从数学神童到 AI 先锋：Scott Wu 的成长之路

你以前喝过吗？

Have you done this before?

我这辈子其实从来没喝过啤酒。

I have actually never had a beer in my entire life.

那你一上来就喝最好的啤酒，挺好的。

Well, you're starting with the best beer. So that's good.

你能用 Devin 订亚马逊包裹吗？

Can you order your Amazon packages with Devin?

对。你就在 Slack 里让它帮你买东西。比如说“Devin，你能再给我们买些白板吗？”

Yeah. You just go in Slack and ask it to buy something for you. Like, 'Devin, can you go buy some more whiteboards for us?'

说说你的成长经历和那些数学的事吧。我觉得你现在以数学闻名。

Tell me about your upbringing and all the math stuff. I feel like you're known for the math stuff these days.

我在巴吞鲁日长大。我父母都是化学工程师，他们从中国移民过来读研。找工作时做空气排放许可之类的事。路易斯安那州有很多石油和天然气，所以我们就在那儿定居了。

Yeah. I grew up in Baton Rouge. My parents were both chemical engineers. They immigrated from China for grad school. When they looked for jobs, they did air emissions permitting and things like that. Louisiana has a lot of oil and gas, so that's how we ended up there.

空气排放也挺有意思的。

Love the air emissions too.

我从小就喜欢数学。我有个哥哥叫 Neil，我们关系特别好。他比我大五岁，中学时开始参加数学竞赛。我当时一年级，作为弟弟就看他做什么，我也学着做。就这样我开始喜欢数学。我发现我真的很享受数学竞赛和比赛。比如我问你 694 的平方是多少……

Yeah. I always loved math as a kid. I had an older brother named Neil, super close. Neil was about five years older than me. He started doing math competitions in middle school. I was in first grade at the time, and as a little brother, I would watch what he was doing and try to learn the same math. That's how I first got into it. I found I really enjoyed math competitions and competing. This is stuff like if I ask you what's 694 squared...

我觉得可能不是那种问题。

I think it's probably not quite things of that nature.

是 481636。但其实是数学谜题：青蛙每天往上爬又掉下来，要多少天？那种需要批判性思维、想出有趣点子的题。我二年级开始参加数学竞赛。当地大学有个面向中学生的比赛，我作为二年级学生参加了七年级组。那是第一次。他们念了第三名、第二名、第一名，都不是我。我当时特别着迷。

It is 481636. But it's things like math puzzles: a frog goes up and falls down a well every night, how many nights? Things where you do critical thinking and come up with interesting ideas. I started math competitions in second grade. There was a contest at the local college for middle and high schoolers. I competed in the seventh grade math division as a second grader. It was my first time. They called out third, second, first place, and none were me. I was just obsessed.

这就是你的超级反派起源故事。

That's your supervillain origin story.

没错。一切就是这样开始的。第二年我三年级，参加了代数 1 比赛并赢了。之后我一直参加数学竞赛。高中最后一年，也就是高二，我提前一年离校。我参加了编程算法比赛。我参加了三次国际数学奥林匹克，都拿了金牌。

Exactly. That's how it all began. The next year, in third grade, I competed in algebra 1 and won. I kept doing math competitions. My last year of high school, my junior year, I left a year early. I did the programming algorithm. I did the IMO three times and got gold.

你上的什么学校？

Where'd you go to school?

我休了一年学。我提前一年离开高中。我可能不太擅长上学。

I took a year off. I left high school a year early. I wasn't that good at school, I guess.

这挺意外的。你不擅长上学。

That's surprising. You weren't that good at school.

我不太擅长完成学业。我有中学文凭，但没读完高中或大学。我提前一年离开高中，在湾区待了一年，在一家叫 Adapar 的公司做软件工程师。那是 2014 年。

Well, I wasn't that good at finishing school. I have a middle school degree, but I didn't really make it through high school or college. I left high school a year early, spent a year in the Bay working at a company called Adapar as a software engineer. That was back in 2014.

你怎么去了 Adapar？他们招一个高中辍学生，真有远见。

How did you end up at Adapar? That's very forward-thinking of them to take on a high school dropout.

那是个有趣的团队。有意思的是，我们四个高中生同时入职：我、Alexander Wang、Eugene Chen 和 Ser R。Alex 和我在中学时通过一个叫 Math Counts 的数学竞赛认识。我们都参加了全国赛，然后在 Google Hangouts 上开始聊天。

Yeah, it was a fun group. Funnily enough, there were four of us who started at the same time as high schoolers: myself, Alexander Wang, Eugene Chen, and Ser R. Alex and I met in middle school at a math competition called Math Counts. We were both at the national competition. We started talking on Google Hangouts.

Alex 现在在 Scale AI 吧。

Alex now of Scale AI, I guess.

对。看来数学和 AI 有点联系。我们那一批人很多后来都创业了。Alex 是我们当中第一个创办公司的，功劳不小。

Yeah. It turns out there's some math and AI connection. A lot of folks from our vintage ended up being entrepreneurial. Alex deserves credit for being the first of our group to start a company.

但还有很多人——比如 Perplexity 的联合创始人 Johnny Ho，Pika 的创始人 Demiggle，Decagon 的创始人 Jesse Zang——我们很多人同年参加了这些数学和编程竞赛，彼此都认识。

But also, you know, a lot of folks—Johnny Ho, who's one of the co-founders of Perplexity, for example; Demiggle, who started Pika; Jesse Zang, who started Decagon—a lot of us were actually competing in these math and programming competitions in the same year, and we all knew each other.

好的，这让我想到一个问题。之前人们讨论过：年轻创始人去哪儿了？过去有很多二十出头的年轻人创办了突破性公司。迈克尔·戴尔 19 岁创办戴尔，23 岁上市。马克·扎克伯格创办 Facebook 时非常年轻，公司真正爆发时他依然很年轻。有一段时间没有年轻创始人，现在又多了很多——比如你提到的那些人。你 28 岁，经营着 Cognition。年轻创始人的出现是否是行业活力的标志？戴尔年轻是在 PC 时代起飞时，扎克伯格年轻是在社交网络起飞时，而现在我们正处于 AI 编程工具起飞的时代。

Okay, so this gets to something I was wondering. There's this topic people talked about a while back: where are the young founders? There used to be people in their early 20s working on breakout companies. Michael Dell was 19 when he started Dell, 23 when he took it public. Mark Zuckerberg was very young when he started working on Facebook, and when it became a real breakout, he was still very young. There was a period with no young founders, and now there are many more—like the people you mentioned. You're 28, running Cognition. Is the presence of young founders a biomarker for industry vibrancy? Michael Dell was young during the takeoff of the PC era, Zuckerberg was young during the takeoff of social networking, and now we're in the takeoff of AI coding tools.

谢谢你说我年轻。不过相比 18、19 岁，还有很大差距。

Yeah, I appreciate you calling me young. I think relative to being 18 or 19, still a long way.

关键是在二十多岁。

The test is like in your 20s.

我对此有看法。我也一直在想这个问题。我的观点是，总的来说，做创始人变得更难了，这可能是最重要的原因。我认为那些非常聪明、坚定的年轻创始人之所以成功，是因为好的第一性原理思维胜过经验。做创始人很大程度上是做前所未有的事，得出自己的结论。问题是，现在有很多人既有第一性原理思维又有经验。这个领域变得更成熟了，所以更难了。现在直接从大学出来的创始人更少了。

So I have a take on this actually. I've been thinking about this question as well. My take is that overall, being a founder has just gotten harder, and that's probably the highest-order bit. I think the reason young founders who are really sharp and determined did very well is because being a good first-principles thinker beats experience. A lot of being a founder is doing something that has never existed before and coming to your own conclusions. The thing is, now there are a lot of people who have both first-principles thinking and experience. Things have gotten more mature as a space, so it's gotten harder. There are fewer literally coming out of college now.

很难说过去创办领先企业很容易。Facebook 面临很多竞争，戴尔也不是唯一的 PC 制造商。所以我不认为他们轻松。但我觉得你说到了一点：如今所有大公司都非常了解并紧密联系着生态系统。看看萨提亚·纳德拉或马克·扎克伯格，他们对 AI 领域的一切都很关注。所以也许没有巨大的机会被大公司遗漏。

It feels hard to make the claim that it was easy to start a leading business in prior eras. Facebook faced lots of competition; it's not like Dell was the only PC maker. So I don't think they had it easy by any stretch. However, I think you are getting at something: clearly all the large companies these days are very aware and connected with the ecosystem. If you look at Satya Nadella or Mark Zuckerberg, they are very aware of everything going on in AI and paying a lot of attention. So maybe there aren't giant opportunities just being left on the ground by big established companies.

是的，也许“更难”这个词不对。更准确地说，这个领域更成熟了，有更多的套路和现有知识。每个企业当然都有独特之处，但很多细节——比如如何设计股权、如何融资、如何组建初始团队——这些都可以借鉴经验。在过去，这些根本没有成文的东西，所以完全取决于你有多敏锐、多善于自己做决定。现在有更多经验可以借鉴。也许这是部分原因。我还有一个理论，我称之为“万物数据化”。比如，我平时玩的一个爱好是打扑克。

Yeah, and maybe 'harder' is not the right word. It's more that the space is a bit more mature, and there's more of a playbook and existing knowledge. There's obviously something unique with every business, but a lot of the details—here's how you should structure equity, here's how you should figure out fundraising, here's how you should hire your initial team—many of these things carry over with experience. In previous eras, the book wasn't written at all, so it really came down to how sharp you were and how good you were at making your own decisions. Now there's a lot more experience to draw from. Maybe that's part of it. I also have a theory I'd call the 'moneyballification' of everything. For example, one thing I do casually for fun is playing poker.

扑克很有趣，实际上比很多人以为的要数学得多。人们通常认为它……

Poker is a very fun game. It's actually much more mathematical than a lot of people realize. People think of it as...

人们知道扑克求解器和赔率表。还是说它比那更数学？

People know about poker solvers and odds tables. Or is it more mathematical than that?

不，不，我觉得没错。我觉得没错。嗯，第一印象是它全在于知道自己的牌。

No, no, I think that's right. I think that's right. Well, I think there's a first-order impression of it's all about knowing what you've got.

没错。玩的是对手。但它显然比那更数学。有趣的是，你可以在顶尖玩家的演变中看到这一点。在 80 年代或 90 年代，顶级职业选手——我并不是说竞争不那么激烈——但让人成为优秀扑克玩家的技能是极强的直觉。他们理解很多数学概念，但处于系统一的层面，能够思考它们，并且对游戏有很好的感觉，知道如何改进自己的玩法。而现在全是数学书呆子。当领域足够成熟时，对于不那么成熟的领域，当人们不知道要问什么问题或如何思考时，拥有敏锐的直觉并得出自己的结论很重要。然后随着这些事物变得更成熟，结论就变成了数学。我觉得很多领域都是这样，创业公司也开始出现这种情况。

Exactly. Play the person on the other side. And it obviously is much more mathematical than that. But one interesting thing is you see it in the evolution of the top players in the space as well. Back in the '80s or '90s, the top pros—again, I don't think it's less competitive—but the skills that made someone a great poker player were just really great intuition. They understood a lot of the mathematical concepts but at a very system-one level, just being able to think about them, and they had a good feel for the game and a good sense of how to improve their own play. And now it's just all math nerds. At some point, when the space gets mature enough, for a less mature space when people don't know what the right questions to ask are or how to even think about it, having a really sharp intuition and coming to your own conclusions matters. Then at some point, as these things get more mature, the conclusion kind of is math. I feel like that's been the case in a lot of different fields, and it's happening a little bit for startups as well.

我看到更多领域已经归结到其底层——就像国际象棋引擎直接判定局面是“白方胜势 41”之类的。

I see more spaces have resolved to their underlying—like a chess engine just deciding that the position is, you know, 'mine 41' or something.

是的。顺便说，国际象棋完全一样。在 19 世纪，人们……

Yeah. And chess is totally the same way, by the way. Back in the 1800s, people...

浪漫主义风格。

The romantic style of play.

是的，没错。浪漫主义风格。而现在就像有一系列正确的走法，你只是看自己离最优有多近。

Yeah. Exactly. The romantic style of play. And now it's kind of like there is a right sequence of moves, and you just see how close you are to that optimum.

是的。还有哪些领域出现了万物数据化？

Yeah. What are other domains where the moneyballification of everything is?

我的另一个爱好，至少在 Cognition 出现之前，是一款叫《任天堂明星大乱斗》的游戏。

One of my other hobbies, which I played at least before the advent of Cognition, was a game called Super Smash Brothers.

我以前参加过《任天堂明星大乱斗》的比赛。你看到非常相似的模式，特别是有一款叫《Melee》的游戏。它是 2001 年出品的 GameCube 游戏，所以是一款很老的游戏，但人们还是一直在玩。在最初的六到八年里，游戏风格非常狡猾，思维敏锐，反应迅速，能想出各种点子。而现在，一切都变成了数学。那些玩得好的人……

I used to play tournaments for Smash. And you saw very much the same pattern where there's a game called Melee in particular. It's for the GameCube which came out in 2001. So it's a very old game, but people just keep playing the same game. For the first six to eight years, the personality was very much wily, sharp thinkers, people who are quick on their feet and coming up with ideas. And now it's just all math. The people who play and do really well are...

我觉得一些即时战略游戏也有点这样。随着玩家水平提高，他们变得不那么有创意了。

I think some of the RTS's are a little bit that way as well. The players have gotten less creative as people have gotten better at them.

是的。有趣的是，其中也有不少书呆子式的美感。只是被最看重的技能不同了，也许可以这么描述。

Yeah. And it's a funny thing where there's a lot of beauty in the nerd side of it too. It's just a difference in what skills get most selected for, maybe that's the way I'd describe it.

好吧，我跑题了，本来要问你关于认知的问题。什么是认知？它需要什么？

Okay. I'm getting distracted from asking you about cognition. What is cognition? What does it take?

我们正在构建 AI 软件工程师。过去一年半我们一直在打造 Devon，最近又收购了 Windsurf。所以 Devon 是智能体，Windsurf 是 IDE。但从高层来看，我们真的想构建软件工程的未来。

So we're building the AI software engineer. We've been building Devon for the last year and a half, and most recently just acquired Windsurf. So Devon the agent and Windsurf the IDE. But at a high level, we really want to build the future of software engineering.

你们有两个品牌：Cognition 公司，以及 Devon 这个有点拟人化的化身，会不会让人困惑？

Is it confusing for people that you have two brands: Cognition the company, and Devon the slightly anthropomorphized instantiation of it?

我们一直在讨论这个。现在又有了 Windsurf，所以是第三个东西。但我觉得做些整合可能是好的。

We've been talking about it. Now there's Windsurf as well, so there's a third thing. But I think some consolidation is probably good.

人们可能熟悉 GitHub Copilot 或 IDE 风格的模式，你在 IDE 里写代码，它帮你自动补全，或者你可以给一些指令。这不是 Cognition Devon 的模式。相反，使用 Devon 时，你在 Slack 频道里和 Devon 对话，提示它去构建某个东西，就像和同事在 Slack 里交流一样。

People are maybe familiar with the GitHub Copilot or IDE style paradigm where you're writing code in your IDE and it helps you autocomplete or you can give some instructions. That is not the Cognition Devon paradigm. Instead, with Devon, you're in a Slack channel with Devon and you're prompting it to go off and build me an X or a Y, but you're talking to it as you would a coworker in Slack.

没错。你可以从 Slack、Linear、Jira 或者 IDE 里调用它，但并非必须。过去，GitHub Copilot 是 IDE 领域最大的开创者。我把它描述为：当你作为工程师在键盘上打字时，它让你更快，给你工具和快捷键。Devon 则是一个非常不同的范式：一种异步体验，你有一个智能体，然后委派任务。Devon 自然更倾向于在工单级别或项目级别运作。你在 GitHub 上有个问题，标记 Devon，然后 Devon 就开始处理。

That's right. You can call it from Slack, Linear, Jira, or your IDE as well, but you don't have to. In the past, GitHub Copilot was the biggest originator of IDEs. I would describe it as basically when you are typing at the keyboard as an engineer, making you a bit faster and giving you tools and shortcuts. Devon is a very different paradigm: an async experience where you have an agent and you delegate a task. Devon naturally operates more at a ticket level or project level. You have some issue in GitHub and you tag Devon, and then Devon gets to work on it.

目前 Devon 在什么级别的任务上表现良好？

What level of task is Devon doing a good job of today?

我们现在把 Devon 比作初级工程师。有些事情 AI 比我们所有人都强，比如百科全书式的知识和事实检索。有些事情它仍然会做出糟糕的决定。但总体而言，这个平均水平是合适的。我们看到人们通常用它来处理 bug、简单的功能请求和修复等。你和团队讨论一个问题，然后说‘嘿，Devon，去做这个’。另一方面，还有很多重复乏味的任务，比如迁移、现代化改造、重构、版本升级、测试和文档。软件工程师花在修复 Kubernetes 部署之类事情上的时间多得惊人，而不是真正有创意的工作。

We like to call Devon a junior engineer today. There are some things an AI is way better at, like encyclopedic knowledge and pulling facts. There are some things it still makes terrible decisions on. But overall, that's the right average. What we see folks typically using it for are bugs, simple feature requests, fixes, and so on. You're talking about an issue with your team, and you just say, 'Hey, Devon, go do this.' On the other hand, a lot of repetitive tedious tasks like migrations, modernizations, refactors, version upgrades, testing, and documentation. It's crazy how much of software engineers' time is spent on things like fixing Kubernetes deployments rather than building something really creative.

依赖管理。对，所有这些事情。

Dependency management. Yeah. All that kind of stuff.

你能分享一些业务指标吗？

What metrics can you share on where the business is at?

Devon 部署在全球数千家公司。我们与一些最大的银行如高盛和花旗银行合作，一直到只有两三个人的初创公司。我们衡量很多事情的方式是看合并的拉取请求。让 Devon 达到在组织中占合并拉取请求的显著比例。通常，在一个成功的组织中，Devon 合并了大约 30%到 40%的所有拉取请求。

Devon is deployed in thousands of companies all over the world. We work with some of the biggest banks like Goldman and Citibank, all the way down to startups with two or three people. A lot of how we look at things is in terms of merged pull requests. Getting Devon to the point where it is a significant percentage of merged pull requests in an org. Typically, in a successful org, Devon is merging something in the range of 30 to 40% of all pull requests.

你谈到了这种异步模式，但 GitHub Copilot、Cursor 和 Claude Code 不也是不完全同步的吗？你提示它们，它们就去执行。这些区别是暂时的吗？它们会不会消失，变成能即时完成时就同步，不能时就异步？这是一个持久的区别吗？

You talked about this async model, but isn't it the case that GitHub Copilot and Cursor and Claude Code are not fully synchronous? You prompt them and they go off and do something. Are these distinctions a moment-in-time thing? Do they kind of go away where everyone is synchronous when they can do it instantly and asynchronous when they don't? Is this a durable distinction?

好问题。我认为这两种体验在接下来一段时间内会继续存在。而且我实际上认为，找出它们之间的共同体验才是真正有趣的事情。这就是我们最近在 Windsurf 上思考的很多内容。我们很兴奋能在不久的将来推出一些东西。你知道本质复杂性和偶然复杂性的概念吗？

It's a good question. I think the two experiences continue to exist for the next while. And then I actually think that figuring out the shared experience between them is the really interesting thing. That's a lot of what we've been thinking about with Windsurf recently. We're pretty excited to ship some things in the near future. Do you know the concept of essential complexity and accidental complexity?

在我看来，软件工程师本质上就是在代码语境中解决问题的人。他们告诉计算机该做什么，并做出各种决策，从大的架构选择，到微小的决策，比如当余额小于零时，是显示错误还是请求其他操作。所有这些决策就是所谓的本质复杂性，即软件底层逻辑的决策。而偶然复杂性则是其他所有事情，比如为了支持扩展而必须做的所有工作，或者那些每个人都知道需要但无需真正决策的标准功能。在 AI 编程出现之前，软件工程的核心在于做决策，但你却把 80% 到 90% 的时间花在了常规实现上。因此，这种融合的体验就是：对于任何需要你参与的事情，你同步地做出高层战略决策；而对于所有纯粹执行的部分，你可以异步地交给智能体。在单个项目中，往往会有很长一段时间是其中一种模式，然后交替。同步体验是 IDE，你直接查看代码；异步体验是智能体，它去执行任务。你希望工程师在重要选择的高影响时刻与智能体互动，而不是在所有的基础工作上。

What it means to be a software engineer in my mind is basically just somebody who solves problems in the context of code. It is somebody who tells the computer what to do and makes all these decisions, from big decisions like what is the right architecture, to micro decisions like when a balance is less than zero, should we show an error or request something. All these decisions are what people typically call the essential complexity of the underlying logic of what the software is doing. The accidental complexity is basically everything else, like all the things you have to do to support scaling, or standard features that everyone knows you need but require no real decision. Up until AI coding came along, the meat of software engineering has been in making decisions, yet you spend 80 or 90% of your time on routine implementation. So this merged experience is basically where for anything that needs you in the loop, you make high-level strategy decisions synchronously, and for all the parts that are pure execution, you hand them off asynchronously. For an individual project, there are long stretches of one or the other, alternating. The synchronous experience is the IDE where you look at code directly, and the asynchronous experience is the agent that goes off and does things. You want the engineer to be interactive with the agent on high-impact moments of important choices, not on all the groundwork.

你如何让大型企业放心地给 Devon 足够的权限以使其有效？比如迁移用例非常枯燥：你修改表，让它与新表通信，最后删除旧表。最后一步有点吓人。人们仍然担心模型会产生幻觉并做错事。你如何让人们放心地给它足够的权力来发挥作用？

How do you get large enterprises comfortable with giving Devon sufficient permissions to be effective? For example, the migration use case is super boring: you change the table, get it talking to the new table, and eventually delete the old table. That last step is kind of scary. People still have fear of the model hallucinating and doing something wrong. How do you get people comfortable with giving it enough power to be effective?

我们强烈建议使用 Devon 的用户不要给它生产数据库的访问权限。我不知道有任何出问题的实例，但你最好还是不要冒这个险。我的观点是，我们已经有处理这些事情的流程，因为人类也会犯错，所以我们有拉取请求、代码审查、持续集成等机制。Devon 自然能融入这些流程。通常，人们会使用 Devon 进行大型代码迁移，将任务分解，比如 5 万个文件需要从 Angular 的一个版本升级到另一个版本。Devon 会逐个处理并创建拉取请求。你审查代码以确保正确，但仍有人的参与。这又回到了偶然复杂性：迁移耗时的原因不是单一的删除步骤，所有的时间成本都花在其他地方。

We strongly recommend that people using Devon don't give it prod database access, for example. I don't know of any instances where it has been an issue, but you'd rather not take that chance. The framing I would give is that we have processes for these things because humans make mistakes too, which is why we have pull requests, review, CI, and all these things already. Devon naturally slots into these. Typically, folks will work with Devon on a big code migration, breaking up the task, maybe 50,000 files that need to upgrade from one version of Angular to another. Devon will go and do each one and make pull requests. You review the code to make sure things look correct, but there's still that human element. It goes back to incidental complexity: the reason a migration is time-consuming is not the single deletion step; all the time cost comes in other places.

是的，没错。在实践中，尤其是在企业迁移中，内部测量显示，使用 Devon 在许多用例上能获得 8 到 15 倍的提升，因为你只需审查代码，而不用逐行编写或逐个检查引用。

Yeah. Exactly. In practice, with folks especially on enterprise migrations, when they measure internally, they see something like an 8 to 15x gain for many of these use cases with Devon, because you're just reviewing the code, not writing every single line or going through every reference.

我们来谈谈这个。我认为全球所有组织都在试图弄清楚 AI 编程对生产力的影响。工程师们当然希望使用 AI 工具，但在每个开发者的拉取请求数量等指标上并不明显。通常你会看到一些增长，但甚至这个指标本身的好坏也不清楚。如果发布了低质量代码，还会有持续的维护成本。每个人都在寻找确凿的生产力数据。你认为生产力影响有多大？它真的可测量吗？

Let's talk about that. I think all organizations around the world are trying to figure out the productivity impact of AI coding. Engineers want access to AI tools, but it's not totally obvious on metrics like PRs per dev. Generally you see some increase, but it's not clear how good even that metric is. There's also ongoing maintenance cost if you ship low-quality code. Everyone is looking for slam dunk productivity data. What's your view on how big the productivity impact is? Is it actually measurable?

当然。这种向智能体的逐步转变实际上会大有帮助。说实话，IDE 的生产力常常被低估，因为很难量化。你看数据，工程师平均每周使用 Tab 补全 238 次。这显然应该有价值，能让你更快，但快多少很难说。另一方面，使用智能体，很多工作流程就是为你完成任务。如果是 Jira 工单或迁移，你通常很清楚需要多少工程工时。因为它端到端地完成整个任务，所以很明显你不再需要做那个迁移了。

For sure. This gradual shift towards agents will actually help a lot. To be honest, IDE productivity is often underrated because it's hard to state. You look at numbers and it's like engineers on average took tab completion 238 times this week. It seems clear that should be worth something and make you faster, but how much faster is harder to say. On the other hand, with agents, a lot of the workflow is doing the task for you. If it's a Jira ticket or a migration, you typically have a good sense of how many engineering hours are needed. Because it's doing the whole thing end to end, it's much clearer that you just didn't have to do that migration anymore.

你五分钟就审完了那个 PR，就全搞定了。是啊。而且我觉得随着时间的推移，这些事情会越来越清晰。有些人认为编程工具只是昙花一现，会被模型性能的提升碾压，比如 GPT-6 或 GPT-7。你大概不这么看吧？你怎么避免被实验室碾压？

You reviewed the PR in five minutes and that's all done. Yeah. And I think as time goes on, these things will become more and more clear. There is a view that some people have that coding tools are a moment in time thing that get run over by increasing model performance, like GPT-6 or GPT-7. Presumably you do not hold this view. How do you avoid getting run over by the labs?

当然。我觉得实验室都是了不起的企业。据我理解，我把这种观点称为“虚无主义的计算机使用论”，就是说知识工作中的所有事情都涉及使用计算机，而 AI 会越来越擅长使用计算机，直到有一天只剩下 AI 替你干活。我觉得这有道理，很难反驳。但实践中我们看到的是，有很多上下文知识和行业细节。比如做 Angular 迁移——不是说这些事不会变好，它们会的。但让模型变好的方法是给它们正确的数据。如果你从没见过 Angular 或没做过迁移，你能有多擅长？这有个上限。同样，用 DataDog 调试错误也是。最重要的是，现实世界中的软件工程非常混乱。大多数领域都是这样——法律、医学等等。虽然通用智能会越来越聪明，但在让你的特定用例变得非常好，以及交付产品体验、带给客户方面，还有很多工作要做。

Yeah, for sure. So look, I think the labs are incredible businesses. As best as I understand it, I would describe this view as the nihilist computer use take, which is that all these different things we do in knowledge work involve using a computer, and AI will get better and better at using the computer until someday there is nothing left except the AI doing your work for you. I see the wisdom of it; it's hard to disprove. But in practice, what we've seen is there is a lot of contextual knowledge, industry details. For example, doing an Angular migration—it's not to say these things can't get better; they will. But the way we make models better is by giving them the right data. How good can you be at Angular migrations if you've never seen Angular or done one yourself? There's a cap on that. Similarly, using DataDog to debug errors. The biggest thing is that software engineering in the real world is so messy. Most disciplines look like this—law, medicine, and so on. While general intelligence will get smarter, there is still a lot of work to make something really good for your particular use cases, and also in delivering a product experience and bringing it to customers.

所以这不是一个通用智能任务。这是一种特定智能——在 Stripe 代码库工作需要一些通用智能，但也需要大量上下文和工作流程。你认为这仍然是一个需要专业化的领域。

So it's not a general intelligence task. It's a specific intelligence—working in the Stripe codebase requires some general intelligence but also a bunch of context and workflows. You think that persists as an area where you need to specialize.

没错。也许可以这样说：这个论点有点像超级智能。从某种意义上说，是的，我们正在朝那个方向前进。通过强化学习，这个东西在进步。我认为强化学习和这种 AI 范式是解决任何基准的柏拉图式理想。你有一个数据集，包含你想要什么、如何衡量成功，然后训练一个模型达到 100%。我们正以比大多数人预期更快的速度接近那个理想——比如 IMO 金牌或 SWE-bench 的分数。但当那发生时，我不认为我们会得到纯粹的超级智能和人类知识工作的终结。相反，难题变成了：基准是什么？在所有领域定义基准都涉及世界的实际混乱。对于软件工程师来说，这关乎你日常使用的工具、如何使用它们、随时间构建代码库的表示、判断发布功能是否成功，以及围绕它们创建合适的环境。

Yeah, exactly. Maybe one way to put it is: the argument is something like a superintelligence. In some sense, yes, we are moving toward that. With RL, this thing is improving. I think of RL and this paradigm of AI as the platonic ideal of being able to solve any benchmark. You have a dataset of what you want, how to measure success, and you train a model to get 100% on it. We're moving toward that ideal faster than most expected—like the IMO gold medal or scores on SWE-bench. But when that happens, I don't think we end up with pure ASI and the end of human knowledge work. Instead, the hard question becomes: what is the benchmark? Defining the benchmark in all these spaces involves the practical messiness of the world. For a software engineer, it's about the tools you interact with daily, how you use them, building a representation of the codebase over time, deciding whether shipping a feature was successful, and creating the right environments around them.

那么，对于 Devin 想做的那些事，能有一个好的模型性能基准吗？还是说 Devin 的商业模式和收入本身就是基准？

So can there be a good benchmark for a model's performance on the kinds of things that Devin wants to do, or is Devin's business model and revenue essentially the benchmark?

好问题。从我们的角度看，我们内部有很多基准。最大的一个叫“初级开发者”，我们可能很快要升级到“高级开发者”。它基本上是完成各种随机真实世界初级开发任务的能力。我们分享过一些例子。

Yeah, it's a good question. From our perspective, we have a lot of benchmarks internally. The biggest one is called Junior Dev, which we might need to upgrade to Senior Dev soon. It's basically the ability to do a variety of random real-world junior dev tasks. We've shared some examples.

显然我们不会发布整个基准测试，因为那样它就会被淘汰，但很多任务是这样的：你需要去修复一个 Grafana 仪表盘，让它运行起来，然后调出结果。这是软件工程师非常常见的工作。难点不在于某种算法编码本身，而是托管这个的服务器运行了错误版本的某个包，所以你必须通过错误信息找出问题，然后说好的我需要把包降级到另一个正确的依赖版本，然后运行它，调出结果，确保数字看起来正确。诸如此类的事情基本上是我们能做的、最接近真实软件工程师日常工作的内容。

Obviously we don't publish the whole benchmark because then it would get obviated, but a lot of the tasks are things like: you need to go and fix a Grafana dashboard, get it going, pull up the results. This is a very common thing that a software engineer does. The hard part is not some algorithmic coding thing itself, but it turns out the server hosting this is running the wrong version of some package, so you have to go through the errors, figure out what happened, say okay I need to downgrade the package to another one which is the right dependency, then run it, pull this up, and make sure the numbers look correct. Things like that are basically as close as we can make them to what real software engineers spend their time on.

那么新发布的 Claude 4.1 和 GPT-5 做过这个基准测试吗？

And have the newly released Claude 4.1 and GPT-5 done this benchmark?

是的，它们两个在这个基准测试上都比我们本周之前见过的任何模型都要好。

Yeah, I mean both of them are better at this benchmark than any of the models we've seen before this week.

当你思考未来 5 到 10 年的 AI 业务和行业时，你会想到堆栈的所有不同层级。有数据中心，然后是实验室，再然后是像你们这样的应用层。谁受益？什么变得更竞争？什么变得不那么竞争？这些都只是经典的竞争性寡头垄断吗？市场结构是怎样的？

As you think about the AI business and industry over the next 5 to 10 years, you think about all the different layers of the stack. You have the data centers, then you have labs, and then you have the application layers such as yourself. Who benefits? What gets more competitive? What gets less competitive? Are all these just classic competitive oligopolies? What's the market structure?

每次我说这个大家都会取笑我，但我认为所有层级都会做得很好。我认为会有大量的 AI，而且价格到处都很便宜。至少在过去 6 到 12 个月里我一直在这么说，我们看到所有这些的价格都有相当程度的上涨，但总体而言，首先，会有大量的 AI。这一点怎么强调都不为过，因为我觉得我们正在走出一个各种 B2B SaaS 盛行的十年。90 年代和 2000 年代初有互联网，然后是 2000 年代末、2010 年代初的移动电话和云。这些是过去 30 年里最重大的事情。在过去大约 10 年里，大多数构建的东西都是增量式的，为特定细分市场或工作流程的一小部分而构建，使其更高效。现在的 AI 则完全相反：我们谈论的是整个知识工作，甚至可能包括整个体力工作，取决于机器人技术的发展。所以首先，就是会有大量的 AI。

So, everyone always makes fun of me whenever I say this, but I think all the layers are going to do very well. I think there's going to be a lot of AI and I think the prices are cheap everywhere. I've been saying this at least for the last 6 to 12 months, and we've seen prices go up a decent bit across all of these, but at a high level, first of all, there's going to be a lot of AI. It can't be understated in the sense that I think we're kind of coming off of a decade of a lot of various B2B SaaS. There was the internet in the 90s and early 2000s, and then the mobile phone and cloud, which were late 2000s, early 2010s. Those were some of the biggest things in the last 30 years. Over the last 10 years or so, most of what was built was a lot more incremental, building for a particular niche or a small part of the workflow and making that more efficient. AI now is the total opposite: we're talking about the entirety of knowledge work and perhaps the entirety of physical work as well depending on what happens with robotics. So first thing is there's just going to be a lot of AI.

关于价值在哪里积累的第二点。我诚实的回答是，价值在有意义差异化的地方积累。简单来说：如果有 Nvidia 和 TSMC，只要 Nvidia 需要与 TSMC 合作，TSMC 也需要与 Nvidia 合作，他们之间会有一些摩擦，但他们都会继续做得很好。你在堆栈的下游也能看到这一点。我认为所有这些层级中解决的问题都是非常不同的问题，具有相当有意义的差异化。

I think second thing about where does the value accrue. My honest answer is value accrues wherever there's meaningful differentiation. Simple: if there's Nvidia and there's TSMC, for as long as Nvidia needs to work with TSMC and TSMC needs to work with Nvidia, there will be some rubbing up on each other's shoulders, but they will continue to do great. You kind of see this down the stack as well. I would argue that the problems being solved in all these layers are very different problems that have pretty meaningful differentiation.

你是说这基本上防止了过多的垂直整合，各层级各自做自己的事情。

You're saying this prevents too much vertical integration basically, where the layers kind of keep each doing their own thing.

正是如此。我认为有一个真正的区别：一旦你从硬件进入基础模型训练，那就是一个完全不同的难题。公司的 DNA 是寻找异常强大的研究人员，给他们尽可能多的 GPU，并建立围绕这一点的文化。而应用层则真正专注于如何让一个用例工作。例如对我们来说，唯一关心的是构建软件工程的未来。人们经常抽象地在真空中谈论 AI 代码。我认为有很多公司在基础模型层思考代码。我认为我们独特地真正思考软件工程，包括随之而来的所有混乱、产品界面、交付、使用模式，当然还有特定的能力。所以每个人都有自己的 DNA 和自己最擅长的事情。

Exactly. And I think there's a real difference where as soon as you go from hardware to foundation model training, it's its whole own can of worms. The DNA of the companies is finding exceptionally strong researchers, giving them as many GPUs as you can afford, and setting up a culture that orients around that. And then the application layer is really focused on figuring out how to make one use case work. For us, for example, the only thing we care about is building the future of software engineering. People often talk about AI code abstractly in a vacuum. I think there are a lot of companies that think about code in the foundation model layer. I think we uniquely really think about software engineering, with all the messiness that comes with it, the product interface, delivery, usage model, and of course the particular capabilities. So everyone has their own DNA and their own things they do best.

我们在 Stripe 一直在思考为 AI 构建经济基础设施以及需要什么。你可以有一个代表人的智能体，你希望能够通过提示或在你的应用中操作，而你的 AI 可以使用的工具之一就是去现实世界中进行商业活动。所以我们正在为此构建基础设施。然后我们注意到，由于 AI 的经济性，每个人都采用基于使用的模式，按 token 或其他单位计费。所以我们正在构建基于使用的计费基础设施。人们在 Stripe 上构建的计费系统与经典的 SaaS 按席位定价非常不同，而 AI 中的一切都是按消耗的单位计费，你还可以考虑智能体之间如何在没有人类参与的情况下进行商业活动。所以我们的产品路线图受到了所有这些方面的启发，但我很好奇你认为 AI 的经济基础设施需要是什么样的？有什么我们应该记住的吗？

We at Stripe have been thinking a lot about building the economic infrastructure for AI and what is required. You can have an agent acting on behalf of a person and you want to be able to just be prompting or doing stuff in your app and part of the tool use that your AI can engage in is going off and conducting commerce in the real world. So we're building infrastructure for that. Then we noticed that because of the economics of AI, everyone has usage based models, per token, per what have you. So we're building out usage based billing infrastructure. The billing systems people are building on Stripe are very different from classic SaaS per seat pricing, whereas everything in AI is per unit consumed and you can get into how the agents engage in commerce with each other where there's no human in the loop. So there are all these ways in which our product roadmap is being informed, but I'm curious what you think the economic infrastructure for AI needs to look like. Are there things that we should be keeping in mind?

当然，肯定有。

Yeah, for sure.

是的，从按席位到按使用量计费绝对是一个大变化，我认为从两方面看都是。一方面，席位模式没什么意义，因为 AI 本身也可以算是席位，它们也做了很多工作。另一方面，使用量显然与成本自然挂钩，因为这很大程度上就是 GPU 的消耗，取决于你运行模型的频率。所以我认为这非常合理。

Yeah, seat-based to usage-based is a big one for sure, I think on both sides. From the perspective of one, seats don't really make sense when the AI themselves are arguably seats as well, you know, they're doing a lot of the labor too. And then on the other side, I think usage obviously goes so naturally with the costs themselves because a lot of this is effectively GPU spend on how much you're spinning the models basically. So I think that makes a ton of sense.

另一个显而易见的大趋势是形成一个完整的智能体经济。我认为目前这还更多是一个话题而非现实，但情况正在迅速变化，很快你的智能体就能做各种事——有趣的是，我们就在用 Devin。Devin 完全专注于软件工程，但我们用它点 DoorDash 外卖，用它下单亚马逊包裹。有些功能意外地好用。

The other big one which comes to mind obviously is for there to be an entire agent economy as well. I think today it's still probably more of a talking point than reality, but things are pretty rapidly changing and getting to the point where your agents are—funnily enough, we use Devin. Devin is entirely focused towards software engineering, but we order our DoorDash on Devin, we order our Amazon packages with Devin. There are pieces of that that turn out to work nicely anyway.

所以你就在 Slack 里让它帮你买东西？

So you're just in Slack and you ask it to buy something for you?

对，就像跟 Devin 说，你能帮我们买些白板吗？

Yeah, like just at Devin, can you go buy some more whiteboards for us or something like that?

在某个时候，你让 Devin 做的现实世界的事情会不会遇到网站试图阻止机器人活动的障碍？

At a certain point, do the real-world things you ask Devin to do run into blockers with sites trying to block bot activity?

Devin 能很好地工作很大程度上依赖于它能完成这些任务并突破障碍。但有些东西对模型来说很自然，比如你经常有 API 密钥或秘密信息，希望 Devin 能保管。信用卡号也一样。现实世界的软件工程并不需要大量浏览网页、找不同网站并点击。即使只是测试自己的前端或发布文档，良好的浏览器使用也是重要的一部分。这只是……

A lot of Devin working really well relies on Devin being able to do these things and get through. But some of these things are quite natural with the model, which is you often have API keys or secrets that you want Devin to be able to hold on to. So that works for credit card numbers as well. Real-world software engineering doesn't involve a lot of just going and browsing the web and finding different sites and clicking around. Even if you're just testing your own front end or putting up documentation, good browser use is an important piece of that as well. It's just something that's...

那你们不应该开发一个消费者应用吗？难道大家不想要一个魔法棒应用，可以拥有虚拟助手吗？有无数虚拟助手初创公司，但似乎没有一个真正达到规模。

So shouldn't you build a consumer app? Doesn't everyone want this magic wand app where you can just have your virtual assistants? There's a million virtual assistant startups. It seems like none of them have really gotten to any scale.

这是个有趣的问题。从我们的角度看，看到 Devin 去点 DoorDash 很有趣。但同时，我们的团队很小，没有精力在软件工程之外再做这个。你打开 Devin 看到这个，另一边还有 IDE，但 Devin 却在 DoorDash 上。这感觉有点格格不入，我觉得我们保持现状就好。

It's a fun question. From our perspective, it's fun seeing Devin go and do these DoorDash things. At the same time, our team is so small, we just don't have the focus to be able to do that in addition to doing software engineering. You're pulling up Devin and you're seeing this and then on the other side there's the IDE, but Devin's just going on DoorDash. It's a very fish out of water experience, and I think it's fine for us to keep.

但你知道很多产品开发都源于人们注意到产品在这些新兴模式中的使用方式。比如 Twitter，人们开始链接站外照片，于是他们构建了原生图片支持。话题标签是社区发明的。类似地，你检查 Devin 的日志，发现人们在大量购买 DoorDash。也许这暗示了产品方向。

But you know the way a lot of product development follows from people noticing how a product is being used in these emerging patterns. Like Twitter, people started linking to photos off site, so they built native image support. The hashtag was invented by the community. So similarly, you're checking the Devin logs and you notice people are buying a lot of DoorDash. Maybe that's a suggestion on the product side.

是啊，挺有趣的。公平地说，主要也是我们自己。产品使用还在萌芽阶段。

Yeah, it's funny. To be fair, it's mostly just ourselves too. Still emerging product usage.

我同意。这很有趣。

I agree. It's a fun one.

我们有个有趣的例子：Walden 的航班取消了，他试图用 Devin 去和航空公司协商退款。Devin 访问了网站，网站自然把你转接到他们的智能体进行对话。Devin 解释了半天没有进展，然后它说：‘这不行，我现在需要和真人说话。’

We had a fun one where Walden had a flight that got cancelled and was trying to use Devin to go and negotiate with the airline to get the refund for it. Devin went to the site and naturally the site forwards you to their agent to have the conversation. Devin was explaining things and wasn't making progress. Then at some point Devin said, 'This is not working. I need to speak to a human right now.'

然后它做到了吗？

And did it?

做到了。它找到了真人客服。客服上线后，它发送了航空公司合同链接，比如‘哦，第 22 条说了这个那个。’然后 Walden 真的拿到了退款。

It did. It got to the human. The human got on the line and then it sent some link to the airline contract, like 'Oh, section 22 says this and that.' And Walden actually did get the refund.

但抱歉，是 Devin 在说话吗？

But sorry, Devin was speaking?

Devin 在和真人聊天，基本上绕过了机器人智能体，然后找到了真人。

Devin was chatting with the human, basically made it past the robot agent equivalent and then got you a human.

它成功拿到航班退款了吗？

And did it successfully get the flight refund?

它拿到了退款。是的。

It got the refund. Yeah.

再说一次，人们想要这个。回到 AI 的经济基础设施，我们想到的另一件事是，信任在网上会变得更加重要。我不太确定这会以什么形式出现，因为互联网长期以来一直很糟糕，有很多诈骗和黑客攻击。但黑客手段变得更复杂，深度伪造等等。所以，清楚谁是可信的个人、谁是可信的企业，在这个世界里似乎变得更加重要。

Again, people want this. Going back to the economic infrastructure for AI, the other thing we think about is it feels like trust is going to become a bigger deal online. I don't quite know what form that takes, because obviously it's been a big bad internet for a long time, a lot of scams, a lot of hacking. But the hacking attempts become more sophisticated, deepfakes and everything. So having a good sense of who is a trusted individual, who is a trusted business just seems to become much more important in this world.

是的。与此相关，我认为 Cloudflare 与智能体等问题是个热门话题。我们来谈谈 Cloudflare 的问题。现在有更多智能体在浏览网页，已经设置了一些保护措施来阻止智能体访问网站。到目前为止的模式往往是，非人类有很多事情是不允许做的。我认为随着时间的推移，我们可能需要更多地看到授权委托，如果这说得通的话。更明确地表明智能体可以代表你做事。在某种意义上，你也把部分声誉附着在上面。这有如何运作的金钱问题，但智能体采取的行动也可归因于你，代表你。

Yeah. Related to that too, I think Cloudflare with agents and everything is a hot topic. Let's bring the Cloudflare issue. There are a lot more agents browsing the web these days and there have been certain protections set up to not give agents access to websites. The paradigm up until now has often been basically like a 'tons of things which you are not allowed to do' as a non-human. I think what we will probably need to see a lot more of over time is basically delegating access, if that makes sense. Making it more clear that an agent can do something on your behalf. In some sense, you're attaching some of your reputation to it too. There's a monetary question of how this works out, but there's also just that actions the agent takes are attributable to you and on your behalf.

说得好。现在我们有机器人 vs 非机器人，允许爬虫 vs 不允许爬虫。而实际上应该是，如果你授权，机器人就应该被允许。

That's a great point. Right now, we have like bots versus no bots, clankers versus clankers not allowed. Whereas instead it needs to be bots allowed if you sign for them.

就像我刚才说的，简单版本就是如果你登录了谷歌 Chrome 邮箱账户并且有验证地址，那么你可以让一个智能体在那个浏览器窗口中运行并执行操作，但你要对其所做的工作负责。

As I was going to say, the simple version is just like if you're signed into your Google Chrome email account and you have a verified address, then you can have an agent run in that browser window and do things, but you're responsible for the work it does.

是的。这有点像 API 密钥权限，但是在大规模消费者层面，涵盖所有网站和一切。

Yes. Yeah. It's sort of like API key permissions, but at the mass consumer scale across everything and all websites and everything.

我喜欢这个。那么 Devin 的存在如何影响你们自己招聘工程师的方式？

I like that. And how does the existence of Devin affect your own hiring of engineers?

是的，从我们的角度来看，我们一直喜欢保持核心工程团队非常精简和精英化。

Yeah, from our perspective, we've always loved keeping the core engineering team very tight and very elite.

多精简？大概 30 人？

What's tight? Like 30 people?

是的。所以直到几周前，我们整个团队大约 35 人，其中——

Yeah. So up until a few weeks ago, our whole team was about 35 people, of whom—

涵盖所有职位。

Across all roles.

涵盖所有职位。是的。其中，几乎所有人都有工程师背景，说来有趣。但我们所谓的核心工程团队大约 19 人。有了 Windsurf 之后，团队规模显然增长了很多，但实际上核心工程本身并没有扩大太多；从 19 人增加到了大约 30 到 35 人。

Across all roles. Yeah. Of whom, almost everyone is an engineer by background, funnily enough. But what we call core engineering was about 19. With Windsurf, obviously the team count has grown a lot, but actually core engineering itself hasn't gotten all that much bigger; it's gone from 19 to something in the range of like 30 to 35.

所以你们保持工程团队较小。那么这些工程师本身与 20 年前组建的公司相比有什么不同？

So you keep the engineering team smaller. And how are the engineers themselves different versus a company being built 20 years ago?

是的。我们需要做的工作性质非常不同，因为有很多执行和实现工作，但 Devin 会做这些，所以人类不需要。因此我们通常寻找的——例如我们的整个面试流程，对于很多职位来说——基本上就是让人们在 8 小时内构建自己的 Devin，看看他们能做到什么程度。

Yeah. It's a pretty different profile of the work we have to do, in the sense that there is a lot of execution and implementation that has to be done, but Devin does that so humans don't need to. So what we typically look for—our whole interview process, for example, for a lot of these, is basically just having people build their own Devin in 8 hours and seeing how far they get with it.

抱歉，是构建他们自己版本的 Devin，还是用 Devin 构建东西？

Sorry, build their own version of Devin or build stuff with Devin?

构建他们自己的版本，他们自己的智能体，比如他们自己的完整端到端智能体，在 8 小时或 6 小时内。是的，我认为我们发现——而且我认为这个趋势在软件工程中会普遍出现——就是知道所有小细节、记住所有事实、或者非常擅长某种语言的语法之类的东西会变得不那么重要。更重要的是大量的高层决策、真正理解技术概念、对产品有良好的感觉、以及具有良好的直觉知道该构建什么和做什么，并且以那种方式成为自我主导者。所以是的，我们团队中很多人实际上都是前创始人。在我们最初的 35 人中，我想有 21 人之前创办过公司。所以密度非常高。

Build their own version, their own agent, like their own full end-to-end agent in like 8 hours or six hours or whatever. Yeah, I think what we find—and I think this trend we'll see generally in software engineering—is that knowing all the little details, memorizing all the facts, or being really good at syntax of some language or things like that are going to be less important. What's going to be more important are a lot of the high-level decision-making, understanding the technical concepts really well, having a good sense of products, and just having a good intuitive sense of what to build and what to do, and being a self-owner that way too. So yeah, a lot of our team actually are specifically former founders. Of our initial kind of 35, I think 21 of us have founded a company before. So it's been a very high density of that.

哇。好的。很好。

Wow. Okay. Good.

是的。也许我应该早点说出来，然后——

Yeah. Like I maybe should have just come out earlier and then—

我正纳闷呢。是的。我觉得那可能会很有趣。

I was wondering that. Yeah. I think that could have been interesting.

好。好游戏。那是个好游戏。是的。

Good. Nice game. That's a nice game. Yeah.

你什么时候会雇佣最后一名工程师？

When will you hire your last engineer?

这是个好问题。我要在这里做个区分。我认为会有一个时刻——我猜这个时刻大概在 2、3、4 年后——我们将不再把代码作为主要界面。基本上，成为一名软件工程师实际上就是指导你的电脑，告诉你的电脑该做什么，然后说，‘哦，你在看自己的产品，你说，嘿，我们需要在这里新建一个页面。顺便，所有这些数据，我们这样保存，然后根据 X、Y、Z 来索引，因为我们需要做这些查找等等。’你知道，做出很多这样的架构决策，但自己不查看代码，至少在大多数情况下是这样。

It's a good question. I'll make a distinction here. I think there will come a point—and my guess on this point is probably in the neighborhood of let's say 2, 3, 4 years from now—where we stop using code as the main interface. Basically, being a software engineer really is just instructing your computer and telling your computer what to do, and saying, 'Oh, you're looking at your own product and you're saying, hey, we need to make a new page here. By the way, all this data, let's save it this way, and let's index this according to X, Y, and Z, because here are the things that we need to do lookups for or whatever.' You know, making a lot of these architectural decisions, but not looking at the code themselves, at least in the majority of circumstances.

你认为从现在起两到四年后，软件工程师在日常工作中不再真正看代码，就像他们今天不看汇编一样。

You think two to four years from now, software engineers are not really looking at code in their day-to-day, just like they don't look at assembly today.

完全正确。是的。所以那会改变很多事情。我认为到那时，显然工作会发生很大变化。有趣的是，我认为如果有任何变化，我们会有更多的软件工程师，而不是更少。而且我认为仅仅因为界面不再是代码，并不意味着软件工程的核心技能会消失。

Exactly. Yeah. So that's going to change things. I think at that point, obviously, the jobs change a lot. Funnily enough, I think if anything, we will have way more software engineers, not fewer. And I think just because the interface is not code anymore doesn't mean that the core skills of software engineering disappear.

人们经常问我们，‘我的儿子或女儿正在上高中或刚上大学，他们应该学习计算机科学吗？’我的回答总是绝对是的。如果有任何变化，有趣的是，我觉得大学计算机科学总是有相反的问题：教了你太多概念——编程是什么、计算机科学是什么——而没有足够地教‘这是你需要用的语法’和‘这是设置 React 应用的意义’等等。我认为我们会达到这样一个点：那些理论概念和高层次理解——也许用一句话说，计算机的模型以及如何用计算机作为工具做决策和解决问题——那就是编程的本质。

People often ask us, 'My son or daughter is in high school or just starting college, should they even be studying computer science?' And my answer is always absolutely yes. If anything, funny enough, I feel like university computer science always had the opposite sin of doing too much of teaching you the concepts—what programming was about and what computer science was about—and not enough of 'here's the syntax you need to use' and 'here's what it means to get a React app set up' and whatever. I think we'll get to a point where those theoretical concepts and that high-level understanding—maybe in one line, the model of a computer and how to make decisions, problem-solve with the computer as a tool—that is what programming will be.

是的。而且如果有任何变化，会有更多的软件工程师。是的。我认为一件好事是大家都在谈论杰文斯悖论以及它如何与 AI 相关。我认为没有比软件更适用这个悖论的地方了，因为我们似乎永远不会耗尽对更多代码的需求。我们写了很多软件。

Yeah. And if anything, there's going to be a lot more software engineers. Yeah. I think one of the nice things is everyone talks about Jevons paradox and how it relates to AI. I think there's nowhere that it's more true than software, because we really never seem to run out of demand for more code. We write a lot of software.

半开玩笑地说：尽管世界上有这么多软件工程师，我们都知道，还有很多产品仍然很糟糕。你登录银行或处理零售结账等等。所有这些事情仍然非常过时、非常漏洞百出。你登录医疗平台，试图点击找到你的——我们还没有写完所有的软件。UI 完全没有改变，这难道不令人震惊吗？我们仍然和 Siri 对话，它的按钮位置和 iPhone 上的品牌与 Transformer 之前的模型一样。我们通过 Slack 提示 Evan。我们在网页浏览器中使用 AI 工具，把它们输入文本框，就像我们在 1980 年代玩 Zork 一样。

The half-joking way to say this is: despite how many software engineers in the world, we all know this, there are so many products out there that are still so bad. You're logging into your bank or dealing with your checkout in retail or whatever. There are all these things that are still super outdated, super buggy. You're logging into your healthcare platform and trying to click around and find your—we haven't finished writing all the software yet. Isn't it shocking that the UIs haven't changed at all? We still talk to Siri, which is the same button placement and the same brand on the iPhone as pre-Transformer models. We prompt Evan via Slack. We use our AI tools in a web browser and enter them into a text box like we're playing Zork in the 1980s or whenever that came out.

也许是 70 年代。

70s maybe.

我不知道 Zork 有多老。你知道 Zork 是什么吗？我不知道。你太年轻了。它是最早的纯文字冒险游戏。

I don't know how old Zork is. Do you know what Zork is? I don't. You're too young. It was like the original text-based adventure game.

哦，我明白了。

Oh, I see. I see.

是啊。但话说回来，我们什么时候能看到 AI 用户界面？因为现在这很复古。

Yeah. But yeah, when are we going to see AI UIs? Because it's very retro right now.

我对此的总体看法是，新技术浪潮总是这样。手机就是个好例子：最初的 App 看起来就像缩在盒子里的网站。随着时间的推移，你仍然能从中获得很多价值。手机的核心价值主张已经存在了。但当然，我们后来构建了许多酷炫的触控界面，并发展了关于什么构成良好 App 用户体验的科学。

My high-level thought on this is, you always see this with new waves of technology. Mobile phone is a great example: the initial apps just looked like websites in a smaller box. Over time, you can still get a lot of value out of those. The core value prop of the phone was already there. But of course, we've built a lot of cool touch interfaces and developed the science of what makes a good app UX.

是啊。但我们没有多点触控，没有弹性滚动。

Yeah. But we've no multi-touch, we've no rubber banding.

是的。我认为我们现在正进入那个阶段。过去几年，我们只是用 AI 替换现有流程并做得更好。现在开始思考更多生成式流程。最简单的例子是很多产品现在底部有个聊天框：不用点菜单，直接问聊天框就行。这是非常简单的版本。但我认为还有更多创新空间。

Yeah. I think we are entering that phase now. For a few years, it was just replacing existing flows and using AI to do that better. Now we're starting to think about more generative flows. The simplest example is many products now have a chat box at the bottom: instead of clicking through menus, you can ask the chat box. That's a very simple version. But I think there's way more innovation to do.

我想到的一个框架是：晶体管和微芯片发明后不久，人们就意识到所有东西都会有一个微芯片。每样东西都能从一个小型计算机中受益。你的车会有一个，你的洗碗机也会有一个。类似地，所有东西在被消费之前都会经过一个 Transformer 模型。

One framing I was thinking about: shortly after the invention of the transistor and microchip, it became clear that everything would have a microchip. Everything could benefit from a small computer. Your car would have one, your dishwasher would have one. There's some equivalent where everything will pass through a Transformer model before it's consumed.

我的一个想法是，AI 在重要方面与之前的浪潮截然不同。个人电脑、互联网、手机——都有很大的硬件组件和网络效应。AI 没有这些问题。结果就是，一旦技术对某人有效，它就是纯软件，可以单人使用并直接带来巨大价值。它几乎对每个人都有效。我们看到每隔几周就有人宣称自己是增长最快的公司，从 100 万到 1 亿，因为 AI 快得多。但我认为产品方面有些滞后。即使今天冻结所有能力，没有新模型，仍然有整整十年的产品进步要做。以前，产品进步与分发同步。现在突然多了：总共两年大家都在思考，而对于智能体能力，还不到一年。我们都在应对这一点，试图找出正确的新产品体验。所以需要更多时间。

One of my thoughts on this is AI is uniquely different from previous waves in an important way. Personal computer, internet, mobile phone—all had a big hardware component and a network effect. AI has neither of those problems. As a result, as soon as the tech works for somebody, it's pure software, it can work single-player and give you a ton of value directly. It kind of works for everyone. We've seen a new person posting they're the fastest company from 1 million to 100 million every couple weeks because AI is so much faster. But I think there's a bit of lag with product. You could freeze all capabilities today and have no new models, and there would still be a whole decade of product progress to make. Before, product progress tracked alongside distribution. Now it's been much more sudden: two years total where everyone's been thinking about it, and for agentic capabilities, less than one year. We are all grappling with that and trying to figure out the right new product experiences. So it's taken a bit more time.

你的 AGI 时间线是什么？

What are your AGI timelines?

我认为我们已经有了 AGI。

I think we have AGI.

好吧，现在。

Okay, now.

嗯，有个笑话：2017 年你问我们有没有 AGI，答案是没有。今天，每个人第一句话就是你必须定义 AGI。

Well, there's this joke: back in 2017, if you asked if we have AGI, the answer is no. Today, the first thing everyone says is you have to define AGI.

是啊，这种支支吾吾。

Yeah, this hemming and hawing.

是的，在某种意义上确实如此。但我诚实的看法是，我认为人们谈论的那种快速奇点超级智能，我猜很难说，没有不可能，但我猜不会在近期发生，尤其是因为很多工作是收集现实世界的问题并定义成功。话虽如此，我认为这不是二元的。我们只会不断推出更多改进，这些东西会越来越强大，但我不认为未来几年会有突然的转变。

Yeah, it's kind of true in some sense. But my honest opinion is I think there is some rapid singularity superintelligence thing that people talk about. I would guess it's very hard to say, nothing's impossible, but I'd guess that's not something that happens in the immediate future, especially because a lot of the work is collecting real-world problems and defining success. With that said, I think it's not so binary. We're just going to keep rolling out more improvements, and these things will be more capable, but I don't know that we have some sudden shift for the next few years.

是啊，很有道理。

Yeah, that makes a lot of sense.

我们得聊聊 Windsurf。事情发展得太快了。给我们讲讲来龙去脉。我们听说消息是谷歌要收购 Windsurf，或者严格来说不是收购，整个交易在那个周五发生，和其他人同时知道。

We got to talk about Windsurf. It played out so quickly. Give us the play-by-play. We heard the news that it was going to be Google buying Windsurf, or not technically buying, and this whole deal that was happening that Friday the same time everyone else did.

好吧，这不是事先就计划好的。新闻出来的那个周五，对我们来说也几乎是突然的。我们可能前一天晚上听到了一些传闻。

Okay, so this was not something that played out in advance. The Friday when the news came out, it was basically just as sudden for us. We heard some rumors maybe the night before.

Deon 在为你刷推特。

Deon was scrolling Twitter for you.

对，没错。Devon 回来说：‘嘿，你们应该看看这个。’

Yeah, exactly. Devon came back and said, 'Hey, you guys should check this out.'

我们可能应该看看这个。所以我们当时听到了消息，自然那天下午我们就在讨论，思考是否应该对此采取行动。AI 领域出现一些疯狂的消息并不罕见，但这次尤其涉及我们的领域。我们讨论了这个想法，当晚就主动联系了他们，见到了 Windsurf 的新领导层 Jeff 和 Graham。那天晚上我们交谈时，共同得出了一个结论：如果真要做什么，那必须在周一早上之前准备好，因为所有客户都慌了，整个团队都在问‘我还有工作吗？’

We probably should look at this. And so we heard the news then and naturally that afternoon we're kind of talking about it and thinking about like is there something that we should do off of this. It's not uncommon that there's some crazy news that happens in AI, but this is especially I think in our space. And we talked about this idea, we reached out to them cold that evening and got to meet the new Windsurf leadership, Jeff and Graham, and that evening as we were both talking about it, I think we kind of came to this conclusion together: if there is something to do here at all, then it has to be ready to go by Monday morning, because everyone, all the customers were reeling, the whole team was like, 'Do I have a job? Do I not have a job?'

就像一块正在融化的冰块。

It was a melting ice cube.

没错。如果等到周四而不是周一，人们就会取消合同，开始去别处面试。所以我们说：‘好吧，这意味着如果我们想探索这个，就必须整个周末不停歇地投入。’有很多有趣的时刻。我们在周六达成了握手协议，然后显然还有所有法律事务要处理。我们周日晚上都熬了个通宵。我们当时有一个非常乐观的计划，以为能在周日晚上签完。

Exactly. And so if it even waited until Thursday instead of Monday, people were going to cancel their contracts, people were going to be interviewing at other places. So we said, 'Okay, this means if we want to explore this, we have to go spend the entire weekend on this non-stop.' A lot of fun moments there. We got to the handshake agreement that Saturday and then obviously there's all the legal and everything to figure out. We all pulled an all-nighter that Sunday night. We had a very optimistic plan that we were going to get it signed on Sunday night.

你周六晚上也通宵了，还是睡了点觉？

You also pulled an all-nighter this Saturday night or did you get some sleep?

我们周六睡了几个小时。特别要感谢 Jeff、Graham 和 Kevin，因为他们之前也经历了相当艰难的几天，所以进来时已经严重缺觉。我们一直在推进。我们乐观地认为能在周日晚上签完，然后就可以专注于拍摄和思考如何向团队传达等事宜。显然这没有发生，我们在周一早上 9 点左右才签完，因为我和律师们几乎整夜都在处理各种细节。幸运的是，我们在 Windsurf 工作室拍摄了 Windsurf 视频。我们说：‘好吧，无论如何还是拍了吧。’

We got a couple hours of sleep on Saturday. It was especially a huge shout out to Jeff and Graham and Kevin because they had had a pretty rough few days as well, so they were already pretty sleep-deprived coming into it. We were going through it. We had this optimistic view that we were going to get it signed on Sunday night and then we could go focus on filming and figuring out how we address the team and everything. Obviously that did not happen, and we got it signed on Monday at like 9:00 a.m. because us and the lawyers were up all night basically just sorting out all these things. We luckily filmed the Windsurf video in the Windsurf studio. We said, 'Okay, we should just film it anyway.'

你意识到你要在没有视频的情况下宣布收购了。

You realize you're gonna announce acquisitions without a video.

是啊，有一个总是好的。然后一旦签完，我们就站在整个团队面前，向他们通报最新情况，随后很快公开分享。这非常有趣。说实话，我活着就是为了这些时刻。

Yeah. Well, it's always nice to have one. And then as soon as we got things signed, we were up in front of the whole team and giving them the update and then sharing that publicly pretty soon after. It was a lot of fun. I live for these moments, honestly.

所以，你在周五看到了新闻。是的。然后你在周一签署并宣布了交易，但这意味着你几乎是瞬间就决定要收购 Windsurf 的剩余部分。

So, you read the news on Friday. Yeah. and you signed the deal on Monday, but that means that you decided more or less instantaneously that you wanted to buy the remaining part of Windsurf.

是的。所以我认为我们在周五晚上讨论过了，从我们的角度来看，这件事有几个好处。首先，我们非常了解这个领域。所以从这个意义上说，我们不需要对产品或客户进行尽职调查，因为我们了解。但当我们了解团队到底发生了什么、还有多少人留下、谁离开了时，我们发现有一种很好的协同效应：核心研究和产品工程团队去了 Google，而所有其他职能都完全保留，包括企业工程、基础设施、部署工程、市场推广、财务、营销、运营等等。有趣的是，对于 Cognition 来说，无论好坏，我认为我们在建立核心研究和产品工程团队方面做得很好，但在发展其他职能方面有点落后。所以我们发现那里也非常自然地契合。而且正如我们所说，他们有摩根大通，我们有高盛，所有这些都非常自然地契合。所以从我们的角度来看，是的，我们知道那里有一些非常有趣的东西，我们想去做，剩下的很多只是解决细节问题。

Yeah. So I think we talked it through on Friday evening and from our perspective there were a few things that were nice about this. First of all, we know the space very well. So in that sense, we didn't really have to diligence the product or the customers because we knew that. But as we were understanding the pieces of what happened exactly with the team, how many of the folks are still there and who left, we found that there was a very nice synergy in the sense that there was a core research and product engineering team that went to Google and all the other functions were entirely intact, which includes enterprise engineering, infra, deployed engineering, go-to-market, finance, marketing, operations, all these various things. And funnily enough, with Cognition, for better or worse, I think we had done a good job of building out this core research and product engineering team, but were a little bit behind on growing all the other functions. So we found a very natural fit there as well. And as we were just talking, they had JP Morgan and we had Goldman Sachs, and there were all of these kind of just very natural ways to fit in. So from our perspective, yeah, we knew there was something really interesting there and we wanted to do it, and a lot of the rest was just figuring out the details.

所以你收购了一群对这个领域非常熟悉的人。他们的产品与 Devin 相邻但不完全相同。因此你可以加速市场推广工作并拓宽产品组合。你是这么想的吗？

So you got to acquire a bunch of people who have lots of familiarity with the space. They have a product offering that is in an adjacent but not identical place to Devin. And so you get to accelerate the go-to-market efforts and broaden out the product portfolio. That's how you think about it.

是的，绝对如此。当然，产品本身，有趣的是，我们一直在思考像 Devin 这样的异步产品与更同步的产品如何交互，我们有一些关于构建某些同步功能的想法。我们不打算完全构建一个 IDE，因为感觉市场上已经有几个玩家了。但事实证明，有了这个想法，实际上与我们考虑的许多同步功能有很多自然的协同效应。一个非常简单的事情是，我们在交易完成后几天就发布了 Wave 11。还有很多基本的东西，比如在 IDE 中访问你的 deep wiki，或者在搜索中使用所有 Devin 代码库表示，或者在那里启动智能体，对吧？所有这些。我认为我们只是感觉到了很多自然的互补，所以从那里开始，感觉如果有合适的人一起合作做这件事，那就是他们。

Yeah. Absolutely. And then of course the products themselves, I think, funnily enough, we were thinking about what does the interaction of an async product like Devin look like with a more sync product, and we had some ideas for certain synchronous things that we wanted to build. We weren't going to build an IDE entirely because it felt like there were a couple players in town already. But as it turns out, having the idea, there actually were a lot of natural synergies with a lot of the synchronous stuff that we thought about. A very simple thing like we shipped Wave 11 a few days later after we closed that deal. And there are a lot of these basic things like being able to access your deep wiki in your IDE or being able to use all the Devin codebase representation in search, or spinning up the agent there, right? All of these things. I think we just felt a lot of natural complements, and so from there it felt like if there was a right person to work with and do this with, it would be them.

那么六个月后，我买 Devin 就能得到 Windsurf 捆绑包吗？还是我单独买 Windsurf，也可以买 Devin？会怎么运作？

So in six months, do I buy Devin and I get Windsurf bundle? Do I separately buy Windsurf and I can buy Devin? How will it work?

是的，还有很多要解决的。我们当然希望保持每个产品的理念不变，就像我提到的。我认为仍然会同时有同步和异步产品，但我认为让它们之间的集成更强大、更容易，我认为这会非常好。所以从客户的角度来看，肯定会有很多变得更简单，但如果出于某种原因他们真的只想用其中一个，我想他们仍然可以做到。

Yeah, a lot to figure out still. We certainly want to keep each of the product philosophies the same, like I mentioned. I think there will still continue to be both sync and async products, but I think making the integrations between them much stronger and much easier, and I think it's going to be really nice. So certainly a lot that'll be much easier from the customer perspective, but if for some reason they really wanted to use one of the two, I'd imagine that they would still be able to do that.

AI 领域有一个有趣的现象，就是出现了很多这种 49% 的许可交易，目的是为了避免收购被阻止的风险。公司购买 IP 许可，然后确保他们想要的人才也随之加入公司。你认为这在 AI 领域会持续存在吗？这有点像特定时期的产物，对吧？

It's obviously been an interesting aspect of the AI space that there's been a number of these 49% licensing type deals to avoid the risk of an acquisition being blocked. Companies buy a license to the IP and then the talent that they want to be sure comes with the company. Do you think that stays as a thing in AI? Like it's a funny moment in time thing, right?

是的，我当然不觉得自己是这方面的专家。我觉得有趣的是，每次都会出现一些新花样。我感觉围绕法律和监管层面，存在一种元游戏。你会看到一种模式，比如现在我们做这种许可交易，然后又是别的。所以我认为围绕这个的元游戏肯定在不断发展。在 AI 领域的顶层，存在一定的两极分化，因为到了一定程度，你希望这些东西随资源扩展，而它们也确实会扩展。所以我认为，基本上游戏规模越来越大，可以这么说。对大多数公司来说，问题在于他们是否认为自己能独自达到那个目标，还是想与另一家公司合作。

Yeah, I certainly don't feel like I'm the expert on this one. It's the thing that I find funny: there's one new bell or whistle each time. I feel like there's a meta-game around the legal and regulatory character scale. You see one thing, like now we do this licensing deal, and then something else. So I think the meta-game around that is certainly developing. There is some amount of polarity at the top level of AI space in the sense that there is a point at which you want these things to scale with resources, and they do scale. So I think basically the games get bigger, I guess is one way to put it. And for most companies, the question is whether they think they will get there themselves or whether they want to work with another company.

所以你是说你会预期更多的并购，无论是传统的并购还是这种新模式，因为在这个游戏中有规模效益。

So you're saying you would expect more M&A, whether it be classical M&A or this new model of M&A, because there are scale benefits in this game.

是的。也许我的一个大胆看法是，对于很多大公司来说，AI 领域当然会有小型或中型的成果，但我认为这个领域比以往更倾向于两极分化：你要么成为超大规模企业，要么就失败。所以对于那些认为这就是他们想要的发展轨迹和登月计划的公司来说，那是一回事。但对其他公司来说，与别人合作是常有的事。

Yeah. Like maybe one of my hot takes is that for a lot of the big ones, of course there will be mini or medium-sized outcomes in AI, but I think in this space a little bit more so than previous ones, it's a little bit more polarized towards you become a hyperscaler or bust. So for some companies that feel like that is the trajectory and the moonshot they want to go for, that's one thing. But for others, working with someone is something that people do.

现在你们把 Windsurf 团队纳入麾下，Cognition 有着非常紧张的文化。你们周末工作，都在这栋房子里工作。而你们正在做这个买断要约。

Now as you're bringing the Windsurf team on board, Cognition has this very intense culture. You guys work on weekends, all work out of this house. And you're doing this buyout offer.

是的。对我们来说，大多数人都非常兴奋地加入并投入工作，只有一小部分人接受了买断。但从我们的角度来看，我们只是想确保这对每个人来说都是自愿选择的情况，因为说实话，这不适合所有人。而且这是一个非常刻意的安排。有这种紧张感。我们希望人们选择接受这种紧张感和新文化。我们将追求一些非常雄心勃勃的目标。按收入标准或随便你怎么说，人们可能会称我们为中期或后期公司，但从我们的角度来看，就接下来的发展以及还有多少东西要构建和完成而言，我们仍然非常早期。显然，在早期阶段，我们都必须接受不确定性，并愿意每周迎接不同的挑战，投入大量时间，并拥有那种文化。这是很重要的一部分。无论发生什么，我们都想确保人们得到很好的照顾。

Yeah. For us, most folks have been really excited to come in and do it, and only a small fraction have taken the buyout. But from our perspective, we just want to make sure it's an opt-in situation for everyone, because let's be honest, it isn't for everyone. And it is a very intentional thing. There's the intensity. You want people to opt into the intensity and the new culture. We're going to be going after some very ambitious goals. By revenue standards or whatever you want to call it, folks might call us a mid or later stage company, but from our perspective, we are still very much early stage in terms of the profile of what happens next and how much more there is to build and do. And obviously at an early stage, we all have to sign up for the uncertainty and the willingness to take on a different challenge every week, to put in a lot of hours and to have that culture. That was a big piece of it. Regardless of what happens, we wanted to make sure people were well taken care of.

每天 Cognition 都是你经营过的最大公司。你正在快速学习如何经营一家公司。我很好奇你是怎么学这些东西的。你怎么用 AI？但更广泛地说，你是怎么学习的？

Every day Cognition is the largest company you've ever run. You're speed running learning how to run a company. I'm curious how you learn this stuff. How do you use AI? But how do you learn more broadly?

是的，我们肯定还有很多要学的。我认为很多职能，如果说有什么的话，我们在很多职能上投资不足，可能是因为它们没有像应该的那样成为我们的首要关注点，而现在我们正在积极努力地加强这些方面。我不相信字面意义上的专业教练或职业教练，但显然你可以从做类似事情的同行和朋友那里学到很多。所以有很多亲密的朋友——显然是你一起上数学营的人——从这些不同的人那里学习。我确实认为，作为创业者，有一群亲密的朋友非常有帮助，你可以非常坦诚地对他们说：‘这件事完全搞砸了，我不知道我们该怎么办，请告诉我你以前是否做过类似的事情’之类的，这真的很有帮助。比如 RAMP 的 Eric 和 Kareem，或者数学竞赛中的各种朋友，或者我之前的联合创始人 Lunch Club 的 Vlad——我和很多不同的人交流寻求建议，我认为这真的很有帮助。

Yeah, we've got a lot to learn still for sure. I think many of these functions, if anything, we have underinvested in a lot of functions maybe because they're not as top of mind for us as they should be, and now that's something we're pretty actively working to do more of. I don't believe in a professional coach or career coach in the literal sense, but I think obviously you learn a lot from your peers and your friends who are doing similar things. So having a lot of close friends who are — people you went to math camp with apparently — learning from all these different folks. I do think as an entrepreneur, it helps a lot to have a close group of friends that you can just be very honest with and say, 'This thing is totally messed up and I have no idea what we're going to do, and please tell me if you have done anything like this before,' or things like that, which has been really helpful. Eric and Kareem from RAMP, for example, or all these various folks from math competitions, or my previous co-founder Vlad from Lunch Club — a lot of different folks that I talk to for advice, and I think it really does help a lot.

最后一个问题。我很好奇你的信息摄入方式是什么样的，你是怎么了解世界的？

Last question. I'm curious how what is your information diet in terms of how you learn about the world?

嗯，很多——我觉得 Twitter 确实是获取科技新闻的地方。我们分享很多——

Yeah, a lot of — I feel like Twitter is really the place to be for tech news. We share a lot of —

现在算法里的视频太多了。

There's too much video in the algorithm these days.

我觉得他们——视频很多，但大部分时候我就是不看视频，或者只看头几秒。这对制作视频的人来说也是一个有趣的点：确保你可以在没有声音的情况下，在前三秒内传达你的观点。尽可能做到这一点，我认为你还能触及另外 5 倍的用户，他们属于那一类。Twitter 的算法就是 AI 影响我信息的程度。

I think they're — there is a lot of video, but then I just don't watch the videos for the most part, or you see the first few seconds. Which is an interesting thing to think about as people who are making videos too: make sure you can convey your point with no sound and in the first three seconds. As much as you can do that, I think there are still another 5x of users you reach that are in that camp. The Twitter algorithm is the extent of how AI affects my information.

但那是你作为 AI 的接收端，而不是你使用 AI 作为工具。

But that's you on the receiving end of AI as opposed to you using AI as a tool.

说得好。我是说，我应该让 Devin 做一个 GitHub Action 的晨间任务，基本上就是让 Devin 去做晨间报告并获取信息。还有很多自动化要做。

It's a good point. I mean, I should have Devin do a GitHub Action morning job basically where Devin just goes and does the morning report and gets that. There's a lot of automation to do still.

总统每日简报。

The president's daily briefing.

是的。

Yeah.

好了，Scott，谢谢你。这太棒了。

Well, Scott, thank you. This is awesome.

非常感谢你邀请我。

Thank you so much for having me.

显然，一旦那变得太容易，自然的事情就是拿八张牌，然后做出当前年份。

The natural thing obviously once that got too easy is then you take eight cards and you make the current year.

而今年是 2025 年。

And this year is now 2025.

所以我们来试试这个，实际上很简单，因为 13 减 12 等于 1，11 减 10 等于 1。1 加 1 加 7 等于 9。结果 9 乘以 9 乘以 5 乘以 5 正好是 2025。

And so we get to try that one and in this case it's actually quite easy because 13 - 12 is 1, 11 - 10 is 1. 1 plus 1 plus 7 is nine. And it turns out that 9 times 9 times 5 times 5 is exactly 2025.

好吧。不过，有时候这太简单了。

Okay. Well, but at some point that gets too easy.

现在，你们可以给我一个最喜欢的四位数，比如说，然后我们试着凑出来。

And now, you know, you guys can give me your favorite four-digit number, let's say, and then we'll try and make that.

那就用 6843 吧。

Let's go with 6,843.

6843。好的。这个稍微大一点。所以我们用九张牌。6843。好。6 加 4 等于 10。12 减 10 等于 2。2 乘以 10 乘以 10 等于 200，加 8 等于 208，乘以 11 等于 2288，减 7 等于 2281，再乘以 3 等于 6843。

6843. Okay. So, it's a little bit bigger. So, we'll use nine cards for that now. Okay. 6843. Right. So 6 + 4 is 10. 12 - 10 is 2. 2 times 10 times 10 is 200 plus 8 is 208 times 11 is 2288 minus 7 is 2281 and times 3 is 6843.

干杯，你正好算出来了。是啊，我还没喝够健力士，所以没注意到。

Cheers while you get exactly that. Yeah. I haven't had enough Guinness yet to be able to notice.

没错。现在我们来测试一下，随着你喝更多健力士，表现会如何变化。

Exactly. And now we test as you drink more Guinness how performance goes.