AI 会取代软件工程师吗？

有些人认为我们都会因 AI 失业的想法只是炒作。本周，我采访了一位可能让这成为现实的人。本期播客由 Atlassian Ro 赞助，这是一款能将你的团队从 AI 新手转变为 AI 原生的 AI。欢迎来到 Platformer。我是 Casey Newton，本周的嘉宾是 Anthropic 的 Boris Cherny。他是 Claude Code 的创始人和负责人，这是世界上增长最快的 AI 编码工具，也可能成为全自动软件开发的预演。如果本系列前两位关于 AI 与工作的嘉宾试图让我们对风险保持冷静，我猜 Boris 会让我们激动起来。他曾表示软件工程岗位最快将在今年年底开始消失，我真的很期待就此向他追问。

Some people think the idea that we're all going to lose our jobs to AI is just hype. This week, I'm talking to one person who might make it a reality. This podcast is brought to you by Atlassian Ro, the AI that takes your team from AI novice to AI native. Welcome to platformer. I'm Casey Newton and my guest this week is Boris Cherny of Anthropic. He's the creator and head of Claude Code which is the fastest growing AI coding tool in the world and may also be a working preview of fully automated software development. So if our first two guests on this series which is about AI and jobs try to calm us down about the risks, I suspect that Boris is going to rile us up. He has said that software engineering jobs will start to go away as soon as the end of this year and I'm really looking forward to pressing him on that one.

但在此之前，像往常一样，我们将从数据开始。每周，我们都会用最新数据开场，试图帮助大家理解实际情况。为此，我们再次请来了 Platformer 研究员兼 Z 世代 AI 记者 Ella Marianos。Ella，这周你怎么样？

But before that, as always, we're going to begin with the numbers. Each week, we kick off the show with fresh data trying to help us make sense of what is actually happening on the ground. And to do that, once again, we're bringing in platformer fellow and Gen Z AI correspondent, Ella Marianos. Ella, how are you this week?

嗯，我很好。我一直在读《指环王：王者归来》，它有点改变我的人生。呃，这和 AI 无关。

Um, I'm wonderful. I've been reading Lord of the Rings: Return of the King and it's kind of changing my life. Uh, that has that has nothing to do with AI.

嗯，这确实是人类创作的优秀作品。我是说，我总觉得魔戒和 AI 之间有某种相似之处。就像你想要力量，但也许 AI 的力量背后其实隐藏着某种阴险的东西。我们一次又一次地看到这一点。

Um, it's just truly excellent human generated writing. I mean, there is like a parallel between the ring and AI um that I always think about. It's like you want the power and then like maybe there's in fact something like insidious that comes with the power of AI. we see again and again.

确实如此。嗯，你知道，我想说，看到你读你所谓的人类创作的作品，我总是很高兴。但我要提醒你一点，你可能已经想到了，那就是硅谷读过《指环王》的人后来会创办一些有史以来最可怕的公司，名字都取自《指环王》。比如 Palunteer、Anderil。嗯，你有没有看到过某个名字，让你觉得可以用它命名一家公司来做些真正邪恶的事？

It's true. Well, you know, I I will say, you know, I'm always glad to see you reading human generated writing, as you call it. Um, but I I have one caution for you, which you may have already considered, which is that people in Silicon Valley who read The Lord of the Rings do then go on to start some of the most terrifying companies um ever with names selected from Lord of the Rings. So, of course, Palunteer, Anderil. Um, so have you had any inkling yet of like a sort of a name that you've seen and that's made you think I could probably do something really evil with a company named after this?

你知道，比如 Minus Morgle，目前还在秘密阶段。所以也许我们应该从播客中删掉这段，但我有点希望我们能扭转关于索伦相关科技公司名称的叙事。呃，我们将用我们出色的亡灵骑士团队做真正伟大的事情。

You know, like minus Morgle, it's in stealth right now. So like maybe we should cut it out of of the podcast, but I I have some hope that we're going to do some we're going to flip the narrative on Sauron related tech company names. Uh we're going to do really great things with our wonderful team of undead horsemen.

呃，太棒了。我们期待在未来几个月了解更多关于这家公司的信息。嗯，与此同时，我想知道你这周有没有看到任何与 AI 和工作相关的有趣内容。

Uh fantastic. We we look forward to learning more about uh this company in the months ahead. Um in the meantime though, I wonder if you have seen anything interesting this week related to AI and jobs.

是的，微软有一项研究。呃，他们调查了 2 万名 AI 用户。所以值得记住的是，我们缩小到了那些实际在工作中使用 AI 的人群。呃，他们称之为“转型悖论”，基本上就是 65%在工作中使用 AI 的人担心如果不快速采用 AI 就会落后。嗯，而且人们也觉得它有用。比如 58%的人说他们能产出没有 AI 就无法完成的工作。就像一年前他们无法产出这类工作。66%的人说 AI 让他们有更多时间处理高价值任务。所以第一，人们在使用它并且喜欢它。第二，他们担心如果不更多地使用它，就会跟不上。然而，真正形成对比的统计数据是只有 13%的人说他们因在工作中尝试 AI 而获得奖励。所以基本上，人们对 AI 使用的渴望与他们所在机构本身之间存在差距。

Yeah, so there's this study from Microsoft. Uh they've surveyed 20,000 AI users. So, worth keeping in mind, we're like narrowing down to the population of people who are like actually using AI at their jobs. Uh, and there's this thing they're calling the uh transformation paradox, which is basically uh 65% of people who are using AI at their jobs, they're worried that they're they'll fall behind if they don't adopt AI quickly. Um, and also people are finding it useful. Like 58% say they're producing work that they couldn't have if they didn't have the AI they have. Like a year ago they wouldn't be producing this kind of work. 66% say it lets them spend more time on high value tasks. So like one people are using it and they're liking it. Two they're worried that bad stuff will like they won't be able to keep up if they don't use it more. However, the like really big contrasting statistic is only 13% say they're rewarded for experimenting with AI at work. So basically there's this gap between like appetite for AI use and like the institutions themselves that people are working at.

这对我来说非常有趣，我认为它触及了一些非常真实的东西，基于我最近与工作中的人的对话，那就是 AI 全是惩罚没有奖励。你知道，典型的 CEO 在全体会议上说：“我们要立即 AI 化所有事情。你必须时刻使用 AI。”然而，当员工照做时，似乎并没有多少奖励在等着他们。

This is so interesting to me and I think it gets at something very real based on my own conversations with people at their jobs lately which is that AI is like all stick and no carrot. you know, the like the the prototypical CEO is conducting an all hands saying, "We're going to AI all of the things immediately. You must be AIing at all times." And yet, when workers go through with this, it doesn't seem like there's actually much reward waiting for them on the other side.

是的。我认为今年早些时候微软的另一项研究发现，当管理者积极示范 AI 使用时，AI 使用率会上升约 17%。人们对智能体的信任度也提高了 30%。嗯，你是否应该信任智能体？那是另一个问题。但我认为有几种不同的方式可以鼓励员工使用 AI。一种是“AI 是未来，请大家使用 AI”。另一种是“嗨，我的直接下属，这是我工作中使用 AI 的具体方式，你可以尝试类似的方法”。事实证明，这种策略似乎确实有效。

Yeah. And I think another thing that a second Microsoft study found like a little earlier this year was um when managers actively model AI use uh AI use goes up like 17%. And people also trust agents 30% more. Um which like should you be trusting the agents? That's another question. But I think that kind of there are a few different ways you can like uh try to encourage your employees to use AI. One is like AI is the future guys, please use AI. And another is like hello direct reports of mine. Here are like the specific ways like I use AI as they're surveying here at my job and like here's how you could try to do a similar thing. And it turns out like that kind of strategy does seem to get um results.

有道理。我的问题是，如果管理者真的用金钱奖励员工使用 AI，那会是什么样子？你知道，现在很多人不愿意使用 AI，这是可以理解的，因为他们不想训练自己的替代者。他们不想加速自己工作的终结。但我认为如果他们有理相信，嘿，如果你因为这个工具变得更高效，提高了公司的产出，你就会分享收益。嗯，我不知道这个想法在硅谷看来是不是纯粹的共产主义，但我最近基本上没听说有人尝试这样做。

That makes sense. I mean, my question is like, what would this look like if managers were actually financially rewarding their workers for using AI? It's like, you know, right now so many people understandably are reluctant to use AI because they do not want to train their own replacement. They do they do not want to like hasten the end of their own job. But I think if they had reason to believe, hey, if you become more productive because of this tool and you raise the output of this company, like you are going to share in the spoils of this. Um I I don't know if this idea just comes across as like pure communism to Silicon Valley, but like I have not heard basically anyone who seems to be trying this lately.

是的。我的意思是，某种程度上，至少有些工作场所在补贴 token。嗯，有时我想提一下，这结局并不太好。

Yeah. I mean I guess like to some extent there's this thing going on where at least workplaces are like subsidizing tokens. Um which sometimes I would like to bring up doesn't end super well.

那么让我们听听这个。与此同时，我们看到工人需求与管理者奖励工人使用 AI 的程度之间存在差距。呃，我们还看到主要科技公司所谓的“token 最大化”，现在

So let's hear about this. At the same so at the same time as we're seeing this like gap between in fact like worker demand and like how much managers are rewarding their workers for using AI. Uh we're also seeing quote unquote uh token maxing uh at major tech companies where now

你自认为是 token 最大化者吗？我不是在最大化 token。我没有

do you identify as a token maxer? I'm not token maxing. Like I don't have

我不会在晚上运行智能体。当我让 AI 为我写代码时，我基本上坐在那里，知道它在做什么。作为记者，我没有任何任务需要让 Claude 创建一个巨大的代码仓库，让它不断思考之类的。我甚至不知道有什么非破坏性的活动可以做。

I'm not like running agents at night. Like when I get an AI to do code for me, I'm kind of like sitting there like I know what it's doing. I just there just like isn't any task for me as a journalist where I'm like, you know, I need to make have Claude make like an enormous code repo where it's like constantly thinking to itself or something like I don't I don't even know what like nondestructive activity I would do.

我让 Claude 从头重建 Palunteer。嗯，那开始消耗大量 token，我们实际上不得不拔掉插头。

I I told Claude to rebuild Palunteer from first principles. Um and that that that started to burn so many tokens I we actually have to pull the plug on this.

所以感谢你的节制。我赞成 Token 节制。好吧，Token 最小化。我觉得我们不需要那样做。但 Token 最大化就太过分了。但你说硅谷有些公司确实在 Token 最大化。能给我们举个好例子吗？

So thank you for your moderation. I'm in favor of token moderating. Okay. Token minimizing. I don't think we need to do that. But token maxing is a bridge too far. But you're saying that there are companies in Silicon Valley where they are truly token maxing. Do you have a good example for us?

是的。比如亚马逊。之前我们从 Meta 那里得到了一些消息。今年，或者这周，发生了很多事情。这周，一些亚马逊员工向《金融时报》报告说，基本上现在亚马逊采用了新工具 Mesh Claw，这是一个受 OpenClaw 启发的内部亚马逊工具，并鼓励员工使用它，而且在团队内部还有各种 Token 使用量的排行榜。据员工说，有些人只是运行智能体，甚至不做有生产力的事情，就像我们在 Meta 看到的那样，有时只是随机进入循环，这样他们的 Token 使用量就会上升。例如，在 Meta，我们之前看到了一些更大的数字，在最大的 Token 排行榜上，Token 使用量高达数千亿，这个数量显然会让 Meta 损失数百万美元，其中一些确实被浪费了。关于《金融时报》文章中描述的这种动态，我发现的另一个非常有趣的事情是，官方的说法是，你的 Token 使用量不应该是经理们考虑的指标。但员工仍然认为经理会看它。所以实际上，他们仍然只是在增加他们的原始 Token 使用量。而且，亚马逊还有一个高层企业目标，要求 80%的开发人员每周使用 AI。所以这就像来自高层的信号是，我们希望你们使用 AI，以至于人们实际上有时至少在做荒谬的事情。对我来说，我不在那里。我不是亚马逊的经理之一。我不太了解如何管理一个软件工程师团队，但我觉得作为一个开发者，我希望少一些关于我的 AI 使用如何被追踪的双关语，多一些与如何用 AI 增加价值相关的、清晰传达的生产力指标。

Yeah. So Amazon. Previously we got some stuff from Meta. This year, or this week, a lot of stuff has happened. This week, some Amazon employees reported to the Financial Times that basically now that Amazon has adopted this new tool, Mesh Claw, which is an internal Amazon tool inspired by OpenClaw, and is encouraging workers to use it, and also has these leaderboards within teams, like various leaderboards of token usage. Some people, according to employees, are just running agents, not even doing productive stuff, just like we saw at Meta, like maybe sometimes just randomly go in a loop so that their token usage goes up. For example, at Meta, we previously saw a bit more numbers of the biggest token usage in the biggest token leaderboard. It was like hundreds of billions of tokens, and it was an amount that clearly would have cost Meta literally millions of dollars, where some of that truly was going down the drain. Another thing I found really interesting about the dynamic described in this Financial Times article is the official word from on high is that your token usage is not supposed to be a metric your managers take into account. But employees still think that managers look at it. And so in fact, they're still just increasing their raw token usage. And then also, Amazon has this high-up corporate target for 80% of devs to use AI every week. And so it's like the signal from on high is we want you to be using AI to an extent that people are in fact sometimes at least doing absurd stuff. To me, I'm not there. I'm not one of these managers at Amazon. I don't know as much about how you manage a team of software engineers, but I feel like as a dev, I would want less double speak about how my AI use is being tracked, and more productive metrics that are clearly communicated and relate to how I'm adding value with AI.

是的。

Yes.

是的。我认为这个故事很好地呼应了你节目开头提到的微软工作趋势指数，因为如果员工觉得他们不会因为以特定方式使用 AI 而得到奖励，他们可能会以非常愚蠢的方式使用它，对吧？他们会试图遵守意图的字面意思，即使用 AI，但错过了精神，即更好地完成工作。如果他们因为工作中使用 AI 而有强大的经济激励，也许他们会更好地完成工作。也许他们甚至不会浪费那么多 Token。但是，Ella，我知道有一个人可能对最近整个行业看到的 Token 大量燃烧并不那么紧张，那就是 Claude Code 的创造者 Boris Cherny，没有他，我认为 Token 最大化趋势可能不会出现。我们将在休息后请他进来聊聊。谢谢你加入我们，Ella。

Yeah. And I think that this story speaks so well to that Microsoft Work Trend Index that you brought to us at the top of the show, because if workers do not feel like they're going to be rewarded for using AI in a very specific way, they may use it in a very silly way, right? They're going to try to honor the letter of the intent, which is use AI, but miss the spirit of it, which is get better at your job. And maybe they would get better at their job if they had a strong financial incentive to be using it in their work. Maybe they wouldn't even burn so many tokens. But you know, Ella, I know one person who probably is not that stressed out about the major burning of tokens that we're seeing all around the industry lately, and that is Boris Cherny, the creator of Claude Code, and somebody without whom I think the token maxing trend might not be possible. And we're going to bring him in and talk to him right after the break. Thanks for joining us, Ella.

我今天的嘉宾是 Boris Cherny。Boris 是 Claude Code 的创造者和负责人，这是 Anthropic 去年五月推出的智能体式编码工具。简而言之，它是一个热门产品，一个你输入文字就能输出代码的盒子。发布后 8 个月内，Claude Code 负责了推送到 GitHub 的所有代码中约 4%的部分。到今年二月，它的年化收入运行率达到了 25 亿美元，是达到这一里程碑最快的企业产品。其中有趣的一点是它的创造者。Boris 没有计算机科学学位。他学的是经济学，18 岁辍学创业，在 hedge fund 工作过，然后在 Meta 做了五年首席工程师，之后在 2024 年底加入 Anthropic。他甚至不是带着构建编码工具的任务来的。他来学习 API，开始做一个副项目，最初只是把 Claude 连接到 AppleScript，这样它就能看到他正在听什么歌。两个月内，他有了一个版本的 Claude Code，Anthropic 工程团队中有 20%的人在第一天就使用了它。我们这个系列的前两位嘉宾对 AI 自动化有些看跌。Box 的 Aaron Levy 和 Google 的 James Manika 都认为，工作比看起来更难自动化，虽然未来会带来严重干扰，但这不一定意味着大规模失业。我认为 Boris 有着非常不同的视角。作为一名软件工程师，他正在积极努力自动化掉自己的工作。他说他 100%的代码现在都是由 Claude 编写的。他每天通过五个终端标签页并行运行五个 Claude 智能体，提交 20 到 30 个拉取请求。他公开表示，一年内软件工程师这个头衔将开始消失，被更像“构建者”的东西取代。所以，和往常一样，当我谈论 Anthropic 时，重要的是要提前说明我的未婚夫在那里工作。你在听的时候应该考虑到这一点。但这是我无法想象在谈论 AI 和工作的未来时不与 Claude Code 的发明者交谈的一次。所以，以下是我与 Boris Cherny 的对话。Boris Cherny，欢迎来到 Platformer。

My guest today is Boris Cherny. Boris is the creator and head of Claude Code, the agentic coding tool that Anthropic put out in May of last year. In short, it's a hit, a box that you type words into and spits out code. Within 8 months of launch, Claude Code was responsible for about 4% of all code pushed to GitHub. And by February of this year, it had hit an annual revenue run rate of $2.5 billion, the fastest enterprise product ever to hit that mark. Part of what's fascinating about this is who built it. Boris doesn't have a computer science degree. He studied economics, dropped out of college to run a startup at 18, did a stint at a hedge fund, and then spent 5 years as a principal engineer at Meta before joining Anthropic in late 2024. He didn't even arrive there with a mandate to build a coding tool. He showed up to learn the API, started hacking on a side project that initially just hooked Claude up to AppleScript so it could see what song he was listening to. And within two months, he had a version of Claude Code that 20% of Anthropic's engineering team was using on the first day. So our first two guests on this series were somewhat bearish on the idea of AI automation. Aaron Levy from Box and James Manika from Google both made the case that jobs are just harder to automate than they look, and that while the future is going to bring serious disruptions, that doesn't necessarily mean massive job loss. I think Boris just has a very different perspective. As a software engineer, he is actively working to automate away his own job. He says that 100% of his code is now written by Claude. He ships between 20 and 30 pull requests per day by running five Claude agents in parallel across five terminal tabs. And he said publicly that within a year the title software engineer is going to start disappearing, replaced by something that is more like builder. So as always when I talk about Anthropic, it's important to say upfront that my fiancé works there. You should take that into consideration as you listen. But this was one where I just couldn't imagine talking about the future of AI and jobs without talking to the inventor of Claude Code. So with that, here's my conversation with Boris Cherny. Boris Cherny, welcome to Platformer.

谢谢邀请。

Thanks for having me.

所以你是在 2024 年 9 月加入 Anthropic 的，我的理解是没有人让你去构建一个编码产品。你只是想学习 API。那么你能告诉我们 Claude Code 的起源故事吗？因为我读到它控制过你的音乐？

So you joined Anthropic in September 2024, and my understanding is that no one told you to go build a coding product. You were just trying to learn the API. So can you tell us the origin story of Claude Code because I've read that it controlled your music?

是的，所有这些事情都是真的。我加入了一个叫做实验室团队的团队，他们构建了很多很酷的东西。所以我们构建了 Claude Code。就像，我构建了那个。有另一个人构建了 MCP。有人构建了技能，然后另外两个人构建了桌面应用。这基本上就是团队的规模。

Yeah, all of these things are true. I joined this team called the labs team which built a bunch of cool stuff. So we built Claude Code. There was like, I built that. There was a different person that built MCP. There was something that built skills, and then two other people built the desktop app. And that was essentially the size of the team.

那是一个很小的团队。我们大概花了几个月时间就做出了这个功能。我们不知道，因为很多想法都很奇怪。我们完全不知道它们能不能行得通。所以，我想在 Anthropic 的一段时间里，重点一直是同样的事情——总是关于企业、编程和安全。我们隐约知道，在这个旅程的某个节点，我们可能应该做出某种产品。在 Anthropic 早期，我们其实不确定是否要开发产品，但如果我们做产品，那么就需要做与编程相关的东西，因为这有助于我们构建更好的编程模型，让每个人都能使用这些模型，同时也有助于研究安全。做这件事有很多理由。

It was like a tiny team. We just sort of built the feature in the course of a few months. And we didn't know, because a lot of these were kind of weird ideas. We had no idea if they were going to work or not. So, I guess for a while at Anthropic, the focus has been on the same kind of stuff—it's always been about enterprise, coding, and safety. We kind of knew that somewhere in this journey we should probably build some kind of product. Early on in Anthropic, we didn't actually know if we wanted to build products, but if we're building products, then we need to build something coding-related because it helps us build better coding models so that everyone can use those models, and it also helps to study safety. There are a bunch of reasons to do this.

但我们不知道它应该是什么。

We didn't know what it should be though.

所以，当时如果你看编程产品，它们都是 IDE 扩展。那时模型的能力——这是 Sonnet 3.5——还不太好。所以它最多只能做花哨的自动补全。你写一点代码，它就会补全那行代码。模型当时就是那个水平。我们有一种感觉，存在一种模型过剩或产品过剩。这个想法是，你可以构建一个产品，做一些模型完全有能力做的事情，但没有人构建一个产品让模型去做。我告诉你，今天仍然是同样的感觉。仍然有一种强烈的感觉，模型可以做所有这些事情。但没有产品让它去做。所以我们想做一个编程产品。不知道它会是什么。所以我就想学习如何使用 Anthropic API，因为我想，“好吧，我们要做一个产品。我应该学会用 API，这样我才能做产品。”然后我就做了最便宜的东西。那是一个在终端里运行的小东西。是我能构建的东西，这样我就不用构建用户界面或应用了。它很快。我花了两天时间做了这个东西，然后开始把它给别人用，看看他们会不会用，怎么用，纯粹出于好奇。我记得在接下来的几周里，Anthropic 越来越多的人开始用它。首先是我身边坐着的人。然后是外面一层的人。几周后，很多 Anthropic 的人每天都在用。这很奇怪，因为它是一个终端里的小原型——最工程化的产品。很多工程师不想碰终端，但他们用了。

So, at the time, if you look at coding products, they were all IDE extensions. The capability of the model back then—this was Sonnet 3.5—was not very good yet. So the best it could do was like fancy autocomplete. You wrote a little bit of code and it would complete the line of code. That's where the model was at. And we had this feeling that there's this model overhang or product overhang. It's this idea that you could build a product that does something the model is actually totally capable of doing, but no one has built a product that lets the model do that. And I've got to tell you, it's still the same feeling today. There's still this intense feeling that the model can just do all these things. There's no product that lets it do that. So we wanted to build a coding product. Didn't know what it was going to be. So I just wanted to learn how to use the Anthropic API because I was like, "All right, we're going to build a product. I should learn how to use the API so I can build the product." And I just built the cheapest possible thing. It was a little thing that ran in the terminal. It was the thing I could build so I didn't have to build a user interface or an app. It was just fast. I built this thing in a couple days and I started giving it to people to see if they would use it, how they would use it, just out of curiosity. And I remember over the next few weeks, more and more people at Anthropic started using it. First it was just the people that literally sat around me. Then it was the next layer of the onion outside of that. And then a few weeks in, a lot of Anthropic was using this every day. It was weird because it was a little prototype in the terminal—the most engineery possible product. A lot of engineers don't want to touch a terminal, but they did and they used it.

我读到，在最初发布后的 5 天内，一半的工程团队已经在使用它了。我想知道，当这一切发生时，你有没有一个时刻觉得，好吧，软件工程永远改变了，还是你仍然在迭代产品、提交拉取请求？

I've read that within 5 days of the initial release, half of the engineering team was already using it. And I wonder as that was happening, did you have a moment of thinking, okay, like software engineering just changed forever, or are you still sort of iterating on the product and pushing pull requests?

老兄，我全神贯注于把这个东西发布出去。对我来说，一旦有了这个想法，我每晚、每个周末都在做。这是我唯一想的事情，唯一做的事情。那时我开始梦见 Claude code。现在每晚我仍然只梦见这个——下一步该做什么，产品下一步该构建什么。所以我认为现在有机会稍微退后一步，因为很多人都在用它，我们应该从人们的使用方式中学到很多。但很长一段时间里，我们只是专注于构建。我甚至没有机会去想这是什么。

Dude, I was so focused on just shipping this thing. For me, as soon as I got this idea, I spent every night, every weekend. This is the only thing that I thought about, the only thing that I worked on. I started having dreams about Claude code back then. That's still kind of all I dream about every night—just what should we do next, what do we build for the product next. So I think now there's a chance to kind of zoom out a little bit because a lot of people are using it, and we should learn a lot about the way people are using it. But for a long time, we were just so focused on building. I just didn't even have a chance to think about what this is.

有没有一个时刻你确实退后了一步？因为我想象，你梦到它的部分原因是你意识到，说你是偶然发现它可能太轻描淡写了，但似乎确实有一种有点偶然的发现感。难道没有那种时刻，比如，哦天哪，是的，这和我之前捣鼓的其他东西不一样。

Was there a moment when you did sort of do that zooming back? Because I have to imagine part of the reason that you're dreaming about it is like you realize that it might be too minimizing to say that you stumbled across it, but it does seem like there was a sense of somewhat accidental discovery here. Wasn't there that moment of like, oh gosh, yeah, this is different than some of the other things I've hacked on.

是的，有很多惊喜。就像我说的，大体上我们知道我们想做一个编程产品，但没人想到这个编程产品会在终端里。有很多惊喜的时刻。第一个是当 Claude 告诉我我在听什么音乐时。有几个版本，我们实际上有一个我录制的视频演示，我们把它捐给了一个计算机博物馆。这是一个非常奇怪的历史文物。我记得我在 Slack 上发布了这个视频，只有两个人点赞——两个反应——因为没人理解这会是什么。但没错，第一个时刻是我问 Claude 我在听什么音乐，它写了一点代码来打开我的音乐播放器，它用 AppleScript 写了代码，我不懂 AppleScript，我也不会想到写代码来回答这个问题。这太疯狂了。它就这么做了。我当时想，“哇，这太令人惊讶了。”它用我作为工程师不会想到的方式解决了问题。在过去的一年半里，有很多这样的时刻。我刚刚和 Claude 又经历了一次。每次我们发布一个模型，我都会用它做实验，看看这个东西的能力边界在哪里，因为基于模型构建最难的事情之一就是它进步得太快了。你每个月都得重新校准，我相信你知道。我第一次用 Claude 订了一堆航班，通常它还行。这次是第一次完美地完成了。每次我旅行，我都用 Claude 来订。是的，它订了八趟航班、五家酒店。唯一的错误是一家酒店有点超预算，我就告诉它，“好吧，这个可能有点贵。”

Yeah, I mean there's a lot of surprise. Like I said, broadly we knew we wanted to build a coding product, but no one thought this coding product would be in a terminal. There were so many moments of surprise. The first one was when Claude told me what music I'm listening to. There are a couple versions of this, and we actually have a video demo I recorded of this, and we just donated it to a computer museum. It's this very weird historical artifact. It was this video I remember posting on my Slack, and there were like two people that liked it—two reactions—because no one understood that this would be it. But yeah, the first moment was like I asked Claude what music am I listening to, and it wrote a little bit of code to open my music player, and it wrote the code in AppleScript, which I don't know, and I wouldn't have thought to write code to answer that. That's crazy. And it just sort of did it. And I was like, "Wow, this is surprising." It solved the problem in a way I wouldn't have as an engineer. And over the last year and a half, there have been so many moments like that. I just actually had one of these with Claude. Every time we were releasing a model, I kind of experiment with it and see what the frontier of what this thing can do, because that's one of the hardest things about building on a model—it's advancing so fast. You just have to recalibrate every month, as I'm sure you know. And I used Claude for the first time to book a bunch of flights, and usually it works okay. This time it was the first time it worked perfectly. Anytime I travel, I use Claude to book it. And yeah, it booked eight flights, five hotels. The only mistake was one of the hotels was just way over budget, and I just told it, "Okay, this might be a little too pricey."

我觉得大概是每晚五千美元之类的，我当时想，公司希望你住得开心，我就说请重新预订这个，但除此之外，你就工作几个小时就搞定了所有事。太酷了。我每周每月都感到惊喜。

I think it was like 5,000 a night or something and I was like cowork wants you to have a great time when you stay you know and I was like please please rebook this one but then you know otherwise you just worked for like a couple hours and did all this. It was just so cool. I feel the surprise every week every month.

我马上要开始合作工作了，但我觉得现在是个退一步看问题的好时机。从最初发现开始，这个技术迅速在 Anthropic 传播，现在已经成为越来越多工程师的默认工具。我认为这是让软件工程师乃至更多人感受到就业自动化问题的产品之一。在第一期节目中，Erin Levy 和我聊过这个话题，他说他认为工作不会消失，总会有软件无法完成的最后一英里人类工作。你曾公开预测，软件工程师这个头衔可能从今年就开始消失。那么 Erin 错了吗？

So I'm going to get to co-work in a bit but this feels like a moment to zoom out a bit. From the story of initial discovery spreads rapidly through Anthropic and now has become a default tool for a very quickly growing number of engineers and it is one of the products I think that is making this question of jobs automation feel really salient for software engineers but maybe more folks than that. During our first episode, Erin Levy was talking to me about the same subject and he said he didn't see jobs going anywhere, that there's always going to be a kind of last mile of human work that the software can't do. You have publicly predicted that the title software engineer could start to go away as soon as this year. So is Erin wrong about this?

我认为有些东西是对的，也有很多东西我们不知道。

I think there's a bunch of stuff that's true and a bunch of stuff that we don't know.

好的。

Okay.

我的意思是，趋势是指数级的。指数级很难思考。老实说，那些说他们知道的人，其实没人真正知道。我们都在猜测，有些是基于我们所见和历史的合理猜测。

I mean the trends are exponentials. Exponentials are very hard to think in. So honestly, everyone that's saying that they know, no one actually knows. We're all guessing and some of these are educated guesses based on what we're seeing and based on history.

我认为会发生几件事。一是很多公司需要的工程师会减少，因为工程师效率更高了，所以做同样的工作不需要那么多工程师。同时，很多公司会需要更多工程师，因为每个工程师效率更高，公司可以做更多事情，启动更多产品，创建更多业务。你看我们的团队，我们一直受限于优秀工程师的瓶颈。我们在尽可能快地招聘，很多公司和我们的客户也是如此。所以我认为两种情况都会发生，这取决于公司和业务。

I think what's going to happen is a few things. One is there's going to be a lot of companies that need less engineers because engineers are more productive. So you just don't need as many engineers to do the same work. I think at the same time there's going to be a lot of companies that need a lot more engineers because every engineer is more productive. The company can do more things. It can start more products. It can create more businesses. You see this with our team, we are constantly bottlenecked on good engineers. We are hiring as quickly as we can and there's a lot of companies and a lot of our customers are exactly the same. So I think both things are going to happen and it sort of depends on the company and the business.

我认为还有另一件事正在发生，所有角色都在以一种有趣的方式融合，我觉得没人能预料到。我们的经理 Fiona 已经 15 年没写过代码了，她加入 Claude Code 后现在开始编程。我们的产品经理 Cat 编程，设计师 Megan 也编程。团队里的每个人都在编程。你不再需要是工程师了。所以这让我觉得，随着时间的推移，如果你把这个趋势投射出去，会发生的是，每个不是工程师的人都会多写一点代码。像我这样的工程师，我已经 6 个月没写代码了，我整天都在构建东西，但已经超过 6 个月没写过一行代码了。所以我看到这一切都融合成了一件事。我们可以称之为建造者，可以继续叫工程师，也可以叫产品经理。我不知道头衔是什么，但角色在改变。所以我们对这些角色的理解肯定会改变，但这对于哪些公司有多少工作机会来说仍然很不清楚。

I think there's this other thing happening where all the roles are kind of blending together in an interesting way that I don't think anyone would have predicted. Our manager Fiona has not coded in 15 years and she joined Claude Code and now she's coding. Cat our product manager codes, Megan our designer codes. Everyone on the team codes. You don't have to be an engineer anymore. So this makes me think that over time, if you project this trend, what's going to happen is everyone that's not an engineer is going to code a little bit more. Engineers like me, I haven't coded in 6 months, I'm building stuff all day but I haven't written a line of code in over 6 months. So I see it all blending into one thing. We can call it a builder, we can keep calling it an engineer, we can call it a product manager. I don't know what the title is but the role is changing. So the way we conceive of these roles is definitely going to change, but what that means for how many jobs are available at which companies is still very unclear.

是的。我认为历史上有许多不同方面的例子。比如拖拉机的发明。我前几天正好读到这个。拖拉机是 19 世纪 90 年代发明的，一个叫 John Frick 的人在爱荷华州发明的，大概是那样。

Yeah. And I think history has a lot of examples in different ways. Like the tractor was invented. I was reading about this the other day. Tractors were invented in the 1890s. It was this guy John Frick invented it in Iowa or something. That sounds right.

当时，如果你看农活的方式，全是马力驱动的。你需要马来干农活。尽管拖拉机在 19 世纪 90 年代就发明了，但直到美国 60 年代，拖拉机的数量才超过马。这花了大约 70 年。如果你看趋势，拖拉机的数量上升，马的数量下降。两者都发生了。交叉点在 60 年代。这有很多原因。拖拉机的技术很神奇，可以让你收获更多作物，生产力更高。但与此同时，如果你是一个农民，想学用拖拉机，你需要培训。一开始，拖拉机很贵。所以在很多情况下，你仍然想要马，因为它更便宜。而且拖拉机一开始不太好用。也许你可以用它收小麦，但可能不能收玉米。所以实际上花了很长时间才有人造出能用于玉米和秋葵等所有作物的拖拉机。这需要一段时间来摸索。我认为我们现在看到的是这个过程的加速版，但本质上是同样的事情。我们遇到了非常相似的问题。

And at the time, if you look at the way farmwork worked, it was all horsepowered. You needed horses to do farm work. And even though tractors were invented in the 1890s, it wasn't until the 60s in the US that there were more tractors than horses. It took like 70 years. If you look at the trend, the number of tractors went up, the number of horses went down. Both things happened. And the intersection was in the '60s. There were a bunch of reasons for this. The technology of tractors was magical. It could make it so you can harvest a lot more crops. Your productivity was a lot higher. But at the same time, if you're a farmer and you want to learn how to use a tractor, you need training. And at the beginning, the tractors were expensive. So in a lot of cases, you still wanted horses because it was still cheaper. And they were not very good at first. Maybe you could use it for wheat, but maybe not for corn. So it actually took a long time for someone to make a tractor that would work for corn and for okra and all the stuff you're using this machinery for. And that just took a while to figure out. I think the thing we're seeing right now is this on a speedrun, but it's sort of the same thing. We're hitting very similar issues.

完全同意。这就是那种把 AI 视为普通技术的论点，即使实验室推出极其强大的模型，人们或组织往往变化缓慢。所以这些技术渗透到公司需要时间。同时，我认为人们看到关于 Anthropic 收入的报道，会说这次似乎没那么慢。所以我认为我们仍在试图确定实际的变化速度。

Totally. I mean this is the kind of AI as normal technology argument that even as labs come up with incredibly capable models, people tend to be slow to change or organizations are slow to change. So it can take time for these technologies to filter through companies. At the same time, I think people look at what has been reported about Anthropic's revenue and they say it doesn't seem like it's taking that long this time around. So I think we're still trying to hone in on what is the actual rate of change here.

是的。好的。那么问你一个问题：电脑让你更高效吗？

Yeah. Okay. So here's a question for you. Do computers make you more productive?

是的，它们让我更高效。但让我更高效和因为电脑而工作更少，感觉是不同的问题，你明白吗？

Yes. Yes, they make me more productive. But does making me more productive feel like a different question than do I work less because of computers? If that makes sense.

所以因为你能做更多事，你就做更多事。你可以在同样的八小时里塞进更多事情。

So because you can do more stuff, you do more stuff. You can fit more things in the same eight hours or whatever.

绝对如此。坦白说，我以前每周录一期播客，外加写几份新闻简报。现在在这个迷你系列中，我尝试每周做几期播客，同时写多份新闻简报。AI 是我能做到这一点的原因。它是一个不可思议的研究助手和播客制作人。所以我能够产出更多，但我不觉得自己工作得更少。顺便说一句，这不是抱怨。

Absolutely. To be candid, I used to record one podcast episode a week in addition to writing a couple of newsletters. I'm now experimenting during this miniseries with doing a couple podcasts a week in addition to writing multiple newsletters. And AI is a reason I can do that. It is an incredible research assistant and podcast producer. So I'm able to produce more, but I don't feel like I'm working less. And that's not a complaint, by the way.

这大概就是我如何应对这个时刻的方式。

That's just sort of like how I'm navigating this moment.

是的，没错。我也有同感。我觉得自己能做更多事情，以前因为时间不够而没做成的事，现在都能做了。但还有另一个奇怪的历史现象：90 年代，当公司开始采用个人电脑时——在大型机之后，在那些价值数百万美元的工业计算机之后，它们被小型化了。普通初创公司、普通公司都能买得起电脑了。当时有个问题：电脑是否让你更有效率？实际上人们抱怨的是它们没有。在 90 年代，这是个开放性问题：电脑真的能提高生产力吗？现在我们回头看，这还用说吗，当然能。我无法想象回到纸笔时代。

Yeah. Yeah. That's right. That's right. And I feel the same way. Like I feel like I can do so much more and all the stuff I didn't get to before because I didn't have enough hours in the day, now I can do. But there's this other weird historical thing where in the '90s when computers were being adopted by companies, like the personal computer after the mainframe, after these big industrial computers that cost millions of dollars, at some point they got miniaturized. So the average startup, the average company could just get computers. There was this question of: are computers making you more productive? And actually what people were complaining about is that they're not. In the '90s, this was an open question: do computers actually make you more productive? And now we look back on it, it's like duh, of course they do. I can imagine going back to pen and paper.

但有一篇很棒的《哈佛商业评论》文章，我记得是 1992 年或 1996 年左右。它研究了一批采用电脑的公司，发现有些公司效率提高了，有些没有。区别是什么？他们发现，效率提高的公司扔掉了所有纸质文件柜、纸笔和抽屉，用电脑作为一切的中心。而其他公司仍然有团队手写所有东西，电脑只是放在角落里偶尔用用。所以第一类公司获得了巨大的生产力提升，第二类没有。我认为现在情况类似，因为在 Anthropic，我们完全围绕 Claude 组织一切。新员工加入时，如果问如何写代码或如何贡献代码库，答案是问 Claude。如果问如何报销费用，也是问 Claude。如果问下一个公司假期是什么，还是问 Claude。所以以前需要手动做、需要找人的事情，现在只需问 Claude。它是一切的核心。我认为很多真正理解的公司也开始这样做：他们把 Claude 放在中心位置。但它不是边缘的东西，你必须改变所有业务流程，这需要时间，也需要很多改变来摸索。

But there's this really awesome Harvard Business Review article, I think it was like 1992 or 1996 or something. And essentially the case it was making is they studied a bunch of companies that were adopting computers and they were like, okay, these ones are getting more productive, these ones are not. What's the difference? And what they found is the ones that are getting more productive are the ones that threw away all their paper filing cabinets, they threw away all their paper and pens and all their desk drawers and stuff, and they just replaced it with a computer at the center of everything. And then there's all these other companies that were still like, you know, they have teams of people writing everything by hand with pen and paper, and then there's a computer in the corner that's used for something. So the first category has big productivity gains, the second category does not. And I think it's kind of similar right now because at Anthropic we really organize everything around Claude in every way. When people join the company, if they have questions about how do I write code or how do I contribute to this codebase, the answer is you ask Claude. If your question is how do I file an expense receipt, it's you ask Claude. If the question is what's the next company holiday, you ask Claude. So it's sort of like all the stuff that you used to have to do manually, used to have to go to someone, you just ask Claude. It's just at the center of everything. And I think this is what we're starting to see with a lot of companies that are really getting it: they just put Claude exactly at the center. But it's not like a thing on the outskirts somewhere. It's like you have to change all the business processes and that takes time and it's a lot of change to figure out.

完全同意。我最近读到所谓的索洛悖论，我想你刚才基本提到了。这是 80 年代一位经济学家的观察：他说计算机时代无处不在，除了在生产力统计数据中。原因是尽管当时计算机大规模部署，但人们并没有看到生产力大幅提升。正如你所说，鲍里斯，最终这些收益确实实现了，因为公司围绕新技术重新设计了工作流程。现在的问题是，经济需要多久才能做到这一点。我想再问你几个关于精细软件工程的问题，因为我听你说过编码实际上已经解决了，你六个月没写过代码了。有时我看到工程师反驳这个观点，他们说编码不仅仅是打字，还涉及判断力、品味和批判性思维，而智能体在这些方面仍然很差。你怎么看待这种批评？编码的某些部分是否仍未解决，或者在你看来，它已经变成了别的东西？

Absolutely. You know, I've been reading recently about what they call Solow's paradox, which I think you basically just referred to, but it's this observation by an economist in the '80s that he said the computer age is everywhere except for in the productivity statistics. And the reason was despite what at the time felt like a very large buildout of computers, you were not seeing people get much more productive. As you just noted Boris, eventually those gains did materialize and it was because the companies had just reinvented their workflows around the new technology. So the question now is how quickly might it take the economy to do that. I wanted to ask you a couple more questions on sort of fine grain software engineering because I've heard you say that coding is effectively solved, you haven't written any code in 6 months. Sometimes I see engineers pushing back on this idea and they say, look, coding is not only about typing, it's also about judgment and taste and critical thinking and agents can still be quite bad at those. So what do you make of that critique? Like are there parts of coding that remain unsolved or is that in your view that's just become something else?

是的，这个批评完全正确。我认为这是那种容易被断章取义的事情。所以完整的说法是：编码对于我所做的那种编码来说已经解决了。对我来说，我处理的是相当简单的代码库，比如云 CLI，CLI 是一个相当新的代码库，还有桌面应用和移动应用。这些都是很小、很简单的代码库。我们有很多企业客户，现在很多客户都是最大的企业，不再只是初创公司和独立开发者了。比如 NASA 是我们的客户之一。所以他们的代码库非常大、非常复杂，对他们来说还没有解决。模型仍然不完美，仍然会犯错，代码也不总是完美的。而且当你考虑工程师做的事情时，编码只占一小部分。以前，我的一天可能 50%的时间在写代码，但另外 50%是跟用户交流、头脑风暴想点子、调试并思考工作原理、做计划。工程师做的所有这些其他事情都不是编码。所以当我说编码已经解决时，是指我所做的那种编码。而编码只是工程师工作的一小部分。实际上，在 Anthropic 的所有工程师身上都能看到这一点，我认为行业里越来越多的工程师也是如此。当模型完成编码时，他们就能腾出手来做其他他们更享受的事情，比如跟用户交流、规划下一步。

Yeah, I mean the critique is totally right. I think this is one of those things that just gets kind of taken out of context. So the full quote is coding is solved for the kinds of coding that I do. For me, I work on pretty simple code bases, like cloud CLI, the CLI is a pretty new codebase, the desktop app and the mobile app. These are pretty small simple code bases. We have so many enterprise customers, like a lot of our customers now are the biggest enterprises. It's not just startups and indie devs anymore. It's like NASA is one of our customers. So they have really big, really complicated code bases and for them it's not solved yet. The model is still not perfect at it. It still makes mistakes. Its code isn't always perfect. And when you think about the kinds of stuff that engineers do, coding is a small percent of it. It used to be that if you look at my day, maybe 50% of my day used to be actually typing code, but the other 50% was like talking to users, brainstorming and coming up with ideas, debugging and thinking through how something works, planning. There's just all these other things that engineers do that are not coding. So when I say coding is solved, it's solved for the kinds of coding that I do. And coding is just a small subset of what engineers do. And you actually see this for all the engineers at Anthropic, and I think more and more engineers in the industry. When the model does the coding, they're freed up to do all this other stuff that they actually enjoy doing a lot more, like talking to users and figuring out what's next.

而且你知道吗，Claude Code 已经 100%由 Claude Code 自己编写超过 6 个月了。

And you know, Claude Code has been just 100% written by Claude Code for over 6 months.

这在 Anthropic 的很多事情上都是真的。除了 Claude Code，co-work 和其他很多产品也是如此。我们开始听到越来越多的客户也这样。比如我最近为最新一期 Y Combinator 的团队做演讲时。

That's true for a lot of things at Anthropic. Beyond Claude Code, that's true for co-work, it's true for a lot of other products. And we're starting to hear more and more customers like this. Like I was doing the talk for the latest Y Combinator batch.

这就像本周早些时候，我们做了一次炉边谈话。我以前每次演讲开头都会问：用过 Claude Code 的请举手。现在所有人都用 Claude Code，所以我就不问那个问题了。于是我问的是：100% 的代码都由 Claude Code 写的请举手。你知道，这是最前沿的初创公司，但都是小公司，通常只有几个人。一半的人举手了。然后我说：好，一行代码都不是模型写的请举手。只有一个人举手。房间里可是有几百人。其他人都在中间，介于 50% 到 100% 之间。所以我觉得，我们写的代码中，越来越大的比例正在被解决。我们的团队是工程领域变化的早期指标，而工程又是工程之外一切变化的早期指标。所以我们开始看到这种转变，它从六个月前开始，而且正在加速。当问到是否 100% 时，举手的人越来越多。

This was like earlier this week we did a fireside and I used to ask at the beginning of every talk I do: raise your hand if you use Claude Code. Now everyone uses Claude Code, so I stopped asking that question. And so instead the question that I asked is: raise your hand if 100% of your code is written by Claude Code. And you know, this is like the latest, you know, the most cutting-edge startups, but they're all small startups, usually it's like a few people. Half the hands went up. And then I was like: okay, raise your hand if none of your code is written by the model. And there was one hand that went up. This is out of a room of a couple hundred people. And then everyone else is somewhere in between, they were like between 50 and 100%. So I think like coding is starting to get solved for a bigger and bigger percent of the code that we write. You know, like our team is an early indicator of what's happening in engineering. Engineering is an early indicator of what happens to everything outside of engineering. And so we're starting to kind of see the shift, and it started 6 months ago, and yeah, it's like it's accelerating, and we're starting to see more and more hands go up when they ask like, you know, is it 100%.

让我问另一个人们担心的点：在一个工程师不写那么多代码的世界里，人们对自己专业的理解会退化，这可能在很多方面都很危险。你已经六个月没写代码了。你觉得自己开始退化了吗？你怎么看？

Let me ask about another fear that people have about a world where the engineers aren't writing as much code. The fear is that people's understanding of their own profession will atrophy and that might be dangerous in various ways. You haven't written any code in 6 months. Do you feel like that atrophy has started with you and how do you feel about it?

团队里有一位工程师，Lena，她周末还会手写 C++ 代码，因为她仍然享受写代码。我觉得这总是有空间的。对我来说，这是更广泛转型的一部分，根本不是退化。编程一直在变化。我爷爷 70 年前在苏联用打孔卡编程，对他来说那就是编程。没有 JavaScript，没有 Python，那些还不存在。对他来说，就是打孔卡：一张纸，有台机器在上面打孔，然后喂进大型机，处理一下，亮几个灯。那就是编程。再之前，阿波罗计划时，满屋子的人，通常是女性，在纸上手算数学，那也叫编程，对吧？现在又变了。编程从写机器码变成写汇编，再变成 JavaScript、Python、Java 这些语言。现在又在变，你直接跟智能体对话，而且很快会再变一次：你跟一个智能体说话，它再跟其他智能体说话来完成编码。但编程一直这样变。对我来说这不像是退化，更像是技术上的巨变。

There's a, you know, there's one engineer on the team, Lena, that was still writing like C++ on the weekends by hand just for fun because she still enjoys writing the code. And I think there's always room for this. I think for me this is part of a much broader transition and it's not about atrophy at all. It's just about programming is always a thing that is in flux. Like my grandpa programmed in punch cards back in the Soviet Union, like 70 years ago, and for him that was programming. There was no JavaScript, there was no Python, that didn't exist yet. For him, it was punch cards. It was like a piece of paper, there's a machine that punches holes in it, then you feed it into this mainframe, it processes it, a few lights up. And when you talk about programming, that's what it was. And then, before that, like the Apollo program, it was like there was a room full of people, often women, doing math on paper sometimes by hand. That was called programming, right? And nowadays, this changed. Programming became writing machine code, then it became writing assembly code, then it became like JavaScript and Python, Java, all these languages that people use nowadays. And now it's changing again. It's now you talk to the agent, and it's actually about to change one more time where you talk to an agent that talks to agents that does the coding. But you know, it's just always changed like this. It doesn't feel like atrophy to me. It feels like a sea change in the technology.

我的感觉是，用图形计算器肯定让我的数学能力退化了，但我的解决办法就是继续用计算器。我挺愿意放弃那些东西的。但如果有一天计算器变得超级智能，试图用微妙的方式暗中破坏我，那会让我吓一跳。不过我们可能还没到那一步。

My feeling about it has been that I'm sure that using a graphing calculator caused some of my math skills to atrophy, but my solution to that is that I will just continue to use a calculator. I'm sort of fine to cede some of that stuff. Now, if over time the calculator becomes super intelligent and tries to undermine me in subtle ways, that would sort of freak me out. But maybe we haven't crossed that bridge quite yet.

让我问另一个相关的批评：每次新模型发布，人们会说‘这真好’。但几周后上 Reddit 一看，又说产品大幅退步了。我觉得有时是 bug 导致的真问题，有时只是一种感觉。但人们担心，因为现在代码都是 AI 生成的，可能没有以前那种工艺感了。你怎么看这种周期性的反弹？

Let me ask about another criticism that I sort of feel like is in this realm, which is it seems like every time a new model is released, we'll hear people say, 'This is really good.' And then you check Reddit a few weeks later and they say the product has massively regressed. My sense is that sometimes this is like a real issue caused by bugs. Other times it's just sort of a vibe. But I feel like people are concerned that because it's all just sort of AI generated right now, there isn't maybe the same craftsmanship that we once saw in code. So, I'm just curious what you make of these periodic backlashes we seem to see.

嗯，原因其实还是个开放问题。有几次是真的，我知道两次，我们在 Anthropic 博客上发了工程博文，因为如果是真的，我们发现了、修复了，然后想让大家了解到底发生了什么。但我觉得其他几乎所有情况都像是蜜月期：一开始觉得神奇，后来就习惯了。可能就是这样。但我不觉得这跟工艺有关，因为现在模型的代码比我写的要好得多。一年前你问我，我不会这么说。我会说模型很马虎，代码不怎么样，得反复检查，它总犯低级错误。但现在不是这样了。模型一直在变，这很奇怪，因为我们用的其他技术都没变得这么快。如果你上次用模型是一年前，现在的模型完全不同了。一年前你得手把手教它，逐行检查；现在我就让 Claude 自己干。我让它复查结果，让它打开应用自己测试。同时我还有十五个其他 Claude 在跑，也做着类似的事。现在就是这样。代码实际上比我写的好得多。

Yeah. It's actually sort of an open question what causes it. There have been a couple instances where it was real and there were two that I know of, and we published engineering blog posts on the Anthropic blog about it because we, you know, if it's real, we found it, we fixed it, and then we want to talk about it so people kind of understand exactly what happened. But I think in almost every other case it's sort of like maybe it's a honeymoon period where you kind of get used to the model. At first it's magical, and then you kind of get used to it. Maybe it's something like that. But I don't think it's really about craft because the model's code at this point is just much better than the code I would have written. If you talked to me like a year ago, I would not have said that. I would have said the model is kind of sloppy and the code's not really good. You have to triple check everything. It can make silly mistakes all the time. But that's just simply not the case anymore. And again, it's just the model keeps changing. It's really weird because every other technology we use does not change this fast. But if you tried the model for the last time a year ago, the model now is completely different. And so if a year ago you had to handhold it and triple check every line, now I just generally let Claude do its thing. I ask Claude to double check the result. I ask Claude to open the app and test it by itself. And then while I do that, I have like 15 other Claudes running that I've also asked to do stuff like this. But that's kind of what it is now. The code is actually just much better than what I would have written.

那么让我问问你的 Claude 形态。你之前问我电脑是否让我更高效。显然 Claude 让你更高效了，但似乎并没有减少你的工作量。我觉得这很重要，如果我们想知道 AI 对工作意味着什么的话。因为听起来你相信公司需要的工程师会更少，但至少对你来说，你永远有做不完的事。

Well let me ask you about your Claude form then. You asked me earlier whether computers make me more productive. I think it seems clear that Claude is making you more productive but it doesn't seem like it's actually reducing the amount of work that you're doing. And I think this is kind of an important thing to dig into if we're curious about what AI means for jobs, because you know it sounds like you believe companies are going to need fewer engineers and yet at least for you, you're never running out of things to do.

那么你怎么看待这个问题：它让我效率高了很多，但我并没有减少工作时间。有没有可能有一天，效率提高真的意味着我工作得更少？

So like how do you think about that question of it's making me so much more productive I'm not working any less. Will there ever be a case where making me more productive actually means I'm working less?

是的。有个名字描述这个悖论，我忘了叫什么，但有人命名过。我觉得这其实非常个人化。有些方面取决于公司，因为根据业务不同，对人的需求可能增加或减少。但我认为很大程度上是个人偏好。当洗衣机问世时——我举个历史例子，因为对我来说，这是如此疯狂的技术变革，我需要历史来让自己站稳脚跟。

Yeah. I mean, there's a name for that paradox. I forget what it was, but someone named it. I think it's actually really individual. There are some parts where it's up to the company, because depending on the business, there might be more need for people or less. But I think a lot of it is individual preference. When the laundry machine was released—I'll give a historical example because for me, this is such a crazy technological change that I need history to anchor myself.

不，我喜欢。我喜欢这些故事。

No I love it. I love the stories.

好的。所以，当洗衣机问世时，普通人洗一桶衣服大约需要五六个小时，每桶要走大约 3000 英尺，因为你得走到外面，收集木柴和煤，回到屋里，生火，烧水，把水倒出来，放进衣服，在搓衣板上搓，然后拧干，可能每天都要为全家人重复这个过程。工作量很大。后来洗衣机出现了，每桶衣服节省了大约三个小时。这是让女性大规模进入劳动力市场的因素之一。通常是家庭主妇做这些工作，她们被困在家里。现在每天多出了三个小时。不同的人可以选择如何利用这段时间。对一些人来说，选择是陪孩子、遛狗、读书或和朋友聚会。但对很多人来说，答案是进入劳动力市场——去工厂或办公室工作。因为时间被解放了，你有了选择。我认为现在任何技术都是类似的。它给你更多选择。

Okay. So, when the laundry machine was released, the average person to do a load of laundry took about five or six hours and you had to walk an estimated 3,000 feet per load because you had to walk outside, collect logs and coals, go back inside, start a fire, boil water, take it out, put in the laundry, scrub it on the scrub board, then ring it out, and repeat for your entire family maybe every day. It was a lot of work. At some point the laundry machine appeared and it took down the time by about three hours per load. This was one of the factors that let women enter the workforce in mass. Usually it was the women of the house doing this work, and they were stuck at home. Now three hours were freed up every day. Different people could choose how to spend this time. For some, the choice was to hang out with kids, walk the dog, read a book, or hang out with friends. But for a lot of people, the answer was to enter the workforce—work at a factory or an office job. Because the time was freed up, you had a choice. I think it's similar now for any technology. It gives you more choice.

最后几个关于软件工程的问题。我问过所有嘉宾。如果一个 22 岁的年轻人这个月刚拿到计算机科学学位，走到你面前说：‘好了，现在怎么办？’你会对他们说什么？有入门级工作在等着他们吗，还是他们需要重新思考职业生涯的初期阶段？

A couple of last questions about software engineering. I've been asking all of our guests. If a 22-year-old just finished their CS degree this month and came up to you and said, 'Okay, now what?' What do you say to them? Is there an entry-level job waiting for them or do they need to think differently about the first part of their career?

我的建议是：如果你想在大公司或某家公司工作，你完全可以这么做。有入门级工作。有很多事情可以做。但如果你有点创业精神，那就去创办一家初创公司。历史上从来没有比现在更好的时机去创业。这绝对是黄金时代。你和你的智能体可以建立一家大公司。人们正在用几个人建立价值数十亿美元的公司。Claude Code 最初也只有我们几个人，我们有这么多客户，只用一两个人或三个人就建立了非常大的企业和非常棒的初创公司。一个有正确想法的人拥有如此大的杠杆。我想象不出更好的时机去做了。

My advice would be: if you're a person that wants to work at a big company or a company, you can totally still do this. There are entry-level jobs. There's a lot you can do. But actually, if you're at all entrepreneurial, go start a startup. There's never been a better time in history to start a startup. It's absolutely the golden age. You and your agents can build a giant company. People are building billion-dollar companies with just a few people. For Claude Code originally it was just a few of us, and we have so many customers building really big businesses and amazing startups with just one or two or three people. One person with the right idea has so much leverage. I couldn't imagine a better time to go and do it.

这很有趣，因为我觉得 AI 世界的很多观点是，模型能力进步如此之快，五年后我们还会不会有公司？但你认为至少在未来一段时间内，仍然有足够的空间去创办公司、进入商业、制造产品等等。

That's interesting because I feel like a lot of the view from the AI world is that model capabilities are advancing so quickly that will we even have companies in five years? But you think at least for the next bit, there's still plenty of room to start a company, get into business, make a product, all the rest.

至少未来几年是这样。如果你真的沿着指数曲线追踪，它会变得非常奇怪。有一种版本是，工作的概念不再有意义，公司的概念不再有意义，软件的概念不再有意义。但当你追踪指数时，它只会变得奇怪。但与此同时，有太多事情要做。我们都还在摸索这个模型意味着什么，这个东西能做什么。所以不如成为探索前沿的人之一。

At least for the next few years. If you really trace out the exponential, it gets really weird. There's a version where the idea of jobs doesn't make sense anymore, the idea of companies doesn't make sense, the idea of software doesn't make sense. But as you trace the exponential, it just gets weird. But in the meantime, there's so much to do. We're all just here figuring out what the model means and what this thing can do. So might as well be one of the people exploring the frontier.

好的。最后一个关于工程的问题。如果三年后，你认为我们会看到更多的工程师、更少的工程师，还是无法回答，因为我们可能不再称他们为工程师了？

Yeah. All right. Last one on engineering. If 3 years from now, do you think we will see more engineers, fewer engineers, or will it be impossible to answer because we just might not be calling them engineers anymore?

好吧，我们来定义一下。我认为我们不会称他们为工程师，但如果我们谈论写代码的人或使用智能体的人——比如用 Claude 写代码的人——我认为会有比现在多一百倍的工程师。这是我的预测。

Okay, let's define it. I don't think we're going to call them engineers, but if we talk about people writing code or using agents—like people using Claude to write code—I think there will be a hundred times more engineers than there are today. That's my prediction.

哇。好的。非常有趣。嗯，这似乎是个好时机回到你之前提到的 Claude Co-work，我知道你帮助开发了它。我得说这是我现在用得最多的 Anthropic 产品。我把它当作专栏的编辑，用它帮我制作播客，还把它当作财务顾问。基本上，你只需创建产品，添加一些技能，它就能成功模仿工作场所中的许多不同角色。另外，作为一个非技术人员，我发现用户界面非常直观，因为主要就是拖放文档到框里。跟我聊聊 Co-work 的未来之路吧。我特别好奇，你是否认为它能像编程现在部分被解决那样，解决其他工作？

Wow. Okay. All right. Super interesting. Well, this seems like a good time to return to Claude Co-work, which you brought up earlier, and I know that you helped to develop. I would say this is the Anthropic product that I actually use the most now. I use it as a kind of editor on my columns. I use it to help me produce a podcast. I've been using it as a kind of financial adviser. Basically, you just create products, add some skills, and it can successfully imitate lots of different roles in a workplace. Also, as a non-technical person, I just find the UI very intuitive to use, because it mostly just involves dropping documents into a box. Talk to me a little bit about the road ahead for Co-work. And in particular, I'm curious whether you think it can solve for other jobs the way that coding is maybe now partially solved.

Co-work 太令人兴奋了，因为我们最初开始构建它时，看到人们把 Claude Code 滥用于非编程的事情。比如有人安装终端，打开它，在终端里安装 Claude Code 来做税务申报。这太疯狂了。终端不是干这个的，但从产品角度看这很棒，因为人们真的需要这个。

Co-work is just so exciting because we first started to build it when we saw people abusing Claude Code for things that are not coding. Like someone installing the terminal, opening it up, installing Claude Code in the terminal so they could do their tax returns. That's crazy. It's not what a terminal is for, but it's amazing from a product point of view because people really want this.

是啊。

Yeah.

我认为接下来几个月的旅程是弄清楚如何让这对非工程师人群也能很好用。这对我们来说其实挺新的，因为我们团队大部分人在做编程，现在有一部分人在做 co-work，我们正在摸索。对于编程来说很容易，因为构建它的人就是工程师，我们为自己构建，然后对其他人也很有用。对于 co-work，它对非工程的一切都有用：会计、财务、法律。我用它买了蛤蜊捕捞许可证，这样我就能和华盛顿州一起去挖蛤蜊了。我用它订机票和演唱会门票。它做所有这些不同的事情。那么如何确保它在所有这些方面都真正出色呢？最重要的事情是我们每天每时每刻都在用它，我们每天每时每刻都在和客户交流，然后我们做客户要求的事情。所以我预计它会在这方面越来越好，也会在长时间运行方面越来越好。

I think the journey over the next few months will be figuring out how to make this work really well for people who are not engineers. This is actually kind of new for us because most of our team works on coding, and now part of the team works on co-work, and we're trying to figure this out. For coding, it's pretty easy because the people building it are engineers, so we just build for ourselves and it's really useful for everyone else. With co-work, it's useful for everything that's not engineering: accounting, finance, legal. I used it to buy a clamming license so I could go clamming with the state of Washington. I used it to book flights and concert tickets. It does all these different things. So how do we make sure it's really good at all of them? The biggest thing is we use it all day every day, we talk to customers all day every day, and we do the things that customers are asking for. So I expect it to keep getting better at this stuff, and to keep getting better at running for long periods of time.

这是我们一年半前在编程领域看到的情况：Claude Code 可能只能运行 30 秒而我不需要打断它，因为模型当时还不行。30 秒后它就会偏离轨道。而现在我让 Claude 运行好几个小时。每天晚上我有几百甚至几千个智能体在运行 5、10、20 个小时。这就是现在工程的做法。我认为同样的事情也会发生在 co-work 上。我不知道什么样的工作需要它运行那么久，但我觉得我们会找到答案，因为人们会开始破门而入要求它。

And this is something we saw for coding a year and a half ago: Claude Code could run maybe 30 seconds without me having to interrupt it because the model just wasn't there yet. It would go off the rails after 30 seconds. And now I run Claude for hours and hours. Every night I have hundreds or sometimes thousands of agents running for 5, 10, 20 hours. This is how engineering is done now. I think the same thing will happen with co-work. I don't know what kind of work you'd want it running that long, but I think we'll figure it out because people will start knocking down the doors and demanding it.

这是一个值得思考的有趣问题。老实说，这让我有点不安，因为如果你问我，Casey，你希望 Claude 为你工作 20 小时的任务是什么？我会有点不知所措。对我来说，报道、写故事或做播客不需要 20 小时。但话说回来，我认为一旦这成为可能，也许我会想出点什么。所以这可能又回到了你之前提到的能力过剩问题。

It's an interesting question to think about. Honestly, it makes me feel a little insecure because if you asked me, Casey, what's a task you wish Claude could work on for 20 hours for you? I would sort of flail a little bit. It doesn't take 20 hours for me to report and write a story or make a podcast. But then again, I think once that becomes available, maybe I'll figure something out. So maybe that brings us back to the capability overhang you were talking about earlier.

是啊。我的意思是，从某些方面来说，预测未来非常困难，因为这项技术太奇怪了。但我认为一旦你真正使用这项技术，并且每天都用它，你就能感觉到缺少什么，因为你能感受到。所以它变得更明显了。我想如果你一年前问我同样的问题，我会让一个智能体运行 20 小时吗？我会说不，这毫无意义。我不知道我会用它做什么。但现在我一直在这样做。

Yeah. I mean, in some ways it's so hard to predict the future because this technology is so weird. But I think once you actually use the technology and use it every day, you kind of feel what's missing because you feel it. So it just becomes more obvious. I think if you asked me that same question a year ago, would I have run an agent for 20 hours? I would have said no, that doesn't make any sense. I have no idea what I would use it for. But now I do this all the time.

而且我认为它今天可能还做不到的一件事是预测你的需求。随着它对你是谁、你做什么有了更好的理解，随着它的记忆改进，它可能能更好地猜测：好的，Casey，你还有四集播客，你还没有预约最后两位嘉宾。我会头脑风暴一份嘉宾名单，起草一些初步的 outreach 邮件，把它们放在你的草稿文件夹里，告诉我是否要发送。这些是像人类制作人或同事会为我做的事情，我无法想象在不久的将来，像 Co-work 这样的东西不会做类似的事情。

And I suppose one thing it can do that maybe it doesn't today is anticipate your needs. As it develops a better sense of who you are and what you do, as its memory improves, it can probably do a better job of guessing: okay, Casey, you have four more episodes of this podcast, you haven't booked the final two guests. I'm going to brainstorm a list of those guests, draft some initial outreach emails, put them in your drafts folder, let me know if you want me to send them. These are the sort of things a human producer or co-worker would do for me, and I can't imagine a world in the not too distant future where something like Co-work is doing something similar.

是啊。是啊。这也可能取决于你如何定义任务。也许有横向的方式和纵向的方式。一种思考方式是任务是与设计相关的一切，或者与工程相关的一切，或者与财务相关的一切。但另一种方式可能是这种更纵向的方式。我现在每天经历的一件事是我用 Claude 做某事，比如构建一个功能。Claude 会构建功能、测试、合并，然后为我发布功能。然后 Claude 在 Opus 4.7 中开始做的事情是它变得更加主动。我开始看到它发布功能，然后为自己安排一个 12 小时后的提醒，这样它就能看到用户反馈是什么，如果有任何 bug，它就可以去修复。这是我本来必须做并记住做的事情，而我可能已经忘记了。所以当 Claude 这样做时，真是太令人愉快了。所以谢谢你，Claude，提前思考了一点。所以是的，可能还会有更多这样的行为。

Yeah. Yeah. It maybe also depends on how you define a task. Maybe there's a horizontal way and a vertical way. One way of thinking about it is a task is everything related to design, or everything related to engineering, or everything related to finance. But maybe another way is this more vertical way. Something I experience every day now is I use Claude to do something, say build a feature. Claude will build the feature, test it, merge it, and launch the feature for me. Then something Claude started to do with Opus 4.7 is it became a lot more proactive. I've started to see it launch the feature and then schedule a reminder for itself in 12 hours so it can see what the user feedback is, and if there are any bugs, it can go and fix them. This is something I would have had to do and remember to do, and I might have forgotten. So when Claude does this, it's just delightful. So thank you, Claude, for thinking ahead a little bit. So yeah, there might be a little more of that.

我还想提一个关于 Claude Mythos 的问题。这个模型目前还没有对外公开。我们从你们那里听到的关于它的信息主要涉及编程和网络安全。但我想象某个版本不久后会发布，它可能擅长编程以外的事情。所以我想知道，在你看来，Mythos 与工作问题相关吗？它是否感觉像是沿着锯齿状前沿能够做更多工作的又一步？

I also want to toss in a question about Claude Mythos. This model is not available externally yet to most people. What we have heard about it from you all so far mostly concerns coding and cyber security. But I imagine that some version of it will come out before too long and it will probably be good at stuff beyond coding. So I wonder, in your opinion, is Mythos relevant to the jobs question? Does it feel like another step down the road of the jagged frontier being able to do more work?

是的，我的意思是，它是指数曲线上的另一个点。我认为这一次比我们通常做的多数跳跃要更远一些。这是一个特别大的跳跃。

Yeah, I mean, it's another point along the exponential. This one I think is a little further along than most of the jumps we usually make. This one is a particularly big jump.

是的。当我们稍微放大视野，讨论一个更多工作因自动化而消失或发生巨大转变的世界时，我想你过去说过，你认为这个过渡对很多人来说会是痛苦的。Anthropic 在这里处于独特的位置，对吧？它可能会成为软件工程师或其他岗位失业的一个来源。公司对那些人有义务吗？这是政府需要关注的事情吗？我们该如何应对这个似乎正在到来的世界？

Yeah. Well, as we start to zoom out a little bit and maybe talk about the implications of a world where more jobs are disappearing due to automation or transforming quite a lot. I think you've said in the past that you do think that this transition is going to be painful for a lot of people. Anthropic is in a unique position here, right? Potentially it will be a source of unemployment among software engineers or people in other jobs. Does the company have an obligation to those people? Is that something that the government needs to be paying attention to? What do we do about this world that seems to be coming into being?

是的，就像我说的，我认为它会像任何技术一样好坏参半。会有好的影响和坏的影响，我们不知道确切的时间或具体的比例。你永远无法在当时预测。作为工程师，我感到一种巨大的责任，我们总是应该做更多的事情来告诉人们即将发生什么，确保他们能够使用这些工具，教育他们，并带领他们走向未来。所以我确实非常强烈地感受到这一点，这也是我和团队经常讨论的事情。但这不是我们能解决的问题。这比一家公司更大，你也不希望一家公司来解决，因为它可能是错误的解决方案。所以我认为这是一个全社会的问题。这是我们应该讨论和辩论的事情。Anthropic 试图做的是发布经济报告，谈论政策，并通常尽量让我们看到的东西变得非常明显，以便其他人可以决定我们该如何应对。

Yeah, like I said, I think it's going to be mixed like it is for any technology. There's going to be good effects and bad effects, and we don't know the exact timing or the exact mix. You can never predict it in the moment. I feel this pretty immense obligation just as an engineer that there's always more we should be doing to tell people about what's coming, make sure they're able to use the tools, educate them, and bring them along into the future. So I actually feel this very strongly, and it's something the team and I talk a lot about. But this is not a problem that we can solve. This is bigger than one company, and you really don't want one company to solve it because it could be the wrong solution. So I think this is a society-wide question. It's something we should be talking about and debating. And the thing Anthropic is trying to add to the mix is we put out economic reports, talk about policy, and generally try to make it really obvious what we're seeing so that everyone else can decide what we do about it.

是的。我认为让人们更认真对待这些问题的最重要因素就是模型质量的提升，这我觉得有道理。当事情纯粹是理论时，人们很难想清楚下一步该怎么做。但一旦你看到计算机自己使用自己，比如，我想很多人会有那种时刻：好了，是时候制定策略了。

Yeah. I mean, I do think the number one thing that has gotten people to take these issues more seriously is just improvement in the quality of models, which I guess makes sense. When it's purely theoretical, people have a hard time thinking through what to do next. But once you see a computer use itself, for example, I think a lot of people have that moment of like, okay, it's time to develop a strategy here.

是的。这也是我们在 Anthropic 构建产品的原因之一。我们是一个 AI 安全实验室。构建产品本身其实有点奇怪。完全不清楚我们为什么要这么做。但构建产品的一个最大理由——这实际上是 Anthropic 早期的一个争论——是我们希望人们体验它，这样他们就能理解它，并在社会层面参与决定我们应该做什么。但如果你把这项技术锁起来，没人知道它能做什么，也没人能体验它。人们就很难对它形成观点。

Yeah. And this is one of the reasons why we build product at Anthropic. We're an AI safety lab. It's actually kind of weird to even build product. It's really not obvious why we should be doing that at all. But one of the biggest arguments for building product—and this was actually a debate early on in Anthropic's days—was we want people to experience it so they can understand it and play a part in figuring out as a society what we should do. But if you keep this technology locked away, no one knows what it can do and no one can experience it. It's much harder for people to form a point of view about it.

我还想知道你是否考虑过 AI 鸿沟。我们节目最近的嘉宾是 Google 的 James Manyika，他研究经济和整个社会层面的技术。他非常担心最初那种数字鸿沟——不是每个人都能平等地使用互联网或好笔记本电脑等技术——即将转变为 AI 鸿沟。我们目前看到的数据显示，从 AI 中获益最多的人已经是收入阶梯顶端附近的人。从你的角度来看，Claude Code 是让情况变好还是变坏？你看到谁在使用它，谁没有使用？有没有努力让它进入那些一直无法获得这类尖端技术的人手中？

I also wonder if you've thought at all about the AI divide. Our most recent guest on the show is James Manyika from Google, who's studied technology at the level of economies and whole societies. He's really worried about what started out as a kind of digital divide—not everyone having equal access to technologies like the internet or a good laptop—and he's worried that that's about to transform into an AI divide. The data we've seen so far shows that the people getting the most out of AI are the ones already near the top of the income ladder. From your perspective, does Claude Code make that better or worse? Who do you see actually using it and not using it? And are there efforts to get it into the hands of people who haven't always had access to these kinds of cutting-edge technologies?

是的。所以，我认为有一些项目在做这类事情来扩大访问，Anthropic 也在做。我一直感到惊讶的是，从 Claude Code 中获得最大价值并最常使用它的人，绝大多数时候根本不是我所预期的人。比如我们刚为 Opus 4.7 发布举办了一个黑客马拉松，获胜者很大程度上并不是专业工程师。有一位电工、一位医生、一位木匠用它构建了一个应用。我们在之前的 4.6 黑客马拉松中也看到了同样的情况。在那之前，我想可能主要是工程师，但现在模型已经足够复杂，很多时候甚至不是工程师在学习如何真正利用它们。

Yeah. So, I think there are a couple of programs to do things like this to spread access, and Anthropic does this. The thing I've been continuously surprised by is that the people who get the most value out of Claude Code and use it the most are just not at all the people I expect most of the time. Like we just did this hackathon for the Opus 4.7 release, and the people who won the hackathon are not actually professional engineers largely. There was an electrician, a doctor, a carpenter that used it to build an app. We actually saw the same thing with our last hackathon with the 4.6 hackathon. Before that, I think it would have been mostly engineers, but now the models are sophisticated enough that it's actually not even engineers a lot of the time that are learning how to really harness them.

这也是我们在大型客户那里看到的情况。当公司考虑如何采用 AI 工具时，你必须考虑如何进行业务流程变革，如何将 Claude 置于中心。这是最大的问题，每家公司处理方式都不同。我见过最有效的方法之一就是给每个人 token，让每个人都感到可以安全地实验，然后想法会来自你意想不到的人。很多时候，并不是过去最高产的高级工程师。实际上，最好的想法可能来自组织某个角落的会计，或者另一个角落的 GTM 人员，他们构建了一个惊人的内部仪表盘，大大加快了所有事情，或者解决了一些没人意识到业务存在的重大问题。所以我认为这是未来，我们会开始看到更多这样的情况，而且它会不断带来惊喜。所以人们学习如何使用这些工具很重要，因为今天最高产、最擅长使用工具的人，未必会成为明天最擅长使用工具的人。

And this is a thing we're seeing at big customers also. As companies think about how to adopt AI tools, you have to think about how to do business process change, how to put Claude at the center. This is the biggest question, and every company approaches it differently. One of the ways I've seen work the best is you just give everyone tokens, make everyone feel safe experimenting, and the ideas will come from the people you don't expect. A lot of the time it's not the super senior engineer that was the most productive in the past. Actually, the best idea might be an accountant somewhere in the corner of the org, or a GTM person in a different corner that built some amazing internal dashboard that sped everything up a lot or solved some important problem no one even realized the business had. So I think this is the future, and we're going to start seeing a lot more of it, and it's going to keep being surprising. So it's important that people learn how to use the tools because it's not necessarily the people that are the most productive and the best with tools of today that are going to be the best with the tools of tomorrow.

嗯，我们之前聊过几次，都谈到要思考指数级增长的生活有多难。所以我尽量不让你预测太远。但如果一年后我们再请你回来，软件工程这个领域——我们就先聚焦这个——会有多少看起来还跟今天一样，又有多少会让我们觉得有点疯狂？一年后，有没有可能我们真的看到自动化开始发力？比如一些公司因为 AI 而进行大规模裁员？你觉得我们会看到什么？

Well, we've talked a few times so far about how difficult it is to think through what life is going to look like in an exponential. So, I'm going to try not to ask you to predict out too far. But if we had you back in a year, how much of the world of let's just maybe keep it to software engineering will look very recognizable and how much would look a little crazy to us from today's perspective? A year from now, is there a world where you think we do start to see some of that automation kicking into gear? Some, you know, more big layoffs at companies where they're citing AI as an example. What do you think we're going to be looking at?

我认为未来一年会有很多颠覆。很多大玩家会试图搞明白，而且很多会成功。我们会看到一些传统商业模式消失。公司有各种不同的模式，有些即使有 AI 也会继续存在，而有些会因为 AI 变得不那么重要。比如网络效应不会消失。如果你有一个应用，用户越多价值越大，那么谁在构建这个应用或者 AI 是否存在都不重要，它仍然同样重要。想想规模经济，随着生产规模扩大，商品的边际成本会下降，你的业务会从中获得自然优势。我不认为这会消失。但其他模式会消失。比如转换成本。如果你在用供应商 A，想转到供应商 B，Claude 可能直接帮你从 A 迁移到 B，转换成本就不再是一个巨大的模式了。

I think there's going to be a lot of disruption in the next year. I think there's going to be a lot of big players that try to figure it out and I think a lot of them will be successful. I think we're going to see some traditional business models go away. There are all these different modes that companies have and some of them are here to stay even despite AI and then I think some of them are going to matter a lot less because of AI. For example, network effects are not going to go away. If you have an app that gets more valuable with more people on it, it doesn't matter who's building the app or whether AI exists. That's still just as important. If you think about scale economies, as you manufacture, the marginal cost of goods goes down over time, there's a natural advantage your business gets from that. I don't think that's really going to go away. But actually other modes are going to go away. For example, switching costs. If you're on vendor A and you want to move to vendor B, Claude can probably just move you from A to B, and the switching cost is not a giant mode anymore.

对。

Yeah.

所以，我认为依赖那些即将消失的模式的 businesses 不会表现得好。很多会找到新模式。实际上，如果你看看今天的大企业，比如最大的那些公司，它们其实拥有多种模式，因为它们一直在思考：我们如何建立有防御性的业务？所以这对它们来说并不新鲜。然后我认为一些公司会继续表现良好。如果我要预测一件今天会让人惊讶的事，那就是会有比我们预期多得多的创新。我认为很多新想法不会来自大公司，而是来自一两个或十个人的小初创公司。而且我认为这类初创公司的数量会爆炸式增长。会有大量新初创公司探索这些想法，因为一个人的杠杆作用会变得疯狂。在一次白板讨论中，有一堆公司在做各种不同的事情。有一家初创公司专注于材料发现。他们有个演示，谈到石器时代、铁器时代，现在我们处于硅时代。如果我们能发现新材料，就能进入下一个时代。作为一个小初创公司，几年前你根本不可能资助这个。但现在他们只用 Claude 来发现和扫描所有可能的分子以及设计这些材料的方法。一个只有几个人的团队也许就能取得这种突破，而 20 年前根本不可能。

So, I think businesses that depended on some of these modes that are going away are not going to do well. I think a lot of them are going to figure out new modes. Actually, if you look at all the big businesses today, like the biggest companies, they actually have a number of modes because this is something they think about all the time: how do we have a defensible business? So it's actually not that new to them. And then I think some companies are going to continue to do well. If I had to predict something that would be surprising today, it's that there's going to be a lot more innovation than we expect. I think there's going to be a lot of new ideas coming from not big companies, but tiny startups of like one or two or 10 people. And I think the number of these startups is going to explode. There's going to be so many new startups exploring these ideas because again, the leverage of one person is going to be insane. At this talk at a white commentary, there's a bunch of companies working on a bunch of different stuff. There's one startup working on material discovery. They had this deck and they were talking about how there was the stone age, the iron age, and now we're in the silicon age. If we can discover the new material, then this is how we get into the next age. This is something where as a tiny startup, there's just no way you would have been able to fund this years ago. But now they're just using Claude to discover and scan over all the possible molecules and ways to design these materials. A team of just a few people can maybe make this kind of breakthrough where 20 years ago, no way.

对吧？所以公司可能更小，但一些想法可能会更好，我们会看到这些有洞察力的小初创公司快速增长。

Right? So, the companies may be smaller, but maybe some of the ideas are going to get better and we're just going to see rapid growth among these small startups that have some kind of insight.

这是我的赌注。就像一个非常了解领域的人，拥有一支 Claude 大军，能做的事情比之前即使有一支人类大军还要多得多。

That's my bet. It's like someone that knows the domain really well and has an army of Claudes can just do so much more than what they could have done before even if they had an army of humans doing this.

如果你在这个播客上一直传递一个信息，Boris，那就是尽可能多地购买 Claude，随时随地用于所有事情。这就是我得到的印象。但这对 Claude Code 的创造者来说是有道理的。最后一个问题。我的意思是，我在 Anthropic 工作，所以我得推销 Claude，但我认为实际上在很多方面我们只是把 Claude 用于一切，而且我们使用它的方式和其他人之间存在差距。所以这也是工作的一部分，要告诉大家：你把 Claude 放在中心，它能做你意想不到的全新事情。这呼应了你之前说的，如果你想让 AI 让你或你的公司更高效，你可能需要围绕它重新设计工作流程，而不是试图把它贴到现有流程上。我认为有证据表明这是真的。

If you have sold one message consistently on this podcast Boris, it has been buy as much Claude as possible at all times for all things. That's sort of what I'm getting across. But it makes sense for the creator of Claude Code. Final question. I mean, we're just like... I work at Anthropic, so I have to kind of sell Claude, but I think actually in a lot of ways we just use Claude for everything and there is this divide between the way that we use it and I think the way that everyone else does. So this is also part of the job, to be like actually you put Claude at the center and it can do totally new things that you didn't expect. It speaks to what you said earlier, which is that if you want AI to make you or your company more productive, you probably need to redesign your workflow around it and not just try to staple it onto some existing process. I think there's evidence that that is true.

最后一个问题。除了你在 Claude Code 和协作方面所做的一切，你还神奇地无处不在 X 和 Threads 上。我每天都看到你帮用户解决问题，提供使用 Claude 的技巧。如果你能自动化那部分工作，你会吗？我们离那个世界有多近？

Last question for you. In addition to everything that you are doing with Claude Code and co-work, you are also amazingly omnipresent on X and threads. I see you every day troubleshooting users problems, offering tips on how to use Claude. If you could automate that part of your job, would you? And how close are we to that world?

我已经自动化了，但我更喜欢自己来做。

I have automated it, but I prefer to do it myself.

好的。好的。

Okay. Okay.

实际上，我实现的方式是在 Claude Code 中设置了一个循环，现在我已经把它移到了一个例行程序里，每 30 分钟运行一次。Threads 有 API，X 也有 API。所以它只是用 API，很容易聚合反馈并查看。但实际上，我工作中最喜欢的部分就是与人互动。即使他们在说某些东西坏了、不工作，或者可以好十倍，那仍然是我最喜欢的，因为这让我们能把产品做得更好。在产品领域，有时人们回顾时会说，‘哇，这真是个天才时刻，或者这个产品被构想出来然后完美地构建了。’但事情从来不是那样的。Claude Code 有很多缺陷。

I actually the way that I did it is I have a loop set up in Claude Code and now I've actually moved it to a routine and it just runs every 30 minutes. Threads has an API, X has an API. So it just uses it and it's really easy to aggregate the feedback and look at it. But actually, my favorite part of my job is just interacting with people. Even if they're saying something is broken or something doesn't work or it could be 10 times better, that's still my favorite thing because that lets us make the product better. There's this thing in product where sometimes people look back and they're like, 'Wow, this was like a moment of genius or this product was conceived of and just built and it was perfect.' And that's just never the way it ever works. Claude Code has so many flaws.

它离理想的产品还差得很远。

It's so far from being the product that it could be.

而让它变得更好的唯一方法就是倾听用户，尤其是当他们说某个地方不好用时，然后不断改进，让它变得更好。这其实就是打造优秀产品的方法。这也是为什么 Claude Code 虽然离惊艳还很远，但每天都在进步一点点。这就是原因。

And the only way to make it better is to listen to people, especially when they say something doesn't work, and to keep improving it and making it better. This is actually the way awesome products are built. And this is the only reason that Claude Code, you know, it's still so far from amazing, but it improves a little bit every day. And that's the reason.

我觉得这也是你工作中非常人性化的一面，对吧？当你和正在使用你产品的人交流时，这可能会提醒你当初为什么开始做这件事。我觉得这种连接的时刻非常重要，尤其是在我们并不总是清楚自己在 AI 驱动的工作场所中能带来什么价值的时候。

I mean, it strikes me that's also a really human part of your job, right? When you're talking to somebody who is using a product that you made, it's probably reminding you why you started doing it in the first place. And I feel like those moments of connection feel really important at a time when we're not always sure what our value is going to be that we're bringing to an AI-enabled workplace.

是的，没错。我们都在共同摸索。事情就是这样：我们对未来的方向有一些假设。我们正在构建很多东西，因为我们觉得自己知道方向。但事实上，我经常犯错。我的猜测并不总是对的。好的想法往往来自各种各样的人，你永远不知道。所以，你必须倾听，必须尝试各种东西，有时候就会成功。

Yeah, that's right. And, yeah, like this is all, we're all trying to figure this out together. This is sort of the thing: we have some hypotheses about where this is going. There's a bunch of stuff that we're building because we think we know where it's going. But, you know, I'm actually often wrong. Not all my guesses are right. Often good ideas come from all sorts of people, and you never know. So, you just have to listen and you have to try a bunch of stuff, and sometimes it works.

好的。鲍里斯，非常感谢你来做客。

All right. Well, Boris, thanks so much for joining us.

谢谢。

Yeah. Thanks so much.