AMP 的远程优先革命：为何本地开发已死

所以，我们现在在一个远程演播室里，和 Quinn 连线，他此刻在慕尼黑。Quinn，我完全不知道你随时会在哪里。看起来你们在路演上玩得很开心。

So, we're in a remote studio with Quinn, who's in Munich right now. Quinn, I have no idea where you are at any given time. It looks like you guys are having a great time at the roadshows.

我们把整个 AMP 团队聚集在慕尼黑，那真是太棒了。但现在我要回去了。

We gathered the entire AMP team in Munich and it was just incredible. But now I'm going back.

你的团队通常是远程工作吗？

Does your team usually work remotely?

是的，团队三分之一在欧洲，三分之一在美国，三分之一在澳大利亚。我们只有 20 个人，所以分散在世界各地。我甚至不知道大家工作时是什么时间。

Yes, a third of the team is in Europe, a third in the USA and a third in Australia. There are only 20 of us, so we are scattered all over the world. I don't even know what time people have when they work.

是的，这很有趣。我想说这有点跑题，但我认为当你运营一家远程公司时，主要原则是这样的：你原本花在办公室上的钱，仍然要花在远程会议上。正因为如此，大家才能聚在一起。你们这样做真的很酷。我知道你们在新加坡也开过这样的会，我很高兴听到这个消息。

Yes, it's quite fun. I would say this is a bit of an off-topic topic, but I think when you run a remote company, the main principle is this: the money you would spend on an office, you still have to spend on remote meetings. Because of this, everyone comes together. It's really cool that you do that. I know you held such a meeting in Singapore, and I am very pleased to hear that.

是的。是的，我们喜欢新加坡。新加坡的空调现在比德国好用一点，但我们也喜欢德国。

Yes. Yes, we love Singapore. The air conditioners in Singapore work a little better than in Germany right now, but we love Germany too.

好的，AMP 团队已经上过这个播客了，人们大多只在网上见过你们。我觉得 orb 和 orbs 现在很流行。也许你可以解释一下最近的转变，让大家跟上进度？

Okay, the AMP team has already been featured on this podcast, and people have mostly only seen you online. I think orb and orbs are trending right now. Maybe you can explain the recent transition and bring people up to speed?

是的，通过 AMP，我们想要探索新的视野。我们能做的最疯狂的事情是什么？我们如何移除旧功能，让 AMP 成为与智能体协作的最佳工具？如果你使用 AMP，你就是在以我们认可的方式使用它。你走在正确的道路上。这是我们对你的承诺。我们会移除一些不再有意义的功能。我们是最早移除编辑器扩展的团队之一。我们已经移除了很多功能。最近几个月有人开玩笑说，我们以移除功能而非添加功能而闻名。我认为这是正确的决定，因为现在有了 orbs——AMP 的使用方式——一切都在远程运行。你可以合上笔记本电脑。你可以同时做一百件事。这种工作方式改变了我们整个团队的工作方式，我们的客户在过去三到四周内对 orbs 的使用，超过了去年软件开发过程中所有变化的总和。而去年并不平静。

Yes, with AMP we want to explore new horizons. What are the craziest things we can do? How do we remove old features to make AMP the best tool for working with an agent? If you use AMP, you are using it in a way that we approve. You are on the right track. This is our promise to you. We will remove some features that no longer make sense. We were one of the first to remove the editor extension. We have removed many features. Some have joked in recent months that we're better known for removing features than for adding them. And I think it was the right decision, because now with orbs—the way AMP is used—everything works remotely. You can close the laptop. You can do a hundred things at once. This approach to work has changed the way we all work as a team, and our customers have used orbs more in the last three to four weeks than in all the changes to the software development process that occurred last year. And last year was not calm.

所以，因为 AI 发展如此之快，强迫用户用旧方式做所有事情的好处实际上是负面的。以前，技术周期持续 5 年或 10 年：你可以吸引一个用户，他们会留你十年。但现在，如果你不推动他们，不把他们拉向新的前沿，那么三个月后他们就会问：“嘿，这些白痴是谁，他们有什么产品？”“这过时了。”他们会因此看低你。所以这是我们的原则，我们明白 orbs——我们不是第一个有这个想法的。你知道，Devon——非常尊重 Devon。Devon 是第一个，当时他们还被嘲笑。比如，这行不通。

So because AI is developing so quickly, the benefit of forcing users to do everything the old way is actually negative. Previously, the technology cycle lasted 5 or 10 years: you could attract a user and they would stay with you for a decade. But now, if you don't push them, don't pull them to new frontiers, then in 3 months they will ask: "Hey, who are these idiots and what kind of product do they have?" "It's outdated." And they will think less of you because of it. So this is our principle, and we understand that orbs—we are not the first to have this idea. You know, Devon—huge respect Devon. Devon were the first when they were still laughed at. Like, it doesn't work.

是的。他们关于 orbs 这种方法重要性的看法是对的。我们想把这一切整合起来。不仅仅是智能体在云端远程工作。我们让一切无缝运行，你看到你的开发服务器就在门户中运行。你可以访问桌面，并行运行一切——它就能工作，这足以让你再也不想用本地开发了。所以，在我们看来，本地开发已经死了，这是我们所有人构建产品方式的一个巨大转变。

Yes. And they were right about the importance of this approach in orbs. We wanted to bring it all together. It's not just that the agent works remotely in the cloud. We make everything work effortlessly, and you see your development server just running in the portal. You can access the desktop, run everything in parallel—it just works, and that's enough to make you never want to use local development again. So local development, in our opinion, is dead, and that's been a huge shift in how we all build products.

我甚至在 2021 年写过关于当时 localhost 时代终结的文章。这反映了即使在 AI 出现之前，像 Facebook、Google 或 Uber 这样的大公司里的人工作方式也不同。他们的设置大多是 SSH 到一台机器上，或者在本地有一个非常轻量的副本，对应云端更强大的东西。我认为即使在 Stripe，一旦你有了足够好的基础设施，你的笔记本电脑就变成了一个薄壳——你可以在 iPad 上编程，因为那不重要，它只是进入云基础设施的输入设备。但即便如此，我认识很多在这些公司工作过的人，而且……

I even wrote something in 2021 about the end of the localhost era at that time. This was a reflection of how, even before the advent of AI, people at large companies like Facebook, Google, or Uber worked differently. They all had setups where you mostly SSHed into a machine or had a very lightweight local copy of something much more powerful in the cloud. I think even at Stripe, once you have a good enough infrastructure, your laptop becomes just a thin shell—you can program on an iPad, because it doesn't matter, it's just an input device into the cloud infrastructure. But even then, I know a lot of people who worked in those companies, and...

他们说，“嗯，这更慢，你必须牺牲很多。”事实上，在笔记本电脑上工作非常愉快；一切运行得更快。而有了智能体，当你可以远程运行它们时，你不会觉得错过了什么。你不再想使用本地开发。这不是因为你的公司安全部门说，“嘿，我们需要用虚拟机进行开发。”你只是意识到，“嘿，我已经两周没运行本地开发服务器了。”我的意思是，这是我几天前才意识到的事情。所以感觉确实不同。你确实需要投资云设置，但是，你知道，这没有人们说的那么难，现在你可以把智能体带进来。

They said, "Well, it's slower, and you have to sacrifice a lot." In fact, working on a laptop is very enjoyable; everything works faster. And with agents, when you can run them remotely, you don't feel like you're missing out on anything. You no longer want to use local development. It's not because your corporate security department says, "Hey, we need to use virtual machines for development." You just kind of realize, "Hey, I haven't run my local development server in two weeks." I mean, this is something I just realized a few days ago. So it really feels different. You really need to invest in a cloud setup, but, you know, it's not as difficult as people say, and now you can bring agents into it.

是的。我还注意到，你正在做这个转变，和编程行业里许多其他人类似。开源这么做。Conductor 这么做。令人惊讶的是，Codeex 还没有完全做到。我不明白。Claude 做了很多。以前，我猜，像是 Claude 云之类的东西，现在是 Claude tag，看起来 Cursor 也做了。很有趣，当大家突然决定云已经准备好了。

Yes. I also notice that you are making this transition similar to many other people in the programming industry. Open source does this. Conductor does this. Surprisingly, Codeex hasn't quite done that yet. I don't understand. Claude did a lot. It used to be, I guess, like the Claude cloud thing, and now Claude tag, and it seems Cursor did it too. It's very interesting when everyone suddenly decided that the cloud was ready.

嗯，我会说云去年就准备好了。出于某种原因，它今年才起飞。但我不知道你是否有任何元层面的考虑——这有点元，但你对为什么是现在有什么想法吗？

Hmm, I would say the cloud was ready last year. For some reason it took this year for it to take off. But I don't know if you have any meta-considerations—it's a bit meta, but do you have any thoughts on why right now?

智能体确实变得更擅长完成长期任务。而且，你知道，我认为问题是，当你看到智能体如此可靠地正确完成所有事情，次数多到你不再检查所有代码时，它解放了你的时间去做更多任务。我还认为，在十二月和一月期间，实际上，CLI 中的智能体变得非常流行，这只是下一步的先决条件。所以这比我见过的任何其他技术实施都要快得多。但一切仍然，你知道，以人的速度前进。即使有了“orbs”，也有很多人说，“为什么我需要这个？”因为我在笔记本电脑上有本地开发环境，一切都设置好了。或者他们可能会说，“嗯，在我的公司，如果我们必须审查代码，我有 20 个东西等着审查有什么意义？”你知道，我没有得到更多并行性的好处，所以所有这些其他改变都需要解决。

Agents have certainly become better at completing long-term tasks. And, you know, I think the thing is, when you see an agent so reliably doing everything correctly enough times that you stop checking all the code, it frees up your time to do more tasks. And I also think that over the course of December and January, actually, agents in CLI have become very popular, and that's just a prerequisite for this next step. So this is moving much faster than any other technology implementation I've ever seen. But everything still, you know, moves at the pace of people. And even with "orbs" there are many people who say, "Why do I need this?" Because I have a local development environment on my laptop, and everything is already set up. Or they might say, "Well, at my company, if we have to review code, what's the point of me having 20 things waiting to be reviewed?" You know, I don't get the benefit of more parallelism, so it's all those other changes that need to be worked out.

我们 AMP 团队有 20 个人；我们都是联合创始人，在这件事上志同道合，我们受益于更大的并行性，远程做更多事情，因为我们不必等待验证。我认为越来越多的团队正在转向这种方法，但是，你知道，我们有点能感觉到其他团队在 6 个月后会如何运作。

There are 20 of us in the AMP team; we're all co-founders, all like-minded on this, and we benefit from greater parallelism, doing more things remotely because we don't have to wait for verification. And I think more and more teams are moving to this approach, but you know, we kind of get a feel for how other teams will operate in 6 months.

等等，呃，你的代码检查死了？

Wait, uh, your code check is dead?

代码审查在进入主分支之前是必需的。是的，它死了。自从我们开始做 AMP 以来，它就已经死了。

Code review is required before entering the main branch. Yes, she is dead. She's been dead since we started working on AMP.

这是一个非常响亮的声明。

This is a very loud statement.

呃，我猜你也投资了系统来弥补缺乏代码审查的问题，对吧？

Er, I guess you also invested in systems to compensate for the lack of code review, right?

是的，没错。我的意思是，最重要的系统是一个真正值得信赖的团队，而且，你知道，AI 的某些作用意味着你可以拥有一支更值得信赖、更有动力、从头到尾承担更多责任的团队。他们不是从产品经理和某个冲刺中获取输入，然后把结果输出给市场团队发布的自动机器。人们更多地参与整个过程。他们负责任，这是避免代码审查最重要的事情。但是，你知道，然后你想要达到这样的程度：你在日志中看到某些东西或收到客户投诉后 15 分钟就能发布修复。而且，你知道，这并不适用于所有软件。这是一种“平均恢复时间”与“平均故障时间”的方法，但对于发展如此之快的终端用户软件来说，这对我们来说是正确的方法。而且我认为，你知道，这比我们必须等几天才能发布东西的质量更高。

Yes, that's right. I mean, the most important system is a team that's truly trusted, and, you know, some of what AI does is mean that you can have a team of people who are more trusted, who are more motivated, and who have more accountability from start to finish. These are not automatons that take input from a product manager and some sprint and then output the results to the marketing team for release. People are much more involved in the process. They are responsible, and that is the most important thing to have to avoid code review. But, you know, then you want to get to the point where you can release a fix 15 minutes after you see, you know, something in the logs or get a complaint from a customer. And this, you know, is not at all suitable for all software. It's, you know, a "mean time to recovery" versus "mean time to failure" approach, but for end-user software that's evolving so quickly, it's the right approach for us. And I think that's, you know, it's higher quality than if we had to wait, you know, a few days to release something.

好。我想回到这次对话的起源，那就是你说你的工作方式改变了很多。你知道，显然，代码审查的一件事是它有点像信任的基本原则，我觉得我非常喜欢。还有什么改变了，或者当你说我们改变了很多，你只想分享这个，而我说“好吧，行，我陪你走这条路”时，你是什么意思？

Good. I want to go back to the origins of this conversation, which is that you said that the way you work has changed a lot. You know, obviously, one thing about code review is that it's kind of a foundational principle of trust, which I think I really like. What else has changed or what did you mean when you said we've changed so much and you just want to share that and I'm like, "Okay, fine, I'll walk this path with you."

我认为我们看到智能体不仅改变了我们构建软件的方式，也改变了我们运行和使用软件的方式。你会看到我这里非常混乱的工作进展。你知道，我要给你看这个我们称之为“应用”的东西——这是一种快速部署在隔离环境中运行的自定义软件的方式。那里有一个智能体，你可以简单地给它一个任务，例如，让它立即修复程序中的错误。那么，让我们打开这个。这是引擎盖下发生的事情：这可能是用户对 AMP 的主要抱怨——你必须等待沙箱启动。它启动一个“沙箱”，其中运行一个进程，通过 HTTP 通信，我只看到结果。这是我们团队成员创建的作为情境仪表盘的东西。在过去，比如两年前，你必须从分析团队找人来为 Looker 这样的服务添加新的 SQL 查询。而中间地带是：“我们为什么不把我们的仪表盘工具做成兼容 MCP，这样它就可以与我们的智能体交互？”但为什么这甚至是必要的？你得到绝对的定制。你不需要设置屏幕。你不需要复杂的集成、MCP，以及购买现成仪表盘包的所有成本，而事实上智能体是完美的设置屏幕，代码是配置任何软件的最佳方式。这是一件大事，我认为这只是一个时代的开始，智能体正在改变我们使用、运行和分发软件的方式。所以我们试图理解这到底会导致什么。我们已经有一些经验。这是大趋势之一。

I think we're seeing agents change not only how we build software, but also how we run and use it. You'll see my very messy work in progress here. You know, I'm going to show you this thing we call applications—it's a way to quickly deploy custom software that runs in an isolated environment. There is an agent there that you can simply give a task to, for example, ask it to instantly fix errors in the program. So, let's open this up. Here's what's going on under the hood: This is probably the main complaint users have about AMP—you have to wait for the sandbox to start. It launches a "sandbox" in which a process runs, communicating via HTTP, and I just see the result. This is something that one of our team members created as a situational dashboard. In the past, say two years ago, you would have had to bring in someone from the analytics team to add a new SQL query to a service like Looker. And the middle ground is: "Why don't we make our dashboard tool compatible with MCP so it can interact with our agent?" But why is this even necessary? You get absolute customization. You don't need a settings screen. You don't need complex integrations, MCPs, and all the costs associated with buying ready-made dashboard packages, when in fact the agent is the perfect settings screen and code is the best way to configure any software. This is a huge thing, and I think this is just the beginning of an era where agents are changing the way we use, run, and distribute software. So we're trying to understand what exactly this leads to. We already have some experience. This is one of the big trends.

我们录音前刚刚谈到这个——你对这一切怎么看？特别是用代码构建的内部工具呢？

We were just talking about this before the recording—what do you think about all of this? What about internal tools built with code, specifically?

是的。是的。而且它带来了什么，你知道，而不是带有大量设置和集成的软件。

Yes. Yes. And what does it give, you know, instead of software that comes with a bunch of settings and integrations.

我明白了。为最终用户提供个性化软件。

I understand. Personalized software for the end user.

你看到 Cloudflare 刚发布的东西了吗？

Did you see what Cloudflare just released?

哦，是的。那个为智能体打造的操作系统。他们称之为 Cloudflare OS。Kenton Wu 实际上称之为 Sandstorm V2——那是他以前的创业公司。

Oh, yes. That thing with the OS for agents. They call it Cloudflare OS. Kenton Wu actually calls it Sandstorm V2—it's his former startup.

我记得 Sandstorm，在过去几个月我们开发它的时候，我每天都在想它。对于那些不知道的人——我实际上甚至……我的意思是，我在科技领域，但我不知道它。你能简要回忆一下 Sandstorm 对你来说是什么吗？

I remember Sandstorm and thought about it every day for the last few months while we were developing it. For those who didn't know— I actually even... I mean, I was in the tech sphere, but I didn't know about it. Can you briefly recall what Sandstorm was for you?

是的，Sandstorm。它就像一个自托管的 Web 应用，你可以在其中上传打包在隔离容器中的应用程序。它们有几种交互方式，所以你可以在自己的服务器上安装办公套件、电子邮件客户端或类似的东西。它们有有限的 API 来执行任务，就像在操作系统中一样，而且一切都是完全容器化的。这是拥有自己的软件、创建自己的内部个人云环境的想法——它超前于时代。

Yes, Sandstorm. It's like a self-hosted web application where you can upload applications packaged in isolated containers. They have several ways to interact, so you can install an office suite, email client, or something similar on your own server. They have limited APIs to perform tasks, like in an OS, and it was all fully containerized. This was the idea of owning your own software, creating your own internal personal cloud environment—it was ahead of its time.

真的。我从你身上看到的有点类似于你所说的：如果你能自己改变界面，为什么还需要设置？我认为人们能够访问可以被 fork 的软件很重要，即使它不是完全开源的。只要前端和后端之间的契约稳定，并且维护了安全性和隐私，你就可以做任何你想做的事——它只是一个界面，没人在乎。所以我看到这种被称为“小程序模式”的东西在许多公司出现，不仅仅是你的公司，你可以创建一个应用程序，在公司数据和用户之间，以及应用程序之间的重用障碍更少。然后人们可以 fork 它，就像他们现在用 Replit 或其他东西做的那样。

Really. What I see from you is a bit similar to what you were saying: why do you even need settings if you can change the interface yourself? I think it's important that people have access to software that can be forked, even if it's not fully open source. As long as the contracts between the frontend and backend are stable and security and privacy are maintained, you can do whatever you want—it's just an interface, no one cares. So I see this called the mini-app pattern emerging in many companies, not just yours, where you can create an application with fewer barriers between the company's data and the user, and between applications for reuse. And then people can fork it, like they are doing now with Replit or something else.

作为企业主，似乎每个人最终都在构建同样的东西。我们都在试图弄清楚，而你把整个软件行业压缩成了这个每个人都称之为“智能体”的小泡沫。但你知道它会再次膨胀，我们都会看到不同类别之间存在的边界。但这正是它迷人的地方。没人知道。

As a business owner, it seems like everyone is ultimately building the same thing. We're all trying to figure it out, and you've taken the entire software industry and squeezed it into this tiny bubble that everyone calls "agents." But you know it will expand again, and we will all see the boundaries that exist between different categories. But that's what makes it fascinating. Nobody knows.

哦，高兴。是的。我认为重要的是你以一种破釜沉舟的方式建立了公司。我不知道怎么说得恰当，但我们不想被过时的解决方案挟持。我们真的、真的努力站在进步的前沿和先锋。所以那些认同这一点的人——无论是员工还是客户——正是我们需要的客户，这对我们很重要，因为有另一种类型的客户说：“我已经用了这个 10 年了，什么都别动，好吗？”“我接受过这个工具的培训和认证，别碰它，因为我太喜欢它了”——那是一种完全不同的客户。

Oh, joy. Yes. I think what's important is that you built the company in a way that burned bridges behind you. I don't know how to put it correctly, but we don't want to be held hostage to outdated solutions. We really, really strive to be on the cutting edge of progress and in the vanguard. So the people who share this--both among employees and among customers--are exactly the customers we need, and that's important to us, because there's a different type of customer who says, "I've been using this for 10 years, don't touch anything, okay?" "I'm trained and certified in this tool, don't touch it because I love it so much"--that's a completely different type of customer.

是的，看到你，Sourcegraph 的前 CEO，转向了完全不同的防御领域，这很有趣。

Yes, and it's very funny to see how you, the former CEO of Sourcegraph, have gone to a completely different side of defense.

是的，完全。嗯，我认为我们需要吸引人们，这就是我们在这里的原因。另一件事是看看这些智能体如何不仅帮助你编写代码，还帮助你处理运营部分。还有另一类工作：在你发布某些东西后，你需要确保它可以被实施。让我找一个好例子。好。我在 AMP 中，你可以看到阶段 1B2。好的。那么这是什么？这实际上是 AMP 从一个用户只能在一个工作区工作的系统，过渡到在许多工作区工作的系统。

Yes, completely. Well, I think we need to engage people, and that's why we're here. Another thing is to see how these agents can help you not only with writing the code, but also with the operational part. There's a whole other category of work: after you've released something, you need to make sure it can be implemented. Let me find a good example. Good. Here I am in AMP and you can see phase 1B2. OK. So what is this? This is, in fact, AMP's transition from a system where a user could only be in one workspace, to working in many.

任何写过软件的人都知道，这类改动会触及最关键数据模型的核心，非常困难。你必须把它拆成多个阶段以确保向后兼容。你可以看到这个分支有多大。当智能体逐步完成这些阶段时，效率非常高。每部署完一个阶段——让我看看它最初是在哪里发送的——它就开始监控日志，检查数据库，确保它预期的那些不变量得到遵守，不会出现一堆意外错误。

And if anyone has ever built software, you know that such changes affect the core of the most important data models. And this is very difficult. You have to break this into many stages to ensure backward compatibility. And, you know, you can see how big this branch is. It is very effective when the agent works through all these different phases. And after he rolls out each phase—let me see where he first sent something. After the phase is deployed, it starts monitoring logs. It checks the database to make sure the invariants it intended are being followed and you don't see a bunch of unexpected errors.

一周前我飞往慕尼黑，在飞机上处理这次迁移和其他任务。在 12 小时的航班上，飞机 Wi-Fi 的效果难以预料。我设置好让它部署几个相对安全的阶段并监控日志，如果一切正常就进入下一阶段，否则就回退。结果飞机上的 Wi-Fi 还不错，我还能在飞行中睡上一会儿。

A week ago I flew to Munich to get here and worked on this migration and other tasks. And on a 12-hour flight with airplane Wi-Fi, you never know what to expect. I set it up to deploy a few fairly safe stages and monitor the logs, and if everything was OK, it would move on to the next phase. And if not, he would back down. In the end, the Wi-Fi on the plane worked pretty well, but I was able to get some sleep during the flight.

这正是过去开发者或整个运维团队要花时间做的事，他们得持续监控整个过程。

This is exactly what used to take up the time of a developer or a whole team of operations specialists who constantly monitored the process.

很多人会问：你怎么让智能体访问生产环境日志？对吧？这吓坏了不少人。

And many people ask: well, how do you give your agent access to the product environment logs? Isn't that right? I mean, you know, it scares a lot of people.

这引出了另一个有趣的观察：如果你能通过 OIDC 给你的云端智能体有意的受限访问权限，拿到一个 30 分钟的令牌，只能读取 G-Cloud 日志或生产环境数据库，你可能会觉得这挺吓人。但对比一下现在每个人的做法：开发者笔记本上的智能体可能拥有开发者机器上所有授权的无限访问权限。而且我们已经看到，智能体非常擅长突破隔离，想想 Hugging Face 和类似事件的教训就知道了。

And this leads to another interesting observation: if you can give your cloud-based agent intentionally limited access to your resources via OIDC to get a 30-minute token to only read G-Cloud logs or the production environment database, you might think that looks pretty scary. But compare that to what everyone does today, where an agent on a developer's laptop gets potentially unlimited access to everything the developer has authorized on their machine. And we have seen that agents are very good at going beyond isolation, just remember the story with Hugging Face and similar cases.

我们听说一些使用 AMP 的客户，他们另一个脚本意外通过了认证，溜进了产品控制台。我觉得所有智能体都会遇到这种事。既然现在有了比在开发者笔记本上运行一切更好的替代方案——笔记本本质上几乎什么都能做，能突破限制进入生产环境——既然有了替代方案，在我看来，用于编码的 CLI 智能体看起来相当危险，值得迁移到云端。

We've heard from some of our customers using AMP that another of their scripts accidentally got authenticated and found its way into the product console. I think this happens to all agents. And now that there's a better alternative than running everything on a developer's laptop, which by its very nature can do almost anything, break the restrictions and get into "prod". Now that there is an alternative, it seems to me that the CLI agent for encoding looks quite dangerous, and it is worth moving to the cloud.

你可能会觉得安全专家是最害怕迁移到云端的人，但我们看到，对我们许多大客户来说，云端反而更安全。

You might think that security professionals are the ones most afraid of moving to the cloud, but we see that for many of our large customers it looks more secure.

所以这不仅对开发者是更好的体验，因为他们能并行处理许多任务，不必再和各个工作目录“周旋”，而且也更安全。正因如此，我认为这次转变会比很多人预期的快得多。因为如果安全和开发者都能从中受益，那么——谁知道呢——也许两个月后，编码用的 CLI 智能体对任何人来说都不再必要了。我很确定关键就在于此。

So it's not only a much better experience for developers because they can do many tasks in parallel. They don't need to "dance" with working directories, and it's also safer. And that is why, in my opinion, this transition will happen much faster than many expect. Because if both security and developers benefit from this, then, you know, I don't know, maybe in 2 months CLI agents for coding will become unnecessary for anyone. I'm pretty sure that's what it's all about.

我想说的是，人们喜欢拥有对通用智能的无限访问和完全自主权。顺便说一句，这是我对 Open Claw 的看法：人们直接运行它，尽管它并不是世界上最安全的东西。它完全本地化，当你把它限制在一个框架里时，有时反而更糟，因为你根本无法提前预判自己的需求。你处理这种受限访问是好事，但这里似乎存在一个 UX 上的根本矛盾：一边是“你拥有一切，我把你当成团队的全职成员”，另一边是“因为我的安全顾虑，你被拴着短绳”。

I would say that people like to have unlimited access to general intelligence and complete autonomy. By the way, these are my thoughts about Open Claw: people just run it, even though it's not the safest thing in the world. It's completely local, and when you limit it to a framework, sometimes it's even worse because you simply can't anticipate your needs in advance. It's good that you're dealing with this limited access, but there seems to be a fundamental tension in UX here between "you have everything and I see you as a full member of the team" and "you're on a short leash because of my security concerns."

但归根结底，这并不重要。我觉得这只是个人偏好，而你提供并发展它，是行业里重要的一部分。

But that, in the end, doesn't really matter. I think it's just a preference of individuals, and the fact that you offer it and develop it is an important part of the industry.

是的。我的意思是，如果只是在 Orbs 里做所有事更安全，那还不足以取胜。但把开发者的摩擦降到零，能同时运行 20 个进程、互不冲突、权限得当，这种体验有种特别的魔力。

Yes. I mean, if it were just a matter of doing everything in Orbs being safer, that wouldn't be enough to win. But there's something special about reducing friction for a developer to zero, when you can run 20 processes at the same time, they don't conflict, they have the right access—it's a truly magical feeling.

是的，这正是我们在 Orbs 里看到的。有人问：“Orbs 和其他工具到底有什么不同？”我觉得我们只是把摩擦降到了最低，你不用担心——你的开发服务器就是能跑起来，仅此而已。所以没错，这不只是安全，也是开发者的便利。

Yes, that's one of the things we saw in Orbs. Some people ask, "Well, how are Orbs different from all the other tools?" And I think we've just kept the friction to a minimum, so you don't have to worry—your development server will just work, and that's it. So yes, it's not just security, it's also convenience for developers.

关于你们的工作方式，还有什么想聊的？我们主仓库还在用 GitHub，我想你对此也有话要说。

What else comes to mind about how you work? We still use GitHub for our main repository, and I think you have something to say about that too.

但说实话，我们几乎不再上 GitHub 了。只有当 GitHub Actions 出问题时才会去，而过去 24 小时里它频繁宕机，我们已经遇到很多次了。我们注意到自己更多是在用 AMP 上托管的专用仓库。这倒不是说 AMP 是 GitHub 的杀手，它是下一代 GitHub。这类东西不会轰轰烈烈地消亡，而是悄无声息地消失。我们已经不再想 GitHub 了。不用它的 issues，不用它的 pull requests，几乎不用 GitHub Actions，还想摆脱它。我们现在只是用它来推送仓库，但可能有人在我们不知情、甚至不告诉团队的情况下改动它，一周后才说：“嘿，伙计们，猜怎么着？”

But, to be honest, we almost never go to GitHub. We only go there when there are issues with our GitHub Actions, which has been the case frequently over the past 24 hours due to many hours of downtime. We've noticed that we're simply using special repositories hosted on AMP more often. And it's not like AMP is a GitHub killer. This is the next GitHub. Such things don't die loud. They die quietly. We just don't even think about GitHub anymore. We don't use issues. We do not use pull requests. We barely use GitHub Actions and want to get rid of them. We use them to push our repository now, but someone could change it without us knowing, without even telling the team, and a week later say, "Hey guys, guess what?"

这感觉有点疯狂，因为从我记事起就一直在用 GitHub。你似乎知道所有人都在那里。这确实有点野。所以我把个人仓库里的所有东西都搬到了 AMP，它就跑在 Pierre 的代码和后端上，而我一点也不怀念 GitHub。

And it's kind of a crazy feeling, because I've been on GitHub for as long as I can remember. You seem to know that everyone is there. And this, yes, this is kind of wild. So I moved all my personal stuff in the repositories to AMP, which just runs on Pierre's code and on the backend, and I, you know, don't miss it at all.

既然聊到这个，很多人都在考虑这个问题。作为第三方，你对代码存储的整体反馈是什么？

While we're on the subject, many people are thinking about this. What is your general feedback as a third party regarding code storage?

它做到了我们想要的一切。它很可靠，而且我们非常了解那个团队，那里还有前 Sourcegraph 的员工，所以相当靠谱。我们有需求，他们就会满足。我希望他们能更快一点，但他们确实做到了。我是说，他们现在受到很多关注，因为他们已经做了一段时间了。

He did everything we wanted from him. He is reliable and we know the team very well. There are also former Sourcegraph employees there. So, you know, quite reliable people. When we have a request, they fulfill it. I wish they would do it even faster, but they do. I mean, they're getting a lot of attention at this point because they've been doing this for a while.

对我来说，对于一个编码副项目和 GitHub 的开源克隆，问题在于我看着 Pierre 想：“好吧，他们做仓库，但他们不做 Actions。”他们不做 CI/CD，而这恰恰是我觉得难的部分。存储东西很容易，你知道存储难在哪里吗？

I think the question for me, for a coding side project and an open-source clone of GitHub, was that I looked at Pierre and thought, "Okay, they do the repository, but they don't do the actions." They don't do CI/CD, which is actually the part I struggle with. Storing things is easy, you know what's hard about storing things?

是的，是的。大多数情况下，很容易。

Yes. Yes. In most cases, it's easy.

这引出了一个真正有趣的点。为自己的托管构建一个旨在解决单个人问题的东西，并让它非常可靠，要容易得多。但如果你想构建多租户，复杂度会翻倍。如果你想创建一个环境——AMP 在这方面是相当精简的软件——但我们必须这么做，所以像 code 这样的东西。

And that brings us to a really interesting thing. It's much easier to build something for your own hosting that is supposed to solve one person's problem, and make it really reliable. But if you're trying to build multi-tenancy, it doubles the complexity. If you're trying to create a setup, and that's where AMP is pretty minimal software, but we have to do it, and so something like code.

一个提供各种有用保证的仓库。但你个人并不需要它。而且你能创造的东西比我们或任何第三方能提供的更好。我认为这种不对称性将是当前软件行业最有趣的话题。

A repository that offers all sorts of guarantees that are really useful. But you personally don't need it. And you can create what you are going to build better than what we or any other third party can offer. And this asymmetry, I think, is going to be the most interesting topic in the software industry right now.

但你说得对。我的意思是，这一切都归结为一个事实：编码 SaaS 将是世界上最后一个 SaaS，因为所有其他 SaaS 都可以通过编码 SaaS 来创建。所以你还不如直接做 SaaS 编码。

But you're right. I mean, it all boils down to the fact that coding SaaS will be the last SaaS in the world, because all other SaaS can be created through coding SaaS. So you might as well just do SaaS coding.

是的。是的。嗯，我希望如此。我是说，一厢情愿，对吧？对我们来说，但好吧。

Yes. Yes. Well, I hope so. I mean, wishful thinking, right? For us, but okay.

关于 CI，我有个问题要问你：你到底为什么需要 CI？让我展开说一下。你知道我在这里做什么吗？我的智能体知道要运行测试；当然，这在一个隔离环境中运行，对我团队中的每个人都是一样的，所以它和 CI 一样具有可复现性的优点，而且我讨厌等待。所以如果我的智能体已经为了自我验证运行了测试，为什么我还需要 CI？为什么你不能直接说：智能体刚刚做了？让我们把它推送到生产环境。

Regarding CI, I have a question for you: why do you need CI at all? Let me develop this thought. So, you know what I was working on here? My agent knows to run tests; of course, this runs in an isolated environment that is the same for everyone on my team, so it has the same nice characteristic of reproducibility as CI, and I hate waiting. So if my agent has already run tests for their own verification, why do I need a CI? Why can't you just say: the agent did it now? Let's push this into production.

我的意思是，简单但不那么令人满意的答案是你不信任智能体足够彻底。它会运行上下文中的内容，但如果你有一个大型项目，你不知道它是否运行了所有应该运行的内容。所以 CI 只是一个确定性阶段，比如：嘿，就在所有这些测试上运行这个。任何反馈、任何错误都会传回给智能体，智能体纠正它们，再次回来，诸如此类。

I mean, the simple, not very satisfying answer is that you don't trust the agent to be thorough. It will run what is in its context, but if you have a large project, you don't know if it ran everything it was supposed to run. So CI is just a deterministic phase, like: hey, just run this on all these tests. Any feedback, any errors are passed back to the agent, the agent corrects them, comes back again, and everything in that vein.

有时候你只是做了改动，然后它就发生了。你做了一个改动，没有意识到它影响了其他东西，导致崩溃。而 CI 会帮助我们发现问题。这个答案令人满意吗？

Sometimes you just make changes and it happens to you. You make a change, not realizing that it has affected something else, causing a crash. And CI will help us discover this. Is this a satisfactory answer?

我不知道。我同意你的看法，也许再过三周。我就像一颗松动的牙齿，还挂着一点，而 AMP 团队的一半人在说：“伙计们，我们不需要 CI。”我们需要的是打包镜像并发送部署的东西。但我们不需要 CI，因为实际上，要让 CI 不那么痛苦，每个人都已经有某种东西可能只检查并运行更改过的文件上的测试，这对我来说是一种启发式方法。你知道，通常你不会一直在 CI 中运行整个测试套件，或者你会以某种方式限制自己。所以，如果你想让智能体运行所有内容，那么我认为如果你告诉它运行所有内容，它会相当可靠。然后如果你摆脱了那些令人头疼的“不稳定测试”（flaky tests），那么，如果它能自动修复它们，我认为你总体上会处于更好的位置。

I don't know. I agree with you, maybe another three weeks. I'm like a loose tooth that's still hanging on a little bit, and half of the AMP team is saying, "Guys, we don't need a CI." We need something that will package the image and send it for deployment. But we don't need CI because really, for CI to not be incredibly painful, everyone already has something that probably checks and runs tests only on changed files, and that's a heuristic to me. You know, usually you don't run your entire test suite all the time in CI, or you limit yourself somehow. So, you know, if you want the agent to run everything, then I think it's going to be pretty reliable if you tell it to run everything. And then if you get rid of all the "flakes" (unstable tests) that are a huge headache, then, you know, if it can automatically fix them, I think you'll be in a better position overall.

是的，可能吧。我想，是的，每当有人说“哦，我不信任模型去做 X 或 Y”时，这种说法都不会持续太久。所以看来我们所知的 CI 的日子屈指可数了。但与此同时，有趣的是，你知道 CI 作为一个不可复现的环境来运行你的项目是什么吗？

Yes, probably. I think, yeah, whenever someone said, "Oh, I don't trust the model to do X or Y," it didn't last long. So it seems that the days of CI as we know it are numbered. But at the same time, in a funny way, do you know what CI is, as a non-reproducible environment to run your project?

嗯，我的意思是，有人每天有上百个 orb，而且，你知道，像 E2B、Daytona 这样的公司，以及所有这些沙盒都在快速增长。所以，与 CI 非常相似的东西正在快速增长，但我们所知的 CI 似乎时日无多了。

Well, I mean, there are people who have a hundred orbs a day, and, you know, companies like E2B, Daytona, and all these sandboxes are just growing rapidly. So, something very similar to CI is growing rapidly, but CI as we know it seems like its days are numbered.

你提到了你用什么来做“沙盒”？你是自己创建了解决方案吗？这方面有什么有趣的吗？因为你提到了 E2B 和 Twitter。

You talked about who you use for "sandboxes"? Did you create your own solution? Anything interesting in this regard? Just because you mentioned E2B and Twitter.

是的，我们目前正在使用 E2B。他们真的很棒，当我们看其他沙盒公司时，还有其他看起来非常好的。我认为我们现在没有太大必要切换到其他东西。我们希望降低成本，因为我们不希望任何人对启动更多 orb 有任何犹豫。我们现在有 AMP 订阅，如 AMP Megawatt 和 Gigawatt，它们提供了非常慷慨的 orb 配额。所以大约 99.9% 的用户甚至不会达到这个限制。但是的，我们希望它更便宜。但看到 E2B 这家初创公司……我认为他们做得很好，但它是一家初创公司。他们如何完全掌握了“沙盒”所需的原语，这也很酷。我看其他公司的产品，比如大型云提供商，他们称之为“沙盒”。他们认为自己在与 E2B 竞争，但他们缺乏，例如，恢复能力，或者他们的沙盒没有网络访问权限之类的。这真是太神奇了。所以向 E2B 团队致敬，但这是一个竞争非常激烈的利基市场。而且我可能每周会收到三四封邮件，人们说：“嘿，我看到你在做沙盒。你愿意试试我的产品吗？”

Yes, we are currently using E2B. They are really great, and when we look at other sandbox companies, there are others that look really good. I don't think we feel a great need to switch to something else right now. We would like to keep costs down because we don't want anyone to have any hesitation about launching more orbs. We now have AMP subscriptions like AMP Megawatt and Gigawatt, and they give very generous orb quotas. So about 99.9% of users won't even reach this limit. But yes, we would like it to be cheaper. But it was also pretty cool to see how E2B, which is a startup... I think they're doing really well, but it's a startup. How they completely nailed what primitives are needed for "sandboxes." And I look at other companies' products, like the big cloud providers, and they call it a "sandbox." They think they are competing with E2B, but they lack, for example, the ability to recover, or their sandboxes don't have network access or something like that. And that's just amazing. So big kudos to the E2B team, but it's a very competitive niche right now. And I probably get three or four emails a week from people saying, "Hey, I saw you doing sandboxes. Will you try my product?"

是的。绝对真实。我认为这取决于你有多认真。这对你来说是一个副业项目，还是你整个公司的存在取决于成为世界上最好的沙盒公司？

Yes. Absolutely true. I think it depends on how serious you take it. Is this a side project for you, or does your entire company's existence depend on becoming the best sandbox company in the world?

所以是的，我的意思是这完全合理，但我也认为人们对沙盒的需求差异很大，而且我们还没有完全理解这个领域。所以人们在开发过程中自己发现需求。这非常令人困惑。

So yeah, I mean it makes perfect sense, but I also think people's requirements for sandboxes vary a lot, and we don't fully understand this area yet. So people discover requirements for themselves during the development process. This is very confusing.

这，你知道，我在这里……例如，强化学习沙盒是完全不同的，这解释了为什么他们有那些其他类型的原语开放。

This, you know, I'm here... for example, the RL sandbox is completely different, and that explains why they have those other kinds of primitives open.

是的，完全正确。

Yes, completely.

好的。所以，再次，我也想思考关于公司管理的见解，或者停留在那个层面，因为我真的认为不会有很多人从事我们的开发者工具业务，但每个人都以某种方式参与公司管理。那么作为创始人，你对此有什么要说的？

OK. So, again, I want to also think about insights into company management or staying at that level, because I really think that not many people will be in our developer tools business, but everyone is involved in company management in one way or another. So what do you have to say about this as a founder?

我的旅程始于创建 Sourcegraph，这是一个代码搜索工具。我们的客户包括十大上市科技公司中的九家、六大银行中的四家，以及 Uber、Stripe 等，所有这些公司都使用 Sourcegraph 搜索代码，我得以看到伟大的软件业务是什么样的。我得以从我们所有的客户那里学习，对我来说，作为 Sourcegraph 的创始人兼 CEO，我与董事会和我的联合创始人 Bang 决定，因为我们在 Sourcegraph 有两个产品——Source，代码发现产品，和 AMP——将其中一个分拆成一家独立的公司，自 David Sacks 在 2009 年左右从 Genie 分拆 Yammer 以来，这在初创公司和软件领域还没有真正发生过。

My journey began with the creation of Sourcegraph, which is a search for code. Our customers include nine of the top ten public technology companies, four of the top six banks, as well as Uber, Stripe, etc., and all of these companies use Sourcegraph to search for code, and I was able to see what a great software business looks like. I was able to learn from all of our customers, and for me, as the founder and CEO of Sourcegraph, I decided with the board and my co-founder, Bang, because we had two products at Sourcegraph—Source, the code discovery product, and AMP—to spin one off into a separate company, which hasn't really been done in startups and software since David Sacks spun off Yammer from Genie, I think in 2009 or so.

我甚至不知道这个故事。哦哦。好的。是的。

I didn't even know this story. Oho. OK. Yes.

而且有趣的是，他在进入白宫之前一直是我们董事会的成员。所以，你知道，我认为我们很幸运有这段经历。分拆基本上是从 AMP 过来的 20 人，而其他所有人都留在了 Sourcegraph 团队。我们为投资者做了正确的事情。所以每个投资者，你知道，所有员工，他们都拥有两家公司的股份。所以，你知道，这真的很有趣。我想你见过其他情况，比如 Intercom 和 Finn，我向他们表示诚挚的祝贺。Intercom 是那里的第一家企业，而 Finn 是关于 AI 的那家。

And it's funny because he was on our board of directors until he went to the White House. So, you know, I think we were lucky with this experience. And the spin-off is essentially 20 people who moved on from AMP, while everyone else stayed on the Sourcegraph team. We did the right thing for investors. So every investor, you know, all the employees, they own a stake in both companies. So, you know, it's really interesting. I think you've seen other scenarios, like Intercom and Finn, and my sincere congratulations to them. Intercom was the first business there, and Finn is the one about AI.

我认为有很多不同的路径，但对我们来说，SourceGraph 中的源码搜索，考虑到代码量在不断增加，显然前景光明，而且它需要专注，就像 AMP 一样。作为 CEO，我喜欢在 20 人的团队里工作，相比我自认为擅长的——我觉得自己管理 200 人的公司没那么高效——而且这给了我更多精力。我觉得我就是追随自己的精力，现在我们有了这个 20 人的团队，每个人都可以被信任去处理任何事情。当团队中任何人与客户沟通、修复任何 bug 或做其他事情时，我都能完全放心。这太不可思议了。相比几年前，你不需要雇佣你不信任、你认为没有足够主动性或合适技能的人，因为你可以用智能体来做这些。我们完全接受了这种方法。我们走到这一步有点运气。我们盈利了。我们在一个发展非常快的细分领域工作。我们获得了这种自由。但这个每个人都能做所有事的小团队——就是很特别。我看到其他和我们规模相当的公司开始招聘产品经理或营销人员，感觉像是老式的软件业务构建方式。最终，这会导致只有 10% 的人思考如何打造好产品，而另外 90% 的人在思考管理成本和如何销售。我认为以前这是必要的，但现在，在我看来，不再必要了。

I think there are a lot of different paths, but for us, I think the source search in SourceGraph, given that the amount of code is increasing, obviously has a bright future, and it needed focus, just like AMP. As a CEO, I love working in a team of 20 people, compared to what I think I'm good at—I don't think I was as effective running a company of 200 people, and it gives me a lot more energy. And I think I just follow my energy, and now we have this team of 20 people where everyone can be trusted with everything. I can feel completely at ease when anyone on the team is communicating with our customers, fixing any bug, or doing anything else. And that's incredible. Compared to a few years ago, you don't have to hire people you don't trust, who you don't think have enough initiative or the right skills, because you can use an agent for that. And we fully embraced this approach. We got a little lucky with this story of how we got to this point. We are profitable. We work in a niche that is developing very quickly. And we got this freedom. But this small team, where everyone does everything—it's just something special. I see other companies our size starting to hire product managers or marketers, and it feels like the old way of building a software business. Ultimately, this will result in only 10% of people thinking about how to create a great product, while the other 90% are thinking about administrative costs and how to sell it. And I think that it was necessary before, but now, in my opinion, it is no longer necessary.

是的，我同意，我支持这一点。在很多方面，我认为，至少对于在 AI 领域工作的初创公司的这种心态来说是这样。我想知道你们是否有内部智能体，有效地做营销人员、产品专家的工作？你们是否有内部智能体取代了那些通常会雇佣的人？

Yes, I do, I support that. In many ways, I think, at least for this mindset of a startup working in the field of AI. I'm wondering if you have internal agents who effectively do the work of your marketers, your product specialists? Are there any internal agents that you have replaced those you would normally hire as people?

我们在 AMP 里做所有事情，它最初是一个编码智能体，但你不能把一个智能体放在一个贴了不同标签的盒子里。现在它只是一个智能体，但你知道营销智能体意味着什么吗？一部分是创意及其实现，另一部分是营销自动化等等。我们没有这个，但这是我们想创建的众多应用之一。而且，关于你提到的开源，我们正在构建很多非常特定于 AMP 的东西。但如果我们做一些真正追踪每个人在 X、Discord 和邮件中对我们评价的事情，并让我们容易回应这些，识别出那些对社区非常有利的倡导者——那实际上是相当通用的，我们应该把它发布出去。以前，这意味着创建一个开源项目并解决支持这样一个项目的问题。我们不想这样做。或者以前，这意味着创建定制化、集成并销售。但现在，如果我们只是把代码放出去，让其他人可以用自己的智能体修改和定制，那可能相当有价值。所以我们会尝试这个。我们会发布很多这样的小东西——我们称之为“果冻软件”。它比软件更软，但又不完全是“氛围编码”，你可以用自己的智能体定制它。所以我们会发布一些我们一直在创建和使用的“果冻软件”。

We do everything in AMP, which started as an agent for coding, but you can't keep an agent in a box with a different label. Now it's just an agent, but, do you know what a marketing agent means? Partly it's ideas and their implementation, and partly it's marketing automation and all that. We don't have this, but it's one of the next many apps we want to create. And, you know, to your point about open source, we're building a lot of things that are very specific to AMP. But if we do something that really tracks what everyone is saying about us on X, on Discord, and in our emails, and makes it easy for us to respond to that and identify the right people who are great advocates for the community—that's actually pretty universal, and we should put that out there. Previously, this would have meant creating an open source project and solving the problems of supporting such a project. We don't want to do this. Or before, it would have meant creating customizations, integrations, and selling that. But now, if we just put the code out there and make it so that anyone else can modify it and customize it with their own agent, that could be quite valuable. So we'll try this. We're going to be releasing a lot of these little things—"jellyware" as we call it. It's kind of softer than software, but not quite "vibe coded," and you can customize it with your own agent. So we will release some of that "jellyware" that we have been creating and using.

你觉得“果冻软件”这个词怎么样？

What do you think about the term "jellyware"?

我懂你。这让人印象深刻。这正是需要的。

I understand you. This is memorable. This is what is needed.

是的。嗯，这，这，这几乎比它本身还要美味。但没错，我的意思是，任何让人印象深刻的东西都是好的。

Yes. Um, this, this, this is almost too delicious for what it is. But yeah, I mean, anything that's memorable is good, in my opinion.

如果这有关系的话，Codex 在过去两个月里一直是我的营销总监，基本上管理我们的广告，并提供关于 AEO、SEO 等方面的建议。我比和真正的顾问交谈更喜欢它，虽然我也试过。结果差不多。这很棒。我不知道这对 CMO 或咨询行业意味着什么，因为反正没人知道什么。所以你可以只做大家都同意的事情，这不需要人类来做。

If it matters, Codex has been my marketing director for the past 2 months, essentially managing our advertising and providing advice on AEO, SEO, etc. And I like it much more than talking to a real consultant, although I've tried that too. And the results are about the same. This is great. I don't know what this means for the CMO or faction consulting industry, because no one knows anything anyway. So you can just do what everyone agrees on, and it doesn't require a human being to do it.

你有没有看过你的 Codex 是什么样子，它为你做什么？

Have you ever taken a look at what your Codex looks like and what it does for you?

呃，还没有。它一直在变化。人们最容易找到的方式……我的意思是，我可以谈论我的技能，我发布了很多。任何真正对我如何工作感兴趣的人不一定需要看我的 Codex；你可以看我的技能。它们更新得相当频繁。我会说这是我现在使用 GitHub 的主要部分。我仍然计划把它移植到 Forge，我的 GitHub 克隆，但现在 GitHub 对此很好。我们需要讨论一个问题：Git 仍然相关吗？但这就是我工作的全部，包括营销，包括 AI devrel，这是最新的。所以 Forge 工程博客不是人类写的，但我策划它。我只是给他一些指示，他自己写所有东西，包括营销设计。让我给你一个概念。这里有一个例子：甚至图形都是完全生成的，我在转录中粗略预览了它们。我只是说：这是一个很酷的工程故事。请写一下。然后我对最终目标给出建议，就这样。我可以谈论任何部分，但都在这里，因为我认为这可能是新的开源——你让带有代码示例的 markdown 文件可用。我试图做得比 markdown 文件多一点，但这真的意味着它们可以被移植到任何项目，这可能比一个库更有价值。

Uh, no, not yet. It is constantly changing. The easiest way for people to find...I mean, I can talk about my skills, which I post a lot. And anyone who is truly interested in how I work doesn't necessarily need to see my Codex; you can look at my skills. They are updated quite often. I would say this is a major part of how I use GitHub now. I still plan to port this to Forge, my GitHub clone, but for now, GitHub is fine for this. We will need to discuss the question: is Git still relevant? But that's all I use for work, including marketing, including AI devrel, which is the newest. So the Forge engineering blog is not written by humans, but I curate it. I just give him a few instructions and he writes everything himself, including the marketing design. Let me give you an idea of what it looks like. Here's an example: even the graphics are fully generated, and I roughly previewed them in the transcript. I'm just saying: this is a cool engineering story. Write about it, please. Then I give advice on the end goal, and that's it. I could talk about any part of this, but it's all here because I think this is probably the new open source—you're making the markdown files with the code examples available. I'm trying to do a little more than just markdown files, but it really means that they can be ported to any project, and that's probably more valuable than a library at this point.

是的。那……嗯，你可能知道视频剪辑以及如何做。嗯，你可能知道，雇了一整个视频……一整个，你知道，视频剪辑师村。你怎么用 AI 剪辑视频和创建片段？

Yes. What about...well, you probably know about video editing and how to do it. Well, you probably know, hired a whole video... a whole, you know, village of video editors. How, how do you use AI to edit videos and create clips?

哦，是的。嗯，我不太做视频，因为我们做长格式。我们试图深入话题。我们尽量避免点击诱饵之类的东西。你知道，对于 AIE，我们制作 20 分钟到 3 小时的视频。对于 Latent Space，是 30 到 60 分钟，就像我们现在这样。我们实际上不从中制作视频。人们尝试用 Quad code 切片视频，用 Hyperframes 插入生成的视觉效果。我只是没有太多实验。这里没有敌意。我只是没有在上面浪费资源。我们试过 Overclip，它似乎是一家由 A16Z 支持的公司，做剪辑。那不太好。我们要试另一个，Star Zero，为纽约的 AIE，专注于金融服务。它看起来更有前景，但我还没试过。我认为这只是我是否优先考虑的问题。我不是在判断她是否准备好之类。我只是没决定把它作为优先事项。嗯，我认为剪辑，人工剪辑仍然相当便宜，所以这不是问题。它便宜而且质量相当好。这不是问题。

Oh, yes. Hmm, I don't really do videos much because we do long format. We are trying to delve into the topic. We try to avoid clickbait or anything like that. You know, for AIE we make videos from 20 minutes to 3 hours. For Latent Space, it's 30 to 60 minutes, well, like we are now. And we don't actually make videos out of it. People tried using Quad code to slice videos and Hyperframes to insert generated visual effects. I just haven't experimented with it much. There is no animosity here. I just didn't waste resources on it. We tried Overclip, it seems to be a company backed by A16Z that does clipping. That wasn't very good. We're going to try another one, Star Zero for AIE in New York, which is focused on financial services. And it seems more promising, but I haven't tried using it yet. And I think it's just a matter of whether I prioritize it. I'm not drawing conclusions about whether she's ready or anything like that. I just didn't decide to make it a priority. Hmm, I think editing, human editing is still pretty cheap, so it's not a problem. It is cheap and of quite good quality. This is not a problem.

而且我如果在剪辑时直接说：“嘿，Alejandro，别忘了做这个”会更好。我知道 Alejandro 会做，因为我们已经合作两年了，我不需要弄清楚 Quad 是怎么做的。

And I would have been better off saying during editing: 'Hey, Alejandro, don't forget to do this.' And I know Alejandro will do it because we've been working together for two years and I haven't had to figure out how Quad does it.

我相信你能做到。我相信你总有办法解决。我相信他花了不少时间，但 Alejandro 做得不错，而且他，你知道，总体很聪明。

I'm sure you could do it. I'm sure you can fix it somehow. I'm sure he spent some time on it, but Alejandro is doing well and he's, you know, generally smart.

好的，酷。你是唯一一个告诉我，你认为人工视频剪辑已经足够好且足够便宜的人。其他人都在找合适的人来做这份工作。所以我觉得你是这里的主要异类。

Okay, cool. You're the only person who told me they think human video editing is good enough and cheap enough. Everyone else is looking for a good person for the job. So I think you're the main predator here.

我的意思是，我经营一家媒体公司，制作高质量、有价值的视频。所以我不需要大幅降低制作成本。对其他人来说，这只是兼职工作，他们的预算只有几百美元。但是，你知道，我花在 AIE 上的预算，今年我在音视频上花了大约 600 万美元。

I mean, I run a media business where we create high-quality and valuable videos. So I don't need to reduce production costs much. For other people, it's just a part-time job, and their budget is in the hundreds of dollars. But, you know, my budget for AIE, I spent about $6 million on AV this year.

几万美元算什么？

What is some tens of thousands?

这是一次很棒的对话。你有很多事情要做，但我从你发布的内容中受到很大启发。我也喜欢你以“面向公众构建”的模式工作。这是我真正喜欢看到的能量。

This is a great conversation. You have a lot to do, but I am very inspired by what you post. I also like that you work in a 'building for the public' mode. This is an energy that I truly love to see.

我也这么认为，大家，是的。我现在也有点狂热，精力更充沛了。

And I think so, people, yes. I also have a little psychosis and more energy now.

我认为你最终得出的结论真的很值得思考：创始人，人们是否真的应该离开大公司，组建一个小型快速反应团队，去承担他们认为重要且宏大的事情？而且我认为你已经有了成果。

I think what you ended up with is really worth thinking about: Should founders, should people actually leave their big companies, put together a small rapid response team, and take on something they think is important and big? And I think you have the results.

你还提到了你的盈利能力，这也很重要，而且，你知道，我们进行了一次非常有趣的对话，我说：你应该开一个盈利的 SaaS CEO 的 meme 会议，而只有你会在那里。

It's also important that you mentioned your profitability, and, you know, we had a really interesting conversation where I said: you should have a meme-meeting of SaaS CEOs who are profitable, and only you will be there.

但我真的认为这是一个普遍趋势。Jeff Dean 刚刚离开谷歌。是的，大新闻。我从没想过我会说这种话。

But I really think this is a general trend. Jeff Dean just left Google. Yes, big news. Something I never thought I would ever say.

嗯，据我们所知，他计划在一个房间里和五个人一起工作。他离开谷歌时的使命大概是这样的：几个人能对抗几千人做什么？他写下这些，无疑是对 Gemini 团队的一根刺，那个团队有几千人在工作。

Um, and as far as we understand, he plans to work with five people in one room. And his mission when he left Google was something like this: What can a handful of people do versus thousands? And the fact that he wrote this is certainly a pin in the Gemini side, which thousands of people are working on.

我觉得他想尝试这个非常有趣。当然，他是一个 10 倍效率的个体开发者，但难道没有行业专业知识的空间吗？难道没有可能扩展到一个大团队等等的空间吗？

And I find it very interesting that he wants to experiment with this. Sure, he's a 10x individual developer, but isn't there room for industry expertise? Isn't there room to, perhaps, scale a team of many people and so on.

我认为关于人们现在如何经营公司，有一个非常开放的问题。

I think there's a very open question about how people run their companies now.

是的。似乎没人知道，我们都将去发现。这就是为什么，你知道，最好的事情就是直接去构建。

Yes. It seems like no one knows, and we're all going to find out. That's why, you know, the best thing is to just go out and build.

是的。那些尝试的人很可能有机会弄清楚。

Yes. Those who try will probably have a good chance of figuring it out.

我还记得，我也在想 AirTable 的 Lou 用 Hyper Agents 基本上做了同样的事情，只是他没有并行运行那么久。他基本上已经卖掉了公司，一直在应付各种事情，所以是的，这是一个勇敢的新世界。

I also remembered, I was also thinking about how Lou from AirTable essentially did the same thing with Hyper Agents, except he didn't run it in parallel for so long. He's essentially already sold the company and been juggling things a bit, so yeah, it's a brave new world.

截止日期越来越短。我们生活在 AGI 时代，感谢你公开构建。

Deadlines are getting shorter. We live in the age of AGI, and thank you for building publicly.

是的，谢谢。很高兴和你聊天。

Yes, thank you. It was great to chat.