The AI Coding Agent Revolution, The Future of Software, Techno-Optimism | Amjad Masad, CEO, Replit
发布时间 2025-02-06 12:00:09 来源
Welcome back to the Matt podcast. I'm your host Matt Turk from FirstMark. My guest today is Amjad Masad, the CEO of Replit, one of the most exciting companies in the Red Hot Space of AI coding agents, and also one of the most visible with over 34 million users around the world. We started the conversation with Amjad's fascinating journey as a founder, including the long and winding road to starting Replit from humble origins as a student project in Jordan. I kind of want to start a company, but I don't want to start this company. These multiple attempts to get into YC. Three or four times, I kind of lost count. How an email from Sam Altman changed everything. He was like, hey, I run YC. We're really impressed by what you're doing. Come to the mission.
欢迎回到Matt播客。 我是你的主持人，来自FirstMark的Matt Turk。今天我的嘉宾是Amjad Masad，他是Replit的首席执行官，这是一家在红火的AI编程助手领域里备受瞩目的公司，在全球拥有超过3400万用户。我们从Amjad作为创始人的精彩旅程开始对话，包括从约旦一个学生项目的起步到创立Replit这漫长而曲折的道路。他曾经想创立一家新公司，但对创立这家公司并不那么热衷。为了进入YC，他进行了多次尝试，三四次，以至于都记不清了。直到有一天，他收到一封来自Sam Altman的邮件，改变了一切。Sam对他说：“嘿，我是YC的负责人，我们对你正在做的事情印象深刻，加入我们吧。”
We then dug into Replit's new hit product, Riplet Agent, which enables anyone to build a fully functioning app using just prompts. We're going to build something that's not going to work today, but at some point, some model will come out and it'll work. And that's exactly what happened. And this goes to the implications of AI coding agents on the software development profession and entrepreneurship. Silicon Valley dogma is like ideas don't matter on the execution. In the future, it'll be more like execution doesn't matter on the ideas.
我们接着深入探讨了Replit的新热门产品——Riplet Agent。这个工具使任何人只需通过提示语就能创建一个功能齐全的应用。我们准备构建一个今天可能无法运行的东西，但在未来，总会有某个模型能让它成功。而这正是所发生的。这也引出了AI编码代理对软件开发行业和创业的影响。硅谷传统观念认为，创意不重要，执行才是关键。而在未来，可能会变成执行不重要，创意才是关键。
We closed with Amjad's thoughts on the techno optimism and the future of Silicon Valley. I came to Silicon Valley so we could yell in cell computers, bringing back the heart charging energy of Silicon Valley. And that doesn't mean that we want to alienate anyone. I think it's going to sell track the best talent from all over the world, but it needs to be energetic and it needs to be exciting. Please enjoy this wide encompassing and very fun conversation with Amjad from Replit.
我们最后以Amjad对科技乐观主义和硅谷未来的看法作为总结。我来到硅谷，希望我们能在手机电脑中大声呐喊，重振硅谷充满活力的核心。这并不是说我们想要疏远任何人。我认为这将吸引世界各地的顶尖人才，但硅谷需要活力，需要令人兴奋。请欣赏这段与来自Replit的Amjad之间广泛且非常有趣的对话。
Amjad, welcome. Thank you. Excited to be here. So I thought that would be a fun place to start, actually, your entrepreneurial story, because it's super interesting. First of all, you've been at this mission for a number of years now and second, so many twists and turns and so many lessons to be learned. So let's start with that. So you grew up in Jordan. You studied computer science. That's right. And that's when you started thinking about Replits or Replits. What is Replits in Replits? You know, for example, a DOS or like a Unix command line is a form of a REPL. So it comes from Lisp, actually, in the 1950s at MIT, where you can create the simplest programming environment with like one command, which is read eval print loop.
阿姆贾德，欢迎你。谢谢你，我很高兴能来到这里。我觉得从你的创业故事开始会很有趣，因为它非常吸引人。首先，你已经为这个目标奋斗多年了；其次，其中有很多曲折和经验教训。那我们就从这开始吧。你是在约旦长大的，并且学习了计算机科学。没错，也就是在那时你开始考虑Replit这个项目。那么，Replit到底是什么呢？比如说，DOS或者Unix命令行就是一种REPL。这个概念其实源自20世纪50年代麻省理工学院的Lisp编程语言，那时可以通过一个简单的命令创建一个编程环境，这个命令就是读-执行-打印循环。
So read reads the command from the command line. Eval evaluates like a Lisp string, print prints out the results, and then goes back to the start. So this is like the simplest IDE in the world, right? And so when I was building, so the way sort of Replit was born, I was going to school in Jordan, saying computer science. And I didn't have a laptop. So every time I would want to do some homework, I would have to reinstall the development environment. And that was like the, you know, it was so annoying, right? You have to download gigabytes of software and packages. And there's always something that goes wrong. And so I was like, you know, I'm using everything in the browser. At the time, Google was putting docs and Gmail in the browser. This is 2008, we're talking about it. Yeah, actually, the eight, nine. Chrome came out, V8, the JavaScript engine, you know, the web was becoming a real application platform. So I was like, why can I code in the browser?
所以，"read" 命令从命令行读取指令，"eval" 像在Lisp中一样评估字符串，而"print"会输出结果，然后返回开始。因此，这就像世界上最简单的IDE，对吧？我在构建它的时候，Replit就是这样诞生的。当时我在约旦上学，学习计算机科学。但我没有笔记本电脑，每次要做作业时，我都得重新安装开发环境，这非常麻烦。你得下载好几GB的软件和包，而且总是有问题出现。当时我想，既然我在浏览器中使用一切功能，Google已经在浏览器中推出了Docs和Gmail，我们说的是2008年到2009年。Chrome浏览器和V8 JavaScript引擎也出来了，网络正逐渐成为一个真正的应用平台。所以我想，为什么不能在浏览器中编写代码呢？
And then I sort of Googled it. And it was surprised that no one had built it. So, okay, I'll build it, how hard could it be? And actually got a prototype pretty quickly, because you can eval JavaScript in the browser, right? So I put a text box, I put a button, you write a little bit of JavaScript, you run it, and it alerts like the results. And I gave it to my friends, and they were all like playing with it and really excited about it, started adding features, like sharing and saving and things like that. And I was like, okay, I need to add more languages. You're studying computer science, you're probably gonna be using a lot of different languages, Python, Java. And that's when I head into problems, because the browser can only run JavaScript.
然后我有点随便地在谷歌上查了一下，发现居然没有人做过这个。于是我想，那好吧，我来做，能有多难呢？实际上很快我就做出了一个原型，因为在浏览器中可以执行JavaScript代码。我放了一个文本框和一个按钮，你可以写一点JavaScript代码，然后运行，结果就会用弹窗的方式显示出来。我把这个东西给我的朋友们，他们都很喜欢，玩得很开心，还很兴奋，并且开始为它添加功能，比如共享和保存等等。于是我想，我需要添加更多的编程语言。你如果学习计算机科学，可能需要用到很多不同的语言，比如Python、Java。然而，这时候我遇到了麻烦，因为浏览器只能运行JavaScript。
And this became like a multi-year sort of research effort, sort of culminating in, you know, having a breakthrough where my friends and I were able to compile a bunch of languages into JavaScript. We used a library from Mozilla. That library became Wasm, which runs sort of native programs in the browser, Figma has built on that, a bunch of really modern applications, wouldn't be possible with that. But at the time, it was just a library. And there wasn't anyone using it, and so we were like the first to kind of use it to run all these programs in the browser. And put it up on GitHub and went viral and hack and news and a bunch of other places. And it was really surreal because like, yeah, I didn't expect it to be this big. And people were talking about it in conferences, Brendan Eich, the CTO of Mozilla at the time and the inventor of JavaScript, like tweeted about it. And that was like the highlight of my life at the time. Yeah, it was a huge deal, Brendan. Yeah. And then one day, you know, as always, I'm like lurking and hacking news.
这变成了一项历时多年的研究工作，最终，我们取得了一项重大突破：我和我的朋友们能够将多种编程语言编译成JavaScript。我们使用了Mozilla的一个库，这个库后来发展成了Wasm，它能在浏览器中运行类似原生程序。像Figma这样的现代应用都是基于此，它们的实现若没有这个技术是不可想象的。但在那时，这只是一个库，没有人使用，所以我们成了首批利用它在浏览器中运行各种程序的人。我们把项目上传到了GitHub，结果在Hack and News等多个平台上迅速走红。这种经历真的很梦幻，因为当时我没想到它能取得这么大的影响。人们在会议上谈论它，Mozilla的首席技术官也是JavaScript的发明者Brendan Eich还发了推文提到此事。这成为了我人生中的一个高光时刻，是件大事啊，Brendan。然后有一天，像往常一样，我在Hack and News上潜水。
And I see, you know, this startup code academy, like maybe a couple months after I open sourced all the infrastructure that we built. And like I click on the site and it was like going viral, right? And they built this like interactive way to learn how to program. Because learning how to program at the time is like, it's like textbook and all of that. But they built this thing where it's like this, you know, like the course kind of talks to you and like gives you these exercises and you have to code with it. It was a very novel idea. And at the time it was like the hype was around MOOCs, massively online, open online courses. I had a feeling that they're using myself. I just like the way the system worked. And then I tried to like, I kind of look behind the scenes and I valid it. So I put a comment on hacking news. It was like, oh, you're using this library, my library. And then two days later I get an email from the CTO, Ryan Obinski. And he's like, I'd love to chat with you. Amazing. And I don't think they knew I was like a kid in Jordan. So we started talking and.
我看到这家初创公司——代码学院，大概是在我开源了我们开发的所有基础设施几个月后。然后我点击他们的网站，发现它已经开始疯传。他们构建了一种互动的方式来学习编程。当时学编程通常就是通过教科书和类似的材料，但他们创造的这个东西让课程好像可以和你对话，并给你提供练习，你需要亲自编码。这是一个非常新颖的创意。当时，大家都在关注MOOCs，即大规模开放在线课程。我有预感他们在使用我的工具。我喜欢他们系统的运作方式。于是我试着进行一些调查，证实了我的猜测。于是我在Hacker News上发表评论，说“哦，你们在使用我的库。”过了两天，我收到了一封来自首席技术官Ryan Obinski的邮件。他表示很想和我聊聊。太棒了。我想他们不知道我只是一个住在约旦的孩子。于是我们开始交谈。
And how'd you graduate? And somebody who had Yahoo in Jordan, right? Yeah, so at that time I had graduated. My time at the university was pretty tenuous because I just like, I didn't like sitting in class all that much. Initially, I also, and I'm sure we'll get into more into that. But I was reading a lot of sci-fi. I was like really into AI and really into, how do you build a brain? And I was like reading a lot about cognitive science and neuroscience. There was into way to study that in Jordan, but that was like a personal interest of mine. And I felt like we're on the cusp of like, you're solving intelligence. And I felt like computers will be able to program themselves.
你怎么毕业的？在约旦还有人使用雅虎，对吧？是的，当时我已经毕业了。在大学时，我的经历有些波折，因为我不太喜欢上课。一开始也是如此，虽然我们后来会谈到更多细节。我当时读了很多科幻小说，对人工智能非常感兴趣，还对如何构建一个“大脑”充满好奇。我大量阅读认知科学和神经科学方面的书籍。在约旦没有途径可以专门研究这些，但这些都是我个人感兴趣的领域。我感觉我们正处在解决智慧问题的关键时刻，我觉得未来电脑将能够自己编程。
So I was like, okay, if computers can program themselves, so what can I do with computers? And I was like, okay, well, someone still needs to build these computers and run them. So I went into electrical engineering, computer engineering, I suppose, to programming. And then like two or three years in, I'm like, you know, I don't think machines are gonna be coding themselves anytime soon. That would be crazy. Yes. Well, and I just switched back to computer science. But yeah, it took me six years. And then I graduated. I was like hacking all the time. And if I may ask, and I'm asking as a, you know, French guy who's been in the US for like decades, at this point and still sounds the way I do, you sound very American. Like how did that happen? Well, we're the YouTube generation, right? Is that right? It's like, you know, YouTube came out what, 2005? Yeah. You know, I started watching it then. And so, you know, I think-
所以我当时在想，好吧，如果计算机可以自我编程，那我能用计算机做些什么呢？然后我就意识到，嗯，总还是需要有人来制造和运行这些计算机的。因此，我选择了电气工程和计算机工程，接着又学编程。
大概学了两三年，我开始觉得计算机短时间内不太可能自我编程，那会太疯狂了。于是，我转回了计算机科学。但整个过程花了我六年的时间才毕业。我那时一直在研究编程。
如果可以问一下，我是一个在美国生活了几十年的法国人，我的英语还是带有法国口音，但听你说话很有美国味，你是怎么做到的？我们是“YouTube一代”，对吧？YouTube差不多是2005年出来的吧？从那时起我就开始看了，所以我想……
So it's not like you went to American school in Jordan or- No, I went to British schools. I was in downtown British. Yeah, because- It's fascinating. So that was YouTube. So Jordan was like a British colony, right? Yeah. And so the influence was more British. And so we did IGCSE and all levels and all levels and things like that. But it was also lame to sound English. Although now it's cool, I think. Yeah. But I did, you know, I did to be a little serious. I did work on the accent a little bit because- I don't know, I guess I just want to fit in. And I thought it's- You know, I really want to be American.
所以你没有在约旦的美式学校上学吗？- 不，我上的都是英式学校。我是在市中心的英式学校。- 哇，这很有趣。所以约旦以前是英国的殖民地，对吧？- 是的，因此影响更多来自英国。我们学的是IGCSE课程，还有O-Levels这样的内容。但说英语口音在那个时候不太受欢迎，虽然我觉得现在可能很酷。我也确实认真练习过口音，因为——我不知道，我只是想融入一些美国文化。我真的很想成为像美国人一样的人。
And I've always kind of wanted to be here and wanted to be in Silicon Valley. I watched- There's a movie, this like very low budget movie called Pirates of Silicon Valley. And it's about the story of Steve Jobs and Bill Gates. And it's like very overdramatized story and half of it is wrong. But it like really inspired me. And I was like, oh, Silicon Valley is like a great place. You can like fight and yell at each other and like build computers. And it's awkward. What could be better? Exactly. So, you know, we grew up with a lot of American culture and there was like a very, you know, interest of mine to kind of really be here.
我一直想来这里，想去硅谷。我看过一部电影，好像是一部低成本电影，叫《硅谷海盗》。这部影片讲述了史蒂夫·乔布斯和比尔·盖茨的故事，情节非常戏剧化，而且一半内容是不真实的。但这部电影真地激励了我，让我觉得硅谷是个好地方。在那里，人们可以争论、吵架，甚至一起制造电脑，虽然有点尴尬，但还有什么比这更好的呢？确实如此。我们是在美国文化的影响下长大的，而我一直对这里非常感兴趣。
So, good Academy was the way in? Yes. But I was also like very, very passionate about this. Because like when I build like this open source version of Replit, it was about education. It was about me, you know, being able to, you know, learn to code and code on the go without needing an expensive computer. And so I was like, OK, you know, this helped me so much. I think it's going to help a lot of a lot of people in the world. And I had, you know, this theory that like programming could be the thing that you democratize access to kind of wealth and entrepreneurship.
所以，好学校是你的起点吗？是的。但我也对这件事非常非常有热情。当时我建立Replit的开源版本，是因为教育，是为了让我自己能够学习编程，并且随时随地编程，而不需要昂贵的电脑。所以我觉得，这对我帮助很大。我相信它也会帮助世界上很多很多人。我有一个想法，就是编程可以成为一种方式，来让财富和创业的机会更加平等地被人们获得。
And I think it is playing out that way. And so I was really passionate about joining Code Academy. And at the time you saw a ton of people come online in Africa and India. And then Android had just started going and it felt like everyone's going to be able to get a computer or a phone. And I always want to make programming work on the phone. And I've tried really hard to make Code Academy work on the phone. Now, Replit, we have a mobile app actually. Yeah. So good Academy actually to these days, I don't know. It's his pronounced Code Academy.
我认为事情的发展确实如此。所以，我非常热衷于加入 Code Academy。当时，你会看到大量非洲和印度的人们开始使用互联网。然后，安卓系统也刚刚开始流行，感觉每个人都会拥有电脑或手机。我一直希望能够在手机上进行编程，并努力让 Code Academy 能在手机上使用。现在，Replit 其实已经有一个移动应用程序了。对了，如今我也不知道好 Academy 究竟应该怎么念。
It was it was it was spelled Code Academy, but pronounced. No, no, it's pronounced Code Academy. It's pronounced Code Academy. OK, but everyone everyone like. Yeah. Well, anyway, it worked out for for Zach. If you remember at the time, it was like the hot thing to do to a startup is misspell your company. Yes. I ran out of letters for the comma. Yeah, exactly. And you mentioned Android, but that that was your next stop. When you joined Facebook, you joined the Android team.
它的拼写是“Code Academy”，但发音上却被有意误读了。没有，没有，它的发音应该是“Code Academy”。没错，大家都是这么以为的。嗯，不管怎样，对扎克来说，这个选择还是奏效了。如果你还记得，当时对于创业公司来说，一个流行的做法是故意拼错公司名字。是的，我没有更多的字母来加逗号。对，确实如此。你提到了安卓，那也是你接下来的工作。当你加入 Facebook 时，是加入了安卓团队。
Yes. So I was inspired by a lot of the open source work that Facebook was doing, like around React.js. And I was very early to that. But I was also inspired by Zucks vision for internet.org. Do you remember that? Yeah. So the idea, which you know, Elon ended up doing it, but the idea is like beaming internet from the sky, essentially. And and again, this like fits into this vision that I had of more people connecting their internets. And I felt that was like such an important thing.
好的，我受到了Facebook在React.js等开源工作上的很多启发，我加入得也很早。同时，我也受到扎克伯格对internet.org愿景的启发。你还记得这个吗？是的，这个想法就是从空中传输互联网信号，这一理念最终被埃隆·马斯克实现了。这其实也契合了我希望更多人可以连接互联网的愿景。我觉得这是一件非常重要的事情。
And so when I went to Facebook, I was like, OK, I'm either going to work on and at the org or or work on on mobile. So when I got there, you know, the the way Facebook worked is like you go through bootcamp for like three months and then you do team selection. Because I don't know what I was going to work on, but I picked up Android tasks. And did did really well at them. And I picked really the hardest task that you shouldn't pick as a as a sort of like a boot camper at some point, I broke the Android app for like 10 million, 10 million users, the beta users.
我去Facebook的时候，我心里想着，好吧，我要么就为组织工作，要么就做移动端的。在我到了Facebook之后，我发现那里的工作方式是：你需要经过大约三个月的新员工培训，然后进行团队选择。因为我不知道自己具体要做什么，所以我选择了Android的任务，并且做得很好。我选择了本不应该由新人处理的最困难的任务。结果有一次，我弄坏了Android应用，影响了大约一千万的测试用户。
And and everyone was surprised that I was able to do that as someone bootcamp, because I was taking a lot of risks to, you know, move fast and break things. And but I couldn't get into the internet or team. And and because we were in New York, almost all the teams wouldn't wouldn't take me. They were like, you have to move to the Bay Area. And I wanted to stay here. So I ended up on the photo steam for about six months. And that was like the only team that would take me essentially. And then on the nights and weekends, I was contributing to react. And I was interested in this team that was that was trying to make mobile programming more accessible.
作为一个训练营的学员，我能够做到这一点，着实让大家感到惊讶，因为我在努力快速行动、敢于冒险。然而，我无法加入互联网或其他团队。由于我们在纽约，几乎所有的团队都不愿意接纳我。他们都说我必须搬到湾区去，但我想留在这里。于是我最后加入了一个摄影团队，待了大约六个月。这几乎是唯一愿意收留我的团队。在晚上和周末，我则在为React做贡献，并对一个致力于让移动编程更易接触的团队很感兴趣。
Because Java was this like heavy, you know, brutal thing. You know, every time you change the line of code, you need to recompile the entire app. So this project called Catalyst at the time, which became react native. And I was the first to I was one of the first like three or four years ago. I was one of the first like three or four people to work on it. And I was tasked by the founder of react and the founder of react native Jordan Walk to build the JavaScript tool chain for react native.
因为 Java 当时被认为是一个比较沉重、复杂的东西。每次你更改一行代码时，都需要重新编译整个应用程序。当时有一个名为 Catalyst 的项目，后来发展成为 React Native。我大约三四年前是最早参与这个项目的人之一。React 和 React Native 的创始人 Jordan Walk 当时给我分配的任务是为 React Native 构建 JavaScript 工具链。
So I started building it here initially on my own and they told me like go out and build a team. So I was here in New York kind of interviewing who's who and like the JavaScript community. And I met a lot of friends that I'm still in touch with today, which was a big thing. It was like at the time there was what was it called? Glitch and Joel Spoles key and doing this big thing around JavaScript.
起初，我开始在这里独自开发这个项目，他们告诉我说要去组建一个团队。因此，我就在纽约面试在JavaScript社区中有影响力的人。我结识了很多朋友，直到今天我们仍然保持联系，这对我来说非常重要。当时有一个叫Glitch的平台，还有Joel Spolsky，他们在进行一项有关JavaScript的大型项目。
Right. Yeah. Yeah. But actually I ended up moving to the Bay Area because I actually, you know, we were just talking about it. The talent. There wasn't a lot of talents here. There's like very competitive. And so I moved to the Bay Area, built the team there and released a react native a year later. And it was this like super viral thing and a ton of people used it.
对，嗯，对。不过最后我搬到了湾区，因为就像我们刚才讨论的那样，这里人才不多，竞争却很激烈。所以我搬到了湾区，在那里组建了团队，一年后发布了一个基于React Native的项目。这个项目非常火爆，有很多人使用。
So to continue the thread, you had started a version of replits back in Jordan and you effectively put it to the side for a bunch of years. And so the next step in the journey is that you decided to pick it up again. What was that that journey to decide to making that decision? You know, I sort of. I sort of wanted to to let go. Like I was like, I did this thing. You know, I kind of like burnt out while working at Code Academy. And I wanted to do to do other things and.
所以，继续这个话题，你在约旦的时候曾经开始了一个类似 Replit 的项目，但实际上搁置了好几年。接下来的旅程就是你决定重新拾起这个项目。是什么促使你做出这个决定的呢？我有点想放手不做了。当时我觉得，自己已经做过这个项目了。在 Code Academy 工作的时候有点精疲力竭，所以我想尝试做其他事情。
But like actually the project kept growing. It was like this open source project and we had like a demo page. And I keep getting emails. And it looked at Google Analytics and like we had like still 10,000 users. And like this thing is like is roof is to die. Yeah, it was to die. So let me fix it a little bit. So I started fixing fixing it. And pretty quickly it started growing again.
实际上，这个项目一直在增长。它是一个开源项目，还有一个演示页面。我不断收到电子邮件，看了下谷歌分析，发现我们仍然有1万用户。这东西简直是不可思议。是的，它的确不可思议。所以我决定稍微修复一下它。我开始修复它，很快它又开始增长了。
10,000, 20,000, you know, 50,000, a hundred thousand users. And at the time I moved the code execution from the browser to the back end. At the time Docker had just come out. And I felt like if we move it to the back end, we can have packages and servers and I can make it into a full development environment. And and so it was costing me a ton of money at the time to just like keep going.
翻译成中文，这段话的意思是：
"1万，2万，你知道的，5万，10万用户。在那时，我把代码的执行从浏览器转移到了后台。那时，Docker刚刚问世。我觉得如果我们把代码转移到后台，就可以有软件包和服务器，并且我可以把它变成一个完整的开发环境。因此，当时为了持续发展，这让我花费了大量金钱。"
It was like the user growth was was insane. I was like, OK, maybe I have to sell it. And so I tried to I tried to sell it. And then I was like my time at Facebook was coming to a natural end. I was like, I kind of want to start a company, but I don't want to start this company. Because I'm already done with it. But but it kept growing and people kept really getting excited about it. And I was like, OK, OK, I guess I guess I need to start this company.
用户的增长简直是疯狂。我心想，好吧，也许我该出售它。所以我尝试去出售。但是与此同时，我在Facebook工作的时间也快结束了。我有点想创业，但我不想成立这家公司，因为我对它已经厌倦了。可是它不断地增长，人们对它也越来越感兴趣。我想，好的，好吧，我想我需要成立这家公司。
And I looked at the landscape of online IDEs and there was like there was a ton of attempts at it. The problem is no one had built like a sort of a web native experience. And the way I define web native experience is that it should be able to load really quickly. It should be able to be URL addressable. So like when I send you your all, you should be able to see my project. It's just like you doing docs or a figment. Yeah. Yeah. And should we multiply her by default? It should be a real time in none of those existing IDs where we're doing that. So incredibly complicated, technical problem. Right. That's why that's why it took four years to be able to figure out before they started taking.
我查看了在线IDE的现状，发现有很多尝试。然而问题在于，没有人构建出一种真正“原生于网页”的体验。我所定义的“网页原生体验”是指：它应该能够快速加载，应该可以通过URL访问，这样当我发送给你一个URL时，你就能立刻看到我的项目，就像你使用在线文档或者Figma一样。此外，它默认应该是多人协作的，并且实时更新。而当时现有的这些IDE都没有做到这一点。这是一个极其复杂的技术问题，这也是为什么花了四年时间才弄明白，在开始之前需要这么长时间。
Yeah, Figma was in stealth over four years. Exactly right. And so yeah, we started Reploed and officially in 2016. Because this thing that just wouldn't die and like kind of exploded itself onto the world and I was like, OK, I was still passionate about the idea, but I just I just, you know, I knew it was going to be super hard.
是的，Figma在隐秘开发阶段持续了四年。这很正确。所以我们在2016年正式启动了Reploed项目。因为这个项目就是无法止步，自己一下子就爆发到了世界面前。当时，我心想，虽然我对这个想法依然充满热情，但我知道这将会非常艰难。
And then in the, you know, in the folklore of entrepreneurial stories, like this, the that's a YC part of the story. And then tell us, tell us a story of trying several times and then getting in and like just just tell us how that went down. So we, we, you know, we try to get in a YC, even before I quit my job at Facebook, because that's that's what you would do to do risk it, essentially. And really why I see do risk a lot of entrepreneurship.
在创业故事的传说中，这是一个关于YC（Y Combinator）的部分。请给我们讲讲你尝试了几次最终进入YC的故事，告诉我们整个过程是怎么发生的。我们尝试进入YC，甚至在我辞去Facebook的工作之前就已经开始了，因为这样做可以降低创业的风险。而YC确实在很大程度上降低了创业的风险。
Mostly for the good. I think something, you know, I think you need to take, I think some amount of risk is good. But we wouldn't get it. We weren't getting in. And I said, when I took the plunge, it was it was just us and we were able to raise money from Bloomberg beta, a bit of capital. And over the next two years, that was your pre-c. The pre-c, yeah. And that was so that was pre-YC pre-YC. So that was like a like a really a pre-c. The kind of the OK.
大部分是好的。我觉得呢，你需要承受一定程度的风险是有益的。不过，我们当时没有机会。所以，我决定冒险，有了这个决定后，我们成功从Bloomberg Beta筹集了一些资金。接下来的两年里，这就算是我们的前种子轮了。是的，前种子轮。在加入Y Combinator之前，这真的是一个前种子轮，差不多就是这样。
OK. And we were like grinding for for two years, building a lot of those, a lot of this infrastructure that that we just talked about. But it was like pretty lonely and we weren't like really able to hire a lot of engineers. We weren't part of the like, you know, Silicon Valley sort of hype or buzz or also the, the, you know, Dev tools at the time were kind of like not an exciting space, right? That was like pre GitHub acquisition. And GitHub acquisition kind of really exchange everything. Yeah. So, you know, I got used to writing about what we're, what we're doing. And so I tried to write like something every couple of weeks about some of the engineering challenges we're facing. And I thought it could be a good recruiting tool, marketing tool. And Hackenoo's really liked my writing. And so it would always make it there. Now they hate me, but that's the, that's what, that's the kind of natural evolution of Hackenoo's.
好的。在过去的两年里，我们一直在努力工作，建设我们刚提到的很多基础设施。但是，这段时间有点孤独，我们并没有能够招聘到很多工程师。我们并不属于硅谷那种热门或躁动的圈子，而且当时的开发工具并没有那么吸引人，那还是在 GitHub 被收购之前。 GitHub 的收购确实改变了一切。
因此，我习惯于写一些关于我们正在做的事情的文章，所以我尝试每隔几周写点关于我们在工程中遇到的挑战。我认为这可能是一个很好的招聘工具和营销工具。而 Hackenoo's 非常喜欢我的文章，因此它总是会出现在那里。现在他们不喜欢我了，但这就是 Hackenoo's 的自然演变过程。
So you tried several times to get how many times? Three or four times. I kind of lost count. Three or four times. And the reasons why you didn't get in was, was the room reached out to us. It was like, I would get your automate email. Yeah. And at the time you are doing a lot of things that were not necessarily, let's say, popular in Silicon Valley, like you're selling to students effectively and hobbyists. So the, the only people that would like give us money, we had a lot of developers use the platform, but they're like, I'm not going to pay for it. It's just, it's a good, it's like a useful thing, but, you know, probably not going to kill me if I don't have it. Yeah.
你试了好几次，总共是多少次？三四次吧。我有点记不清了。三四次。而你没有进入的原因是，有人联系我们了。就像，我会收到你的自动邮件。而当时你做的很多事情在硅谷并不算热门，比如你主要面向学生和爱好者销售。因此，唯一愿意给我们钱的人中，很多是开发者，但他们觉得这个平台有用，但并不愿意付费，因为即使没有这个东西，他们也不会有什么损失。
But the people that were really dependent on it were the students. Again, like not really surprising because, because that's what I built it for essentially. And hobbyists and, and, and, and people who wouldn't know how to set up a programming environment essentially. And, and so we started monetizing that. So we started selling to schools. We sold the API as well and other, other companies built, built on top of it to kind of sell to schools as well. But selling to the education market is not necessarily the number one thing that gets excited. It's horrible. Excited about it. I mean, I understand it. It's, it's, it's really, I mean, I, you know, I have tons of respects for education entrepreneurs. It's just like, it takes a lot of grits and like, to be about the mission. Yeah. Okay. We get checks for like $50. And then the, so to close on the YC thing. So how did, how did the, how did the, how did it come? How did it eventually, how did you get in? How did it come about?
但是，真正依赖这个系统的人是学生。这并不令人意外，因为这本来就是我为他们开发的。另外还有业余爱好者以及那些不会搭建编程环境的人。所以，我们开始从中盈利，向学校出售产品。同时也出售API，其他公司基于我们的产品进行开发，也向学校销售。然而，向教育市场销售并不是一件让人兴奋的事情。尽管我理解它的重要性，并且我非常尊重投身于教育事业的创业者，但这需要极大的毅力和对使命的执着。我们有时收到的支票金额只有50美元。至于Y Combinator的事情，最后是怎么进的？这其中是怎么回事？
So, so Paul Graham had already retired. If you remember, he passed the mantle to, to Sam in like 2015. And so 2017, 18, he was Sam Altman. Now the CEO of OpenAI for anyone that doesn't follow the YC history. Yeah. You know, by bit, yes. And so Paul Graham, PG, the founder of YC. Reads at the time reads Hackers every day. Um, and so he found one of my posts. Um, and, uh, and like, uh, you know, a few days later, you know, wake up one morning. And I have a message from Sam Altman. Uh, it's probably still in my DMS. I could look it up, but, um, yeah, I probably want to save that one. Yeah. It was like, Hey, um, I run YC. Really impressed by what you're doing. Uh, it's like, dude, I know how you are. Uh, it was like, uh, you know, I'd like to meet you. And, um, so I, uh, he's like come to the mission. I was like, that's not YC's office. So I go to the mission and it was, uh, your link on the right. And it was, um, opening eye on the left. And so I went to open and I was like, why am I here? Yeah, but both part of the Elon empire at the time. Yeah. I mean, uh, Elon just, uh, I had the investors then. No, well, Elon just retook the building. I'll retook that building. Okay. So now I see, okay. Okay. Okay. Fantastic. That's hilarious.
保罗·格雷厄姆已经退休了。如果你还记得，他在大概2015年把权杖传给了萨姆。那么在2017、18年的时候，萨姆·阿尔特曼成为了OpenAI的CEO，给那些不熟悉YC历史的人解释一下。是的，你知道的。保罗·格雷厄姆，也就是PG，是YC的创始人。他当时每天都在看《黑客与画家》。然后，他看到了我其中一篇帖子。几天后，我早上醒来时，收到了萨姆·阿尔特曼的一条消息。那条消息可能还在我的私信里，我可以找出来，但我可能想留着它。内容大概是这样的：“嘿，我负责YC，非常对你做的事情印象深刻。”就像是，“伙计，我知道你是谁。”然后说，“我想见见你。”所以他说让我去Mission。我当时想，那不是YC的办公室啊。我去了Mission，发现右边是你的链接，左边是OpenAI。我走进OpenAI时心想，我为什么在这里？不过在当时，这两者确实都属于埃隆的帝国。是的，埃隆只是当时的投资者。不，好吧，埃隆又重新拿回了那栋楼。哦，我明白了，好吧，这太搞笑了。
Yeah. I think there's some, uh, you know, there's some, uh, sort of, you know, message underneath. Yeah. That's like, we're taking it. Um, so, um, so he's like, um, you know, uh, PG found your, your company. And, and he said, he said it was something that he was, he's been looking for for a long time. It's like, why don't you go visit him in, in the UK? And I'm like, well, yeah, let me get my private jacket. At the time I was an American to get a visa as a Jordanian to the UK. What have taken me another year? Yeah. I was like, how about I email him first? And, um, and so I started this email relationship with Paul, by the way, PG doesn't hop on, on zoom calls or very early does. That's right. Yeah. Yeah. I really, um, so we started this, uh, which was awesome because I'm, I'm getting personalized PG essays.
好的。我认为这里面有一些隐藏的信息。是的，就是我们正要采取行动。所以他说，嗯，PG找到了你的公司。他说那是他一直在寻找的东西。就像，他建议说你为什么不去英国见见他。我当时觉得，好啊，那我就开着我的私人飞机去。然而，当时我还是美国人，要作为一名约旦人获取去英国的签证可能需要另一年。所以我想，不如我先给他发封邮件。于是，我开始与Paul进行邮件交流。顺便说一下，PG很少参加Zoom会议。是这样的，对。我真的很开心，因为我在收到专属的PG文章。
Yeah. It felt such an honor. And we just talked about, um, you're programming what's hard about it. And he had told me that one of the insights from via web, via web is the online store that PG built in Solter, Yahoo. Was like, Oh, people want to customize the stores and they want to put it on a piece of code. And he always imagined like if he built like a browser environment where you can code and like easily generate apps that that could be amazing for entrepreneurship. And so we really connected on that. And that's always been my passion. It's like, how do you get more people to participate in the amazing wealth creation engine? That's the internet. Um, and we also connected on the technical aspect of it and programming languages. And, and I really enjoy the emails and he's like, look, I'm going on a trip now, but, um, you know, why she's about to start. I was like literally a few days later, where I see you would start. Why don't you join the batch? And then he went off flying.
是的，真的觉得非常荣幸。我们谈到了编程的难点，他告诉我他从Viaweb获得的一些启示。Viaweb是PG在出售给雅虎前构建的一个在线商店。比如，人们想自定义他们的商店并把它变成一段代码。他总是想象，如果他能创建一个可以编程的浏览器环境，并且能轻松生成应用程序，那对创业来说将是很棒的。我们在这个话题上有很深的沟通，这其实一直是我的热情所在——如何让更多人参与到互联网这个奇妙的财富创造引擎中来。我们也在技术层面和编程语言方面建立了共鸣。我非常享受这些邮件的交流。他说：“我现在要去旅行，但几天后Y Combinator就要开始了，为什么不加入这一期呢？” 然后他就飞走了。
So I forwarded the message to sound was like, Hey, you know, PG saying we should do I see. And my co-founder, uh, my wife, she's, she's still with the company. Had a, had a design. Um, you have a, you have a brother as well. My brother is a family affair. My brother, uh, leads a portion of engineering.
所以我转发了那条消息，意思是：“嘿，你知道吗，PG 说我们应该这样做，我明白了。” 我的联合创始人，我的妻子，她还在公司工作，并且负责设计。你也有个兄弟，我的兄弟是我们公司的一部分，他负责领导一个工程团队。
Um, and, uh, you know, I was like, you know, should, should we, should we do? I see. And it's like, you know, we got rejected all these times. We're starting to make money. Should do any of the given I was having percent. And. And so I, um, I emailed a salmon poll and just say, like, we'll do I see as long as we get access to you to, we don't want to go in and like become some random startup.
嗯，你知道的，我当时就在想，我们应该做些什么呢？我们之前被拒绝了很多次，但现在开始赚钱了。在这种情况下，我是不是应该采取行动呢？于是，我给Sam和Paul发了一封邮件，说我们会参与，只要能得到他们的支持和建议。我们不想进去之后就变成一个默默无闻的初创公司。
And they said, yes. Um, it was like, I was like, okay, we'll do it. And then Sam is like, Hey, just a formality filling this application. I fuck, you know, I've done this application so many times. I was like, you know, it was kind of like a triggering some trauma. I mean, it's like record the video and like talk about it and send it. And you never hear back.
他们说，好。嗯，就像，我当时就想，好吧，我们就做吧。然后山姆就说，嘿，这只是个形式，填这个申请表。我心想，该死的，我已经填过这个申请表很多次了。我觉得，这有点触发了一些创伤。就是说，要录个视频，谈谈这个事，然后发出去。但你从来都不会收到回复。
So we did this application very quickly. And the next day I go to YC, it was literally the kickoff day. So we would have a dinner or sitting outside. I think we're literally the last interview. So we sat outside for two, three hours and you see this late interview startups that are going in and out. Um, and so they, they call us in and I go in and it was, uh, Gustav and Jared and all this amazing YC partners.
我们非常迅速地完成了这个申请。第二天，我去了YC，那天可以说是正式启动的日子。我们要么吃晚餐，要么坐在外面。我记得我们真的是最后一个面试的。所以我们在外面坐了两个小时左右，看着那些晚面试的创业团队进进出出。然后，他们叫我们进去，我进去时，见到了Gustav、Jared和其他那些了不起的YC合伙人。
And at the end, it was Michael all the time he was the CEO. Uh, Sanal, what was the present? He was the CEO. Michael is a big guy. And so I, you know, I shake his hand and I felt a squeeze. It's like, what's wrong with this guy? I like, why is it? And like, I barely sat on the chair and he was like, you know, his face was really angry. I was like, why did you recroll us?
最后，事实是一直以来都是迈克尔担任CEO。呃，Sanal，那时候的情况是他就是CEO。迈克尔是个大人物。所以，我和他握手时，我感受到了一种用力的握手。我心想，这家伙有什么问题？为什么会这样？我刚坐下，他的脸看起来非常生气。我在想，为什么你要用"Rickroll"来整我们？
And so, uh, the, the, because I was so sick of making these YC videos. And I thought it was really just a formality. I put their recroll song, their recast, the never going to give you a song. And the YC application, actually, if you go to our YC profile, now they publish the videos, click on the video. It is still a recroll. People discover it. And I agree. Everyone said a while.
因为我实在是厌倦了制作这些 YC 的视频，所以我觉得这只是个形式而已。我在他们的申请视频中加入了 "Never Gonna Give You Up" 的恶搞版。如果你现在去我们的 YC 个人资料页面，他们会发布视频，点击那个视频，还是会看到这个恶搞版本。人们发现后，我也认同他们说的话。
And, uh, and you imagine what's happening because they're reviewing the application just before the interview and they're all sitting together, huddling in the computer and they click the video. They can recroll. So they're all really pissed out of me. And they gave us the toughest YC interview, probably the history of YC. It was really tough and pressuring a question from a hair, one question from there, because I asked my friends from the batch and no one had such an intense interview.
他们在面试前刚刚审核完申请材料，然后想象一下，他们全都聚在电脑前，点击观看了我们的视频，所以他们对我感到非常不满。他们给了我们可能是 YC 历史上最难的面试。面试真的很困难，压力很大，一个问题接一个问题。我问过同批的人，没人遇到过这么激烈的面试。
Um, so, you know, the interview is done. We leave. I tell hi, I was like, okay, I don't think we're going to get out and we fucked it up. We fucked up our only chance. That I call an Uber and then I get, I get this phone call from a random number and pick it up and it's like, um, it's that door of Chang. Uh, she was in the room as well. And she's like, um, you got into YC. How does that sound?
嗯，所以，你知道，面试结束了。我们离开的时候，我对他说，好吧，我觉得我们可能搞砸了，错过了唯一的机会。我叫了一辆Uber，然后接到一个陌生号码的电话，我接起来，电话那头是张女士。她也在面试房间里。她说，你们进入了YC（创业孵化器）。听起来怎么样？
It's like, are you sure? Like I just offended all of you. Like, and you were really angry. And it was like, no, come back. Do the paperwork quickly. Cause the dinners about to start. That's how we got into it. Fantastic. Yeah. Which by the way, it says really good things about YC because if you're Sam or PG and you got inundated by applications and you'd be in a completely inbound kind of like, uh, mode, uh, but very much for the credit that actually went outbound to you, uh, based on that's very true. Actually, which is interesting. Yeah.
这就像是：你确定吗？好像我刚刚冒犯了你们所有人，你们真的很生气。但结果却是，你们说来吧，快点把手续办完，因为晚餐快要开始了。我们就是这样进入这个过程的，太棒了。顺便说一句，这其实对YC来说很有利。如果你是Sam或PG，你会被大量申请淹没，基本上是在处理所有进来的申请，但实际上他们主动找到了你，这确实是很有趣的事情。
So you know, YC and, um, so let's talk about the, the journey to today, the, the final stage and especially, uh, with a machine learning, an AI angle because, um, pretty much from the beginning, uh, of replicas, a commercial entity. So not the, the, the project before that, but the commercial entity. Sounds like machine learning was a key part of it. When you raise your seat around, it was one of the core principles of what you wanted to do.
翻译成中文，这段话可以这样表达：
所以你知道，YC和，嗯，那么我们来聊聊到今天这一刻的旅程，特别是谈谈从机器学习和人工智能的角度，因为几乎从复制品作为商业实体开始的时候，机器学习就是其中的关键部分。我说的不是之前的项目，而是商业实体。听起来在你们进行种子轮融资时，机器学习就是你们想要做的核心原则之一。
Yeah. Uh, we published the, the seed deck at somewhere on there and you can see there was like this master plan and AI was a part of it. So as part of my work at both Facebook and code academy, I worked in compilers and interpreters and there are very, you know, kind of finicky programming exercise. So taking a piece of code and parsing it, that's how compilers work. You parse it into a structured tree and then the way, you know, you traverse the tree, that's how you evaluate a program essentially.
好的。嗯，我们在某个地方发布了种子计划，你可以看到里面有一个总体规划，而人工智能是其中的一部分。 在我在Facebook和Code Academy的工作中，我处理过编译器和解释器，它们往往涉及一些复杂的编程任务。编译器的工作原理是解析一段代码，将其转换成一个结构化的树，然后通过遍历这棵树来实现程序的评估。
Um, and I was, I always figured that like, um, that you could use machine learning instead of like doing all of this by hand. And NLP was getting better and better. Actually, I don't know if you remember, but 2015, 16, we had a chat bot hype. Yeah. Period. Yeah. Cause there was some NLP unlock at the time, but it wasn't good enough. Everyone was faking it. There was like Facebook M, which was like, yeah, it was like a bunch of people. Yeah. Um, and, uh, and so I had this intuition that, you know, we're going to be able to run machine learning and code to evaluate code, to help people learn how to code to, um, just, just like this intuition that like deep learning on code will unlock a lot of avenues. Um, and I felt like if we built this like large community, we're also going to be able to collect a lot of data that's going to help with that. Um, and so every year we, we would prototype something since the start and it was just never, it was never good enough to, to ship anything.
嗯，我一直认为可以使用机器学习来代替手动完成这些工作。自然语言处理（NLP）技术也在不断进步。不知道你是否还记得，在2015年和2016年，我们经历了一波聊天机器人热潮。当时NLP取得了一些突破，但还不够好，很多东西都是假的，比如Facebook M，其实就是一群人在后面操作。我有一个直觉，觉得我们能用机器学习和代码来评估代码，从而帮助人们学习编程。我觉得在代码上应用深度学习会开启很多可能性。如果我们建立一个大型社区，就能收集大量有助于实现这一点的数据。所以从一开始，我们每年都会做一些原型，但这些原型一直都不够好，无法推出任何产品。
And then GPT two came out. And that was the first time when I felt like, okay, this, this feels like the unlock because GPT two can write in my coherent code. Uh, and then when GPT three came out, I think we were one of the first companies to build anything on top of it. The first thing we built was like you highlight a piece of code and you explain it. And I was, yeah, I was taking notes. So that was, um, explain code. That was an important product. Yeah. So what did that do? So, uh, you know, still a lot of people, you know, hobbyists, students that are using the platform. And so, uh, yeah, a lot of code, you're just copy basing from SAG or a flow or library code and you ought to be able to understand it. And so highlighting a bit of code and explaining it is actually very useful. Still having the product people still use it a lot. And you wouldn't have been possible with that. Like it wouldn't have been, you wouldn't have been able to even conceptualize it. Without all of those.
然后，GPT-2 发布了。当时我第一次觉得，这就像是一个突破，因为 GPT-2 能写出连贯的代码。然后，当 GPT-3 发布时，我认为我们是最早在其基础上构建产品的公司之一。我们构建的第一个功能是，用户可以高亮一段代码，然后对其进行解释。我还在记笔记，所以这是一个重要的产品，叫做“解释代码”。这个功能有什么作用呢？很多人，包括爱好者和学生，使用这个平台，所以就有很多代码是从 Stack Overflow 或者一些库里复制粘贴过来的，大家需要能够理解它们。因此，能够高亮代码并进行解释是非常有用的。这个产品依然被广泛使用。没有这些技术，你甚至无法想象能做到这一点。
But what I really wanted to build was, um, like a true autocomplete code generator system. And, um, it was really hard to build anything like that with GPT three. It was expensive. It was slow. And we started talking to open AI by doing something with, um, and then Microsoft had just come out with, with co-pilot and beta. Um, and it was like, you know, we really need to do this. Um, you know, uh, this, this is the future. Like our company would like die if we, if we don't do this, because I was like, all of coding will be that. Mm hmm. Uh, there was a lot of skepticism even inside the company, but I felt like that was the big bet that we're going to have to do. And, um, so, so at the time it was like maybe 22.
我真正想要构建的是一个真正的自动完成功能的代码生成系统。但是，用GPT-3来做这样的东西实在太难了，它既昂贵又慢。我们开始和OpenAI讨论合作的可能性，这时微软刚刚推出了预览版的Copilot。我们意识到，我们确实需要这么做，这就是未来。如果我们不去做，我们的公司可能会面临危机，因为我觉得编程的未来就是这样。公司内部对此存在很多怀疑，但我认为这是我们必须下的一个大赌注。当时大概是2022年。
And so we're, we're, you know, between a rock and a hard place, we were like, okay, we either build on GPT three and it really sucks. Cause co-pilot was using some kind of fine-tuned version, distill version of a GPT three. So that was, uh, it was really fast. We couldn't come to any sort of agreement with open AI at the time. And so I was determined that we're going to have to go build it on our own. Um, and no one at the time had done that. Everyone was relying on open AI, open source AI was not really a thing. And, and we were really plugged into the news and we found that, um, uh, sales force had a, had a small research team and they worked on code models. Um, and they produced this model that was still very, uh, very slow because of the model architecture.
我们当时真是在进退两难的境地。我们想，要么在GPT-3的基础上开发，但这个效果很糟糕，因为Co-pilot使用的是某种经过调优和精简的GPT-3版本，非常快。而我们无法与OpenAI达成任何协议。所以我决心我们必须自己独立构建。当时没有人这样做，所有人都依赖于OpenAI，开源AI还不太流行。我们紧跟新闻，发现Salesforce有一个小型研究团队，他们在研究代码模型。他们确实生产出一个模型，但由于模型架构的原因，速度还是很慢。
But it was, it was decent. It could like generate some code. Um, and so we took that. And that was code gen. That was code gen. Yeah. So we took that. We, uh, sort of rewrote the code to make it a lot faster. We did a lot of engineering work on top of it to be able to ship it. So that was the beta version of what we called ghost rider. So we built the first sort of co-pilot alternative, essentially. Um, and it was super viral because, um, because it really. A lot of people to imagine that there's a, there's a world in which you can build AI as a startup without depending on an open AI. Because at the time the feeling was like, Oh, open AI invented this thing. And it is this magical thing.
但是，那确实是不错的。它可以生成一些代码。嗯，所以我们利用了这一点。这就是代码生成。这就是代码生成。对，所以我们利用了那一点。我们对代码进行了重写，使其运行得更快。我们在此基础上做了很多工程工作，以便将其发布。所以，那就是我们称之为“Ghost Rider”的测试版。基本上，我们创建了第一个类似于“共同体”的替代品。它非常具有传播力，因为它真的让很多人想象，可以在不依赖OpenAI的情况下，作为一家初创公司来构建AI。因为当时的感觉是，哦，OpenAI发明了这个东西，而且这简直是个神奇的东西。
Uh, it's just like no one will be able to replicate it. It's like this dark art and open AI had stopped publishing research and models. And so I think we inspired a lot of people that, Oh, okay. Open source is possible. Open models are possible, but still code gen wasn't, wasn't really great. So I was determined to train, train our own model. And so that was the next phase. And I wanted to use our data to do that. Um, and again, we were in your data being what users do, uh, when the build and, uh, what, what works, what doesn't work and use that to feed the, the LLM. Yeah. By the way, when we released code gen, it was, um, me and like an intern working on it.
嗯，这就像是一种没有人能轻易复制的技术，就像是一门神秘的艺术。当时，OpenAI已经停止发布研究和模型了。所以，我觉得我们激励了很多人，让他们觉得，哦，开源是可能的，开放模型是可能的。但即便如此，当时的代码生成技术还不是特别好。所以我下定决心要训练我们自己的模型，这成了下一个阶段。我希望能够利用我们的数据来进行这项工作。我们的数据就是用户在构建产品时的行为，哪些有效，哪些无效，把这些用来训练大型语言模型。顺便说一下，当我们发布代码生成时，是我和一个实习生一起完成的。
Uh, I, I like recruited an intern. Um, and, and that what, that was it. Um, when it came time that we needed to like train a model, we had to, we had to have like more and better people. And by the way, other people on the company had joined and started learning how to do, how to do machine learning and some people on the data science team joined. Um, we had started talking to Google, Google, we were, we've been always building on top of the Google platform and, you know, they reached out and they're like, look, we're building like, you know, GPT three alternative and we want to see if you want to use it. I was like, Oh, yeah, it looks great.
呃，我招募了一名实习生。就是这样。嗯，当我们需要训练一个模型的时候，我们需要更多更优秀的人才。另外，公司里的其他人也加入进来，开始学习如何进行机器学习，还有一些数据科学团队的成员也加入了。我们开始和谷歌进行对话，因为我们一直在基于谷歌平台进行开发。谷歌联系了我们，他们说他们在开发一个类似GPT-3的替代品，问我们是否有兴趣使用。我当时心想，哦，听起来不错。
Palm at the time. We would love to use it. And, you know, the, the, the talk, the talk progressed, the conversation progressed. And then it was blocked. They're like, we're not going to release it. The search team said it was like too much of a legal risk. So we're not going to release any alarms. Our counterparts, an engineer on the other researcher on the other side of Google, uh, was part of this conversation. He was so frustrated by it and he wanted to leave anyway. So he left, uh, he left Google. His name is Michaela Catasta, a researcher who was on the Palm team. And, um, I was like, you know, we're going to train this model.
当时的Palm技术。我们很想使用它。谈话进行得很顺利，交流在不断进展。但后来被叫停了。他们说不打算发布，因为搜索团队认为这存在太大的法律风险，所以不会发布任何相关警报。谷歌那边的另一位研究员，也是工程师，参与了这次讨论。他对此感到非常挫败，他本来就有离开的想法，所以他最终离开了谷歌。他的名字是Michaela Catasta，是Palm团队的研究员。我当时就想，我们要训练这个模型。
Why didn't you, uh, at least part time kind of join us to, to train this model. Um, and, and he was like, yeah, let's, uh, let's do something great. And I was like, look, we're, we're going to be constrained by capital by talent. And so we have to work within these parameters. Um, and so we decided to train a super small model. At the time, Lama had just come out and, uh, in the research paper, they're talk about how you can train a model, train a small model a lot longer, get the same performance as a much bigger model. So it's a lot more cost efficient.
你为什么不至少兼职加入我们来训练这个模型呢？他回答说，好啊，让我们做些出色的事情。我当时想，我们会受到资金和人才的限制，所以我们必须在这些条件下工作。因此，我们决定训练一个非常小的模型。当时，Lama模型刚刚推出，研究论文中提到可以通过更长时间地训练一个小模型，获得与大模型相同的性能。因此，这样做的性价比更高。
Um, and we trained this three billion parameter model. It was the first three billion parameter model that was state of the art encoding and open source coding. It did better than all the other open source models. It actually did better than Palm when coding. Uh, and, um, and, uh, and yeah, that was another kind of big moment in, uh, in the AI community. And then at a time, a ghost writer was finding a product that we're proud of and we could sell and, uh, and we put it on the market. And that's commercially started being a business. 2023 was very early 23. Yes, very early 23.
嗯，我们训练了一个拥有三十亿参数的模型。这是第一个在编码和开源代码方面达到最先进水平的三十亿参数模型。它表现得比所有其他开源模型都要好，甚至在编码方面超过了Palm。这对人工智能社区来说是另一个重要的时刻。当时，一个名叫Ghostwriter的产品正在寻找我们引以为豪并能够销售的方案，所以我们将其推向了市场，从而正式启动了商业运作。在2023年初，这一切开始进行得非常早。
I mean, around, around the, you know, the time on Chagipati was, was going to go viral. Yeah. And then in the last few months, you've had a huge moment at the company right that feels at least from the outside as, um, you know, like a real hit, uh, that you've had, which is your new product, new ish, but I guess it's still very new, uh, called replete agent. Uh, so maybe walk us through that story. I read somewhere that you had, uh, uh, you know, to the PG thing, a little bit of, uh, founder mode moment when, uh, you decided, okay, this is the time when a lot of the thing that you've been thinking about all along, uh, become possible.
这段话的大意是这样的：差不多是在Chagipati走红的时候，在最近几个月，你们公司也迎来了一个巨大的时刻。从外界来看，这感觉就像是一个真正的成功，那就是你们的新产品，虽然算是最近推出的，但我想它依然很新，叫做Replete Agent。那么，能不能和我们讲一下这个故事？我曾在某处读到，当你决定这是实现你长久以来想法的时刻时，你有过一个类似于创始人模式的时刻。
So what, what happened then? Yeah. So when we, um, uh, after, after released a, uh, a ghost writer and we got a lot of, uh, excitement, we did this, like really big deal with, with Google, uh, on the cloud and AI side and raised a really big round. Uh, and I was like, okay, you know, it's time to mature as a company. I'm going to like go hire all these executives and we're going to, you know, grow the business and, and we're going to go into enterprise and we're going to obviously get to do all the things you you're supposed to do. You're supposed to do, I guess, paper. I did all of that.
那么，后来发生了什么呢？对，当我们发布了一个“鬼才写手”项目后，我们得到了很多关注和兴奋。然后我们与谷歌在云计算和人工智能领域达成了一项非常重大的合作，并成功筹集了一大笔资金。那时我就想，好吧，现在是时候让公司更加成熟了。我打算去招聘一些高管，我们要发展业务，进军企业市场，显然要做所有应该做的事情。于是，我就去做了这些。
At the same time, I sort of like neglected the product a little bit. Um, and also the culture a little bit at the time up until that point for the past six years, that at the time I was interviewing everyone. And when I would like, I want to put my big boy hat on, I was like, okay, I'm going to delegate all of these things. Um, and I'm going to just like be more higher level and, um, add more structure management and things like that. And we, we were suddenly behind, uh, you know, at the time, like the auto complete stop being the state of the art, have people, um, use AI for coding, cursor, it just come out. And, and there were a lot of other explorations around agents and, and, and different things like that.
同时，我有点忽视了产品和企业文化。在过去的六年里，我一直亲自面试每个人。当时，我想要展现成熟的一面，于是决定把这些任务委派出去，自己则专注于更高层次的事务，加强结构化管理等等。突然之间，我们的进度落后了。那时，自动补全已不再是最先进的技术，人们开始使用AI进行编码，Cursor刚刚问世，还有许多关于智能代理和其他技术的探索。
And we had a lot of these explorations already on, but we, we started like lagging behind a little bit because we were optimizing. And this is a lesson, uh, I like to talk to right now, we were optimizing for the current generation of, of models. We're, we're building like, you know, ghost writer V two, right? As opposed to like figuring out what's next. And we'll get back to that in a second. And so, uh, so we were like, we started 2024. We're like 130 people burning, uh, a ton of money. Um, the commercial aspect kind of lagged behind how much we were burning and how big the team was.
我们确实已经进行了许多探索，但我们开始有些落后了，因为我们在进行优化。我现在想谈的是一个教训，我们当时在为当前一代的模型进行优化，比如说我们在构建“Ghost Writer V2”，而不是去思考下一步的发展。接下来我们会回到这个话题。所以在2024年初，我们已经有大约130人的团队，花费了大量资金。商业方面的发展有些跟不上我们的开销和团队规模。
Um, a lot of the engineers were really miserable, even the early engineers, people started leaving because it just felt started to feel bureaucratic. And you know this, like a lot of companies go through this. Um, and, uh, and also we weren't also defining our, uh, customer set, uh, clearly. Um, you know, Reploed was always about democratizing programming. Our mission was always like, um, our first mission was, uh, making programming more accessible. And then we kind of updated it to make it a little more ambitious, which is, um, uh, empowering a billion software creators. Um, but people at the company were confused because, oh, are we building for developers or building for no code people? Are we building for?
嗯，很多工程师都非常痛苦，即使是早期的工程师。人们开始离开，因为这个公司开始让人感觉过于官僚化。这种情况很多公司都会经历。此外，我们也没有非常明确地定义我们的用户群。Reploed一直致力于编程民主化，我们的使命一直是让编程变得更容易接触。后来，我们把这个使命调整得更有野心一些，就是希望能够赋能十亿的软件创作者。但是公司内部的人感到困惑，因为他们不确定我们是在为开发者服务，还是为那些不需要编程基础的人服务。
And so, you know, I, I felt we needed a radical change and I felt we need to bring the burn down and I felt that we needed a reset. And I felt we needed to innovate again and we need to build the AI product that wasn't possible today. And the thought experiment, uh, I, I went through was like, um, let's build the things such that the next generation models will make possible, which actually turns out to be a really hard thing. Like how do you, how do you project forward? What is the capability of these models? And, um, and, uh, you know, when are they going to land? And there's, there's so much risk to that. Yeah. Cause you had a TED talk in 2023 where you said, uh, that some of the stuff that you're actually doing today was possibly not going to be possible for a decade. Right. Is that, is that right?
我觉得我们需要做出一些彻底的改变，需要降低消耗，并且重启很多东西。我认为我们需要重新创新，打造那些今天还不可能实现的人工智能产品。我做了一个思想实验，就是建造一些让下一代模型成为可能的事物，这实际上是非常困难的。要如何预测这些模型的能力呢？它们何时会面世？这其中有很大的风险。是的，在2023年的TED演讲中，你提到自己现在正在做的一些事情可能在十年之内都无法实现，对吧？这样说对吗？
I said, I said, uh, so I give this big talk about how, um, uh, programming agents are going to, uh, change the, uh, the industry and make it so that anyone, um, even people with no technical skills would be able to build software by just commanding these agents, uh, and put up, uh, like a very ambitious vision, uh, for, for replant and for the world to see. And I felt like, oh, maybe we'll get there by the end of the decade. Um, and, uh, and early 2024 when I felt like, um, when I did, when we kind of went to defounder mode, laid off, you know, half the team, uh, leaned up the executive team and, uh, the management structure and brought us back to kind of like a, a small team with a very intense focus on one thing.
我说，我说，呃，我做了一个关于编程代理如何改变行业的演讲，说明即使是毫无技术能力的人，也可以通过指挥这些代理来构建软件。我为Replant和全世界描绘了一个非常宏伟的愿景。我觉得，到这个十年末我们也许能实现这个目标。然而，在2024年初，我们进入了创始人模式，裁掉了一半的团队，精简了执行团队和管理结构，把我们带回到一个小团队，专注于一件事情。
And that, that one thing was like, let's, let's, let's build some version of agent. Um, and so we created this thing called the agent task force. Uh, and that was basically the only thing we were doing at the company. And, uh, that was like March or April, 2024. And we, we got something done really quickly and it was starting to work, but it was very clumsy. And then the big unlock was June, July, 2024. Once on it, 3.5 came out. And throbbing and throbbing son at 3.5. And that was basically the bad. It was like, we're going to build something that's not going to work today, but at some point, some model will come out a little more. And that's exactly what happened. And then he did very quickly.
翻译成中文并易读：
那时候，我们决定要做一件事，就是创建某种版本的代理。于是我们成立了一个叫做“代理工作组”的团队。这几乎是公司唯一在做的事情。大概是在2024年3月或4月，我们很快开发出了一些东西，它开始运作起来了，但是很笨拙。真正的突破发生在2024年6月或7月，当时出了一个叫做3.5的版本。一旦这个版本推出，一切都改变了。虽然那时候我们的目标是构建一个可能暂时不起作用的系统，但我们相信将来某个新的模型会让它发挥作用。而事实也确实如此，进展很快。
Yeah. Yes. You know, the, the product was barely usable with four. Oh, it was like very buggy. It was like very random. They could like produce something. It was very slow. It would take, um, like $100 to make it like a to do app or something like that. Uh, and if you, if you like watched some other videos, like Devin was, like, for example, working on, they had all these problems as well. Yeah. It's very slow, very expensive. So Devin being, uh, another competing, uh, coding agent created by a company called cognition. That's right. That's right. Um, and when, um, Sonic came out, we, uh, we plugged it in and we're like the product immediately got better.
对。是的。你知道，那个产品以前只有四个功能时，几乎没法用。问题很多，运行起来很随机，效果不稳定。想要做出点什么非常慢，要花大约100美元才能做个类似备忘录的应用程序。而且如果你看其他视频，比如德文也在处理类似的问题。是的，速度很慢，成本很高。德文是由一家叫Cognition的公司开发的另一个竞争性编程代理。没错。当Sonic发布时，我们接入它，产品立马变得更好了。
Um, we did also cut scope quite a bit. Uh, so with, with Devin and others, a bunch of them got funded. Um, their idea is like fully automated agents. So you can put it in the background and I'll, I'll do its thing with, with us, the kind of intuition that I had is we want to make it so that it's like this. You know, man, machine symbiosis. Um, and that's something I actually wrote about in, uh, back in 2017. Um, we, we didn't think we were at a place in the technology was not our place to create fully autonomous systems.
嗯，我们确实缩小了项目的范围。和德文及其他人一起，一些项目获得了资助。他们的想法是创建完全自动化的代理程序，你可以将它放在后台自行运行。对于我们来说，我的直觉是，我们希望实现人机共生。我在2017年就写过这个想法。我们并不认为当时的技术水平能够支持我们创建完全自主的系统。
So let's build this like human in the loop system where the agent goes and like could do 10 iteration or 20 iterations, but we'll come back to the user and say, um, was what I built, uh, what you expect. Can you test it for me? cause agents are very good, very bad now at having eyes and, and being able to test something. And that was another kind of piece of product innovation that we did. So we got to September, we launched it in early access. And at the time, there was no other programming agent on the market. So it was really the first and, um, and to capture everyone's imagination.
好的，让我们建立一个“以人为中心”的系统，其中智能代理可以进行10次或20次迭代，但会回到用户这里，询问“我构建的东西是否符合您的预期？您能帮我测试一下吗？”因为目前智能代理在观察和测试方面做得不够好。这是我们进行的另一个产品创新。因此，到9月时，我们在早期阶段就推出了这个产品。当时市场上还没有其他类似的编程代理，所以我们是第一个，这激发了大家的想象力。
Uh, and, and the video of the launch went viral. We were, we thought we'd get like some users to test it and, and we got a ton of users, which was good and bad. It wasn't ready. And so that there's always this, um, you know, you want to set expectations correctly. And I think, uh, we quickly made it a lot better. Like right now, um, right now it's like, I would say like five X better than what, uh, what we launched. And in September, and I feel like it's going to get 10 X better even this year. With those, you also have five X revenue, right? Yes. More. Well, now it's more, I was more than six or seven X, uh, this been a big year. So like you're reset and, and, and all of things.
呃，发射的视频走红了。我们原本以为会吸引一些用户来测试，但结果用户蜂拥而至，这既是好事也是坏事。因为产品还没准备好，所以在这个过程中，你总是希望能正确地设定期望。我认为我们很快就改进了很多。现在，我可以说我们的产品比初次发布时好五倍。而且我感觉今年它会提升到初版的十倍。收入的增长也很明显，对吧？是的，甚至更多，现在已经超过六到七倍了。这真是了不起的一年，就像重新开始一样，各方面都在进步。
And, uh, just to, to bring it home for anyone that may not follow the sort of coding, um, AI for coding space. So there's the, uh, the, the co-pilot model, which is what we talked about with GitHub and I guess ghost writer, which is the AI helps you, I could assist you as you, as you code. And then there's the agent model, which like in its purest form is, uh, an autonomous agent that does the whole thing for you. I would say there's something also between the, the assisted and the autonomous. Um, and that's like the cursor model, which is it's not like, you know, a type ahead kind of like co-pilot.
翻译并简化这段话为中文：
对于那些可能不太了解代码或人工智能代码领域的人，我来简单说明一下。有一个叫做“联合编程模型”的东西，这是我们与 GitHub 讨论的，它类似于“幽灵写手”，可以在你编程时为你提供帮助。然后，还有一个叫做“代理模型”的东西，基本上是一个可以完全自主完成任务的智能代理。我想说，还有一种介于辅助与自主之间的模式，叫做“光标模型”，它不像“联合编程模型”那样只是输入预测。
But it's not fully, it's not like an agent. It is somewhere in between where you can generate entire files. Right. And that's the thing that made them very successful. So the, the magic of Ripley, the agent, um, is that, um, you can prompt it, right? So you, you, you basically say, build this app or even I saw you can even give it a screenshot of what, and the app of your dreams looks like. And you say, do this. That's exactly right. yeah. We wanted so that, um, you're gonna have to be a programmer.
这段文字可以翻译为中文并表达其意思如下：
但是它并不完全是一个代理。它介于两者之间，能够生成完整的文件。对，这正是使他们非常成功的原因。Ripley这个代理的神奇之处在于，你可以对它进行提示。你可以基本上告诉它，构建这个应用程序。甚至我看到你可以给它一个截图，展示你理想中应用的样子，然后你说，做成这样。没错。我们想要的就是这样，这样你就不需要成为一个程序员。
Um, I think you still need to be technically savvy, which is, I think the kind of people we're getting is Silicon Valley professionals and, um, it largely sort of finance and kind of these kind of industries, although we're starting to break out into, um, into more sort of traditional industries that are able to, to use agent, but our, um, our vision. And again, it's like, I think we're going to be able to get to at some point, like a billion users and the kind of mental model that I have, you know, a fix cell has like a billion users. Anyone who's able to use Excel should be able to use agents. Mm hmm.
嗯，我认为你仍然需要具备一定的技术能力。我觉得我们现在吸引到的人大多是硅谷的专业人士，主要是在金融等领域。不过，我们开始向更多传统行业拓展，他们也能够使用这种代理工具。我们的愿景是，我相信我们最终可以达到十亿用户。就像Excel有十亿用户一样，我的想法是，任何能使用Excel的人都应该能够使用这种代理工具。
You're interesting. So yeah, you write a prompt. It, um, it quickly goes into generating almost like a full application for you. uh, it'll provision a database. It'll do whatever it needs, like a program we would need, it'll create a developing environment. We'll bring in the libraries. That's a big thing that that's everything that you had built at Replitt up until now. Right. So, so versus, uh, so it's not just a magical cloud ID. It's, it's, it's quite literally is going to build a live, fully functioning application for you without you as a software creator needing to know anything about databases or cloud. Or any of that stuff.
你很有趣。那么，是这样的，你写一个提示，它很快就能为你生成几乎完整的应用程序。它会配置一个数据库，做任何程序所需的事情，比如创建一个开发环境，导入库。这就是你在 Replitt 上构建的一切，对吧。所以，这不仅仅是一个神奇的云 IDE，而是能够实实在在地为你构建一个现场运行的完整应用程序，而你作为软件创建者不需要了解关于数据库或云的任何知识，或者其他这些东西。
And I think it's all to say, we're like, we're the only, uh, company that that's done that, which is like the full stack experience from the code to the databases to the, to the cloud deployment. Amazing. And so did you rip and replace everything that you had done with, uh, code gen and your own model, we saw it 3.5. We actually, um, stopped caring and perhaps removed a lot of features of people going to the ID and coding. It's still possible. But our view in the way we track our metrics now is that when you go into the editor and code, it's almost like the Tesla disengage. Right. Like, so what is the main metric that Tesla tracks miles between disengagements? Right? And like the V13, I don't know if you saw the chart, but it goes, you know, it's like, you know, 10 X, V12 in terms of like how many miles to go between. And that's my experience to a Tesla now. Like a go to work without disengaging ones.
我认为可以这么说，我们大概是唯一一家做到了从代码到数据库再到云部署的全栈体验的公司。真是太棒了。那么，你们是不是把之前用代码生成和自己的模型做的东西都换掉了呢？我们看到版本3.5的时候，其实我们开始不太关心这些了，甚至可能去掉了很多让人们进入集成开发环境编码的功能。当然，这仍然是可能的。但我们现在衡量指标的方式是，当你进入编辑器编码时，这就几乎像特斯拉的自动驾驶接管被解除一样。特斯拉跟踪的主要指标是什么？是两次接管之间的行驶里程，对吧？我不知道你有没有见过那个图表，但在V13版本中，它的表现比V12好十倍，行驶距离更长。这就是我现在的特斯拉体验，比如可以一路开到公司而不需要手动接管一次。
So when you disengage, they consider it a bug and they're like, you know, give us feedback. So that's how we're thinking about it. So all the stuff that we built around coding AI, I don't think it matters. What really matters is being able to chat with an agent and get an application at the other end. And again, you have this concept of, um, it's that call a control panel or whatever, but like you see the code being written and perhaps in the YouTube video, we can show a screenshot of that. Yes. Um, but so you're able, uh, that's how you achieve your goal of being, um, a product that can be used by non-technical software creators, but also technical users.
所以当你停止使用时，他们认为这是一个问题，就像是在说，请给我们反馈。因此，这就是我们对这件事的看法。所以我们围绕编码AI构建的所有东西，我觉得是不太重要的。真正重要的是能够与一个代理对话，然后得到一个应用程序。同样，你可以想象有一个控制面板或者其他东西，你能看到代码的编写过程，也许我们可以在YouTube视频中展示一个这样的截图。是的，这样你就能够实现自己的目标，即成为一个非技术软件创作者和技术用户都能够使用的产品。
The technical users can get into the code and if it's stuck, sort of has, right? Yes. And, and, um, the multiplier nature of RAPlet, what we're seeing is, um, uh, you know, you can bring in an engineer into the project so you can invite an engineer. Uh, and you have a bounty system that you never needed to do that. Yes. Uh, so let's say you're a solo entrepreneur and you're trying to build an application and you get, you get 80% done, but then you get stuck on this one problem. What do you do? Um, you could potentially get like, go to Op-Worke with some of these places, but they may not know RAPlet and they might like confuse you even further. So we built like the small community where you can hire someone from the RAPlet community to get your application to the, they were asked the finish line.
技术用户可以进入代码解决问题，对吗？是的。而且，RAPlet 的多重特性允许你邀请工程师参与项目，即使你不需要使用赏金系统来实现这一点。假设你是一名独立创业者，正在尝试构建一个应用程序。你已经完成了80%，但遇到了一个难题。你该怎么办？你可以尝试去一些平台，比如 Op-Worke，但他们可能不了解 RAPlet，并可能让你更加困惑。因此，我们建立了一个小型社区，你可以从 RAPlet 社区中聘请某人，帮助你的应用程序顺利完成。
So what, what does this stack look like around? Uh, Sun at 3.5 is there, uh, you know, how do you think about, uh, fine tuning, uh, prompt engineering, rag, uh, like any of those now AI engineering kind of, um, uh, you know, Lego blocks. Where do they fit in? Yeah. So, uh, at the bottom layer is the RAPlet container. And there's a process that we call Pidwan, which is like the first process in the container that orchestrates, uh, all these things. It, um, it has services such as like install this package. Um, uh, write this file, read this file, uh, but also has services where it is like watching the file system and then every edit is updating a vector DB, uh, there. So rag built into that.
这段文字大致的意思是描述一个技术堆栈的结构，以及其中不同模块和组件的位置和用途。可以翻译为：
"那么，这个堆栈长什么样呢？呃，Sun 3.5在哪里，你对微调、提示工程、RAG（检索增强生成）等AI工程模块是怎么考虑的？这些像乐高积木一样的组件怎么适配进来呢？是的，在底层是RAPlet容器。我们有一个叫Pidwan的进程，它是容器中的第一个进程，负责协调所有这些东西。它提供了一些服务，比如安装软件包、写文件、读文件等，同时也有监控文件系统的服务，每次编辑都会更新向量数据库，RAG就是内置在这其中。"
这样翻译能使读者清楚了解不同组件的角色和它们如何在整个系统中协同工作。
And so this, this like fully, uh, encapsulated system that it kind of looks like an operating system. So, uh, you know, uh, you know, manages the entire system and exposes this very clean interface to the agent. Where it can, uh, we expose all these tools. We can install a package, read a file, delete a file, write a file. Um, you know, a provision of database, do all these things that a human using the RAPlet editor and IDE could use. Um, and then in terms of the architecture of the, um, of the agents, um, it's a multi-agent system. So when you're talking to the agent in what we call the lobby, it was like when you're starting to the agents hash on.
这个系统就像一个完整的操作系统一样，它管理整个系统，并向代理提供一个非常干净的界面。在这个界面中，我们可以提供各种工具，比如安装软件包、读取文件、删除文件、写入文件，还有数据库配置等，所有这些是通过RAPlet编辑器和IDE的用户可以使用的功能。至于代理的架构，它是一个多代理系统。当你在所谓的大厅与代理交互时，就像是开始与代理进行讨论。
Um, it is like mostly a conversational agent that is like building a plane, uh, plan for you. Uh, it doesn't have a lot of tools, only tools around, um, um, around a retrieval. Um, and so that builds a plan for you. And then when you start with a plan, it goes into, uh, sort of a prototyping agent, uh, as it were, where we're like, uh, this agent is like responsible for like generating the initial application. Uh, it tries to, it tries to kind of be as complete as possible, but it tries to kind of, you know, reduce the ambition of, and this cope of the program because, you know, people put PRDs that are like really big.
嗯，这就好比是一个对话助手，为你制定计划。它的工具不多，主要是和信息检索有关的工具。它会为你制定一个计划，当你开始执行这个计划时，它就变成一个原型制作助手。这个助手负责生成初始应用程序。它尽量做到完整，同时也努力降低程序的野心和范围，因为人们常常提出的产品需求文档（PRD）会比较庞大。
And so we're trying to get the first thing done. Uh, and then it kicks it off to like the project management agent. And so that becomes the primary agent where you're conversing with. And that agent is the agent that has the memories, uh, has the connection with the user. And then it is the, um, router agent. It will kick off other agents, right? So you're, you're, um, you're, um, you're, um, you give it a new task. You know, add this feature. It will kind of provision editor agent. So the editor agent gets some context, gets the files, you know, the rag, the files. And goes into its own iteration loop where it's like writing the files and trying to edit a file.
我们正在努力完成第一项任务，然后它会交给项目管理代理。这个代理成为你主要交流的对象，并且存有用户的记忆和连接。接下来，它担任“路由代理”的角色，会触发其他代理。比如说，当你给它一个新任务，比如添加一个新功能，它会启用一个编辑代理。这个编辑代理获取一些相关背景和文件信息，然后进入自己的循环过程，进行写作和编辑文件的操作。
By the way, editing files with models is actually quite a tough problem. Turns out they don't know how to generate diffs. They don't work with line numbers very easily. So you have, you need a quite a complicated system for edits to work. So when the editor finishes, it kind of goes back to the project manager, gives it a summary of what it was able to do and that gets committed to memory. And the project manager kicks off a bunch of processes to restart the application and, um, and get all the way to presenting the application to the user. Now along this way, it might run into an error, right?
顺便说一下，用模型编辑文件实际上是一个相当困难的问题。事实证明，它们不知道如何生成差异（diffs）。它们也不容易处理行号。所以你需要一个相当复杂的系统来使编辑工作。编辑完成后，它会将能做到的内容总结反馈给项目经理，并将其记录下来。然后项目经理启动一系列流程，重新启动应用程序，直到应用被展示给用户。在此过程中，有可能会遇到错误，对吧？
We resource the application, runs into compiler. The editor maybe ran into a syntax error or runs into a runtime error or runs into database error. So now you need to kick off to some debugging loop to be able to debug that and the debugging loop might come back and go back to the editor. And so this is where the autonomy is happening, right? It is, it is going through these different states in order to get to the application to the sort of final state to be able to present to the user. So it's very much a family of agents. It's a multi agent system. Yes. Yes. And that way you can encapsulate every agent and this clean interface because the contacts get pretty messy if you're like doing everything in, in, in one place.
我们将应用程序资源化，运行到编译器。编辑器可能遇到了语法错误、运行时错误或数据库错误。因此，现在你需要启动一个调试循环来调试这个问题，而调试循环可能会返回并重新进入编辑器。所以这就是自主性发生的地方，对吗？它正在通过这些不同的状态，以便将应用程序呈现给用户。在这个过程中，它就像是一个代理的大家庭，这是一个多代理系统。没错，这样你可以为每个代理封装一个干净的接口，因为如果你在一个地方处理所有事情，内容会变得非常混乱。
And what we found is, you know, all these LMS are getting marketed as, um, very long contacts models. What we found is reasoning over long contacts. Actually, not very good. It's like not good at all. Like, once you cross 32,000, uh, tokens, the performance of reasoning, you know, just goes down a hell like very, very quickly. So chunking the problem into smaller bits, uh, helps with that. And, uh, being able to audit the memory and being able to, yeah, like, um, making sure not everything goes into the memory. Like when you, when you have a sub agent going and doing its thing and not everything comes back and pollutes your memory and contacts and kind of really having this isolation.
我们发现，所有这些学习管理系统（LMS）都被宣传为可以处理非常长上下文的模型。然而，实际上它们在处理长上下文上的推理能力并不是很好。比如，当处理超过32000个词元时，推理性能就会急剧下降。所以，将问题分解为更小的部分有助于解决这个问题。此外，能够审查记忆并确保并非所有信息都进入记忆中也很重要。比如，当一个子代理执行其任务时，并不是所有的内容都会返回并污染你的记忆和上下文。保持这种隔离非常重要。
And how do you think in the general context about, um, evaluation, uh, so, you know, for hallucination and all the things, but I can, in general, like evaluation is like the big problem of those, um, of one of the big problem of AI engineering. So what does that fit there? So software agents, um, there's a, there's this benchmark called, uh, sweet bench. So software engineering, uh, bench. Uh, and that's what everyone's competing on.
在整体环境中，你如何看待评估的问题，尤其是关于幻觉和类似现象的评估？评估通常被认为是人工智能工程中的一个大问题之一。那么这在这个领域中是什么位置呢？对于软件代理，有一个名为“Sweet Bench”的基准，这相当于软件工程的基准测试，大家都在以此为竞争标准。
We don't really compete on that. And the reason is because we are building for a different audience. The people who are building sweet bench, um, are building agents that go from, uh, issue or a ZR ticket to a per request. What we go from is a high level product description to an app. And that's like a fundamentally different problem. And so the, uh, the tough thing for us is that there isn't academic benchmarks, which gives you a head start, right? So we've had to build our own benchmarks.
我们实际上并不在那个领域竞争。原因是我们针对的受众不同。那些在做Sweet Bench的人，正在开发从问题或ZR工单到请求的解决方案。而我们是从一个高层次的产品描述开始，直接构建应用程序。这是一个根本不同的问题。因此，困难在于没有现成的学术基准来给我们提供一个起点。因此，我们不得不建立自己的基准。
And, you know, as you get users, you get more data and you're, you're able to get some of that data and create benchmarks from it. And, and that's like a quite a tough problem. So we have our own internal emails and benchmarks. I heard you somewhere talk about, um, ACI versus HCI. Um, can you go into that? That was, uh, basically what you learn about interacting with machines versus, uh, humans.
随着用户的增多，你会获得更多的数据，从而能够利用这些数据创建基准。这是一个相当棘手的问题。我们有自己的内部邮件和基准。我听你在某个地方谈论过关于 ACI 和 HCI 的区别。你能详细讲一下吗？这主要是指你在与机器互动与人与人互动中学到的东西有什么不同。
Yeah. So, uh, when we were building, uh, the rapid environment for humans, um, you, you're like, okay, this is the editor. This is where you go and type code. This is the UI for the editor. There's the line numbers and this is where you see. There's the console. This is where you run the code and here's the UI for the, for the terminal and everything. Here's the files and here's the UI. Here's how the files work and and then you want to deploy or you want to add a database. Here's how you do it. Here's how you provision the database. Here's the UI for it. Here's the interaction, right?
好的。那么，当我们为人类构建快速开发环境时，你会这样理解：这是编辑器，这是你输入代码的地方，这是编辑器的用户界面。这里是行号，你可以在这里看到。这里是控制台，这是你运行代码的地方，这是终端的用户界面和一切。这里是文件，这是它们的用户界面，这就是文件的工作方式。如果你想部署或者添加数据库，这就是如何操作的方法。这是配置数据库的方法，这里是它的用户界面，以及与之互动的方法，对吧？
Um, when you start building, uh, something like the agents, your expectation is like, okay, I'll just present all these things as APIs and I'll just work. What you find pretty quickly is that if you just do that kind of verbatim, you have, you don't have the best performance and there was a paper, uh, I think from like Open Dev in or some of the people that are working on agents in open source, we're introduced the terms ACI. So the term HCI is human computer interaction. ACI is like agent or AI computer interaction.
嗯，当你开始构建一些像代理这样的东西时，你可能会期望：“好吧，我只要把所有这些东西呈现为API，然后它们就会正常工作。” 但你很快会发现，如果只是照本宣科地这样做，效果并不理想。有一篇论文，好像是来自Open Dev，或者是从事开源代理研究的人的，他们引入了“ACI”这个术语。HCI代表人机交互，而ACI类似于“代理或AI计算机交互”。
And the main observation from that paper is that actually you need to create tools that has like that almost has a UI but a UI built for language models. So the interesting philosophical thing here is that language models are actually kind of human imitation machines because the train on the train on all our crap on the internet. So they become this like, you know, you know, it's like very much like us and they say they understand, um, their view of the world is like closer to us than to like, uh, you know, pure program.
从那篇论文中得出的主要观察是，实际上你需要创建一种工具，这种工具几乎就像具有用户界面一样，但这是为语言模型构建的用户界面。这里有趣的哲学观点是，语言模型实际上是一种“人类模仿机器”，因为它们是在互联网上训练的，吸收了我们所有的信息。所以它们变得非常像我们，它们所谓的理解，对世界的看法比起纯粹的程序更接近我们。
And um, so you're in that paper, for example, they say like, um, it's better to give the, um, instead of giving it the entire file, the editor, it's better to give it a view of the file and have it navigated to page up and page down. So you're actually building an editor for the agent. And this is, this goes across, um, the entire stack, for example, we give it eyes, right? We give it a screenshot tool. So again, go screenshot the, you know, the browser and see what's going on in the application.
翻译如下：
比如说，在那篇文章中，他们提到，与其将整个文件交给编辑器，不如给一个文件的视图，这样就可以通过上下翻页进行导航。这样实际上是在为代理构建一个编辑器。而这个概念贯穿于整个技术栈。比如说，我们给它"眼睛"，即一个截图工具。这样它就可以截取浏览器的截图，查看应用程序中发生了什么。
Um, and the kind of feedback, you know, as it's editing the code, it gets like compile errors and things like that. And so, uh, now we spend a lot of time crafting these tools for the AI, which is kind of like a weird surreal thing to think about. What, uh, doesn't work yet? The main problem is, is reasoning. And the reason why I sort of predicted that is going to take a while to get reasoned correctly. And I think I was kind of, it happened faster, but like, I still think the reasoning is like not there. Um, obviously like all one and these test them compute, uh, models might change things.
嗯，当我们在编辑代码时，AI 会得到各种反馈，比如编译错误之类的。因此，我们花了很多时间来为 AI 打造这些工具，这种想法有点奇怪和不真实。目前还存在的主要问题是推理能力。我曾经预测，推理能力的提升需要一些时间。我觉得它发展的速度比我预想的快，但推理能力还是不到位。显然，像所有的一体化测试和计算模型可能会改变这一点。
Um, but, uh, uh, sort of what, what models are still are today is they're like, they're sort of a completion engines, right? That's, that's how LMs are trained, the sort of order aggressive models where they try to predict the next token. They're still next to open prediction machines. Um, reasoning is, is, is different. Uh, next token prediction is more like intuition. It's more like, oh, I know what you're going to say next, right? Whereas reasoning is more like, um, it is more like, you know, this happened and potentially this caused this and, and, well, here's a conjecture. There's a hypothesis and let me, let me think about that. Oh, but you know, I'm wrong. Let me backtrack and, you know, think about a different hypothesis. And so it is like qualitative, very different than, uh, you know, intuition and you know, in the sort of literature, it's called system one versus system two, you know, the Daniel Kahneman book, the fast and thinking slow.
嗯，目前的模型基本上还是一种“完成引擎”。这是因为大型语言模型（LMs）是通过预测下一个词来训练的，这是一种顺序性的预测方式。这种方式更像是一种直觉，就像是“哦，我知道你接下来要说什么”。而推理则是不同的，它更像是“这个事情发生了，可能导致了这个结果，这里有一个猜测或假设，让我来仔细想想”。然后，如果发现错误，再换个角度思考其他假设。
这种思考方式与直觉非常不同。在学术界，它被称为"系统一与系统二"，就像丹尼尔·卡尼曼的书《快思慢想》中提到的那样，直觉属于快速思维，而推理则是慢速思考。
And so models today think fast. We don't, I mean, you could argue old one is that, but still like, and it's your reason out, like whether they can do whether you can build really good agents with them. We're starting to see results from that, but still kind of early. Inevitable question and to the, you know, point about clickbait, the headlines is, um, where, what does all of this mean for professional software developers? Where if we build agents that can empower non-technical people to build fully functioning apps, what is the future for software developers? I think the best way to think about it is to think about other industries where that happened.
如今，模型的反应速度很快。我们并不是说过去的模型不行，但就算是那样，我们还是需要考量是否能用它们构建优秀的智能代理。目前我们已开始见到一些成果，但仍处于早期阶段。一个不可避免的问题是，与点击诱饵（clickbait）相关的新闻标题，以及这些发展对专业软件开发人员意味着什么。如果我们创建的智能代理能够使非技术人员也能构建完整功能的应用程序，那么软件开发人员的未来会是怎样的？我认为最佳的思考方式是参考那些已经发生类似变革的其他行业。
Um, you know, right now it's very easy to take an amazing photo with an iPhone. Right. Um, and that a lot of like, you know, I mean, Apple did this whole campaign around, you know, to get rid of an iPhone. There are a lot of people that are just like making really great content with just, uh, iPhones. Um, professional photographers like didn't die. Maybe the population is not growing as fast, but they exist and they have their own equipment. And, um, it tends to be better. I don't know if the sort of the market size of professional photographers have like shrunk. Uh, but it's kind of probably state constant.
嗯，你知道，现在用iPhone拍出一张很棒的照片非常简单，对吧。而且苹果公司还专门做了一个宣传活动，就是为了用iPhone拍摄。有很多人仅仅使用iPhone就能够制作出很出色的内容。但这并不意味着专业摄影师就不存在了。也许他们的数量增长得不那么快，但他们依然存在，并且拥有自己的专业设备。这些设备通常要更好。我不确定专业摄影师的市场规模是否缩小了，但大概是保持相对稳定的。
Um, and I think that's probably what's going to happen to software engineering. The growth that we've seen in computer science, um, over the past 10 years, I think we'll slow down. Uh, as the, um, need for them, uh, as well as them because companies will be more efficient. Suck was unrogan recently and he talks in 2025. We're going to have a mid-level software engineer equivalent of a coding agent that's going to be able to commit in code. I think it's a little overstated, but, um, I don't think he's, he's wrong.
嗯，我认为这可能会是软件工程的发展趋势。在过去的十年中，我们在计算机科学领域看到了快速增长，我认为这种增长将会放缓。因为随着公司效率的提高，对软件工程师的需求会减少。不久前，Suck在一场活动中提到，到2025年，我们将会有相当于中级软件工程师水平的编码代理人，他们能够直接参与编码。我觉得这个说法有些夸大，但我认为他的观点并非全无道理。
I think we're on the trajectory where you can do a lot of. Sort of junior to mid-level engineering tasks, uh, automatically and that's like refactoring, testing, I think initially. Uh, but pretty soon, just like some basic features and moving UI around and things like that. So the need for software engineering will. Will, will go down. Um, and as, uh, as, you know, non-professional software engineers are able to build their own software, um, I think that. The, uh, demand for the, you know, plethora of SaaS that we have today will also go down and that will reduce because you'll be able to generate software on the fly that like really fits your use case and, um, I don't think SaaS is going away anytime soon because there's all these ways in which, um, companies can can support enterprises.
我认为我们正处在一个可以自动化完成许多初级到中级工程任务的轨道上，比如重构和测试这样的任务可能是开始，然后很快还可以自动化实现一些基础功能、移动用户界面等。所以对软件工程的需求可能会下降。非专业软件工程师能够自己构建软件，我觉得这样一来，对现有大量SaaS（软件即服务）的需求也会减少，因为你可以即时生成真正适合自己需求的软件。不过，我不认为SaaS会在短期内消失，因为企业在许多方面仍然需要这些公司的支持。
Um, that's like quite important to security and all that stuff, but there are a lot of long tail SaaS that I think will just go away and people will be able to generate the software that will reduce demand on, on for software engineering. Uh, so I, I think, um, I think, I think the population of software engineering will probably be constant and not grow all that much. So to unpack a couple of the things, so the, the, the future of software developers and then the, the future of the SaaS and software industry. Uh, so on the developer front, um, so there's a couple of, a couple of things. Uh, one is the often asked sort of obvious question of, um, you know, if you, if you remove the stage where you're a junior and then a mid level that does, you know, very basic tasks.
这个问题有点重要，涉及到安全性和相关领域。不过，我认为许多小众的SaaS（软件即服务）可能会消失，人们将有能力生成软件，从而减少对软件工程师的需求。所以，我觉得软件工程师的数量可能会保持稳定，而不是大幅增长。让我展开谈谈几个方面，一个是软件开发者的未来，一个是SaaS和软件行业的未来。在开发者方面，有几个问题值得讨论。其中一个经常被问到的明显问题是，如果你移除初级和中级阶段，去做那些非常基础的任务会怎么样。
Uh, how do you become senior and what does it become? What does it mean to be a senior software developer when you never were a junior? That's the question. Well, I mean, you can generalize it. How do you become an account executive if there's no AI SDRs? Yeah. And it feels like there's not going to be any. Hundreds. Yeah. Right. Yeah. I think it's just a general question about, uh, entry, entry jobs. I don't have an answer. Yep. What do you think? No, fascinating, uh, fascinating future. Uh, yeah. I mean, I think as, as much as we're creating an alien form of software, I think we're creating and we're going to create an alien form of what it means to be a professional, right? And then, you know, uh, we'll have hopefully superpowers that we can't imagine just yet.
呃，怎么才能成为资深人士，这意味着什么呢？如果从来没有做过初级软件开发人员，怎么能成为高级软件开发人员？这是个问题。嗯，我的意思是，你可以把它推广一下。比如说，如果没有AI销售开发代表，怎么能成为一名客户主管？是的。而且感觉好像永远也不会有。成百上千的。对，我觉得这只是一个关于入门级工作的普遍问题。我没有答案。你怎么看？嗯，这是个令人着迷的未来。我想，就像我们正在创造一种陌生的软件形式一样，我们其实也在创造，且将继续创造一种陌生的职业定义，对吧？希望到时候我们会拥有无法想象的超级能力。
Well, you can imagine something like that or, you know, uh, what you hear often as well is, uh, which is what I think as well, uh, that, uh, intuition, salesmanship ability to network, create contacts, uh, will matter as much as technical skills. Well, you could also imagine, um, sort of, uh, AI's teaching humans. Um, like, I feel like, you know, education sucks, right? It's something to education sucks. But it, you know, it is one of those applications of AI that hasn't been fully explored. Um, I think using AI to teach those junior programmers and accelerate them to become a senior pretty quickly is a possibility.
你可以想象类似的情况，或者说，你也常听到的一种观点，我也是这么认为的，就是直觉、销售能力、建立人脉的能力等，和技术技能一样重要。你也可以想象，人工智能可以用来教人。我觉得教育有些地方做得不好，这是教育的一个缺点。但是，教育是人工智能尚未被充分挖掘的一个应用领域。我认为可以利用人工智能来培训初级程序员，并迅速提升他们成为高级程序员。
Assuming that any run of the meal, uh, programmer with replant, with coding agents can create amazing applications, then what's the next bottleneck? Is it? Creativity is that what is it? Well, I think if you think about it as a sort of a, as a, like a factory line or a pipeline, um, you know, uh, the bottleneck has always been sort of like the, uh, the making of the thing, right? So you have an idea, um, and, uh, you need to make, make the thing, uh, and then you need to sort of sell it and distribute it. And we have all these ways where sales and distribution has gone easier over time. Like PLG is an example of one, um, adds, you know, things like that.
假设任何普通的程序员配合编码代理都能开发出令人惊叹的应用程序，那么下一个瓶颈是什么呢？是创意吗？究竟是什么？嗯，我认为如果把这看作一个类似于工厂流水线的过程，那么瓶颈一直以来都是在于产品的制作。你有一个想法，需要将其实现，然后再销售和分发它。随着时间的推移，我们在销售和分发上都有了很多便利的方法，比如产品主导增长（PLG）就是一个例子，还有各种广告等。
So there's a lot of innovations and ways in which companies can grow fast today. Um, so the bottleneck is like, oh, how do you, you know, how do you create the thing? And even if you have the capital, it's actually still quite hard to make things, right? And you know that as an investor, sometimes you invest in a company. It's still not working. You could give it as much money as you can. I mean, what is it like all these companies, um, magic leaves and equi, uh, movie or whatever. And like, uh, there's a ton of examples of, uh, companies that sucked up a lot of capital and win nowhere. Um, but you know, uh, if, if in fact, uh, soft recreation gets dramatically easier and better, then I think, uh, the bottleneck will shift somewhere else.
今天有很多创新和方法可以让公司快速成长。然而，瓶颈在于如何创造产品。即使你有资金，制造东西实际上还是很困难。作为投资者，你可能会遇到这样的情况：即便你投资了一家公司，它仍然不见成效。你可以投入大量资金，比如那些拥有大量资本却没有取得成功的公司，例如某些高调的创业公司。这类例子有很多。然而，如果软件开发变得显著简单和高效，那么瓶颈将会转移到其他地方。
Uh, and my, uh, prediction, uh, is that the bottleneck will become on the idea side of things, which is counterintuitive because the, um, Silicon Valley, um, dogma is like ideas don't matter, only execution. And, and maybe in the future, it'll be more like execution doesn't matter. Only ideas. Um, because, you know, if you think about it, there aren't a lot of great ideas in the world. And a lot of times you should, you know, it takes, um, a lot of times like it takes this novel idea and all these ideas and retrospect seem great. But, um, but, you know, sometimes a thing could have happened 10 years ago. And it just happened today because someone had like a really great idea. And that's true of companies, but even at a sort of personal level, or that's a problem of like broadly horizontal platform, which is okay. All right.
呃，我的预测是，瓶颈将会出现在创意层面，这有点反直觉，因为硅谷的传统观念是，创意不重要，只有执行力重要。也许在未来，情况会变成执行力不重要，只有创意重要。因为，如果你仔细想想，世界上并没有很多伟大的创意。通常需要一些新的想法，而很多在回顾时显得很棒的想法，有时可能是在十年前就能实现的，但它们直到今天才出现，因为有人有了一个真正伟大的想法。这种情况不仅适用于公司层面，也适用于个人层面。或者说，这也是广泛横向平台的问题。
So now tomorrow morning I can know today who's replete, I can build any app. Okay. What is it? What is it that I need? Right. That's almost a problem of a chat GPT when we were first all confronted to it. It's like, okay, you got this box. And like, I think, you know, we two years in and people are still basically just starting to understand the range of things that they can do. It's a skill.
所以现在，明天早上，我就能知道今天谁是充实的，我可以开发任何应用程序。好的。那么我需要什么呢？可以说，这就像我们第一次接触到ChatGPT时所遇到的问题。这就像得到了一个工具箱，两年后的今天，人们基本上才开始真正了解他们能够做的各种事情。这是一种技能。
Yes. Right. Like, um, finding problems solve is a skill. Problem solving is a skill, but like finding, like, if you know how to code or if you know how to make things with replete agent, you still need to develop the skill of like, potting problems in the world that you're going to be able to solve.
是的，没错。发现问题并解决问题是一项技能。解决问题是一个技能，但就像编程或使用Replete Agent制作东西一样，你仍然需要培养一种技能，就是能够在现实世界中识别出你能够解决的问题。
And by the way, back to the bottleneck. Um, that bottleneck is also true of science. And math and every part of a human invention, um, uh, assume that we have the, uh, you know, labs are fully automated and we have robotics and, um, and so what becomes the bottleneck to running experiments, ideas? Mm. Um, you know, I assume we have this like amazing, uh, theta improvers and things like that. And then you can try things really fast.
顺便说一下，回到瓶颈问题。嗯，这个瓶颈在科学领域也是存在的。在数学和人类所有发明的各个部分也是如此。假设我们的实验室是完全自动化的，我们还有机器人。那么，进行实验的瓶颈就变成了想法。嗯，你知道，我假设我们有一些很棒的理论改进工具之类的东西，然后能非常快速地进行尝试。
Um, and then what is the bottleneck to coming up with new mathematical insights? It's going to be new things to try and ideas to try. It brings up the question of like, um, how do you, how do you teach the next generation of people? How do you teach your kids, um, to become more creative and I'm not sure how you do that.
嗯，那么产生新的数学见解的瓶颈是什么呢？关键在于尝试新的方法和想法。这引出了一个问题：我们该如何教育下一代，如何教导孩子们变得更加有创造力？对此我也不太确定要怎么做。
And then going, going back to the point you were making about, um, software. So indeed, um, this, this idea that all of this could be profoundly disrupting, uh, to the, uh, I guess what, what has been known as a SaaS industry, uh, and software in general. And, um, you know, just like a couple of tweets, um, I saw like as I was, you know, prepping for this.
然后回到你刚才提到的软件这个话题。确实，这种想法可能会对所谓的SaaS行业以及整个软件行业产生深刻的冲击。我在为这个做准备时看到了一些推文，提到了类似的观点。
So that's the one from PG precisely that, uh, uh, went, uh, very, uh, viral where he said, I talked to the CEO of a moderately big tech company who said that replaced Figma with Replit. This surprised me because I don't even think of them as being in the same business. But he said Replit is so good at generating apps that they, uh, just, uh, go straight to Port-O-T-A-P now. So that was, that was one thing.
这句话来自PG，他提到有件事非常火。他说，他和一家中型科技公司的CEO交谈，那位CEO说他们用Replit替代了Figma。这让他感到惊讶，因为他从未认为这两个公司属于同一领域。但那位CEO说，Replit在生成应用程序方面非常出色，他们现在直接使用Port-O-T-A-P。因此，这就是其中一件事。
And then, um, another, uh, tweet from, uh, Chris Branridge who said that, um, he had just built a type form clone in 20 minutes for $3.50 using Replit AI agent. Uh, so just a couple of examples, but, um, I can indeed, uh, you know, if you can build any software and what does that mean for entrepreneurship? What does that mean for venture investment? Uh, does any company have any, uh, mode at least from a technical standpoint? Uh, and then, you know, is that a good thing? Is that a bad thing? It's, it's so huge as a concept that it's hard to wrap your mind around it.
然后，还有一条来自克里斯·布兰里奇的推文，他说他刚刚用Replit AI代理在20分钟内花了3.50美元构建了一个Typeform的克隆。这只是几个例子，但确实，如果你可以随意构建任何软件，这对创业意味着什么？这对风险投资意味着什么？至少从技术角度来看，任何公司还有护城河吗？那么，这是一件好事还是坏事？作为一个概念，它是如此巨大，以至于很难理解。
I feel like venture, uh, capitalists are already intuiting that the, um, sort of golden age of offsasses on its way out. And, uh, you can look at the, uh, sort of the American dynamism trend, uh, like trend to going to the world of atoms away from bits. Uh, you have the, you know, after Andrew, like defenses is a big thing that all these VCs are getting into, uh, government in general, obviously like AI, AI chips. Um, and, uh, sort of what we used to call heart tech is, is becoming trendy. Yeah.
我感觉到风险投资家们已经在暗暗察觉到，那种所谓的“办公室即服务”的黄金时代正在消退。可以看看美国当前的动态趋势，似乎正从“比特”世界转向“原子”世界。比如，自从安德鲁（也许是指某个重要人物或事件）之后，国防领域变得很热门，吸引了很多风险资本家。还有就是政府领域的投资，当然还包括人工智能和人工智能芯片。另外，我们以前称为“硬科技”的东西也开始变得流行起来。
And maybe it becomes the majority of venture capitalist, uh, venture capital investment, as opposed to the, to the other way around. Mm. Um, uh, I think there are still like network effect businesses and software that are quite defensible, like even if you generate Facebook, you know, you're going to have to get a billion users, right? And so, um, you know, consumer, social, um, those were probably become more, more important as categories for, for VCs.
也许，风险投资主要会集中在这一部分，而不是相反的趋势。我认为仍然有一些依靠网络效应的业务和软件非常具有防御性。就像如果你创建了一个类似于Facebook的平台，你还需要获得十亿用户。因此，消费类的社交软件可能会成为风险投资中更重要的类别。
Um, and perhaps there's, there's probably going to be a consolidation of, of, uh, venture capitalists, especially there's like a bit of a squeeze where like the, uh, you know, um, do a lot of the early stage rounds are like, there's like a lot of angels and solo, solo capitalists that are able to do those. And, and the later stage rounds, they're like so expensive and, and capital intensive that like only a few could do.
嗯，可能会出现风险投资家的整合，特别是在目前这种紧缩的情况下。就是说，很多早期融资轮次都是由天使投资人和个人投资者来完成的。而在后期融资轮次中，由于成本和资金需求非常高，因此只有少数几个人能够参与。
And in terms of entrepreneurship, the other thing that brought up, I actually think it's going to be net positive for entrepreneurship. The way we use the way we define entrepreneurship today in Silicon Valley is like you build a venture scale business. The way kind of America defined entrepreneurship pre-s Silicon Valley is like anyone be able to build a business.
在创业方面，提到的另一件事是，我实际上认为这将对创业产生积极的影响。今天我们在硅谷定义创业的方式是指建立一个具有风险投资规模的企业。而在硅谷兴起之前，美国对创业的定义则更广泛，就是任何人都可以创办一个企业。
Um, if you look at the, uh, number of firm creation, uh, it's like been trending down. So this is like whatever 1960s or 70s, which is kind of crazy because like everyone's talking about startups today, but like new firm creation has been trending down. And actually there was like an uptick with COVID. I don't know if it stayed, but COVID, um, got a lot of people to, you know, quit the jobs and start online businesses.
嗯，如果你看看公司创立的数量，它近年来一直在下降。这有点疯狂，因为不论是1960年代还是1970年代，大家现在都在谈论创业，但实际上新公司创立的数量一直在减少。实际上，在新冠疫情期间有过一次增加。我不确定这种趋势是否持续下去了，但疫情确实促使很多人辞去工作，开始在线上创业。
So I think the form of entrepreneurship will come back as the sort of more small business entrepreneurship. Okay. So maybe too close, uh, as I was, uh, you know, researching this and even before, I was, I, you know, saw you online on all the things, one of the things that I've found fascinating about you as a, as a founder, um, is that there is a whole side of you that's deeply thoughtful about topics, uh, philosophy and politics and, you know, all those, uh, all the things.
所以，我认为创业的形式会重新回归到更多的小型企业创业。好的。在我研究这个问题期间，甚至在此之前，我在网络上关注到你，令我着迷的一件事是，作为一个创始人，你在哲学、政治等话题上有非常深入的思考。
So, uh, maybe for the last few minutes or this kind of thing, I wanted to get into some of this, um, and to start with your described, uh, to, to me and also online as a techno optimist, which is like part of that, uh, you know, uh, I guess movement into like in value. Like is that, is that fair? Is that how you, uh, think about yourself? yeah. So, um, so, so I think that, uh, this is actually, um, there's actually two ways to view the world. Uh, you, you're either, uh, a primitivist or a techno optimist because I guess they're European style of like, you know, we're done, we're just gonna, just get a distribute the wealth and, and, uh, regulate, uh, regulate.
所以，呃，也许在最后几分钟或者在这种情况下，我想谈谈一些事情，呃，首先你曾经描述过自己是一个科技乐观主义者，你也在网上这样描述过自己，这也是这类价值观的一部分。我想问，这样形容你是否合适？这真的是你对自己的看法吗？对。所以，我认为，其实看待世界有两种方式，你要么是原始主义者，要么是科技乐观主义者。我想这是欧洲风格，就是我们已经完成了，只需要分配财富和进行监管。
I don't think it works at all because capitalism is based on growth. And when you don't have growth, you actually have decline. Um, and so you can say we're actually gonna, gonna go back, uh, you know, and, and, you know, and like actually technology is bad and deep growth is actually, I think, a more coherent view than, uh, than we're done. Uh, and so you're either deep growth or for growth. There's nothing in the middle. And I think for growth is it has its downsides. Like everything that we talked about, it's scary because you just don't know what to teach your children.
我认为这根本行不通，因为资本主义是建立在增长基础上的。如果没有增长，就会出现衰退。所以你可以说我们实际上要倒退，认为技术是不好的，而“去增长”实际上是一种比“完成增长”更为连贯的观点。因此，你要么支持去增长，要么支持增长，没有中间状态。而且我认为增长也有其缺点，就像我们谈到的一切，都令人害怕，因为你不知道该教你的孩子什么。
And you're just like, things are gonna rapidly change. But I, you know, I'm much more optimistic about that and about the future of humanity than, uh, than a world where we sort of go back to a more primitive. You are a self-described civilizationist. Yes. Not even truck and pronounce the word. Uh, what, what is, what is that and why is it, um, important to you?
你就像在说，事情会迅速改变。但我对这种改变和人类的未来持更加乐观的态度，而不是回到更原始的社会。你自称是一个"文明主义者"。是的，甚至都不知道怎么正确发这个词。这是什么意思，为什么对你来说很重要？
Yeah, I've always kind of struggled to kind of describe my, my world viewer outlook. And the, I think the best way I could, I could think about it is like, um, I'm like, in awe of civilization. And I feel like if you restart the world, um, you know, hundred times, perhaps, uh, only the minority of time will get to the level of civilization that we're in. Um, you know, it's, it's quite hard, uh, you know, to go from a single cell organisms to, to, to, to, to multiple cell organisms. And it's quite hard to get to more complex. And there's all these ways in which, uh, evolution could have stopped.
是的，我一直以来都有点难以描述我的世界观或者说我的人生观。我觉得最好的方式来表达它就是，我对文明感到十分敬畏。我认为，如果让世界重新开始一百次，也许只有少数情况下能发展到我们现在这样的文明水平。从单细胞生物发展到多细胞生物是很困难的，进化到更复杂形式更是如此。在这个过程中，有许多地方进化都可能停滞不前。
But I do think no one talks about how, you know, you could have, you know, a human civilization could have stopped at, um, sort of the stone age or the agricultural age. There's no reason for it to actually continue developing further. Um, which he did for centuries. Yeah. And I think, uh, we take that for granted. And I think this is the sort of the de-growth versus growth mindset. Um, COVID showed how fragile civilization is. In some sense, you know, civilization is, is, is very robust.
我确实认为，没有人讨论过人类文明本可以停留在石器时代或农业时代的可能性。事实上，并没有理由一定要继续发展下去。而人类文明确实这样发展了几个世纪。我觉得我们对此习以为常。这体现了一种关于退化与增长的思维方式的区别。新冠疫情暴露了文明的脆弱性，但从某些方面来看，文明又是非常强大的。
But, uh, but, you know, um, you had the, the, the disruption and the, and the, and shipping and, and that caused all sorts of like downstream effects. And, um, and there were like a few instances where I just like felt like, oh, like actually like, you know, you could easily disrupt the global supply of food and like, people will die. Right. Or, or like just the fact that there's like a virus going around and like spread super fast and, and any governments are dysfunctional. They're not how to deal with it.
但是，你知道，由于运输出现了中断，这导致了各种下游影响。有几次我感觉其实很容易就能扰乱全球的食品供应，而这可能会导致人们死亡。还有一个事实就是，有一种病毒在快速传播，各国政府却没有能力有效应对。
And they either become super draconian and, or they, they're kind of, you know, just like, put their hands up and we're like, we're not going to deal with it. And, um, and, and, and, and there was like this massive distrust and governments and institutions and, um, and I think going through that phase, I just felt like, um, the thing I, I value the most and the, the thing that, um, I think we should be, um, we should, uh, protect and, um, and this also gets the techno optimist to continue to develop as human civilization.
他们要么变得非常严苛，要么就是举手放弃，表示不愿处理这个问题。这导致了对政府和机构的极大不信任。在经历这个阶段时，我觉得我最看重的是，我们应该保护的东西，同时这也是让科技乐观主义者继续推动人类文明发展的动力。
And to the government dysfunction and, and, and all the things seem to be, um, very energized by the new administration and that, uh, the next reason what, what is the, what is your hope? What, what, um, what does success look like for this new administration? So if you look at the Biden administration and you, if you look at what happened to the world, just look at geopolitics, um, uh, you have the Ukraine-Russia war, which, you know, I know people are very sensitive about this, but I do believe that, you know, we, we bear some responsibility to, to kind of exaggerate in this issue and make it, and like we made sure that there wasn't, um, any resolution to it and, and, and, and there was such a, um, motivation to kind of keep it going.
政府功能失调等问题似乎在新的行政当局上得到了很大的推动。那么，请问您对这届新政府的希望是什么？对于拜登政府，看看全球局势，包括地缘政治，比如俄乌战争。尽管我知道人们对此非常敏感，但我确实认为，我们在某种程度上对这个问题的扩大化负有责任，我们确保没有解决方案，并且有很强的动力去维持这种状态。
Our relationship with China got, got a lot worse. It's kind of surprising because like Trump, uh, sort of started that, um, and, but Biden like really took it to the next level and, uh, and there were moments where they could have gone on the, on the right track and there's like small things where, you know, President Xi visited us and Biden same day kind of insulted him and, um, uh, and the situation in the Middle East, I think could have been resolved if Biden was stronger and put some pressure on, on both parties to kind of resolve the situation. And America actually has a lot of power to be able to do that. But, um, but the Biden administration was a combination of both being, um, weak and, um, and so it was just like not having a vision for the world that's, that's coherent.
我们与中国的关系变得糟糕了很多。这有点令人惊讶，因为像特朗普那样的人是最初开始这一切的，但是拜登实际上将其推向了新的高度。有些时候，本可以走上正确的轨道，但却发生了一些小事情，比如习近平主席访问我们，而拜登同一天有些冒犯了他。我认为，中东的局势如果拜登更强硬一些，对双方施加压力，可能已经得到解决。美国实际上有很多能力可以做到这一点。但是，拜登政府在这方面显得软弱，同时又缺乏一个连贯的全球愿景。
Um, obviously Trump, there's all these downside to him, but I think I do think that he's a, uh, powerful force in, in G of politics because, um, I think he's both a deal maker and, uh, projects American strength, uh, such that, um, uh, a hostage and ceasefire deal was done even before he took office exactly what he said. And, and I think, you know, he remains to be seen. There's so many ways it could go wrong. And from a, from a tech and sort of Silicon Valley, uh, perspective what, um, uh, what do you think is important? Yeah.
嗯，显然，特朗普有很多负面之处，但我确实认为他在全球政治中是一个强有力的力量，因为我觉得他既是一个善于交易的人，也展现了美国的实力，以至于在他上任前就达成了一个人质和停火协议，就像他说的那样。我认为还有很多不确定性，事情可能会有很多不如意的地方。从科技和硅谷的角度来看，你觉得什么是重要的？
Um, uh, I think, I think, you know, without a stable world, there's no tech. And I think, uh, if you want techno optimism, you have to really care about geopolitics. You have to care about a stable world because, because, you know, um, it's, it's all connected, you know, that's the world that we built. It is a global world. It is globalists, right? So you can't return to this idea of like, oh, you know, I'm a big fan of reindustrializing the US, but like the dependence on China is, is a reality. And, um, and there's no reason for them to be really an adversary.
嗯，呃，我觉得，你知道，没有一个稳定的世界，就没有科技。我认为，如果你想对科技持乐观态度，你必须真正关心地缘政治。你必须关心一个稳定的世界，因为，这一切都是相连的，你知道，这就是我们建立的世界。它是一个全球化的世界，对吧？所以你不能回到那种认为我们可以完全依赖自己的想法。我很支持重新工业化美国，但对中国的依赖是一个现实，而没有理由让他们成为一个真正的对手。
Um, and, and so in terms of like, uh, the Trump's, uh, kind of view on AI, um, I think, uh, you know, just, you know, they're ready, they're ready to hammer to the regulate to regulations with executive orders. Um, I think the Biden administration was heavily influenced by the sort of effective altruism movement. And they, you know, spent a lot of money lobbying and everything. And, um, they were really trying to pause or slow down AI development that resulted in the executive order. And I think a lot of other things that the Biden administration was excited about doing.
嗯，关于特朗普对人工智能的看法，我觉得，他们已经准备好通过行政命令来加强监管。我认为，拜登政府受到了有效利他主义运动的很大影响，他们花了很多钱进行游说，试图暂停或减缓人工智能的发展，这最终导致了这项行政命令。我认为，这也是拜登政府对其他许多事情充满热情的原因。
And, uh, you know, now we have projects to our gate. 500 dollars recording this, uh, the day after the announcement, yeah, 500 billion, uh, uh, announcement. Yes. Yeah. And, um, and, you know, uh, it seems like Trump played a big role in that and in attracting investments and, um, and I think it sounds like, uh, you know, there's going to be a lot of, um, energy around space and getting to Mars. And I think that'll also help with industrialization and it'll help with, um, with the U S becoming like competitive again with, with the kind of what Elon talks about all the time, which is you need to be able to look at the future and be excited about it. Um, and, um, and yeah, those were the, the primary things.
嗯，你知道，现在我们在我们的项目上有了一些进展。录制这段话的时候，刚好是在宣布后的第二天，那笔投资是5000亿美元。是的。看起来，特朗普在吸引投资方面发挥了重要作用。我认为，这似乎意味着将来会有很多关于太空和前往火星的热情。我觉得这也有助于工业化，并帮助美国再次成为竞争强国。就像埃隆常说的，你需要对未来感到兴奋。这就是主要的内容。
I think that, um, I think just generally, the mood in Silicon Valley is also better. Um, I think we got to a point where, um, it got very focused on identity politics. And I think that's something also, you know, the Democrats, really pushed and excelled at. Um, and you see, Zuck, you know, uh, sort of bringing fast Facebook back or meta back to the hacker culture, uh, and, uh, and reversing a lot of this sort of, identity based hiring and identity based, um, uh, sort of, uh, uh, promotions and, and all that stuff. Uh, and he went on a road gun and he said, uh, you know, we need more masculine.
我觉得，总体而言，硅谷的氛围也变得更好了。我认为我们曾经非常关注身份政治，而这是民主党推动并且很擅长的一件事。你可以看到，扎克伯格试图把Facebook或者Meta带回黑客文化，逆转很多基于身份的招聘和晋升等等。他曾说，希望看到更多的阳刚之气。
And that company, uh, which was a amazing meme fodder. Yes. Uh, but you could tell that, you know, you know, I came to Silicon Valley so we could like, you know, yell and sell computers, right? And, and so just like, like bringing back the hard charging energy of Silicon Valley. And that doesn't mean like we want to alienate anyone in fact, you know, uh, I think it's gonna still be, uh, still attract the best talent from, from all over the world. But, um, but it needs to be energetic and it needs to be exciting. And, and, and the, uh, side effect of this administration is actually making it possible for, for, for, for entrepreneurs to, to be, to be in founder mode. It feels like a wonderful place to live it. Yeah. I'm glad. Thank you so much.
那家公司，嗯，曾一度成为网络笑谈的素材。是的。你可以看出，我来到硅谷的目的是为了激发我们的活力，大声宣传并销售电脑。就像要重新带回硅谷那种拼劲十足的能源。这并不意味着我们想疏远任何人，事实上，我认为它仍将吸引来自世界各地的顶尖人才。但它需要充满活力和令人振奋。而这一管理变革的附带效果实际上是让企业家们更容易进入创始人模式。这感觉就像是一个美好的生活地方。是的，我很高兴。非常感谢。
This was terrific. Of course. My pleasure. Thank you. Hi, it's Matt Turk, again. Thanks for listening to this episode of the mad podcast. If you enjoyed it, we'd be very grateful if you would consider subscribing if you haven't already or leaving a positive review or comment on whichever platform you're watching this or listening to this episode from this really helps us build a podcast and get great guests. Thanks and see you at the next episode.
这真是太棒了。当然。这是我的荣幸。谢谢你。你好，我是Matt Turk，又来了。感谢你收听这一集的疯狂播客。如果你喜欢这一集，我们将非常感激，如果你考虑订阅（如果还没有订阅的话），或者在你收听或观看这一集的平台上留下积极的评价或评论。这对我们扩展播客、吸引优秀的嘉宾非常有帮助。谢谢你，我们下一集再见！