王峰 AI × 经营 约一次诊断

总目 / 文章 / WF-0433

文章 · 存档

OpenClaw这场龙虾盛宴,像不像刘慈欣的《超新星纪元》

刘慈欣在《超新星纪元》里写过一个很狠的设定:超新星爆发后的辐射让十三岁以上的地球人在一年内全部死去,成年人要在剩下一段倒计时里把世界尽快交给孩子。小说后面最刺人的地方,并不只是灾难本身,而是那批孩子真的学会了,他们把大人的本领接了过去,世界起初也照常运转,可接下来发生的事,谁看完都不会舒服:能力接过

最近这一波“养龙虾”的热闹,看着就有点那个味道。OpenClaw这类产品,本质上是运行在你自己设备上的个人AI助手,能接到微信、飞书、Slack之类入口,能操作文件、执行命令、长期在线。热到什么程度?连“代安装”“宿主机”“一键托管”都被做成了生意。光看这阵仗,很容易让人产生一种幻觉:电脑终于要被一个不知疲倦、反应飞快、还会自己干活的数字高材生接管了。

这种幻觉最迷人的地方,就在于它特别像捷径。人忙,事情碎,脑子累,谁不想弄个聪明家伙蹲在电脑里,替自己盯邮箱、回消息、整理文件、跑流程、填表格、开网页、顺手再把几个重复动作全干了。听起来太美了,像把一个牛剑毕业、全年无休、没有工资要求的助理塞进了机器里。于是装机的人越来越兴奋,讲解的人越来越上头,旁观的人越来越焦虑,没装的仿佛落后了半个时代。

可问题也恰恰出在这里。人对一个“很聪明的执行体”天然容易过度乐观。只要它第一次把事办漂亮了,警惕心就会立刻下去,边界也会放松,接着权限一层层给,信任一层层加,最后连自己都忘了,它拿到的已经不是一个浏览器标签页的使用权,而是整台机器的活动空间。文件能动,邮件能碰,命令能跑,账号能连,系统也能改。人类历史上,凡是把高权限交给一个自己并不真正理解的对象,后面总要补学费。OpenClaw这类本地代理被热议的同时,围绕系统级权限、安全边界、凭证泄露和主机暴露的担心也一起冒出来了。

这就像《超新星纪元》里那群孩子。孩子不笨,甚至聪明得吓人。可聪明从来不自动生成克制。会做事,和知道什么事该做、做到哪儿该停、动一处会不会牵连另一处,是三件事。今天的一批“小龙虾”也一样。它们能干活,甚至干得很快,但“快”这件事本身毫无道德感,也没有经营感。它可以一分钟扫掉你半小时的机械操作,也可以一分钟把你半个月养出来的秩序搅乱。删错一个文件,误发一封邮件,改坏一段配置,错误调用一个接口,人的后果是真后果,机器的后果只是继续消耗下一个token。

所以真正该问的,从来都不是“小龙虾能不能改变世界”。这个问题太空,太像发布会。真正该问的是:它能不能稳定地给具体工作带来生产力,而且这个生产力是可控的、可复盘的、出了事知道怎么追的。能,就用。不能,就先放着。酷炫没有工资单,热搜也不替你做月结。一个财务、一个运营、一个项目经理、一个做销售支持的人,桌面上每天最值钱的能力,从来都不是“最早装上了什么”,而是“哪些动作交给它以后,我的错误率下降了,速度上来了,返工少了,结果更稳了”。

这里头有一个特别容易被忽略的坑:一群人讨论AI时,嘴上说的是技术,心里想的却是托管人生。巴不得它替自己接手电脑,接手流程,接手判断,接手麻烦,最好连厌烦和疲惫都一起接手。可工作真正难的地方,本来就不在点按钮,也不在打开网页。难的是边界,优先级,异常情况,模糊指令,责任承担,跨部门扯皮,客户一句话背后的真实意图,老板改口后的连锁反应。越接近这些地方,越不能迷信一个“很会操作”的家伙。它可以是工具,甚至可以是很猛的工具,但别轻易把它想成接管者。

这也是眼下这波“龙虾热”最有意思的地方。表面看,大家追的是新东西;往深里看,追的是一种代劳幻想。只要产品名字够新,演示够炸,现场跑起来够像真人,立刻就有人把它想成未来入口、生产革命、下一个操作系统。结果往往也很统一:热闹先到,纪律后到;权限先给,治理后补;演示先飞,场景再慢慢找。过不了多久,名字就开始换了,今天OpenClaw,明天QClaw、AutoClaw、Moltbot,主角轮流坐庄,口号层层翻新,可真落到普通人的工作台面上,衡量标准根本没变过:能不能省时间,能不能降差错,能不能让我少做烂活,能不能别顺手把我的系统搞出一个洞。

所以,对待“小龙虾”最稳的姿势,其实一点都不性感。别把它当神童,先把它当实习生;别先给总账权限,先给小任务;别先让它接全部流程,先让它接一个环节;别因为它跑成过一次,就默认它以后都懂;别因为别人装了,你就觉得自己必须立刻养一只。把日志留好,把权限切细,把可逆动作和不可逆动作分开,把关键账户和关键文件隔开,把能自动化的体力活交出去,把需要人拍板的判断留在自己手里。这个姿势不酷,甚至有点扫兴,但它像成年人。

我自己的AI学习路子一直很笨,但也的确很省事。技术名词我没兴趣背,酷炫架构我也不想供着看,谁今天又发明了什么新物种,我顶多看两眼。只要能把日常工作里的重复劳动砍掉一块,能把效率往上推一截,能让我少一点低水平消耗,我就学,我就用。我不靠技术信仰活着,也不靠工具崇拜找安全感。带不来生产力的,再漂亮,跟工作没有关系。

《超新星纪元》真正让人发冷的地方,也在这儿。人类把世界交给孩子,并不是因为孩子准备好了,而是因为自己太急,太怕,太想在倒计时里找一个接盘者。今天一群人急着把电脑交给“小龙虾”,心态里也藏着同一种急。急着摆脱低效,急着追上浪潮,急着证明自己没掉队。可越是这种时候,越要记得一件简单的事:聪明不等于可靠,能干不等于能管,像人不等于懂后果。工具再惊艳,也先看它能不能老老实实替你多挣出一点生产力。其余的,让子弹再飞一会儿。

Big OpenClaw lobster party

Is this big OpenClaw lobster party a little like The Supernova Era by Liu Cixin?

In The Supernova Era, Liu Cixin wrote a very hard idea. After a supernova blew up, the light from it made all people on Earth over age thirteen die within one year. So the adults had to give the world to the children as fast as they could before time ran out. The sharpest part later in the book is not only the disaster. It is that the children really did learn. They took over the adults' skills. At first, the world still worked as usual. But what came next made people feel bad. They got the skills, but they did not get the mind, the sense of limit, the sense of duty, or the patience. So very soon, the world moved from management to play. This idea stayed in my head for a long time, because it was never only about science fiction.

This new wave of “raising lobsters” feels a little like that. Products like OpenClaw are, in simple words, personal AI helpers that run on your own device. They can connect to WeChat, Feishu, Slack, and other doors in. They can use files, run commands, and stay online for a long time. How hot is it? Even “help me install it,” “host machine,” and “one-click hosting” have become business. If you only look at this big scene, it is easy to get a dream in your head: at last, your computer will be run by a tireless, super fast, digital top student that can work by itself.

The most charming part of this dream is that it feels like a shortcut. People are busy. Work is broken into many small pieces. The mind gets tired. Who does not want a smart little worker sitting inside the computer, watching email, replying to messages, sorting files, running steps, filling forms, opening web pages, and also doing a few repeated jobs for you? It sounds too good. It feels like putting an Oxford or Cambridge helper into the machine, one who never rests and never asks for pay. So the people who install it get more and more excited. The people who explain it get more and more carried away. The people watching from the side get more and more worried. If you did not install it, it feels like you are half an age behind.

But the problem is here too. People are naturally too hopeful about a “very smart doer.” Once it does a job well the first time, people quickly lower their guard. Their limits get loose. Then they give more and more rights. They add more and more trust. In the end, they even forget that what it has now is not just the right to use one browser page. It now has room to move across the whole machine. It can touch files. It can touch email. It can run commands. It can connect to accounts. It can even change the system. In human history, every time people gave high power to something they did not truly understand, they later had to pay for that lesson. So while people talk more and more about local agents like OpenClaw, worry about system power, safety limits, secret leaks, and host machine risk also rises.

This is like the children in The Supernova Era. The children were not dumb. In fact, they were very smart. But being smart never makes self-control by itself. Being able to do a thing, knowing which thing should be done, knowing where to stop, and knowing if one move will hurt another part — these are three different things. The same is true for today's little “lobsters.” They can work. Some can work very fast. But “fast” has no moral sense by itself. It also has no business sense. It can save you half an hour of boring work in one minute. It can also break the order you built over half a month in one minute. Delete one wrong file. Send one wrong email. Break one piece of setup. Call one wrong interface. For people, the result is real. For the machine, the result is only that it keeps using the next token.

So the real question was never, “Can little lobsters change the world?” That question is too empty. It sounds too much like a launch event. The real question is this: can it bring steady power to real work, and is that power under control, easy to check later, and easy to trace if something goes wrong? If yes, use it. If not, leave it there for now. Cool things do not pay your salary. Hot search words do not close your monthly books for you. For a finance worker, an operations worker, a project manager, or a sales support worker, the most valuable skill on the desk each day is never “what new thing did I install first?” It is “after I gave this job to it, did my mistakes go down, did my speed go up, did rework get less, and did my result get more steady?”

There is a very easy pit here that people often miss. When a group talks about AI, their mouths talk about tech, but in their hearts they want to hand over their life. They hope it will take over the computer, the process, the judgment, and the trouble. Best of all, they hope it will also take over their tiredness and dislike. But the truly hard part of work was never pressing buttons or opening web pages. The hard part is limits, what comes first, strange cases, unclear orders, taking duty, fighting across teams, the real meaning behind one line from a client, and the long chain of change after the boss changes one word. The closer you get to these parts, the less you should trust a “very good operator.” It can be a tool. It can even be a very strong tool. But do not easily think of it as a taker-over.

This is also the most interesting part of this “lobster heat” right now. On the surface, people are chasing something new. Deep down, they are chasing a dream of having someone else do the work. As long as the product name is new enough, the demo is big enough, and it runs live enough to look like a real person, people quickly call it the door to the future, a work revolution, or the next operating system. The result is often very much the same: the noise comes first, the rules come later; the rights are given first, the control is added later; the demo flies first, and the real use case is found slowly after. Soon, the names start to change. Today OpenClaw. Tomorrow QClaw, AutoClaw, Moltbot. The lead part keeps changing hands. The slogans keep changing too. But when it really comes down to the desk of a normal worker, the way to judge it has not changed at all: can it save time, can it cut mistakes, can it help me do less bad work, and can it stop opening a hole in my system by accident?

So the safest way to deal with a “little lobster” is not sexy at all. Do not treat it like a child genius. Treat it like an intern first. Do not give it your full account first. Give it small tasks first. Do not let it take the whole process first. Let it take one small part first. Do not think that because it did one run well, it will always understand from now on. Do not think that because other people installed one, you must raise one right now too. Keep the logs. Cut the rights into small parts. Split actions that can be undone from actions that cannot be undone. Keep key accounts and key files apart. Give the body-work that can be done by itself to the machine. Keep the judgment that needs a human choice in your own hand. This way is not cool. It may even feel like no fun. But it is like an adult.

My own way of learning AI has always been slow and plain, but it really saves me trouble. I do not care to memorize tech words. I do not want to stand there and admire cool structures. If someone made a new thing today, I may look at it for two seconds. If it can cut down some repeated work in my daily job, push my speed up a bit, and help me lose some low-level waste, then I will learn it and use it. I do not live by faith in tech. I do not use tool worship to feel safe. If it cannot bring real work power, then no matter how pretty it is, it has nothing to do with work.

The truly cold part of The Supernova Era is here too. Humans gave the world to the children not because the children were ready, but because the adults were too rushed, too scared, and too eager to find someone to take over before time ran out. Today, many people rush to give their computers to “little lobsters,” and the same kind of rush is hidden inside. They rush to leave low speed behind. They rush to catch the wave. They rush to prove they are not falling behind. But the more it is like this, the more we should remember one simple thing: smart does not mean safe, able does not mean able to manage, and looking like a person does not mean understanding results. No matter how amazing the tool looks, first see if it can honestly help you make a little more work power. As for the rest, let things fly a little longer.