新创业指南 | Box 首席执行官亚伦·莱维

初创公司实际上可以超越这些大公司，这在以前是根本不可行的。

Startups can actually outrun these big companies in a way that was just never economically feasible before.

你几乎可以在任何领域竞争，因为人们忽略了整个这件事。

You could actually go and compete in almost any category on the dimension of people are missing this whole thing.

我真的认为会有数百家公司，市值都在10亿、50亿、100亿美元甚至更高，基本上就是……

I really think that there will be hundreds of companies that are all $1 billion market cap, $5 billion market cap, $10 billion market cap, and beyond, that basically are just...

基于……，新初创公司有近乎无限的机会涌现。

There's near infinite opportunity for new startups to emerge on the basis of

人工智能给了小型初创公司前所未有的能力，那就是以曾经只有大公司才有的规模来部署资源。

AI is giving small startups something they've never had before, the ability to deploy resources at a scale that once belonged only to big companies.

在本期《Speedrun Office Hours》中，Box首席执行官Aaron Levie解释了为什么一个三人、五人或者十人的团队，能够越来越多地产生一个规模数倍于它的公司的产出。

In this episode of "Speedrun Office Hours," Box CEO Aaron Levie explains why a three, five, or 10-person team can increasingly produce the output of a company many times its size.

Aaron分析了创始人如何成为AI代理的管理者，为什么随着执行成本降低，策略和品味变得更加重要，以及为什么他预计专业代理将在几乎每个行业和工作职能中出现。

Aaron breaks down how founders are becoming managers of AI agents, why strategy and taste matter more as execution gets cheaper, and why he expects specialized agents to emerge across nearly every industry and job function.

但初创公司剧本中有一点没有改变，那就是分销。

But one part of the startup playbook hasn't changed, distribution.

正如亚伦所说，"如果你建好了，他们不一定会来。"

As Aaron puts it, "If you build it, they will not necessarily come."

亚伦，欢迎。

Aaron, welcome.

谢谢你来参加我们的节目。

Thanks for joining us.

谢谢邀请我。

Thanks for having me.

有意思的是，你其实还很年轻，只比我大几岁，可能和安德鲁差不多，但你从2005年就开始经营Box了，那差不多是你人生的一半时间了。

It's kind of interesting that you're actually still a young guy, only a few years older than me, and probably similar to Andrew, but you've been running Box since 2005, which is almost exactly half your life.

如果你把现在的创业世界和那时候相比——

If you could compare the world of startups now to the world as it was back then-

你觉得创业体验有什么变化？

... what do you think about the startup experience has changed?

我最大的感受其实是，我真的觉得自己老了。

My biggest takeaway, actually, is I feel really old.

所以我挺感激你用你的年龄来定义什么是年轻。

So I appreciate that you're using your age as the anchor for what young is.

但我觉得，不幸的是，我现在常常觉得自己很老。

But, I think unfortunately, I tend to feel very old now.

而且，我对现在这么多年轻的创始人感到非常佩服和兴奋。

And, I am so impressed and excited by so many of these young founders right now.

有趣的是，2005年的时候，我们大三中途辍学。

And, what's interesting is that in 2005, so we dropped out of college, middle of our junior year.

在2005年，辍学然后20岁就创业是相当罕见的。

And in 2005, it was fairly rare to drop out and be 20 doing a startup.

那时候硅谷大概只有几十个这样的人。

There were a few dozen of us in Silicon Valley at that time.

我觉得这主要是受Facebook的影响，辍学成了一种潮流，但当时还是比较少见的事情。

I think prompted probably mostly by Facebook had made dropping out a thing, but it was still a relatively rare event.

而现在最让人着迷的是，现在辍学的20岁年轻人数量之多。

And what is so fascinating now is the amount of now 20-year-olds that are dropping out.

对这些孩子来说，现在这简直成了标准操作流程。

It is just like now a standard playbook, for these kids.

所以，我觉得特别兴奋的是，现在有这么多年轻人开始创业，直接投身到创业公司里。

And, so I think it's super exciting just how many young people are getting started and jumping right into building startups.

所以我确实觉得自己挺老的，不过，现在的玩法确实大不一样了。

So I do tend to feel pretty old, but yeah, the playbook seems incredibly different.

现在创业公司的运作方式，是我见过的所有技术变革时期里，变化最大的。

The way that startups are being run right now is the most different than any other period of technological change that I've ever seen.

如果你看看2005年我们怎么运营Box，再和十年前的公司比，比如拿1995年的公司和2005年的创业公司比，其实差别没那么大。

Where if you looked at how we ran Box in 2005, and you compared us to a startup 10 years prior, so let's say you took the 1995 company and the 2005 version of a startup, it actually wasn't that different.

至少在2005年，那会儿云计算浪潮还没完全到来。

At least in 2005, it was a little bit right before the cloud wave.

我们还得去数据中心。

We still went to data centers.

我们还得自己装服务器。

We racked servers.

我们没日没夜地写代码。

We coded morning and night.

我们是一家初创公司。

We were a startup.

但如果你拿2005年的我们和95年的公司比，运营方式几乎一模一样。

But if you looked at how we ran the company in 2005 compared to '95, it would almost look exactly the same.

就像，95年每家初创公司的起步方式和我们当时起步、运营业务的方式完全一样。

Like, the way that every '95 startup got started looked exactly like how we got started and ran the business.

如果拿2025年的初创公司和2005年的我们比，或者拿2015年的初创公司比，也就是说，按十年一个跨度来看，这是我在整个成年生涯中见过的最大的不同，我从事科技行业已经27年了。

If I compare the 2025 startup to us in 2005, or let's say the 2015 startup, so if you just use 10-year increments, it is the most different that I've ever seen in the history of my adult life and doing technology now for 27 years, let's say.

这完全是因为这种以AI为先的导向，你可以用三、五、十个人的团队，产出五十、一百甚至两百人团队的成果。

And it's simply because of this AI-first orientation where you can be a three, five, 10-person team and have the output of a 50 or 100 or 200-person team.

这彻底改变了运营这些公司的方式。

And that just changes everything about how you run these businesses.

所以我对这些初创公司印象非常深刻。

So I'm wildly impressed by these startups.

我也觉得有些公司简直是疯狂得离谱。

I also think some of them are just totally crazy and unhinged.

但看着挺有意思，也能学到东西。

But it's fun to watch and fun to learn from.

但这是我见过的，如今公司构建方式最不一样的一次。

But it is the most different that I've ever seen how to build a company today.

关于这一点，挺有意思的，对吧？

On that point, it's interesting, right?

就拿你和你的联合创始人来说，你们辍学了。

In the case of it, where you and your co-founder, you guys dropped out.

如果你们当时能用AI代理来成倍放大能力，比如。

If you had the ability to force multiply with AI agents, for example.

我很好奇你看到了什么。

I'm curious what you're seeing.

我觉得你最近写过相关文章，特别是有些初创公司，在后台做了一些事情，或者用AI代理的方式挺有意思的。

I think you wrote about this recently, how some of these startups are, in particular, I think, doing things in the background, or they're using AI agents in sort of interesting ways.

到目前为止，你见过的最疯狂的事情是什么？

What's been the most craziest thing that you've ever seen so far?

嗯，如果你只是从基本前提出发，然后把它展开——

Well, if you just take the basic premise, and then draw it out-

……当我们创办Box的时候，我们作为初创公司的进步速度是——

... when we started Box, our rate of progress as a startup was-

……100%取决于我们能在键盘上打字多少小时。

... 100% correlated to how many hours we could be on a keyboard and type things.

我们唯一的杠杆点就是我们打字输入键盘的速度有多快。

The only point of leverage we had was how fast we could type text into a keyboard.

我们可能是在写代码，可能是在写营销材料，也可能是在写博客文章。

And we could be typing code, we could be typing marketing assets, we could be typing blog posts.

我们当时在管理服务器。

We were managing servers.

但我们四个人都辍学了。

But the four of us, we all just dropped out.

我们住在一个改造过的车库，工作和生活都在那里。

We lived and worked in this renovated garage, live-work kind of thing.

那时候，我们执行的速度完全取决于我们在电脑上工作的速度。

And that was the only speed at which we could execute, was how fast we could work on computers.

我现在接触的初创公司，虽然不全是这样，但可以说越来越多是这样。

The startups that I'm talking to now, this is not exclusively the case, but let's say it's increasingly the case.

现在似乎有一种新的感觉，初创公司的创始人和早期员工，而且我认为这基本上会随着时间扩展到每个行业——

It seems like there's now a new sense where the startup founders and the early employees, and I think this basically scales to To every industry over time-

他们更像是工作的管理者，而工作是由智能体完成的。

... they're more managers of work, and the work is being done by agents.

所以他们不再受限于打字速度、写代码的速度，或者制作营销素材的速度。

And so they're no longer rate limited by how quickly they can type and how quickly they can write code and how quickly they can come up with marketing assets.

他们部署智能体去做很多这样的工作。

They are deploying agents to go and do a lot of that work.

我最近和一个初创公司创始人聊过，他声称，而且我基本相信，他每天只是审查智能体写的几千行代码。

And I was talking to a startup founder recently that claims at least, and I basically believe it, he's just reviewing thousands of lines of code a day that the agent did.

而作为编辑或审查者的能力，其杠杆效应是实际创造事物之人的十倍。

And the ability to be an editor or a reviewer is like 10X the leverage of the person who actually has to create the thing.

所以，如果你以此为思维模式，即这个人现在是代码的编辑者、协调者和审查者，他们可以继续部署多个代理，因此他们不再受限于单一终端或单一用户界面的交互速度。

And so if you have that as your mental model where this person now is the editor, orchestrator, reviewer of code, and they can go on and deploy multiple agents, so they're no longer rate limited by, again, one terminal, one user interface that they can interact in.

作为个体程序员，你一次只能在一个IDE中打字。

You can only type in one IDE at a time as individual programmer.

但现在我可以在后台运行许多代理，当它们全部完成时，我可以去审查它们的工作。

But I can have lots of agents now running in the background, and I can go review their work when they're all done.

这完全是一种全新的杠杆形式。

That's just a totally new form of leverage.

所以现在的瓶颈只是审查代码、协调、集成、启动代理。

And so now the bottleneck is just reviewing code, orchestrating it, integrating it, kicking off agents.

瓶颈不再仅仅是你写了多少代码。

The bottleneck is not just how much code have you written.

所以我主要在工程方面看到这一点。

So I'm mostly seeing it on the engineering side.

所以最大的好处是，它给初创公司带来了一种以前在创业界从未有过的杠杆。

So the big benefit is it's giving startups a new form of leverage that never existed in the startup world.

通常，初创公司有个劣势，就是只有两三个人，却要对抗那些有百倍人马的巨头。

Usually, you have this deficiency as a startup, which is you have two or three people, and you're attacking some big incumbent that has 100 times more people.

你虽然能比那些巨头跑得快，但他们总能投入更多资源。

And you can basically move faster than that incumbent, but that incumbent can always deploy more resources.

而我们现在正通过智能体来逐渐缩小这个差距。

Well, we are increasingly neutralizing that delta because of agents.

我们还没完全做到，因为智能体会犯错。

We're not fully there because agents make mistakes.

它不是万能的。

It's not a panacea.

很多人把它形容得像一台老虎机。

A lot of people describe it like a slot machine.

你得不断尝试——

You have to keep trying-

一遍又一遍地试，看看智能体会带回什么结果。

... over and over again and seeing what the agent comes back with.

但你可以稍微想象一下这可能会变成什么样子，那就是想象一个世界，智能体能够写出几千行高质量代码，符合你的代码标准，而且真的能运行，你审查代码，只需要改动几处。

But you can squint and see what this could turn into, which is imagine a world where the agent could do a few thousand lines of code that was really high quality and it adhered to your code-based standards and it really worked, and you reviewed the code, and you only changed a few things.

那样的话，作为初创公司，你就拥有了十倍的工程能力，这意味着你几乎可以在任何类别上，在规模和速度的维度上真正去竞争。

Well, that would mean that you do have now 10X the engineering capacity as a startup, which mean you could actually go and compete in almost any category on the dimension of scale and speed.

这会彻底改变竞争格局。

Which would change everything about competitive dynamics.

另外，有趣的一点是，大公司很难做到同样的事情，因为他们有很多工作流程和流程，这些都已经深深嵌入他们现有的运作方式中，他们无法足够快地做出改变。

And then as an interesting just asterisk, the big companies find it very hard to actually go and equivalently do that because they have a lot of workflows, a lot of processes, that are so wired into how they already operate, that they won't be able to make the change fast enough.

所以，你有一个窗口期，初创公司可以以前所未有的经济可行性超越这些大公司。

So you have a window where startups can actually outrun these big companies in a way that was just never economically feasible before.

而这实际上完全归功于AI智能体。

And it's effectively all due to AI agents.

所以，再次强调，从代码开始，但我认为最深入整合的公司正在研究如何在营销领域做到这一点。

So again, starting in code, but I think the most wired-in companies are figuring out how to do this in marketing.

他们会想办法在法律工作上应用。

They're going to figure out how to do in legal work.

他们会想办法在对外销售上应用。

They're going to figure out how to do in outbound sales.

所以这将是一个非常有趣的演变——

So that will be a pretty interesting evolution-

……创业公司的形成过程。

... of the startup formation.

我想知道的是，Aaron，我们之前总是说每个人都想以策略的速度工作，对吧？

I guess the thing I'm wondering about, Aaron, is then we've always talked about before that sometimes everyone wants to work at the speed of strategy, right?

他们希望能够以创造力和思考的速度工作，对吧？

And they want to be able to work at the speed of creativity and thought, right?

没有人想花所有时间写营销文案，除非你是硅谷最受欢迎的互联网喜剧演员Aaron Levie。

Nobody wants to spend all their time writing marketing copy, unless you're Silicon Valley's favorite internet comedian like Aaron Levie.

但我确实认为有一个问题，如果你现在可以更加注重策略，那么你做的决定实际上非常重要，对吧？

But I do think that there is this question, though, that if you are now can be more strategic, though, the decisions you make actually matter quite a bit, right?

对。

Yes.

你可能会搞砸。

You might really mess up.

不过幸运的是，成本不高，但你仍然会出错。

And then luckily, it doesn't cost that much, but you're still going to mess up.

实际上，创造力的速度和棋盘布局更重要。

And in fact, really the speed of creativity and the chessboard actually matters more.

是的，我觉得有趣的一点是，现在至少在初创公司里，构建软件的方式似乎是先有非常清晰的提示或规格。

Yeah, I think one thing that is interesting is so the tendency appears to be that the way you're building software now, at least in these startups, is you get really clear prompts or specs.

你会有非常明确的方式与代理协作，理解你的代码库，在合适的位置构建子代理。

You get very clear ways of working with agents to understand your code base, to build sub-agents at the right parts of your code base.

而且越来越流行这种规格驱动的开发，也就是你写非常详细的——

And there's an increasing view of this spec-driven development, which is you write very in-depth-

……关于你要构建什么的规格说明。

... specifications about what you build.

差不多就是产品工程师这个概念，对吧？

Sort of the idea of the product engineer, right?

对。

Yes.

其实更准确地说是产品经理。

The product person, actually.

那个像迪尔伯特漫画里尖头发产品经理那样的人，现在真的有能力自己动手做事情了。

The pointy-haired Dilbert product manager now truly actually has the power to do their own stuff anyway.

你只要等得够久，他们就会变得非常强大。

You wait long enough and they become, yeah, very powerful.

但结果就是，如果你真的很擅长写规格说明，并且很清楚自己想构建什么，那现在就是极大的杠杆。

But so it turns out if you're really good at writing a spec and you're really good at knowing and having clarity about what you want to build, that's actually extreme leverage now.

所以现在非常看重的是，你是否真正了解市场机会，是否擅长制定策略。

So there's a huge premium on knowing, on actually being very good at strategy, being very good at what is the market opportunity for what you're building.

以前可能更看重的是你能多快把东西拼凑出来。

Previously, there used to be probably a premium on just how quickly can I hack something together.

不过现在呢，因为快速拼凑东西已经变得很普遍了，重点在于你有没有想清楚你到底要构建什么。

Well, now actually, because the hacking something together is fairly commoditized, the premium is have you thought through what you're actually trying to build?

是不是真的有——

Is there a real-

……市场机会？

... market opportunity for it?

这是个好主意吗？

Is it a good idea?

看起来好看吗？

Does it look good?

有没有品味？

Does it have taste?

这些现在才是创造价值的新方式。

Those are now the new forms of value creation.

所以我认为，那些真的非常执着于把这件事做对的公司在未来会领先。

And so I do think that the companies that are really religious about getting that right will be the ones that get ahead in the future.

我想这背后的问题是，好吧，假设你是个初创公司，你手头上基本上有所有这些工具。

I guess the question behind this is, okay, so let's assume that if you're a startup, you effectively have all these tools at your disposal.

你显然可以采纳那些大公司的方案，对吧？

You could obviously adopt one of the bigger folks, right?

然后问题在于，你之前提到过AI上下文腐烂这个概念，基本上解决这个问题的办法就是采用这些点解决方案，或者设计良好的智能体，而不是只有一个终极的万能工具。

And then the issue with that is you've argued around this idea of AI context rot, which basically the way to solve that is to actually have these point solutions potentially, or great well designated agents, as opposed to having the single ultimate fighter.

你实际上有一支复仇者联盟团队在AI世界里帮你。

You actually have a team of Avengers to go and help you in the world of AI.

我想如果你是那个初创公司，对吧，你正在试图弄清楚你首先推出的智能体架构是什么，快速上手的方法是什么？

I guess if you're that startup, right, and you're trying to figure out what's the agent architecture that you ship with first, what's the fast way to get off the rails?

这就是那种情况，你什么时候会放弃一切？

It's one of these things where it's at what point do you just give up everything?

你什么时候开始填补...

And at what point do you start filling out...

比如说，你只是从一个基础工具包开始吗？

Do you just start with a basic toolkit, for example?

嗯，我在那个语境里想说的重点是——

Well, the point I was making in that context-

——其实不在于你用多少工具，而在于需要专业化——

... is not so much about the number of tools you use as much as the need for specialization-

——以某种虚拟的或非常具体的方式。

... in some In some virtual or very tangible way.

比如，Claude Code 就通过让你创建子代理来实现这种虚拟专业化，这些子代理可以——

And so for instance, Claude Code has created a way to have that virtual specialization by just letting you go and create sub-agents that-

——负责你代码库的某一部分或工作流程的某一部分，这样就能在 Claude Code 的范式内解决上下文腐烂的问题。

... own a particular part of your code base or a particular part of your workflow, and that's how you resolve the context rot issue within the Claude Code kind of paradigm.

所以，重点不是你需要为每件事都准备一个工具，而是我们面临一个非常现实的问题：超级强大的超级代理并不存在。

So, it's not so much that you want a tool for everything, it's that we have a very real issue, which is this super powerful, super agent doesn't exist.

它行不通。

It doesn't work.

因此，你确实需要一定程度的专业化，而这种专业化可以通过多种形式来实现。

And so you do have to have some degree of specialization, and that specialization can come in a variety of packages.

我觉得关于怎么解决这个问题，YouTube上每个方法都有视频。

I think there's a YouTube video on every way to solve this.

你要是问三家或五家初创公司，会得到三到五种不同的答案。

If you ask three startups or five startups, you're going to get three or five different answers.

这完全没问题，因为现在大家都还在摸索最佳实践，一切都还在发展中。

And it's totally fine because this is the stage where everybody's kind of figuring out the best practices, and it's all emerging right now.

但在这方面我给不出什么通用原则，因为每个团队、每家公司、每种工具都不一样。

But I can't give any kind of general principle on that because it's just different by team, by company, by what tool you choose.

但至少从你的观点来看，这肯定不是只会让现有的大公司受益。

But it would seem that at the very least, your perspective is there's definitely not like, this is not just going to be able to go and benefit incumbents.

这里有个独特的机会。

There's this unique opportunity.

如果你能作为初创公司成为那些专业工具之一，

If you could be one of those specialized tools as a startup-

是的

Yes

……那你的处境就非常非常好了。

... you're in a really, really good spot.

就像你基本上可以——

Like you can basically-

完全同意

Totally

……重新打造整个技术栈的领域。

... remake the whole area of the stack.

就像很多——

Like the number of-

对，好的。

Yeah, okay.

不错

Cool

……很多初创公司都在重新打造技术栈的每一个部分。

... startups that are remaking every part of the stack.

所以现在，如果我稍微换个角度，因为我之前主要是从普遍的角度来回答这个问题，AI初创公司能比老牌公司快10倍，这太棒了。

So now, if I flip it a little bit, because I was mostly answering the question through the lens of generically speaking, the fact that AI startups can move 10 times faster than an incumbent, that's fantastic.

现在，如果你再想想，好吧，这对你应该构建什么意味着什么？

Now, if you think about it now, okay, what does that mean in terms of what you should build?

你应该如何在这个市场中竞争？

How should you compete in this market?

我认为有无限的机会，这显然有点夸张，但确实有近乎无限的机会让新的初创公司涌现，因为我们将看到软件格局的彻底转变，因为许多以前没有软件的行业和业务领域，现在将首次以AI代理的形式拥有软件。

I think there's infinite, that's obviously a little hyperbolic, but there's near infinite opportunity for new startups to emerge on the basis of we are going to see a complete shift in the software landscape because so many industries and parts of business that previously didn't have software now will have software in the form of AI agents for the first time ever.

而且，如果我看看大多数知识工作领域，都会有某种与该类知识工作相关的代理。

And if I look at most domains of knowledge work will have some kind of agent that is associated with that type of knowledge work.

老牌公司会占据一些类别，但在每一个类别中，基本上都有初创公司的机会来构建做知识工作的代理。

Incumbents will take some categories, but there'll be startup opportunities in every single category for basically building agents that do knowledge work.

而最先达到一定规模的是编码代理。

And the first ones that have reached some degree of critical mass are coding agents.

而且，在Replit、Cursor、Windsurf、Lovable、Cognition这些形式中，已经出现了非常明显的初创公司机会。

And there's been very clear startup opportunities in the form of Replit, Cursor, Windsurf, Lovable, Cognition.

天哪，如果十年前有人告诉我们，编程领域会出现两三百亿、三四百亿甚至四百亿美元的市值，我们可能会说，嗯，GitHub 大概是唯一一个真正在编程领域取得成功的公司。

Like holy crap, if we had said 10 years ago that there'd be $20, $30, $40 billion of market cap in the coding space, and we would have been like, "Well, kind of GitHub is the only thing that's ever really even exited in the coding space.

怎么可能在短短三年内就出现三四百亿、四五百亿美元的市值呢？

How could there be $30, $40, $50 billion within just a three-year period?"

这听起来不太可能。

That wouldn't sound plausible.

但事实不仅发生了，而且这还只是冰山一角。

And yet that's not only happening, but it's actually just scratching the surface.

我完全有信心认为，AI 编程领域的市值能达到两三千亿美元，而且这些市值不会都归老牌公司所有，纯粹是因为这个领域太大了，市场空间太广阔了。

I am perfectly comfortable underwriting a couple of hundred billion dollars of market cap in the AI coding space that don't go to incumbents purely because there's just so much surface area, so much TAM that needs to exist.

好，现在把这种思路应用到每一个领域。

Okay, now do that for every single field.

生命科学、医疗保健、法律工作、咨询行业，都是如此。

Do that for life sciences, do that for healthcare, do that for legal work, do that for consulting.

每一个领域都会出现新的 AI 代理公司。

There is going to be new AI agent companies in every single one of those spaces.

而且它们不仅会超越传统的软件公司，还会超越那些领域的传统服务公司。

And not only will they outrun the traditional software incumbent companies, they're going to outrun the traditional services companies in those spaces as well.

这对初创公司来说，是一个全新的环境。

And it's just a completely new environment for startups to emerge.

你在这次对话开头说的一个观点我觉得很有意思，就是当你和创始人交流时，你感觉大约有70%的时间是在学习。

You said something at the beginning of this conversation that I thought was interesting, which was when you talk to founders, you feel like you're doing the learning about 70% of the time.

我很好奇，有没有什么让你印象深刻的？比如，当你和新生代创始人交流时，你从他们身上学到了什么？

I was curious if there's anything that comes to mind, like what are you learning from kind of the new generation of founders when you do speak with them?

对，我主要是在每次对话中重新思考智能体带来的杠杆效应。

Yeah, I mostly just have to rethink the leverage that you get from agents in every conversation.

比如，如果拿一年前我和创始人的一次典型对话来说，我会问：“嘿，你是怎么使用AI的？”

So if I took an example conversation I would have had with a founder a year ago, and you say, "Hey, how are you using AI?"

那时候的标准答案还是GitHub Copilot，可能还有少数人在用Cursor。

And again, the standard at that point would've been GitHub Copilot, maybe a few users on Cursor.

他们会说：“对，我大概有10%到20%的生产力提升。”

They would say, "Yeah, I'm getting 10 or 20% productivity gains.

它真的帮我实现了自动补全和自动完成功能。

It's really helping me with type ahead, auto-complete functionality.

它让我不用再去谷歌搜索编码问题了。

It's replaced me needing to Google when I have a coding question."

现在，如果你问最先进的初创公司创始人他们如何使用AI，又回到了这一点。

Now, if you ask the most advanced startup founders how are they using AI, it's back to this point.

这不是20%或30%的杠杆效应。

It's not a 20, 30% form of leverage.

这可能是3倍、5倍甚至10倍的杠杆效应，因为有了后台代理的能力。

It's possibly a 3 to 5 to 10 X form of leverage because of this ability to have background agents.

但问题是，在你现有的团队、现有的组织、现有的工程方式中，并没有为启动大量后台代理去执行代码库而做好准备。

But the problem is, again, in your existing teams and your existing organizations, your existing way of doing engineering is not wired up for kicking off lots of background agents to go and execute on your code base.

你的代码库也不是这样设计的。

Your code base isn't designed in that way.

很多时候，你的代码库中并没有足够的文档来支持代理去操作代码库。

You don't have the documentation in the code base in a lot of times to support agents to go and operate on the code base.

你大概也没有合适的工具栈来做这件事。

You don't probably have the tool stack to be able to do that.

你还没有改变工程方法论，让几十上百人能够同时去做这些。

You haven't done the engineering methodology change to enable dozens or hundreds of people to go do all this at once.

实际上，这对小团队是有利的，因为他们不会在所有代理之间一直遇到这些冲突。

It actually is something that benefits a small team where they're not going to have all these conflicts happening at all time across all these agents.

所以，这对小型初创公司来说，自然是有优势的，能够快速行动。

And so this is something that naturally is advantaged for the smaller startup to be able to move quickly on.

所以当我跟初创公司交流时，我主要是在了解你们是怎么做到的。

And so when I'm talking to startups, I'm mostly just learning how are you doing that?

然后想办法找出我们公司如何尽快做到这一点，以便我们能充分利用这些经验，并把它们应用到我们运营Box的方式中。

And then trying to figure out what's my way of bridging our company to do that as quickly as possible so we can take a lot of the benefit of these learnings and port them into how we run Box.

这很有道理。

That makes so much sense.

好的。

All right.

回到我们最开始的问题，我很好奇你是怎么想的，如果有的话，你觉得大家需要知道哪些重要的事情。

Circling back to our very first question, I'm curious how you're thinking about, if any, important things you think folks need to know.

显然，我们谈到了杠杆、智能体，基本上就是两个人就能创业的能力。

Obviously, we've talked about leverage, agents, the ability to basically be, you're a two-person start.

所以实际上，选择你的联合创始人变得重要得多。

So actually, choosing your founder really matters a lot more.

因为那基本上相当于两千人。

Because effectively, that's equal to 2,000 people.

对，完全正确。

Yeah, exactly.

是的。

Yes.

但想想那个模式，想象一下现在的Aaron辍学。

But thinking through that paradigm now, picture Aaron from dropping out now.

你现在会对Aaron说什么？

What would you tell Aaron now?

我们已经谈了很多了，但更基础的东西是什么？

We've already talked a lot about it, but what are the more basic building block things?

有什么变化吗？

Does anything change there?

我觉得公司建设的很多核心原则是相对永恒的，它们超越了任何特定的技术浪潮。

I think a lot of the core principles of company building are relatively timeless, and they transcend any particular technology wave.

正如你说的，团队可能比以往任何时候都更重要，因为联合创始人或创始团队带来的杠杆效应。

To your exact point, team probably matters more than ever before because of the leverage on your co-founders or co-founding team.

所以，我觉得这是个很好的见解。

So, I think that's a great insight.

那是你的见解。

So that was yours.

嗯。

Huh.

所以，我只是想强调这一点。

So, I just would underscore that point.

那市场呢，比如说？

What about market, for example?

有些人说过——

Some folks have said-

……试着找到最专业、最无聊的市场——

... try to find the most highly specialized, really boring market-

对

Yes

……在这个星球上，也许是为了找到监管漏洞。

... on the planet to find regulatory capture, maybe.

或者你会说，"别管那个。

Or are you like, "Forget about that.

那对你没用，伙计。"

That's not going to work for you, man."

不。嗯，有些事情是在你有了想法之后才发生的，但如果你想一路追溯到还没有想法的时候，那也许我可以从那里开始。

No. Well, there's some things that are downstream from you have the idea, but if you want to go all the way up to you don't have the idea yet, then maybe I'll start there.

我真的认为将来会有几百家公司，市值都在十亿、五十亿、一百亿甚至更高，它们基本上就是选取经济中的一个工作职能，然后为这个职能打造一个智能体，再把这个智能体卖给企业、服务公司、系统集成商，卖给所有人。

I really think that there will be hundreds of companies that are all one billion market cap, five billion market cap, 10 billion market cap, and beyond, that basically are just taking a job function in the economy, and they're building an agent for that job function, and they're selling that agent to corporations, to services firms, to system integrators, to everybody.

而且我觉得你完全可以做一个矩阵，也许安德森就有一个这样的，你可以把每个行业和每个行业里的每个工作职能都列出来，那大概会有五千个格子，就像生命科学领域有销售代表，金融服务领域有律师，有些市场可能总市场规模有限，但有些市场会大得多。

And I think you could literally do, maybe Andreessen has one of these, you could literally do a matrix that is every industry and every line of every job function per industry, and it would be literally 5,000 cells, and it would be like, there's sales reps in life sciences, there's lawyers in financial services, and some of the markets probably will be TAM constrained, but some will be much bigger.

有些可能需要这样，"好，我们得拿下整个横向市场。"

Some need to be like, "Okay, we have to get the whole horizontal."

有些就得拿下整个纵向市场。

Some will have to be we'll have to get the whole vertical.

但我认为你完全可以遍历整个经济，为其中任何一个位置来评估AI智能体的可行性。

But I think you could very easily go through the entire economy and underwrite AI agents for any one of those slots in the economy.

而且，如果我真的要从零开始，我可能会看着那个矩阵说，"好，有没有一个我非常熟悉的行业，我可以深入进去，真正逆向工程，看看那个行业的工作是什么样的，然后为它打造一个智能体？"

And, if I had to literally start from scratch, I'd probably be looking at that matrix and saying, "Okay, is there an industry I know extremely well, and that I can go in and really reverse engineer, what does the work look like in that industry and build an agent for that?"

杰瑞德·弗里德曼发过一条很棒的推文。

Jared Friedman had this great tweet.

他在Y Combinator，他基本上是说，"去做你想让智能体自动化的工作，成为做那项工作的专家。"

He's at Y Combinator, and he basically said, "Go do the job of what you want the agent to go and automate, and be the expert in how to do that job."

然后你就能非常得心应手地去创建一个代理来做那份工作。

And then you're going to be very potent at being able to go and create an agent to do that job.

所以，最理想的情况是你以前已经做过那份工作，然后你直接为你已经知道怎么做的工作构建一个代理。

And so, probably the best case scenario is you've already done that job in the past, and you just go and build an agent for the work you already know how to do.

次优的情况是，有一个联合创始人做过那份工作。

Second-best scenario, have a co-founder that has done that.

第三种最好的情况是，直接想办法去做那份工作，跟着别人学习，整天做那件事。

Third-best scenario, go literally find a way to do that, shadow somebody, do it all day long.

然后机会就在于，你能不能做一个生命科学FDA监管审查代理？

And then the opportunity is, can you make an agent that is the life sciences FDA regulatory review agent?

你能不能去做，我是法律审查案例代理？

Can you go do the, I am the legal review case agent?

再说一次，这个领域会有很多失败者，他们就是没有把握好规模。

Again, there's going to be a lot of flame outs, in this space that are, again, they just didn't nail the scale.

市场契合度不对。

It wasn't the right market fit.

市场规模不够大。

The TAM wasn't big enough.

这种情况会有很多。

There'll be lots of that.

但也会有很多惊喜，因为我们会发现，实际类别比我们想象的要多。现在经济中的供需平衡其实并不像我们想象的那样完美。如果有人能以原来十分之一的成本引入劳动力，那么对这种劳动力的需求实际上可能会是原来的50倍。

But then there's going to be a lot of surprises where it'll turn out that there's more categories than we think, where right now the supply-demand equilibrium in the economy is actually not as perfect as we would've imagined, where if somebody could have introduced labor at a tenth of the cost that it previously was priced at, that there would be actually 50x the demand for that particular form of labor.

没错。

That's right.

然后你会突然觉得，'好，我是一个法律审查代理。'

And so you'll all of a sudden be like, "Okay, I'm a legal review agent."

结果发现，传统上，法律审查的费用是每小时2000美元或1000美元，用于高端工作。

And it turns out that traditionally, legal review costs $2,000 per hour or $1,000 per hour for high-end work.

现在，代理可以以每小时5美元或10美元的价格完成这项工作，结果发现，我们实际上并没有拥有世界所需数量的律师。

Now, the agent can do that for $5 or $10 an hour, and it just turns out that we didn't actually have as many lawyers as the world needed.

只是因为他们太贵了，所以没人能负担得起。

It was just that they were so expensive, so nobody could actually afford them.

然后砰的一声，这类新工作就爆发了。

And then boom, there's an explosion of this new category of work.

现在大家都在申请专利。

Everybody's filing patents now.

每个人都会成为某种发明家。

Everybody is going to be an inventor of some sort.

我甚至能说出公司内部的一堆用例。

I can name- ... a bunch of use cases even internally.

我不打算——

I'm not going to-

……在这里透露太多创业点子，但是……

... give too many startup ideas out there, but- ...

我可以说出公司内部的一堆想法，如果某项功能的成本只有原来的十分之一，我们就会需要更多全职人员来承担这项功能。

I can name a bunch of ideas internally where if the cost of the particular function was a tenth of the price, we would have many more full-time people equivalent of that function.

只是现在效率太低，所以它总是排在我们预算规划的末尾，我们永远轮不到它。

It's just that it's too inefficient right now, and so it is always at the bottom of our stack ranking budget planning list, so we never get to it.

但如果是一个AI代理，价格能压缩到十分之一，那我们实际上就会在那个代理上花钱了。

But if it was an AI agent that could be compressed at a tenth of the price, then we actually would be spending on that agent.

而且很惊人——

And amazing-

因为没有上限。

Because there's no ceiling.

投资回报率没有上限，对吧？

There's no ceiling to the ROI, potentially, right?

投资回报率没有上限，而且起步也没有最低门槛。

There's no ceiling to the ROI, and there's no minimum floor price to get started.

大家都忽略了这一点。

People are missing this whole thing.

雇一个人要花10万、20万、30万美元，或者更多，取决于职位。

To hire a person costs $100,000, $200,000, $300,000 or whatever, depending on the function.

引入一个代理可能只要1000美元。

To bring in an agent might cost $1,000.

所以你会开始把它们用在那些你从来没雇过人来做的事情上。

And so you'll start using them in use cases that you never hired the person for.

然后这会突然减少你业务中的瓶颈，以你意想不到的方式，这又会促使你进一步投入这些领域。

And that will then, all of a sudden, it'll reduce the bottlenecks in your business in ways that you didn't anticipate, which then will cause you to actually go and fund those areas even more.

所以我觉得会有很多这样的机会。

So I think there's a lot of opportunity that will look like that.

但再说一次，对于一个行业、一个职位来说，我认为在那个大领域里有无限的机会。

But again, for an industry, for a job function, and I think there's endless opportunity in that general space.

我期待在这之后，突然会有大量二十多岁的年轻人，最终成为FDA监管的后台办公室的管理员。

I look forward to after this, all of a sudden, there's this giant surplus of young 20-somethings who end up as admins in FDA-regulated back offices.

那肯定会有的。

It's going to be there.

然后就像，"这些人都是从哪来的？"

And like, "Where did all these people come from?"

然后他们说，"哦，是Aaron让我来的。"

And they're like, "Oh, Aaron sent me."

好的。

Okay.

他们就在这里。

They're just here.

他们真的很兴奋能来这里工作。

They're real excited to go and work here.

我觉得——

I think-

我们真的很热爱我们的行业。

We really love our industry

……

...

我们可能需要的是一种配对机制，也许你们可以先启动。

what we probably need is this pairing mechanism between maybe you guys start it.

你是个很典型的CTO创始人，非常明显。

It's like you're a strong, you're very clearly central casting CTO founder.

你从MIT辍学，然后把他们和某个刚从辉瑞挖来的人配对，你会说：'你知道怎么做监管审查流程。'

You dropped out of MIT, and then you pair them with somebody who just like, you pluck them out of Pfizer, and you're like, "You know how to do regulatory review processes."

我们就是要配对。

We're just going to match.

然后这就像是一个配对服务，一边是知道怎么构建代理的人，另一边是工作流程的专家，我们就去创建100个这样的配对。

And then it's just a matchmaking service between the person that knows how to build agents and the experts in the workflow, and we just go create 100 of these.

嗯，我很好奇想听听这个。

Well, so I'm curious to hear about this.

因为Box是从消费者领域起步的，然后你们转向了企业领域。

Because Box started out in the consumer world, then you guys went to enterprise.

过去，连Andreessen Horowitz的马克自己都说过，他说：'真正伟大的创始人其实是从消费者领域起步的，因为做消费者领域太难了。

And in the past, even Mark himself over at Andreessen Horowitz once said, he's like, "Really great founders actually start out in consumer just because it's so hard to do consumer.

你必须构建产品。

You have to build products.

这非常困难。'

It's very difficult to do."

但在这个浪潮里，显然AI代理更侧重于企业领域。

But in this wave, obviously with AI agents, it's way more focused on enterprise.

我们正在转向，基本上就是往企业方向推进。

And we're making a tack where we push toward basically enterprise.

你怎么看待消费者市场？

How do you think about consumer?

你还觉得那是块试验田，还是说那完全是另一个世界，没人知道AI会怎么改变它？

Do you still think that that's a proving ground, or is it one of those things where it's like, hey, that's just a completely different world that no one knows how that's going to change as a result of AI?

我觉得消费者市场机会多多。

I think there's plenty of consumer opportunities.

我不会预先计划转型。

I don't think I would pre-plan a pivot.

所以就像——

So like-

那可不是计划好的，Aaron。

That was not planned, Aaron.

这其实就不太符合转型的定义了。

It kind of breaks the definition of a pivot.

没错。

That's right.

所以，好吧，如果六个月后我们要转型公司，那我们会先从一个不同的领域开始。

So, okay, in six months from now, we are going to pivot the company, but we're going to first start in a different space.

所以，我觉得如果人们真的需要转型，这个建议可能仍然很有用，但我认为消费者领域会有很多机会，而且——

So, I would say that it's probably still very good advice if people do need to pivot, but I think there's going to be plenty of consumer opportunities, and-

对。

Yep

……

...

我比较偏向企业领域，因为那是我花了很多时间的地方。

I'm partial to the enterprise because it's the space that I spend a lot of time in.

从市场规模来看，我基本上会押注AI对企业市场会是一个突破，因为企业市场的规模实在太大了。

I think by TAM, I would largely bet on AI being a breakthrough for the enterprise market by just sheer size of the TAM.

简单来说，AI本质上是一种生产力工具，而生产力通常在企业领域能产生显著的影响。

Simply because, AI is fundamentally a form of productivity, and productivity generally shows up in a meaningful way in the enterprise.

在消费领域，我们不太会考虑这个。

In the consumer space, it's not a thing that we think about as much.

所以我觉得大多数应用场景都偏向企业。

So I just think most of the use cases tend to lean enterprise.

有意思的是，Anthropic的Dario最近在一个播客里提出了一个观点，我觉得这是最贴切的思考方式，那就是——

And it's interesting, Dario from Anthropic had a point on a recent podcast that I think is the most salient way to think about it, which is-

如果Anthropic把AI模型从化学专业的本科生提升到研究生水平，消费者在Claude.ai上根本感觉不到。

... if Anthropic improves the AI model from basically being an undergrad in chemistry to a grad student in chemistry, the consumer won't notice in Claude.ai.

你在个人生活中不会问那种需要博士级别才能回答的化学问题。

There's no question you're going to ask about chemistry in your personal life that sort of is going to require that PhD level.

除非你有某些严重的健康问题。

Barring you have some deep healthcare related issue.

但博士级别的提升对辉瑞和生命科学行业来说，将是巨大的收益。

But the PhD jump is going to have massive gains for Pfizer and the life sciences industry.

所以，如果你想想我们现在在AI发展中的位置，我觉得我大部分……

And so just if you think about where we are in the progress of AI, I think most of my...

我就拿她来说吧，因为这正好能说明问题。

I'm going to just pick on her just because it kind of makes the point.

我妈用AI的场景就是——

My mom's AI use cases are like-

……已经完全解决了。

... they're fully solved.

我们搞定了。

We're done.

她不会——如果GPT-5是AI最后产出的东西，她也会觉得没问题，因为它已经——

She doesn't- She would be fine if GPT-5 was just the last thing that AI ever produced, because it's already-

……解决了所有问题。

... solving every problem.

因为她不会再问更多化学问题了。

Because she's not going to know more chemistry questions.

没有了……对。

There's no... Yeah.

她的化学问题已经问完了。

She's done with her chemistry questions.

她又不是要研发新的生命科学药物。

She's not trying to create a new life sciences drug.

就算在我个人生活里，比如育儿方面。

Even in my personal life, like for parenting, let's say.

我觉得AI现在停止发展也没关系，对我来说已经是巨大的正面影响了，我们还能给孩子们创作歌曲。

I think AI could stop right now, and it would already have been a huge net positive for me and, we'll make songs for the kids.

我们画画。

We draw things.

我们生成故事。

We get stories made.

对我六岁的孩子来说，故事不需要再好了。

I don't need stories to be any better for my six-year-old.