Ramp 押注百倍效率员工

Ramp's Bet on the 100x Employee

埃里克·格莱曼 Eric Glyman · Y Combinator · 2026-11-18 · 约 54 分钟 · 原视频 ↗

打开互动全文版(中英对照 + 问答)→

本期速览 · Overview

Ramp 联合创始人兼 CEO Eric Glyman 讲述这个估值 600 亿美元的平台如何借助 AI 智能体重塑企业的财务运营方式。

Eric Glyman, co-founder and CEO of Ramp, explains how the $60 billion platform is using AI agents to reshape how companies run their finances.

要点 · TL;DR

核心观点 · Key points

反共识 · Contrarian takes

本期章节 · Chapters(共 31)

全文 · Full transcript(中英对照)

引言 Introduction

Host

今天在《如何构建未来》节目中,我与 RAMP 的联合创始人兼 CEO Eric Glyman 坐下来聊聊。RAMP 七年前以一张公司卡起步,承诺只有一个:为公司节省时间和金钱。如今,它已是一个价值 600 亿美元的平台,为数以万计的企业运行财务运营。而且越来越多地,它使用智能体来完成所有这些工作。

Today on How to Build the Future, I'm sitting down with Eric Glyman, co-founder and CEO of RAMP. RAMP started seven years ago as a corporate card with one promise: save companies time and money. Today, it's a $60 billion platform that runs the financial operations of tens of thousands of businesses. And increasingly, it does all of that work using agents.

Host

我们的发展速度,我想大多数人看了都会说:“这有点了不起。”这就是 Ramp 如何从底层重塑公司构建方式的故事。Eric,感谢你来。

We're sort of at a pace that I think most people would look at and go, "This is kind of remarkable." This is the story of how Ramp is reshaping the way companies are built from the ground up. Thanks for being here, Eric.

Eric

说实话,Gary,很高兴回来。感谢邀请。

Honestly, Gary, it's great to be back. Thanks for having me.

Host

欢迎回来。你和你的联合创始人 Kareem 打造了 Parabus,我们在 YC 2015 年夏季班一起工作过。实际上你把它卖给了 Capital One,然后你继续创办了 Ramp。跟我们讲讲这段旅程。那是什么感觉,你知道,决定申请 YC,加入,做起来,然后你怎么走到下一步的?

Welcome back. You and your co-founder, Kareem, built Parabus and we got to work together YC Summer 15. Actually you sold it to Capital One and then you went on to start Ramp. Walk us through that journey. What was that like, you know, deciding to apply to YC, coming in, doing it, and then how'd you get to your next thing?

Parabus与YC历程 Parabus and YC Journey

Eric

十年前我们有一个很简单的想法:如果你每次在网上买东西后价格下降了,你能自动拿回那笔钱会怎样?于是我们打造了 Parabus,当时是一个省钱应用,但我真的认为它是史前 AI 智能体之一,为你做这件事。所以,它会关联你的 Gmail 或 Yahoo 账户。它会扫描你的收件箱找收据,结构化数据,去查价格,如果你有资格获得退款,它就从你的账户发邮件要求退还差价。而且效果非常好。我们发布了。一个月后,我们就来到了 YC。我们实际上申请了两次。所以,对于任何考虑申请的人,如果你第一次没进,再试一次。我们很高兴我们来了,因为它让我们走上了完全不同的轨迹。从 YC 毕业一年内,我们每天有近一百万用户,而且我们收到了 Capital One 的收购邀约。所以,你知道,我们最终在 Capital One 度过了接下来的两年半。我们看到了一个世界级组织在很多方面是如何运作的,它如何利用信息来革新其信用卡行业,学到了很多了不起的东西。也学到了一些反面教训。你知道,看到一个组织有那么多好的流程,却常常扼杀伟大的想法,看到激励措施有时并不总是与客户一致,导致结果本可以更好。最终在那里的岁月让我们质疑:如果你的金融产品不是仅仅为了公司赚更多钱而设计,而是真正帮助个人少花钱呢?你可以靠价值竞争,而不是靠价格,你可以建立一种非常不同的公司。

A decade ago we had a pretty simple idea: what if every time that you bought something online and it dropped in price, you could get that money back automatically? And so we built Parabus, which was back then a savings app but I really think was one of the prehistoric AI agents that did that for you. So, it would link up with your Gmail or Yahoo account. It would scan your inbox for receipts, structure the data, go check the price, and if you were eligible for a refund, it just emailed from your account asking for a refund of the difference. And it worked really well. We launched. A month later, we were here at YC. We actually applied twice. And so, for anyone thinking about it, if you don't get in the first time, try again. And we're very happy we came because it set us off on a completely different trajectory. Within a year of graduating from YC, we had almost a million users every day and we had a budgeting offer from Capital One to buy the company. And so, you know, we ended up spending the next two and a half years just over that at Capital One and we got to see how in many ways a world-class organization operated, how it used information to revolutionize its industry of credit cards and learned a lot of incredible things. Also learned some anti-lessons. You know, seeing an organization that had so many good processes that often killed great ideas, saw how incentives sometimes that are not always aligned with customers led to outcomes that could be much better. And ultimately the years of being there led us to question what if your financial products were not designed just to go and make the company more money but were actually designed to help individuals spend less. You could compete on value not on the basis of price and you could build a very different kind of company.

从零到一与首次创业 Zero to One and First Startup

Host

这很迷人。第一个创业想法的从零到一是什么样的?比如你是怎么想到的?你知道,你需要转型吗,还是说这就是我们想要的,这是我们为自己想要的,然后人们也想要它。

That's fascinating. What was like the zero to one like on the first startup idea? Like how did you come up with it? You know, did you need to pivot or was it sort of like this is what we wanted, this is what we wanted for ourselves and then people wanted it.

Eric

我毕业时,似乎只有两份工作:银行家或顾问,所以我在一家银行工作,Kareem 是顾问,因为那有助于他的签证,我们喜欢我们的工作,但不是热爱,我们总是梦想着,我们会读 Paul 写的一些文章,看一些早期的创业学校视频。让我们印象深刻的是,你知道,不要只是抽象地想出想法,但我记得那篇关于如何获得想法、如何获得创业想法的文章。我从中得出的结论是,你应该只是注意你日常生活中的问题。在办公室工作了一年,每天直到外面天黑才离开,终于有一群朋友累得不行,我们订了一张机票。我们以为自己很聪明。我们用了无痕模式。这样我们就能得到非个性化的价格。第二天,另一个朋友想加入。我看了航班信息发给她,我说:“这很奇怪。航班便宜了 100 美元。这是对的吗?”结果是真的。我非常沮丧。在我的工作中,我会阅读信贷文件和契约。所以我只是读了政策。结果发现,如果你在 24 小时内通知航空公司,他们会给你退款。所以我打了电话,一小时内就完成了。你知道,我和朋友们拿回了近 1000 美元。我就坐在那里。我想,“这可能一直在发生。这可能一直在发生,而且”你知道,我学过足够的计算机科学,知道哦,如果你给电脑一个网站去检查,你可以去抓取并检查价格,政策相对结构化,我们能做些什么呢?所以这导致了一年的捣鼓,试图弄清楚什么是对的形态,你能从信用卡拉数据吗?信息不够,没有逐项收据,可以用电子邮件,有一些生产力应用帮你整理收件箱或写得更快,但从来没有人把电子邮件作为收集信息的方式,然后发送去构建退款请求。所以我们开始买一些东西。结果它们降价的速度比我们想象的要频繁得多。我们尝试构建这些触发器。最初是用 VBA 作为脚本构建的。那是我们所知道的。我们开始收到退款。我们问朋友,“嘿,我知道我们是你的朋友。这有点奇怪,但我们能拿到你的电子邮件登录信息,试着帮你省点钱吗?”我们就一直拉。这就是我们走到那里的方式。我想当我们见面时,是这样的:每 20 个人中就有一个人在你注册 Parabus 后的几分钟内就能省钱。

When I graduated, it seemed like the only two jobs were banker or consultant and so I worked at a bank and Kareem was a consultant cuz it helped with his visa and we liked our jobs, didn't love it, and we were always kind of dreaming and we would read some of the essays that Paul would put out and watch some of the early startup school videos. And what stuck with us was, you know, don't just try to come up with ideas in the abstract, but I remember the essay of like how to get ideas, how to get startup ideas. And the conclusion that I took from it was you should just notice problems in your regular life. And after a year of working in an office and not leaving until it was dark outside, finally a group of friends got exhausted enough that we booked a flight. We thought we were smart. We use incognito mode. So we'd get a non-personalized price. And the next day, another friend wanted to join. And I looked at the flight to send her the details and I was like, "This is weird. The flight is $100 less. Is this right?" And turned out it was. And I got really frustrated. In my job, I would read credit documents and indentures. And so I just read the policy. It turned out the airline would give you a refund if you let them know within 24 hours. And so I called, it was done within an hour. You know, my friends and I had almost $1,000 back. And I just sat there. I'm like, "This could be happening all the time. This could be happening all the time and like" you know we had taken I had taken enough computer science to know like oh if you give a computer a website to go check you can go scrape and check the price and policies are relatively structurable like what can we do about it and so it led to a year of kind of tinkering and trying to figure out what's the right form factor could you pull the data from credit cards not enough information there doesn't have the itemize your receipt could use email there was these productivity apps to help you organize your inbox or write faster but no one had ever used email as a way to collect information and then send you know to go and build refund requests. And so we just started buying some things. Turns out they were dropping in price a lot more often than we thought. We tried building these triggers. Initially was built in VBA as a script. That was what we knew. And we started getting refunds. We asked our friends, "Hey, I know we're your friends. It's a little weird, but can we get your email login and try to save you some money?" And just we kept pulling. And so that was how we got there. And I think when we had met, it was such that for one in 20 people within minutes of you signing up for Parabus, you would save money.

早期增长与用户获取 Early Growth and User Acquisition

Host

太棒了。

Amazing.

Eric

我们就想,要么我们相信这个,要么不信。让我们试试这个。

And we're like either we believe in this or we don't. Let's try this.

Host

获得第一批用户是什么感觉?你知道,你做了什么来获得你的前一千或一万用户?

What was it like to get first users? And you know, what did you have to do to get your first thousand or 10,000 users?

Eric

我们什么都做了。我们在纽约举着牌子。我们和朋友们交谈。我们没有太多朋友,所以很快就用完了。好笑的是,我们为 Parabus 获得很多第一批用户的地方是在 Reddit 上。有一个叫 Beer Money 的子版块。那里的人想要 5 美元周末买啤酒。所以,Kareem 就复制了它,写了一个帖子:“嘿,我们有这个应用能帮你省钱。你可以连接你的电子邮件。”然后,我们醒来,有了几百个用户,然后就顺利开始了。但是,你知道,一切都是艰苦争取的。我认为关于把所有复杂性限制在一个指标上的建议,对我们来说就是用户。我们想每周增长 10% 以上,我们做了各种权衡。

We did everything. We carried signs around New York. We talked to friends. We didn't have too many friends, so we ran out of them very quickly. Hilariously, where we got a lot of our first users for Parabus was on Reddit. There was a subreddit called Beer Money. Where it was people who wanted $5, to buy a beer on the weekend. And so, Kareem just ripped it and just wrote a post of, "Hey, we have this app that saves you money. You can connect your email." And, we woke up and we had a few hundred users and was kind of off to the races. But, you know, it was everything was hard fought. I think the advice around just try to constrain all the complexity to one metric for us it was users. We wanted to grow that 10 plus% per week and we made various trade-offs.

简化价值主张 Simplifying the Value Proposition

Eric

我们学到了很多关于如何简化故事的方法:就是“商家欠你钱,我们去帮你拿回来”。那个业务的大问题是它看起来好得令人难以置信。那么,如何帮助人们克服那种“等一下,你在帮我省钱,只有省了钱才收费,但你要访问权限”——陷阱是什么?当我们和客户交谈时,我们试着听他们向朋友解释。就像我们以前做的 Craigslist 测试,我们假装是市场研究人员,为我们自己的公司做研究,说“是啊,我们不知道这个应用是干什么的”,然后你问人们觉得它怎么样。他们会解释他们喜欢什么,我们后来看到他们并解释“其实,我们是创始人”。我们会问:“你有没有告诉过别人,只是为了推销它?”他们会说:“有个叫 Paribus 的应用,真的很棒,每当你买东西并且它打折时,它就能帮你省钱。不过你得知道,它会保留省下的钱的 25%。”所以自然,当人们向朋友介绍东西时,他们想展示正面和负面——你要注意这一点。

We learned a lot about how to simplify the story: just 'stores owe you money, we're going to get it for you.' The big problem we had with that business was it seemed too good to be true. What would help you overcome something that seemed like, 'Wait a minute, you're saving me money, you only charge if you save me money, but you want access'—what's the catch? When we talked to customers, we tried to listen to them explaining it to friends. Like we used to do the Craigslist test where we'd pretend to be market research people doing research for our own company and be like, 'Yeah, we don't know what this app does,' and you'd ask people what they think of it. They'd explain what they like, and we'd see them later and explain, 'Actually, we're the founders.' And we'd ask, 'Did you ever tell other people about it just to pitch it?' They would say, 'There's this app called Paribus, it's really great, it saves people money whenever you buy things and it goes on sale. You should know though, it keeps 25% of what it saves you.' So naturally, when people tell their friends about things, they want to show you the positives and the negatives—you want to look out for that.

Host

真的,真的,确实如此。

It's real, it's real, that it's real.

Eric

我们听了足够多这样的对话,然后想:“嗯,我们能不能给人们一些东西,让他们更容易克服负面因素?”我们意识到,其实我们应该加入一个双向推荐协议,激励人们告诉更多朋友。如果每次你推荐一个人,Paribus 保留更少——所以不是收 25%,我的费率会是 20%,你的费率也会是 20%——然后我们看到推荐量爆炸式增长。我们找到了让新闻报道它的方法,随着我们推进,增长率非常快地持续指数级增长。

And we listened to enough of these conversations and we're like, 'Huh, could we give people something that would make it easier to overcome the negative?' And we realized actually we should put in a two-sided referral deal that would motivate people to tell more friends. If every time you refer someone Paribus keeps less—so instead of charging 25%, my rate will be 20%, your rate will be 20%—and we saw referrals explode. We figured out ways to get the news to talk about it, and as we got that going, very quickly the growth rate just continued to grow exponentially.

简洁的价值主张 The Simple Value Proposition

Host

太棒了。我的意思是,这很酷,因为当你开始做 Ramp 时,听起来你推出时也是类似的氛围。你说:“你知道,节省时间和金钱。”你是怎么做到的?

That's amazing. I mean, it's cool because when you set out to do Ramp, it sounds like that was sort of the same vibe when you came out. You said, 'You know, save time and money.' How do you do that again?

Eric

也许我们只有一个想法。

Maybe we've only ever had one idea.

Host

是啊。我的意思是,有时候如果它是个好想法,那就足够了。

Yeah. I mean, sometimes if it's a good one, that's all you need.

Eric

我认为这就像人类深层的欲望:用更少的钱得到更多,你知道,也许无论贫富,每个人都喜欢划算的交易。

I think it's just like a deeply human desire to get more for less, you know, and maybe rich or poor, everyone likes a good deal.

Host

是啊。是啊。

Yeah. Yeah.

Eric

我们有这样的洞察:如果你有一个非常简单的价值主张,你就能非常清晰地传达它。人们想使用这类产品。当我们在 Capital One 时,我们了解了金融机构如何运作。这些是令人难以置信的商业模式,利润极高。并不是任何个人是坏人,而是组织上的激励措施使得人们设计的产品让客户稍微吃亏,让公司受益。我们认为,如果我们真正想要同样的东西,那将非常强大。

We had that insight that if you had a very simple value proposition, you can communicate it very clearly. People want to use those types of products. And when we were at Capital One, we learned about financial institutions, how they work. These were incredible business models that were incredibly profitable. It wasn't that any individual person was bad, but just organizationally the incentives were such that people were designing products that made their customers just a little bit worse off to make the company better off. And we thought it could be so powerful if we actually wanted the same thing.

激励错位的例子 Examples of Misaligned Incentives

Host

你们是从那里开始的吗?有哪些例子看起来特别令人惊讶?

Did you start there? What were some examples that seemed particularly surprising?

Eric

当你有一把锤子时,每个问题看起来都像钉子。所以第一,在制作新银行账户或新信用卡方面,很少有新的竞争对手。

When you have a hammer, every problem looks like a nail. So one, there's very few new competitors in making new bank accounts or perhaps in making new credit cards.

Host

我的意思是,这些在某种程度上都是监管的、被俘获的东西。

I mean, these are all regulatory, sort of captured things at some level.

Eric

我们这么认为。我的意思是,当时我们的大多数竞争对手,他们的创始人戴着高顶礼帽,不是因为他们古怪,而是因为他们出生在 1800 年代。就像没有很多新的竞争,对吧?大多数人传统上拥有的工具是你可以有返现卡或积分卡。为了分解逻辑树,人们知道的公司的工作是每个季度赚更多利润,快速增长,最终长期利润金额变得非常可观。所以每个季度你的团队都在这个汽车业务的激烈竞争中,为了增加购买量,让人们花更多钱,推动更多购买,从而推动更多顶线收入。然后你想保持更高的净利润率。信用卡公司最大的成本之一就是那些积分和奖励计划的成本。所以机构上的激励变成了:如果你能说服人们积分值很多钱,他们就会花更多钱

We think so. I mean, most of our competitors then, their founders wore top hats, not because they're eccentric, they just were born in the 1800s. Like it's just like not a lot of new competition, right? And the tools that most people had conventionally were you could have cards with cash back or with points. And to kind of break down the logic tree, the job of companies people know is make more profit each quarter, grow quickly, and eventually long-term the profit dollars become really great. So every quarter your team is on this rat race in the car business to grow purchase volume to get people to spend more money and drive more purchases which drives more topline revenue. And then you want to keep a higher net margin. One of the largest costs for credit card companies was the costs of those points and reward schemes. And so the incentive institutionally became well if you could convince people that the points were worth a lot, they were going to go spend more

Host

并让他们不使用它们。

and get them to not use them.

Eric

没错。或者贬值它们,或者做这类事情。而且他们在不同的部门。所以每个人都在优化不同的东西,但不知何故没有对齐。我们是这个集团内部和卡部门内部的产品人员,我们过去做各种用户研究,你可以把一些练习归结为问客户:什么会让你更多使用我们的产品?你想要积分还是返现计划?如果你问人们并且真正倾听他们说的话,不仅仅是 A 或 B,而是他们反馈给你的信息。他们会说类似“我其实只是想要银行账户里有更多钱”。如果你深入挖掘,你就会想:“嗯,也许如果你不花 100 美元在那个你付费但已经忘记的订阅上,并在月底取消它,你的银行账户里可能比只得到 1% 返现多出 99 美元。为什么没有一张卡或软件能做到这一点?”如果你能做好这一点,你就能为客户提供不是 10 倍而是几乎 100 倍更好的价值主张。那个业务的另一个大问题是人们害怕这是一场逐底竞争,如果你只是分割一个小利润,那么你知道这会被蒸发掉。如果你能真正在价值基础上竞争——我说的价值是指你账户里实际有更多美元——但你通过技术做到这一点,你不是通过说“我只是要给更大的返利”来做到,你就能有更多美元投资于新事物。所以某种程度上,早期的洞察甚至还不是技术洞察,而是商业模式的洞察:如果你能把这些有时是金融产品的利润池,用来投资于伟大的软件,就导致了我们的一些早期产品,比如那张不仅想让你少花钱而且帮你做费用报告的卡。它让我们能够投资于伟大的账单支付软件,我们实际上比只做这类软件的企业有更多的美元。现在你看到这个平台,我们不仅寻求与最终客户对齐,而且我们衡量他们实际少花了多少美元和多少小时。

That's right. Or devalue them or do these kind of things. And they were in different departments. So everyone was optimizing different things but somehow wasn't aligned. And we were product people inside of this group and inside of the card division and we used to do all kinds of user research and you could boil some of these exercises down to asking customers what would get you to use our products more. Do you want a point or cashback program and if you ask people and you really listen to what they said not just was it A or B, but what they were telling back to you. They would say something along the lines of I actually just want more in my bank account. which if you pull on that, then you're like, "Huh, well, maybe if you don't spend $100 on that subscription that you're paying for that you've forgotten about and you cancel that at the end of a month, you might have $99 more in your bank account than if you just got one% back. Why isn't there a card or software that does that?" And if you could do that well, you could have a not a 10x but an almost 100x better value proposition for customers. The other large problem in that business was people were terrified that it was a race to the bottom that if you're just slicing up a small margin well it's you know this gets boiled away. And if you could actually compete on the basis when I say value of actually more dollars in your account but you do this through technology you do this actually not by saying I'm just going to give a larger rebate you could have more dollars to invest in new things and so in some ways the early insight was not even yet a technological insight but a business model insight of if you could take these profit pools of sometimes financial products and use them to invest in great software led to some of our early products like the card that not only wants you to spend less but does your expense reports. It led us to be able to invest in great bill payment software where we had a lot more dollars actually than businesses just doing that type of software. And now you see this platform where not only are we seeking and pursuing to be aligned with our end customers. and so we measure how many fewer dollars and hours they actually spend.

在Ramp从0到1 From 0 to 1 at Ramp

Host

但因为你更全面地考虑了模式,我们比许多竞争对手有更多资金,而他们只是在投资不足,没有真正让客户的业务变得更好。所以 Ramp 从 0 到 1 的关键就是:你如何真正介入,类似于 Paribus,你花了这么多钱在订阅上——你在用它们吗?让我们让取消订阅变得容易。这就是最初奏效的从 0 到 1 吗?

But because you've thought about the model more holistically, we have more dollars than many of our competitors who are just underinvesting in actually making their customers' businesses better. So the zero to one for Ramp really was: how do you actually go in and, similar to Paribus, you're spending all this money on subscriptions—are you using them or not? Let's make it easy to unsubscribe. Was that sort of the initial zero to one that worked?

Eric

说得太对了,还有一个额外的转折点,对于考虑构建 AI 业务的人来说很重要:行动或做事的能力。我们在 Paribus 感受到的挑战之一是,我们可以看到并代表客户被授权查看每天数亿封电子邮件。其中一小部分是网上零售购买或航班或酒店预订。但我们相信,我们可以节省数十亿美元。但我们唯一的行动机制是发邮件给客服,希望他们给最终客户退款。我们着迷于这样一个想法:你能否真正将其嵌入资金的流动中?你能否足够接近底层——直接说,看到你在浪费钱是一回事。能够阻止购买并真正防止你不想保留的订阅流出的钱是另一回事。

It's perfectly said, and there's one extra twist which is important for people thinking about building AI businesses: the ability to act or do something. One of the challenges we felt at Paribus was we could see and were authorized on behalf of customers to look at hundreds of millions of emails per day. A small amount of them were online retail purchases or flight or hotel bookings. But there were, we believe, billions of dollars we could be saving. But our only mechanism to act was email customer service and hope they give a refund to the end customers. We were fascinated with the idea of could you actually embed this into the movement of money? Could you be close enough to the metal that—to put it directly, it's one thing to see that you are wasting money. It's another thing to be able to stop the purchase and actually prevent dollars going out for subscriptions you don't want to leave.

Host

没错。我们进一步深入。所以这就是为什么它从仅仅构建智能转变为真正进入运营层。

Exactly. We're taking it down further. So that's what part of pulled it from just building the intelligence to actually could you be at the operational layer.

Eric

这很合理。你知道 Paul Graham 经常谈论这个。就像当你从 0 到 1 时,向初创公司销售实际上是一个优势,而不是向最大的企业销售。你以后会得到最大的企业。这对你来说是真的吗?

That makes sense. You know Paul Graham talks about this a lot. It's like when you're going zero to one, it's actually a feature to go sell to startups as opposed to the biggest enterprises. You'll get the biggest enterprises later. Was that true for you?

Eric

是的,有一些非凡的公司,多年前还是小初创公司,现在已经是巨头,我们与他们一起成长。我想到 Cursor 和 Cognition。这两个组织我们服务时它们只有几个人,后来他们建立了非凡的公司,所以这绝对正确。我认为如果你能为未来构建,也能为过去构建,而你必须选一个,我每天都会选未来。

It was, and there are just extraordinary companies that years ago were small startups and now they're huge companies and we've grown with them. I think about Cursor and Cognition. Both organizations we served when they were just a handful of people to they built extraordinary companies and so that's absolutely right. I think if you can build for the future, build for the past and you had to pick one, I'd pick the future every day.

Ramp的600亿估值与序幕终结 Ramp's $60B valuation and the end of the beginning

Host

好吧,快进一点。Ramp 刚刚达到 600 亿美元估值。恭喜。

Well, fast forward a little bit. Ramp has just reached $60 billion valuation. Congratulations.

Eric

谢谢。

Thank you.

Host

你称这是开始的结束。这是什么意思?如果你必须说,好吧,那是第一章。那是什么,然后当然下一步是什么?

And you called this the end of the beginning. What does that mean? And what if you had to say, okay, well, that's the first chapter. What was that and then of course what's the next?

Eric

我从产品层面和我们的运营方式来谈,写那篇文章时我们确实在考虑两方面。第一,如果你在公司内部运营过,你有一边是薪资系统,谁向谁汇报、部门位置,很多层级。另一边是你的 ERP,你的记录系统,存储这里是我们这个月花了什么,这里发生了什么,这里是我们收到的所有钱。而中间对财务团队来说通常是一团糟,因为银行可以转移资金,软件人员可以构建奇怪的后市场软件。如果你看看公司现实——我的意思是,今天大多数公司仍然如此,但对于非 Ramp 客户,你有一个系统用于信用卡,另一个用于费用,另一个用于账单,另一个用于采购,另一个用于差旅,另一个用于——你可以一直列举下去。真的很笨重。所以,如果你真的想甚至只是教育,哦,如果你是员工想买软件,你可能必须教他们使用多个系统。你得到不完整的视图,最终导致浪费、浪费时间、浪费金钱等等。所以第一章的一部分是构建完整的网状层,几乎像吃豆人一样吞噬所有不同类型的工具,这样你实际上可以使用一个非常简单和精简的盒子来确保离开组织的钱更少。今天,中位数公司使用 Ramp 每年在费用上节省 7.5%,这是交换费甚至可能的多倍,并帮助他们增长得快得多。从产品角度,我们做得很少,试图在——如果我们的目标是帮助人们削减成本,我们在软件中关于如何帮助更快增长收入方面做得较少。所以第二章的一部分不仅是如何帮助确保离开组织的钱更少,而且实际上对于你试图花费的钱,我们如何让它们有更高的乘数?你如何收取收入?所以我们最近推出了应收账款。Glass 和 Inspect 以及我们的一些新 AI 工具实际上明确设计为让团队每个成员更高效。所以从产品视角,这不仅仅是减少浪费,更多的是更快增长、时间和金钱走得更远。从公司视角,大约不到四年前,ChatGPT 推出。这是 GPT-3.5。它是世界上最先进的模型,你我都记得它真的不会算术,对吧?2 加 4 等于谁知道。

I talk about from a product level and how we operate is we were really thinking about both when we wrote that post. One, if you've operated inside of a company you have, on one side you have your payroll system, who reports to whom, department locations, a lot of hierarchies there. On the other side you have your ERP, your system of records where you store here's what we spent this month, here's what happened, here's all the money we received. And in between is like a total cluster for the finance team, classically, because the banks could move money and the software people could build weird aftermarket software. If you looked at companies and realities—I mean, it's still true for most companies today, but for non-Ramp customers, it's you have one system for credit cards, another for expenses, another for bills, another for procurements, another for travel, another for—you can go on and on and on. It's really cludgy. And so, if you actually want to go and even just educate, oh, if you're an employee and you want to buy software, you might have to teach them to use multiple systems. You have an incomplete view and you end up with waste, wasted time, wasted money, all that. And so part of the first chapter was building kind of the complete mesh layer almost Pac-Manning up all of the different types of tools so you could actually use one very simple and streamlined box to make sure there's fewer dollars leaving your organization. And today the median company uses Ramp and saves 7.5% per year on their expenses, which is multiple times what interchange even could be, and is helping them grow a lot faster. We did very little from a product perspective trying to be disciplined on—if our goal was to help people cut costs, we did less in our software of how do you help grow revenue faster. And so part of this chapter 2 is not just how do you go and help make sure there's fewer dollars leaving organizations, but actually for the dollars you are trying to spend, how do we make them have a higher multiplier? How can you collect revenue? And so we've launched accounts receivable recently. Glass and Inspect and some of our new AI tools are actually expressly designed to make every member of teams more productive. And so from a product lens, it's really about not just less waste, but it's more about faster growth, time, and money going further. From a company perspective, roughly a little less than four years ago, ChatGPT launched. This is GPT-3.5. It was the most advanced model in the world and you and I both remember it couldn't really do arithmetic, right? 2 plus 4 equals who knows.

Host

基本的事情它都做不了。

Basic things it couldn't do.

Eric

是的。我的意思是,它是个很酷的玩具。

Yeah. I mean, it was a cool toy.

Host

它很神奇。你知道,有时它们有有意义的段落。你来回互动,然后——

It was amazing. You know, sometimes they have paragraphs that made sense. You go back and forth and—

Eric

那时它还是个随机鹦鹉,但它进化得很快。

That was when it was a stochastic parrot still, but it evolved quickly.

Host

它进化到了解决纳维-斯托克斯方程的地步,这是过去 200 年世界上最顶尖的数学家都无法解决的问题。然后我想 3 周后在一个随机的周四晚上,它又给出了数百个问题的证明。我认为大多数人仍然没有内化进步的速度及其意义。虽然大多数人还没有在自己的工作流程中采用 AI,但我认为更少的组织试图问:在存在这种智能的世界里,我是否以正确的方式设计了我的组织?所以第二章的一部分是关于这种持续重新想象未来组织可以是什么样子,以及我们如何构建工具来推动它。

And it's evolved to this place where it solved Navier-Stokes, a problem that mathematicians, the best in the world for the past 200 years, could not solve. And then I think 3 weeks later on a random Thursday night it dropped proofs to several hundred more problems. I think most people still haven't internalized just the pace of progress and what it means. And while most people still haven't yet adopted AI in their own workflows, I think even fewer organizations have sought to ask: in a world in which this kind of intelligence exists, have I designed my organization in the right way? And so part of chapter 2 is about this continued reimagination of what an organization of the future can look like and how we build the tools to power that.

Host

是的,似乎坐在我们现在的位置,递归自我改进正在实验室内部发生,至少据报道如此,未来 9 到 24 个月将是相当惊人的时刻,模型能力将起飞。然后与此同时,你发布了一个疯狂的数据:目前前 1% 的支出者花费了 80% 的 AI token 支出。对吗?

Yeah, it seems like sitting where we're at now, recursive self-improvement is happening inside the labs, at least reportedly, and the next 9 to 24 months are going to be pretty astonishing moments where model capabilities just take off. And then simultaneous to that, you released this crazy stat about the top 1% of spenders spending 80% of the spend on AI tokens right now. Is that right?

Eric

完全正确。

That is exactly right.

AI采用的不均衡 The uneven adoption of AI

Host

所以看起来,在这个疯狂的事情发生的同时,甚至初创公司彻底改变做事方式的实际能力却停滞不前。我的意思是,那些打算采用它的人,至少在这 24 个月的时间框架内,似乎已经接受了它并且正在做,但有 99% 的人目前说,实际上我们没有——这像是某种技能问题之类的。这真的很疯狂。这不是边缘现象。这实际上是影响社会、影响 GDP 的那种使命。这将是我一生中经历过的最具变革性、最奇异的几年,也许也是我们所有听众在思考和谈论这件事的几年。

And so it seems like simultaneous to this crazy thing that's happening, the actual ability for even startups to radically change the way they do things is flat. I mean, it's the people who were going to adopt it at least in this 24-month time frame seemingly have embraced it and they're doing it, but there's this large 99% that currently says actually we have no—this is like some kind of skill issue or something. It's very wild actually. This is not a fringe thing. It's actually kind of a societal, GDP-affecting sort of mission. It is going to be the most transformative, bizarre years I've lived through in my life and maybe all of us listening and thinking about this.

Eric

未来已来,只是分布不均,这是肯定的。

The future is here, it is not evenly distributed, that is for sure.

Ramp的内部工具 Internal tools at Ramp

Host

所以在 Ramp 内部,你一直在做这些。你有很多内部工具:有 Inspect,一个编码智能体;有 Research,是你的内部 AI 分析师;然后 Glass 算是你的知识和共享工作流系统。这些东西为你做了什么,或者它们是怎么来的?你是怎么组建团队的?你自己就是那 1% 的人。所以对于那些属于 99% 的人,告诉我们你自己迄今为止的转型故事。

So internal to Ramp, that's what you've been doing. You have a number of internal tools: you have Inspect, a coding agent; you've got Research, which is your internal AI analyst; and then Glass is sort of your knowledge and shared workflow system. What have those things done for you and/or how do they even come about? Like how did you structure the team? You're one of the 1% yourself. So for the people who are part of the 99, tell us the story of your own transformation so far.

改变对人来说很难 Change is hard for people

Eric

改变对人来说非常难。也许举个例子:我们在金融领域待了很长时间。所以有时我们会从投资银行或私募股权公司以外的地方招人,他们来到公司后,金融团队——也许你的金融团队,如果你在听这个——会刁难这些新人,方式就是他们可能已经非常擅长在 Windows 上用 Excel,知道所有快捷键,非常厉害。然后你刁难他们,给他们一台 Mac,再给他们 Excel,他们就说,‘我不知道怎么用这个。太复杂了。’某种程度上这很有趣,但它展示了一个基本事实:即使是学习使用你可能已经用了多年的工具,在一个稍微不同的操作系统上,对人们来说也可能很难做出改变。这就是为什么许多金融组织仍然是初创公司里唯一还在 Windows 机器上运作的团队。所以某种程度上,去对你团队里的任何成员,或者你生活中的家人说,‘嘿,我知道你这样工作 20 年了。你现在应该直接用 AI。’这是一个非常大的要求。这比同一个程序、不同的操作系统要大得多。真的很难。所以某种程度上,真正去思考并给一些人宽容,让他们弄清楚如何真正重新学习他们所知道的一切。

Change is very hard for people. Maybe to give an example: we've spent so much time in the finance world. So sometimes we'll hire people outside of an investment bank or a private equity firm and they'll show up at a company and one of the ways that finance teams—maybe your finance teams, if you're listening to this—will haze those people is they probably have become really good at Excel on Windows and they know all the hotkeys and they're great at it. And then you haze them by getting them a Mac and you give them Excel and they're like, 'I don't know how to use this. It's too complicated.' And in some sense it's funny, but it sort of exhibits this basic truth of even learning how to use the tool that you've maybe used for multiple years in a slightly different operating system can be hard for people to make that change. And that's why many finance orgs still—they're like the only team still operating on a Windows machine inside of a startup. And so in some sense to go to any member of your team or life, your family and say like, 'Hey, I know you've been doing work this way for 20 years. You should just use AI now.' It's like a really big ask. It's like a lot bigger than just same program, different OS. It is really hard. And so in some sense actually thinking through and giving some people the grace to figure out how do you actually relearn everything that you kind of know.

1%的人用80%的token The 1% using 80% of tokens

Host

你必须在团队层面这样做吗?你说,‘好吧,我们要从——’我的意思是工程是显而易见的起点,你一直有不可思议的,你知道,IMO 金牌,最优秀的工程师。他们说,‘嘿,实际上我们可能就开始这么做了。’公司内部发生了什么?

Did you have to do that at the team level? You're like, 'Okay, we're going to start with—' I mean engineering is an obvious place to start and you've always had incredible, you know, IMO gold, the best of the best engineers. They were like, 'Hey, actually we're just going to start doing this probably.' What happened inside of the company?

Eric

公司内部发生的就是世界上正在发生的事:1% 的人使用了 80% 的 token,他们同时也成为最高效的人。你会看到,不夸张地说,100 倍工程师,这些人效率极高,他们更快,而其他人落后,一些职能进展快,开始破坏组织的某些部分。例如,你有工程师可能效率提高 10 倍或 100 倍,设计师效率提高 2 倍,产品人员可能也跟上了,效率提高 5 倍,而公司传统上有这些比例,比如六到八个工程师对一个产品经理对一个设计师,如果一个职能的生产力跟不上,就会有大问题。所以你开始有所有这些痛点,工程师跑得比其他部门快。

What happened inside of the company is what is happening in the world, where 1% of the people were using 80% of the tokens, who were also becoming the most extraordinarily productive people. You'd have, not exaggerating, 100x engineers, people who were just so radically productive and they're being faster, and others who are behind, and some functions moving fast and it would start breaking parts of the organization. For example, you'd have engineers who maybe were 10 or 100x as productive, designers who were 2x as productive, product people maybe got with the program and they were 5x as productive, and companies traditionally have these ratios, like six to eight engineers to one product manager to one designer, and it's a big problem if the productivity of one function or another doesn't get there. And so you'd start having all these pain points where engineers were outrunning other parts of the company.

招聘冻结与AI采用 Hiring freeze and AI adoption

Host

有意思。

Fascinating.

Eric

你会看到有人突然增长更快,对于容量模型,我们不得不雇所有这些人人,你会收到这些招聘请求,存在这种反相关。你用 AI 越多,通常效率越高,所以你不需要那么多人,如果你根本不用,也许你会说你是——

You'd have people who suddenly were growing faster and for capacity models like we had to go hire all these people and you'd get these hiring requests where there was this inverse correlation. The more AI you used, often you were radically more efficient so you didn't need as many people, and if you weren't using it at all maybe you would say that you were—

Host

是的。

Yeah.

Eric

而你会申请很多人。所以到了今年早些时候,我们就说——顺便说一下,这个业务每年增长超过 100%——我们说整个季度不招聘。我们这样做的部分原因是,首先招聘需要大量工作:找人、面试、入职,在变得更有生产力之前你会先降低生产力。但第二,我不希望学习 AI 只占你 5% 的时间。它需要成为你的核心能力。整个问题是如何让人们达到那个顿悟时刻,最终看到并理解它,再也不想回去。所以一部分是如何创造这样的条件?

And you would request a lot of people. And so it got to this point where earlier this year we just said—and by the way this business is growing more than 100% per year—we said we're not hiring for a whole quarter. And part of why we did that is first of all hiring takes a lot of work: to find people, interview people, onboard people, you get less productive before you get more productive. But two, I didn't want learning AI to be a 5% of your time. It needs to be your core competency. And the whole question was how do you get people to this aha moment where finally you see it and you understand it and you never want to go back. And so part of it was how do we create the conditions to do that?

AI与数据之间的墙 The wall between AI and data

Eric

另一方面实际上是——你在小型初创公司不会遇到那么多这些问题,那里你刚起步,整个环境可能都在本地设备上。但在大型组织内部,情况是你说去用 Claude Code,然后销售团队的人说好,我来做,然后他们下载它,然后他们必须打开终端,他们就说终端是什么,我必须安装这些包,他们终于完成了,然后他们必须问它或某人,我能把它连接到 Databricks 或 Snowflake 或 Google 或 Slack 或 Stripe 吗,你什么也做不了。就像我认为大多数人对 AI 的体验是,这是一种难以置信的外星智能,有一堵墙,像一个小纸条,你可以写个便条聊天,你可以传个便条给它,然后它传个便条回来给你,这与你看那些不可思议的工程师获得 100 倍生产力提升相去甚远。

Another side of it is actually just—you don't encounter these problems as much in small startups where you're just getting going and your whole environment maybe is on your local device. But inside of large organizations what happens is you say go use Claude Code, then someone from your sales team says fine I'll do it and then they download it and then they have to open terminal and they're like what's terminal and I have to install these packages and they finally get it done and then they have to ask it or someone of can I connect it to Databricks or Snowflake or Google or Slack or Stripe and you can't do anything. It's like I think most people's experience with AI is it's this unbelievable alien intelligence and there is a wall and like a little slip and you can write a note chat and you can pass a note to it and then it can pass a note back to you which is so far from when you looked at how these unbelievable engineers who were getting 100x productivity gains.

Inspect及其影响 Inspect and its impact

Eric

我想到 Inspect,它为什么这么有用,它占我们在 Ramp 所有拉取请求的 85% 以上。每个工程师的拉取请求数量仍然每月增长约 40%,所以这是速度的巨大提升。它能看到工程师能看到的所有数据。它能访问代码。它有自己的虚拟机、浏览器,什么都有。它不仅仅是一个伟大的智能,它与你的团队融合在一起。

I think about Inspect, what makes it so useful, and it's 85% plus of all the pull requests that we have at Ramp. The number of pull requests per engineer is still growing like 40% every month roughly, and so it's a radical velocity improvement. It can see all the data that the engineer can see. It has access to the code. It has its own virtual machines, its browser, it's everything. It's not just a great intelligence, it's merged with your team.

矩阵愿景 The Matrix Vision

Eric

我一直记得《黑客帝国》里的那个场景,基努醒来后说‘我会功夫了’。或者如果你看过《阿凡达》,整个前提就是主角知道他的祖先所知道的一切。所以它几乎有一种超级视野、超级洞察力。

I always remember that scene in The Matrix where Keanu wakes up and he's like, 'I know kung fu.' Or if you've watched Avatar, the whole premise was that the main character knows all the things that his ancestors know. And so it has this almost super vision, this super insight.

Eric

几乎有一种愿望是,实际上我们如何让每个人都能轻松做这些事情,知道公司所知道的一切?当你收到客户关于某个 bug 的投诉时,Inspect 可以去读那封邮件。它知道所有最近的代码变更,它能定位问题。它可以去写更改,然后给客户回邮件。所以与其回到我们开始时的那些容量模型,你得去雇 30 个人,你可以在 15 分钟内修复它,所有这些复合优势就都来了。

There was almost this desire to say, actually, how do we make it easy for everyone to do these things, to know all the things that the company knows? When you get a complaint from a customer about a bug, Inspect can go and read that email. It knows all of the recent code changes, it can locate it. It can go write the change and it can email the customer back. And so instead of going back to where we started with these capacity models, you have to go hire 30 other people, you can just fix it in 15 minutes and all these compounding advantages just go.

构建框架与集成 Building Harnesses and Integrations

Host

听起来像是在这些工具周围构建你自己的 harness。

Sounds like building your own harnesses around these things.

Eric

不仅仅是 harness,还有集成,比如身份验证是内置的,然后整个公司的标准化,就像,哦对,我们只用 Inspect。你知道,有些人在这里用这个工具,另一些人在那里用另一个工具。我不强制人们使用任何工具。你可以用任何你想用的。我们确实如此,有一种令人难以置信的实验文化。Anthropic 有一篇关于 Claude Code 早期日子的精彩文章,Ramp 是,我相信,第一个 alpha 客户,并尝试与他们一起修补。Cognition,我们是 Devon 的第一个客户。Cursor,最早获得这些伟大工具的人之一,我们让人们选择他们自己认为最好的。这就是目标。

It's not just the harness, it's also the integrations like the authentication is built in and then standardization across the whole company like, oh yeah, we just use Inspect. You know, have some people over here using this thing and some other people using this other thing. And I don't mandate any tools for people. You can use whatever you want. And we do, like, there's an incredible culture of experimentation. There was a great post from Anthropic about the early days of Claude Code and Ramp was, I believe, the first alpha customer and like trying to tinker with them. Cognition, we were the first customer of Devon. Cursor, one of the first like people have access to great tools and we let people choose what in their own opinion is the best for them. And that's the goal.

AI ROI与衡量 AI ROI and Measurement

Host

那么你会推荐什么?我的意思是,作为人们,你知道,外面有观众可能管理着千人组织,但他们告诉他们的工程师,你知道,你们可以有两个 Claude Max 账户,就这样。比如,你知道,你的政策是什么?你会推荐什么?比如,如何,你知道,99% 如何成为 1%?

So what would you recommend? I mean, as people, you know, there are people out here watching who actually probably run thousand-person organizations, but you know, they're telling their engineers, you know, you can have two Claude Max accounts and that's it. Like, you know, what's your policy? What would you recommend? Like, how, you know, how does the 99% become the 1%?

Eric

并不是说我们的做法适合所有人。但这就是为什么我们发布了 Inspect 和 Glass。我坚信如果其他组织能够更高效,世界会更好。所以有些人欢迎看看我们在内部构建的东西是否适合你。但更重要的是,我认为你提出的问题背后的问题是,我实际上认为 AI 的 ROI 时代已经到来,对吧,在某种意义上,当你设定一个固定预算,你只说我们只花 1000 美元或只花 100 美元或不管什么数字,就这样。在我看来,这有点愚蠢。我这样说吧:有些人在花 1000 美元,我为此感到兴奋。另一些人,我则无比愤怒。如果有人花 1000 美元赚 10000 美元,我会想方设法给他们钱。如果有人花 1000 美元做每周准备却毫无价值,我不知道,也许我们不应该那样做。所以只问我们花了多少钱,这是一种完全不完整的看待世界的方式。这就是 Ramp 现在痴迷的地方:你如何真正去衡量,甚至在语义上,不仅仅是誰花了钱,并在所有工具中在一个地方展示,而是理解,好吧,我们在一个新的着陆页上花了 5 美元,那个新着陆页转化率提高了 3%,每月多带来 5000 美元或不管什么。所以我们的构建方式可能与他人不同的一部分是,我们不卖 token,我们买它们。我们寻求衡量它们,我们帮助公司将他们的支出与结果联系起来。

Not that the way that we do things is right for everyone. But this is why we've released Inspect and Glass. I believe really strongly if other organizations are able to be more productive, the world would be better. And so some folks just welcome to see if what we built internally is great for you. But more than that, and I think like the question behind the question you're getting at is like, I actually think like the ROI era of AI is here, right, in some sense where when you put like a flat budget or something and you just say we're going to spend only $1,000 or only $100 or whatever the number is and that's it. It's sort of in my opinion like a bit of a silly way of doing things. Like I put it this way: there are some people where I'm thrilled that they're spending $1,000. Others where I couldn't be more angry. If someone spends $1,000 to make $10,000, I'm trying to throw money at them. If someone's spending $1,000 to do their weekly prep for no value, I don't know, like maybe we shouldn't do that. And so like it is a completely incomplete way of looking at the world of just asking how much did we spend. And this is where this is now the obsession of Ramp of how can you actually go measure even semantically not just like who spent and show this in one place across all of your tools but understand, okay, we spent $5 on a new landing page, that new landing page picked up, you know, converts 3% better, which generates $5,000 more per month or whatever it is. And so part of where we build in maybe a different way than others is we don't sell tokens, we buy them. We seek to measure them and we help companies connect back what they're spending to the outcomes.

YC广告时间 YC Ad Break

Host

YC 的下一批正在接受申请。你有一个创业想法吗?在 ycombinator.com/apply 申请。永远不嫌早,填写申请会让你的想法升级。好的,回到视频。

YC's next batch is now taking applications. Got a startup in you? Apply at ycombinator.com/apply. It's never too early and filling out the app will level up your idea. Okay, back to the video.

Glass作为公司大脑 Glass as Company Brain

Host

我想是的。特别是 Glass,听起来它像是你的,嗯,其他人可能只是把人扔给 Codex 或 Claude Code,然后你得想办法连接系统,而 Glass 是你的公司大脑,就像,你知道,你可以用它搭配前沿模型,但它有所有的集成,然后它是不是像所有关于公司的信息,比如分析、财务、客户支持,所有关于运营 Ramp 的事情都在里面,然后谁可以获得访问权限,你知道它是如何工作的?

I guess so. Glass specifically, that's it sounds like it's kind of your, um, other people might just throw people at Codex or Claude Code with and it's like you figure out how to hook up the systems, whereas Glass is your company brain like it, you know, you can use a frontier model with it but it's got all the integrations and then is it like all info about the company like analytics like finance you know customer support like all the things about running Ramp are like all in there and then who gets access and you know how does it work?

Eric

你知道我们有 1930 人。我看了这里过去 24 小时,1925 人使用了它。

You know we're 1930 people. I was looking here last 24 hours, 1925 used it.

Host

所以人们用它来工作。

So people use it to do their work.

Eric

你知道,完成他们的整个工作。

You know do their whole work.

Host

并且了解正在发生什么,并且

And to know what's going on and

Eric

正是如此。这里面有很多设计决策。首先,正如你所说,它自动连接到公司使用的所有核心系统,对吧?所以无论是你的数据仓库、分析套件,也许你用 Zendesk 做客户支持,所有数据都被拉进去。所以在公司层面,可以更完整地了解实际发生的情况,你可以在数据集之间建立连接。但对于个人,你拥有你应有的访问权限,这因公司而异。有共享技能。所以如果销售团队的人找到了一种更好的研究和场景转化的方法,它会显示出来并弹出,成为其他人的默认。有模型选择。所以如果一个开源模型系列更好,或者 Claude 或 OpenAI,你可以使用任何你想要的。有更智能的默认值。我敢说,也许大部分在前沿模型上的花费可以用更简单的模型来完成。有智能的权衡内置其中。例如,我们的后台编码智能体,有时我们乐意让过程和推理多花 10% 的时间,以换取价格降低 50%。所以有更好的默认值。我认为当你开始获得所有这些数据时,你不仅不用为每个人反复学习同样的教训支付 10 倍的费用,而且你实际上能够在组织层面发展和开始看到,是什么决策、什么推理让这家公司与众不同。不仅仅是个人在做什么,而是 Ramp 运营的独特视角和方式。在功能上,你在教你的系统这样做。所以即使一个很优秀的人做了很棒的工作然后离开了,

Exactly. And there's a lot of design decisions that went into this. First of all, to your point, it connects automatically to all of the core systems that a company uses, right? So whether it's your data warehouses, your analytic suites, maybe you're using Zendesk for customer support, all that data is getting pulled into it. And so at a company level, there's much more of a complete view of what actually is occurring and you can make connections between data sets. But for an individual, you have the access permissions that you should have and that can vary by company. There are shared skills. So if someone from the sales team figures out a better way to do research and a scene conversion, it shows up and pops up becomes the default for others. There's model selection. So if an open model series is better or Claude or OpenAI, you can use any one that you want. There's smarter defaults. I would venture to say that maybe the majority of spend on frontier models could be done by a simpler model. There's smart trade-offs built into it. For example, our background coding agent, sometimes we are happy to have the process and reasoning take 10% longer for a 50% cut in price. And so there's better defaults. I think when you start to get all this data, not only are you not paying 10 times for everyone to keep relearning the same lesson, but you're actually able to develop and start to see at the organizational level, what is the decision-making, what is the reasoning that makes this company different. Not just like what is individual person doing, but like what is the unique viewpoint and way that Ramp operates. And functionally you're teaching your system to do that. So even if someone great does great work and leaves,

Host

它就在一个技能文件里。

It's in a skill file.

Eric

它就在技能文件里。它在那里。

It's in the skill files. It's there.

Host

我想其中一些是,嗯,这不仅仅是为了让每个人都能更好地工作。它实际上对你和 Kareem 作为联合 CEO 非常有用。

I guess some of this is, um, it's not just for everyone to be able to do their work better. It's actually really useful for you and Kareem as co-CEOs.

流程可见性与标准化 Process Visibility and Standardization

Host

你能够真正深入查看,比如,那个流程是什么?Glass 里是否有东西会说,好吧,对于这个或那个工作流,退款给客户只有一个流程?当人们只是拼凑东西时,你可能得不到那种可见性,做一件事可能有五种不同的方式。

You're able to actually dive in and see, like, well, what is that process? Is there something in Glass that says, like, okay, well, there's one process for refunding a customer for this or whatever workflow? When you have people just cobbling things together, you might not get that visibility, and there might be like five different ways to do something.

Eric

是的,完全正确。你实际上能够理解标准操作程序是什么,以及事情能够完成的原因是什么。但你也可以开始看到存在差异。说到 Glass 如何帮助我们开发出一款已经落地的出色产品,也许在财务方面我举个例子。每家公司——Ramp 做了 YC——不幸的是都需要有费用政策,这源于萨班斯-奥克斯利法案和审计要求等等。但这实际上意味着每次有人购物时,都应该有人来审核——那个人不是你——而在大型组织中,这最终导致高薪人员只是为所有人审核费用。这有点毫无意义,这些人并没有在思考费用或政策内容,或者哪里不完整或错误,而且实际上经常出错。他们要么直接提交并批准一切,要么浪费时间去做。我们为政策智能体构建了一个类似的产品,它实际上了解你完整的费用政策。它能够以 99% 的准确率去批准、拒绝,或者说政策不完整。每次事情被批准、更改或覆盖时,它都会学习。如果事情有冲突或不完整,它会随着公司的变化而编写和调整政策。所以正如你所说,公司不是离散的规则引擎——它们会进化。思维风格和推理可以进化,但你可以追溯血统并理解决策过程,而有时大公司是官僚的、难以理解的机构。

Yeah, that is exactly right. You functionally are able to understand what are the standard operating procedures, what are the reasons of how something is able to get done. But you also can start to see that there are discrepancies. To speak to how Glass helped us develop a great product that has come to life, maybe on the finance side of the house I'll give you an example. Every company—Ramp does YC—unfortunately needs to have expense policies, kind of falls out of Sarbanes-Oxley and audit requirements and all kinds of things. But it functionally means that every time someone makes a purchase, someone is supposed to review it—that's not you—and it's culminating in large organizations where you have highly paid people just reviewing the expenses for everybody. And it kind of makes no sense, and these people aren't thinking about expenses or what the policy says or where it's incomplete or wrong, and are actually wrong quite often. Either they'll just submit and approve everything or waste their time doing it. And we had built a similar product for the policy agent where functionally it knows your full expense policy. It's able to, at 99% accuracy, go and either approve it, decline it, or say policy is incomplete. And every time that things are approved or changed or overridden, it learns. And if things are in conflict or incomplete, it writes and adjusts the policy as the company changes. And so to your point, companies are not discrete rule engines—they evolve. The style of thinking and the reasoning can evolve, but you can trace the lineage and understand the decision-making in a way where sometimes large companies are bureaucratic, inscrutable organizations.

透明度与组织开放性 Transparency and Organizational Openness

Host

是的。我想对于所有选择采用这种公司大脑模型的组织来说,这是一个非常有趣的透明时刻,因为它非常开放。就像以前是孤岛,副总裁或高管可能能够非常严格地控制信息的流动,而这通常是你在金融机构可能看到的事情之一,并不太能很好地服务用户。对吧?

Yeah. I guess this is a very interesting moment for transparency inside all organizations that choose to do this company brain model because it's very open. Like before it was silos, and a VP or an executive might be able to have very tight control over what information goes, and it usually is probably one of those things you might have seen at a financial institution that doesn't quite serve its users well. Right?

Eric

绝对如此。其中一些是根本性问题:一个组织处于什么模式——你是在寻求创造价值吗?我认为对于一家刚起步的公司,没有收入,没有资产,没有代码,没有知识产权,所以一切导向都是我们必须去创造,我们必须去做事情。我认为这些组织默认往往非常开放。你雇佣聪明人,你为他们的判断力付钱。

Absolutely. Some of these are fundamental questions of what mode is an organization in—are you seeking to create value? I think at a company just getting started, there's no revenue, there's no asset, there's no code, there's no IP, and so everything is oriented on we must go create, we must go do stuff. And I think these organizations tend to be very open by default. You hire smart people, you pay them for their judgment.

Host

初创公司很棒——这就是为什么初创公司很棒。

Startups are great—that's why startups are great.

Eric

这就是为什么有创造性破坏,我想是这样。我认为这样更好。

That's why there's creative destruction, I think so. I think it's better.

Host

嗯,我的意思是,流程最终会使公司僵化,因为它们有东西要保护。

Well, I mean, the processes end up ossifying companies because they have something to protect.

Eric

没错。而且有趣的是,它们也是 20 世纪的一种现象——这是对的——你知道人类每周只有 40 小时,我们大脑在任何时刻只能保持七加减二件事。然后,我的意思是,今天的模型,即使更便宜的也能处理一百万 tokens——那就像三本《哈利·波特》。

That's right. And then they're also interestingly a 20th century sort of phenomenon in that—this is right—you know human beings only have their 40 hours a week and you know we can keep seven plus or minus two things in our brain at any given moment. And then I mean, the models today even the cheaper ones can do a million tokens—that's like three Harry Potter books.

Host

是的,没问题。

Yeah, no problem.

Eric

即使在纯粹的记忆案例中,这也像是超人的。所以组织当然会改变,它们不会成为这些孤岛式的、受保护的信息孤岛,只向一个积累信息的人汇报。我的职业生涯始于微软,所以那是我对大型组织的体验——你有这些领地,有一个控制中心,有长篇的书讲述那个时代的微软多么僵化和糟糕。那现在可以结束了。

That's like superhuman even in a pure rifle shot memory case. So of course the organizations are going to change, and they're not going to be these siloed, protected information silos that roll up to a single person who's then accumulating. I started my career at Microsoft, so that was my experience with large organizations—you had these fiefdoms that had a locus of control, and there's long books written about how ossified and terrible Microsoft of that era was. That can end now.

AI助力大公司快速行动 AI Enabling Large Companies to Move Fast

Host

我的意思是,即使在大型组织中,你认为至少那 99% 拒绝这样做的人中的一部分——你认为那些人实际上是这些领地和信息孤岛的受害者吗?这需要自上而下的文化变革,要么 CEO 不愿意采取,要么董事会不愿意支持 CEO,要么高管们集体相互争斗,所以他们就是不做?

I mean, even in large organizations, do you think that at least some percentage of the 99% that refuses to do it—do you think that those are actually victims of these fiefdoms and sort of the information silos? It requires a top-down cultural change that either the CEO is not willing to take, the board is not willing to support the CEO, or the executives en masse are at war with one another and so they just won't do it?

Eric

我认为这完全正确。我确实认为这是一个有趣的问题:如果你看看谁幸存下来,企业如何继续并转型。历史非常清楚:公司的几率并不好。默认情况下,在足够长的时间范围内,你会破产。在足够长的时间范围内,如果你不创新,你就会死。有趣的是,哪些企业继续生存以及为什么。我确实认为这些组织的主要死因是某种消化不良:协调成本太高,信息过于孤岛化,他们不做艰难的决定。你说得对,我实际上认为有了 AI,一些重量级公司可以开始像轻量级一样行动。你开始看到迹象,大公司突然增长更快,这在功能上违背了重力。你越大,速度越慢。你看到这些公司的拐点。这就是 Ramp 发生的一部分,现在我们年收入超过 20 亿美元,而且增长显著更快——现在比一年前快了 25% 以上。我谢谢你,但我认为许多其他组织也能做到。

I think that is exactly right. And I do think it is this interesting question of if you look at who survives and how do businesses go on and transform. The history is very clear: the odds are not good for companies. By default, over a long enough time horizon, you will go bankrupt. Over a long enough time horizon, if you don't innovate, you will die. And it's interesting to question what businesses continue to survive and why. And I do think the primary cause of death of these organizations is some kind of indigestion: the coordination costs get too high, information is overly siloed, they don't make the hard calls. And you're right, I actually think with AI, there are heavyweight companies that could start moving like lightweights again. And you're starting to see signs of that where large companies are suddenly growing faster, which functionally defied gravity. The larger you'd get, the slower you would get. And you're seeing these inflection of companies. And that's part of what happened at Ramp, where now we do over $2 billion a year in revenue and we're growing significantly faster—more than 25% faster now than we were a year ago. And I thank you, but I think a lot of other organizations can do this.

形态与销售信任 Form Factor and Selling Trust

Eric

我还认为许多组织陷入了形式因素的困境:他们卖什么以及如何围绕它组织。有时 Ramp 被比作美国运通,因为我们从发卡开始。我觉得那家公司的历史非常迷人。它始于 1850 年。他们的第一个产品基本上就像驿马快信。你可以把东西放在马车后面,他们会送到俄亥俄州或其他地方。那到底和 21 世纪的全球金融集团有什么关系?在某种意义上,我认为他们不像我们不记得的其他公司,他们某种程度上理解他们实际上在做的是出售信任。如果你给他们黄金或一些银行票据,它会到达你需要它去的地方。

And I also think a lot of organizations get stuck in the form factor of what do they sell and how do they organize around that. Sometimes Ramp is compared to American Express because we started by issuing cards. And I find the history of that company very fascinating. It started in 1850. And their first product was basically like the Pony Express. You could put stuff in the back of a wagon and they'd get it to Ohio or whatever. What in the heck does that have to do with a global financial conglomerate of the 21st century? And in some sense, I think that they, unlike the other companies that we don't remember, sort of understood that what they actually were doing was selling trust. That if you gave them gold or some bank note, it would get to where you needed it to be.

公司的使命与演变 The Mission and Evolution of Companies

Eric

所以当然合理的是,随着世界变大,他们应该进入旅行支票行业。他们有卡吗?这样如果你去商家而现金用完了,没问题,他们会照顾你。许多伟大的公司理解他们的使命是什么,他们在追求什么。最终,每个公司都在超越他们的使命,试图把 1 美元变成更多,他们对如何实现这一目标的细节持开放态度。

And so of course it's logical that as the world became bigger, they should go into the traveler's checks industry. Do they have cards so that if you go to a merchant and you're out of cash, no problem, they'll take care of you. And a lot of great companies understand what their mission is, what they are in pursuit of. Ultimately, every company is trying beyond their missions to convert $1 into more, and they're open to the details as to how they're going to get there.

Eric

我认为这是一个需要极大重新想象我们为什么这样工作的时期。现在 AI 的争论非常有趣,因为基本上可以提出两种论点,而且两者都非常极端。因为公司的规模是由公司不同部分之间的摩擦决定的。我想这就像科斯教授有科斯定理,对吧?然后基本上它可能朝两个方向之一发展。我的意思是,它可能在不同行业朝两个方向都发展。我不太确定。比如一种观点是快速起飞和末日论 AI 的角度,认为不可避免会有一个实验室遥遥领先,然后他们会捕获所有未来价值的光锥,对吧?所以那种论点认为实际上理想的公司数量是一个,它将满足人类的所有需求之类的。然后另一端是可能出现超级专业化,你知道世界通常是垄断或双头垄断,因为人们某种程度上受制于品牌和七加减二件你能记住的事情,但智能体实际上可以带来超级专业化。所以没有理由不能有成千上万甚至数万家公司,每家赚十亿或数十亿美元,超级专业化地提供正好适合某一群人或公司的东西,你知道这是 YC 所希望的光明面版本。

And I think this is a period that requires enormous reimagination of why do we work this way? The AI argument is so interesting right now because you could make basically two arguments and both of them are very extreme. Because the size of a firm is defined by the friction of the different parts of the firm. I think it's like this professor Coase has Coase's theorem, right? And then basically it could go one of two ways. I mean, it could go both ways in different industries. I'm not really sure. Like one view with the fast takeoff and sort of the doomer AI angle is like it's inevitable that one of the labs will run away with it and then they will capture a light cone of all future value, right? And so that's the argument that actually the ideal number of firms is one and it'll serve all the needs of humanity or something like that. And then way on the other end is like there could be hyper specialization and you know the world is often like monopoly or duopoly because people are sort of beholden to brands and seven plus or minus two things you can keep in your head, but agents actually can bring about hyper specialization. So there's no reason why there can't be thousands or tens of thousands of firms each making a billion or billions of dollars like hyper specializing in providing exactly what is right for a particular set of people or companies, and you know that's sort of the white pill version of this that YC is hoping for.

AI能力与经济影响 AI Capabilities and Economic Impact

Host

哦,我想谈谈这个,因为如果你看看数据中发生的情况,我实际上认为过去两个月非常迷人,而且我非常强烈地支持第二种世界观。

Oh, I want to talk about this because I actually think the last two months have been incredibly fascinating if you look at what's occurred in the data, and I actually believe support worldview number two very strongly.

Eric

我也是。我的意思是我们正在押注它,但你知道我们也必须为它而战。我们必须为它而战。而且看,我认为人们对实验室将变成什么的许多先验和想法都基于一个时期,那时 GPT-3、3.5,我不知道,很难将其智能直接转换为 IQ,但肯定远低于人类平均水平,也许四系列开始达到略低于或接近人类平均水平,五和六已经超越了。这太疯狂了。

Me too. I mean we're betting on it, but you know we got to fight for it too. We have to fight for it. And look, I think that so much of people's priors and ideas about what the labs are going to turn into were based in a time period where GPT-3, 3.5, I don't know, it's hard to translate its intelligence directly to IQ, but certainly far below the average human, to maybe the four series started to get to a little bit below or around average human, five and six have surpassed. It's insane.

Host

如果你有一份需要 120 IQ 的工作,每次你从 70 到 90 到 115,回报都会非常惊人。但问题是,当你从 120 到 150 对于不需要它的任务时会发生什么。换句话说,你知道,会计可能用 120 模型就能解决。我们没问题。我们在会计方面没问题。我有朋友解雇了他们的注册会计师,现在 Claude 做他们的税。

And if you have a job that requires 120 IQ and every time that you go from 70 to 90 to 115, the returns to that are going to be incredible. But there's this question of what happens when you go from like 120 to 150 for tasks that don't require it. Like put differently, you know, it might be that accounting is solvable with 120 model. Like we're good. We're good on accounting. I have friends who fired their CPA and Claude does their taxes now.

Eric

注册会计师 Claude。我要直接说的是,没有人谈过这个,但在数据中非常清楚,那就是过去两个月杰文斯悖论打破了。

CPA Claude. And what I would say really directly that no one has talked about this but it's very clear in the data, which is over the last two months Jevons' paradox broke.

Host

太迷人了。

Fascinating.

Eric

就像过去那段时间,这些是编造的数字,假设你将智能的价格降低 10%,它变得便宜 10%,你会看到智能使用量的增加超过抵消。

Like it used to be for that period of time, these are made up numbers, let's say you discount the price of intelligence by 10%, it becomes 10% more affordable, you would see the increase in consumption of quantity of intelligence used more than offset.

Host

就像铁路价格下降,铁路的使用量激增。就是这样。

Like railroads came down in price and they spiked up the usage of railroads. This is it.

Eric

你知道我想象铁路、电信,就像任何网络化并将人们聚集在一起的事物都有重复效应,而 AI 直到最近都是这样。

And you know I imagine railroad, telecom, like there's a repeating effect for anything that networks and brings people together, and AI was that until recently.

Host

直到最近。谁知道呢,也许新的模型家族可以再次打破它。它可能不是永久的,也许我们现在处于局部最大值,还有方法推动采用。我的意思是我认为这可能与我们刚才谈到的有关。就像递归自我改进来了,模型能力这样发展,然后突然社会吸收它的能力或多或少是一个限制。

Until recently. And who knows, maybe new model families can go and break that again. It may not be permanent, maybe we're at a local maxima for now, and there's ways to go and drive adoption. I mean I would argue it's probably related to what we just talked about. It's like RSI is here, the model capabilities are going like this, and then suddenly the capability of society to absorb that is more or less a limit.

Eric

就像所有将要成为 AI 狂热者的人都已如此,他们将从中受益。是的。但然后你知道那 99% 呢?它有点被内部政治或非我发明或很多只是我们现在太舒适了,不想处理它。有些事情会改变。有些事情会改变,这给了我对世界的希望,但尤其是对初创公司。有趣的问题是现在,假设智能的成本应该非常明显地继续下降,也许未来智能几乎免费,那么什么是可能建造的?什么样的企业,什么样的问题可以解决,一旦你跨过这个成本阈值,你可以从不可能到可能?你能从零到一,突然新的创新企业可以存在吗?经典地,我认为资本主义的故事是发明的生态系统的价值远大于这些发明的创造者所捕获的价值。所以我认为未来将是一个完全美好而迷人的十年。

Like all the people who were going to become AI psychos have done it and they're going to benefit from this. Yes. But then you know what about the 99%? It's a little bit captured by internal politics or not invented here or lots of just like we're too comfortable right now, we don't want to deal with it. Something's going to change. Something's going to change and it gives me hope for the world but also especially for startups. Like the interesting question is now, let's say that the cost of intelligence should very clearly continue to drop, maybe intelligence will be almost free in the future, what is possible to build? What kind of businesses, what kind of problems can be solved that once you've crossed this threshold of cost you can go from impossible to possible? Can you go from zero to one and suddenly new innovative businesses can exist? Classically, I think the story of capitalism is the value to the ecosystems of inventions are much larger than the value captured by the creator of these inventions. And so I think it's going to be just a completely wonderful and fascinating decade ahead.

内部工具与产品影响 Internal Tools and Product Impact

Host

我想回到你刚才说的关于 Glass 的事情,因为你知道我们之前谈到 Glass 就像你改变了所有事物的运作方式。它让你的增长提升了 25%,超过了之前的速率,这在大型规模上相当显著。制造 Glass 和这些内部工具在多大程度上实际上显著改变了使用 Ramp 的人们的核心产品?在第二季度,有 7,000 个拉取请求是由非工程师创建的,这很棒,对吧?有趣的部分是,大概两年前,非工程师创建的拉取请求数量实际上为零。

I want to get back to what you were saying about Glass, because you know I think we were talking about Glass like you changed the way everything works. It popped your growth by 25%, like in excess of what you know the previous rate, which is pretty significant at large scale. How much of that, like making Glass and these internal tools, has actually changed the core product for people who use Ramp significantly? In the second quarter, there were 7,000 pull requests that were created by non-engineers, which is great, right? And the interesting part is like probably two years ago the number of pull requests created by non-engineers was functionally zero.

Eric

是的。

Yeah.

Host

你就是不能,我的意思是你得会用 Git,你知道谁会让你提交?

You just can't, I mean you're going to like get Git and you know who's going to let you commit?

Eric

是的,没错。谁会让你,谁会批准这个人的代码?其中大部分来自客户体验团队。

Yeah, exactly. Who's gonna let you, who's going to approve like this person's code? The majority of those are coming from the customer experience team.

Host

是的。

Yeah.

Eric

这在公司里经常发生。有人报告了一个 bug,你怎么办?也许你发送一个回复,希望它能工作。你收到 10 个这样的东西。有人在写用户故事。

And this happens in companies all the time. Someone reports a bug, what do you do? Maybe you send a response, hope it works. You get 10 of these things. Someone's writing a user story.

构建软件的成本 The Cost of Building Software

Eric

你进入一个传统组织。也许他们还在用敏捷开发或某种糟糕的方法论。他们写一张工单,试图说服产品经理这值得做。产品经理排定优先级,交给设计师,也许最终到工程师手里,两个月后你才得到这个东西。

You go in a classic organization. Maybe they're still on agile or some horrible methodology. And they write up a ticket, they go try to convince the product manager it's worth it. They prioritize it. They give it to a designer. Maybe it makes it to an engineer and two months later you have this thing.

Host

是啊。

Yeah.

Eric

或者更常见的是,它进入 bug 数据库,被标记为 P3 不予修复,就像再见。它永远不会被修复。

Or more often like it goes into a bug database and gets marked P3 won't fix like goodbye. It'll never get fixed.

Host

去某个地方等死。

Go somewhere to die.

Eric

schwab.com 上有些东西永远都是那样。简直难以置信。

There are whole things in schwab.com that like will always be like that. It's like unbelievable.

Host

有趣的是,在一个做某事的成本从 3 个月降到几分钟的世界里,事情开始不断变好。你开始快速行动,软件开始变得更直观易用。人们能真正感受到更多的自主权。你不只是把奶酪从这里移到那里,而是可以真正去构建它。听起来有点像 80 年代的丰田。

What is interesting is in a world where the cost of doing something goes from 3 months down to minutes, things start getting better all the time. You start moving quickly like the software starts getting more intuitive to use. People can actually feel a lot more agency. You're not just move the cheese from here to there, but actually you can go build it. Kind of sounds like Toyota in the '80s.

Eric

是的。记得突然之间这些日本车开始涌入美国市场,它们是更好的车,不会经常抛锚,这一切都来自流程创新,你知道,还有改善,就是工厂流水线上的工人可以随时改变他们的流程,这就像让你那个正在与经历真正痛苦的人交谈的客户支持人员,进去按一个按钮,然后实际上,哦,那件事现在完成了,就像字面上同一天。其影响是深远的。第一,人们对彼此的期望提高了。你知道,说不,或者我们将在周一而不是今天做这件事,就变成了你为什么要那样做?你今天就可以做。所以每个人的水平都提高了,你开始看到,我想想过去五个月,每个月我们都有一个重大的产品发布。过去在 Ramp,我们行动非常快。我们计算我们存在了多少天。如果幸运的话,你一年会有一两个重大发布。你看到这体现在更快的增长率上。你看到这体现在销售人员的配额上升上。所以不要为他们感到难过。他们赚得更多。所以当你把这些事情做对时,你知道,我不认为这是 Ramp 独有的现象,但我觉得公司不仅赚更多钱,而且你更快地向更多人交付产品,人们的生意变得更好。我喜欢的一个数据,我更自豪的不仅是 25% 的增长率。而是一年前 Ramp 的中位客户能够每年削减约 5% 的费用。现在是 7.5%。如果你能少浪费 50% 的美元,这意味着什么?顺便说一句,美国公司的平均利润率是 8%。你能否加速的不仅是你服务的公司,而是希望整个世界的增长?

Yeah. Remember like suddenly these Japanese cars started flooding the American market and they were way better cars and they didn't break down all the time and it all came from like a process innovation with you know and Kaizen like just this idea that the person on the factory line can change their process on the fly that's the same as like letting you know your customer support person who's like talking to someone experiencing a real pain go in and press a button and then actually oh that thing's done now like literally literally like that same day and the implications are profound. One, people's expectations of each other go up. You know, saying no or we're going to do this, you know, on Monday instead of today like becomes why would you do that? You can go do this today. And so everyone like their game is raised and you start seeing like I think about the last five months every single month we've had a major like product release. It used to be at ramp and we move very fast. We count the days of how old we've been. You would have one or two major releases a year if you were lucky. And you see this expressed in faster growth rates. You see this expressed in terms of for sales people their quotas go up. So then don't feel bad for them. They're making more. And so when you get these things right, you know, I don't think this is a phenomenon unique to RAMP, but like I just think company not just makes more money, but you deliver faster products to more people and people's businesses get better. And I like the stat which I'm more proud of is not just the 25% growth rate. It's that the median customer on ramp a year ago is able to cut their expenses by about 5% per year. It's 7.5%. And what does it mean if you can waste 50% fewer dollars? By the way the average company in the US has a 8% profit margin. Can you actually accelerate the growth of not just the companies you're serving but hopefully the whole world?

Host

是的。你知道,我当时发了一条有点争议的推文。就像要么你作为记录系统公司死去,要么你活得足够长成为 AI 框架。这引起你的共鸣吗?你们是否已经为客户发布类似 AI 框架的东西?你认为那条推文是对还是错?

Yeah. You know, I tweeted something that like was a little controversial at the time. It's like either you die a system of record company or you live long enough to become an AI harness. Does that resonate with you? Like are you guys releasing AI harness like things for your customers yet? Do you think that that tweet is correct or false?

Eric

我认为是对的。实际上,我同意。在我们这个行业,我们有点奇怪,有时人们认为 Ramp 是一家金融科技公司,但那不是我们。我们更像一家生产力公司。就像如何少浪费美元?如何让你的美元和小时发挥更大作用?如果你相信 AI 是这种激进的生产力工具,如果你相信有 100 倍工程师和 10 倍工程师,你怎么能不尝试让每个人都成为 100 倍工程师,你必须开始考虑 AI 框架。Glass 就是一个 AI 框架。

I think it is correct. Actually, I'm going to agree with it. We are strange in our industry of like sometimes people think of ramp as a fintech and like that's just not us. We're more of a productivity company. It's like how do you waste fewer dollars? How do you make your dollars and hours go further? And if you believe that AI is this radical tool for productivity, if you believe that there are 100x engineers and 10x engineers, how could you not try to make everyone 100x engineers, you must start thinking about AI harnesses. Glass is an AI harness.

Host

但客户还不能在 ramp.com 界面内使用他们自己的 Glass。或者可以吗?

But customers can't use their own glass yet inside the ramp.com interface. Or can they?

Eric

所以 Ramp 有一些部分,比如如果你想以智能体方式交互,如果你想调整政策,如果你想做定制研究,

So there's pieces of ramp where like if you want to interact agentically, if you want to go and adjust policies, if you want to go and do custom research,

Host

我确定,就像 MCP 一样。

It's like an MCP, I'm sure.

Eric

没错。顺便说一句,MCP 的使用量现在每周增长约 10%。我们已经有了 2 年。使用量正在爆炸式增长,我认为在不久的将来,它将成为 Ramp 的主要使用方式。但你可以实际使用,截至本周,我们正在为客户的自己的公司引入 Glass。

Exactly. And by the way, MCP usage is growing about 10% per week right now. We've had it for 2 years. It is exploding in terms of usage and I think will become a dominant way that ramp is used in the not too distant future. But you can actually use as of this week, we're onboarding customers onto Glass for their own companies.

Host

哦,那很酷。

Oh, that's cool.

Eric

是的。所以,它正在发生。

Yeah. So, it's happening.

Host

它正在发生。

It's happening.

招募最优秀的人才 Hiring the Best and the Brightest

Host

我们稍微谈谈招聘。你说过,因为人们在青少年时期在 Minecraft 上建立业务而雇用他们,这是一种好方法。你还有哪些其他方法能够雇用最优秀和最聪明的人?

Let's talk about hiring a little bit. You said, you know, hiring people because they built businesses on Minecraft as teenagers is, you know, one of the good ways. What are some of the other ways that you've been able to sort of hire the best and the brightest?

Eric

你听说过萨顿定律的故事吗?

Have you ever heard the story of Sutton's Law?

Host

嗯。

Mm-m.

Eric

有一个银行劫匪,应该是在 30 或 40 年代,故事是这样的,你知道,他会去抢银行,然后进监狱,又抢银行,又进监狱,有一天一个记者抓住他。他在法院外面,他们说:“你为什么这么做?你为什么抢这家银行?”他回头看着记者说:“因为钱在那里。”

So, there was a bank robber and it must have been in like the 30s or 40s and the story goes, you know, he would go rob banks and he'd go to jail and he robbed banks again and go to jail and one day a reporter, you know, catches him. He's outside of the courthouse and they're like, "Why'd you do it? Why did you rob this bank?" And he looks back at the reporter and he says, "Cuz that's where the money is."

Host

这有点好笑。就像,你知道,你还能抢哪里?

Which is sort of funny. It's like, you know, where else were you going to rob?

Eric

和任何好的需求函数一样。如果你想找到擅长构建东西的人,你去那些展示他们喜欢构建东西的地方。Minecraft 就是一个很好的例子。如果你在 14 岁时在私人服务器上建立业务,其他人会使用并付费,这清楚地表明你能做出人们想要的东西。如果你想找到擅长解决算法问题的人,那就去看看谁在 16 岁时赢得数学竞赛或 IMO 或 IOI。我的意思是,说出那个函数。有点像人们只是卡在看简历和人们如何展示东西上。我只是对此不那么感兴趣,我更感兴趣的是

Good demand function as any. And if you want to find people who are good at building stuff, you go to places that demonstrate that they like building things. Minecraft is a great example. If you are building a business as a 14-year-old, in private servers that other people will use and pay money for, it is a clear sign that you can make things that people want. If you want to find people who are good at solving algorithmic problems, well go look and see who at the age of 16 is winning at you know math counts or the IMO or the IOI. I mean name the function. Like it's a little bit of people just get stuck at looking at resumes and how people present stuff. And I'm just not as interested in that as I am of

Host

有没有证据表明某人过去独自做过伟大的工作?

Is there any proof that someone has done great work in the past on their own?

Eric

看,我认为有些人没有推动自己去做,你仍然可以通过更标准的机制找到人,但某种意义上,萨顿定律有点聪明,而且在研究中,你知道,招聘也是一样的

And look I think there are people who haven't pushed themselves to do it and you still can find people through more standard mechanisms but in some sense Sutton's law is sort of brilliant and under you know study it's the same thing for hiring you know

Host

是的,为什么经常我认为你知道我们做我们必须在 10 分钟左右做出选择,然后有趣的意识到在某个时刻就像人们一直在广播他们是谁,他们说注意我注意我,所以你知道有时你只需要看

Yeah why often I think you know we do we have to make choices in 10 minutes or so and then the funny realization at some point was like people are just broadcasting who they are at all times and they're like notice me notice me so you know sometimes you just have to look

Eric

你就会看到他们是什么。

And you'll see what they are.

给年轻自己的建议 Advice to Younger Self

Host

你会看到他们的本质。回顾所有的成功和所有的逆境,你建立了多家伟大的企业,如果有一件事——如果你能通过时光机给 18 岁或 20 岁的自己发送一条信息,你会发送什么?

You'll see what they're about. Looking back on all the success and all the adversity you've gone through, building multiple great businesses, what's one thing—if you could send a time machine back, or just one message to the 18 or 20-year-old version of yourself—what would you send?

Eric

我记得,这也是 YC 对我们如此具有变革性的部分原因:我从未创办过初创公司。我不太理解它们是什么、如何运作——看起来如此陌生。无论你是刚刚准备开始的人,还是正在创业路上的创始人,你都会不断遇到障碍和僵局。我认为在生活中——无论是因为学校里的成绩,还是当你打开 Instagram 时,人们试图展示他们很酷、他们知道自己在做什么——每个人都在努力表现得一切尽在掌握,这导致你不去寻求帮助。我认为我通过 Parabus、通过 RAMP、通过试图改变自己和他人生活得到的最有力的建议之一是:作为创始人、CEO 或普通人,你能说的最有力的话之一实际上是:“我需要你的帮助。我正在试图弄清楚这件事。这是我遇到的问题。”这太罕见了。我认为对于创始人来说,去说“我确实有这个问题。我需要这个建议。”我记得非常清楚的一件事是,当我想到办公时间过去和现在对我来说意味着什么——其实就三个问题:你这周增长了多少?你最大的问题是什么?你打算怎么解决?我认为这实际上是试图将人们拉入一种思考模式:我需要帮助的地方在哪里?展示你的推理,但要寻求帮助——对处理这个问题感到自在。如果你在早年或任何年龄就习惯了这样做,你可以通过直面问题来解决问题。所以我认为我花了很多时间等待寻求帮助,而它已经走了很长一段路。

I remember, and this is part of why YC was so transformative for us: I had never started a startup. I didn't really understand what they were, how they operated—it seemed so foreign. Whether you're someone who's just waiting to get started or you're a founder on your journey, you're going to encounter obstacles and impasses all the time. I think that in life—whether it's because of the grades you get at school or when you open Instagram, people are trying to show that they're cool, they know what they're doing—everyone's trying to look like they've got it all together, and it leads you to not ask for help. And I think one of the most powerful pieces of advice I've gotten through Parabus, through RAMP, through just trying to change my life and other people's is: I think one of the most powerful words you can say as a founder, CEO, or just a person is actually, "I need your help. I'm trying to figure this out. Here's the problem I have." And it's so rare. I think for founders to go and say, "I actually have this problem. I need this advice." And I think it was one of the things I remember vividly when I think about what office hours were and are for me—it's really just three questions: How much did you grow this week? What's your biggest problem? And what are you going to do about it? I think it's actually trying to pull people into a mode of thinking: where do I need help? Show your reasoning, but ask for it—feel comfortable just working through that. And if you get used to doing that early in your life or at any age, you can solve things by confronting it. So I think I spent a lot of time waiting to ask for help, and it's gone a long way.

Host

Eric,非常感谢你加入我们。

Eric, thanks so much for joining us.

Eric

谢谢,Gary。

Thanks, Gary.

互动版:中英对照全文 + 针对本期提问 →