平衡 AI 领导力：布雷特·泰勒谈 OpenAI、Sierra 与生活

大家下午好。我们站在你们和欢乐时光之间，也站在漫长一天的结束之际。所以，感谢你们留下来。太棒了，这个房间几乎坐满了，这在下午晚些时候可不常见。嗯，我想你们很多人今天早上都在。我是 Aaron Katz。我和 Alex、Yuri 一起创办了 ClickHouse 公司。嗯，我非常高兴地向你们介绍 Bret Taylor。我觉得 Bret 不需要太多介绍。事实上，如果我把他的履历过一遍，可能会占掉我们分配的 45 分钟的大部分时间。但是，你们知道，他是过去 20 年科技界最具影响力的人物之一。我们都在用 Google Maps，这要感谢 Bret，因为他负责了那个产品的开发和发布。呃，Bret 曾是 Facebook 的首席技术官，这源于他创建的一家初创公司被收购，他还因 Facebook 的“点赞”按钮而广为人知。所以，开发一个全球数十亿人使用的功能，不是我们所有人都有机会做到的。嗯，Bret 创办了一家叫 Quip 的公司，2017 年被 Salesforce 收购。我在 Salesforce 待了 12 年，2014 年离开。呃，我 2002 年加入 Salesforce，同年他加入 Google。他从 2002 年到 2007 年在 Google。嗯，Salesforce 收购 Quip，很大程度上是因为 Mark 想收购 Bret。我不知道他对产品有多大兴趣，但别往心里去。

Good afternoon, everyone. We are standing between you and happy hour and the end of a really full day. So, thank you for staying with us. It's amazing the room is almost entirely full, which you often don't see late in the afternoon. Um, so I think many of you were here this morning. I'm Aaron Katz. I started the company ClickHouse with Alex and Yuri. Um, and I am thrilled to introduce you to Bret Taylor. I don't think Bret needs much of an introduction. In fact, if I went through his credentials, it would probably occupy most of the 45 minutes we have allocated. But, you know, one of the most influential figures in technology over the last 20 years. We all use Google Maps and we have Bret to thank for that as he oversaw the development and launch of that product. Uh, Bret was the chief technology officer of Facebook, due to an acquisition of a startup that he created, and is widely recognized for the like button in Facebook. So, developing a feature that billions of people around the world use is not something that all of us get to do. Um, Bret started a company called Quip that was acquired by Salesforce in 2017. I spent 12 years at Salesforce. I left in 2014. Uh, I joined Salesforce in '02 at the same year he joined Google. He was at Google from 2002 to 2007. Um, Salesforce acquired Quip largely because Mark wanted to acquire Bret. I don't know how much interest he had in the product, but don't take that personally.

我喜欢这个产品。我不……

I like the product. I don't

呃，而且 Bret 很快成为 Salesforce 的联合 CEO，那是 Bret 在任期间最具影响力的企业软件公司之一。然后最近，Bret 创办了一家叫 Sierra AI 的公司，刚刚完成了一轮定价融资，我想是 150 亿。呃，在很短的时间内，并且担任 OpenAI 的董事长。呃，在 Sam Altman 离开公司又迅速回归期间，他被任命为董事会主席。所以，我对你有太多问题要问，Bret，但首先，让我感谢你今天下午加入我们。

Uh, and Bret quickly became the co-CEO of Salesforce, one of the most consequential enterprise software companies while Bret was there. And then most recently, Bret started a company called Sierra AI, which just did a priced funding round, I think at 15 billion. Uh, in a very short period of time, and serves as the chairman of OpenAI. Uh, and was appointed to chairman of the board during the time when Sam Altman came out of the company and quickly came back into the company. And so, I have so many questions for you, Bret, but first off, let me just thank you for joining us this afternoon.

谢谢邀请。

Thanks for having me.

是啊。嗯，让我们从最明显的问题开始，你经营着一家高增长的初创公司，同时担任董事会主席，这家公司可以说是我们一生中不可思议的增长故事，而且也是一家非常复杂的公司。嗯，

Yeah. Um, let's start with the most obvious one, which you're running a high-growth startup and you're serving on the board as the chairman of arguably incredible growth stories of our lifetime, and also a very complicated company. Um,

我们容易有戏剧性。

We're prone to drama.

嗯，首先，祝贺诉讼的结果，我想。

Well, first off, congratulations on the outcome of the lawsuit, I guess.

谢谢。

Thank you.

这对 OpenAI 来说是个有利的结果。呃，我相信我们所有人都关注了。嗯，但也许谈谈你怎么……然后你有三个孩子，和我两个孩子差不多大，我知道这本身就是一份全职工作。那么，你如何在日常基础上平衡这一切，然后你如何划分这些对你时间的需求？

It's a favorable outcome for OpenAI. Uh, I'm sure all of us followed that. Um, but maybe talk about how do you... And then you've got three kids around the same age as my two kids, which I know can be a full-time job in and of itself. So, how do you balance all of this on a day-to-day basis, and then just how do you compartmentalize all of these demands for your time?

呃，我不会说我的生活特别平衡，先说明一下。拥有两份工作，你知道，这是一种生活选择。不过我不介意。我基本上没有爱好，所以我和家人在一起，然后工作。我对此很满意。呃，有一句 Alan Kay 的名言。Alan Kay 是 Xerox PARC 的研究员之一，他发明了很多后来成为个人电脑的东西，还有其他许多东西。他有句名言：“预测未来的最好方式就是创造未来。”这真的让我产生共鸣。我认为我首先是一名工程师，当一项技术让你早上醒来或晚上睡觉时都不由自主地思考它，我不知道我能不能不在这个领域工作。我就是需要，而且我会去做。我会免费做开源。我非常相信资本主义的力量，它能激励一群人努力解决人类一些最困难的问题。所以，呃，我想说选择在这些领域工作，我认为这是一种特权。只是……我对 Sierra 的成功非常感激。我认为 OpenAI 的使命，即确保通用人工智能造福人类，是当今世界上最重要的使命。所以，嗯，这是一种特权。嗯，我实际上怎么做呢，就是很多自律。你知道，我起得很早，然后试着在早上 8 点之前处理 OpenAI 的事情。呃，比如 Sara Fryer 很多早上 5:30 就给我打电话。所以，我说：“嘿，Sara。”她就像……我们都在 Peloton 上或者什么的。这样就行。你骑车的时候还能谈财务。嗯，所以就是非常严格地分配我的时间。嗯，但这是特权，我临时担任主席，那已经是两年半前了。呃，所以也许我不认为这会永久，但呃，我很高兴能参与其中。

Uh, I wouldn't say my life is particularly balanced, just to be clear. It's sort of a life choice to have, you know, to have two jobs. I don't mind, though. I basically don't have hobbies, so I spend time with my family and I work. And I'm pretty happy with that. Uh, there's an Alan Kay quote. Alan Kay was one of the researchers at Xerox PARC who invented a lot of what became the personal computer among many other things. And he has this quote, "The best way to predict the future is to invent it." And it's something that really resonates with me. I think I'm an engineer first and when a technology is something that you can't help but think about when you wake up in the morning or you're going to sleep at night. I don't know if I couldn't not work in the space. I would just need to and I would do it. I would do open source for free. I'm a huge believer in the power of capitalism to inspire a group of people to work to solve some of the hardest problems in humanity. So, uh, I would say like the choice to work in these areas, I've considered it a privilege. Just to... I'm so grateful for the success of Sierra. I think the mission of OpenAI, which is to ensure that artificial general intelligence benefits humanity, is the most important mission in the world right now. So, um, it's a privilege. Um, how I do it practically is a lot of discipline. You know, I wake up really early and then try to do OpenAI stuff kind of before 8:00 a.m. Uh, like Sara Fryer calls me at like 5:30 a lot of mornings. So, I'm like, "Hey, Sara." It's like she's like we're both at our Peloton or whatever. Like, it works out. You can still talk about finance while you're biking. Um, and so just be really regimented about just how I allocate my time. Um, but it's a privilege and I took the job of chair temporarily and that was like 2 and 1/2 years ago. Uh, so maybe I don't think it'll be permanent, but uh, I'm really happy to be a part of it.

太棒了。你不得不和很多大人物同台。嗯，你知道，我和 Mark 共事了 12 年，现在和 Sam Altman。呃，你如何处理这些关系，确保你能完成你的议程，同时也要迎合那些我确信是非常大的个性和自我？

Awesome. You've had to share the stage with a lot of pretty big personalities. Um, you know, I worked with Mark for 12 years and now with Sam Altman. Uh, how do you navigate those relationships to make sure that you can get your agenda accomplished while also catering to what I'm sure are some very big personalities and egos?

我以前讲过这个故事，但我职业生涯中最感激的导师之一是 Sheryl Sandberg。所以，当 Facebook 收购我的公司 FriendFeed 时，我们是一家 12 人的公司，而我在 Google 时有点像中层人员，发布了 Google Maps，我为此非常自豪，但我没有管理过大型团队，你知道。Google Maps 发布时只有五个人，你知道。所以，只是较小的团队。当我成为 Facebook 的 CTO 时，我最终管理了一个超过一千人的团队。我不记得具体是多少，但这是我第一次真正管理一个非常大的团队，而且我犯了你能犯的每一个新手错误。呃，我是一个相当聪明的人，也是一个相当有同理心的人，但当时 Facebook 并没有真正……当时 Facebook 的中位年龄大概是 23 岁，而且并没有真正的基础设施来支持你完成这种转变，除了 Sheryl。所以，我遇到了这种独特的时刻。嗯，我很乐意谈细节，但我就是有点搞砸了，她把我拉到一边，带进一个会议室，然后把我训了一顿，基本上说：“你搞砸了。嗯，原谅我的语言。我不知道我能不能在这里那样说话。”

I've told this story before, but one of the mentors I've been most grateful for in my career is Sheryl Sandberg. So, when Facebook acquired my company, FriendFeed, we were a 12-person company, and I had been kind of a mid-level person at Google, launched Google Maps, which I'm really proud of, but I didn't manage big teams, you know. It was Google Maps when we launched had five people, you know. So, it was just smaller teams. When I became CTO of Facebook, I ended up with a team of over a thousand. I don't remember what it was, but it was the first time I was really managing a very large group of people, and I was making every rookie mistake you could make. Uh, I'm a reasonably intelligent person. I'm a reasonably empathetic person, but there wasn't really a Facebook at the time was like the median age was like 23, and there wasn't really sort of an infrastructure to support you in making that transition, except for Sheryl. So, I ended up with this kind of unique moment. Well, I'm happy to talk to the details, but I just sort of was like had sort of stepped in it, and she took me aside, brought me to a conference room, and kind of lit me up, and just sort of said, like, "You're it up. Um, excuse my language. I don't know if I'm allowed to talk that way here."

你可以说。

You're good.

基本上，她给了我很多逆耳忠言，以一种关心你的人才能做到的方式，然后我在接下来的 12 个小时里经历了悲伤的所有阶段。你知道，那是什么感觉？否认、悲伤、愤怒。我全都经历了。

And basically, just had a... she gave me like a lot of hard truths in a way that like someone who cares about you can, and then I went through every stage of grief over the next 12 hours. You know, what is it like? Denial, sadness, anger. I went through all of it.

然后第二天早上我醒来，心想：她基本是对的。我意识到，我每天醒来都在想我是谁，我是个工程师，我之所以被放到这个位置，是因为我的技术能力和产品直觉。所以我就打算把时间花在这上面，因为那才是我被任命为 CTO 的原因。但我发现，实际上 Facebook 需要的、以及我需要做的让 Facebook 更成功的事情，往往是别的东西。招聘，当时我们在平台上有大量的合作复杂性，所以有很多合作方面的工作。甚至像政府关系、品牌，所有这些都和我对自己身份的认知毫无关系。而谢丽尔的反馈，某种程度上让我意识到了这一点。所以，我在那一刻决定，我每天醒来就只问自己：“今天我能做的最有影响力的事情是什么，来真正让 Facebook 更成功，并且不那么执着于自己的身份认同。”就像，我是谁并不重要，我只去做最重要的事。然后由此产生的正向反馈循环，结果真的非常惊人。我就是更擅长我的工作了，而且我发现……这听起来有点显而易见或老生常谈，但做好自己的工作感觉真的很好，你知道吗？你会得到所有这些正面反馈，然后影响力就像……我们开始让事情运转得更好了。所以，我最终对没有强烈的自我认同感这件事上了瘾，因为我就想：“这太棒了。如果我每天醒来只做最重要的事，好事就会发生。这太神奇了。”所以，回到你的问题，我发现和非常强势的人合作相对容易，因为他们通常在自己领域里非常出色。

And then I woke up the next morning, and I was like, she's basically right. And I realized for myself what I was doing was I was waking up every day and thinking about who I was and I'm an engineer and I was put in this role for my technical ability and my product sense. So, that's what I'm going to spend time on because that's why I was made CTO. And what I realized is that actually the things that Facebook needed and the things that I needed to do to make Facebook more successful were often other things. Hiring, we had a lot of partnership complexity on the platform at the time, so a lot of partnership work. Even things like government relations, brand, all these things that had nothing to do with my own conception of my own identity. And what Sheryl's feedback was sort of implicitly made me come to realize that. So, I decided at that moment I was just going to wake up every day and say, "What is the most impactful thing I can do today to actually make Facebook more successful and be less wedded to my own sense of identity." Like, it doesn't matter who I am, I'm just going to work on what's most important. And then the positive feedback loop from that turned out to be really amazing. I just was better at my job and it turned out I didn't... It sort of sounds obvious or trite, but like it feels good to be good at your job, you know? Like, you get all these positive feedback and then the impact like we just started like things started to work better. So, I ended up sort of hooked on not having a strong sense of my own identity just because I was like, "This is great. If I just could wake up every day and work on what's the most important thing, good stuff happens. This is amazing." So, then going back to your question, I've found it relatively easy to work with really strong-willed people because usually they're pretty remarkable in their own way.

我认为除了马克·贝尼奥夫，没有人能让 20 万人出现在一个企业软件大会上。这就像某种宗教复兴。没人能理解。你提到 Salesforce，大多数人都会提到 Dreamforce。人们在那边做什么？没人真正知道。实际上，这已经持续很久了，就是这样。有摇滚音乐会。我想也有一些软件在卖，谁知道呢？我们甚至不知道。而对我来说，因为有了更灵活的身份认同，我就想：你需要什么？我来做。结果，我发现我非常喜欢和强势的人合作……我的意思是，不是因为强势的性格，而是我喜欢和真正有才华的人一起工作。我爱伟大的工程师，我爱伟大的设计师，如果有人想抢走房间里的氧气，我也不会介意，因为我不再需要它了，这我要感谢谢丽尔。

I think there's no one other than Marc Benioff who could get 200,000 people to show up to an enterprise software conference. It's like a religious revival or something. And no one understands it. You talk if when you mention Salesforce, most people mention Dreamforce. Like, what do people do there? Like, no one really knows. Actually, this has been going on and it's just it's this is there. There's rock concerts. I think some software's being sold, who's to say? We don't even know. And for me, by having a more flexible sense of my own identity, I was like what do you need? I'll do it. And as a consequence, I found I really enjoy partnering with people with strong... I mean, it's not for the strong personality, but like I enjoy working with really talented people. And I love great engineers, I love great designers, and it doesn't bother me if there's no oxygen you can take out of the room I'm in because I don't need it anymore, and I thank Sheryl for that.

好的。在硅谷，你经常听到人们被描述为“连续创业者”。大多数时候，人们认为这是赞美。你对此有不同的看法。你愿意和大家分享一下吗？

All right. Often times you hear people in Silicon Valley be described as a serial entrepreneur. And most of the time people perceive that to be complimentary. You've got a slightly different take on that. You want to share that with the group?

是的，马克·扎克伯格不是连续创业者，因为他的公司成功了。所以，你知道，这很有趣。我有……对于你们中那些当过创业者的人来说，如果你卖掉你的公司，你会期望人们真的很棒。比如，“恭喜”。但没有人创办一家公司，或者我认为大多数严肃的创业者不会为了想“我创办这个是为了把它卖给另一家公司，最终它会关闭它，这样我就能成为另一家公司中层管理者”而创业。这显然不是建立收购公司的方式。但这正是很多收购中发生的事情。所以，我为我参与过的产品感到自豪，但我也认为，你知道，创业者们的使命是创造一家价值超越你而存续的公司。我想，在某种程度上，成功的定义是你真正看到一家公司实际上比你活得更久。能做到这一点的科技公司并不多。IBM、微软肯定会做到。我认为谷歌也会。但其他很多公司，即使是像 Sun Microsystems 这样的大公司，甚至像雅虎，现在多亏了一位出色的新 CEO 而有点复苏，但科技公司的默认状态是被颠覆和死亡。然后，通过收购或成为连续创业者，意味着你多次没有实现这一点。所以，我为我所取得的成就感到自豪，但这是我诚实的看法。我已经失败过两次了。第三次会成功。

Yeah, Mark Zuckerberg is not a serial entrepreneur because his company worked. So, I you know, it's interesting. I have... for those of you who've been entrepreneurs, it's interesting if you sell your company or something what you'd expect people are really great. Like, "Congratulations." But no one starts a company, or I think most serious entrepreneurs don't start a company to think, "I'm starting this so I can sell it to another company that will eventually like shut it down so I can become a middle manager at another company." That's not exactly how you build acquiring companies, obviously. But that's sort of what happens with a lot of these acquisitions. And so, I'm proud of the products I've worked on, but I also think that, you know, entrepreneurs like their mission is to create a company whose value endures beyond you. I would like to think like the to some degree the definition of success is you actually see a company that literally outlives you. There's not that many technology companies that do. IBM, Microsoft, will certainly meet that. I think Google will. But then many others, even big ones like Sun Microsystems, all the way through to even, you know, Yahoo sort of coming back to life now thanks to a great new CEO, but the default state of a technology company is disrupted and dead. And then with an acquisition or serial entrepreneur, it means that you didn't achieve that multiple times. So, I'm proud of what I've accomplished but that is my honest perspective. I've failed twice. Third time's the charm.

好的。那我们谈谈这个。你和联合创始人 Clay 创办了 Sierra 公司。我们对这家公司非常了解，因为你们在后端使用了 ClickHouse Cloud。而且我们对这个类别也很了解，因为和我们在 Salesforce 一起做的事情有些重叠。为什么这次会不同？我听说你的团队把你们所处的类别描述为他们参与过的最激烈竞争的市场机会。意思是你们不是轻松就能进入客户的。要赢得这些客户，完全是一场白刃战。那么，为什么你会在这个非常拥挤的类别中成为赢家？为什么这有潜力成为一家世代相传的公司？

All right. Well, let's talk about that. So, you started the company Sierra with your co-founder Clay. We've gotten to know that company very well because you use ClickHouse Cloud on the back end. And then we know the category quite well because there's a bit of overlap with what we did at Salesforce together. Why is this going to be different? I've heard your team describe the category that you're in as the most fiercely competitive market opportunity they've ever been part of. Meaning you're just not getting walked into accounts. It's just a complete knife fight for you to win these logos. So, why are you emerging as the winner in the category that's very crowded? And why does this have the potential to be a generational company?

首先，我们是 ClickHouse 非常非常满意的用户，而且我们是一个非常有鉴别力的团队。我们不用很多第三方软件。所以，我真的很高兴被邀请来这里演讲，因为我们大部分后端都运行在它上面。所以，这很了不起。我们就是非常真诚地喜爱这个平台。回答你关于 Sierra 的问题，我们是帮助公司为他们的客户体验打造 AI 智能体的领先平台。这意味着，如果你打电话给诺德斯特龙，你会和一个叫 Nora 的 AI 智能体通话，而不是按 2 键转服务。或者，如果你想在 Sirius XM 订阅上获得更低的价格，你可以打电话和他们的 AI 智能体 Harmony 交谈，协商一个更好的费率，Harmony 会和你协商，希望能为你找到你非常看重的东西。我们的平台支撑着数十亿通电话和数十亿次聊天，从像 Cigna 这样的医疗保健公司，到像 SiriusXM 或 Singtel 这样的电信公司，再到世界各地的银行。这是一个非常令人兴奋的市场。我会说，它可能是仅次于编码的第二大应用 AI 市场。而且我认为编码……我甚至不知道那个市场有多大，但可以说有数万亿。再往下，就是客户体验。而这个市场之所以大，原因非常简单。

First, we are very very happy users of ClickHouse and a very discerning team. We don't use a lot of third-party software. So, really I was happy to be invited to talk here just because we run most of our back end on it. So, it's remarkable. So, just like quite authentically love the platform. To answer your question about Sierra, so we are the leading platform that helps companies make AI agents for their customer experience. So, that means if you called Nordstrom on the phone, you'll talk to an AI agent named Nora rather than press two for service. Or if you want a lower price on your Sirius XM subscription, you can call and talk to Harmony, who's their AI agent, and negotiate a better rate on your plan, and Harmony will negotiate with you and hopefully find you something that you find a lot of value in. We power billions of phone calls and billions of chats on our platform from healthcare companies like Cigna to telecommunications companies like SiriusXM or Singtel to banks around the world. And it's a really exciting market. I would say it's probably the second largest applied AI market after coding. And I think coding and I don't even know how big that market is, but call it trillions. Go way down, and then there's customer experience. And the reason for it is really simple why the market's big.

消费者不喜欢排队等待，即使你终于接通了电话，也很少能有愉快的体验。通常，接电话的是个承包商，对你的产品组合或账户几乎没什么了解，他们只是在一个你无权操作的屏幕上打字。所以，这既慢又低效。而且，那通电话的全部成本，包括人力和软件成本，通常超过 10 美元，因为维持整个运营很昂贵。因此，几乎不可能联系到任何消费品牌。我开过这个玩笑，但这是 100% 真实的：你们所有人联系桑达尔·皮查伊，都比联系谷歌客服更容易。我觉得联系谷歌助手根本不可能。所以，从定义上讲，桑达尔有电话，所以有一条理论上的路径可以联系到他，而联系客服则完全没有理论路径。原因很简单：谷歌的每客户平均收入低于一通电话的成本。这取决于，你知道，不是谷歌云，而是谷歌消费者业务。所以，实际上，对于消费公司来说，你根本负担不起联系他们的费用。这就是为什么你能联系到的公司，比如有线电视公司，往往是高客户终身价值的服务，但因为每次联系他们都会降低他们的利润率，所以他们必须让联系变得困难。

Consumers don't like waiting on hold, and even once you get off hold, you rarely have a delightful experience. Usually, it's someone who's a contractor with very little context on the product portfolio or your account, and they're just typing into a screen that you don't have the privilege to type into. So it's slow and inefficient. And typically, the price of that phone call, all in with labor and software costs, is greater than $10, just because it's expensive to power that whole operation. As a consequence, it's almost impossible to talk to any consumer brand. I've made this joke, but it's 100% true: it would be easier for all of you to talk to Sundar Pichai than it would be to talk to Google customer support. I don't think it's possible to talk to Google assistant. So definitionally, I think Sundar has a phone, so there's a theoretical path to him, and no theoretical path to the alternative. The reason is simple: Google's average revenue per customer is lower than the price of a single phone call. Depending on, you know, not Google Cloud, but just Google consumer. So literally, for consumer companies, you could not afford to talk to them. And it's why the companies you can talk to, like your cable company, tend to be very high LTV services, but because every time you talk to them it's essentially bringing down their margins, by definition, they have to make it hard.

AI 首次将这一渠道数字化了。所以，我们——虽然还没到网页浏览的成本，但电话成本正在下降，不是下降一个数量级，而是两个数量级。所以这有点像互联网——你不只是把以前做的事情数字化，你实际上是在让某样东西从成本中心变成你负担得起的事情。所以现在，突然间，如果你经营一家 B2C 公司，你第一次能负担得起与客户交谈了。我认为这是一个巨大的变革，就像网站首次对银行一样重要。比如，像查余额这样简单的事情，突然变得免费，而以前这很繁琐。所以我认为我们正处于起点，就像 1995 年的网站建设一样，关于大规模降低与客户个人互动成本的意义。我认为这将产生与互联网一样重大的影响，我真心相信这一点。因此，花在客户服务上的数千亿美元正在转移到智能体上，而且市场正在扩大，因为杰文斯悖论，就像当某样东西从 10 美元降到 50 美分或 10 美分时，你会做更多。它会开辟新的市场。

AI has digitized that channel for the first time. So we've—it's not quite the cost of a page view yet, but those costs of a phone call are going down by not one, but two orders of magnitude. So it's sort of like the internet—it's not like you're just digitizing what you did before, you're actually making something go from a cost center to something you can afford to do. So now, all of a sudden, if you run a particularly B2C company, for the first time ever, you can afford to talk to your customers. And I think that's a really humongous deal, as significant as a website was to a bank for the first time. Like, just simple stuff like checking your balance suddenly became free in a way that was quite tedious before. So I think we're at the start, just like the 1995 of website building, of what it means to bring down the cost of having personal engagements with customers at scale. I think we'll have as significant an impact as the internet, and I truly believe that. As a consequence, the hundreds of billions of dollars that have been spent on customer service are moving to agents, and it's expanding the market because of the Jevons paradox thing, which is like when something that costs $10 suddenly costs 50 cents or 10 cents, you'll do a lot more of it. It'll just open up new markets.

每个应用 AI 市场都极具竞争力。这有点像，回到 1995 年的比喻，每个人都知道电子商务会很重要，每个人都知道搜索会很重要，每个人都知道。有一些新颖的东西，比如 eBay，但大多数市场在最终主导者出现之前就已经为人所知。比如，谷歌成立于 1998 年，搜索在那之前就是一个大类别，问题不是‘嗯，搜索所有网页有用吗？’当然有用。问题是商业模式，比如 AdWords，以及谁能在那个市场赢得胜利的权利。所以，这就是我们现在在应用 AI 领域所处的境地。在编程领域，你有 Codex，这显然是我最喜欢的。你有 Claude Code，你有 Cognition，你有 Cursor，在我看来，这些都是相当高质量的公司。在客户体验领域，我们是遥遥领先的领导者，但我们有很多高质量的竞争对手。然后还有其他一些领域，比如法律，你有 Harvey 和 Legora。我的观点是，在每一个这样的市场中，需求如此之大，以至于你最终会进入增长非常快、竞争非常激烈的公司。

Every applied AI market is extremely competitive. It's a little bit like, going back to the metaphor of 1995, everyone knew that e-commerce would be a big deal, everyone knew that search would be a big deal, everyone knew. There were some novel things like eBay, but most of the markets were known before the eventually dominant player was created. Like, Google was started in 1998, search was a big category before that, and it wasn't a question of 'Hmm, is searching all the web pages useful?' Of course it is. It was a question of business model, things like AdWords, and it was a question of who would earn the right to win in that market. So that's what we're in in applied AI right now. In coding you have Codex, which is obviously my favorite. You have Claude Code, you have Cognition, you have Cursor, all of which are pretty high-quality companies in my opinion. In customer experience we're the leader by a fairly large margin, but there are a lot of very high-quality competitors. And then you have a few others like legal where you have Harvey and Legora. My view is that in each of these markets there's so much demand that you end up in very fast-growing companies with very intense competition.

我们为什么赢？我认为我们有最好的产品。虽然这就像——我想每个人可能都会这么说自己的产品。

Why we're winning? I think we have the best product. Though that's like—I think everyone would say that probably about their product.

我们的产品其实很烂。不，我开玩笑的。

Ours sucks actually. No, I'm just kidding.

我们有一个非常刻意的合作伙伴战略。我们有一个排行榜，记录客户从我们开始接触他们到上线所需的天数，值得注意的是，这是以天为单位衡量的，而不是以周或月。举个例子，诺德斯特龙，从我们开始到他们的电话线 100% 上线，只用了 35 天。而且我们与财富 50 强中超过 40% 的公司合作。我们与信诺在 6 周内上线，举个例子。所以，优秀的产品和能让财富 100 强快速成功的合作伙伴模式的结合，是非常非常独特的。所以这真的是我们的秘诀。哦不，我告诉所有人了。我开玩笑的。我们的增长速度是第二大竞争对手的两倍多。所以我希望，凭借这种势头，随着时间的推移，这些市场会出现自然的硬币效应。但地面战斗是激烈的，这毫无疑问。但我认为这是一种特权，因为卖过更难卖的软件，我宁愿要一个竞争激烈、需求旺盛的环境，也不愿独自在一个没人关心的市场里，我也经历过那种情况。所以，我任何时候都会选择前者。

We have a very intentional partnership strategy. We have a leaderboard of the number of days it takes for our customers to go live from the moment we start talking to them, and it's notable it's measured in days, not in weeks or months. Just as an example, Nordstrom, from the time we started to 100% of their phone line, was 35 days. And we work with over 40% of the Fortune 50. We went live with Cigna in 6 weeks as an example. So the combination of a great product and a partnership model that enables the Fortune 100 to actually be successful quickly is very, very unique. And so that's really been our secret sauce. Oh no, I told everyone. I'm just kidding. We're growing more than twice as fast as the next biggest competitor. So I'm hoping that with that momentum, over time there'll be the natural coin effect that happens in these markets. But the ground game is intense, there's just no doubt. But I think it's a privilege because having sold software that's a little harder to sell, I'll take a competitive high-demand environment over being alone in a market that no one cares about, which I've also experienced. So I'll take it any day of the week.

呃，所以，这个问题——我并不是想让它有争议或挑衅，但是呃，你知道——

Uh so, this question—I it's not intended to be controversial or provocative, but uh you know—

很少有争议性不大的问题会以这种话开头。

Rarely is an uncontroversial question preceded by that.

呃，好吧，我们得让观众保持参与。嗯，所以，现在全国各地，据报道 AI 正在产生负面——你知道，全国 70% 到 75% 的人对 AI 有负面联想。呃，原因有很多，其中之一是失业。所以，你所描述的，自然引出了你正在取代客户支持代表，并将这些工作从行业中移除的概念。嗯，然后，显然，还有所有围绕 OpenAI 的担忧。所以，你——我觉得没有人比你更能近距离看到 AI 在硅谷这个小圈子之外是如何被构建和接受的。你有什么想分享的观点吗？

Uh well, we got to keep the audience engaged. Um and so, right now across the country, it's reported that AI is pulling negative—you know, 70 to 75% of the country has a negative association with AI. Uh for a variety of different reasons, one of which is job displacement. And so, what you're describing naturally leads to the concept of you're taking customer support reps and you're removing those jobs from the industry. Um and then, obviously, all of the concern around OpenAI. So, you—I feel like nobody's got a better front row seat to how AI has been essentially framed and received outside of this small microcosm of Silicon Valley. Do you have a point of view on that you'd want to share?

我有。嗯，我对 AI 领袖们总体上谈论 AI 的方式感到非常沮丧。这可能有点简化，但大致形式是：我们可能都会死，但我们肯定都会失业，而我正因此致富。你会想，你到底在干什么？如果你这么想，为什么还要做这个？这毫无意义。

I do. Um, I've been really frustrated with how AI leaders writ large have been talking about AI. This is maybe a reductive, but it's sort of a rough view of the form is: we may all die, but we're definitely all going to lose our jobs, and I'm getting rich on it. And you're like, what the hell are you doing? Like, if that's what you think, why are you working on it? It makes no sense.

我觉得这其实不是很多人的真实想法，但这是他们嘴上说的。所以，并不是媒体在唱衰 AI。不，媒体只是在报道人们说的话，大致就是这样。如果我不是身处这个圈子中心的人，我也会想，那我们为什么还要做这个？

I don't think that's actually what a lot of people think, but that's what they're saying. And so, it's not even like the press is down on AI. Like, no, the press is just reporting quotes of what people are saying, and it's roughly that. And if I were like just someone who wasn't in the center of this, I would also think like, why are we doing this then?

那么让我从我所认为正在产生的价值说起。我觉得，当你走进一个房间打开开关时，很难去思考电的价值。但我们正在让许多今天稀缺且极其昂贵的东西变得充裕。总的来说，我们正在让专业知识和智能变得充裕。我想这意味着——我就从一件事说起。我们有孩子。辅导可能是世界上最奢侈的事情之一。谁能负担得起给孩子请家教？只有富人，只有富人。现在，世界上每个孩子，无论说什么语言，都能拥有一个私人导师。无论他们的学习风格如何，无论他们的节奏如何，这是多么大的民主化啊。

And so let me start with just like what I believe to be the value that's being produced. And I think it's hard to think about like the value of electricity when you just walk in a room and turn on a switch. But we're making a lot of things that are scarce today that are extremely valuable abundant. In general, we're making expertise and intelligence abundant. And I think that means I'll just start with one thing. We have kids. Tutoring was probably one of the most exclusive things in the world. Like, who can afford having a tutor for their kids? It's the wealthy, exclusively the wealthy. Now every child in the world in every language can have a personal tutor. No matter what their learning style, no matter their pace, how democratizing is that?

医疗建议。每次我拿到化验结果，在和我主治医生谈之前，我都会先上传到 ChatGPT。他甚至都不觉得烦，至少他是这么告诉我的。这不像上 WebMD，你上去之后会想，我要死了，我什么病都有，这是每个人上 WebMD 时的想法。实际上，我只是问它我应该问什么问题，然后我带着更多信息去就诊。我有几个家人正在应对癌症和其他疾病。像这样，用一个超级智能体来应对这些，真是太棒了。因为以前这太吓人了。你甚至不知道术语，不知道该问什么问题。

Health care advice. Every time I get a lab result, I upload it to ChatGPT before I talk to my primary care physician. And he's not even annoyed by it, or at least that's what he tells me. And it's because it's not like going to WebMD where you go and you're like, I'm going to die, I have every disease, which is what everyone thought when they went to WebMD. It's actually I'm just I ask it what questions should I ask, and I come in more informed. I have a few family members dealing with cancer and other things. Like, navigating that with a super intelligent agent is so great. Just because it was so intimidating before. You don't even know the nomenclature, you don't know what questions to ask.

财务建议。想想那些年长的家庭成员，那个沉迷快速致富计划的表亲。我们家里都有这样的人，对吧？现在，每个人都能获得最高质量的财务建议。这些东西通常——基本上，如果你在这方面有天赋，你倾向于服务更富有的客户。现在，我们正在让这些变得人人可得。所有这些我认为通常只属于全球前 1% 或前 0.1% 的东西，现在将广泛提供给全世界。下一个拉马努金将能获得世界上最高质量的数学陪练伙伴。他不需要通过一封信去发现某个不知名的英国教授。

Financial advice. Just think about like the elderly family member, the cousin who's on the get-rich-quick schemes. Like, we all have those people in our families, right? We will now have like everyone will have access to the highest quality financial advice. The stuff that's typically society is just sort of like basically if you're talented in that, you tend to serve wealthier clients. Like, now we're democratizing access to that. All of these things that I think generally go to the top 1% of the world or top 0.1% of the world will now be broadly available for the world. The next Ramanujan will have access to the highest quality, you know, math sparring partner in the world. He doesn't need to discover some obscure British professor to through a letter.

我认为这将是一股令人难以置信的民主化力量。除此之外，我们还将有科学发现。也许我们会在聚变领域取得突破。也许我们会找到治愈阿尔茨海默病的方法。也许我们会找到，你知道，所有困扰我们这个社会、实际上拉低预期寿命的东西。我认为我们有一条真实可信的路径去取得真正的突破。我认为这非常强大。

And I think this will be an incredibly democratizing force. On top of that, we'll have scientific discoveries. Maybe we'll find breakthroughs in fusion. Maybe we'll find a cure for Alzheimer's. Maybe we'll find, you know, all the things that vex us in this society that actually bring down life expectancy. I think we have a real credible path to find real breakthroughs. And I think that's incredibly powerful.

同样，我是一个坚定的资本主义者。我真心相信资本主义在面对技术变革时的力量。在人类历史上，我们从未有过一项技术取代了一个工作，而另一边的工作却更少。如果在 1776 年，这个国家 90% 多的人口是农业人口——我猜超过 95%——他们是农民。我们所有时间都花在种粮食上，这样我们才能吃。如果你把本杰明·富兰克林带到这里，让他坐在这儿，那会挺酷的，顺便说一句。

Similarly, I am a card-carrying capitalist. I really believe in the power of capitalism in the face of technology change. And we have never in the history of humanity had a technology that displaced one job and had fewer jobs on the other side. If in 1776, 90-something percent of the country was agrarian—I assume it was like over 95%—they were farmers. And we spent all of our time like growing food so we could eat it. If you brought Benjamin Franklin and sat him right here, that'd be pretty cool, by the way.

我打赌我——我们要花多长时间才能向他解释清楚我们是怎么获得食物的？

I bet I—how long would it take us to just explain to him how we get our food?

那大概要花一天时间，他还是不明白。你们都是靠什么谋生的？我不明白你们盯着发光的屏幕。他不会明白的。我们已经取代了农业，现在农业占劳动力的多少？大概 3% 左右。我们经历了工业革命，经历了全球化和服务业经济。我们向银行引入了自动取款机。尽管出纳员已成为过去，但银行网点仍然更多，网点里的人更多。我坚信，我们缺乏想象力，无法想象在这项技术面前我们将要做的一切。但认为技术存在后我们就坐享其成，认为我们已经做完了所有能想到的事情，我觉得这很可笑。我会说，你对人性没有信心。几百年来我们一直在这样做。

It would probably take like a day and he still wouldn't get it. What do you all do for a living? I don't get you staring at glowing screens. You would have no idea. We've displaced farming, which is what? 3% of the workforce right now or something like that. We went through an industrial revolution, went through globalization and the services economy. We introduced the automated teller machine to banks. There's still more bank branches and more people in those bank branches despite tellers being a thing of the past. I firmly believe we lack the creativity to imagine all the things that we will do in the face of this technology. But the idea that it will exist and we'll just sit around like done. We've done all the things that we can ever think of doing I find laughable. I'm like you have no faith in humanity. We've been doing this for hundreds of years.

我觉得这很吓人，因为，你知道，如果你看看工业革命时期卢德分子之类的早期故事，当机器变得比人类更强大时，或者如果你像我一样认同自己是一名程序员，一名计算机程序员，而我认为这个职业在未来不会以那种形式有价值。这就像，我不知道。我对此感到不安。就像我必须改变才能在这个新世界中实现自我。但认为事物会消失，这是一种零和的世界观，我根本不相信。所以，我坚信这将真正造福人类。

I think it is scary because, you know, if you look at sort of the early stories of Luddites in the Industrial Revolution and other things, when machines became more powerful than humans, or if you identify like I do as a programmer, a computer programmer, and that is not a profession that I think will be in that form valuable in the future. It's like I don't know. I feel insecure about it. It's like I have to change to be self-actualized in this new world. But the idea that things will go away is such a sort of fixed sum view of the world that I just don't believe in. And so I, my firm belief is that this will really benefit humanity.

但我认为我们必须重视安全。我认为这正是 Anthropic 和 OpenAI 大部分对齐研究的方向，这至关重要。这有点像 AI 的希波克拉底誓言。但总的来说，我不同意这个领域许多人的那种极度悲观。我认为这可能源于这样一个事实：许多从事 AI 工作的人只做过 AI，而这当然是一个会被 AI 自动化的职业。因此，你可能很难对在市场街开花店的人产生同理心。或者，天哪，我想湿实验室要完全自动化还需要一段时间。哦，那药物审批流程呢？嗯，那不会真正受到 AI 的影响。然后你把这些都加起来，你会意识到这个世界上有很多原子，不是所有东西都能靠按一下按钮解决。所以我非常乐观。我认为我们需要改变这种叙事。

I think it's very important we focus on safety though. I think that's where most of the alignment research at Anthropic and OpenAI is, it's vitally important. It's sort of the Hippocratic oath of AI. But I think broadly speaking I don't agree with the abject pessimism from a lot of folks in the space. And I think it might come from the fact that many people working in AI have only worked on AI, and that is certainly one profession that will be automated with AI. And as a consequence, maybe you have a hard time having empathy for someone who runs a flower shop down on Market Street. Or gosh, I guess it will take a little while for that wet lab to be automated. Oh, and what about the drug approval process? Well, that's not really going to be impacted by AI. And then you all start adding it all up and you realize that there's a lot of atoms in this world and not everything can be solved with the push of a button. So I'm very optimistic. I think we need to change the narrative on it.

我还认为公司需要为负面的二阶效应承担责任。我确实认为，人们因为数据中心导致电费上涨而感到不满是合理的。我认为公司应该承担这部分成本。我不认为你——考虑到这项技术创造的财富量，公司应该主动承诺去消除负面的二阶效应。但总的来说，我也认为叙事上存在重大问题。

I also think companies need to take accountability for the negative second order effects. I do think it's reasonable for people to be upset that their power bill goes up because of a data center. I think companies should eat that cost. I don't think you—I think given the amount of wealth being created by this technology, companies should sign up to actually actively remove the negative second order effects. But broadly speaking, I also think there's significant narrative problems.

我同意。我今天早些时候和观众分享过，我们公司 ClickHouse 在去年 12 月那个月在编码智能体上花了 2 万美元。

I agree. I shared earlier today with the audience that our company ClickHouse back in December spent $20,000 that month on coding agents.

上个月我们花了 30 万美元，在四到五个月的时间里。假设我们公司大约有 300 名工程师在使用这些编码智能体。那么，在那个月里，每个工程师大约就是 1000 美元。年化下来，如果没有增长，就是 12,000 美元。但假设我们的消费持续增长，那么可能会到 15,000 到 20,000 美元。我们取上限。所以，像我们这样的公司——我认为我们在采用这些技术方面可能领先于大多数——今年每个工程师大约会花费 20,000 美元，这大约占工程师总薪酬的 5%，包括福利和其他运营成本。有传言说，OpenAI 和 Anthropic 要实现他们融资时的估值，这个比例需要接近 20%，而不是 5% 的总薪酬需要被重新分配。

And last month we spent $300,000 in a single month over the course of four or five months. And let's say we have roughly 300 engineers in the company who are using these coding agents. So, that equates to about $1,000 in that given month per engineer. So, you annualize that without any growth, it would be 12,000. But, let's say we're continuing to grow our consumption, so let's say that goes to 15 to 20,000. Let's take the high end of the range. So, that a company like ours, and I think we're probably ahead of most in terms of our adoption of these technologies, will spend roughly 20,000 per engineer uh this year, which let's just say that accounts for 5% of a fully loaded engineer's uh compensation when you load in benefits and all of the other costs of doing business. It's rumored that for OpenAI and Anthropic to realize the valuations that they've raised at, that needs to be closer to 20%, not 5% of fully loaded salary needs to be redirected.

关于这背后的数学有很多问题，但我明白了，不过这很有说明性。

A lot of questions on the math behind it, but I got it, but it's illustrative.

是的。嗯，那么，你如何看待 Sierra 的运营，比如你的工程师使用这些编码智能体？嗯，然后我想稍微过渡一下，谈谈 OpenAI 和 Anthropic 的动态，你知道，你被广泛认为是行业内杰出的产品领导者之一。从外部看，OpenAI 似乎如此专注于消费者，以至于在企业方面没有追赶上来。所以也许你可以从这两个方面中选一个，给观众一些建议，关于如果他们经营一家风险投资支持的初创公司，应该如何为这些编码智能体做预算。

Yeah. Um so, how do you think about it running Sierra in terms of uh like your engineers using these coding agents? Um and then I want to transition a little bit and talk about the OpenAI Anthropic dynamic, and you know, you're widely perceived to be one of the prominent product leaders in the industry. And from the outside looking in it looked like opening I was so focused on the consumer that it's not playing catch up on the enterprise and so maybe you can take either one of those two advice for the audience in terms of how they should be thinking about budgeting for these coding agents if they're running a venture back startup.

是的，所以我们在 Sierra 的顶级工程师每年消耗的 token 接近 40 万，这很多，但我们认为这转化为了生产力。显然，这不应该是一个排行榜，但我们确实相信这一点。我现在想，嗯，有一个著名的轶事，关于一位 Apple 工程师，他们来了一位新经理，他们制定了一个公式来计算你那天写了多少行代码，这是一种官僚主义的尝试，旨在推动生产力。所以他做了重构，带着很多讽刺，写了一个负数，因为他删除的代码比添加的多，然后交上去，他说，你知道，解雇我吧，你知道，然后

Yeah, so our top engineers at Sierra are spending closer to 400,000 in tokens annualized which is a lot but we think it's it is translates to productivity. It shouldn't be a leaderboard obviously but we really do believe in it. I I think right now um there's this famous anecdote of an Apple engineer who they had a new manager and they created a formula to write how many lines of code you wrote that day in some bureaucratic attempt to drive productivity. So he had done this refactoring and like with a lot of like snark wrote a negative number in there cuz he had deleted more code than he had added and handed it and he's like you know fire me you know and

证明了他的观点，这现在成了民间传说的一部分。

Proved his point and that became now a part of folklore.

我不认为编写代码或消耗 token 本身具有内在价值。Token 是智能的单位，但我认为，驾驭工程的行为——你基本上创建一个驾驭框架，让 AI 智能体能够积极参与到你的产品开发中——我认为这现在是一门艺术形式，需要变成一门科学，这将导致你更高比例的 token 被利用，以积极的方式做出贡献。我认为把 token 消耗视为价值，就像把代码行数视为价值一样，我认为世界上大多数人都会说这是一个愚蠢的指标。你必须编写代码来产生价值，但有时两者之间存在负相关。所以我认为，与其仅仅把它看作一个输出指标而不是输入指标，你不应该为它设立排行榜。但我认为，就像现代的 CI/CD 流水线一样，你知道，你有代码仓库，你有拉取请求，你有持续集成，你有基础设施的不可变快照。就像所有这些东西，我猜你们大多数人都在做。那是在过去 15 年里发明出来的。我不知道“不可变快照”这个短语是什么时候第一次被说出来的，但不是 20 年前。是后来。我的意思是，这个概念存在，但当它变成每个人都这样做的时候。我认为最优秀的工程团队会花大量时间在元任务上，让他们的工程师用来为代码库做贡献的驾驭框架变得极其高效和有效。这意味着你如何自动化测试？你如何自动化代码审查？你如何为编码智能体高效地提供上下文，针对你正在处理的代码库和产品的部分，这样你最终会减少来回的轮次。这些类型的事情将价值连城，就像三年前让单元测试快速运行对人类工程师来说是一种生产力提升一样。如果运行集成测试需要两天，你可能会不运行它，你知道，我们都经历过糟糕的工程环境，对吧？所以我想说的是，我对这些 token 的看法是，我们肯定会进入一个过剩时期，因为它很容易。只需按一下按钮，token 就会出现。但我认为伟大的团队会把它变成一门科学，并真正思考它。这是定性和定量的结合，实际上归结为一种开发者生产力，通过编排智能体的视角，而不是手工编写代码。而且我认为最优秀的工程团队已经这样做了。就像大多数伟大的工程文化，比如 Stripe、Google 这样的公司，他们在开发环境上投入了大量时间。我认为这只是那个的新时代。这就是我认为将会发生的事情。

I don't think like authoring code or consuming tokens has innate value. Tokens are units of intelligence but I think the act of harness engineering where you essentially create a harness so that AI agents can actively contribute to the development of your product I think is a art form now that needs to turn into a science that will result in a higher percentage of your tokens being utilized contributing in a positive way and I think thinking of token consumption as value is like thinking of lines as code as value and I think most people in the world would say that's a silly metric. You have to write code to produce value but there's no there's at times an inverse correlation there. So I think that rather than just thinking of it's an output metric not an input metric you shouldn't have a leaderboard of it. But I think in the same way where like the modern CI/CD pipeline of, you know, you have code repository, you have pull requests, you have continuous integration, you have immutable snapshots for your infrastructure. Like all the things my guess is most of you do. That was sort of invented called over the last 15 years. Like I don't know when the first immutable snapshot phrase was said, but it wasn't 20 years ago. It was later. I mean, the concept existed, but when it became like everyone does it that way. I think the best engineering teams will be spending a lot of time on the meta task of making the harness that their engineers use to contribute to their code base extremely productive and efficient. And that means how can you automate testing? How can you automate code review? How can you provide context to the coding agent efficiently for the part of the code base and product that you're working on so that you do end up with fewer turns back and forth. Those are the types of things that will be worth their weight in gold in the same way making your unit test run quickly was a productivity enhancement for human engineers 3 years ago. And if it took 2 days to run your integration test, you either wouldn't run that, you know, we've all been in bad engineering environments, right? So I would just say the way I think about these tokens is we'll definitely enter a period of of excess because it's easy. It's just push a button and tokens emerge. But I think the great teams will make it a science and really think about it. And and it's a mix of qualitative and quantitative and it actually comes down to a form of developer productivity through the lens of orchestrating agents rather than authoring code by hand. And I would argue it's something that the best engineering teams already did. Like most of the great engineering cultures, the Stripes, the Googles of the world spend a lot of time on their development environment. I think this is just a new era of that. And that's what I think will happen.

如果你愿意，我也可以回答第二个问题。

I can go to the second question, too, if you want.

好的，我们开始吧。

Yeah, let's do it.

嗯，我，呃，OpenAI 一直是生产最佳前沿模型的先驱，从 GPT-4 开始，我认为这是人类的一个开创性成就。我认为 GPT-4 是许多人第一次体验到某种在智能上感觉像人类的东西。我不认为有人会说它是 AGI。01 是第一个推理模型，我认为这是下一个架构上的突破，导致了我们现在拥有的智能。嗯，我认为我们在 5.5 上看到的东西，如果你还没试过，它非常了不起。嗯，根据我的经验，它确实是一个阶跃函数。阶跃函数是一种奇怪的说法，但它在很多方面超过了一些实用性的阈值，我认为这使得它执行长时间运行任务的能力非常出色。我认为我们做得不好的地方，正如你提到的，是在产品方面。嗯，我认为 Claude Code 作为产品在 11 月份是一个更好的产品，嗯，我不会说模型一定更好，但实际使用它来对代码库进行修改的人体工程学要好得多。呃，我认为这是一个合理的批评，就像我们只是没有足够关注它，嗯，Greg Brockman，他是联合创始人之一，有点进入了创始人模式来抓这个产品，我喜欢这样。每次我打开 Cursor，我都必须更新它。我就像，“哦，天哪，又来了？”我就像，“实际上，我对此有点高兴。”

Um I uh OpenAI has been the pioneer in producing the best frontier models from GPT-4, which I think was I think a seminal achievement in humanity. I think GPT-4 is like the first time many people experienced something that felt human-like in its intelligence. I don't think anyone would have said it was AGI. 01, which was the first reasoning model, which I think was the next sort of architectural breakthrough that led to the intelligence that we have now. Um and I think uh what we're seeing with 5.5, if you haven't tried it, it's remarkable. Um it's truly a step function in my experience. Step function sort of is a weird way of saying it, but it's exceeded some thresholds of utility in a lot of ways that I think have made it its ability to do long-running tasks remarkable. I think where we haven't done a great job is, as you alluded to, sort of on the product. Um I think Claude Code as a product in November was just a much better product than um I wouldn't say the models were better necessarily, but the actual ergonomics of using it to actually author changes to your codebase were were just much better. Uh and I think it it is like a legitimate criticism, like we just weren't focused enough on it, and um Greg Brockman, who's one of the co-founders, has sort of gone founder mode on the product, and I love it. Every time I open Cursor, I have to update it. I'm like, "Oh my god, again?" I'm like, "Actually, I'm kind of happy about this."

这太棒了。简直像闪电一样快，现在至少在我这个圈子里——我觉得我还不算太偏颇，因为我跟很多工程师聊过——大家都在用 Codex 了。过去一个月真是了不起。所以我过去和现在都对那个领域缺乏高质量开发者产品感到沮丧，但我对现在的状况非常满意。比如我现在很难从 Codex 换回 Claude Code，因为改进质量摆在那里。所以，你知道，我相信迭代的速度。这其实很大程度上要归功于 Greg，他现在干得特别出色，在公司里推动专注。总的来说，我认为非常好。这个领域有很好的竞争。大多数时候，伟大的创新不会在只有一家玩家时出现。所以，你其实希望你的竞争对手都烂掉。资本主义很少允许这种事情发生。而我最喜欢 OpenAI 的一点是，我们拥有 AI 领域排名第一的消费品牌，也就是 ChatGPT。我们拥有我认为最好的模型，尽管这显然是一场持续追赶的竞赛。而且随着 Greg 在产品方向上推动，把 ChatGPT 和 Codex 整合成一个统一的应用体验，我真的觉得我们拥有非常独特的产品组合和价值主张。所以我对前进方向感到非常兴奋，尤其是过去几个月看到的变革速度让我很激动。

This is great. You can't like it's like a blistering pace, and now, at least in my circles, and I don't think I'm that biased, like I talk to a lot of engineers, like everyone's on Codex now. It's like the past month has been remarkable. So, I was and am frustrated with the lack of just great quality product for developers in that space, but I'm really happy with the state of things now. Like I would have a hard time using Claude Code over Codex right now, just given the quality of the improvement. And so, you know, I believe in the pace of iteration. It's actually largely thanks to Greg, who's just killing it right now, driving focus at the company. Broadly speaking, I think it's really good. There's good competition in this space. Most of the time great innovation doesn't exist when there's only one player. So, you sort of want all your competitors to stink. Rarely does capitalism allow for such a thing to happen. And what I really love about OpenAI is we have the number one consumer brand in AI, which is ChatGPT. We have, I think, the best models, though obviously that's a constant race to keep up with that. And with the direction Greg's taking the product to have a unified app for ChatGPT and Codex in one unified experience, I really think we have a really unique portfolio of products and a value proposition. So, I'm really excited for the direction of travel and excited mostly for the pace of change that I've been seeing over the past couple months.

是的。很明显，编程已经成为 AI 第一个新兴的大规模应用场景，可以通过 Claude 和 Codex 变现。你提到了 Sierra 在客户服务或客户体验方面的市场机会。你还谈到了像 Harvey 这样的法律科技公司。你对大型模型提供商接下来的几个重大应用领域有什么预测吗？

Yeah. So, it's clear that coding has become the first emerging big use case for AI that can be monetized with Claude and Codex. And you talked about the market opportunity that Sierra's in with customer service or the customer experience. You talked a bit about legal tech with companies like Harvey. Do you have a prediction on what the next few big use case areas will be for the large model providers?

所以，对于大型模型提供商，我不太确定。我实际上广泛看好应用型 AI 市场。我的直觉是，来自模型提供商之外的公司会比内部产生更多创新。这有点像，你知道，你和我都在 Salesforce 工作过。某个时候，我猜如果你有个工程师朋友，他们会问你：“这不就是云上的数据库吗？”是不是有人这么问过你？你会说，嗯，差不多吧。

So, I don't know for the large model providers. I actually believe in the applied AI market broadly. And my intuition is it will see more innovation from companies outside of the model providers than inside. It's a little bit, you know, you and I both worked at Salesforce. At some point, my guess if you had a friend who's an engineer, they'd ask you, "Isn't it just a database in the cloud?" Didn't someone ask you that at some point? You're like, well, kind of.

但它还做了很多其他事情。

But it does a lot of other stuff, too.

而且我认为，自从云基础设施和云应用出现以来，几乎对任何应用，都会有工程师说“我周末就能搞定”。但通常你最终看到的是，这些公司专注于公司的某个单一部门，比如财务部门的 ERP 系统、销售部门的 CRM 系统、营销技术（Adobe 为 CMO 做的），往往就是这样。原因在于，商业用户的体验——他们在购买体验和用户体验上的需求——远不止技术本身。尽管 AI 很神奇，我不一定认为这会改变。换句话说，如果有一家公司向财务部门销售审计软件，这样你在季度结束后可以审计你的财务数据，一边是某个熵增的 L8 工程师发布到 GitHub 仓库的技能目录里的 markdown 文件，另一边是网站上列出了他们支持的所有会计准则的公司，你会选哪个？每个人都会选那家公司。我认为这只是因为还有很多其他因素。你想想，你的 CFO 在购买这项技术时会怎么想？这有点像你在买一个可以掐脖子的对象，你知道，你在买一个“这确实是正确的”的责任。你在买所有围绕它的东西。我认为这是解决方案和软件中无聊但重要的部分，30 年前如此，现在依然如此。所以，我的希望是，对于许多极其有价值但极其重要的业务流程，我们会开始看到供应商出现，他们帮助公司不是部署 AI，而是部署解决非常昂贵、非常有价值的问题的方案，用 AI 以极低的成本完成。我提到审计是因为你将来某个时候会成为上市公司，当你上市时，可能在那之前，你季度结束后，你会付给审计公司数百万美元来检查你所有的东西。你会想，不需要是火箭科学家就能想到，天哪，AI 会很擅长检查东西。事实上这正是它最擅长的，而且我们可能会更好、更便宜、更快。这正是你对技术想要的一切。据我所知，世界上没有公司做这个。为什么？因为懂会计和懂 AI 智能体的人的交集是空的。圈子里没有人。所以我们还处于早期阶段，我相信这些类别中都会出现公司。我可能错了。也许世界上只有两家公司，但我就是不相信。我不认为世界是这样运作的。

And I think almost for any application since the advent of cloud infrastructure and cloud applications, some engineer would say I could build that in a weekend. And often though, what you end up seeing is these companies oriented towards a single department at a company like finance for ERP systems, sales for CRM systems, marketing technology what Adobe did for CMOs, often what it went out. And the reason for it is that there's a lot more to the experience of a business user, what they need in both a purchasing experience and user experience than just the technology itself. And despite how amazing AI is, I don't necessarily see that changing. Put it another way, if there was a company that sold auditing software for finance departments, so after you close your quarter, you can audit your financials, the markdown files in a skills directory published to a GitHub repo by an entropic L8 engineer versus a company that had all the accounting standards they support on their website, like which one would you choose? And everyone's going to choose the company. And I think it's just because there's a lot more. You think of like, what would your CFO think when buying that piece of technology? There's a little bit of it is like you're buying the throat to choke, you know, you're buying the accountability that this actually is correct. You're buying all the things that go around it. And I think that is the boring but important part of solutions and software that I think was true 30 years ago and it's true now. So, my hope is that for a lot of the extremely valuable but extremely important business processes, we'll start to see vendors emerge that help companies not deploy AI, but deploy solutions to very expensive, very valuable problems that use AI to do it incredibly cheaply. I brought up the auditing thing because if you'll be a public company at some point in the future and when you are and probably before that you'll be done with the quarter and then you pay auditing firms just millions of dollars to look over all your stuff. And you're like, you don't have to be a rocket scientist to be like, gosh, AI would be very good at looking over stuff. That's in fact what it does best and in fact we probably be better, cheaper, faster. It's like all the things you want with technology. As far as I know, there's no company in the world that does that. Why is that? Because the Venn diagram of people who understand accounting and understand AI agents does not intersect. There's no one in the circle. And so we're just in these early stages and I believe companies will exist in all of these categories. I might be wrong. Maybe it'll just be there's only two companies in the world, but I just don't believe it. I just don't think that's the way the world works.

我还有最后一个问题，然后我们时间就到了。一年前我坐在我们的会议上，当时 OpenAI 和 Anthropic 在谈论他们如何使用 ClickHouse。Anthropic 说他们问 Claude 应该为这个特定用例使用什么技术，Claude 建议了 ClickHouse。我当时有一种觉醒，我说——这听起来有点抽象和牵强，但我不这么认为——那就是，要过多久模型才能不仅仅是推荐应用开发所基于的基础设施，而是实际上在没有任何人工干预的情况下选择和配置那个基础设施。就像现在，房间里的一位开发者需要说，嘿，我想在这个客户体验应用的后端使用 ClickHouse Cloud，你知道，这是我的用例类型，等等。这是我想要摄取数据的方式。这是我想要可视化数据的方式。我在思考这样一个世界，智能体实际上在做选择。所以智能体必须有一个身份。智能体需要有预算权限。我不认为我们离那个现实很远。而且我认为，对于像我们这样的基础设施公司，如果你不走在前面，你就会——因为当人类实际上不再做关于应用背后实际基础设施的选择决策时，你花多少钱都追不上。

I've got one more question and then we're unfortunately at time. I was sitting in our conference a year ago when OpenAI and Anthropic were talking about how they were using ClickHouse. And Anthropic said they asked Claude what technology they should use for this specific use case and Claude suggested ClickHouse. And I had this like awakening where I said, and it sounds a little abstract and far-fetched, but I don't think it is, which is how long is it going to take before the model is not actually just recommending the infrastructure that an application is developed on, but is actually selecting and provisioning that infrastructure without any sort of human intervention. Like right now a developer in the room needs to say, hey, I want to use ClickHouse Cloud on the back end for this customer experience application and you know, this is the type of use case I have, etc. This is how I want to ingest data. This is how I want to visualize the data. And I'm thinking about this world where agents are actually making the selection. So the agents got to have an identity. The agent needs to have budget authority. I don't think we're that far off from that reality. And I think for an infrastructure company like ours, if you're not ahead of it, you're because there's no amount of money you can spend to catch up when a human is actually not making that selection decision on what the actual infrastructure is that sits behind an application.

我们会面对这样的现实吗？如果是的话，会在什么时间范围内？

Is that a reality that we're going to be looking at? And if so, in what time frame?

嗯，答案是“我不知道”。所以，我给你我的经验法则。我喜欢用案例研究来尝试戳破一些想法，因为我觉得旧金山有些人会说：“智能体会买下一切。智能体，智能体。我们就是，不知道做什么。坐在沙滩上，机器人给我们端玛格丽塔酒之类的。”如果，当我在 2003 年谷歌的时候，你问我：“谷歌购物会非常流行吗？”我会说绝对会。航班搜索、购物，所有这些，因为为什么你要为一个你正在搜索的东西用另一个搜索框呢？我怀疑你们中没有人用过谷歌购物。我没用过。它很糟糕。不过有点意思，因为谷歌真的很受欢迎。所以，我们都在笑，但实际上这有点令人惊讶，因为在 2003 年，谷歌将主导购物是非常可信的传统智慧。那为什么呢？我可以为自己表达，就是，你知道，我知道如果我从沃尔玛或亚马逊买，我知道我可以信任产品不是欺诈的，会按时送达。总的来说，我不信任运费。最后一公里真的有问题。你不知道代发货之类的，所有这些事情。所以，归结为信任，基本上最后一公里的物流实际上是主导决策因素，而不是实际的分配，那是谷歌所代表的。

Um, I don't know is the answer. All my end. So, I'll give you my rules of thumb on this. I like to use case studies to try to poke, because I think there are some people in San Francisco who say, 'Agents will buy everything. Agent, agent. We'll just, I don't know what we'll do. Sit on the beach and robots will serve us margaritas or something.' If, when I was at Google in 2003, if you'd asked me, 'Will Google Shopping be really popular?' I would say absolutely yes. Flight search, shopping, all this, because why would you want a different search box for something you're searching for? I doubt any of you have ever used Google Shopping. I haven't. It stinks. Kind of interesting though, because Google's really popular. So, we're all sort of laughing, but actually it's kind of surprising because it would have been extremely credible conventional wisdom in 2003 that Google was going to dominate shopping. Well, why is that? I can express for myself, which is, you know, I know if I buy from Walmart or Amazon, I know I can trust that the product's not fraudulent, that it'll be delivered on time. In general, I don't trust shipping costs. The last mile is really problematic. You don't know about drop shipping, all this stuff going on. And so, it comes down to trust and basically the last mile logistics actually were the dominant decision-making factor, not the actual distribution, which was what Google represented.

我认为 ChatGPT，在那种环境下简化来看，你可以把它想成另一个分销渠道。就像你有搜索、有信息流、有应用商店，现在你有 ChatGPT 和 Gemini。所以，问题是，你会在哪里委托 ChatGPT 和 Gemini 代表你做这些决定，哪里会有其他考虑？我认为我们的青少年可能不会把他们的时尚选择委托给 ChatGPT，因为他们更在乎学校朋友对他们的衣服的看法，而不是 AI 对他们的衣服的看法。我相信这是真的。我认为手提包可能也是这样。你知道，没有人因为一个深刻的 AI 分析而买手提包。这里面有更多因素。谁做的、你对它的感知、它的独特性，所有这些其他因素。你的车险会是这样吗？这是个有趣的问题，对吧？比如，那是你可能会选择最低成本选项的地方吗？嗯，我本来会说那里是的，但如果你看看现在的航空业，曾经有一个传统智慧，没人关心他们坐哪家航空公司，所以很多航空公司基本上都在试图降低成本。如果你看看航空业的利润去向，都流向了美联航和达美。他们采取了完全相反的方法，不是试图拥有最低的票价，而是转向更忠诚的商务旅客，这实际上在面对 AI 时感觉更持久。

I think ChatGPT, reductively in that environment, you can think of as another distribution channel. Like you had search, you had news feed, you have app stores, now you have ChatGPT and Gemini. And so, the question is, where will you delegate to ChatGPT and Gemini to make these decisions on your behalf, and where will there be other considerations? I think our teenagers probably will not delegate to ChatGPT their fashion choices, because they care more about what their friends at school think of their clothes than what AI thinks of their clothes. I believe that to be true. I think that's probably true of like handbags. You know, no one bought a handbag because of an insightful AI analysis. There's a lot more that goes into it. Who made it, your perception of it, its exclusivity, all these other factors that go into it. Will that be true of your car insurance? It's kind of interesting question, right? Like, is that where you might just go to the lowest cost option? Well, I would have said yes there, but if you follow the airline industry right now, there was a conventional wisdom that no one cared about the airline they flew, and so you had a lot of airlines go to basically trying to lower cost. If you look at where profits in the airline industry are going to, it's all going to United and Delta. And they've taken the exact opposite approach of trying to have the lowest cost fares, and they're going towards more loyal business travelers, which just feels durable in the face of AI, actually.

所以，我要告诉你的是，我认为我可以简化地说，我确信会有人就是不在乎或者比其他人更信任 AI，但在某种程度上，最有价值的企业是在那些不成立的地方的边缘创造出来的。就像现在航空业的每个人都希望自己是美联航和达美。每个人。所以，在某种程度上，如果这是真的，问题就会是，为什么人们会选择 ClickHouse？比如，是什么，你知道，在某种程度上，参与其中将几乎是一个不可避免的结果，你会将自己的品牌商品化。我认为消费者行为和分销中最有趣的问题是在这些规则有例外的地方产生的。

So, the thing I would tell you is, I think I could reductively, I'm certain there will be people who just don't care or trust AI more than others, but to some degree the most valuable businesses are created in the margins where that's not true. Like everyone in the airline industry is wishing they were United and Delta right now. Everyone. So, to some degree, if that's true, the question will be like, why will people choose ClickHouse? Like, what is, you know, to some degree playing into that will almost be an inevitable outcome where you will commoditize your own brand. And I think the most interesting problems in consumer behavior and distribution are made where there are exceptions to those rules.

好的。

Okay.

好了。嗯，不幸的是我们时间到了。所以，代表所有人，Brett，我非常感激你花了一个小时和我们在一起。我们知道你的时间多么宝贵，所以，我们能感谢 Brett 吗……

All right. Well, we are unfortunately out of time. So, on behalf of everybody, Brett, I'm so grateful that you spent an hour with us. We know how demanding your time is, and so, can we thank Brett for...

谢谢你，Brett。

Thank you, Brett.

太棒了。

Amazing.