TED:How I created OpenClaw, the breakthrough AI agent[我是怎么做出这个突破性的AI智能体的.]

讲演人：Peter Steinberger     翻译：鹿之

What’s OpenClaw? I’ve been programming since I was 14. Building software felt like playing a video game. I couldn’t stop. Until I did. I created a company, I poured a decade of my life into it. No venture capital, every weekend, and then I sold the dream. And I felt absolutely nothing.

OpenClaw是什么？我从14岁就开始写代码了，写软件对我来说就跟玩电子游戏一样，根本停不下来，直到有一天我真停了 – 我创办了一家公司，并投入了十年的人生，没拿过风头，我还全年无休，后来我把这个梦想卖掉了，而我的内心毫无波澜。

For three years, I was wondering, did therapy, traveled, I changed countries twice. Nothing, nothing clicked. Now I wake up every morning with everything I was supposed to want and no reason. No reason to get out of bed.

整整三年，我一直在想啊，我做过心理治疗，去旅行散心，还搬了两次国家，但这些都没用，一点感觉都没有。现在我每天早上醒来，我想要的我全都有了，却还是提不起一点劲儿，就是不想起床，没由来的不想。

And then in early 2025, I tried an experiment. I wanted to see what these new AI coding agents are all about. And I had what I can only describe as a “holy shit” moment.

然后在2025年初，我做了一个试验，我想看看这些新的AI编程智能体到底有多厉害，结果看到成果的那一刻，我的脑子里只有一个词 – “卧槽”！

The boilerplate, the plumbing, all the boring parts that was software development, AI could do all of it. The bottleneck is no longer typing. It’s thinking. And thinking was the part I did [for] 25 years. Building software felt like playing a video game again. I was back. I built 44 projects in just a few months.

那些模版代码、底层框架 – 软件开发里所有无聊的部分 – AI都能搞定。瓶颈不再是敲键盘，而是动脑子。然而动脑子这事，我已经干了25年，写软件重新变得像玩电子游戏一样。就这样，从前的那个我又回来了，在短短几个月里，我做了44个项目。

And the latest one was a WhatsApp bot. I put it on my computer, it talks through the apps that you already know. And then I took it on a trip to Marrakesh. You know, just to navigate around, to find restaurants, do translations. And at first, it didn’t really feel right. It felt too much like a tool, not like a friend. You know, too many bullet points, too many tables.

最新做的一个是WhatsApp机器人。我把它装在我的电脑上，它会通过我们已经熟悉的应用程序对话。我带着它去了马拉喀什旅行，你知道的，一开始我就拿它来导航、找餐厅、做翻译，但总感觉哪里不对，这么用起来就只是一个工具而已，不像一个朋友。你懂的，它给的基本都是那种一条条罗列的要点啊还有一堆表格啊。

So I told it, because those modern models, they are so smart, they know what WhatsApp is, they know how people talk. I just had to tell it. And then it felt right. And you know how you talk to friends — at some point I was walking around and I was sending it a voice message. And then I froze because I actually, I didn’t build voice in there. I added support for images, yeah, but even that took hours. So I was looking at the typing indicator, and then the agent responded. And I very vividly remember this situation where I was standing there and I was like, how did you do that? And the agent replied, I’m not kidding you, “The mad lad figured it out on its own.”

所以我提示了它，因为那些现代模型非常聪明，它们知道WhatsApp是什么，也知道人跟人之间是怎么聊天的，所以我只需要告诉它就行，它修正之后这个感觉就对了。你知道你跟朋友是聊天是什么样的吧，然后走着走着我给它发了一条语音消息，发完我愣住了，因为我根本没给它加语音功能，我只加了支持图片功能，是的，但那也花了我几个小时。所以我看着那个“正在输入”的提示，很快它回复了我。我清楚地记得那一刻，我站在原地纳闷：“你是怎么做到的？”然后智能体给了我答案，我没有跟你开玩笑，它说：“这个疯小子自己搞定了。”

And then it walked me through every single step: how it got a message from me but there was no file ending, so it inspected the file, it found that it was audio, but a weird format, so it converted it. And then it was looking for something to translate the audio, but I didn’t have it installed. But then it found an OpenAI key, it sent the whole thing to the server, it got it back, and then it replied. All of that in nine seconds. Can you imagine? I didn’t build any of that.

然后它一步步给我讲了一遍：它是怎么从我这收到了一个没有文件扩展名的消息，然后检查文件，发现是格式奇怪的音频后，就自己转换了格式，之后它找东西来转录音频，可我没有安装相关软件，它就找到了一个OpenAI密钥，把整个文件发送到了服务器，得到结果后，回复了我的消息。所有的这一切是在九秒内完成的，你能想象吗？我压根没写这些功能。

For me, this was the moment where I thought, this is something new. This is not a chatbot. Chatbots give up. Agents improvise. And, you know, I was sold. I wanted to share this. I wanted to like, tell people on Twitter. And nobody really got it. It’s like, it’s almost like you have to experience an agent. It’s kind of hard to explain. It took me a few weeks.And then I did something stupid. And remember, this agent, by default, can do anything that you can do on your computer. So obviously I put it in a public Discord, and I invited random people.

对我来说，就是在这一刻，我意识到这是一件全新的事物。它不是一个聊天机器人，聊天机器人做不到这点，而智能体是会随机应变的。真的，我被彻底征服了，我想分享出来，想在推特上告诉所有人，结果没有人真的能理解其中奥妙，这就好像……就是你必须要亲自体验一下智能体才能明白它的神奇，我很难跟你解释，我花了好几周才意识到这点。于是我干了件蠢事，别忘了，这个智能体在默认情况下，你在电脑上能干啥它就能干啥，所以很自然地，我就把它扔进了一个公开的Discord频道，还随机拉了一群陌生人进来。

And I was looking at it the whole night. People were talking with it, people were having fun with it, people tried to hack it. And when my eyes were like, falling off almost, I exited the process, and I went to bed. Though I forgot — I built a system to be resilient. So while I was walking to the bedroom, the agent happily booted up again and talked to everyone in the world. The next morning I woke up — over 800 messages. I panicked. I pulled the plug, I read every single message just to see if the agent leaked my private life.

我整晚盯着它，大家都在跟它聊天，有人跟它聊得很开心，有人试图黑掉它，等后面我困到不行的时候，我就退出了进程去睡觉了。不过我忘了 – 我设计的是一个灵活的系统，所以在我回卧室的路上，这个智能体又开心地自己重启了应用，并继续和世界各地的人聊天。第二天早上我起来一看 – 800多条消息，我慌了，赶紧拔掉了电源，我读完了每一条信息，就想看看智能体有没有泄漏我的私生活。

Nothing happened. But it could have. But that was the moment where it went viral.

好在什么都没发生，但本来是有可能出事的。也就是在那一刻开始，它火了。

Today, the project is called OpenClaw. It’s pretty much the fastest-growing open-source project. Its mascot is a lobster. It claws into your machine. Jensen Huang calls it the operating system for personal AI. But by far, my favorite quote is from a friend who looked at this statistic and said, “Peter, this is not hockey-stick growth. This is stripper pole.”

现在，这个项目叫Open Claw。他基本上算是目前增长最快的开源项目。它的吉祥物是一只龙虾，他会用钳子伸进你的电脑。黄仁勋说它是个人AI的操作系统。但到目前为止我最喜欢的一句评价是来自一个朋友，他看到统计数据后说：“Peter，这不是曲棍球棒式增长，这是钢管舞杆式的增长。”

And I, I was not ready. You know, when something blows up like this, everything explodes. Hundreds of messages, reporters calling me in the middle of the night, security vulnerabilities. And then the AI company whose model most of my users loved sent me a trademark claim. I had to rename the whole thing while it was taking off. You know, they even tried to push me away from the lobster.

其实我吧…..我根本没准备好，你知道的，当某样东西像这样突然爆火的时候，一切都会失控：数百条信息、记者半夜的电话还有安全漏洞……紧接着，那个我家大多数用户最喜欢用的模型的AI公司给我发来了商标投诉，以致我不得不在项目起飞的时候整个改名。你懂吧，他们甚至想让我把龙虾换了。

Like, I was staring at the message. It was like, it’s not even the same animal. And then they also cut off the model most of my users loved. You know, first the name, then the lobster, then the model. I was that close to just deleting the whole thing. But then I learned what people are building with it.

呃…我看着那条信息，它俩甚至都不是一种动物，然后他们还切断了我家用户最喜欢的那个模型。看吧，先是名字、然后是吉祥物龙虾，接着是模型，我差一点点就把整个项目删除了，但后来我看到了大家用它造出来的东西。

So at ClawCon in Vienna, because yes, we have conferences already — and people wore lobster headbands, I met Stefan and his 60-year-old dad, Gerhard, a beer sommelier who never wrote a single line of code. They connected OpenClaw via Bluetooth, sent it one prompt, and the agent dd the whole 90-minute brew: temperature ramps, hop additions, everything. And then they were like, what are we doing with all this beer? And the agent was, “Let’s make a website.” So they built a website, and then they added payments, and now they have a real product. And almost all of it was just done via the phone.

所以在维也纳的ClawCon – 是的没错，我们已经开线下聚会了 – 并且参会的人都戴着龙虾头箍。我遇到了Stefan和他60岁的父亲Gerhard，老爷子是个啤酒品酒师，一行代码都没写过，他们用蓝牙连上了OpenClaw，只发了一条指令，智能体就自己把整个90分钟的酿酒过程搞定了：温度上升过程、酵母添加量等所有步骤，然后他们问道：“我们该拿这么多啤酒怎么办？”智能体说：“那让我们做个网站吧。”于是他们做了一个网站，并且加了支付功能，现在他们有了一个真正的产品，而这一切几乎全部通过手机完成。

In China, installing OpenClaw is called raising lobsters.Thousands of people were lining up at the Tencent office in Shenzhen to get their lobster installed. Shenzhen even gives out subsidies for people running businesses on OpenClaw. Now if you install OpenClaw on your work machine, at least with the default settings, you might get fired. And then I met an entrepreneur in China who showed me a spreadsheet. Every employee, every day, one task automated by OpenClaw. If you miss too many days, you’re fired. So fired for using it, fired for not using it.

在中国，安装OpenClaw叫“养龙虾”。数千人在深圳腾讯办公楼里排着队等待安装他们的龙虾，甚至深圳政府还会给使用OpenClaw做生意的人发补贴。现在你要是在工作电脑上安装OpenClaw，特别是用默认设置的话，很可能会被解雇，后来我在中国遇到了一个企业家，他给我看了一个表格：每个员工每天都必须用OpenClaw自动化一项任务，如果缺勤太多天，就会被解雇。所以用OpenClaw会被解雇，不用也会被解雇。

After Marrakesh, I thought, this is incredible. And it’s also a little bit scary. How can we make it more scary?

在马拉喀什旅行回来之后，我就在想：这太不可思议了，但这也有些细思极恐。那我们该怎样让它变得更可怕呢？

So I added a new feature, a heartbeat. You know, by default, the agent would only wake up when you send it a prompt. With the heartbeat, the agent would just wake up periodically, check your emails, check your calendar, follow up on loose ends. My initial prompt was simply “surprise me.” And yes, that’s kind of as scary as it sounds. So no large company would ship something like that. But I’m a random builder from Austria. I don’t have a legal department.

于是我加入了一个新功能：“心跳”。你看，在默认情况下，智能体只会在你发送指令时被激活，而有了“心跳”之后，智能体会定时激活来检查你的邮件、确认你的日历还有跟进未完成的事情。我最初的指令只是“给我惊喜”，没错，这听起来有多可怕它就有多可怕。所以没有大公司会发布这种功能，但我是奥地利的一名自由开发者，我哪有什么法务部门啊。

I built this sandbox, this sandcastle, for me, and I made it open-source so other people could play with it and other people could raise their imagination.Imagine putting an agent into a meeting. Not for notes, we figured that out. A bidirectional model that can listen and hear at the same time. Somebody mentions a statistic, a subagent can spin off and check it for you. A decision is made, the agent can send a follow-up before the meeting even ends.

我搭建了这个沙盒、这座城堡，是给我自己玩的，然后我把它开源让别人也能玩，也能放飞想象力。来想象一下，把一个智能体放进会议中，不是为了做笔记（我们已经解决这种问题了），这是一个双向模型，能够同时倾听和接收信息。一旦有人提到了某个数据，立马就能分离出个子智能体来为你核实，做出决定后，这个智能体甚至可以在会议结束之前就把后续通知发出去。

In the future, we’re not just going to have one agent. You might have your work agent, your personal Claw, maybe one for health, maybe one for relationship. And they all should work together in a secure way. Because how did humanity level up? By specializing and collaborating. And agents are about to do the same. Imagine a company, a small business that has ten agents that are all specialized, taking over various parts of the business. We don’t even have a name for what it might become, but we are about to find out.

在未来，我们不会只有一个智能体。你可能会拥有一个用来处理工作的，一个处理个人事务的智能体；也许还有一个管健康的，一个负责情感方面的智能体；它们都应该以安全的方式协同工作。因为人类是如何升级的呢？通过专业化和协作，智能体也是一样的。想象一家公司拥有十个各有所长的智能体，分别负责生意的不同业务板块。我们甚至还没给这种模式作定义，但很快就能见分晓了。

So I created the OpenClaw Foundation, a nonprofit, open-source forever. Because what OpenClaw did for many people was it moved AI from this scary, nebulous thing into something that is fun and useful and maybe a bit weird. You know, lobsters and headbands and beer businesses. Because what we need in the future is more people spending more time with AI to better understand how powerful and transformative this technology really is.

所以我成立了OpenClaw基金会，一个非营利组织，永久开源。因为OpenClaw为很多人做的事都是让AI从一个可怕、模糊的东西变成了一个有趣、有用可能还会有点奇葩的东西。你知道的，龙虾、头箍还有啤酒生意。因为我们在未来需要更多人去花更多时间在AI上，才能更好的理解这项技术有多么的强大和具有颠覆性。

In New York, at ClawCon …Yes, they’re everywhere now. There were thousands of people that were discussing what the lobster did this week. A retiree in Shenzhen who automated their groceries. A teenager in Sao Paulo who built a tutoring business on OpenClaw. Gerhard and his beer machine. None of them are programmers. All of them are builders. Because that’s the real transformation. It’s not a technology. It’s the access. Agents change who can build things, and that door is not closing again. Because when you can prompt a prototype into existence in one hour, anything is possible.

在纽约的ClawCon上…没错，现在到处都在开线下聚会。成千上万人都在讨论龙虾这周又做了什么。在深圳的一位退休老人用它把买菜给自动化了；圣保罗的一个青少年在OpenClaw上创办了一家辅导公司；还有Gerhard和他的啤酒生意。这些人没一个是程序员，但他们都是创造者。而这才是真正的变革，这不是一种技术，而是一个通道。智能体改变了“谁能创造东西”这件事，而这扇门不会再关上了，因为当你能够在一个小时内用提示词生成一个原型时，一切皆有可能。

The next breakthrough can come from anyone, any country, any cafe. When even a burnt-out founder, staring at the screen, wondering if his spark is gone can do something like that. It’s not gone. It’s just waiting. The lobster is loose, and it’s not going back into the tank.

下一个突破可以来自于任何人、任何国家，甚至在任何一个咖啡馆里。当一个燃尽了的创始人，盯着屏幕怀疑自己的灵感是否已经枯竭了，都能做出这样的事，那就说明它就还没有消失，只是在等待。龙虾已经跑出来了，它不会再回到缸里。

Thank you very much.

非常感谢。

Chris Anderson: Come here. I mean. I think I want to say something personal to you. With love, but truth. You really terrify me.

Chris Anderson: 来这边，我想说，我想对你说些真心话。带着爱，但也是大实话 – 你真的吓到我了。

CA: I’m serious. If Hollywood was to ever make a movie in which humanity opened Pandora’s box, and everything went crazy, like, you seriously could be cast as the star character.

CA: 我认真的，如果好莱坞想拍一个人类打开潘多拉魔盒然后一切都乱套了的电影，真的，你绝对能当男主角。

CA: Because the story we’re told is that AI researchers are doing all this great stuff, but they’re taking all these great efforts to ensure safety and make sure nothing bad could happen. You take glee in seeing what might happen if you just put it out there. Like, is any part of you feeling that that’s a little bit reckless?

CA:因为我们听到的故事是，AI研究员正在做各种了不起的事情，但他们也付出了巨大的努力来确保安全，避免坏事发生。可你却在为“如果就这样把它放出去会发生什么”而幸灾乐祸，难道你不觉得这有点鲁莽吗？

Peter Steinberger: I wouldn’t say so. I see my work as a … I see it as a window into the future. Like, in the very beginning, there were all these scary moments. Now we have proper security layers. You can have your sandbox, you can put your lobster into a very small, tiny box and really control what it can do. There are still some issues that we need to figure out, but the fact that so many people want this now will help to figure this out much faster.

Peter Steinberger:我不这么认为，我把我的工作看作是…是一扇通往未来的窗户。最开始的时候确实有很多细思极恐的时刻，但现在我们已经有了合适的安全层。你可以有自己的沙盒，把你的龙虾放进一个非常小的盒子里，并真正去控制它能做什么，尽管仍旧还有一些问题需要解决，但事实是现在如此多的人想要使用它，反而会帮助我们更快地解决这些问题。

CA: So I’m glad you mentioned the security layers. I mean, you’re making a huge bet on human ingenuity using this. This is an incredible tool that suddenly can maximize the power of what any human can do. How many people in your community are taking the safety issue seriously and want to use OpenClaw, for example, to find smarter ways? Just checking whether anything might be going wrong and giving an early alarm to someone or something like that?

CA:很高兴你提到了安全层，我的意思是，你在赌人类的聪明才智，赌大家会怎么用它。这是一个不可思议的工具，可以在顷刻间最大化任何一个人的能力。那你的社群里有多少人在认真对待安全问题，并想用OpenClaw来找到更智能的方法？像是检查是否有任何出错的可能性，并提前给人发出警报之类的事？

PS: Most people are not as reckless [as] — number one, putting it into a public Discord — strongly don’t recommend. Number two, I think I single-handedly increased Mac mini sales by multiple percent. So most people give it their own little Mac mini. Mine’s a little princess. Mine got a Mac studio. It calls it “The Castle.” And that greatly reduces the actual risk because you can only access what’s on that computer. And maybe all your pictures are not there.

PS: 大多数人都不会像我这样鲁莽 – 第一，把它放进公开的Discord – 强烈不推荐；第二，我觉得我一己之力就把Mac mini的销量拉高了好几个百分点。所以大多数人都会给它配备一个小的Mac mini，我的龙虾是个小公主，我给她配的是一台Mac studio，她管那个叫“城堡”，这样一来就大大降低了实际风险，因为你只能访问那台电脑的东西。也许你所有的照片都不在那里。

CA: Well, definitely, if humanity goes down, I’ll be very grateful for at least the rise in Mac mini sales for a period of time. You are amazing. You are amazing, and I think you’re actually right at the cutting edge of whether AI is going to be the biggest boon ever or possibly a serious problem. And, you know, I hope you continue conversations with people here and just help us get smarter on how to do this the right way because it is absolutely incredible what you’ve built. Thank you for sharing so honestly.

CA: 好的，说真的，要是人类就此完蛋，至少我会感谢那段时间内Mac mini销量的增长。你太了不起了，真的太了不起了，我觉得你恰恰就站在那个分水岭上 – AI到底会成为史上最大的福音还是变成严重的问题。希望你能继续跟这里的人聊聊，帮助我们更机智地找到正确的玩法，因为你搞出来的东西实在太牛了。谢谢你如此坦诚地分享。

PS: Thank you so much.

PS：非常感谢。