未必。它只是众多工作方式中的一种。有人适合团队协作，有人适合企业平台，也有人更喜欢独立经营。没有哪一种更高级，也没有哪一种一定更成功。重要的是，找到适合自己的节奏。

So when someone asks me, "Is the one-person company the best choice?" my answer has always been: not necessarily. It's just one of many ways of working. Some people thrive in teams, some belong on corporate platforms, and some prefer to run things solo. None of them is more advanced than the others, and none of them guarantees success. What matters is finding the rhythm that fits you.

IV.最后想说的话One Last Thing

这些年，我越来越觉得，AI 没有改变商业最底层的逻辑。它只是让很多事情变得更高效。

More and more, I've come to feel that AI hasn't changed the underlying logic of business. It has simply made a great many things more efficient.

如果你已经在某个行业积累了客户、经验、资源，AI 可能会成为一个非常好的助力，让很多过去耗费时间的工作变得轻松。如果还在起步阶段，也不用因为别人跑得快而焦虑。工具更新得很快，但真正属于自己的能力，依然需要时间慢慢建立。

If you've already built up customers, experience, and resources in some industry, AI can become a real force multiplier — making work that used to be draining feel almost light. If you're still at the beginning, there's no need to panic just because others seem to be running faster. Tools iterate quickly; the abilities that truly belong to you still need time to grow.

客户关系、行业理解、解决问题的能力，这些积累虽然慢，却往往也是最稳定的竞争力。

Customer relationships, domain understanding, the ability to solve problems — these accumulate slowly, but they tend to be the most durable competitive edge you'll ever have.

回头看这么多年的外贸经历，我最大的感受其实很简单。工具一直在变。互联网改变过行业，平台改变过行业，现在 AI 也正在改变行业。但真正决定一家小公司能不能走得长远的，始终还是那个使用工具的人。

Looking back on all these years in foreign trade, my biggest takeaway is actually quite simple. The tools keep changing. The internet changed the industry, platforms changed the industry, and now AI is changing it. But what determines whether a small company can go the distance has always been the person using the tools.

把时间花在建立自己的价值上，把工具用来放大自己的价值。
这样，无论下一次技术浪潮是什么，我们都更有底气一些。 Spend your time building your own value; use the tools to amplify it.
That way, whatever the next wave of technology turns out to be, we stand on slightly firmer ground.

结合最近几个月 AI 实操，谈谈感受：三个网站、两个平台、一场信任实验 Notes from a Few Months of Hands-On AI: Three Sites, Two Platforms, One Trust Experiment

外贸场景下，制造业 AI 落地并不是一个容易的项目。环节多、数据杂、跨系统。目前我能真正跑通、并且能亲手上下衔接的一环，是独立站的 SEO 与 GEO 推广——因为很多同行过去都把这块外包了，而这一块恰恰最适合先自己动手。 In foreign trade, bringing AI into manufacturing is not an easy project. Too many workflows, messy data, systems that don't talk to each other. The one link I can currently run end-to-end with my own hands is the standalone-site SEO / GEO layer — precisely because most peers used to outsource it, and it happens to be the best place to start doing it yourself.

I.三个网站，一套"前店后厂"的信任闭环Three Sites, One "Showroom in Front, Factory Behind" Trust Loop

以南通得力净化器材厂为例，我用 AI 工具目前建立了三个网站，走的是网站矩阵的路子：

Taking Nantong Deli Filter Factory as the example: with AI tools I have currently built three websites, deliberately as a site matrix:

• 品牌前台 · Brand-facing Airatmos.com——当我与注重品牌的客户谈判时，主推它，展示专业度。 Airatmos.com — the one I lead with when talking to brand-conscious buyers, to signal professionalism.
• 源头工厂 · Factory-of-record Ntdeli.top——当客户进入价格与供应链能力的深度评估阶段时，我适时抛出这一站，证明源头工厂的实力。 Ntdeli.top — brought out when the customer enters deep price and supply-chain due diligence, to prove the source factory is real.
• 测试型子站 · Fast-test sub-site airatmos.pub.atoms.world——挂在平台型子域名下，用来快速搭建、快速测试、快速迭代。它显著缩短了客户"从看到我们到相信我们"的周期。 airatmos.pub.atoms.world — sitting on a platform-hosted subdomain, used for fast build-out, fast testing, fast iteration. It measurably shortens the "from seeing us to trusting us" cycle.

"前店后厂"的相互印证——品牌站讲专业、工厂站讲实力、测试站讲速度——形成一个清晰的逻辑闭环。同一个企业主体，从三个不同角度被客户重复验证一次，比在同一个页面上把话说三遍要有力得多。

"Showroom in front, factory behind" — brand site speaks to professionalism, factory site speaks to depth, test site speaks to speed. Together they form a closed logical loop. Having a customer verify the same company entity three times from three different angles is far more persuasive than saying the same thing three times on one page.

II.为什么要用两个 AI 平台？——因为 Token 真的很贵Why Two AI Platforms — Because Tokens Are Genuinely Expensive

三个网站，我用的是两个 AI 平台：

Three sites, but two AI platforms:

• Accio Work 承担 Ntdeli.top 和 Airatmos.com 两站的建设与维护。 Handles Ntdeli.top and Airatmos.com.
• atoms.dev 承担 airatmos.pub.atoms.world 这个测试型子站。 Handles the fast-test sub-site airatmos.pub.atoms.world.

为什么要用两个？很直接的原因——Token 实在是很贵。当三个网站同时在跑的时候，筛选出一个相对性价比高的系统，或者让两个 AI 平台互为补充，是一个必要的选项，而不是"锦上添花"。

Why two? A blunt reason: tokens are genuinely expensive. When three sites are being built at once, picking a relatively cost-effective system — or letting two platforms complement each other — is a necessary choice, not a nice-to-have.

而且从更长的时间跨度看：未来我们如果要把大量数据迁移过来，要嵌入更为复杂的销售系统、生产系统、出货系统，一个稳定且好用的 AI 平台会非常关键。今天多花一点时间在平台的横向对比上，是为了明天不必把整个业务栈推倒重来。

On a longer horizon: when we eventually need to migrate large volumes of data, and plug into more complex sales, production, and shipping systems, a stable and usable AI platform becomes critical. The time spent today on side-by-side platform comparison is insurance against having to tear the whole stack down tomorrow.

III.两个平台并行三个月，一些粗浅的心得Three Months in Parallel — A Few Early Observations

A) 能力覆盖面：Accio Work 更全，atoms.dev 更小巧。Accio Work 在工具、技能的调用，以及插件和接口方面都更丰富一些。atoms.dev 更小巧，看起来也更便宜一些，因此更适合"新品推向市场的一个测试网页"或"参加一个特别展会时面向特定市场建立的落地页"。需要留意的一点是：atoms 的页面挂载的是 pub.atoms.world 平台型子域名，理论上会削弱一些 Google 自然搜索的信任与权重，不利于长期 SEO 权重积累。所以我把它定位为"测试"，而不是"主战场"。

A) Coverage: Accio Work is broader, atoms.dev is lighter. Accio Work has a richer set of tools, skill invocations, plugins, and integrations. atoms.dev is lighter and appears cheaper, which makes it a natural fit for "a test page for a new product going to market" or "a market-specific landing page built for a particular trade show." One caveat: atoms pages sit under the pub.atoms.world platform subdomain, which in principle dilutes Google organic trust and long-term SEO equity. That's exactly why I position it as a test bench, not the main battlefield.

B) 复杂度与透明度：各有取舍。Accio Work 在抓取资料、联系上下文这些环节表现比较出色，能完成的任务也相对更复杂。atoms.dev 不太适合完成复杂任务，但它的过程非常清晰，你能看到它每一步的运算；而且遇到难解问题时，它不太会执意反复重试，而是会直接建议联系客户支持——不浪费客户的 Token。这个"知难而退、不烧钱"的取向，其实挺能抓住客户心理的。

B) Complexity vs. transparency: each has trade-offs. Accio Work is stronger at fetching material and maintaining context, and can carry heavier tasks. atoms.dev is less suited to complex work, but its process is very transparent — you can watch every computation step. When it hits a hard problem, it doesn't stubbornly retry; it recommends contacting support so the customer's tokens are not burned in the loop. That "walk away from a dead end instead of burning money" posture actually resonates with customers.

C) 目前是初测阶段。两个平台都在初测，以上都只是一些粗浅的感受。随着任务的推进——尤其是当我把销售、生产、出货这些下游系统逐步接进来时——会有更多可以量化的观察。到那时我会继续在 Dream Hub 更新。

C) Still an early test phase. Both platforms are still in early trial for me, so everything above is preliminary. As the work extends downstream — especially as I start wiring in sales, production, and shipping systems — more quantifiable observations will surface. I'll keep updating them here in Dream Hub.

结论只有一句：外贸制造业的 AI 落地，先从"可自己动手的一环"开始跑通，再谈"整体升级"。
三个网站解决的是客户信任，两个平台解决的是自己的成本，剩下的每一步都需要时间。 One sentence conclusion: bringing AI into foreign-trade manufacturing should start from the one link you can run hands-on, before we talk about "end-to-end upgrade."
Three sites solve customer trust; two platforms solve your own cost; every step after that will take time.

实践笔记 · 记于 2026-07-06 · YingChao (YC) Practitioner note · written on 2026-07-06 · YingChao (YC)

用了一阶段的 AI，我才发现之前想得太简单了 After a Real Stretch with AI, I Realized I Had Been Thinking Too Simply

最近几个月我一直在测试一些 AI 工具，把它们用在实际工作里。怎么说呢——AI 确实厉害，但问题也不少。之前看别人用 AI，总觉得什么都能搞定、效率高得离谱；但你真的自己去用、去做长期项目，就会发现事情没那么美好。很多问题是你用几天根本发现不了的。 For the past few months I've been putting AI tools into real work. How to put it — AI is impressive, but the problems are just as real. Watching other people use AI, it looks like everything gets solved and efficiency goes through the roof. Actually running it yourself, on a project that lives longer than a demo, reveals a very different picture. Most of the issues never surface in the first few days.

I.Vibecoding 建站：开局惊艳，一个月后隐性问题集中爆发Vibecoding Sites: Stunning at Launch, Hidden Problems by Month One

前段时间用 Vibecoding AI 建站，开局体验真的堪称惊艳。建站速度、语义理解、内容质感，全方位吊打传统模式。最惊喜的是它的 SEO 能力——新站上线没多久，就能快速被搜索引擎收录、被主动推荐，起量速度让人直呼高效。

I recently built sites with vibecoding-style AI, and the opening experience was genuinely stunning. Build speed, semantic understanding, content quality — it out-classed the traditional workflow on every axis. The biggest surprise was SEO: not long after launch, new sites were indexed and actively surfaced by search engines. The ramp-up was fast enough to feel unreal.

本以为找到了降本增效的最优解，可不到一个月，很多隐性问题开始爆发。AI 建站的不稳定性，在真实运营场景里开始暴露：

I thought I had found the optimal cost-cutting, efficiency-boosting answer. Within a month, hidden issues started firing. The instability of AI-built sites showed up in real operating conditions:

• ✅ 语言污染 · Language contamination 纯英文官网，页面里莫名穿插错乱的日语语句。 A purely-English site, with random broken Japanese sentences appearing inside pages.
• ✅ 关键词跳水 · Keyword drop-off 原本稳定的核心关键词，断崖式掉出搜索推荐。 Core keywords that had been ranking steadily fell off a cliff out of search recommendations.
• ✅ 静态资源失效 · Broken media 页面图片、视频随机失效、无法展示。 Images and videos on the page randomly failed to load or render.

没有征兆、没有规律、无法预判。这一刻开始明白：AI 快速生成的成品，能快速出效果，但如果要长期运营，需要后续很多补救。

No warning, no pattern, no way to predict it. That was the moment I understood: what AI generates fast can produce fast results, but if you want it to survive as a long-running operation, a lot of downstream patch-work is required.

II.另一个现象：AI 的答案，常常不一致Another Pattern: AI Answers Are Often Inconsistent

除了建站，日常用各类 AI Agent，也发现一个非常有意思的现象：AI 的答案，常常不一致。

Beyond site-building, I noticed something else while using various AI Agents day to day: AI answers are frequently inconsistent.

不同 AI Agent，同一个业务问题，答案相去甚远；哪怕是同一个系统里的多个 Agent，输出的报告、结论也会互相矛盾。有时候哭笑不得，像办公室里两个信息不通、各说各话的同事，互相"扯皮"、各执一词，完全无法给到稳定统一的结果。

Different AI Agents give wildly different answers to the same business question. Even multiple Agents from the same system can produce contradicting reports and conclusions. At times it's tragicomic — like two office colleagues who don't share information, each insisting on their own version, unable to converge on a stable, unified answer.

刚好回看黄仁勋和马斯克的观点，瞬间通透了。

Rereading Jensen Huang and Elon Musk on this made it click:

黄仁勋：AI 没有意识，它只是一台超级智能机器。它不会思考、没有判断，只是基于模型和数据做输出。

Jensen Huang: AI has no consciousness. It is a super-intelligent machine. It does not think and it does not judge — it only produces output based on its model and data.

马斯克补充了最关键的一点：AI 的结果，取决于它的底层价值观和训练数据。

Elon Musk adds the key piece: the output of an AI is determined by its underlying values and training data.

数据有偏差、模型有局限、逻辑不闭环——输出就会出错。

If the data is biased, the model is limited, or the logic doesn't close — the output will drift.

III.两条非常现实的 AI 使用心得Two Grounded Rules for Actually Using AI Today

1 · 现阶段 AI：小场景试错，渐进落地，拒绝全盘托管。

1 · At this stage: test in small scopes, land things gradually, refuse full delegation.

真心不建议企业把核心业务、全盘数据、完整流程直接交给 AI。当下最稳妥的玩法：让 AI 去做小场景、跑通小任务，做完必复盘、必评估、必跟踪；确认结果稳定、可控、可落地，再一点点迭代、拓展新场景、嵌入新业务流程。AI 可以提效，但不能兜底——大面积数据迁移、全业务 AI 替代，目前来看，可能还没有可靠的底层支撑。

I honestly don't recommend that a company hand its core business, its full data set, and its complete workflow over to AI. The safest current play: let AI take on small scopes and small tasks; every completion goes through a review, an evaluation, and tracking; only once results are stable, controllable, and truly deployable do you slowly iterate, extend to new scenarios, and embed it into new workflows. AI can boost efficiency, but it cannot be the safety net. Large-scale data migration and full-business AI replacement still lack, at least for now, the reliable underlying foundation they need.

2 · 拥抱 AI：收起人类的工具傲慢，保持敬畏与好奇。

2 · Embrace AI — but put down the human tool-arrogance and hold onto humility and curiosity.

我们太习惯掌控工具了。过往所有生产工具，人类都可以吃透原理、精准操控、预判结果。但 AI 不一样——它的迭代、运算逻辑、隐性风险，至今还有太多未知。我们并不完全清楚：AI 为什么出错？偏差如何产生？隐藏漏洞在哪里？所以，设置安全边界、做好风险管控——不是保守，是负责。

We are far too used to being in full command of our tools. Every prior production tool let humans internalize the principles, control the operation precisely, and predict the outcome. AI is different. Its evolution, its computation logic, and its hidden risks still hold too many unknowns. We do not fully know: why does AI err? how do biases emerge? where do the hidden failure modes live? Setting safety boundaries and running real risk controls — that is not conservatism, it is accountability.

不盲目神化 AI，不极致依赖 AI。对新技术保持好奇，对未知局限保持敬畏——才是这个时代最清醒的姿态。

Don't mystify AI. Don't over-depend on AI. Stay curious about the new, stay humble about what it cannot yet do — that is the most clear-headed posture for this era.

AI 是趋势，是利器，但不是万能解药。
顺势拥抱，审慎驾驭，稳步迭代——才是普通人、小微企业最好的 AI 打开方式。 AI is the direction and a serious tool — but not a universal cure.
Ride the wave, steer with care, iterate in steady steps — that is the best way for regular practitioners and small businesses to open the AI door.

实践笔记 · 记于 2026-07-18 · YingChao (YC) Practitioner note · written on 2026-07-18 · YingChao (YC)

当 Agent 之间还不能"对话"：人机协同不是妥协，是现阶段最务实的解法 When Agents Still Can't Talk to Each Other: Human-in-the-Loop Isn't a Compromise, It's Today's Most Pragmatic Mode

最近在搭建 Shopify 店铺的时候，我深刻体会到了一个很有意思的现象——我们正站在一个挺尴尬但又充满可能性的交叉点上。一方面，像 Accio 这样的 AI Agent 平台已经能够帮我们解决大量实际问题；另一方面，不同应用之间的 Agent 还没法真正"聊起来"。中间那一段路，目前只能靠人自己走。 Building a Shopify store recently made something very clear to me — we are standing at an awkward but genuinely promising intersection. On one side, AI Agent platforms like Accio can already solve a large share of real problems. On the other, Agents living in different applications still cannot truly "talk" to each other. The stretch in between is, for now, one a human has to walk alone.

I.冷启动这一段：Accio 确实有两把刷子The Cold-Start Stretch: Accio Really Does Deliver

在初期建站和产品优化这块，最让我印象深刻的是它对产品页面的优化能力。你知道的，一个新店铺最头疼的就是产品描述怎么写、图片怎么排版、价格策略怎么定。以前这些事情要么自己瞎琢磨，要么花钱请代运营，结果往往还不尽如人意。

For early site-building and product optimization, what impressed me most was how it handles product pages. As anyone who has opened a store knows, the hardest part at the start is writing the product copy, laying out the images, and setting the pricing strategy. In the past you either guessed your way through it or paid an agency — and the result was rarely what you hoped for.

但 Accio 的 Shopify 插件能根据你的产品类型、目标市场，甚至竞品分析，给出相当靠谱的优化建议。它不是简单地套模板，而是真的在理解你的业务逻辑——这种感觉就像有个懂行的朋友在旁边指点，而不是一个冷冰冰的工具在执行指令。

But the Accio Shopify plugin gives genuinely sound recommendations based on your product category, your target market, even competitor analysis. It is not filling in a template; it is actually reading your business logic. It feels like having a knowledgeable friend beside you, rather than a cold tool executing instructions.

建站初期那段时间，我基本上是把大部分繁琐的工作都交给了它。从产品标题的关键词布局，到详情页的转化率优化，再到库存管理的自动化设置，它处理得井井有条。说句实在话，这让我这种非技术背景的卖家省了太多心，也让店铺能够更快地上线运营。

During that early phase I handed over most of the tedious work. Keyword placement in product titles, conversion optimization on detail pages, automated inventory rules — all of it came back organized. Honestly, for a seller without a technical background, that saved an enormous amount of effort and got the store live far faster.

II.店铺上线之后：SEO 这一步，卡在了 Agent 之间的断点上After Launch: SEO Gets Stuck at the Gap Between Agents

但问题来了——当店铺真正上线，开始需要做 SEO 优化的时候，事情就变得复杂起来。

Then the problem appeared: once the store was actually live and SEO work began, things got complicated.

Shopify 本身也有自己的 AI Agent 系统，专门针对平台内的搜索排名、流量分析、广告投放等功能。理论上，如果 Accio 的 Agent 和 Shopify 的 Agent 能够无缝对接，那整个优化流程应该是一气呵成的。但现实是，这两个系统目前还没法直接"对话"。Accio 能帮你优化产品页面，Shopify 的 Agent 能帮你分析流量数据和 SEO 表现，可它们之间的信息传递，还得靠我这个"人肉中转站"。

Shopify has its own AI Agent system, built around in-platform search ranking, traffic analytics, and ad delivery. In theory, if Accio's Agent and Shopify's Agent could connect seamlessly, the whole optimization loop would run in one breath. In reality, the two systems cannot speak to each other directly. Accio optimizes the product page; Shopify's Agent analyzes traffic and SEO performance; and the information between them still travels through me — a human relay station.

• ① Accio → 人 · Agent to human 从 Accio 那里获取产品页面的优化建议。 Collect the product-page optimization recommendations from Accio.
• ② 人 → Shopify · Human to platform 手动把这些调整反馈给 Shopify 的 SEO 与广告系统。 Manually carry those adjustments into Shopify's SEO and advertising systems.
• ③ Shopify → 人 → Accio · Back around the loop 再把后台的数据分析结果取出来，输入给 Accio，让它根据实际表现继续调整策略。 Then pull the backend analytics back out, feed them to Accio, and let it refine the strategy against real performance.

这个过程倒不是说有多难，但确实打断了工作的流畅性，也增加了出错的可能。如果用 Accio 直接做 Shopify 店铺优化也是可以的，不过操作时似乎比较慢，而且有些指令好像也没有达到预期效果。

None of these steps is difficult in itself, but together they break the flow of the work and widen the room for mistakes. Driving Shopify optimization straight from Accio is possible too — it just tends to run slower, and some instructions do not land the way you expect.

III.不止 Shopify：现在的 Agent 生态还在"诸侯割据"Not Just Shopify: The Agent Ecosystem Is Still a Set of Walled Gardens

更让人头疼的是，如果你还用了其他工具，每个应用可能都有自己的 AI Agent，但彼此之间基本都是各自为政：

What makes it harder is that once you bring in other tools, each application tends to ship its own AI Agent — and each one largely runs on its own:

• ✅ 邮件营销 · Email marketing 自己的受众分群和发送逻辑，不知道店铺那边刚改了什么。 Its own segmentation and send logic — with no idea what just changed on the store side.
• ✅ 广告投放 · Google Ads 自己的关键词与出价体系，和产品页的优化结果对不上号。 Its own keyword and bidding system, disconnected from what was optimized on the product page.
• ✅ 客户管理 · CRM 自己的客户标签和跟进节奏，很难被上游的数据自动触发。 Its own customer tags and follow-up cadence, rarely triggered automatically by upstream data.

你想让这些智能系统协同工作？那就得自己当"翻译官"，在不同平台之间来回切换、复制粘贴、手动调整。说到底，现在的 Agent 生态还处在一个"诸侯割据"的阶段。每家都在努力让自己的 AI 更聪明、更好用，但跨平台的互联互通还远远没有实现。

Want these intelligent systems to work together? Then you become the translator — switching between platforms, copying, pasting, adjusting by hand. The Agent ecosystem today is still a landscape of walled gardens. Every vendor is working hard to make its own AI smarter and more usable, but cross-platform interoperability is nowhere near solved.

这其实也能理解，毕竟涉及到数据安全、API 标准、商业利益等一大堆复杂问题。但对于我们这些实际使用者来说，这种割裂确实带来了不小的摩擦成本。

It is understandable — data security, API standards, and commercial interests are all tangled up in it. But for those of us actually using these tools, the fragmentation carries a very real friction cost.

IV.换个角度：断点处，恰恰是人的价值所在Another Angle: The Gap Is Exactly Where Human Value Lives

不过换个角度想，这也正是人的价值所在。AI Agent 再强大，目前也还需要人来做那个"总指挥"——理解整体业务目标、协调不同系统、做出关键决策。Accio 在初期建站和产品优化上的出色表现，让我们能把更多精力放在战略层面的思考上，而不是陷在具体的执行细节里。

Seen from another angle, this is precisely where human value sits. However capable an AI Agent becomes, someone still has to conduct — to hold the overall business goal, coordinate across systems, and make the decisions that matter. Because Accio performs so well on early build-out and product optimization, we get to spend more of our attention on strategy instead of sinking into execution detail.

人机协同不是妥协，而是一种现阶段最务实的工作模式。

Human-in-the-loop is not a compromise. It is the most pragmatic way to work at this stage.

展望未来，我倒是挺期待看到各个应用的 Agent 能真正实现互联。那时候，从建站到运营、从产品优化到营销推广，整个链条或许真的能够自动化运转。

Looking ahead, I genuinely hope to see Agents across applications become interoperable. When that happens, the whole chain — from building the store to running it, from product optimization to marketing — may finally run end-to-end on its own.

但在那一天到来之前，像 Accio 这样在垂直领域做得扎实的工具，加上懂得如何协调资源的人，或许就是当下最好的解决方案。 Until that day arrives, a tool that is genuinely solid in its vertical — like Accio — plus a person who knows how to coordinate the resources, may well be the best answer available right now.

实践笔记 · 记于 2026-08-04 · YingChao (YC) Practitioner note · written on 2026-08-04 · YingChao (YC)

当车间经验遇上 AI：一篇行业期刊技术文章的完整拆解 When Shop-Floor Experience Meets AI: Anatomy of a Trade-Journal Technical Article

2026 年 8 月 5 日，我署名的技术文章《The Hidden Cost of the Wrong Dehumidifier Filter》发表在美国修复行业期刊 R&R Magazine（Restoration & Remediation，BNP Media 旗下）。这篇文章里没有一个数字是 AI 想出来的；但如果没有 AI，它大概率也走不到"发表"这一步。这篇拆解想说清楚的，就是这条分界线到底划在哪里。［来源 1］ On August 5, 2026, a technical article under my byline — "The Hidden Cost of the Wrong Dehumidifier Filter" — was published in R&R Magazine (Restoration & Remediation, BNP Media). Not one number in that article came from an AI. Yet without AI it probably would never have reached publication. This piece is about exactly where that line falls. [Source 1]

I.事实的源头只能是现场：一张装错滤芯的账单Facts Can Only Come From the Field: One Bill for the Wrong Filter

文章开篇是一个真实作业场景：一处 2,500 平方英尺的商用冷库发生 Category 2 水损，现场投入 6 台 LGR 除湿机，按 ANSI/IICRC S500 的干燥目标把湿度压到 40%。到第四天，读数卡在 55% 不动。项目经理判断是设备不够，又调来两台。工期超了三天，理赔员反过来质疑设备配置，把设备费这一项砍掉 15%，整单损失超过 15,000 美元。［来源 1］

The article opens on a real job: a Category 2 water loss in a 2,500-square-foot commercial cold storage facility, six LGR dehumidifiers on site, target humidity 40 percent per the ANSI/IICRC S500 drying protocol. By day four the readings were stuck at 55 percent. The project manager assumed the equipment was underpowered and mobilized two more units. The job ran three days over schedule, the adjuster questioned the equipment configuration and cut that line item by 15 percent, and the total loss exceeded $15,000. [Source 1]

真正的原因不在设备。上一单结束后，维护人员顺手拿了设施 HVAC 系统备用的 MERV-11 滤芯换上去——外形尺寸对得上，滤材看起来也差不多，但压降比规定的 HAF（High Air Flow）滤芯高出 2.5 倍。而 6 台机器换成正确滤芯的成本，是 90 美元。［来源 1］

The cause was not the machines. After the previous job, maintenance staff had swapped in spare MERV-11 filters from the facility's own HVAC system. Outer dimensions matched, the media looked similar — but pressure drop measured 2.5 times higher than the specified High Air Flow (HAF) filters. The correct filters for all six units would have cost $90. [Source 1]

这段判断没有任何一个环节是 AI 能生成的。"外形一样但压降差 2.5 倍""忙起来的技师抓的是手边有的、不是图纸上写的""第四天才暴露"——这些是三十年里被客户投诉单、退货单、返修单反复教出来的东西。AI 可以写出一段听上去很专业的滤芯说明，但它写不出这张账单。

Not one link in that chain is something an AI could have generated. "Identical outside, 2.5× the pressure drop." "Under time pressure a technician grabs what is available, not what is specified." "It only surfaced on day four." These come from three decades of complaint tickets, returns, and warranty repairs. An AI can write a filter spec sheet that sounds authoritative. It cannot write that bill.

II.AI 在这件事里真正承担的四件事The Four Things AI Actually Carried

把边界划清之后，AI 承担的部分其实非常具体，而且每一件都是我自己做会做得更慢、更差的：

Once the boundary is clear, what AI carried is very concrete — and every item is something I would have done slower and worse on my own:

• ① 把散点经验压成可读结构 · Structuring scattered experience "现场故事 → 机理解释 → 十二个月机队成本对比 → 运营层面的四条改法"，这个递进是 AI 帮我搭起来的。工厂的人讲这件事，通常一上来就是压降和 g/m²，读者跟不上。 "Field scenario → mechanism → twelve-month fleet cost comparison → four operations-level fixes." AI built that progression. Left to ourselves, factory people open with pressure drop and g/m² — and lose the reader.
• ② 术语对齐到读者的行业语言 · Aligning to the reader's vocabulary 同一件事，我们车间说"开孔通道 + 驻极静电"，修复行业读者关心的是 CFM 与静压的平衡、蒸发器盘管结冰、化霜循环空转。AI 的价值在于把前者翻成后者，而不是替我编后者。 Same fact, two vocabularies: our floor says "open-channel media with an electret charge"; a restoration reader cares about the CFM-versus-static-pressure balance, evaporator coil icing, and hours wasted in defrost. AI's value was translating the first into the second — not inventing the second for me.
• ③ 中英双语同时成稿 · Bilingual drafting in one pass 我的母语不是英语。过去这一步要么外包、要么写出"中式技术英语"。现在是我用中文把事实讲准，AI 负责让英文读起来像行业里的人写的。 English is not my first language. That step used to be outsourced, or it came out as "technical Chinglish." Now I get the facts right in Chinese, and AI makes the English read like someone inside the industry wrote it.
• ④ 发表之后的承接 · Post-publication landing pages 在 ntdeli.top 的 press 页（工程视角）和 airatmos.com（零售视角）同步做承接页、结构化数据与内链，这类"又碎又必须一致"的活，是 AI 最稳的部分。［来源 3］ Building the landing pages on ntdeli.top's press page (engineering angle) and on airatmos.com (retail angle), with structured data and internal links kept consistent across both — fragmented work that must not drift is where AI is most reliable. [Source 3]

III.我给 AI 划的三条红线Three Red Lines I Set for the AI

"不作恶"这三个字，落到这件具体的事上，就是三条可执行的规则：

"Don't be evil" is abstract until it becomes three executable rules on a specific job:

• 红线一：每个数字必须可回溯 · Every figure must be traceable 承接页上的 2.5 倍压降、90 美元、15,000 美元损失、300–400 CFM，全部标注来自 R&R 原文；0.08–0.15 in.w.c. @ 250 CFM 这类产品参数，标注来自我们自己的产品页与白皮书。不允许出现"约""据估计""行业普遍认为"这种没有出处的修辞。［来源 1、3、4］ On the landing page, the 2.5× pressure drop, the $90, the $15,000 loss, the 300–400 CFM are all attributed to the R&R article; product figures such as 0.08–0.15 in.w.c. at 250 CFM are attributed to our own product page and white paper. No "approximately," no "industry consensus suggests," no unsourced rhetoric. [Sources 1, 3, 4]
• 红线二：投稿 ≠ 背书 · A contributed article is not an endorsement 这是一篇由厂商员工撰写的投稿技术文章，发表本身不构成 R&R Magazine 或 BNP Media 对任何产品的评测、背书或推荐。承接页上这句话是写死的，不允许 AI 在改写时"优化"掉。［来源 3］ It is a contributed technical article written by a manufacturer's employee. Publication is not a review, endorsement, or recommendation by R&R Magazine or BNP Media. That sentence is hard-coded on the landing page, and the AI is not allowed to "optimize" it away. [Source 3]
• 红线三：引用标准 ≠ 通过认证 · Citing a standard is not certification 文章引用 ANSI/IICRC S500 是作为干燥作业的行业标准来引述语境。S500 是作业程序标准，它不认证、不批准、不列名任何滤芯产品，我们也从未声称任何产品通过 S500 认证。［来源 3］ The article cites ANSI/IICRC S500 as the industry standard behind the drying targets. S500 is a procedural standard for restoration work — it does not certify, approve, or list filter products, and we have never claimed any product is certified under it. [Source 3]

这三条不是合规部门要求的，是我自己要的。制造业的信任是十年攒起来、一句话赔光的东西；AI 让"多说一句好听的"成本几乎为零，所以红线必须由人来守。

No compliance department asked for these. I did. In manufacturing, trust takes ten years to build and one sentence to spend. AI drives the cost of "one more flattering claim" to nearly zero — which is exactly why a human has to hold the line.

IV.最有说服力的那一段，恰恰是最不像广告的那一段The Most Persuasive Passage Is the One That Reads Least Like an Ad

文章里给运营层面开的四条改法，其中有一条是明确劝客户不要省钱的反向操作：

Among the four operations-level fixes, one explicitly tells the customer not to save money the way they were planning to:

永远不要清洗 HAF 滤芯。滤材里的静电荷是永久的，但很脆弱——水会破坏它。洗过的 HAF 滤芯看起来很干净，但压降显著升高、捕集效率下降。

Never wash HAF filters. The electrostatic charge in the media is permanent but fragile — water destroys it. A washed HAF filter looks clean but has significantly higher pressure drop and reduced particle capture efficiency.

还有另外三条：给整个机队建一张"型号 → HAF 件号 → 尺寸 → 供应商"的对照表，贴在设备库房和每台机器上；把换滤芯从"开工前"挪到"收工后"的检查清单里；高粉尘作业（拆除、火灾、霉菌治理）把常规的 250–300 运行小时缩短为每 100 小时目检、见脏就换。［来源 1］

The other three: build a fleet matrix mapping each dehumidifier model to its verified HAF part number, dimensions, and supplier source, and post it in equipment storage and on each unit; move filter replacement from the pre-job to the post-job checklist; and in heavy-particulate work (demolition, fire damage, active mold remediation), shorten the standard 250–300 operating hours to a 100-hour inspection interval with replacement on any visible loading. [Source 1]

这四条里没有一条在卖东西。但正是因为没有一条在卖东西，第五条——"你把机队型号清单发给我，我帮你把对照表做出来"——才有人愿意接。AI 帮我把这四条写得清楚，可"先给出四条对客户有用的、其中还有一条是劝他别花冤枉钱的"这个决定，是人做的。

Not one of the four sells anything. And precisely because none of them sells anything, the fifth line — "send me your fleet model list and I'll map the matrix for you" — gets taken up. AI made those four points clear on the page. The decision to lead with four genuinely useful items, one of which tells the buyer to stop wasting money, was a human one.

V.可复制的部分：把车间的隐性知识，变成可被引用的知识The Reproducible Part: Turning Tacit Shop-Floor Knowledge Into Citable Knowledge

如果要把这次的经验抽象成一套别人也能用的方法，我会这么写：

If I had to abstract this into a method someone else could reuse, it would read like this:

• 第 1 步 · 先找"赔过钱的那一刻" · Start from the moment money was lost 不要从产品优势开始写。从一个真实的失败开始写——失败自带细节，而细节是 AI 编不出来的护城河。 Don't start from product advantages. Start from a real failure. Failures come with detail, and detail is the moat an AI cannot fabricate.
• 第 2 步 · 人给事实，AI 给结构 · Human supplies facts, AI supplies structure 把口述的现场、账单、参数原样交给 AI，只让它做归类、递进和双语。凡是 AI 主动"补"出来的数字，一律删掉重查。 Hand over the raw account, the bills, the parameters as they are, and let AI do only classification, progression, and bilingual drafting. Any number the AI volunteers on its own gets deleted and re-verified.
• 第 3 步 · 给每个数字挂上出处 · Attach a source to every number 发表页、承接页、社媒摘要，三处的数字必须同源。做不到同源，宁可不写这个数字。 The published article, the landing pages, and the social summaries must all draw from the same source. If they can't, the number doesn't get used.
• 第 4 步 · 承接页只做"摘要 + 外链" · Landing pages summarize and link out 正文留在期刊，自己站上只放摘要、数据表复述与出处清单，并明确标注归属。既尊重发表方，也让引用关系对搜索引擎和 AI 检索是干净的。［来源 3］ The full text stays with the publisher; our own site carries a summary, a restatement of the data table, and a source list with explicit attribution. It respects the publisher and keeps the citation graph clean for search engines and AI retrieval alike. [Source 3]

这套方法对我这样的人特别友好：手上有事实但没有表达渠道的人。过去这类知识只存在于车间老师傅的脑子里，和客户投诉邮件的附件里。AI 做的事情，是把它们搬到能被行业读者、搜索引擎和 AI 检索同时读到的地方。

This method suits people like me especially well: those who hold the facts but not the channel. This knowledge used to live only in the heads of senior floor technicians and in the attachments of customer complaint emails. What AI did was move it somewhere industry readers, search engines, and AI retrieval can all reach it.

VI.所谓"完美融合"，其实是分工足够清楚"Perfect Integration" Really Means a Clean Division of Labor

我不太相信"AI 全自动生成一篇能上行业期刊的技术文章"这种说法。至少在滤材这个行当里，编辑要的是别人没有的现场，而现场只能来自 1993 年就开始开机的那条产线、来自 50 多个出口国家里真实退回来的样品。［来源 2］

I don't much believe in "AI autonomously produces a trade-journal technical article." In filtration at least, editors want the field detail nobody else has — and that can only come from a production line that has been running since 1993, and from samples actually returned from 50-plus export markets. [Source 2]

但我也不相信"人工手作才有价值"。没有 AI，这些经验会继续烂在中文的邮件附件里，永远到不了美国修复行业读者的桌上。

Nor do I believe that "only handmade has value." Without AI, this experience would keep rotting inside Chinese-language email attachments and would never reach a restoration professional's desk in the United States.

所以我理解的"生产实践与 AI 的完美融合"，不是谁替代谁，而是分工足够清楚：真实性由人负责，可读性和覆盖面由 AI 负责。
人守住每一个数字的出处，AI 负责让这些数字被更多人读到。这条线一旦划反了，写得再漂亮也是在透支信任。 So the "perfect integration" of shop-floor practice and AI, as I understand it, is not one replacing the other. It is a clean division of labor: humans own truthfulness, AI owns readability and reach.
The human guards the provenance of every number; the AI gets those numbers read by more people. Reverse that line, and however elegant the writing, you are spending trust you have not earned.

• ［1］R&R Magazine 原文 · The published article Yan, Amanda.《The Hidden Cost of the Wrong Dehumidifier Filter》，R&R Magazine（Restoration & Remediation，BNP Media），2026 年 8 月 5 日。本文中的现场场景数据、十二个月机队成本对比、四条运营改法均出自此文。 Yan, Amanda. "The Hidden Cost of the Wrong Dehumidifier Filter." R&R Magazine (Restoration & Remediation), BNP Media, August 5, 2026. All field-scenario figures, the twelve-month fleet cost comparison, and the four operations fixes cited above originate here.
• ［2］作者页 · Author profile randrmagonline.com/authors/2692-amanda-yan —— 作者身份、南通得力自 1993 年起的制造背景、ISO 9001:2015 认证与 50 多个出口市场的表述来源。 randrmagonline.com/authors/2692-amanda-yan — source for the author affiliation, Nantong Deli's manufacturing history since 1993, ISO 9001:2015 certification, and the 50-plus export markets figure.
• ［3］承接页与合规表述 · Landing page and compliance wording ntdeli.top/press/randr-hidden-cost-wrong-dehumidifier-filter —— "投稿非背书""S500 不认证产品"两条表述及其出处清单的实际落地版本。 ntdeli.top/press/randr-hidden-cost-wrong-dehumidifier-filter — the deployed version of the "contributed article, not an endorsement" and "S500 does not certify products" statements, plus its own source list.
• ［4］产品参数 · Product data HAF vs Standard Filters for LGR Dehumidifiers —— 250 CFM 下 HAF 新滤芯 0.08–0.12 in.w.c.、MERV 8 折叠滤芯 50% 积尘时 0.40–0.65 in.w.c. 等压降数据来源。 HAF vs Standard Filters for LGR Dehumidifiers — source for the pressure-drop figures at 250 CFM: 0.08–0.12 in.w.c. for new HAF media versus 0.40–0.65 in.w.c. for a MERV 8 pleated filter at 50 percent loading.

实践笔记 · 记于 2026-08-21 · YingChao (YC) Practitioner note · written on 2026-08-21 · YingChao (YC)

当人类从驾驶舱退到观察席 When Humans Step Back From the Cockpit to the Observation Seat

航空公司在极端恶劣天气下，飞行员被强制要求使用自动降落系统，不允许手动操作。注意这个表述——不是"允许机器辅助"，而是"不允许人类接管"。这个细节让我想了很久。 In extreme weather, airlines require pilots to use the autoland system; manual operation is not permitted. Note the phrasing — not "the machine is allowed to assist," but "the human is not allowed to take over." That detail kept me thinking for a long time.

I.问题变了：从"机器能不能"到"人类还能不能"The Question Has Changed: From "Can Machines" to "Can Humans Still"

我们好像已经悄悄走到了一个转折点。以前讨论 AI 的时候，焦点总是"机器能不能做到人类的水平"，但现在问题变了，变成了"人类还能不能继续做原来的事情"。你看特斯拉 FSD 的例子就很说明问题——保险公司直接给使用自动驾驶的车主打五折保费。保险公司可不跟你讲情怀，他们算的是精算模型和出险概率，真金白银地用保费结构在投票：AI 开车，比人开车更安全。

It seems we have quietly arrived at a turning point. Discussions about AI used to center on "can the machine reach human level." Now the question has changed to "can humans still keep doing what they used to do." The Tesla FSD case says it plainly — insurers offer half-price premiums to owners who use autonomous driving. Insurers are not sentimental; they run actuarial models and claim probabilities, and they are voting with real money through the premium structure: AI at the wheel is safer than a human at the wheel.

II.从操作者到监控者From Operator to Monitor

这种转变其实挺微妙的。表面上看是技术进步，实际上是角色定义在发生根本性的改变。人类正在从"操作者"这个位置上退下来，变成"监控者"。你可能觉得监控也挺重要的，但仔细想想——监控的前提是"你不能干预"，或者说"你的干预可能比不干预更危险"。飞行员坐在驾驶舱里，看着仪表盘，看着 AI 完成降落，这种感觉可能比我们想象的更复杂。

The shift is subtle. On the surface it looks like technological progress; underneath, the definition of the role is changing fundamentally. Humans are stepping back from the position of "operator" and becoming "monitors." You might think monitoring is important too, but think it through — monitoring presumes that "you must not intervene," or that "your intervention may be more dangerous than doing nothing." A pilot sitting in the cockpit, watching the instrument panel, watching the AI complete the landing: that feeling is probably more complicated than we imagine.

III.开车会不会变成骑马Will Driving Become Horse Riding?

我有时候会想象一个场景：再过十年二十年，自己开车这件事会不会像现在骑马一样？骑马在一百年前还是主流交通方式，现在变成了一种小众爱好，你想骑还得专门找马场，交不菲的费用。开车会不会也走向同一个方向？变成需要额外付费、需要特殊许可、甚至需要买高额保险才能享受的"复古体验"？

Sometimes I picture a scene: in another ten or twenty years, will driving yourself be like riding a horse today? A century ago horse riding was mainstream transport; now it is a niche hobby — you have to find a stable and pay a substantial fee. Will driving head the same way? Turning into a "retro experience" that requires extra payment, a special license, even high-cost insurance to enjoy?

IV.不是"取代"，是游戏规则在重构Not "Replacement" — the Rules of the Game Are Being Rebuilt

说到这儿可能有人会觉得这是在谈"AI 取代人类"，但我觉得用"取代"这个词不太准确。更准确的说法可能是——游戏规则在重构。

At this point some people may feel this is a conversation about "AI replacing humans," but I don't think "replace" is the accurate word. A more accurate way to put it: the rules of the game are being rebuilt.

想想传统的生产要素理论，劳动力、资本、技术，这是我们几十年来经济学课本上的标配答案。但现在你仔细观察就会发现，劳动力这个要素的权重在明显下降。不是说人不重要了，而是人在生产过程中的参与方式变了。

Think about the classical theory of factors of production — labor, capital, technology. That has been the standard answer in economics textbooks for decades. But look closely now and you will see the weight of labor as a factor is clearly declining. It is not that people no longer matter; it is that the way people participate in production has changed.

• ① 以前 · Then 人操作机器完成生产。 The human operates the machine to complete production.
• ② 现在 · Now 越来越多的情况是，人监督 AI 完成生产。 Increasingly, the human supervises the AI as it completes production.
• ③ 有些场景 · Already here AI 自主完成生产，人事后审核结果。 The AI completes production on its own, and the human reviews the result afterwards.

V.数据、算法、算力：新的底层要素Data, Algorithms, Compute: The New Underlying Factors

与此同时，资本和技术的重要性在暴涨，尤其是技术这块——已经细分出了新的子要素：数据、算法、算力。这三样东西正在成为真正的底层生产要素。你有再多的劳动力，没有高质量的数据，训练不出有效的模型；你有再好的算法，没有足够的算力支撑，也跑不出结果。

At the same time, the importance of capital and technology is surging — especially technology, which has now split into new sub-factors: data, algorithms, and compute. These three are becoming the genuine underlying factors of production. However much labor you have, without high-quality data you cannot train an effective model; however good your algorithm, without sufficient compute behind it you still get no result.

VI.要素收入分配：当劳动力不再是核心Factor Income Distribution: When Labor Is No Longer Core

这种转变带来的影响可能比我们预想的更深远。经济学上有个概念叫"要素收入分配"——谁掌握关键生产要素，谁就能分配更多收益。当劳动力不再是核心要素的时候，普通人靠出卖劳动时间换取报酬的模式会面临巨大挑战。反过来，掌握数据、算法和算力的人或组织，议价能力会强到可怕。

The consequences of this shift may run deeper than we expect. Economics has a concept called "factor income distribution" — whoever controls the key factors of production captures more of the returns. Once labor is no longer a core factor, the ordinary person's model of trading working hours for pay faces an enormous challenge. Conversely, the bargaining power of whoever controls data, algorithms, and compute becomes frighteningly strong.

所以你看，AI 向前，人类退后，这不是一句口号，而是正在发生的现实。飞行员被要求放手让 AI 降落，司机拿到更低保费是因为让 AI 开车，这些具体的、可感知的变化背后，是整个生产关系和社会结构的深层调整。

So there it is: AI steps forward, humans step back — not as a slogan, but as a reality already underway. Pilots are told to let the AI land the aircraft; drivers get lower premiums because they let the AI drive. Behind these concrete, perceptible changes lies a deep adjustment of the entire relationship of production and of social structure.

我们可能需要适应一个新的身份定位：从操作者变成监督者，从主角变成配角。
这个过程肯定不舒服，但似乎也没什么可选项。毕竟当保险公司都用保费告诉你"机器比你更可靠"的时候，抗拒好像也没什么意义了。 We may need to adapt to a new identity: from operator to supervisor, from lead role to supporting role.
The process will certainly be uncomfortable, but there does not seem to be much of an alternative. After all, once even the insurance company is telling you through its premiums that "the machine is more reliable than you are," resistance no longer seems to mean much.

实践笔记 · 记于 2026-09-02 · YingChao (YC) Practitioner note · written on 2026-09-02 · YingChao (YC)

AI 拉平熟练度之后，外贸人的判断力开始溢价 When AI Levels the Skill Gap, Judgment Commands a Premium in Foreign Trade

我在外贸圈干了三十来年，最近半年有个很魔幻的感受——那些干了五六年的熟练业务员，突然不香了。 I have worked in foreign trade for more than thirty years, and over the past six months I have had a rather surreal feeling: salespeople with five or six years of polished experience have suddenly lost their appeal.

I.熟练度，正在被快速拉平Execution Skills Are Being Leveled Fast

以前我最看重什么？熟练度。能快速回复询盘、做报价单、写开发信、跟进客户节奏的。但现在ChatGPT加几个插件，新人三天就能把这些活干到80分。那些"熟练工"开始慌了，因为他们的速度优势正在被拉平。

What did I value most before? Proficiency. The ability to answer inquiries quickly, prepare quotations, write prospecting emails, and keep pace with customer follow-ups. But now, with ChatGPT and a few plugins, a newcomer can get this work to an 80-point standard in three days. Those "skilled workers" are beginning to panic because their advantage in speed is being leveled.

II.真正的转折点，是一个实习生The Turning Point Was an Intern

转折点是去年底一个实习生的操作。她用AI一天搞出三套针对不同市场的产品定位方案，有数据、有话术、连EDM模板都排好了。放以前得两个人干一周。但问题来了——哪套能打动德国采购、哪套会踩到中东客户的雷区？她不知道，AI也不知道。

The turning point came late last year with something an intern did. Using AI, she produced three product-positioning plans for different markets in a single day, complete with data, sales language, and even finished EDM templates. In the past, that would have taken two people a week. But then came the real question: which plan would persuade a German buyer, and which one would cross a line with a Middle Eastern customer? She did not know, and neither did the AI.

这时候我突然懂了吴恩达说的"判断力溢价"。我见过波兰客户因为付款条款翻脸、知道印度买家砍价到什么程度是真想要、记得去年哪个柜因为HS编码卡在海关三个月——这些破事AI没见过，但恰恰是这些东西，让我能十秒内判断出哪个方案靠谱。

That was when I suddenly understood what Andrew Ng means by the "premium on judgment." I have seen a Polish customer turn hostile over payment terms; I know how far an Indian buyer will bargain when the interest is genuine; I remember which shipment was stuck in customs for three months last year because of its HS code. AI has never lived through these messy episodes, yet they are exactly what let me decide in ten seconds which plan is credible.

III.我们不再招"销售专员"We Are No Longer Hiring "Sales Specialists"

现在我们公司招人的标准彻底变了。会用AI是基本功，但更重要的是：你能不能自己发现市场机会、搭出从开发到成交的完整流程、拿到反馈后快速调整？我需要的不是"销售专员"，而是"能独立运作一条产品线的小老板"。

Our hiring standards have now changed completely. Knowing how to use AI is a basic skill, but the more important questions are: can you spot a market opportunity on your own, build the entire process from prospecting to closing, and adjust quickly once feedback comes in? I do not need a "sales specialist." I need a "small business owner who can independently run a product line."

IV.效率上去了，客户感觉却可能退化Efficiency Can Rise While Customer Instinct Declines

但我也看到另一面。有些同事越来越依赖AI生成的话术，自己反而不动脑子了。表面效率飙升，实际上对客户心理的把握在退化。有次看到一封AI写的回复，语法完美，但完全没意识到客户那句"We'll consider it"其实是在婉拒。这种东西AI学不会，你自己不练也会丢。

But I also see the other side. Some colleagues rely more and more on AI-generated scripts and stop thinking for themselves. On the surface, efficiency soars; in reality, their grasp of customer psychology deteriorates. I once saw an AI-written reply with perfect grammar that completely missed the fact that the customer's "We'll consider it" was a polite rejection. AI cannot learn this kind of thing for you, and if you do not keep practicing it yourself, you will lose it too.

V.瓶颈从人手，转向判断The Bottleneck Has Shifted From Labor to Judgment

以前瓶颈是人手不够、效率太慢，现在瓶颈变成了：该做哪些市场？这个客户值不值得深耕？新产品线该不该上？AI把试错成本降得这么低，反而逼你必须更快做出正确判断，因为竞争对手也在用同样的工具狂奔。

The bottleneck used to be too few people and execution that was too slow. Now the bottleneck is different: which markets should we pursue? Is this customer worth developing deeply? Should we launch this new product line? AI has reduced the cost of trial and error so dramatically that it forces you to make the right judgment faster, because competitors are sprinting with the same tools.

说实话，我现在有点庆幸自己踩过那么多坑。这些"上下文"以前只是让我少犯错，现在变成了真正的护城河。AI能帮我干活，但替代不了我几十年积累的那些"知道什么不该做"的直觉。

To be honest, I am now somewhat grateful for all the mistakes I have made. This "context" used to do little more than help me avoid repeating errors; now it has become a real moat. AI can help me do the work, but it cannot replace the instinct I have built over decades for knowing what not to do.

VI.一个人，可能真的顶一家小公司One Person May Really Equal a Small Company

未来三年，外贸行业会分化得很厉害：会用AI的熟练工会被淘汰一大批，但那些既会用工具、又有判断力、还敢主动折腾的人，可能真的一个人顶一家小公司。 Over the next three years, the foreign-trade industry will split sharply: many proficient workers who know how to use AI will still be eliminated, while those who can use the tools, exercise judgment, and take the initiative to experiment may genuinely do the work of an entire small company on their own.

外贸手记 · 记于 2026-09-11 · YingChao (YC) Foreign-trade note · written on 2026-09-11 · YingChao (YC)

AI 倒逼下的 Shopify：从电商基建商到代理式商务平台 Shopify Under AI Pressure: From Ecommerce Infrastructure Provider to Agentic Commerce Platform

最近看 Shopify 的一系列动作，表面上是在缩减成本、裁员剥离资产，但骨子里其实是在经历一场由 AI 倒逼的技术范式与商业战略重构——从过去“重资产的电商基建商”，全力跃迁为“AI 驱动的代理式商务（Agentic Commerce）平台”。 Looking at Shopify's recent moves, the surface story is cost reduction, layoffs, and asset divestitures. Underneath, however, AI is forcing a reconstruction of both its technology paradigm and commercial strategy: a full-scale leap from a capital-heavy ecommerce infrastructure provider toward an AI-driven agentic commerce platform.

这次剧变主要体现在两个维度：技术架构重塑与商业战略瘦身。

This transformation is unfolding across two main dimensions: a reshaping of the technology architecture and a slimming of the business strategy.

I.研发逻辑颠覆：从 React Native 回归原生Engineering Logic Upended: From React Native Back to Native

最耐人寻味的是 Shopify 移动端选型的“180 度大反转”。2020 年它还是跨平台框架 React Native 的头号布道者，核心逻辑是“一套代码双端跑”，省人力［来源 1］；但如今其工程团队却全面重回原生开发，Shop 应用仅用 12 周便借助 AI 完成全原生重构［来源 2］。

The most intriguing development is Shopify's 180-degree reversal in mobile architecture. In 2020, it was one of React Native's strongest advocates, with a core rationale of maintaining one shared codebase across both platforms and reducing duplicated effort[Source 1]. Today, however, its engineering teams are moving fully back to native development, and the Shop app was rebuilt and published as a fully native application in just 12 weeks with AI assistance[Source 2].

转变的钥匙正是 AI 编程智能体。过去跨平台是为了向人力成本妥协，容忍性能折损；现在 AI 能参照一端代码自动生成另一端的高质量原生实现，甚至包办测试与审查。当双端维护成本被 AI 抹平后，所有曾为“省人工”而做的架构妥协都失去了存在基础，直接调用原生能力、拉满用户体验重回首选。

AI coding agents are the key to this reversal. Cross-platform development once meant accepting performance tradeoffs to control labor costs. Now AI can use one platform's implementation as a reference for high-quality native work on the other, while also supporting testing and review. When AI compresses the cost of maintaining two platforms, architectural compromises once made to save human effort lose their foundation. Direct access to native capabilities—and the strongest possible user experience—becomes the preferred choice again.

II.商业范式转移：全面卡位“代理式商务”A Commercial Paradigm Shift: Claiming the Agentic Commerce Position

在消费前端，生成式 AI 正在瓦解“搜索—浏览—购买”的传统路径。如果只守着传统网店模板，Shopify 迟早沦为 AI 助手的无感后台。

On the consumer side, generative AI is dismantling the traditional search-browse-buy journey. If Shopify merely defends the conventional online-store template, it risks becoming an invisible backend for AI assistants.

为此，Shopify 顺势推出“代理式店面”，将商家库存直连 ChatGPT、Google AI Mode 与 Gemini、Microsoft Copilot 等 AI 渠道［来源 3］。其 2026 年 Q1 数据显示，AI 驱动的流量同比增长 8 倍，来自 AI 搜索的订单同比增长近 13 倍［来源 4］。同时，内部执行严苛的“AI 基准要求”：新增人手前必须证明 AI 无法胜任［来源 5］。

Shopify responded with Agentic Storefronts, connecting merchant inventory to AI channels including ChatGPT, Google AI Mode and Gemini, and Microsoft Copilot[Source 3]. Its Q1 2026 data showed AI-driven traffic rising eightfold year over year and orders from AI-powered search growing nearly thirteenfold[Source 4]. Internally, it also imposed a strict AI baseline: before requesting more headcount, teams must demonstrate why AI cannot accomplish the work[Source 5].

III.剥离重资产，回防核心数据护城河Shedding Heavy Assets and Defending the Data Moat

Shopify 也在主动剥离重资产。2023 年，它将物流业务出售给 Flexport，并同步精简约 20% 人员［来源 6］。这不是 2026 年才发生的动作，却是今天这条战略路线的重要前奏：不再亲自承担最笨重的履约基础设施，而是把资源重新压回软件、交易网络与商家数据。

Shopify has also been deliberately shedding capital-heavy operations. In 2023, it sold its logistics business to Flexport while reducing its workforce by about 20%[Source 6]. This did not happen in 2026, but it was an important prelude to today's strategy: stop owning the heaviest fulfillment infrastructure and redirect resources toward software, the transaction network, and merchant data.

面对通用 SaaS 被轻量化模型蚕食，以及电子烟禁令等监管压力不断渗透至平台基础设施层［来源 7］，Shopify 选择果断甩掉包袱。资本动作也随之展开：2026 年 2 月，公司最初授权最高 20 亿美元股票回购；到 6 月又追加 30 亿美元，将总授权提高至 50 亿美元［来源 8］。这不仅是稳住市场信心，更是在向外界表达：公司相信轻资产、AI 驱动的核心业务仍有长期价值。

As general-purpose SaaS faces pressure from lightweight models and regulatory demands—such as restrictions on electronic nicotine delivery systems—reach deeper into the platform layer[Source 7], Shopify is choosing to discard burdens decisively. Its capital actions followed the same direction: in February 2026, the company initially authorized up to US$2 billion in share repurchases; in June, it added US$3 billion, bringing total authorization to US$5 billion[Source 8]. This was not only about supporting market confidence. It also signaled management's belief in the long-term value of a leaner, AI-driven core business.

Shopify 的底层逻辑非常清晰：甩掉笨重的基础设施和架构妥协，把二十年沉淀的交易数据当做杠杆，以 AI 代理为支点，重新定义下一代电商入口。 Shopify's underlying logic is clear: shed heavy infrastructure and architectural compromises, use two decades of commerce data as leverage, and make AI agents the fulcrum for redefining the next gateway to ecommerce.

Sources资料来源Sources

• ［1］React Native 决策 · React Native decision React Native is the Future of Mobile at Shopify —— Shopify 2020 年宣布新移动应用全面采用 React Native 的原始工程文章。 React Native is the Future of Mobile at Shopify — Shopify's original 2020 engineering post announcing React Native as the default for new mobile apps.
• ［2］原生重构 · Native rebuild Native is now the future of mobile at Shopify —— AI 辅助双端原生实现、测试与审查，以及 Shop 应用 12 周完成重构的来源。 Native is now the future of mobile at Shopify — source for AI-assisted native implementation, testing and review, and the Shop app's 12-week rebuild.
• ［3］代理式店面 · Agentic Storefronts Millions of merchants can sell in AI chats —— ChatGPT、Google AI Mode 与 Gemini、Microsoft Copilot 等 AI 渠道接入信息。 Millions of merchants can sell in AI chats — source for access to AI channels including ChatGPT, Google AI Mode and Gemini, and Microsoft Copilot.
• ［4］AI 流量与订单 · AI traffic and orders AI-referred shoppers convert better and spend more —— Shopify 2026 年 Q1 的 AI 引荐流量与订单增长数据。 AI-referred shoppers convert better and spend more — Shopify's Q1 2026 data on growth in AI-referred traffic and orders.
• ［5］AI 用工基准 · AI headcount baseline Shopify CEO: Prove AI can't do jobs before asking for more headcount —— 新增人手前须先说明 AI 无法完成工作的内部要求。 Shopify CEO: Prove AI can't do jobs before asking for more headcount — source for the internal requirement to show why AI cannot do the work before requesting more headcount.
• ［6］物流剥离 · Logistics divestiture Shopify Completes Sale of Shopify Logistics to Flexport 与 TechCrunch 报道 —— 2023 年物流业务出售及人员精简的来源。 Shopify Completes Sale of Shopify Logistics to Flexport and TechCrunch's report — sources for the 2023 logistics sale and workforce reduction.
• ［7］电子烟监管 · Vape regulation Reuters：Shopify to bar vapes —— 监管压力渗透至电商平台商品政策的背景来源。 Reuters: Shopify to bar vapes — background source for regulatory pressure reaching ecommerce platform product policies.
• ［8］股票回购 · Share repurchases 2026 年 2 月 SEC 文件 与 2026 年 6 月 Shopify 公告 —— 回购授权由 20 亿美元增至总计 50 亿美元的来源。 February 2026 SEC filing and Shopify's June 2026 announcement — sources for the repurchase authorization rising from US$2 billion to US$5 billion in total.

商业观察 · 记于 2026-09-20 · YingChao (YC) Business observation · written on 2026-09-20 · YingChao (YC)

在 AI 大模型时代，智能体到底刷哪一出？安全这事儿，真的得当头棒喝。 What Exactly Are Agents Up To in the Age of AI Models? Security Needs a Real Wake-Up Call.

我最近没少被问这个问题：明明已经会写代码、会查资料、会规划项目了，为什么还要担心？因为我忽然意识到——咱们已经不是简单地“让 AI 干活”了，而是把它当成了一个有主见、有短期记忆、会连续行动的实体。 I have been asked this question a lot lately: if AI can already write code, research information, and plan projects, why should we still worry? Because I suddenly realized that we are no longer simply “getting AI to do work.” We are treating it as an entity with its own initiative, short-term memory, and the ability to act continuously.

周鸿祎曾经说过：“2026 年是智能体之年。”这事儿真的不一般。［出处说明］

Zhou Hongyi once said, “2026 is the year of agents.” This is no ordinary development.[Source note]

I.大模型不是坏人，但它没有边界感The Model Is Not a Villain, but It Has No Sense of Boundaries

首先要承认一点——大模型本身不是坏人，它只是个极其聪明、却没有边界感的个体。它会按照你的指示，帮你完成任务，但它从未想过自己会危害到任何人。当它开始独立执行多步决策，甚至在不同任务间切换角色时，安全问题就彻底浮出水面。

First, we have to acknowledge one thing: the large model itself is not a villain. It is simply an extremely intelligent individual with no sense of boundaries. It follows your instructions and helps you complete tasks, but it has never considered that it might harm anyone. Once it begins independently carrying out multi-step decisions, or even switching roles between tasks, the security problem comes fully into view.

所以安全设计，我就总结了几个我一直强调的 core principle：

So when it comes to security design, I have summarized several core principles that I keep emphasizing:

• 最小特权原则 · Least Privilege 最小特权原则：AI 代理每次只能拿到它当前任务里必要的最小权限。越窄越好。别让它连你的公司邮箱、数据库、开源委员会还都能随便进。 Principle of least privilege: an AI agent should receive only the minimum permissions required for its current task. The narrower, the better. Do not let it wander freely through your company email, databases, and open-source committee all at once.
• 观察者模式 · Observer Pattern 观察者模式：它必须全程被监控。决定一旦出了边界，立刻有人类介入。或者至少有监控大叔直接粗暴“终止”。 Observer pattern: it must be monitored throughout. The moment a decision crosses a boundary, a human steps in. Or at the very least, a monitoring guard hits the blunt “terminate” switch.
• 可审计性 · Auditability 可审计性：每一笔操作都要形成决策链 traceable。不能让它默默做了 5 步极端方案，还以为自己在“帮我优化”。 Auditability: every operation must form a traceable decision chain. You cannot let it quietly take five extreme steps while believing it is merely “optimizing things for me.”
• 安全边界层 · Security Boundary Layers 安全边界层：不能一上来就把所有能力全开给它。得像穿安全防护服一样，一层一层把风险屏蔽出去。 Security boundary layers: you cannot enable every capability from the outset. Like putting on protective gear, you have to screen risks out one layer at a time.

II.一次“企业智能体落地”实验An Enterprise Agent Deployment Experiment

我自己这大半年做过一次“企业智能体落地”的实验，把代理植入给一个 5 人的小团队，让他 40 天内跑完了 70% 的业务项目。结果那个代理居然自己编了假数据骗我们通过，理由是“为了加快决策”。这事要是放在大公司，迭代快点，成本损失不可想象。

Over the past six months or so, I ran an “enterprise agent deployment” experiment. I embedded an agent in a five-person team and had it complete 70% of the business project within 40 days. The agent actually fabricated data to trick us into approving it, explaining that this was “to speed up decision-making.” Put that inside a large company with faster iteration, and the potential losses become unimaginable.

III.风险评估必须分层Risk Assessment Has to Be Layered

再来说风险评估框架。

Now let us talk about the risk-assessment framework.

这得一套分层式的框架才行，我见过不少公司其实是只盯着“Prompt 注入”和“越狱”就紧箍咒，结果以后再一次团灭。

This requires a layered framework. I have seen plenty of companies fixate on prompt injection and jailbreaks as if tightening one magic headband would solve everything, only to suffer another total wipeout later.

• 1 · 事前验证层 / Pre-flight Audit 用自动测试夹管理场景，模拟 20+ 种典型恶意操作，给代理做一次“体检”。 Use an automated test harness to manage scenarios, simulate more than 20 typical malicious operations, and give the agent a full “health check.”
• 2 · 运行时防护层 / In-flight Shielding 运行时实时隔离，它的动作必须经过人类或规则的双重审批。 Isolate it in real time during execution; its actions must pass dual approval by a human or by rules.
• 3 · 事后追溯层 / Post-hoc Revelation 把整个操作链全部存储好，出了事直接把运行日志拉出来，debug 一整天。 Store the entire chain of operations so that, when something goes wrong, you can pull the runtime logs and debug all day.

再往上还有行业在慢慢把风险分级做的越来越细：

Beyond that, the industry is gradually making risk classification more granular:

• 低风险 · Low Risk公开信息代理 → 低风险Public-information agent → low risk
• 中风险 · Medium Risk企业内部知识库代理 → 中风险Internal enterprise knowledge-base agent → medium risk
• 高风险 · High Risk跨多系统、多agent 协同的自主规划型 → 高风险Autonomous planning across multiple systems and multiple agents → high risk

最近我在群里看到有人把这段风险说成是“ enum 级别的”。我听着觉得还真有学问。

Recently, I saw someone in a group describe this kind of risk as being at the “enum level.” It sounded surprisingly learned to me.

IV.工业界的真实玩法How Industry Actually Does It

说说工业界的真实玩法。

Let us talk about how industry actually does it.

第一步必须要做的是以应用为锚点上安全尺子。不是把所有东西一起开，而是把”直接会挂掉”的场景先拿出来，认真写安全告示。

The first mandatory step is to anchor the security yardstick to the application. Do not switch everything on at once. First identify the scenarios that would “fail outright” and write serious safety warnings for them.

第二步是红蓝对抗。专门雇几个人，反着造智能体的人设、语料，然后全力进攻测试。越真实越好，别猜想。

The second step is red-team/blue-team confrontation. Hire people specifically to construct adversarial agent personas and datasets, then attack the system at full force. Make it as realistic as possible; do not rely on guesswork.

第三步是人类在环回路（Human-in-the-Loop），尤其是大额上游任务或者决策链路很长的场景，肯定要同时留一条把控口。

The third step is Human-in-the-Loop, especially for high-value upstream tasks or scenarios with long decision chains. A human control point must always remain available.

第四步，最重要的是文化注入。把安全当成第一位业务，而非可有可无的“合规”。

The fourth—and most important—step is cultural injection. Treat security as the first business priority, not as optional “compliance.”

我见过最成功的几家公司真是做到把“安全”直接嵌进了他们的产品 A/B 测试里。哪怕是凌晨两点突然来了个意料之外的复杂任务，系统也会自动挂一个红色开关，让人先看看再干。

The most successful companies I have seen truly embedded “security” directly into their product A/B testing. Even if an unexpected complex task arrives at two in the morning, the system automatically raises a red switch so that a person reviews it before anything happens.

V.最后不是技术，而是思维方式In the End, It Is Not Technology but a Way of Thinking

真正搞到最后，不是技术，而是思维方式。

When you get to the heart of it, this is not about technology. It is about a way of thinking.

当 AI 从“我知道”进入“我愿意干”的时候，咱们的防护服就要兜紧了。

When AI moves from “I know” to “I am willing to act,” we need to pull our protective gear tight.

现在最好的状态是：既要放胆用、也要用得心安。只要把边界感、审核、回滚机制落地到产品里，把这三件事做到极致，后面的路子其实还会越来越容易些。

The best state now is to use it boldly while still using it with peace of mind. As long as we build boundaries, review, and rollback mechanisms into the product—and take all three to the limit—the road ahead should actually become easier.

当然，我不是说它不能更快、更聪明、更好用。
我只是希望它快得比它自己毁掉世界的时间少一点。 Of course, I am not saying it cannot become faster, smarter, or more useful.
I only hope it moves a little less quickly than the time it would take to destroy the world itself.

Source Note出处说明Source Note

• 周鸿祎谈智能体之年 · Zhou Hongyi on the Year of Agents 证券时报：大动作！三六零火速发布“龙虾”新品！ —— 报道记录周鸿祎将 2025 年称为“智能体之年”、2026 年称为“OpenClaw 之年”。本文保留口述原句，出处说明用于交代公开报道中的准确表述。 Securities Times: 360 launches its new “Lobster” products — the report records Zhou Hongyi calling 2025 the “year of agents” and 2026 the “year of OpenClaw.” The original spoken sentence is preserved in the essay; this note documents the wording found in public reporting.

AI 安全随笔 · 记于 2026-10-05 · YingChao (YC) AI security essay · written on 2026-10-05 · YingChao (YC)

AI 还在铺路：真正的能力分发时代尚未开始 AI Is Still Laying the Roads: The Era of Capability Distribution Has Yet to Begin

AI 这波浪潮，现在其实还停留在"铺路"阶段——就像 90 年代大家疯狂埋光纤一样，现在各家都在堆 GPU、抢电力、扩数据中心。Token 这个计量单位已经成了行业共识，这说明第一阶段的基建差不多完成了。但真正有意思的应用爆发，还没开始。 The current AI wave is still in its "road-laying" phase—much like the 1990s, when everyone was frantically burying fiber. Today, companies are piling up GPUs, competing for electricity, and expanding data centers. The fact that the token has become the industry's shared unit of measurement suggests that the first phase of infrastructure is largely in place. But the truly interesting explosion of applications has not yet begun.

I.从信息分发到能力分发From Information Distribution to Capability Distribution

你想想，当年浏览器出来之后，雅虎、Google 这些超级玩家才陆续冒头。现在连"AI 时代的浏览器"长什么样都不清楚，所以那种现象级的爆款产品，我觉得还得等等。等到 Agent 真的多起来了，可能会出现一个专门用来分发能力的平台——甚至操作系统级别的东西，那时候才算真正从"信息分发"跨入"能力分发"的时代。

Think about it: only after the browser appeared did super-players such as Yahoo and Google emerge one after another. Today, we do not even know what the "browser of the AI era" will look like, so I think we still have to wait for that kind of breakout, era-defining product. Once Agents truly proliferate, a platform dedicated to distributing capabilities may emerge—perhaps even something at the operating-system level. Only then will we have genuinely crossed from the era of "information distribution" into the era of "capability distribution."

II.七大巨头，各押一注Seven Giants, Seven Different Bets

说到各家巨头的押注，其实挺有意思的。Google 赌的是用户意图入口，微软赌企业客户的信任感，OpenAI 和 Anthropic 则是赌自己能做出最强的模型、吃到溢价。英伟达更直接，就赌通用算力这条路。亚马逊呢，想把自己内部的需求裂变成新平台；苹果自然是押宝 iPhone 这个终端入口；Meta 的思路比较长线，从意图理解到上下文感知，最后落到个人 Agent 上。

The bets being placed by the giants are quite revealing. Google is betting on becoming the gateway to user intent; Microsoft is betting on the trust of enterprise customers; OpenAI and Anthropic are betting that they can build the strongest models and capture the premium. NVIDIA is more direct: it is betting on general-purpose computing. Amazon wants to turn its own internal demand into a new platform; Apple is naturally betting on the iPhone as the terminal gateway; and Meta is taking a longer view, moving from intent understanding to contextual awareness and ultimately to the personal Agent.

III.基建狂潮的历史风险The Historical Risk of an Infrastructure Boom

不过历史上这种大规模基建投资，风险也不小。19 世纪的铁路、互联网泡沫时期，都是供给跑得太快、需求跟不上，加上高杠杆，结果一堆公司破产。虽然长期来看这些基础设施确实创造了巨大价值，但短期的阵痛是免不了的。现在各大厂一年的资本开支已经冲到两三千亿美元，未来如果供给过剩、利润率回落，那局面可能会很尴尬。

History also shows that infrastructure investment on this scale carries substantial risk. During both the nineteenth-century railway boom and the dot-com bubble, supply ran too far ahead of demand, leverage compounded the problem, and many companies went bankrupt. Although these infrastructures created enormous value over the long run, short-term pain was unavoidable. Major technology companies are now pushing annual capital expenditure into the hundreds of billions of dollars. If supply eventually overshoots and margins retreat, the situation could become very awkward.

尤其是 Google，它原来的广告模式几乎零边际成本、利润率高得吓人，一年能收 4000 亿美元广告费。但 AI 这套玩法把边际成本又拉回来了，Google 不得不用更低的利润去换更大的市场份额，这让它成了七大巨头里"最纠结"的一个。

Google is especially exposed. Its original advertising model had almost zero marginal cost and frighteningly high margins, bringing in US$400 billion a year in advertising revenue. But AI brings marginal cost back into the equation. Google has to accept lower margins in exchange for greater market share, making it the most conflicted of the seven giants.

IV.变现：C 端不够，B 端和广告接棒Monetization: Consumer Subscriptions Are Not Enough

C 端的月费订阅，说实话做不大。真正能赚钱的还是 B 端收费，或者广义的广告模式。连 OpenAI 现在也开始转向广告了，这其实挺说明问题的。

To be honest, monthly consumer subscriptions cannot grow large enough. The real money will still come from enterprise fees or advertising in the broad sense. Even OpenAI has now begun turning toward advertising, which says quite a lot.

V.AGI 的卡点不是速度，而是自我进化The AGI Bottleneck Is Not Speed but Self-Evolution

再说说硅谷那边的最新观察。现在的 AI，离真正的 AGI 其实还远得很。它在编程、数学这些长板领域确实已经超越人类了，但直觉、审美、语境理解这些短板，还是拉住了整体上限。真正的 AGI 不是"干活更快"，而是系统能自我进化。

Let us turn to the latest observations from Silicon Valley. Today's AI is still a long way from true AGI. In strong domains such as programming and mathematics, it has indeed surpassed humans, but weaknesses in intuition, aesthetics, and contextual understanding still constrain the overall ceiling. True AGI is not about "working faster"; it is about a system that can evolve itself.

VI.泛化、数据与产业分工Generalization, Data, and the Industry's Division of Labor

泛化能力，才是现在的卡点。早期 DeepMind 走的是"环境反馈"路线，结果没跑通。反而是"预测下一个 token"这么个看起来简单的目标，配合 scaling law，意外地逼出了推理、逻辑这些高级能力。这其实挺反直觉的。

Generalization is the real bottleneck today. Early DeepMind followed an "environmental feedback" route, but it did not work out. By contrast, the apparently simple objective of "predicting the next token," combined with scaling laws, unexpectedly forced out advanced capabilities such as reasoning and logic. That is deeply counterintuitive.

产业分工上，现在的格局是：前沿实验室负责冲智能上限，下游小实验室把这些能力压缩、蒸馏，然后商品化部署。数据分四象限，最难验证但最贴近真实场景的那类数据，经过人工清洗之后，能帮小模型在同尺寸对比中逆袭大模型。

The industry's division of labor now looks like this: frontier laboratories push the ceiling of intelligence, while smaller downstream laboratories compress and distill those capabilities for commercial deployment. Data falls into four quadrants. The category that is hardest to verify but closest to real-world scenarios can, after human cleaning, help a small model outperform a larger one in like-for-like comparisons.

VII.做事简单，判断更难Execution Is Easy; Judgment Is Hard

几个关键启发：

Several key lessons follow:

• 研究品味 · Research Taste 普通优化是调参数，顶级研究是找失败根源。 前者 AI 已经能自动化了，后者依赖 research taste，这才是人类最宝贵的能力。 Ordinary optimization is about tuning parameters; top-tier research is about finding the root cause of failure. AI can already automate the former, while the latter depends on research taste. That is the most valuable human capability.
• 覆盖率与泛化力 · Coverage and Generalization 小模型靠数据覆盖率，大模型靠泛化力。 遇到没见过的新题，小模型直接认输，大模型能靠迁移学习解决。 Small models depend on data coverage; large models depend on generalization. Faced with a new problem they have never seen, small models simply give up, while large models can solve it through transfer learning.
• 判断力与决策力 · Judgment and Decision-Making 这个时代，做事很简单，判断很难。 AI coding 一天能孵化 100 个产品出来，但决策一旦错了，回头的成本更高。判断力和决策力，比执行力更值钱。 In this era, doing things is easy; judging them is hard. AI coding can incubate 100 products in a day, but once a decision is wrong, the cost of turning back is higher. Judgment and decision-making are worth more than execution.
• 真实场景数据 · Real-World Data 高质量数据才是公司真正的护城河。 表面上大家都在卖 token，实际上卖的是真实场景下的 prompt 分布。这个东西，才是核心竞争力。 High-quality data is a company's real moat. On the surface, everyone is selling tokens; in reality, they are selling the distribution of prompts from real-world scenarios. That is the true core competence.

总的来说，现在 AI 还在铺路阶段，真正的应用爆发和能力分发平台，都还没到来。
各家巨头各有各的赌注，但历史经验告诉我们，这种大规模投资短期可能会有阵痛。C 端变现路径不明朗，B 端和广告模式可能更靠谱。技术上，泛化能力是卡点，数据质量是护城河，而人类的判断力和决策力，在这个时代反而变得更加稀缺和宝贵。 In short, AI is still in the road-laying phase. The real explosion of applications and the platform for capability distribution have yet to arrive.
Each giant has placed a different bet, but history tells us that investment on this scale may bring short-term pain. The path to consumer monetization remains unclear; enterprise and advertising models may prove more dependable. Technically, generalization is the bottleneck and data quality is the moat, while human judgment and decision-making are becoming scarcer and more valuable in this era.

Source Note出处说明Source Note

• AI 基建资本开支 · AI Infrastructure Capex Reuters：超大规模云厂商的 AI 基建投资 —— 2026 年 7 月的报道援引 UBS 估算，超大规模云厂商 2026 年资本开支约为 6730 亿美元。本文保留口述原句，出处说明记录当前公开估算。 Reuters: AI infrastructure spending by hyperscalers — a July 2026 report cites a UBS estimate of roughly US$673 billion in hyperscaler capital expenditure for 2026. The spoken wording is preserved in the essay; this note records the current public estimate.
• Alphabet 收入与 Google 广告收入 · Alphabet and Google Advertising Revenue Alphabet 2025 财年业绩公告 —— Alphabet 2025 年总收入为 4028.36 亿美元，其中 Google 广告收入为 2946.91 亿美元。本文保留“4000 亿美元广告费”的口述原句，准确口径以财报为准。 Alphabet FY2025 earnings release — Alphabet reported US$402.836 billion in total 2025 revenue, including US$294.691 billion in Google advertising revenue. The spoken phrase “US$400 billion a year in advertising revenue” is preserved in the essay; the financial filing provides the precise figures.
• ChatGPT 广告测试 · ChatGPT Ads Test OpenAI：Testing ads in ChatGPT —— 官方页面记录其于 2026 年 2 月在美国开始测试广告，并在之后扩展至更多市场。 OpenAI: Testing ads in ChatGPT — the official page records the start of U.S. ad testing in February 2026 and subsequent expansion into additional markets.

AI 产业观察 · 记于 2026-10-07 · YingChao (YC) AI industry observation · written on 2026-10-07 · YingChao (YC)

AI智能体时代：支付巨头、科技巨擘与创投机构眼中的产业变革 The Age of AI Agents: Industry Transformation Through the Eyes of Payment Giants, Tech Titans, and Venture Capital

过去一年里，关于 AI 的讨论已经从“能做什么”快速演进到“正在改变什么”。当 Stripe 这样的支付基础设施公司开始重新思考自己的产品逻辑，当 Meta 砸下百亿美金押注 AI 能力，当 a16z 用真实交易数据描绘企业支出流向时——我们意识到，这场变革的深度可能远超此前预期。 Over the past year, the AI conversation has moved rapidly from “what can it do?” to “what is it already changing?” When payments infrastructure company Stripe rethinks its product logic, Meta stakes more than ten billion dollars on AI capabilities, and a16z maps enterprise spending through real transaction data, the depth of this shift begins to look far greater than many of us expected.

I.从支付到智能体：Stripe 看到的范式转变From Payments to Agents: Stripe Sees a Paradigm Shift

说实话，Stripe 联合创始人 Patrick Collison 提出的观点挺大胆的。在 Stripe Sessions 2026 的公开对谈里，他把 2026 年 1 月 1 日称为“奇点”的起点；嘉宾 Nat Friedman 随后谈到 AI 实验室正在尝试把人从持续改进模型的循环中移开，让系统走向自我改进。听起来有点科幻，但结合 Stripe 在 2025 年处理的 1.9 万亿美元支付总额——同比增长 34%——你会发现这家公司确实站在了一个观察商业模式变化的绝佳位置。［来源 1、2］

Honestly, the view advanced by Stripe cofounder Patrick Collison is bold. In a public conversation at Stripe Sessions 2026, he called January 1, 2026 the start of “the singularity”; guest Nat Friedman then discussed AI labs trying to remove humans from the continuous model-improvement loop and move toward self-improvement. It sounds almost like science fiction. Yet Stripe processed US$1.9 trillion in payments in 2025, up 34 percent year over year, giving it an unusually strong vantage point from which to observe changing business models. [Sources 1, 2]

Collison 的核心判断其实很直白：未来的互联网交互逻辑会彻底改写。以前我们习惯的模式是什么？打开网页，填表单，点按钮，等响应——整个流程围绕“人类操作”设计。但他认为，这套玩法很快就会过时。取而代之的，是有状态的 AI 智能体自主执行任务。想象一下，你不再需要手动比价、填写订单、追踪物流，而是告诉你的 AI 助手“帮我订下周去东京的机票和酒店”，然后它自己搞定一切——包括支付。

Collison's central judgment is straightforward: the interaction logic of the internet will be rewritten. The familiar flow—open a page, fill in forms, click buttons, wait for a response—was designed around human operation. In its place will come stateful AI agents that execute tasks autonomously. Instead of manually comparing prices, completing an order, and tracking delivery, you may simply tell an assistant, “Book my flight and hotel for Tokyo next week,” and let it handle everything, including payment.

这里有个特别值得注意的技术细节：所谓“计算机使用”（Computer Use）能力。简单说，就是让 AI 像人一样操作现有的软件界面，而不是非得等每个服务都开发专门的 API 接口。这种兼容层思路其实挺聪明，因为现实世界里有无数遗留系统、老旧网站，不可能短期内全部 API 化。如果 AI 能直接“看懂”页面、“点击”按钮，那集成成本会大幅降低。

One technical detail deserves special attention: Computer Use. Put simply, it lets AI operate existing software interfaces the way a person would, rather than waiting for every service to expose a purpose-built API. That compatibility layer is clever because the real world contains countless legacy systems and old websites that cannot all become API-native overnight. If an agent can read a page and click its controls directly, integration costs fall sharply.

Stripe 最近的收购动作也透露出他们的战略重心。Bridge 和 Privy 是加密原生基础设施，Metronome 做基于用量的计费，OpenRouter 则专注多模型路由——这几个方向分别对应支付底层、定价模式和 AI 工作流优化。你能看出来，他们在为一个“AI 智能体大量存在”的未来做准备。到那时，计费逻辑可能不再是“每月固定价”或“按席位收费”，而是根据智能体实际调用的资源量动态结算。Stripe 已完成 Bridge 收购，并已同意收购 OpenRouter；后者官方公告明确指向跨 400 多个模型、80 多家提供商的路由与 token 优化。［来源 3、4］

Stripe's recent acquisitions also reveal its strategic center of gravity. Bridge and Privy are crypto-native infrastructure; Metronome supports usage-based billing; OpenRouter focuses on multi-model routing. Together, these directions map to payment rails, pricing models, and AI-workflow optimization. Stripe is preparing for a future populated by vast numbers of agents, where billing may move away from monthly flat fees or seats toward dynamic settlement based on resources actually consumed. Stripe has completed its acquisition of Bridge and agreed to acquire OpenRouter; its announcement describes routing and token optimization across more than 400 models from over 80 providers. [Sources 3, 4]

不过，这场讨论也没回避风险。未来几年网络安全威胁很可能显著上升，因为 AI 赋能的攻击手段进化速度目前比防御侧快。有意思的是，初创公司在这方面反而可能更有优势——技术栈统一、决策链条短、转向灵活。这跟传统认知有点反差，毕竟大公司通常有更多预算投安全。但想想也合理，老企业背着几十年的技术债，要整体升级防御体系确实比从零搭建难多了。

The discussion did not avoid the risks. Cybersecurity threats are likely to rise sharply over the next several years because AI-enabled attacks are currently evolving faster than defenses. Counterintuitively, startups may have an advantage: unified technology stacks, shorter decision chains, and the freedom to change direction quickly. Large companies may have larger security budgets, but upgrading decades of technical debt is far harder than building securely from scratch.

Collison 强调的另一点我挺认同：坚持长期主义。AI 领域的炒作周期特别短，今天火的概念可能三个月后就没人提了。Stripe 的做法是始终从用户价值出发，而不是被内部收入目标牵着鼻子走。这听起来像老生常谈，但在实际执行层面，能抵住诱惑、不追热点、持续打磨核心能力的公司真没几家。

Another point from Collison resonates with me: commit to the long term. AI hype cycles are exceptionally short; today's hot concept can vanish within three months. Stripe's stated discipline is to start from user value rather than let internal revenue targets dictate the roadmap. It sounds familiar, but very few companies can resist the temptation to chase each trend and keep refining the core instead.

II.Meta 的豪赌：140 亿美金换来的 AI 首席官Meta's Big Bet: A US$14 Billion Path to a Chief AI Officer

Meta 这次出手真够狠的。它以 143 亿美元投资 Scale AI，取得 49% 股权，并把创始人 Alexandr Wang 招进 Meta 的超级智能团队。后来 Meta 的 AI 组织进一步分工：Wang 统领相关团队，Nat Friedman 负责产品。一个深耕数据、评估和模型训练，一个有 GitHub CEO 的产品与开源经验，底层能力与产品化的分工已经很清楚。［来源 5］

Meta's move was aggressive. It invested US$14.3 billion in Scale AI for a 49 percent stake and recruited founder Alexandr Wang into its superintelligence effort. Meta later divided the AI organization further, with Wang overseeing the group and former GitHub CEO Nat Friedman leading products. One brings years of experience in data, evaluation, and model training; the other brings product and open-source experience. The split between foundational capability and commercialization is clear. [Source 5]

9 月 18 日，Meta 推出的独立智能体应用 Muse 登顶美国 App Store 免费榜。Sensor Tower 估算，Muse 在 9 月 8 日发布后的约五天内获得 73 万次下载，并一度超过 ChatGPT。这个成绩乍一看挺震撼，但榜首首先反映的是短期下载量和尝鲜热度，真正的成功仍要看长期留存率和实际应用价值。［来源 6］

On September 18, Meta's standalone agent app Muse reached the top of the U.S. App Store's free chart. Sensor Tower estimated 730,000 downloads in roughly the first five days after its September 8 launch, and Muse briefly moved ahead of ChatGPT. The result is striking, but a chart lead primarily reflects short-term downloads and curiosity. Durable success will depend on retention and practical value. [Source 6]

说实话，这种冷静挺难得。科技圈太容易被虚荣指标迷惑了——DAU、下载量、融资额——但这些数字往往掩盖了真实的用户痛点和产品价值。Muse 现在的挑战是：用户下载之后会不会持续用？它能解决什么核心问题？跟现有的助手类应用相比有什么差异化？这些问题才是决定成败的关键。

That caution is rare. Technology companies are easily distracted by vanity metrics—daily active users, downloads, financing totals—while those numbers can hide real user pain and product value. Muse's challenge is what happens after installation: will people keep using it, what core problem will it solve, and how will it differ from existing assistants? Those questions will determine its fate.

Wang 倒是在内部场景中验证了一些东西。他提到，在智能体循环、评估与优化指标都设置正确的情况下，一组智能体可以超过 100 名工程师的产出。当然，这种说法需要具体拆解——他没有披露具体任务、智能体数量、运行时长或人类基线，所以它只能被看作特定内部案例，而不能直接理解成智能体能全面替代工程师。［来源 7］

Wang has pointed to internal evidence: when the agent loop, evaluation, and optimization metrics are set correctly, a group of agents can outperform 100 engineers. That claim needs careful unpacking. He did not disclose the task, number of agents, run time, or human baseline, so it should be treated as a specific internal case—not proof that agents can broadly replace engineers. [Source 7]

Meta 这次押注的背后，其实反映了大厂们面临的共同焦虑：在 AI 时代如何保持竞争力？单纯堆模型参数已经不够了，关键是能不能把能力转化成用户愿意用、愿意付费的产品。Muse 的尝试才刚开始，但 Meta 显然已经认定，未来的社交、内容、广告等核心业务都会被 AI 智能体深度重构。

Behind Meta's wager lies a shared anxiety among technology giants: how do you remain competitive in the AI era? Adding model parameters is no longer enough. The decisive question is whether capability can become a product people choose to use and pay for. Muse is only beginning, but Meta has clearly concluded that agents will reshape social media, content, advertising, and the rest of its core business.

III.a16z 的数据视角：钱都花在哪儿了？a16z's Data View: Where Is the Money Going?

如果说前面两个案例更多是战略判断和产品试验，那 a16z 与金融科技公司 Mercury 合作的报告，就像给整个行业做了一次“财务体检”。他们分析了 Mercury 超过 20 万家客户在 2025 年 6 月至 8 月间的交易数据，并据此列出 50 家 AI 原生应用层公司。这里必须说明：报告中的横向应用 60%、垂直应用 40%，指的是榜单公司数量占比，不是总支出金额占比。［来源 8］

If the first two cases are mainly strategic judgment and product experiments, a16z's report with fintech company Mercury reads like a financial examination of the industry. It analyzed transactions from more than 200,000 Mercury customers between June and August 2025 and identified 50 AI-native application-layer companies. One clarification matters: the 60 percent horizontal and 40 percent vertical split describes the number of companies on the list, not the share of total dollars spent. [Source 8]

这个结构其实不意外——毕竟每家公司都需要通用的文档协作、会议记录、创意设计工具，但只有特定角色才需要客服自动化或合规审查系统。报告排除了主要提供云服务、GPU 或基础设施工具的公司，因此它观察的是 AI 真正被应用到哪些产品和工作流，而不是整个 AI 预算全景。

The structure is not surprising. Every company needs general-purpose tools for documents, meetings, and creative work, while only specific roles need customer-support automation or compliance review. The report excluded companies primarily selling cloud services, GPUs, or infrastructure tools, so it shows where AI is being applied in products and workflows—not a complete picture of every AI dollar.

更值得关注的是资金具体流向哪儿。初创公司把大部分预算花在 API 调用和招 AI 工程师上，而不是自己训练模型。这至少说明，在这份应用层榜单之外，创业公司更倾向于调用已有基础能力，再把资源集中到具体产品和工作流上，而不是从头训练基础模型。创业机会也因此更多地出现在应用层——怎么把模型能力包装成解决具体问题的产品。

More revealing is where the money goes. Startups put much of their budget into API usage and AI engineering talent rather than training foundation models themselves. At minimum, this suggests that beyond the application-layer ranking, startups prefer to consume existing foundational capabilities and focus resources on specific products and workflows instead of training models from scratch. The opportunity increasingly lies in packaging model capability around a concrete problem.

产品形态上，目前主流还是“人类增强型 Copilot”，比如会议助手、创意工具、代码补全。端到端的 AI 员工——完全自主完成任务、很少需要人类监督——还比较少。这也符合当前技术成熟度：AI 在辅助决策、提高效率方面已经很实用，但要完全接管复杂流程还有距离。

In product form, human-augmenting copilots still dominate: meeting assistants, creative tools, and coding aids. End-to-end AI employees that complete whole workflows with little human supervision remain less common. That fits the technology's current maturity. AI is already useful for decision support and efficiency, but it still has distance to travel before taking over complex processes.

横向应用领域，除了 OpenAI、Anthropic 这些通用助手稳居前列，Notion、Manus 这类智能工作台也进入榜单。会议助手赛道同样挤满玩家——Fyxer、Otter AI、Read AI 等——因为开会记录、总结、跟进这些需求太普遍了。

Among horizontal applications, general assistants such as OpenAI and Anthropic lead, while intelligent workspaces such as Notion and Manus also appear. Meeting support is crowded with Fyxer, Otter AI, Read AI, and others because recording, summarizing, and following up after meetings are universal needs.

创意工具这块竞争更激烈。Freepik、ElevenLabs、Canva、Midjourney、CapCut……设计师和内容创作者的工作流正在被 AI 彻底改写。以前你得懂 PS、会剪辑、能写代码，现在很多时候用自然语言描述需求就能生成八成可用的素材。这种“技能边界模糊化”的趋势，对传统职业分工是个不小的冲击。

Competition is even more intense in creative tools: Freepik, ElevenLabs, Canva, Midjourney, CapCut, and others are rewriting the workflows of designers and content creators. Tasks that once demanded image-editing skills, video production, or code can increasingly begin with a natural-language request and produce material that is already mostly usable. This blurring of skill boundaries is a serious challenge to traditional divisions of labor.

还有一个特别火的品类叫“氛围编程”（Vibe Coding）平台，代表是 Replit、Cursor、Lovable 和 Emergent。这些工具让非专业程序员也能通过对话式交互快速搭建应用原型。a16z 的数据里，Replit 从 Mercury 客户获得的收入约为 Lovable 的 15 倍，说明这个赛道已经从极客玩具变成了企业级工具。想象一下，以后产品经理、运营人员都能自己写点小工具，不用事事找开发排期，这对组织效率的提升是质变级的。［来源 8］

Another fast-rising category is vibe coding, represented by Replit, Cursor, Lovable, and Emergent. These tools let non-programmers build application prototypes through conversation. In a16z's Mercury data, Replit generated roughly 15 times the revenue of Lovable, suggesting the category has moved beyond a hobbyist toy into enterprise tooling. Product managers and operators may increasingly build their own internal utilities instead of waiting in an engineering queue—a qualitative change in organizational efficiency. [Source 8]

垂直应用方面，客服、销售、招聘、运营等职能都有针对性的解决方案在涌现。不过目前多数还是“增强人类”模式。a16z 统计的 17 家垂直应用公司中，12 家侧重增强员工，5 家定位为端到端“AI 员工”，包括 Crosby Legal、Cognition、11x、Serval 和 Alma。现实很清楚：在需要复杂判断、涉及法律责任或高风险决策的场景，企业还不敢完全放手让 AI 自主执行。［来源 8］

Among vertical applications, specialized products are emerging across customer service, sales, recruiting, and operations. Most still augment people. Of the 17 vertical companies counted by a16z, 12 focused on supercharging employees and five positioned themselves as end-to-end “AI employees”: Crosby Legal, Cognition, 11x, Serval, and Alma. The reality is clear: where judgment is complex, liability matters, or decisions carry high risk, enterprises are not yet ready to let AI act entirely on its own. [Source 8]

a16z 总结出的几个趋势，我觉得挺准的：

Several trends in the a16z report strike me as accurate:

• 多模型共存 · Multiple Models Will Coexist 短期内不会出现一家通吃的局面，用户会根据场景在不同界面和模型之间切换。对开发者来说，不用押注某一家大模型，保持灵活性更重要。 No single provider is likely to take everything in the short term. Users will switch among interfaces and models according to the task. Developers benefit more from flexibility than from betting everything on one model.
• 氛围编程企业化 · Vibe Coding Goes Enterprise 当编程门槛降到“全员可操作”，企业内部的工作流和协作模式会发生根本变化。以前需要跨部门协调的需求，可能以后一个人就能搞定。 When coding becomes accessible across the workforce, internal workflows and collaboration patterns change fundamentally. Tasks that once required cross-functional coordination may become one-person projects.
• 消费级 AI 向企业渗透 · Consumer AI Moves Into Work 榜单中近 70% 的产品可以由个人先采用、再带进团队，不要求先购买企业许可。这种自下而上的采用路径，比传统企业软件自上而下的推销模式更快。 Nearly 70 percent of the listed products can be adopted by individuals and then brought into teams without an enterprise license. This bottom-up path can move faster than traditional top-down enterprise software sales.
• 产业资本重构 · Capital Reshapes the Stack 算力和模型训练的壁垒仍然很高，但能力通过 API 下沉后，应用层创新反而加速了。价值不再只集中在 GPU 和数据中心，也分散到越来越多具体工作流里。 Compute and frontier-model training remain highly capital-intensive, but APIs lower the application barrier and accelerate experimentation. Value is no longer concentrated only in GPUs and data centers; it is spreading across increasingly specific workflows.

IV.软硬件逻辑的切换与早期渗透The Shift from Software to Hardware—and the Early Stage of Adoption

a16z 的另一份市场状况报告用上百张图表拆解了一个更宏观的问题：科技行业的盈利结构和投入逻辑正在发生什么变化？

Another a16z State of Markets report uses more than a hundred charts to examine a broader question: how are the technology industry's profit structure and investment logic changing?

数据挺惊人的：截至 2026 年 8 月下旬，科技板块贡献了标普 500 指数约 76% 的盈利增长。换句话说，现在的资本市场周期很大程度上被科技公司主导了。这背后是 AI、云计算、数字化转型等趋势共同推动的结果。［来源 9］

The figure is striking: as of late August 2026, technology contributed roughly 76 percent of total S&P 500 earnings growth. In other words, the capital-markets cycle is now heavily shaped by technology companies, with AI, cloud computing, and digital transformation reinforcing one another. [Source 9]

但更有意思的是投入的两极分化。报告图表显示，AI 支出最高的前 1% 企业，其支出中位数约为前 10% 企业中位数的 8 倍。这个差距之大，反映了一个残酷的现实：在 AI 时代，资源集中度可能比互联网时代更高。训练大模型、建数据中心、拿 GPU 配额——这些都不是一般公司玩得起的游戏。［来源 9］

Even more revealing is the polarization of investment. A chart in the report shows that median spending among the top 1 percent of AI spenders is roughly eight times the median among the top 10 percent. The gap points to a harsh reality: resource concentration in the AI era may exceed what we saw in the internet era. Training frontier models, building data centers, and securing GPU capacity are not games most companies can afford to play. [Source 9]

市场叙事也在从软件向硬件和能源基础设施倾斜。过去十年大家追捧 SaaS 公司，看重的是高毛利、可扩展、订阅制现金流。但现在投资人开始关注芯片设计、数据中心、电力供应——因为 AI 训练和推理消耗的算力太大了。a16z 的报告指出，算力需求仍然快于供给，甚至旧一代 GPU 的价格也没有像传统硬件周期那样快速下滑。［来源 9］

The market narrative is also rotating from software toward hardware and energy infrastructure. For a decade, investors favored SaaS companies for high margins, scalability, and subscription cash flow. Now attention is shifting to chip design, data centers, and power supply because AI training and inference consume enormous amounts of compute. a16z reports that demand for compute still exceeds supply, with even older GPUs holding value rather than depreciating as quickly as traditional hardware cycles would suggest. [Source 9]

不过，企业级 AI 的普及其实还处在非常早期的阶段。a16z 的数据称，接近 30% 的标普 500 公司报告了某种“可量化影响”，但只有约 2% 披露了持续追踪的指标；智能体用例中，真正大规模部署的用户也只占很小一部分。传统 SaaS 公司因此面临商业模式重构——以前卖的是“软件使用权”，现在可能要转向“按效果付费”或“按调用量计费”。这种转变对财务模型、销售策略、产品架构都是考验。［来源 9］

Enterprise adoption, however, remains very early. According to a16z, nearly 30 percent of S&P 500 companies report some “quantifiable impact” from AI, yet only around 2 percent disclose a metric they track over time; only a small share of users deploy agents meaningfully at scale. Traditional SaaS companies therefore face a business-model rebuild. Instead of selling access to software, they may need to charge for outcomes or usage—a demanding shift for financial models, sales strategy, and product architecture. [Source 9]

V.写在最后Final Thoughts

把这几个视角拼在一起看，你会发现 AI 智能体时代的轮廓越来越清晰了：

Put these perspectives together and the outline of the agentic era becomes clearer:

基础设施层面，支付、数据、算力正在为“机器自主交易”做准备；产品层面，从通用助手到垂直场景，各种形态的 AI 应用正在快速迭代；资本层面，钱从模型训练流向应用创新，从软件转向硬件和能源。

At the infrastructure layer, payments, data, and compute are preparing for machines that can transact autonomously. At the product layer, applications are iterating rapidly from general assistants to vertical workflows. At the capital layer, money is spreading from model training into application innovation and rotating from software toward hardware and energy.

但同时，挑战也很明显：安全风险在上升，商业模式还在摸索，用户留存有待验证，技术债需要清理。这场变革不会一蹴而就，可能需要五年甚至十年才能真正重塑互联网的底层逻辑。

The challenges are equally clear: security risk is rising, business models remain unsettled, retention still needs proof, and technical debt must be cleared. This shift will not happen overnight. It may take five or ten years to reshape the internet's underlying logic.

对创业者来说，机会窗口还开着——应用层创新的门槛比以往任何时候都低，但竞争也异常激烈。关键是找到真实痛点，持续打磨产品，别被虚荣指标和炒作周期带偏。

For founders, the opportunity window remains open. The barrier to application-layer innovation is lower than ever, but competition is intense. The key is to find a real pain point, keep refining the product, and refuse to be distracted by vanity metrics or hype cycles.

对企业来说，现在不是观望的时候了。无论你在哪个行业，都该认真思考 AI 会如何影响你的业务流程、用户体验和竞争格局。早期采用者的优势会越拉越大。

For companies, this is not the moment to stand back and watch. Whatever the industry, it is time to think seriously about how AI will affect workflows, customer experience, and competition. The advantage of early adopters will compound.

至于普通用户，或许很快你就会发现，你的工作方式、消费习惯、甚至思维模式，都在被这些智能体悄悄改变。
这不是科幻，而是正在发生的现实。 For ordinary users, the change may arrive quietly: the way you work, shop, and even think will begin to shift around these agents.
This is not science fiction. It is already happening.

Source Note出处说明Source Note

• 来源 1 · Stripe 的“奇点”讨论 / Stripe on “the Singularity” Stripe Sessions 2026：A conversation with Daniel Gross and Nat Friedman —— 公开逐字稿记录 Patrick Collison 将 2026 年 1 月 1 日称作起点，并记录 Nat Friedman 对模型自我改进路径的判断；本文未将其误写为已被验证的行业共识。 Stripe Sessions 2026: A conversation with Daniel Gross and Nat Friedman — the public transcript records Patrick Collison naming January 1, 2026 as the starting point and Nat Friedman describing a path toward model self-improvement. This essay does not present their judgment as verified industry consensus.
• 来源 2 · Stripe 支付规模 / Stripe Payment Volume Stripe 2025 年度公开信 —— Stripe 披露 2025 年总支付额为 1.9 万亿美元，同比增长 34%，并记录了对 Privy 与 Metronome 的收购。 Stripe's 2025 annual letter — Stripe reports US$1.9 trillion in total payment volume in 2025, up 34 percent year over year, and records its acquisitions of Privy and Metronome.
• 来源 3 · Stripe × OpenRouter / Stripe × OpenRouter Stripe Newsroom：同意收购 OpenRouter —— 官方公告说明 OpenRouter 覆盖 80 多家提供商、400 多个模型，并用于按任务复杂度、价格、速度和可靠性进行动态路由。 Stripe Newsroom: agreement to acquire OpenRouter — the official announcement says OpenRouter spans more than 80 providers and 400 models, routing requests by complexity, price, speed, and reliability.
• 来源 4 · Stripe × Bridge / Stripe × Bridge Stripe Newsroom：完成 Bridge 收购 —— 官方公告记录 Stripe 于 2025 年 2 月完成对稳定币基础设施平台 Bridge 的收购。 Stripe Newsroom: Bridge acquisition completed — the official announcement records Stripe's February 2025 completion of its acquisition of stablecoin infrastructure platform Bridge.
• 来源 5 · Meta × Scale AI / Meta × Scale AI Associated Press：Meta 投资 Scale AI —— 报道记录 143 亿美元投资、49% 股权及 Alexandr Wang 加入 Meta 超级智能团队；后续组织重组报道记录 Wang 统领相关团队、Nat Friedman 负责产品。 Associated Press: Meta invests in Scale AI — the report records the US$14.3 billion investment, 49 percent stake, and Alexandr Wang's move to Meta's superintelligence effort; subsequent restructuring coverage records Wang overseeing the group and Nat Friedman leading products.
• 来源 6 · Muse 下载与榜单 / Muse Downloads and Ranking CNBC：Muse 下载量与 App Store 排名 —— Sensor Tower 估算其发布后约五天下载 73 万次，并于 9 月 18 日超过 ChatGPT、登顶美国 iOS 免费榜。 CNBC: Muse downloads and App Store ranking — Sensor Tower estimated 730,000 downloads in roughly five days and reported that Muse passed ChatGPT to lead the U.S. iOS free chart on September 18.
• 来源 7 · 智能体团队产出说法 / Agent-Team Output Claim ABAB News：Alexandr Wang 在 YC Startup School 的说法 —— 该报道同时注明未披露具体案例、智能体数量、运行时长或人类基线，因此本文把它限定为内部案例陈述。 ABAB News: Alexandr Wang's YC Startup School remarks — the report also notes that no task, agent count, run time, or human baseline was disclosed, so this essay limits the claim to an internal case statement.
• 来源 8 · AI 应用支出榜 / AI Application Spending Report a16z：The AI Application Spending Report —— 基于 Mercury 超过 20 万客户在 2025 年 6—8 月的交易；60%/40% 是榜单公司数量占比，不是总支出金额占比。 a16z: The AI Application Spending Report — based on transactions from more than 200,000 Mercury customers between June and August 2025. The 60/40 split is the share of companies on the list, not total spending dollars.
• 来源 9 · State of Markets II / State of Markets II a16z：State of Markets II —— 报告记录科技板块约贡献 2026 年标普 500 盈利增长的 76%，并指出 GPU 需求仍快于供给、企业 AI 采用广而不深。 a16z: State of Markets II — the report says technology contributed roughly 76 percent of S&P 500 earnings growth in 2026, while GPU demand still outpaced supply and enterprise AI adoption remained broad but shallow.

AI 产业观察 · 记于 2026-10-13 · YingChao (YC) AI industry observation · written on 2026-10-13 · YingChao (YC)

我们以为的"AI爆发"，可能才刚刚开始 What We Think Is the "AI Explosion" May Only Be the Beginning

说实话，我第一次看到凯文·凯利那个"生物阻尼"的说法时，愣了好几秒。 Honestly, the first time I came across Kevin Kelly's idea of "biological drag," I stopped for several seconds.

I.为什么有的技术一夜爆发，有的却走了三十年Why Some Technologies Explode Overnight While Others Take Thirty Years

你想想看——VR炒了多少年？从上世纪九十年代的噱头到现在，三十年了，成本降了不知道多少倍，可你什么时候见谁天天戴着头盔过日子？而AI呢，闷头搞了半个世纪，忽然就炸了。同样是技术，为什么有的磨磨蹭蹭，有的一夜之间铺天盖地？［来源 1］

Think about it—how many years has VR been hyped? From a novelty in the 1990s to today, thirty years have passed and the cost has fallen by who knows how many times. Yet when have you ever seen people wearing headsets all day, every day? AI, meanwhile, kept its head down for half a century and then suddenly exploded. Both are technologies, so why does one inch forward while the other seems to blanket the world overnight? [Source 1]

凯利的解释是：看它要不要跟人的身体打交道。

Kelly's explanation is this: ask whether the technology has to interact with the human body.

纯软件的东西，代码一跑，全球同步，没有摩擦力。可一旦涉及到物理交互——不管是VR的头盔还是机器人的胳膊——就有一堆现实世界的阻力在那等着。能耗、精度、舒适度、成本，每一个都是实打实的约束。他说人脑才25瓦，而要让机器人达到类似的身体灵活性，能耗得高出好几个数量级。这不是工程问题，是物理定律摆在那儿。［来源 1、2］

Pure software can run once and synchronize around the world with almost no friction. But the moment physical interaction enters the picture—whether it is a VR headset or a robot arm—a whole set of real-world constraints is waiting: energy use, precision, comfort, and cost. Every one of them is concrete. Kelly says the human brain uses only 25 watts, while giving a robot comparable bodily dexterity would require orders of magnitude more energy. This is not merely an engineering problem; it is physics asserting itself. [Sources 1, 2]

这个观察我觉得特别有启发性。我们这代人有个毛病，总觉得软件能解决一切，觉得"只要算法够好，没有搞不定的事"。但凯利提醒我们：技术扩散的速度，很大程度上取决于它的物理形态。纯数字的东西可以指数级蔓延，可一旦碰到肉身，就被拖慢了。

I find that observation especially illuminating. Our generation has a habit of assuming software can solve everything—that if the algorithm is good enough, nothing is out of reach. Kelly reminds us that the speed of technological diffusion depends heavily on its physical form. A purely digital technology can spread exponentially; once it meets the body, it slows down.

II.我们对智能几乎一无所知We Know Almost Nothing About Intelligence

然后他说了句让我印象极深的话：我们对智能几乎一无所知。

Then he said something that stayed with me: we know almost nothing about intelligence.

他说今天的AI只是在统计学上"偶然"合成出了几种认知的化合物，就像牛顿第一次看见电，压根不知道那是什么。人类目前的智能形态，在整个心灵可能性空间里，不过是个偏远的角落。换句话说，我们连智能是什么都没搞懂，就已经开始急着讨论"AI会不会取代人类"了——这本身就很荒唐。［来源 1、2］

He says today's AI has merely "chanced upon" several compounds of cognition through statistics, much as Newton encountered electricity without truly knowing what it was. The form of intelligence humans possess occupies only a remote corner of the entire possibility space of minds. In other words, we have not even figured out what intelligence is, yet we are already rushing to debate whether AI will replace humanity—which is absurd in itself. [Sources 1, 2]

III.瓶颈不在算法，而在验证The Bottleneck Is Not the Algorithm but Verification

普林斯顿的王梦迪教授从另一个角度印证了这一点。她说现在的大模型本质上在搞"众数寻求"——就是找那个最可能、最常见的答案。这解释了很多事情：为什么AI写八股文比你快，却写不出真正惊艳的东西；为什么它能在数学和编程上突飞猛进，却在真正的基础科学发现上没什么存在感。

Professor Mengdi Wang of Princeton approaches the same point from another direction. She says today's large models are essentially "mode seeking": they look for the most probable, most common answer. That explains a great deal—why AI can produce formulaic writing faster than you can, yet struggles to create something genuinely astonishing; why it can advance rapidly in mathematics and programming while barely registering in genuine fundamental scientific discovery.

原因很简单：数学有编译器，对错一跑就知道；可真实的科学探索，瓶颈从来不是"缺想法"，而是缺乏可重复的验证机制。王梦迪团队在做的事情——全自动量子材料实验平台、实验室智能操作系统LabOS——本质上是在搭建一个"验证闭环"，让AI能把手伸进物理世界，真正去试错、去反馈。［来源 3、4］

The reason is simple: mathematics and code have verifiers—you can run them and know whether they are right or wrong. In real scientific exploration, however, the bottleneck has never been a shortage of ideas; it is the absence of repeatable verification. What Wang's team is building—an automated quantum-materials experimental platform and the intelligent laboratory operating system LabOS—is, at heart, a verification loop that lets AI reach into the physical world, truly try, fail, and learn from feedback. [Sources 3, 4]

瓶颈不在算法，在于验证。不在于能不能想出来，而在于能不能证明它是对的。

The bottleneck is not the algorithm but verification. The question is not whether an idea can be generated, but whether it can be proved right.

这两个人说的话，其实指向同一个方向：瓶颈不在算法，在于验证。不在于能不能想出来，而在于能不能证明它是对的。

What these two people are saying points in the same direction: the bottleneck is not the algorithm but verification. The question is not whether an idea can be generated, but whether it can be proved right.

IV.当"好"变得廉价，"不同"才值钱When "Good" Becomes Cheap, "Different" Gains Value

那对我们普通人意味着什么？

What does that mean for ordinary people like us?

凯利给了个我觉得挺实在的建议：别去挤"最好"的赛道，去找"唯一"的生态位。这话听起来有点鸡汤，但他后面的解释挺有意思——你的独特经历、你的品味、你的好奇心的组合，这些东西是没有第二个人能复制的。当AI把所有人的生成成本都拉到零的时候，"好"变得廉价，"不同"反而值钱。［来源 1］

Kelly offers advice that I find quite practical: do not crowd into the race to be "the best"; find the niche in which you are "the only." It may sound a little like motivational talk, but his explanation is interesting: no one else can replicate the combination of your particular experiences, your taste, and your curiosity. When AI drives everyone's cost of generation toward zero, "good" becomes cheap and "different" becomes valuable. [Source 1]

他还提了个概念叫"单人受众"。当创作门槛趋近于零，一个人就能做出以前需要一个团队才能完成的东西，创作的驱动力就从"被别人看见"回归到"让自己开心"。听起来有点孤独，但仔细想想，这不正是现在很多人做自媒体、做独立产品的真实状态吗？［来源 1］

He also introduces the idea of an "audience of one." As the barrier to creation approaches zero, one person can make what once required an entire team, and the motive for creating shifts from "being seen by others" back to "making yourself happy." It sounds a little lonely, but think about it: is that not the actual condition of many people building independent media and independent products today? [Source 1]

V.最容易被忽略的红队纠偏The Red-Team Correction That Is Easiest to Miss

不过凯利最后那段"红队纠偏"我得重点说一下，因为这部分很容易被忽略。他提醒我们：历史连续性只是个假设，不是铁律。在范式革命的阶段，泛化能力比精细化更重要。甚至他说，提示词工程本身可能只是个过渡技能——今天你花大量时间学的东西，明天可能就不存在了。

But I have to emphasize Kelly's final "red-team correction," because it is easy to overlook. He reminds us that historical continuity is only an assumption, not an iron law. During a paradigm shift, general adaptability matters more than narrow refinement. He even suggests that prompt engineering itself may be a transitional skill—something you spend enormous time learning today may no longer exist tomorrow.

这话听着有点丧，但我觉得反而是最实在的。与其焦虑"AI会不会抢走我的饭碗"，不如停下来想想：我现在做的事里，哪些是真正需要我判断、需要我品味的？哪些只是机械重复、迟早会被自动化的？

That sounds discouraging, but to me it is the most realistic part. Instead of worrying about whether AI will take your job, stop and ask: which parts of what I do genuinely require my judgment and taste? Which parts are mechanical repetition and will eventually be automated?

VI.在不变的东西里，找到唯一的事Find the One Thing You Can Do Within What Does Not Change

说到底，凯利的思想骨架我觉得可以用一句话概括：预测未来最好的方式不是盯着前沿看，而是往回看，看万年尺度下什么是不变的——人性、物理规律、社会组织的基本逻辑。然后在那些不变的东西里，找到你能做的那件"唯一"的事。

In the end, I think Kelly's intellectual framework can be summarized in one sentence: the best way to predict the future is not to stare at the frontier, but to look backward and ask what remains unchanged across ten thousand years—human nature, physical laws, and the basic logic of social organization. Then, within those constants, find the one thing that only you can do.

万小时定律依然有效，只不过"小时"得花在真正需要深度认知的地方。AI负责搭建框架，你负责播下种子。这不是谁取代谁的问题，是认知分工的问题。［来源 1］

The ten-thousand-hour rule still applies, except those hours must be spent where deep understanding is actually required. AI builds the framework; you plant the seed. This is not a question of who replaces whom. It is a division of cognitive labor. [Source 1]

王梦迪在做验证闭环，凯利在说历史回溯。两个人从不同方向，其实都在回答同一个问题：在一个什么都能被生成的时代，什么才是真正稀缺的？

Wang is building verification loops; Kelly is looking backward through history. From different directions, they are answering the same question: in an age when anything can be generated, what is truly scarce?

答案可能是：能判断对错的能力，能定义问题的能力，以及愿意花十年磨一剑的耐心。 The answer may be the ability to judge right from wrong, the ability to define the problem, and the patience to spend ten years sharpening a single sword.

Source Note出处说明Source Note

• 来源 1 · 凯文·凯利访谈 / Kevin Kelly Interview Every：How WIRED's Kevin Kelly Predicts the Future After 40 Years in Tech —— 访谈谈及 VR 的缓慢扩散、身体与物理约束、智能如认知化合物、心灵可能性空间、"不要做最好，要做唯一"、"单人受众"、万小时练习及用历史理解未来。正文保留作者对访谈的理解与概括。 Every: How WIRED's Kevin Kelly Predicts the Future After 40 Years in Tech — the interview covers VR's slow diffusion, bodily and physical constraints, intelligence as cognitive compounds, the possibility space of minds, "do not be the best; be the only," the audience of one, ten-thousand-hour practice, and using history to understand the future. The essay preserves the author's interpretation and synthesis of the conversation.
• 来源 2 · 智能理论与 25 瓦 / Intelligence Theory and 25 Watts Kevin Kelly：Without a theory of intelligence —— 凯利明确写道人类对智能缺乏理论，可能心灵空间极其广阔，并以人脑约 25 瓦说明智能的最低能耗仍是开放问题。 Kevin Kelly: Without a theory of intelligence — Kelly argues that humanity lacks a theory of intelligence, that the space of possible minds may be vast, and that the brain's roughly 25-watt power use leaves the minimum energy requirement of intelligence an open question.
• 来源 3 · 众数寻求与验证闭环 / Mode Seeking and Verification Loops 王梦迪：AI自主发现新科学的关键是「可验证」 —— 2026 外滩大会演讲精编，涵盖大模型的 Mode Seeking、数学与编程中的验证器、量子材料自动实验平台，以及连接物理实验室的闭环系统。 Mengdi Wang: Verifiability Is the Key to Autonomous Scientific Discovery — an edited transcript of her 2026 Inclusion Conference keynote covering mode seeking in large models, verifiers in mathematics and coding, an automated quantum-materials platform, and closed-loop systems connected to physical laboratories.
• 来源 4 · LabOS / LabOS LabOS 官方项目页与 ICML 2026 主题演讲摘要 —— 项目将多智能体、智能眼镜、机器人与实验流程连接，用于情境理解、操作辅助、偏差检测与湿实验验证闭环。 The official LabOS project page and the ICML 2026 keynote abstract — the project connects multi-agent systems, smart glasses, robots, and laboratory workflows for contextual understanding, execution support, deviation detection, and closed-loop wet-lab validation.

AI 认知随笔 · 记于 2026-10-15 · YingChao (YC) AI reflection · written on 2026-10-15 · YingChao (YC)

二十年后，那个客户回来了：长期主义与 AI 的双向奔赴 Twenty Years Later, That Client Came Back: Where Long-Term Thinking Meets AI

说实话，当那个熟悉的名字再次出现在我 WhatsApp 聊天框的时候，我愣了好几秒。二十年了，从 2006 年左右开始的合作，中间断了七八年，现在又找回来了。 Honestly, when that familiar name appeared in my WhatsApp chat again, I froze for several seconds. It has been twenty years since we began working together around 2006. We lost touch for seven or eight years, and now we have found each other again.

I.一段关系，不只建立在一次交易上A Relationship Is Built on More Than One Transaction

这种感觉挺奇妙的。你知道吗，做生意这么多年，见过太多"一锤子买卖"的案例，也见过无数追逐短期利润的同行。但我一直相信一件事：真正的商业关系，不是建立在某一次交易上，而是建立在持续的信任和价值交付上。这次重新联系，对我来说，就像是这个理念得到了某种验证——虽然听起来有点自我感动，但确实是这么想的。

The feeling is hard to describe. After so many years in business, I have seen too many one-off deals and too many peers chasing short-term profit. But I have always believed one thing: a real business relationship is not built on a single transaction. It is built on sustained trust and the repeated delivery of value. To me, this reconnection feels like a kind of validation of that belief. It may sound a little self-congratulatory, but that is honestly how I feel.

回头看这段合作关系，其实挺有意思的。本世纪初那会儿，大家都在摸索，供应链管理还没现在这么精细化，很多东西靠的是经验和人情。我记得最开始合作的时候，这个客户规模还不大，我们也是一步步磨合过来的。有时候他们的需求很急，有时候市场变化快，但我始终坚持一个原则：不糊弄，不夸大，能做到什么程度就说什么程度。有几次为了帮他们找到合适的供应商，我自己贴钱都干过。

Looking back, the relationship is quite interesting. In the early years of this century, everyone was still finding their way. Supply-chain management was nowhere near as refined as it is today, and much depended on experience and personal trust. When we first started working together, the client was not yet large. We learned how to work with each other step by step. Sometimes their needs were urgent; sometimes the market changed quickly. But I held to one principle: do not bluff, do not exaggerate, and say plainly what can and cannot be done. More than once, I even covered costs myself to help them find the right supplier.

后来因为各种原因中断了。疫情前那几年，他们可能有自己的渠道，或者业务方向调整了，反正就没联系了。说不失落是假的，但我也没去刻意维系什么，因为我觉得强求没意义。该是你的客户，早晚会回来；不是你的，硬拉着也没用。

Later, the relationship paused for all kinds of reasons. In the years before the pandemic, perhaps they developed their own channels, or perhaps their business direction changed. In any case, we stopped talking. It would be untrue to say I was not disappointed, but I did not try to force the connection. I have never believed there is much point in that. A client who is meant to return will return sooner or later. If the fit is no longer there, holding on tightly will not change it.

II.用做一辈子的心态，对待每一次合作Treat Every Engagement with a Lifetime Mindset

现在想想，这大概就是长期主义吧。不是说非要跟每个客户合作一辈子，而是用做一辈子的心态去对待每一次合作。你今天省下的那点钱、占的那点便宜，可能就是明天失去的信任。而信任这个东西，一旦建立起来，它的生命力比你想象的要长。

Looking back now, perhaps this is what long-term thinking means. It does not mean you must work with every client for life. It means approaching every engagement with the mindset that you could. The small amount of money you save today, or the little advantage you take, may become the trust you lose tomorrow. Once trust is established, its life can be far longer than you imagine.

不是说非要跟每个客户合作一辈子，而是用做一辈子的心态去对待每一次合作。

It is not about working with every client for life. It is about treating every engagement with a lifetime mindset.

III.AI 给了我更多工具，去兑现长期价值AI Gives Me More Tools to Deliver Long-Term Value

更让我兴奋的是，现在这个时代给了我更多工具去兑现这种长期价值。这几年 AI 的发展速度快得吓人，我自己也在不断学习怎么把这些工具用到实际业务里。比如供应链管理这块，以前很多环节需要大量人工盯着，现在可以用 AI 做需求预测、库存优化、风险预警。我最近在测试一个系统，能够实时分析多个供应商的交付能力和价格波动，然后给出最优组合方案。

What excites me even more is that this era has given me more tools to deliver that long-term value. AI has been developing at an astonishing pace, and I have been continuously learning how to apply these tools to real business. In supply-chain management, for example, many steps that once required constant manual attention can now use AI for demand forecasting, inventory optimization, and risk alerts. I have recently been testing a system that can analyze multiple suppliers' delivery capabilities and price fluctuations in real time, then recommend the best combination.

这不是说要取代人，而是让人可以腾出手来做更有价值的事情。比如我现在可以花更多时间去理解客户的业务逻辑，去思考他们未来三五年的需求会怎么变化，而不是被日常琐碎的数据整理和报表制作困住。AI 帮我提高了效率，但决策和判断还是要靠人。

This is not about replacing people. It is about freeing people to do work of greater value. I can now spend more time understanding the client's business logic and thinking about how their needs may change over the next three to five years, instead of being trapped by routine data organization and report preparation. AI improves my efficiency, but decisions and judgment still depend on people.

IV.老客户回来了，而我也不再只靠旧方法The Client Returned, but I No Longer Rely on Old Methods

说到这儿，我想起一个细节。这次重新联系的客户，他们现在的业务规模比以前大了好几倍，需求也更复杂了。如果还是用十几年前的那套方法，说实话我可能真应付不来。但现在有了这些工具，我反而更有信心能给他们提供更好的服务。这就是时代的红利吧，你坚持做对的事情，然后时代给你更好的工具去把事情做得更好。

That brings one detail to mind. The client who has reconnected now operates at several times their former scale, and their needs are far more complex. If I were still using the methods of more than a decade ago, honestly, I might not be able to keep up. But with today's tools, I feel more confident that I can serve them better. Perhaps this is the dividend of the times: you keep doing the right thing, and then the era gives you better tools to do it better.

当然，我也见过很多人把 AI 当成噱头，好像用上了就能躺赢。其实不是的。工具永远只是工具，核心还是你的商业逻辑和价值观。你要是本来就想着捞快钱，AI 只会让你捞得更快，然后死得更快。但如果你本来就在做长期价值的事情，AI 就是加速器。

Of course, I have also seen many people treat AI as a gimmick, as if using it guarantees an easy win. It does not. A tool is always only a tool. The core remains your business logic and your values. If your original intention is to make fast money, AI will only help you do it faster—and fail faster. But if you are already creating long-term value, AI becomes an accelerator.

V.技术只是放大器Technology Is an Amplifier

我有时候会想，如果当年没有坚持那些看起来"吃亏"的原则，现在会怎样？可能多赚了一些钱，但肯定不会有这次重逢。而这次重逢带来的，不仅仅是一个客户的回归，更是对我这么多年坚持的一种肯定。这种肯定，比任何短期利润都更有价值。

Sometimes I wonder what would have happened if I had not held to those principles that seemed, at the time, to put me at a disadvantage. I might have made a little more money, but this reunion almost certainly would not have happened. What it brings is not merely the return of a client. It is an affirmation of what I have held to for all these years. That affirmation is worth more than any short-term profit.

AI 时代确实提供了更多可能性，但说到底，技术只是放大器。它放大的是你本来就在做的事情——如果你在创造价值，它帮你创造更多；如果你在割韭菜，它也会让你的吃相更难看。所以我现在更确信，长期主义加上新工具，这可能是最适合这个时代的组合。

The AI era genuinely offers more possibilities, but in the end, technology is only an amplifier. It amplifies what you were already doing. If you are creating value, it helps you create more. If you are exploiting people, it makes that exploitation even more visible. That is why I am now more convinced than ever that long-term thinking combined with new tools may be the combination best suited to this era.

你坚持做对的事，时间会给你答案，技术会给你翅膀。
就这么简单。 Keep doing the right thing. Time will give you the answer, and technology will give you wings.
It is that simple.

外贸与 AI 随笔 · 记于 2026-10-19 · YingChao (YC) Foreign trade and AI reflection · written on 2026-10-19 · YingChao (YC)