Replit CEO Amjad Masad 谈 AI 末日论、软件与氛围编程

大家好，欢迎收看 Times Tech 的访谈。我是 Katie Prescott。

Hello and welcome to the interview from Times Tech. I'm Katie Prescott.

我是 Danny Forson。本周我采访了 Replit 的创始人兼 CEO Amjad Masad。

And I am Danny Forson. And this week I have been speaking to Amjad Masad, the founder and CEO of Replit.

Replit 在硅谷可是个大热门，对吧？他们是一个 vibe coding 平台，有点像我们这边的 Lovable。所以我想它的理念就是，你告诉他们你想构建什么，一个应用或一个软件，然后它就能自动生成。

And Replit is a really big deal in Silicon Valley, right? They're a vibe coding platform, a little bit like our lovable over here. So I guess the idea is you can tell them what you want it to build, an app or a piece of software, and then it just spins it up.

是的，没错。Amjad 经常谈到我们如何失去了计算的乐趣，因为现在没人编程了。事实上，他说我们都“受制于硅谷科技”。他对当前盛行的这种末日论氛围有一些非常有趣的看法。而且他很有发言权，他在这个领域已经很久了。他还解释了为什么他认为当科技公司的智能体失控时，它们应该面临处罚，这又是一个多么新奇的概念。但我想你会很享受这次对话。Amjad 很棒。那么，有请 Replit 的 Amjad Masad。

Yes. Exactly. And Amjad speaks a lot about how we've kind of lost the fun in computing because no one programs anymore. In fact, he says we are all quote at the mercy of Silicon Valley tech. And he has some really interesting things to say about, you know, this whole doomer vibe that is so dominant right now. And he knows of what he speaks. He's been in this world for a very long time. And he explains why he thinks tech companies should face penalties when their agents go rogue, which again is just like what a concept. But I think you're really going to enjoy the conversation. Amjad's great. So, here he is. Amjad Masad from Replit.

Amjad，非常感谢你抽出时间。我因为常驻硅谷并报道科技，所以知道 Replit。但对于不了解的人，Replit 是什么？你在这个领域深耕多久了？

Well, Amjad, thank you so much for taking the time. I know Replit because I'm based out here in Silicon Valley and I cover tech. But for those who don't, what is Replit and how long have you been hacking away at this particular coalface?

Replit 是将软件想法变为现实的最佳方式。我想如果你曾经有过一个应用、自动化或软件智能体的想法，现在你可以构建它，你可以构建任何复杂度的应用，而且在 Replit 上特别重要的是你可以扩展它们，你可以通过它们盈利，普通人和软件工程师、受过训练的软件工程师之间的能力差距越来越小。我作为一家公司已经做了 10 年，但作为一个想法可能已经 15 年了。它始于我对编程工具如此难用、如此难以触及、如此古老的沮丧。令人惊讶的是，即使有了 AI，它们仍然如此。

Replit is the best way to bring software ideas to life. I think if you've ever had an idea for an app or an automation or software agent, you know now you can build it and you can build apps of any complexity and especially important on Replit you can scale them you can monetize them and there is increasingly little difference between what any ordinary person is capable of and a software engineer and a trained software engineer. I've been working on this as a company for 10 years but as an idea perhaps for 15. It started with my frustration with programming tools being so hard to use, so inaccessible, so archaic. Surprisingly, they still are even with AI.

我的意思是，你必须打开终端才能使用像 Claude Code 这样的东西，我认为这简直是个悲剧。终端，或者说是终端模拟器，对吧？所以它们在模拟 1950 年代的技术。所以这就像搞什么鬼，对吧？

I mean, the fact that you have to open a terminal to use something like Claude Code, I think is just a travesty. Terminals or as they're like terminal emulators, right? So they're emulating 1950s technology. So it's like what the hell, right?

所以软件开发一直停留在过去，这很奇怪，因为软件工程师在构建未来。所以我一直对软件工具有一种挫败感，软件编程工具应该是令人愉悦的。它应该是协作的。应该是有趣的。它不应该和 Canva 或使用设计工具有什么不同，终于我们做到了。我认为 Replit 用起来很有趣。

Um and so like software development has always been stuck in the past, which is weird because software engineers are building the future. And so there's always been a frustration that I had with like software tools, software programming tools should be joyful. It should be collaborative. Should be fun. It shouldn't be any different than Canva or using design tools and finally we're there. I think Replit is a lot of fun to use.

那么我们能回到你刚才说的，你思考这个想法已经 15 年了。你为此工作了 10 年。我们都知道 ChatGPT 时刻以及现在发生的事情，当你听到 Anthropic 谈论基本上编码人员做的编码越来越少，感觉就像已经越过了卢比孔河，但允许像我这样完全不懂技术的人或其他任何人创建软件、创建工具、做他们想做的事情而不需要专业知识。10 年前那是什么样子？因为感觉那几乎是不可能的，对吧？如果你回到 ChatGPT 之前的时间。

So can we go back to you know you said you've been thinking about this idea for 15 years. You've been working on it for 10. And we all know the ChatGPT moment and kind of what's happening now when you hear Anthropic talking about basically you know coders are doing less and less coding and it does feel like you know that Rubicon has been crossed but like the idea of allowing me deeply untechnical or anyone else kind of create software, create tools, do whatever they want to do without that expertise piece. What did that look like 10 years ago? Cuz it feels like that's like an almost an impossibility, right? If you kind of go back in time before ChatGPT.

实际上，我们可以追溯到更早的时候。如果你想想早期的计算先驱，他们从不区分受过训练的软件工程师和计算机用户。使用计算机的理念就是你要编程。这就是计算机的用途。这在 80 年代你买电脑时是真的，比如 Commodore 64，当你买它时，它启动到一个 BASIC 解释器，你被期望编程才能使用它。我认为从史蒂夫·乔布斯普及的这个想法开始，计算机应该更直观地使用，这是一个合理的想法，但它的意外后果是创造了两个计算机用户阶层。有超级计算机用户，即程序员，还有计算机消费者，即世界上的其他人，你依赖计算机程序员来构建你使用计算机所需的所有软件。这有各种各样的后果，比如商业经济后果，我们有一个非常庞大繁荣的软件产业。你可能认为这是好是坏取决于你的观点，但这是一个。但它创造了我认为可能是社会和政治影响，我认为是负面的，即操纵人们及其观点和看法的想法，我认为这是人们没有编程自由、受制于硅谷大型科技公司给你提供所有信息和软件的直接后果。所以我绝不是第一个阐述这些问题的人，但你知道我强烈地感受到了它们，我觉得使用计算机的方式实际上是编程，我就是这样使用计算机的，然后当 Windows 和 Mac 出现时，我只是觉得非常不确定，当 iOS 和 iPhone 出现时。你会说，我甚至不拥有这台电脑。它是一台电脑，但我不能编程。我必须去这个应用商店下载应用，而且它们大多很烂。

Actually, we can go back way more in time. And if you think about the early computing pioneers, they never saw a difference between a trained software engineer and a computer user. The idea of using computers was that you would program them. That's what computers are for. And that was true when you bought a computer, I think in the 80s, like at a Commodore 64, when you bought it, it booted up to a basic interpreter and you're expected to program it in order to use it. And I think starting I think popularized by Steve Jobs this idea of like computers should be more intuitive to use which is I kind of it was a reasonable idea but the unintended consequences of it that it created two classes of computer users. There's supercomputer users, the programmers, and there's computer consumers, the rest of the world where you're reliant on computer programmers to build all the software you need to use your computer. And this has all sorts of consequences like there's commercial economic consequence where we had like a very big thriving software industry. You might think that's good or bad depending on your view, but that's one. But it created I think potentially social and political impact that I think is negative where the idea of manipulating people and their opinions and their views and their I think is a direct consequence of people not having the freedom to program their computers and they are at the mercy of kind of Silicon Valley big tech companies to kind of feed you all the information and the software you need to do something and so you by no stretch of the imagination I'm the first one to kind of articulate these problems but you know I felt them sort of intensely and I felt like the way to use a computer was to actually to programming and that's how I used computers and then when I when like Windows and Mac you know came on the scene I just it felt very indefinitely when iOS and iPhone came on the scene. You're like, I don't even own this computer. Like it is a computer, but I can't program it. Like I have to go to this like app store thing to download apps and they're mostly drunk.

所以 2016 年我们创办公司时的样子是，让我们一次移除一个障碍。第一个障碍是设置开发环境，这意味着打开终端，下载 Python 或 Java 或其他东西并设置它。这实际上是一个相当大的障碍。你必须下载千兆字节的代码。

And so what it looked like in like 2016 when we started the company was let's actually remove one barrier at a time. So what's the first barrier is setting up the development environment and that meant opening the terminal downloading Python or Java or whatever and setting it up. That's actually quite a big barrier. You have to download gigabytes of code.

是的。

Yeah.

而且你必须学会如何使用终端。所以我们说好吧，让我们创建一个网站。我们称之为 Replit。它实际上是对第一个交互式编程环境的致敬，那个环境叫做 REPL。我们把它放在浏览器里，你去这个网站。最初，我们甚至没有登录屏幕。你只需去这个网站，那里有一个代码编辑器，你可以运行代码。

And you have to learn how to use a terminal. And so we're like okay let's create a website. So we call it Replit. It's a homage to actually the first interactive programming environment that's called a REPL. And we put it in the browser and you go to this website. Initially, we didn't even have a login screen. You just go to this website and there's like a code editor and you can run the code.

我其实没有太多预期。我不知道人们会怎么用它。我以为人们会像我一样，开始探索编程语言之类的东西。但最初的产品市场契合点其实是教育。

I actually didn't have a lot of expectations. I didn't know how people would use it. I think, like me, people will just start exploring programming languages and things like that. But the initial product-market fit turned out to be education.

所以是学编程的人，你消除了那道门槛。

So people learning to code, you remove that barrier.

想学编程的人就那么多，很多人注册 Replit 来学编程。然后下一件事就是，好吧，你写了个程序，怎么托管它？怎么部署它？当时有一些创新。你可能还记得 Heroku，Salesforce 收购了它，最后又把它关掉了。但那就像是，嘿，你在本地写了个程序，敲几个命令，突然它就在网上了。像 AWS，但简单得多。所以我们就在浏览器里构建了类似的功能。你写个程序，点一下部署，突然它就在一个公开域名上，你就有了一个应用。

So only so many people want to learn to code, and a lot of people signed up to Replit to learn to code. And then the next thing was, okay, you wrote a program, how do you actually host it? How do you deploy it? There was some innovation. You might remember Heroku, which Salesforce bought and eventually killed. But that was like, hey, if you wrote a program locally, you can type a few commands and suddenly it's on the web. It's like AWS but a lot simpler. And so we sort of built a similar functionality right in the browser. So you write a program, you hit deploy, and suddenly it's on a public domain and you have an app.

我就想，好吧，现在你有了一个应用，你需要存储数据，然后你需要做用户认证。所以我们一路做下来，基本上解决了所有这些不同的问题，包括安全和滥用问题——现在智能体正在突破沙箱，这些问题其实非常应景。我们做这件事已经很久了。

I was like, okay, now that you have an app, you need to store data, and then you need to do user authentication. So we went down the line and we've sort of solved all these different problems, including security and abuse problems, which are actually very topical now that agents are breaking out of sandboxes. So we've done that for a long time.

然后当 AI 出现时，我算是预测到了——其实我们发布了我 2016 年的幻灯片。我谈到 AI 将如何改变软件，让它更容易上手。但我在 2019 年看到 GPT-2 的那一刻，就觉得这很明显了。我们开始构建 GPT interlet。2023 年我们有了一个类似 ChatGPT 的聊天界面，你可以和你的 IDE、你的编程环境对话。2024 年，我们发明了现在所谓的 vibe coding（氛围编程），我们做了一个激进的决定，当时非常有争议。就是你甚至不需要看代码。我记得网上的人对此议论纷纷。

And then when AI came on the scene, I sort of predicted that — actually, we published my slide deck from 2016. I talk about how AI is going to change software and make it more accessible. But it was obvious to me the moment I saw GPT-2, actually in 2019. We started building GPT interlet. In 2023 we had a chat interface similar to ChatGPT where you can talk to your IDE, to your programming environment. In 2024, we invented, I think, what's called now vibe coding, which we made this radical decision, which was very controversial at the time. It's like you don't even need to see the code. And I remember people online about it.

在那之后三四个月，也许五个月，Karpathy 出来，算是创造了 vibe coding 这个词。然后——我记得那条推文，大家都说，天哪，世界变了。现在每个人都能 vibe code 了。

And three, four months, maybe five months after that, Karpathy came out and kind of coined the term vibe coding. And then sort of — I remember that tweet and everybody was like, oh my god, the world has changed. Everybody can vibe code now.

是啊。

Yeah.

没错。但这很棒。我是说，这是我长久以来一直梦想的。但现在 Replit 的使命演变了。我认为我们已经实现了「任何人都能编程」的使命。现在更多是，你如何部署你的作品？如何分享它们？如何扩展它们？如何在企业里维护它们？如何治理它们？如何控制成本？把这种能力赋予更多人，会带来太多二阶效应和复杂性，现在我们正在解决这些问题。

Exactly. But it's great. I mean, that's what I've been dreaming of for a long time. But now the mission of Replit sort of evolved. I think we've achieved the mission of anyone can program. I think it is more like, how do you deploy your creation? How do you share them? How do you scale them? How do you maintain them in the enterprise? How do you govern them? How do you control costs? There are so many second-order effects and complications that come with giving that power to a lot more people, and now we're solving those.

我很好奇，因为你做这件事这么久了，让我印象深刻的是——我们待会儿会聊一点 AI 末日论之类的东西，以及它对社​​会和就业意味着什么——但 AI 最大的解锁似乎是，它就像是随取随用的专业知识，对吧？就像你不需要——所以你不必是程序员才能编程，对吧？所以我很好奇，第一，你现在还写代码吗？你觉得这对软件意味着什么？因为叙事的一部分是，看，没人需要再写代码了。但另一方面，我猜你还是需要程序员，因为你也需要有人能说，嗯，这里出问题了。有些东西不工作了。我需要——你需要那个修理工。你需要有人打开引擎盖，然后说，哦，好吧，嘿，是你的化油器，或者你的风扇皮带之类的。但我只是好奇你怎么看这件事，以及你在世界上看到了什么。

I'm curious, because you've been doing this for so long, and the thing that strikes me — and we'll get into a little bit of the AI doomerism and all of that stuff and what it means for society and jobs and all that stuff — but it seems like the big unlock of AI is that it's kind of expertise on tap, right? It's like you don't — and so you don't have to be a coder to code, right? And so I'm curious, one, do you code anymore? How do you think this means for software? Because part of the narrative is like, look, no one needs to code anymore. But also, you still need coders, I would guess, because you also need to be like, well, something's going wrong here. Something's not working. I need to — you need the mechanic. You need somebody to open the hood and be like, oh, okay, hey, it's your carburetor or it's your fan belt or whatever. But I'm just curious how you think about that and what you're seeing out there in the world.

我认为你还是需要程序员，因为存在一个风险——我觉得这个风险没被怎么讨论过——就是我们不再理解这些系统。我认为你仍然需要有人理解这些系统。我认为如果你完全用 vibe coding 做了一个应用，而且就像你说的，有些东西真的出了大问题，我觉得你仍然可以通过足够的提示、AI 和足够的 token 投入来摆脱困境，但如果有能掀开引擎盖看一看的人，可能会更好。但我不知道这能持续多久，因为我认为在某个时候——现在我们有一点低效，就是 AI 在写高级编程语言。高级编程语言是为人类发明的。而在某个时候，我认为它会跳过中间人，直接生成目标代码或其他更高效的发明。所以我认为我们正处在某个节点，AI 编程的整个方式将会改变，然后也许你根本不需要程序员了。

I think you'd need coders because there is a risk — I don't think that's talked about much — but there is a risk in which we stop understanding the systems. And I think you still need people to understand the systems. I think that if you fully vibe coded an app and, like you said, there's something that's really deeply wrong, I think you can still get out of it with enough prompting and AI and enough tokens at the problem, but it's probably better if you have people that can peer under the hood. But I don't know for how long that will last, because I think at some point — right now we have this bit of inefficiency where AI is writing high-level programming languages. High-level programming languages were invented for humans. And at some point I think it cuts the middleman and goes straight to object code or some other invention where it's more efficient. And so I think we're at some point that the whole way the AI codes will change, and then maybe you don't need coders at all.

你不再写代码了。

You don't code anymore.

不，不，我很多东西都只是通过提示来构建。

No, no, I build a lot of things just by prompting.

我很好奇，作为专业人士和经营这样一家公司的人，这种交互对你来说是什么样的。OpenAI 最近发布了他们 Astra 模型的广告，我们前几周在播客里聊过，很有意思的是，他们做了一个微妙但又不那么微妙的选择——他们展示人们与模型互动。没有人坐在电脑前。每个人都只是在和机器说话，对吧？没人在打字。我觉得这非常有意思，因为语音、转录和即时转录现在太好了，所以他们像是在说，是的，越来越多的人只是在和机器说话，这打开了一个全新的宇宙：计算机是什么，我们如何与它互动。我只是想知道，这是你看到的吗？你们公司内部是这样发生的吗？

I'm curious how that interaction is for you again as a professional and somebody building a company like this. OpenAI recently came out with their ad for their Astra model and we were talking about it on the pod the other week, and what was really interesting was kind of a subtle, not-so-subtle choice of like they were showing people interacting with the model. No one was sitting at a computer. Everybody was just talking to the machine, right? No one was typing anything. And I thought that was a really interesting thing because the voice, transcription and instant transcription is so good now that they're like, yeah, a lot more people are just talking to their machines and it opens up this whole new universe of what a computer is and how we interact with it. I'm just wondering if that's what you're seeing. Is that how things happen inside your company or not?

有一些语音，但我没看到很多。我认为很多人会更愿意直接和他们的电脑说话。我的意思是，你开始从 Siri 看到这一点。我记得见过一些老年人完全通过 Siri 来使用他们的 iPhone，虽然它一半时间都不好用，但对他们来说就是容易得多。我认为会有很大一部分人只是和他们的设备说话。而且我认为会有新的设备。我不知道是不是 OpenAI 的新设备，但肯定会有完全由语音操作的新设备。这是肯定的。我认为电影《Her》是对未来走向的一个很好的预测，尤其是对于深度个人化的设备。

There's some voice, but I don't see a lot of that. I think a lot of people would prefer to just talk to their computers. And I mean, you start to see that with Siri. I remember seeing just older people who interact with their iPhone entirely through Siri, although it doesn't work half the time, but it's just much easier for them. And I think there's going to be large parts of the population that will just talk to their devices. And I think we're going to have new devices. I don't know if it's the new OpenAI device, but there's going to be new devices that are entirely operated by voice. That's for sure. I think the movie Her is a good prediction of where the future is headed, especially for deeply personal devices.

我觉得专业人士永远都会需要键盘。尤其是在开放式办公室里——想象一下你四处走动，所有人都在说话。

I think professionals will always have a keyboard. Especially with open offices — imagine you walk around and everyone's talking.

对，对，对。

Yeah, yeah, yeah.

这不太合理。你肯定不想在办公室里大声给 HR 口述那封邮件。

Doesn't make a lot of sense. You don't want to dictate that email to HR out loud in the office necessarily.

对，没错。

Yeah, exactly.

但我觉得你的问题背后有一个重要的点——关于 Astra，我认为没有被很好覆盖的一件大事是：计算机操作模型终于好用了。我们从 2024 年就有计算机操作模型了，但一直很糟糕。现在 Astra 第一次真正能操作 Excel 表格、能操作 PowerPoint、能操作所有这些。所以最大的变化，无论是语音还是聊天，就是越来越少的人会去打开 App。这会是一个非常有意思的变化，因为你可以直接找智能体。

But I think there's an important point underlying your question, which is — the big thing about Astra that I think wasn't covered very well is that finally computer-use models are good. We've had computer-use models since 2024 and they've been terrible. Now for the first time Astra can actually operate an Excel sheet, can operate PowerPoint, can operate all these things. And so the big change, whether it's voice or chat, is that less and less people will open apps. And that's going to be a very interesting change, because you can just go to the agent.

对，那是因为 App 经济规模相当大。你看苹果的服务业务单独拿出来，就相当于一家《财富》50 强公司，而它基本上就是 App Store。

Right, that's because the app economy is quite large. I mean, if you look at the Apple services arm alone, it's like a Fortune 50 company, and it's all basically the App Store.

百分之百。我们也在思考这对我们的业务意味着什么，因为我们以做 App 闻名。我觉得我们越来越多地开始让智能体直接解决问题。所以很多时候你可能输入一个像 App 一样的提示词，但实际上你是想让智能体直接去做。最近我用 Replit 做了一个 App 来管理我的邮件，因为邮件变得非常难管理，尤其是人们发给我的那些 AI 垃圾内容。后来我实际上不再用那个 App 了，而是直接用 Replit 对着我的 Gmail 写代码，通过编码智能体和其他计算机操作智能体完成我需要做的一切。

100%. And we're thinking about what that does to our business, because we famously make apps. And I think increasingly we're starting to get our agents to solve the problem directly. So often you might put a prompt that's shaped like an app, but actually you want the agent to just do it. Recently I built an app with Replit to manage my email, which has become really hard to manage, especially with all the AI slop that people send me. And I actually stopped using that app and started using Replit directly to write code against my Gmail and do everything I need to do via the coding agent and the other computer-use agents.

对，对。

Right, right.

再聊聊公司——你们融了不少钱。我记得是 4 亿美元，对吗？那是你们最新一轮。

And just a bit about the company — you raised a bunch of money. I think it was 400 million. Is that right? That was your latest round.

最新一轮，对。

Last round, yeah.

我当时很震惊。不是因为数字——这年头这算小钱了。但 4 亿美元是很多钱。而且这是 D 轮，也就是你们十年里的第四轮融资。我当时就想，哇。因为对一些更大的公司来说，我们都到 H、I、J、K 轮了。我看了些报道，看起来你们正朝着大约 10 亿美元的年化营收迈进。你看到的采用情况是怎样的？有没有一个例子能体现你们正在释放的东西？比如——我随口编的——路易斯安那州的一个渔民说，让我做个网站或 App 帮我卖鱼之类的。这在现实中到底是什么样子？

And I was stunned. Not at the number — that's small potatoes these days. But 400 million is a lot of money. And it was a Series D, i.e. your fourth funding round in a decade. Which I was like, wow. Because we're getting into the H, I, J, K's for some of these bigger companies. And I was looking at some of the coverage and it looks like you guys are headed toward about a billion dollars in annual run rate for revenue. What are you seeing in terms of just the uptake of this? And is there an example that kind of captures what you're unleashing? Like — I'm making this up — a fisherman in Louisiana who's like, let me create a website or an app that helps me sell my fish or whatever. What does this actually look like on the ground?

其实，英国有一家钓鱼运动公司完全是用 Replit 搭建的，据创始人说，现在大概值 2000 万美元。

Actually, there is a fishing sports company in the UK that's entirely built on Replit that's, according to the founder, probably worth $20 million now.

哇。谁能想到？

Wow. Who knew?

对。你就给我一个词——像即兴表演一样。给我一个词，我就给你讲一个我见过的公司的例子。

Yeah. So just give me a word — it's like improv. Give me a word and I'll tell you an example of a company I've seen.

说到英国，因为我们刚在那里开了办公室……

Speaking of the UK, because we were just there opening our office...

有位老师为她的学生做了一个 App，但结果发现它其实远不止对她的学生有用，现在上架 App Store 第一天就有 2000 次下载。我觉得现在多得多。我想她还没开始变现，但我觉得她会的。我们也有用户赚了几百万美元。不知道你有没有看到《纽约时报》关于那两兄弟的报道，他们正朝着十亿身家去。

There's a teacher that built an app for her students, but it turns out it's actually much more useful than just for her students, and now on the first day on the App Store, 2,000 downloads. I think a lot more right now. I don't think she's monetizing it yet, but I think she will. And we've had people who made millions of dollars. I don't know if you saw the New York Times coverage of the two brothers that were headed to a billion.

至少……哦，这是卖 GLP-1 的吗？

At least like... Oh, is this selling GLP-1s?

对，至少其中一部分是通过 Replit 管理的。

Yeah, at least a portion of that is managed through Replit.

对。

Right.

然后还有很多不一定变现的社会影响类 App。加拿大最近有火灾，有人做了一个社区 App，让人们协调互助。还有一个事件——我记得是达吉斯坦的洪水——人们为此做了一个 App。加沙船队项目，那个 App 就是在 Replit 上搭的。还有很多政治、公民参与类的应用——这挺酷的。然后是企业应用。在企业里，用例往往是：我们要构建一个全新的 App，所以想让产品团队去快速做原型，弄清楚我们下一个要做的业务或应用是什么。或者我们想让业务更高效，所以我们在构建很多内部工具或自动化，让业务运转得更好。我知道我们的一些客户有 Visa、AAX、Zillow、Meta，还有其他几家。

And then there's a lot of social impact apps that are not necessarily monetized. There were recent fires in Canada and someone built a community app where people coordinated to help each other. There was also an issue — I think a flood in Dagestan — people built an app for that. The Gaza flotilla project, the app was built on Replit. And there's a lot of political, civic engagement type of applications — that's pretty cool. And then there's the enterprise applications. At an enterprise, the use cases tend to be: we're building a fundamentally new app, so we want our product people to go prototype and figure out what's the next business or application we're going to build. Or we're trying to make our business more efficient, so we're building a lot of internal tools or automations to make the business run better. I know some of our customers are Visa, AAX, Zillow, Meta, and a few others.

为什么选伦敦？我问这个是因为我的搭档主持 Katie 就在伦敦，而且伦敦一直有点沮丧——就像，哦，我们只是 AI 列车上的行李，我们没什么掌控权，很多人担心人才流失。但为什么，全世界那么多地方，你们偏偏在伦敦开办公室？

Why London? I ask because my co-host Katie is based in London, and there's been a bit of a bummer in London — like, oh, we're just luggage on the AI train, we don't have a lot of control, there are a lot of worries about talent drainage. But why, of all the places in the world, did you open in London?

我觉得有一种观念是，如果你想参与 AI，你就得训练基础模型。但实际上基础模型正在某种程度上商品化，你可以直接拿一个开源模型，想的话微调一下就行。我觉得这种非此即彼的二分法不健康——要么你在训练基础模型，要么就不是。我认为基础模型会到处都能获得。美国还有开源 AI 公司在不断涌现。有 Thinking Machines，显然还有 Reflections 要推出新模型，再加上中国和其他地方。所以显然伦敦有原创的 AI 研究，通过 DeepMind。所以有很多 AI 人才在构建可能是下一波的基础模型公司。但也有增长非常快的 AI 应用公司，比如 Granola 就在伦敦。所以就 AI 而言，伦敦似乎至少是第二名，仅次于湾区。所以那里有很多人才。有一流的学校。它也有一些湾区没有的优势。

I think there's a sense in which if you need to be a participant in AI, you need to train the foundation models. In reality the foundation models are sort of commoditizing, and you can pick up an open source model off the shelf and fine-tune it if you'd like. I think there's this unhealthy dichotomy that you're either training the foundation models or not. I think there are going to be foundation models accessible everywhere. There are open source AI companies in the US that are still coming out. There's Thinking Machines, obviously Reflections coming out with a new model, in addition to China and other places. And so obviously there is original AI research happening in London via DeepMind. So there's a lot of AI talent that's building potentially the next wave of foundation companies. But there are also AI application companies that are growing really fast, like Granola is in London. And so as far as AI goes, London seems to be at least number two, after the Bay Area. So there's a lot of talent. There are top-tier schools there. There are also things it has over the Bay Area.

我觉得这里有更好的多样性，不只是背景或出身方面的。我就是喜欢想法稍有不同的人。湾区已经变成了一种单一文化。

I think there's better diversity, and not just of background or origin. I just like people who think a little differently. The Bay Area has become such a monoculture.

它是个回音室，非常——就像一个村庄和回音室合二为一。

It's an echo chamber, a very—it's like a village and an echo chamber all wrapped up in one.

是的，是的，完全同意。所以我觉得 AI 真的能从想法稍有不同的人身上受益，所以能招到更多这样的人会很棒。最后一点，我们在欧洲的企业业务不断增长，我们想有个地方可以称为总部，一个欧洲总部来服务这些客户，而设在英国感觉是理所当然的，那里的政界和政府部门一直非常非常欢迎我们。我的意思是，我对英国也有好感，因为我是个计算机历史迷，而计算技术可以说就是在那里由阿兰·图灵发明的。

Yeah. Yeah. Totally. And so I think AI really benefits from people who think a little differently, and so it'd be cool to hire more of these people. And finally, we have a growing enterprise business in Europe and we wanted a place to call headquarters, a European headquarters to service these customers, and it just felt like a no-brainer to be in the UK, and the politicians and the administration there have been very, very welcoming. I mean, also I have a soft spot for the UK because I'm a big computer history nerd and computing was sort of invented there with Alan Turing.

是的。

Yeah.

布莱切利园。

Bletchley Park.

我很好奇——你知道，黄仁勋今年早些时候说过一句让我印象深刻的话，他说如果他手下某位年薪 50 万美元的工程师没有用掉至少价值 25 万美元的 token，他会非常担心。这让我想到一个问题：好吧，如果我每月在某个模型上花个几百美元，如果我是一家公司，这能替代一个真人吗？换句话说，token 能取代人吗？我具体说的是软件领域，但如果你退一步看法律、白领工作和客服方面的就业担忧，它很快就变得非常庞大、非常可怕。我很好奇你的想法，以及你看到了什么。

I'm curious—you know, Jensen Huang earlier this year said something that has stuck with me, which was that if one of his $500,000 engineers is not using at least $250,000 worth of tokens, he'd be really concerned. And it kind of led me to this question of: all right, if I'm spending, I don't know, a couple hundred bucks a month on Model X, if I'm a company, can that stand in for an actual human? In other words, can tokens replace humans? And I'm talking specifically about the software context, but then if you step back and look at the job worries in law, in white-collar work, and customer service, it quickly gets very big and very scary. I'm just curious what your thought is and what you're seeing.

我觉得 Satcha 说得最好，他说有 token 资本，也有人力资本。我不认为它一定是替代。我认为它是补充。这也是我们在 Replit 的理念：我们打造的工具是为了让智能变得有用。我认为 token 本身并不有用。我们见过很多公司，CFO 和 CTO 公开说我们只是在毫无理由地烧 token。而你很容易就会这样。我的意思是，凭借快速的——我们会给你加上护栏。我们把智能体设计成一种方式，同时也在系统里内置了很多预先写好的代码和库，而如果你下载 Claude Code 之类的东西，你完全可以空转，花掉几千美元却什么也没做成，你在公司里经常能看到这种情况。我跟朋友们聊，他们说，是啊，有人用掉了 1 万美元的 token。他们完全不知道，因为他们让智能体跑起来就去睡觉了，第二天就发现价值 1 万美元的 token 就这么进了垃圾桶。所以我认为 token 资本需要非常理性地管理。

I think Satcha put it best when he says there's token capital and there's human capital. I don't think it's necessarily a replacement. I think it's a complement. And this is our philosophy at Replit: the tools we build are to make intelligence useful. I think tokens on their own are not useful. We've seen a lot of companies where CFOs and CTOs have gone publicly and said we're just burning tokens for no reason. And you can easily do that. I mean, with the rapid—we put you on guardrails. We put the agent in a way where we also provide a lot of pre-written code and libraries built into the system, whereas if you download Claude Code or something like that, you can totally spin your wheels and spend thousands of dollars doing nothing, and you see it in companies all the time. I talk to my friends. They're like, yeah, someone used $10,000 of tokens. And they had no idea because they put the agent to run and they went to sleep, and the next day it was like $10,000 worth of tokens just in the garbage. And so I think token capital needs to be managed very rationally.

嗯，这正是我觉得当下这一刻如此迷人的地方，尤其是随着 IPO 即将到来，感觉人们很快就变聪明了，因为你刚才提到的那点——Uber 说我们在四个月里花光了全部预算，做了一堆东西，结果对利润毫无影响。所以很多人从疯狂堆 token 变成了，等一下，我需要 AI 做的 98% 的事，我不需要花最高价去做。

Well, that's what I find so fascinating about this moment, especially with the IPO or IPOs coming, is it does feel like people have gotten very smart very quickly, because you had that—to your point around Uber being like, we spent our whole budget in four months and we built a bunch of stuff and it's not made a difference to our bottom line. And so you've had a lot of people go from token maxing to like, wait a minute, 98% of what I need AI to do I don't need to spend top dollar on that.

是的。

Yes.

我可以用开源模型。

I can use an open-source model.

是的。

Yeah.

所以你觉得前沿实验室会不会迎来当头一棒，因为越来越多的人、越来越多的公司说，你知道，我不需要在你那些 token 上花几十万美元。我只需要把它们用在非常具体、有针对性的用途上。

So do you think the frontier labs are in for a rude awakening as more and more people, as companies, are like, you know, I don't need to spend hundreds of thousands of dollars on your tokens. I just need to use them for very specific, targeted purposes.

我认为 OpenAI 明白这一点。我不确定 Anthropic 是否明白，你可以从他们的行为看出来。OpenAI 有个 Luna 模型，比 DeepSeek Flash 更便宜，目前也更好。所以他们有——顺便说一句，谷歌也有类似的东西，Gemini Flash 就是那种主力、便宜到不用计量的模型，但 OpenAI 现在在这个领域夺得了桂冠。而且我觉得当你跟 Sam 聊时，他会说存在这种分化：人们会用最便宜的模型处理大约 80% 的工作，而剩下 20% 的工作你需要最聪明的模型。这也类似于我们如何看待劳动力或人力资本。

I think OpenAI understands that. I'm not sure if Anthropic does, and you can tell by how they act. OpenAI has this Luna model, which is cheaper than DeepSeek Flash and better at this point. So they have the—I mean, by the way, Google had something like that, like Gemini Flash was this workhorse, too cheap to meter type model, but OpenAI has taken the crown there right now. And I think when you talk to Sam, what he'll say is that there's this bifurcation where people will go to the cheapest model for like 80% of the work, and for 20% of the work you need the smartest possible model. And that also resembles how we think about labor or human capital.

是的。

Yeah.

在任何公司里，你都有水平一般或熟练度普通的人。但接着你有超级专家，他们通常不超过公司的 20%。公司叫他们老虎团队之类的。我觉得模型也是一样，区别在于路由器——你可以自动完成，有一个智能模型在说，哦，这个任务我需要最智能的东西，这个任务我可以用最便宜的东西做。

In any company, you have people of average training or level of proficiency. But then you have the hyper-experts, and they're typically not more than 20% of the company. A company calls them the tiger team or what have you. I think the same thing with models, and the difference is the router—that you can do it automatically, where there's an intelligent model saying, oh, this task I need the most intelligent thing, this task I can do with the cheapest thing.

对，对，有道理。末日论是当下的流行口味。

Right. Right. That makes sense. Doomerism is the flavor of the day.

是啊，你知道，你对这一切都深有研究。你怎么看正在发生的事情？我要说——我这么说是因为我们大多数凡人根本不懂。你知道，在我看来这很疯狂，你就像在说，嘿，我们在造一个末日装置，同时还要以 2 万亿美元上市。这一切，你知道，我们可能都会死。这简直——就像，什么？所以我很好奇你觉得我们现在处在什么位置，而且鉴于你在这个行业待了很久，这一刻与以往相比有多独特、多不同？

Yeah, you know, you're deeply in the weeds in all of this. What is your view of what is happening? I will say—and I say that because most of us mere mortals just don't understand. You know, it seems insane to me that you're like, hey, we're building a doomsday device and we're also going to go public at $2 trillion. It's all, you know, we might all die. It just—it's like, what? So I'm just curious where you think we are, and given that you've been in the industry for a long time, how unique and different is this moment from previous ones?

是啊，这个问题有很多种切入方式。问题在于，如果你相信 AI 会杀死所有人，那你为什么还想造它？他们会告诉你，如果我们不造，那中国就会造，而中国不会在乎这些东西。因此，我们必须造。OpenAI 最初的做法是，你需要让它开放，这点我有点同意。但我认为公司内部很多人不这么想——众所周知，Dario 反对发布 GPT-2，因为 GPT-2 会杀死所有人。顺便说一句，GPT-2 现在能在你手机上运行。然后他离开，创立了 Anthropic，而在 Anthropic，他们说，好吧，我们要尽量不卷入这种竞赛动态，转而解决核心的对齐问题。结果他们实际上与 OpenAI 一起极大地加速了竞赛动态，所以他们做的和他们说的正好相反。

Yeah, there's so many ways to approach this question. And so the question is like, if you believe AI is going to kill everyone, why do you want to build it? Well, they'll tell you that if we don't build it, then China's going to build it and China's not going to care about this stuff. Therefore, we need to build it. And OpenAI's original approach is like, you need to make it open, which I sort of agree with. But I think a lot of people inside the company don't—famously Dario was against releasing GPT-2 because GPT-2 was going to kill everyone. GPT-2 now can run on your phone, by the way. And then he left and created Anthropic, and with Anthropic they're like, okay, we're going to try not to get into these race dynamics and instead we're going to solve the core alignment problem. And instead they actually hyper-accelerated the race dynamics with OpenAI, and so they did the opposite of what they said they were going to do.

These companies ultimately are commercial entities and they'll just follow the commercial logic, but they're dressing it up in this effective altruism sort of language. But there's deep tensions inside this. There's a contradiction where your employee base actually truly believes that AI is going to kill everyone — they've kind of come out publicly and said it — but you're also a commercial entity that's bound by fiduciary laws and you can't do crazy things. Therefore the tensions are now being heightened and I think these companies are going to have a lot of problems. They might have unionization at these companies. They might have sabotage at these companies. And I think eventually we'll get to a point where, similar to what happened in the so-called walk era of Silicon Valley where Brian Armstrong said if you believe in these certain things, you should leave the company. And I think at some point these companies will have to entirely decouple from the EA doomerism type of thing.

现在回到客观论断上，我认为没有好的论证能说明 AI 会杀死所有人。

Now to get to the objective claims, I think that there's no good argument for how AI would kill everyone.

那我们就此确认一下，你不认为我们都会死？

So just to cover that off, you don't think we're all going to die?

No, I do worry about the risk. I do worry about the risk of AI. I think AI used in warfare is already causing human suffering and destruction. There are cyber security concerns across the world, which is why I agree that they need to slow down and they need to fix their systems. They need to harden their systems, or the government needs to start applying penalties when an AI from one of these companies goes out and hacks a system. They need to at least get a slap on the hand, if not actually open a criminal investigation. So we need to take it seriously that agents are really powerful in certain training environments that are going to have objectives that are very dangerous, and that needs to be taken seriously. But we have existing laws to handle that. I mean, we should just apply the laws, which is to say if your company through negligence creates external liability, basically —

需要有责任追究。

There needs to be liability.

是的。

Yeah.

If you come back in, let's say it's 2026, September 2026, September 2028, what's the world look like for Replit if all — if you know, wind at your back? Assuming there is still a world, first of all, that we inhabit. Crystal ball it a little bit for what you guys — you just raised a bunch of money. You've hit this inflection point. You're going up like this. You have, I think, 50 million users, etc. What does it look like?

I would love for our impact to be an increased efficiency in the economy broadly. But one thing I've been really excited about is we have government customers, and I think seeing the impact of Replit and agents inside the government, you can imagine a lot more state capacity and efficiency, and I think the world could run in a better way. We'd have better products and services, our taxes would go a long way. There's a lot of entrepreneurial people in government that have not had the power to make change, and I think AI could potentially do that. I think AI could have the potential for people to be more civically engaged. I think it could improve democracy if people are empowered by AI agents to build things that could have impact on their community, on the country at large. I would love for that future to happen. I think that's really what I'm working towards: AI as a tool that accentuates human creativity and allows great people to have great impact on the world.

好吧，祝你好运，把我们推向积极而非消极的一面。

Well, I wish you luck nudging us toward the positive rather than the negative.

谢谢。

Thank you.

好吧，那是一次精彩的采访。我得说，我觉得最劲爆的部分是关于 Anthropic 的那些内容——你知道，就是那种末日论被包装在，我记得他说是，有效利他主义的语言里。

Well that was a fascinating interview. I have to say the spiciest bit I thought was the stuff about Anthropic — you know, this idea that it's sort of the doomerism is dressed up in this, I think he said, effective altruistic language.

是的，是的。

Yes, yes.

这正好契合我们目前讨论的很多内容，他们的 IPO 即将到来，你知道，那种让世界恐慌的感觉。

Which just plays into so much of what we're talking about at the moment with their IPO on the cards and you know the sort of freaking the world out.

没错。而且我确实认为，从商业角度看他们，我们在 Lovable 上也看到了。有趣的是，这类工具在积极的一面，让完全非技术的人也能做真正技术性的事情。

Exactly. And I do think, you know, also just like looking at them as a business, and we saw it with Lovable as well. It is interesting that these types of tools, you know, on the positive side of things, allow people completely non-technical people to do really technical stuff.

是的，这很有趣，因为这不是——这是我们甚至不再谈论的事情，你知道。但现在这就是事实，真的很有趣，就像任何人——你的孩子、我的孩子、你或我都能——

Yeah, which is funny because it's not — it was something we don't even talk about anymore, you know. But it's just a fact now, which is just really interesting that like anybody — like your kid, my kid, you or I can just like —

我说你和我，更不用说孩子们了。

I say you and I, let alone the children.

没错。现在都能做东西了。所以这挺酷的。但有趣的是，这已经变成了甚至不再被谈论或真正处理的事情。我们都在谈论的是我们都会怎么死，以及那种末日、末日论的叙事。

Exactly. Can like make things now. So that is quite cool. But it is funny that like that has become something that's not even talked about or addressed really. What we're all talking about is how we're all going to die and that doom, that doomerism kind of narrative.

但先把这个放一边。你说得对。我们现在都能编程。我们现在都能画画。

But put that aside. You're right. We can now all code. We can now all draw.

嗯。

Mhm.

在某种程度上还能写作。我不认为你和我用 AI 写作，但它确实让人们能起草相当不错的基本文档。

And to an extent write. I don't think you and I would write using AI, but it does allow people to draft pretty decent basic documents.

你知道，它给了我们所有这些非凡的技能。

You know, it's given us all these extraordinary skills.

是的。所以，非常有趣的采访。我真的很享受做这个。我想这周就到这里了。所以，这是你这周的第二集，因为这就是我们为你们，为我们的听众所做的。

Yeah. So, really fun interview. I really enjoyed doing it. And I think that's it for us this week. So, your second episode of the week, because that's what we're doing for you, for our listening public.

买一送一。或者免费。#free

Two for the price of one. Or for the price of zero. #free

感谢收听本期 Times Tech 的访谈，我们下周五将带来所有最新科技新闻。

Thank you for listening to this episode of the interview from Times Tech and we'll be back next Friday with all of the latest tech news.

下周见。

See you next week.

再见。

Goodbye.