Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next
频道: Sequoia Capital
视频: https://www.youtube.com/watch?v=SlGRN8jh2RI&t=659s
原文语言: en
统计: 共 27 轮 · Asia 1 · Boris 18
[0:02] Asia
Okay, I'm excited to introduce our next speaker. Show of hands, who here uses Claude code? Okay, show of hands, who here has Claude code psychosis? Come on guys, [clears throat] it's okay. It's okay. Um my my my team lovingly says I have Claude code psychosis, which may or may not be true. Um we are delighted to have Boris Cherny with us today. Uh Boris is the creator, the father of Claude code. Um and uh in the process of doing that has just had a front row seat to to reinventing uh the modern way of of software development. Um and we're we're really grateful to you, Boris, for taking the time to speak with us today. We know that um the entirety of software development kind of rests on your shoulders. So, thank you for taking it out of your time to be with us today. And interviewing Boris is Lauren Reader from our team. Thank you.
好，我很高兴介绍我们的下一位嘉宾。举个手，在座有谁用 Claude Code？好。再举个手，有谁得了「Claude Code 上瘾症」？来吧各位，没关系的，没关系。我团队就特别有爱地说我得了 Claude Code 上瘾症——这话可能是真的，也可能不是。今天我们非常高兴请到了 Boris Cherny。Boris 是 Claude Code 的创造者、它的「父亲」。在做这件事的过程中，他也算是坐在头排，亲眼见证了现代软件开发方式被重新发明。Boris，真的非常感谢你今天抽空来跟我们聊。我们都知道，整个软件开发某种程度上就压在你肩上，所以特别感谢你挤出时间来参加。今天负责采访 Boris 的，是我们团队的 Lauren Reader。谢谢。
[0:48]
[applause]
（掌声）
[0:52] Boris
Giving our chairs. Um you took my you took my opening line, Asia. We asked who here uses Claude code. There's a lot of hands. That's awesome. Thank you for joining us, Boris. It's very special to have you here. Um as a roomful of builders, I think you are changing building entirely. And so, I'm very curious to explore how you think about the future of software, coding, and what we should spend all of our free time on. Um but I'll give you a me a tiny bit more background on you so that everyone has a little bit more context. So, beyond creating Claude code, Boris is very much an engineer's engineer. You were writing a lot of code through your whole career, writing textbooks about code, including programming in TypeScript. Um and I think last time we chatted you hadn't written a single line of code in the last year, or at least so far in 2026, which is quite the change. Um There's also a a little known thing back in middle school, I wrote a guide about uh writing BASIC for TI-83 Plus calculators. And I I just I I searched for it, it's actually still on the internet. It's extremely embarrassing, so please don't search it. But it [laughter] exists. We will definitely be finding that. Um so, we're going to do I'm going to start with a few questions here. Maybe we'll start with a little bit of the history of Claude code, how you started it, and then we're going to have a lot of audience Q&A for this one. And so, start thinking about your questions in the back of your head, uh and would love to turn it over to you all soon. Yeah. Um and also real quick, so for people that use Claude code, do people use the CLI mostly? Like okay, majority CLI? Okay. That's a lot. Majority desktop? Okay. Majority VS code or JetBrains IDE? Okay. That's actually not a lot. Okay. Other? I'm like iOS mostly these days. Yeah.
我们先入座吧。Asia，你把我的开场白给抢了。我们刚问在座有谁用 Claude Code，好多人举手，太棒了。谢谢你来，Boris，能请到你真的很特别。这屋子里坐的都是 builder，而我觉得你正在彻底改变「做产品、写代码」这件事本身。所以我特别好奇，想跟你聊聊你怎么看软件和编程的未来，以及我们这些空闲时间到底该花在什么上。不过我先再多介绍一点你的背景，让大家有点上下文。除了创造 Claude Code，Boris 还是个不折不扣的「工程师中的工程师」。你整个职业生涯都在大量写代码，还写过讲代码的书，包括《Programming TypeScript》。我记得上次我们聊的时候，你说过去一年——至少 2026 年到目前为止——你一行代码都没亲手写过，这变化挺大的。还有件鲜为人知的事：早在初中的时候，我写过一篇指南，讲怎么在 TI-83 Plus 计算器上写 BASIC。我搜了一下，它居然现在还挂在网上。特别尴尬，所以拜托大家别去搜。但它（笑声）确实还在。我们肯定会把它扒出来的。好，下面我会先问几个问题。我们大概先从 Claude Code 的来历讲起，你是怎么开始做它的，然后这一场我们会留很多观众 Q&A 的时间。所以大家可以开始在脑子里默默准备问题了，待会儿就交给你们提问。对了，再快速问一句，用 Claude Code 的各位，大家主要是用 CLI 吗？好——大多数用 CLI？这么多。主要用桌面端的？好。主要用 VS Code 或者 JetBrains IDE 的？好，这个其实不多。好，还有别的吗？我自己最近基本上都是在 iOS 上用。对。
[2:37]
[laughter]
（笑声）
[2:37] Boris
Okay. Cool. Um yeah, so I started Claude code kind of accidentally in a in a lot of ways. Um I joined this team back in late 2024. It was a sort of this incubator within Anthropic called Anthropic Labs. And uh the team kind of served its purpose. Um we created Claude code, uh MCP, and the desktop app. It was a team it was just a few of us. So, very much like innovation team. We built the thing that we wanted to build, we disbanded the team. Uh now the team's actually back together for round two. Mike Krieger, who's the you know, like the chief product officer at at Anthropic and used to be one of the founders at Instagram, so he's leading that right now. Um so the kind of the the the the reason that I started to work on coding is we felt like there was this product overhang. And I I'm guessing people here use that word a lot. Uh but we definitely use this word a lot in kind of within the lab. Uh there's this idea that the model can do all the stuff that no product has yet captured. And in late 2024, when we were looking at coding, the way that we did coding, the state of the art at the time was type ahead. It was you open your IDE and you press tab and you can like complete like one line at a time. And that was the thing that Sonnet 3.5 enabled for the first time. But the feeling was we could actually go a lot further than that. And the model was almost ready for the next big step. So, we don't have to do type ahead anymore, we can just have the agent write all of the code. And so, I built it, and it just really didn't work for the first 6 months. It was like not very good. It was barely usable. I wrote it from I used it for maybe 10% of my code or something like that. And even after we released Claude code initially, it was not a hit. There's a lot of people that used it, but it did not have this exponential growth that it has today. Um that started with Opus 4 in May. And I I remember that very clearly. That's like when the exponential growth started, and then it kind of inflected with every model release. Uh like it started with Opus 4, then 4.5, then 4.6, now 4.7. It just kind of keeps inflecting. But essentially, we were trying to build this thing that was like pre-PMF, and we knew that it wouldn't have PMF for 6 months because we were building for the next model. And that was the idea the pretty much the whole time.
好，酷。对，其实在很多意义上，我是有点「误打误撞」做出 Claude Code 的。我是 2024 年底加入这个团队的，当时是 Anthropic 内部一个类似孵化器的部门，叫 Anthropic Labs。这个团队后来算是完成了它的使命——我们做出了 Claude Code、MCP，还有桌面 app。团队规模很小，就我们几个人，非常典型的创新小队：把我们想做的东西做出来，然后就把团队解散了。现在这个团队其实又「合体」做第二季了。Mike Krieger——就是 Anthropic 的首席产品官、也是 Instagram 当年的联合创始人之一——现在在带这个团队。我当初会开始做编程，原因是我们觉得这里存在一个「产品红利尚未兑现」（product overhang）的空间。我猜在座各位经常用这个词，反正我们在 Lab 里是天天挂在嘴边。意思就是：模型已经能做很多事了，但还没有任何产品把这些能力真正接住、变现出来。2024 年底我们看编程这件事的时候，当时的「最高水平」就是自动补全（type ahead）——你打开 IDE，按一下 Tab，它一次帮你补全一行代码。这是 Sonnet 3.5 第一次让人能做到的事。但我们的直觉是：其实可以走得比这远得多，而且模型马上就要准备好迈出下一大步了。所以我们不必再停留在自动补全，完全可以让 agent 把所有代码都写了。于是我就把它做了出来——结果头 6 个月真的根本不好用。质量很差，勉强能用。我自己大概也就 10% 的代码会用它来写。哪怕在我们最初发布 Claude Code 之后，它也算不上爆款。是有不少人用，但完全没有今天这种指数级增长。那种增长是从 5 月份的 Opus 4 开始的，我记得特别清楚。指数增长就是那时候起来的，之后基本每发一个模型就再「拐」一次：先是 Opus 4，然后 4.5、4.6，现在 4.7，就这么一路不停往上拐。本质上，我们当时是在做一个「还没到 PMF（产品市场契合）」的东西，而且我们心里清楚它 6 个月内都不会有 PMF——因为我们是冲着下一代模型在做。这个思路基本贯穿了整个过程。
[4:46] Boris
And you know, for Anthropic in general, we've always just been very focused. We've always cared about business and enterprise and safety and coding. That's just always been kind of the way that we wanted to build. And so, at some point we kind of knew that we wanted to build a product. We didn't know exactly what we wanted. So, this kind of ended up being the the product bet. It's an incredible story, especially that it was an accident. Um so, you've said on the record that you think coding is solved. Uh if this is one of the three best from Anthropic, can you tell us more about what you mean by that, and what might still not be solved, or what second-order problems might come? All right. I can ask another question for the room. Who writes 100% of their code by hand? Who writes 100% of their code using a agent like Claude code? Okay. Who's like somewhere in between? Okay. So, like 50% solved.
你知道，对 Anthropic 整体来说，我们一直都非常聚焦。我们一向看重商业、企业级、安全，还有编程，这一直就是我们想要的做法。所以到某个时间点，我们大致知道自己想做一个产品，只是不太确定具体要做什么。结果这个就成了我们押下的那个「产品赌注」。这故事太精彩了，尤其它还是个意外。对了，你公开说过你认为「编程已经被解决了」。如果这算 Anthropic 最棒的几件事之一，能不能多讲讲你这话到底是什么意思？还有什么可能仍然没被解决？或者会冒出哪些二阶问题？好，那我再向全场问一个问题：有谁 100% 的代码是纯手写的？有谁 100% 的代码是用 agent、比如 Claude Code 写的？好。有谁是介于两者之间的？好，那大概就是「解决了一半」。
[5:38]
[laughter]
（笑声）
[5:41] Boris
I mean, for me it's for me it's like for me it's 100%. Like the the Claude code code base, um you know, it leaked, so you know, people know. Uh it's pretty simple. It's just like TypeScript and it's React. Like there's no big secret. There's there's nothing really complicated. The the reason we picked TypeScript and React is it's very on distribution for the model. So, when we started, you know, building the code base, the model was not as intelligent as it is today, so the language and the framework mattered a lot. Nowadays, you know, it can write whatever, and it can pick up new languages, new frameworks it hasn't seen. But back then, you wanted to use something pretty on distribution. Because of that, I think fairly early we got to the point where the model just wrote 100% of the code. And for us, this happened sometime in October, November last year. And so, for me today, you know, like the model writes 100% of my code. I write somewhere, you know, usually a few dozen PRs every day. Uh there was a day last week I did like 150 PRs in a day. That was like that was a record. I was just trying to kind of push to see how far I can get it. Um but yeah, it's like for me for me it's just solved. Um but this is not the case everywhere. There's very big complicated code bases. There's kind of weird languages the model's not good at yet. Um and you know, as everyone here knows, it's it's getting there. Usually the answer is just wait for the next model. Can you actually tell us about your personal setup? You walked us through it the other day. It is pretty wild. Yeah. Um so, I shared my personal setup like 6 months ago or something on on Twitter. And it it's funny, I actually I shared it I didn't realize that it would be surprising for anyone. That was just like the way that I coded.
对我来说，几乎是 100%。Claude Code 的 codebase 之前泄露过，所以大家都知道——其实特别简单，就是 TypeScript 加 React，没什么大秘密，也没什么真正复杂的东西。我们当初选 TypeScript 和 React，是因为它们对模型来说非常「在分布内」。我们刚开始搭这个 codebase 的时候，模型还没今天这么聪明，所以用什么语言、什么框架很关键。现在它想写什么都行，能上手没见过的新语言、新框架。但在当时，你会想用一些比较主流、在分布内的东西。正因为这样，我觉得我们挺早就走到了模型把代码 100% 写完的地步。对我们来说大概是去年十月、十一月发生的。所以现在对我来说，我的代码 100% 是模型写的。我每天通常会提几十个 PR。上周有一天我一天提了大概 150 个 PR，算是个纪录。我当时就是想推一推，看看能把它逼到什么程度。所以对我来说，这事儿基本就解决了。但不是哪儿都这样。有些 codebase 非常大、非常复杂，还有些比较冷门的语言模型现在还不太擅长。不过在座各位都知道，它正在补齐——通常答案就是等下一个模型。能不能跟我们讲讲你个人的 setup？你那天给我们演示过一遍，挺野的。嗯，我大概六个月前在 Twitter 上分享过我的个人 setup。有意思的是，我当时分享的时候根本没意识到这会让谁觉得意外，那就是我平时写代码的方式而已。
[7:07]
[laughter]
（笑声）
[7:08] Boris
And it's changed since then. It's changed. Um and so, now actually most of my work I do from my phone. Um and so, I don't know if like you guys won't be able to see this, but I have um so, I have like the Claude app, and if you open the Claude app, on the left-hand side, there's this little code tab, and I just have a bunch of sessions going. Um you you probably can't see it.
从那之后又变了。又变了。现在其实我大部分活儿都是在手机上干的。我不知道你们这边看不看得到——我手机上有 Claude app，打开 Claude app，左边有个小小的 code 标签页，我就同时开着一堆 session 在跑。你们大概看不清。
[7:33] Boris
How many sessions? Uh usually have like maybe like five to 10 sessions. Uh and then the sessions usually have a bunch of agents, so I think currently probably like a few hundred agents going. Um usually every night I have like a few thousand that are doing kind of deeper work. There's a few ways to manage it. One is that you ask Claude to use a bunch of sub-agents to do work. Actually, the the thing that I've been finding myself using more and more is the loop. So, this is {slash} loop, and it's just like the coolest thing. It's like the simplest thing that works. All it is is you have Claude use cron to schedule a job for some point in the future, and it's a repeat job. And it can run every every minute, every 5 minutes, every day, kind of however often you want to schedule it. And at [snorts] this point, I have like dozens of loops that are running for stuff. So, I have one that's babysitting my PRs, like fixing CI, auto-rebasing. I have another one that keeps CI healthy. So, like if there's like a flaky test or whatever, it'll it'll go and fix it. Um I have another one that grabs uh feedback from Twitter and kind of clusters it for me every 30 minutes. So, I just have a bunch of these loops running at any time. I sort of feel like loops are the future at this point. If you haven't experimented with it, highly highly recommend it. And we also just launched routines, which is the same thing but kind of on the server. So, even if you close your laptop, it it keeps going. So, that's your personal setup. Tell us about what you think teams will look like in the future. How do you extrapolate from all the work you're doing to keep everyone on the team moving forward, understanding the context, or do you think we need to let go of a lot more to agents to make it work? Um I think so I you know, it's like it's so hard to make predictions, but um I'm here to make predictions, so I'll try to make some. I I I feel like the way that things are going is generally there's going to be a lot more generalists than there are today. And today when we talk about generalists, I think largely we're talking about people that are still engineers. So, they're still writing code, but maybe they're kind of product engineers. So, maybe when we say generalist, it's like a you know, they do iOS and web and server, for example. That's like a generalist in engineering.
（开多少个 session？）通常大概五到十个 session 吧。然后每个 session 底下一般又挂着一堆 agent，所以我现在估计同时有几百个 agent 在跑。一般每天晚上还有几千个在做那种更深的活儿。管理它们有几种办法。一种是让 Claude 用一堆 sub-agent 去干活。不过我发现自己越来越常用的，其实是 loop。就是 /loop，这玩意儿太酷了，是那种「最简单但就是管用」的东西。它本质上就是让 Claude 用 cron 在未来某个时间点排一个任务，而且是重复任务，可以每分钟、每五分钟、每天跑一次，你想多久跑一次都行。到现在，我手头有好几十个 loop 在跑各种东西。有一个在帮我看着 PR，比如修 CI、自动 rebase；另一个专门保持 CI 健康，要是有个 flaky test 之类的，它就去把它修了；还有一个每隔 30 分钟从 Twitter 上抓反馈、帮我聚类。所以我随时都开着一堆这种 loop。我现在多少觉得 loop 就是未来。如果你还没试过，强烈强烈推荐。我们还刚上线了 routines，跟 loop 是一回事，但是跑在服务端的。所以哪怕你把笔记本一合，它还接着跑。这就是你的个人 setup。能不能讲讲你觉得未来团队会是什么样？从你现在做的这些活儿往外推——你怎么让团队里每个人都跟着往前走、都掌握上下文，还是说你觉得我们得把更多东西放手交给 agent 才能转得起来？我觉得吧——做预测真的太难了，但我今天就是来做预测的，那我就试着说几个。我感觉事情大方向上是：以后通才会比今天多得多。今天我们说通才，我觉得多半还是指那种本质上还是工程师的人——他们还在写代码，只是可能算产品工程师。所以说通才的时候，比如指一个人同时搞 iOS、web 和 server，这算是工程里的通才。
[9:35] Boris
But I think the thing that we're going to start to see a lot more of is generalists that are cross-disciplinary. So, this is engineers that are really good at product engineering, but also really great at design. Or really great at product and data science and engineering. Um I don't know. It's it's something that we're starting to see on our team. So, actually like a lot of people on the Claude code team are generalists across disciplines. Everyone on our team codes. So, like our engineering manager, our product manager, our designers, our data scientist, our finance guy, our user researcher, every single person on our team writes code. And so, you know, like they're specialist in something, but now also everyone's just coding. And you know, I'm seeing some nods, but I bet also it's actually not that surprising to people in this room cuz I bet you're seeing the same things. Um [clears throat] I'll have one more favorite questions then we'll open up to the audience. So, we talked a bit about what's changing with coding. I'm curious about what you see changing in the world of software or software products. Um I think as we see AI making writing code 10 or 100x cheaper, what happens to the value of the products that are produced with software? Do we have a SAS apocalypse on our hands? How do you think this plays out? And again, you're going to have to make another prediction. The SAS apocalypse question is my favorite question then. Um I think there's two things that are going to happen and I I don't think either of them is the thing that people have been talking about. I think one is Is anyone here an acquired listener? Like the acquired podcast? Yeah, it's like the best podcast. Uh I actually I I got to do a unplugged with them the other week and I I just I I felt like I got to like meet my heroes cuz they're they're just like the hosts are the best. So, they have this idea of uh seven powers and and this is a this is like Hamilton. He kind of wrote he wrote a book about this and this is kind of the seven modes in business. And I think what's going to happen is because of AI, some of these modes are going to get more important and some are going to get less important. And so, like for example, one that gets less important is uh switching costs because you can just use the model and you can kind of port from one thing to a different thing.
但我觉得接下来会越来越多见到的，是那种跨学科的通才。就是说，工程师既很擅长产品工程，同时也非常擅长设计；或者既精通产品，又懂数据科学和工程。我也说不准——这是我们团队里已经开始看到的现象。其实 Claude Code 团队里很多人都是跨学科的通才。我们团队每个人都写代码。我们的工程经理、产品经理、设计师、数据科学家、做财务的、用户研究员——团队里每一个人都写代码。所以他们各自在某个领域是专家，但现在大家也都在写代码。我看到有人在点头，不过我猜这对在座各位其实也没那么意外，因为你们八成也在看到一样的事。我再问一个我最喜欢的问题，然后就开放给观众。我们刚聊了一点编程正在发生的变化。我好奇你看到软件、或者说软件产品的世界正在发生什么变化？随着 AI 让写代码便宜 10 倍、100 倍，用软件做出来的产品，它们的价值会怎么样？我们是不是要迎来一场 SaaS 大崩塌？你觉得这事儿会怎么演变？同样，你又得做一个预测了。SaaS 末日这个问题，那它就是我最喜欢的问题了。我觉得会发生两件事，而且这两件我都觉得不是大家一直在讨论的那回事。第一件——在座有听 Acquired 的吗？就那个 Acquired 播客？对，那是最棒的播客。我前阵子还跟他们做了一期 unplugged，那感觉就像见到了我的偶像，因为两位主持人真的太牛了。他们提过一个概念叫「七种力量」（seven powers），这是 Hamilton 写的，他写了本书讲这个，讲的是商业里的七种护城河。我觉得接下来会发生的是：因为 AI，其中有些护城河会变得更重要，有些会变得没那么重要。比如说，变得没那么重要的一个是「转换成本」（switching costs），因为你现在用模型就能很轻松地从一个东西迁到另一个东西上。
[11:41] Boris
Another one that gets less important is process power because for companies whose mode is like workflows and process and things like this, Claude is getting really good at figuring out process. And especially with 4.7, it can just hill climb anything. So, if you give it a target and you tell it to iterate until it's done, it will just do it. I think this is the first model like that. So, I think these are going to get less important, but I think the previous modes actually still matter. So, this is like network effects, uh scale economies, cornered resources, things like that. These are not really changing with AI. I think the second thing is if you look at the number of startups today or like maybe in the next you know, the past 10 years, I think the number of startups in the next 10 years that are just going to like disrupt everything is going to increase like 10x. Because right now you can be a tiny startup, you could build a thing that's as valuable as a large company and you can actually compete head-to-head because the large company has to evolve their business process, they have to evolve the way they work, they have to retrain everyone to use technology, they're going to face a lot of internal resistance to that. But you know, no one here has that problem. If you're starting fresh, then you can kind of build with AI natively from the ground up. So, I don't know. I I think it's the best time to build. It's the best time to be a startup. It's there's so much disruption coming. So, there is hope for us after all. Thank you, Boris. Um I would love to open up to audience questions if anyone has anything they would like to ask. Dan? I Yeah, I'm curious. Um you said that you built uh 6 months before there was product market fit, but now given that the models are good enough, how much do you attribute the success of Claude code to the model versus like product decisions in the the like field of product? Uh I think it's probably a mix. Yeah, I think it's a mix. I think I think if you asked me maybe a year ago, the ratio was maybe something like 50/50. Um maybe I don't know. If you asked me 6 months ago, the mix would be 50/50.
另一个变得没那么重要的是「流程力量」（process power），因为对那些护城河是工作流、流程之类东西的公司来说，Claude 在摸清流程这件事上越来越强了。尤其是到了 4.7，它几乎能爬坡（hill climb）任何东西。你给它一个目标、让它一直迭代到搞定为止，它就真能做到。我觉得这是第一个能做到这样的模型。所以我觉得这些护城河会变得没那么重要，但前面那几种其实还是有用的——比如网络效应、规模经济、稀缺资源（cornered resources）这些，它们并没有真的因为 AI 而改变。第二件事是：你看今天、或者说过去十年里创业公司的数量，我觉得未来十年那种能把一切都颠覆掉的创业公司，数量会涨大概 10 倍。因为现在你可以是一家小小的创业公司，做出一个跟大公司一样有价值的东西，而且能跟它正面硬刚——因为大公司得改造它的业务流程、改造它的工作方式、重新培训所有人去用新技术，内部会遇到一大堆阻力。但在座没人有这个问题。如果你是从零开始的，那你就能从底层开始原生地用 AI 来搭。所以我觉得现在是最好的创业时机，是当创业公司最好的时候，有太多颠覆正在到来。所以说，我们终究还是有希望的。谢谢你，Boris。我很想开放给观众提问，有谁想问点什么的话。Dan？嗯，我挺好奇的。你说你在产品市场契合（PMF）之前六个月就开始搭了，但现在既然模型已经足够好了，你觉得 Claude Code 的成功，有多少要归功于模型、有多少要归功于产品决策，也就是产品本身这一块？我觉得大概是两者兼有吧。对，是混在一起的。我觉得，要是你大概一年前问我，这个比例可能差不多是五五开。可能吧，我也说不准。要是你六个月前问我，也是五五开。
[13:35] Boris
What about in 2 years? Oh, 2 year I don't know, dude. We plan in like we plan in 1 week out.
（那两年后呢？）噢，两年后……我不知道啊老兄。我们做计划基本是按一周来规划的。
[13:40] Boris
months. Sometime in the future.
几个月。未来的某个时候吧。
[13:41]
[laughter]
（笑声）
[13:42] Boris
And by the way, I think the reason it was 50/50 is um you know, I I I like I I did YC back in the day. I was like the first hire at a YC company and like I did a bunch of startups. And in startups like the thing that they drill into you and then especially in YC over and over is build something people love. And so, it it doesn't matter what the product is, it doesn't matter like the model and all this stuff. You still in the end have to build a thing that people love. And I think that's that's why the product matters is we we pay so much attention to the little details so that as you use it all day, it's a really great experience. I think as the model's gotten better, the harness kind of gets less important. And I I think like I think that we're thinking about right now is like how do we evolve the harness? So, like how do we make loops more of a first class thing? How do we make it easier to run a lot of agents? Uh you know, beside you know, like sub agents is one idea. There's a bunch more stuff that we're cooking. But I think in a year, the model will be much better aligned. And so, all the safety mechanisms that we have today around uh prompt injection and kind of static verification of commands and uh permission modes, human in the loop, all this kind of stuff is just going to be less important cuz the model will just do the right thing. Um So, yeah, that's that's my prediction. Thank you. You want to toss the box, Dan?
顺便说一句，我觉得当初之所以是五五开，是因为——你知道，我早年做过 YC，是一家 YC 公司的第一号员工，还折腾过好几家创业公司。在创业里，他们反复给你灌输的一件事，尤其是在 YC 里被一遍遍强调的，就是「做一个用户真心喜欢的东西」。所以不管产品是什么，不管模型怎么样、那些东西怎么样，你最终还是得做出一个让人真心喜欢的东西。我觉得这就是产品为什么重要的原因——我们会在那些小细节上下足功夫，这样你一整天用下来，体验就特别好。我觉得随着模型越来越强，外层的 harness 反而会变得没那么重要。我们现在在琢磨的是：怎么让 harness 进化？比如怎么把循环（loop）变成一等公民？怎么让同时跑很多 agent 变得更容易？你知道，sub agent 是其中一个思路，我们手里还有一堆别的东西在酝酿。但我觉得一年之内，模型会对齐得好得多。所以我们今天围绕 prompt injection、命令的静态校验、权限模式、human in the loop 这些搞的所有安全机制，到时候都会没那么重要，因为模型自己就会做对的事。嗯，这就是我的预测。谢谢。Dan，要不要把话筒抛过去？
[14:57]
[snorts]
（嗤笑）
[14:59] Boris
Great. Um To zoom to zoom out a little bit from software, I think Claude code did a cultural change a few months ago where it democratized like building software. You can see uh shop owners building their own um software for themselves or even uh programming microcontrollers to control the light when someone opens the door. Um do you see in the future um building software becoming a skill like uh I know uh Microsoft Office? Um so, it's a thing that ev- everybody can do, not just people in the tech industry? Oh my god, yes. Yes. Yes. I I I think it's going to be even more than that. I think it's going to be I don't know. It's going to be a skill like yeah, like I know how to send a text message. I I I think um you know, like I I read a my my two genres are essentially sci-fi and tech history. This is what I read a lot of. I I think in tech history, there's one thing which I think to me is the clearest parallel for what's happening right now. And this is in the 1400s, the printing press in Europe. And what what happened was before the printing press, essentially 10% of the European population was literate. They knew how to read and write. They were often employed by like kings and lords that were not literate. And their job was to you know, their their job was to read and write and this is not something that everyone knew how to do.
太好了。从软件这个话题往外拉一点——我觉得 Claude Code 几个月前带来了一场文化上的转变，它把「开发软件」这件事民主化了。你能看到店主自己给自己做软件，甚至有人去给微控制器写程序，让灯在有人开门的时候亮起来。你觉得未来「开发软件」会不会变成一种像 Microsoft Office 那样的技能——一件几乎人人都会的事，而不只是科技行业的人才会？天啊，会的。会，绝对会。我觉得还远不止于此。我觉得它会变成一种技能，就像……怎么说呢，就像我会发短信一样自然。你知道，我读的书基本就两类：科幻和科技史，这是我读得最多的。在科技史里，有一件事在我看来是当下正在发生之事最清晰的对照。那就是 15 世纪欧洲的印刷术。在印刷术出现之前，欧洲大概只有 10% 的人识字，会读会写。他们往往受雇于那些不识字的国王和领主，他们的工作就是读和写，而这并不是人人都会的本事。
[16:14]
[snorts]
（嗤笑）
[16:14] Boris
The printing press was invented, then there were two more presses and in the 50 years after the first printing press, there was more literature published in Europe than in the thousand years before. And over the same period, the cost of literature, the cost of a book went down like a 100x. And then, you know, it took a couple hundred years cuz you know, learning to read and write is hard. You need education systems and government and everyone can't be working on farms and so on. But over the next few hundred years, literacy globally went up to like 70%. And so, you know, now we can all read and write and you don't need a a degree in reading and writing to know how to read and write. Although still there are professional writers and that is a thing that you can do. So, I I think the thing that's about to happen and it's going to be much faster than 50 years is software will be a thing that is fully democratized, that anyone can do. And you know, there's a lot of corollaries to this. So, for example, let's say you're writing accounting software. The best person to write accounting software, I think maybe even today, is not an engineer, it's a really good accountant because they know the domain really well and coding is the easy part. It's knowing the domain that's the hard part. And I I think this is just obviously the the future. So, uh one of the things Greg said was that you guys are living in the future a little bit cuz you get to have access to the models and the agents. Claude code was an internal tool before you released it. Um is the gap between where you guys are in engineering and the rest of the world, is that a month? Is it 3 months? Is it 6 months? And is that is that gap getting bigger or smaller over time? Yeah, so so internally, we use the same models everyone else does. Um for us, the dog fooding is really really important. So, we use the thing that everyone else here does. Um you know, we use like a little bit of mythos to try it and then we use a lot of Opus 4.7 to to dog food it and to write most of our code. Um I think on the model side, there isn't really a gap. Um you know, it's like it's pretty much mythos and you know, that will become some version of some descendant of that will become available at some point to everyone.
印刷机被发明出来后，又造出了第二台、第三台，而在第一台印刷机问世后的 50 年里，欧洲出版的文献比此前一千年加起来还要多。同一时期，文献的成本、一本书的成本下降了大约 100 倍。然后，你知道，又花了几百年——因为学会读写本来就很难，你需要教育体系、需要政府，还得让大家不必都去种地等等。但在接下来的几百年里，全球识字率涨到了 70% 左右。所以你看，如今我们都会读会写，而你并不需要一个读写专业的学位才会读写。当然，专业作家依然存在，那也是一门可以从事的职业。所以我觉得，接下来要发生的事——而且会比 50 年快得多——就是软件会被彻底民主化，变成任何人都能做的事。这里头有很多类比的推论。比如说，假设你要写一套财务软件，写财务软件最合适的人——我觉得也许今天就已经是了——不是工程师，而是一个很厉害的会计，因为他们对这个领域了如指掌，而写代码反倒是容易的部分，难的是懂这个领域。我觉得这显然就是未来。对了，Greg 之前说过一句话，说你们某种程度上是活在未来的，因为你们能提前用上这些模型和 agent。Claude Code 在对外发布之前，本来是个内部工具。那你们在工程上所处的位置，和外界之间的差距，是一个月？三个月？还是六个月？这个差距随着时间是在变大还是变小？是这样，在内部，我们用的模型跟所有人用的是同一批。对我们来说，dogfooding（自己吃自己的狗粮）特别特别重要，所以我们用的就是在座各位用的同一个东西。我们会先用一点 mythos 来试，然后大量地用 Opus 4.7 去 dogfood、去写我们大部分的代码。我觉得在模型这一侧其实没什么差距。差不多就是 mythos，而它——它的某个版本、某个后代——会在某个时间点向所有人开放。
[18:10] Boris
I think on the product side, there's probably a far larger gap. And that's just related to us changing all of our processes. Like if you talk to people at Anthropic, we use Claude for literally everything. And our Claudes are talking all day like as as I'm coding, as my Claudes are coding in a loop, they will communicate over Slack to talk to other people's Claudes that are also running in a loop to kind of figure out unknowns. We have no more manually written code anywhere at the company. All of the SQL is written by uh by models. Everything is just built by the models. So, I I I think actually the place that we're ahead is not the technology cuz the same technology available to us is available to everyone here because fundamentally, we are building a platform. And so, for us, it's really important that developers can use the same thing that we're using and that we we dog food everything that we put out there. But I think there's actually a far bigger weed in kind of the organizational structure and organizational process. And this is a place where you know, hopefully we can talk about it in places like this and uh everyone can kind of learn from it and and also evolve. Yeah, and I think that's one of the advantages startups have. It's so much easier to start there. Jared? Yeah, um last time we talked, I think I think you'd mentioned we talked a little bit about multi-agent and it was very in code at the time at a prior Sequoia event and you mentioned that there were some things going down the pipeline and thing you're talking you're thinking about. Now obviously there's slash batch, there's slash loop, there's sub teams, there's teams. Can you speak some to either at the model level and at the harness level, how you're injecting priors in the harness level, how the objective function is changing the model level to kind of make this experience around delegating work, spinning up agents better? Cuz so much of the work is parallelizable. You can do so many things so much faster and I feel like I have to overlay my own intuition for when to parallelize things rather than the model kind of understanding that you can spin up 10 sub agents for something. Yeah, I mean on on the product side, it really just comes down to prompting.
我觉得在产品这一侧，差距可能要大得多，而这完全是因为我们整个工作流程都变了。如果你去问 Anthropic 的同事，我们现在真的是什么都用 Claude。我们的 Claude 一整天都在「对话」——我在写代码的同时，我的那些 Claude 也在循环里跑着代码，它们会通过 Slack 跟别人手里同样在循环里跑的 Claude 沟通，一起把各种未知的问题搞清楚。公司里已经没有任何一行是人手写的代码了，所有 SQL 都是模型写的，所有东西都是模型搭出来的。所以我觉得，我们真正领先的其实不是技术本身——因为我们手上的技术，在座各位也都能拿到，毕竟我们本质上是在做一个平台。对我们来说很重要的一点，就是开发者能用上跟我们一样的东西，我们要把自己发布出去的所有东西都拿来 dog food（自己先吃自己的狗粮）。但我觉得，真正更大的「坑」其实是在组织结构和组织流程上。这恰恰是个我们可以拿到这种场合来聊一聊的话题，希望大家都能从中学到点东西，也能一起进化。对，我觉得这也是创业公司的一个优势，从头开始要容易得多。Jared？对，上次我们聊的时候，我记得你稍微提到过 multi-agent，那会儿在之前的一次 Sequoia 活动上还都停留在代码层面，你说有些东西已经在路上了、你正在思考。现在显然已经有 slash batch、有 slash loop、有 sub teams、有 teams 了。你能不能讲讲——无论是在模型层面还是在 harness 层面——你们是怎么在 harness 层面注入先验、怎么在模型层面调整目标函数，来把这种「委派任务、批量起 agent」的体验做得更好？因为太多工作其实是可以并行的，你能同时做一大堆事、快非常多，可我现在感觉得靠自己的直觉去判断什么时候该并行，而不是模型自己就明白它可以为某件事起 10 个 sub agent。是这样，我觉得在产品这一侧，归根结底其实就是 prompting（提示词）。
[19:53] Boris
That's That's how it is. And so, you know, we we tweak prompts to kind of help the model do stuff in parallel more. But also, honestly, as the model gets better, it just naturally does this. And so, something like loop, I found actually 4.7, it just starts doing. Uh which is really cool. It's like it does something like uh you know, I'll I'll I'll tell it, "Go uh pull this data query." And it's like, "Hey, I noticed that the data is changing over time. I'll start a loop and I'll give you a report every 30 minutes." And I'm like, "Great. Can you send it to me over Slack?" And then it uses the Slack MCP to do that. So, so I think actually over time, it's not on users to figure out how to hold the tools better. And if that's the case, it's actually a product design problem and like I'm not doing a good job. It's really on the model to do this stuff better and on us kind of prompting it so it naturally does this. Um so, right now it seems like a lot of us use um like Claude or Codex or these uh tools in the cloud to do a lot of our computing. But then, there are some very vocal advocates of uh having your AI be local. And I could imagine over time as um open way models and other things catch up that this could be more of a possibility for people get really high-quality coding assistance. So, I'm curious your vision of say over the next like years or something like that. Do you see the trajectory of everyone still really relying on the like cloud centralized compute or uh is there a pivot to oh, we all just have our local agents that we can rely on and they don't get throttled and other benefits? Yeah, I think it um I don't know. There's maybe a few ways to answer that. I think maybe like kind of the the most fundamental way to answer that is it doesn't matter. Cuz Cuz I think now we're getting to the point where the model is just able to figure it out. So, I think like by a couple years from now, the model is just going to be doing all the code. It's going to be starting the agents. It's going to be building the environments. And so, like if it decides like actually I'll use like local models to do this, then you know, that's what it'll do. These I I don't think these will be decisions that we are making as engineers anymore.
就是这么回事。所以我们会去调 prompt，帮模型更多地并行干活。但说实话，随着模型越来越强，它自然而然就会这么做。像 loop 这种，我发现 4.7 已经会自己主动用了，这点特别酷。比如我会跟它说「去把这份数据查询拉一下」，它就会回我「嘿，我注意到这份数据是会随时间变化的，那我起一个 loop，每 30 分钟给你出一份报告」，我就说「太好了，能通过 Slack 发给我吗？」然后它就用 Slack MCP 把报告发过来了。所以我觉得，长远来看，并不是要靠用户去琢磨怎么把工具用得更顺手。如果真得靠用户去琢磨，那其实就是产品设计出了问题、是我没做好。真正应该是模型自己把这些事做得更好，再加上我们在 prompt 上稍微引导一下，让它自然而然就这么干。那，现在看起来我们很多人是用 Claude、Codex 这类云端工具来做大量计算的，但也有一些声音很高的人主张让你的 AI 跑在本地。我可以想象，随着开放权重模型之类的东西慢慢追上来，本地化对想要拿到高质量编程辅助的人来说会越来越可行。所以我挺好奇你对未来——比如说接下来几年——的设想：你觉得大家还是会一直高度依赖这种云端、集中式算力，还是会转向「我们都用自己本地的 agent，它们不会被限速，还有别的好处」？嗯，我觉得……我也说不准，这个问题可能有好几种答法。我想最根本的一种答法是：这其实不重要。因为我觉得现在我们正走到一个临界点——模型自己就能搞定。所以我估计再过几年，模型会把所有代码都写了，它会去起 agent、会去搭环境。那如果它判断「其实我用本地模型来做这件事」，那它就会那么做。这些已经不再是我们工程师要去拍板的决定了。
[21:51] Boris
We have time for a couple more questions, so I can toss this out. Jamie. Nester. Thank you. It feels like one of the great uh decisions with Claude Code was making use of the fact that a lot of developers' tools and workflows are local. But um that isn't necessarily always the case for sort of general knowledge work with, you know, cloud tools. I'm curious how you're thinking about this with Co-work of how do you give Co-work enough access to the tools that we use to be powerful the same way that Claude Code is for developers? Yeah, it's That's a really great question. Um I know I know when I was uh when I was at a big company, we took like 5 years moving all the environments to remote. It's just like so much work, especially at a big scale. Um but for knowledge work, largely, it's there already with like Salesforce and Docs and things like that. Um for us, it's always just the simplest answer. It's just MCP. So, the same MCP connector that you have in Claude AI, you hook up like, you know, Salesforce, you hook up Google Docs, Google Calendar. Uh and then Co-work can use that. Claude CLI can use it. Claude Code everywhere can use it. And for the for the systems that don't have MCPs, like do you think that's where computer use is going to be a big opportunity? Yeah, I think computer use is kind of a catchall. Um so, I think currently, for as far as I know, I think Anthropic is like pretty far ahead on computers. And so, like if you use it through Co-work, it's quite good. Um so, it's able to use pretty much any piece of software that you have on your computer. It's very slow, but it does it quite well now, especially with 4.7. Um Yeah, but I think I think otherwise like MCP is is kind of the answer. It's And you know, all this stuff just doesn't matter that much. It could be MCPs, APIs, just some sort of programmatic access cuz the the model doesn't care. It's to mo- To the model, it's just tokens. All right, we have time for one more question. Um Ryan. Sean, do you want to toss the Thank you. Um you've kind of alluded to this, but if like sometime ago you saw the probabil- the product overhang and thought to build a product that would then become more interesting once models got better, could you just talk even in vague terms about the shape of a product you'd build today that you think could becomes a much more interesting as models get better in 6 months to a year?
我们还能再问几个问题，那我先抛出去。Jamie。Nester，谢谢。感觉 Claude Code 当初一个很棒的决定，就是充分利用了「很多开发者的工具和工作流本来就是在本地」这一点。但对一般的知识工作来说，情况就未必如此了，那些往往是云端工具。我很好奇你们在 Co-work 上是怎么想的——怎么给 Co-work 足够的工具访问权限，让它能像 Claude Code 之于开发者那样强大？对，这是个特别好的问题。我记得我以前在大公司的时候，我们光是把所有环境迁到远端就花了 5 年，工作量实在太大了，尤其是在那么大的规模上。但对知识工作来说，大体上它已经在云上了，比如 Salesforce、各种文档之类的。对我们来说，答案永远是最简单的那个——就是 MCP。你在 Claude AI 里用的同一个 MCP 连接器，你接上 Salesforce、接上 Google Docs、Google Calendar，然后 Co-work 就能用了，Claude CLI 能用，到处的 Claude Code 也都能用。那对于那些没有 MCP 的系统，你觉得这是不是 computer use 的一个大机会？对，我觉得 computer use 算是个万能兜底。就我所知，目前 Anthropic 在 computer（操作电脑）这块是相当领先的。所以你要是通过 Co-work 来用，它表现挺好的，基本上你电脑上任何一个软件它都能操作。虽然很慢，但现在它做得相当不错，尤其是配上 4.7。不过我觉得，除此之外，MCP 基本就是那个答案。而且说真的，这些东西其实没那么要紧——可以是 MCP、是 API、是某种程序化的访问方式都行，因为模型根本不在乎。对模型来说，这些全都只是 token。好，我们还有时间再问最后一个问题。Ryan。Sean，你要不要把（话筒）抛过去？谢谢。你前面其实已经隐约提到过：如果说在某个时间点，你看到了那种「产品的悬置空间（product overhang）」，于是决定先做一个产品，等模型变强之后它会变得更有意思——你能不能哪怕含糊地讲讲，今天你会去做一个什么形态的产品，是你觉得在 6 个月到一年里随着模型变强会变得有意思得多的？
[24:03] Boris
Yeah, Claude design I I think is a really good example. It's uh it's pretty good today. It's going to get a lot better. Um there's also a few things that we're cooking up for Claude Code uh that are going to be landing over the coming weeks. So, you'll see those. Um and then I think uh I think loop and batch and things like this around like massively parallelizing agents, that's going to get better. And computer use is another good one. All right, Boris. Thank you so much for joining us. I think we'll be here for a little longer if anyone has questions.
嗯，Claude 的设计能力（Claude design）我觉得就是个很好的例子。它今天已经挺不错了，而且还会变好很多。另外我们还有几样东西正在给 Claude Code 捣鼓，接下来几周就会陆续上线，到时候你们就能看到了。还有我觉得 loop、batch 这一类围绕「大规模并行 agent」的东西也会变得更好。computer use 也是其中很不错的一个。好，Boris，非常感谢你来参加。如果有人还有问题，我们应该会再待一会儿。
[24:30]
[applause]
（掌声）
[24:32] Boris
Thanks, guys.
谢谢大家。