Mercor 以 200 亿美元估值完成 7.5 亿美元融资:CEO Brendan Foody 谈重塑人类工作方式

Mercor Raises $750M at $20B Valuation: CEO Brendan Foody on Reinventing How Humans Work

布伦丹·富迪 Brendan Foody · Sourcery with Molly O'Shea · 2026-11-10 · 约 54 分钟 · 原视频 ↗

打开互动全文版(中英对照 + 朗读 + 问答)→

本期速览 · Overview

Brendan Foody 讲述 Mercor 如何打造强化学习环境与 AI 项目经理,重新定义人类参与经济的方式。

Brendan Foody explains how Mercor is building RL environments and AI project managers to redefine how humans contribute to the economy.

要点 · TL;DR

核心观点 · Key points

反共识 · Contrarian takes

本期章节 · Chapters(共 28)

全文 · Full transcript(中英对照)

引言与增长 Introduction and Growth

Host

Brendan,欢迎来到 Sorcery。

Brendan, welcome to Sorcery.

Brendan

很高兴来到这里。谢谢你邀请我,Molly。

Excited to be here. Thanks for having me, Molly.

Host

我的意思是,我们刚刚还在聊这个,但我们第一次一起散步大概是一年多以前,从那以后你们增长了多少?

I mean, we were just talking about this, but we first went on a walk probably over a year ago and how much have you grown since?

Brendan

我们增长了,我想那是去年九月。所以我们在中九位数的基数上增长了超过 5 倍。

We've grown I think this was in September of last year. So we've grown by more than 5x on a mid nine-figure basis.

Host

这太疯狂了。

That's crazy.

Brendan

确实非常疯狂。

It's been pretty wild, yeah.

万亿美元公司愿景 Vision for a Trillion-Dollar Company

Host

那么你们打算如何成为一家万亿美元公司?

So how are you going to become a trillion-dollar company?

Brendan

好问题。我认为首先要从人类将如何为经济做贡献倒推。因为现在大多数工作相对重复,人们会重复地回复客服工单,或重复地审阅法律合同,或重复地编辑播客以及他们习惯的各种格式,而未来将更像是一次性训练一个智能体来自动化这些重复性任务。因此,我们在企业中看到的是,从所有这些重复性任务和过去的工作方式转变,转向人们通过创建评估、更新技能文件、训练智能体来自动化所有这些重复性工作来为经济做贡献。所以,如果我们认同人类为经济做贡献的方式正在彻底转变,那么显然引领这个行业、定义这个行业,并围绕这项工作如何发生构建所有技术,就创造了打造一家传奇的品类定义型公司的机会。这就是我们从第一天起一直专注的事情,就是解决这个人类如何融入快速演变的经济的问题。

Great question. Well, I think it starts with working backwards from how will humans contribute to the economy? Cuz right now, most work is relatively repetitive where people will repetitively respond to customer support tickets or repetitively redline a legal contract or repetitively edit a podcast and all the different formats that they're used to when in the future, it's going to much more look like training an agent one time how to automate these repetitive tasks. And so what we see emerging within enterprises is the shift away from all these repetitive tasks and how work has been done in the past and moving towards people contributing to the economy through creating evals, updating skills files, training agents how to automate all this repetitive work. And so if we subscribe to the fact that the whole way that humans contribute to the economy is transforming radically, clearly leading that industry, defining that industry, and building all of the technology around how that work happens creates the opportunity to build a legendary category-defining company. And that's what we've always been focused on from day one is just solving this problem of how do humans fit into the economy as that is evolving ever so quickly.

Mercor运营结构 Mercor's Operational Structure

Host

让我们分解一下。让我们分解一下 Mercor 的运营结构。我知道你想谈谈强化学习环境,也称为 RL 环境。那么如果你分解一下,Mercor 的产品是如何工作的?它们如何协同?

So let's break that down. Let's break down how Mercor is operationally structured. So I know you want to talk about reinforcement learning environments a bit, also known as RL environments. So if you break that down, how the products of Mercor work? How does it all fit in?

Brendan

是的,所以为了生产给定的数据类型,我们大致有三个主要产品。第一个产品是我们最初构建的人才平台,我们有超过 500 万人在积极寻找工作,每天支付超过 400 万美元。70% 的人来自推荐。因此,在调动软件工程师、医生、律师等为我们想要在经济中创建的任何特定类型数据方面,存在很多网络效应。然后,我们将这些专家引入 Studio,我们的数据平台,在那里他们有一个 AI 副驾驶以及所有智能体式质量控制和自动化能力,用于创建 RL 环境,填充所有反映他们在工作中会做的上下文、所有任务、所有验证器等。第三个平台是我们的 AI 项目经理,我们基本上自动化了与如何管理这些项目中的所有专家相关的所有手动工作。所以类似于我们会有项目经理为人们设定 KPI,说谁做得好,谁做得不好,给他们反馈,现在我们有了智能体自主运行这些项目,招聘人员,确定数据质量,并最终对客户结果负责。

Yeah, so there's broadly three main products that we have to produce a given data type. The first product is what we built initially, our talent platform, where we have over 5 million people actively looking for work, payout over $4 million a day. 70% of the people come from referrals. So there's a lot of network effects associated with how we can mobilize software engineers, doctors, lawyers, etc. for any specific type of data that we want to create in the economy. From there, we onboard those experts to Studio, our data platform, where they have an AI co-pilot and all of the agentic QC and automated capabilities for how they would create an RL environment, populating all of the context that mirrors what they would do in their job, all the tasks, all the verifiers, etc. And then the third platform is our AI project manager, where we basically automate all of the manual work associated with how we would manage all the experts on these projects. So similar to how we would have project managers setting KPIs for people, saying who's doing a good job, who's doing a bad job, giving them feedback, we now have agents that are autonomously running these projects, hiring people, determining the data quality, and ultimately being accountable to the customer outcomes.

支付与规模 Payments and Scale

Host

所以,你们每天支付 500 万美元?

So, you're doing 5 million in payouts a day?

Brendan

是的。呃,现在远远超过那个数了,是的。

Yeah. Uh now now well over that, yeah.

Host

那是如何运作的?

How does that work?

Brendan

它使用 Stripe Connect 作为后端。呃,但是团队非常出色。我的意思是,这太不可思议了,因为我们早期意识到的一件事是,我们当时与这个领域的另一个参与者合作,他们公开出现了很多支付问题,而我们为他们提供人才。我们收到了大量支持工单,我们为他们雇佣了超过 1000 人。人们通常来自第三世界国家。他们没有拿到工资,难以养家糊口。我们真的收到了很多关于此的支持工单。所以这真的是一个巨大的催化剂,我们需要与实验室建立直接关系。我们需要能够对我们平台上专家的体验负责,并确保他们按时获得报酬。我们需要过度投资于如何建立一个世界级的支付团队,从不犯错,始终确保人们得到支付。到目前为止,你知道,我们历史上已经支付了大约 10 亿美元,这是一个相当陡峭的增长曲线。而且我们没有遇到任何重大问题或类似情况。所以这是一个不可思议的过程,我们如何扩展支付团队。实际上,我写了一些关于此的博客,因为那里有很多有趣的技术问题。

It uh uses Stripe Connect as the back end. Uh but but but the the the team is spectacular. I mean it's been incredible cuz one of the things we realized early on is we were working with another um uh player in the space that was having a lot of payout issues very publicly where we were providing talent to them. And we got flooded with support tickets that uh we'd hired over 1,000 people for them. People often in third world countries. And they were not getting paid and struggling to put food on the table for their families. It's literally we were receiving many support tickets about that. And so really that was one of the large catalysts for we need to have a direct relationship with the labs. We need to be able to be accountable for the experience of the experts on our platform and making sure that they get paid on time. And we need to over invest in how do we have a world-class payments team that never makes mistakes and always makes sure that people are paid out. And so to date we've you know paid out just about a a billion dollars uh historically um which is on a pretty steep ramp. And uh we we haven't had any major issues or uh of of that sort. And so it's been an incredible process of how we scale the payments team. Actually, I've written some blogs about this um cuz there's a lot of interesting technical problems uh in there.

商业模式与RL环境 Business Model and RL Environments

Host

哇。我的意思是,对于一些不了解商业模式的人来说,扩大到 Mercor 的更广泛范畴,你们不关注 GMV。这不是一个市场。那么,商业模式如何适应?

Wow. I mean for some people who don't understand the business model going out to like the broader umbrella of Mercor, you're not focused on GMV. It's not a marketplace. So, how does the business model fit?

Brendan

是的。所以,我认为关于这个业务最大的误解是,人们把我们看作像 Uber 这样的市场,我们只是将专家与工作匹配,然后收取费用,而实际上专家是我们交付给实验室的数据集——RL 环境的一个组成部分。所以一个 RL 环境包括三个部分。第一部分是所有应用,就像 Google Workspace 和其他流行应用的克隆,人们希望教智能体在经济中使用这些应用。第二部分是世界,即这些应用中的所有内容。想想消息、幻灯片、文档、表格等。第三部分是任务,即提示和评分标准,类似于教授创建评分标准来写论文,模型可以尝试问题,然后根据表现获得评分,从而从中学习。而自动化整个经济的障碍在于,我们如何覆盖所有 RL 环境的完整分布,以反映经济分布中的一切。所以覆盖人们使用的所有不同应用、所有不同世界和所有不同任务。因此,最终这些 RL 环境就是我们交付给客户的最终数据。

Yeah. So, I think the largest misconception about the business is that people think about us as a marketplace like Uber where we would just match experts with a job and then take a fee on when actually the experts are one component of the data sets, the RL environments that we deliver to the labs. And so uh an RL environment includes three parts. The first part is all of the apps, which are like clones of Google Workspace and other popular applications that uh people would want to uh teach agents how to use within the economy. The second is the worlds, which is all the content within those apps. So think the messages, the slides, the docs, the sheets, etc. And then the third part is the tasks, which are prompts and rubrics, similar to how a professor would create a rubric to create an essay where the model can attempt the problem and then get a grade on how well it did that it's able to learn from. And the barrier to automating the entire economy is how do we cover the full distribution of all of the RL environments that mirror everything in the distribution of the economy. So cover all of the different apps people use, all of the different worlds, and all of the different tasks. And so ultimately these like RL environments is the end data that we deliver to our customers.

智能体与专家生产力 Agents and Expert Productivity

Host

显然,这在很大程度上依赖于模型,但智能体进入整个 AI 系统后,是如何改变、甚至可能增加你们正在做的事情的?

Obviously a lot of this is predicated on models, but how has the advancement of agents coming into the entire AI system changed and maybe increased what you're doing?

Brendan

是的,我认为一大好处是,智能体让花时间创建数据和训练智能体变得高效得多。举个例子,以前一个人一小时能带来 X 量的模型改进,现在一小时可能带来三到五倍的改进,因为他们可以使用 AI 副驾驶来提高构建所有这些数据集的效率。但最终,随着我们让专家更高效,出现了杰文斯悖论的情况:需求呈指数级扩张,因为归根结底,99% 的知识仍然存在于人们的头脑中。如果我们只是让人类教模型变得更有效、更高效,那只会增加整个经济中人们教模型的量,因为这最终回到我一开始提到的原则:让人们做重复性工作没有意义,他们应该直接训练模型如何做。这是一种更具可扩展性和生产力的贡献经济的方式。

Yeah, I think one of the large benefits is that agents have made it much more productive for someone to spend time creating data and training an agent. As an example, say an hour of someone's time would previously get X amount of model improvement. Now an hour of their time might get three to five X that, because they're able to use an AI copilot to make them more productive in how they build out all of these data sets. But ultimately, what's happened as we've made the experts more productive is that it's been this case of Jevons paradox where the demand just expanded exponentially, because ultimately 99% of knowledge still lives in people's heads. And if we just make it more effective and productive to have humans teaching models, that just increases the amount that people teach models throughout the economy, because it ultimately ties back to the principle I mentioned at the beginning, which is that it doesn't make sense for people to do repetitive work. They should just train the model how to do it. And that's a far more scalable and productive way to contribute to the economy.

品类持久性 Durability of the Category

Host

那么,对这个类别的一个普遍评论是,它是一个非持久性类别。训练模型、利用所有这些知识工作、所有投入其中的东西,只会持续这么长时间。你看到这一点了吗?显然你不认为这会发生,但你的反驳论点是什么?

So one comment on this category in general is that it's a non-durable category. That training models, using all this knowledge work, all this kind of stuff that goes into it is only going to last for so long. Do you see that? Obviously you don't see that happening, but what are your arguments against that?

Brendan

我认为这最终取决于你的超级智能时间线,因为人类教模型如何做事的市场受限于人类能在数字世界或物理世界中做而模型不能做的事情的数量。所以只要经济中还有人类能做的事情——可能是加入 Google Meet 并进行有说服力的对话,或者在湿实验室进行科学实验,或者修理汽车,无论是什么——我们就需要人类来测量那个前沿。模型改进的主要方式是,它们需要某个停滞点来知道它们在哪里犯了错误。它们如何测量这一点?然后如何将其作为可以学习的奖励?这有点类似于人类学习的方式。想象一下,你在英语课上学习如何写出更好的文章,而老师只是说你需要给自己的所有文章评分。没有其他人能给你反馈。你只能自己评分,对吧?很难衡量超出你自己能力前沿的东西。这就是为什么我们需要反馈。可能在某些领域,你可以基于清晰的信号学习。数学就是一个好例子,因为有计算器可以检查你的工作。但在经济中的绝大多数领域,我们学习和改进的方式是从老师、教授、经理、同事等那里获得反馈。这与模型的学习方式非常相似,也是这些实验室的大多数顶尖研究人员预计它们在未来十年将继续学习的方式。

I think it ultimately comes down to your superintelligence timelines, because the market for humans teaching models how to do something is bound by the amount of things that humans can do that models can't do across either the digital world or the physical world. So as long as there are things in the economy that a human can do—and it might be something like joining a Google Meet and having a persuasive conversation, or running a scientific experiment in a wet lab, or repairing a car, whatever it is—we need humans that measure that frontier. The primary way that models improve is that they need some stasis point for what is the place where they're making a mistake. How do they measure that? And then how do they use that as a reward that they can learn from? It's sort of analogous to the way humans learn. Imagine if you are in an English class learning how to write better essays and the teacher just says that you need to grade all your own essays. And no one else can give you feedback. You just got to grade your own essays, right? It's pretty difficult to measure what is beyond the frontier of your own capabilities. And that's why we have feedback. There might be some domains where you can learn based on a clear signal. Like math is a good example because there's calculators where you can check your work. But in the vast majority of domains in the economy, the way that we learn and the way that we improve is getting feedback from our teacher, our professor, our managers, our colleagues, etc. And that's very similar to how models learn and how most of the top researchers at these labs expect that they'll continue to learn for a decade to come.

赞助商信息 Sponsor Messages

Host

本集由 Brex 赞助,我的最爱。你成为你所花费的。我拒绝把时间花在不应该存在的工作上。费用报告、追收据和手动结账。从 Vercel、OpenAI、Anthropic、Granola 和 Deepgram 这些构建未来的公司都做出了同样的选择。它们都运行在 Brex 上。Brex 是智能金融平台,将卡、费用和银行整合到一个堆栈中,内置智能体式金融。AI 智能体自动处理费用,在支出发生前执行政策,并在几分钟内完成结账。这就是为什么 Sorcery 运行在 Brex 上,这样我就可以花时间在构建上而不是忙碌的工作上。是时候使用 Brex AF 了。了解更多在 brex.com/sorcery。那是 brex.com/s o u r c e r y。再见。

This episode is brought to you by Brex, my favorite. You become what you spend on. And I refuse to spend my time on work that shouldn't exist. Expense reports, receipt chasing, and manual closes. The companies building what's next from Vercel, OpenAI, Anthropic, Granola, and Deepgram all made the same call. They all run on Brex. Brex is the intelligent finance platform that combines cards, expenses, and banking into a single stack with agentic finance built in. AI agents that handle expenses automatically, enforce policy before spend happens, and close your books in minutes. That's why Sorcery runs on Brex so I can spend time on building and not busy work. It's time to get Brex AF. Learn more at brex.com/sorcery. That's brex.com/s o u r c e r y. Bye.

Host

从握手到签名更快,Harvey,为法律工作构建的 AI 平台,受到 80% 的 Am Law 100 和超过 20% 的财富 500 强的信任。今天在 harvey.ai 了解更多。那是 harvey.ai。

From handshake to signature faster, Harvey, the AI platform built for legal work, is trusted by 80% of the Am Law 100 and over 20% of the Fortune 500. Learn more today at harvey.ai. That's harvey.ai.

Host

AI 在真实专业工作中的表现如何?Mora 的 AI 生产力指数(APEX)通过每次主要前沿模型发布来跟踪它。与 Harvey 和 Cognition 等合作伙伴一起构建,在 mercor.com/apex 查看前沿的实际水平。

How good is AI at real professional work? Mora's AI Productivity Index, or APEX, tracks it with every major frontier model release. Built with partners like Harvey and Cognition, see where the frontier actually stands at mercor.com/apex.

专家寻找与评估 Finding and Evaluating Experts

Host

我不得不问,你们如何找到这些专家,以及如何评估他们?

I have to ask, how do you find these experts and how do you evaluate them?

Brendan

所以,70% 的专家通过推荐进来。我们也使用其他渠道,但这是我们最关注的,因为我们发现被推荐的人往往质量最高。然后评估他们是一个非常酷的过程。这实际上是我们最早的巨大突破之一,因为当 Adarsh Suri 和我在大学时,在 2023 年 4 月,就在 ChatGPT 问世后,我们构建了第一个 AI 面试官。第一个有大约七到八秒的延迟。它会向 GPT-4 发起 API 调用,基于简历做一些预处理,发起 API 调用,说出问题,询问人们,然后根据评分标准或一组标准评估面试,看看人们表现如何。随着我们扩大规模,我们意识到劳动力匹配中的许多挑战都归结为可扩展的数据收集。一个类比是,如果你招聘斯坦福 AI 博士,你可能会见五个人,其中一个人会相当不错;而如果你从印度偏远地区招聘软件工程师,你有时需要见数百或数千人才能找到真正世界级的工程师,但他们确实存在,而且许多人在我们的团队中。因此,当我们能够扩大面试数量以解决这些人才密度差异时,这确实是我们如何获得出色人才的主要驱动力之一。还有分析人们可能没有考虑的所有不同过滤器,比如谁推荐了他们,他们在 GitHub、Figma 或其他地方做的个人项目是什么。所有这些事情都被拉入上下文然后分析。所以我们现在做了超过五百万或我们已经做了超过 700 万次面试,看到事情规模化发展真是疯狂,因为我们现在创造了历史上增长最快的职业类别。

So, 70% of them come in through referrals. There's also other channels that we use, but that's the one we're most focused on because we find that people that are referred tend to be the highest quality. And then evaluating them is a really cool process. And this was actually one of our first large breakthroughs because we built the first AI interviewer when Adarsh Suri and I were in college in April of 2023, right after ChatGPT came out. And the first one had like seven or eight second latency. It would make an API call to GPT-4 and that based on do some preprocessing based on the resume, make an API call, say the question, ask people, and then evaluate the interviews according to a rubric or a set of criteria to see how well people do. And the realization we had as we scaled that up is that so much of the challenges in labor matching just come down to scalable data collection. And an analog for that is that if you're hiring Stanford AI PhDs, you can probably meet with five people and like one of them's going to be pretty good versus if you're hiring software engineers from a remote place in India, you need to meet with sometimes hundreds or thousand people to find like the really world-class engineers, but they are there and many of them are on our team. And so as we were able to scale up the amount of interviews that we did to solve these differences in talent density, that was really one of the large drivers of how do we get spectacular people. And also analyzing all the different filters people might not be considering of what are who referred them, of what is their personal projects that they do on GitHub, on Figma, or whatever it is. All of these things getting pulled into context then analyzed. So we now do over five or we do we've done over 7 million interviews and it's been pretty crazy to see things scale up as we have now created the fastest growing job category in history.

Host

太疯狂了。

It's crazy.

Brendan

是的。

Yeah.

Host

真的太快了。

Literally so fast.

Brendan

非常超现实。

It's very surreal.

快速增长与市场领先 Rapid Growth and Market Leadership

Brendan

是的,我的意思是,作为一个参照,现在我们每天支付大约 500 万美元,而两年前,24 个月前,我们一年才支付 100 万美元。

Yeah, I mean, as a frame of reference, so now if we're paying out, call it $5 million a day, as of two years ago, 24 months ago, we were paying out a million dollars a year.

Host

我的天啊。

Oh my god.

Brendan

所以就像,嗯,我觉得公司里很多人并没有完全意识到,这种情况以前从未发生过,嗯,这是一个非常特殊的机会,我们比任何人都先发现了这个巨大的市场,我们成为了品类领导者,以及我们如何,嗯,创造这个非常重要的新品类。

So it's like, um, I think that many people at the company don't fully realize just that, like, that's never happened before, um, and that it's this very special opportunity where we found this giant market before anyone else and we've become the category leaders and how we, um, create this really important new category.

Host

是的,太棒了。我的意思是,我有很多问题。嗯,其中一个我超级好奇的问题,稍微岔开一下,就是看看你的团队。你们都是二十出头。

Yeah, that's amazing. I mean I have so many questions off of that. Um one of the questions that I was like super curious about just to deviate for a second is like you look at your team. You're all early 20s.

Brendan

不是所有人,但我们很多早期员工,是的。

Not all of us, but a lot of our early people, yeah.

Host

二十出头,对吧?你们是大学辍学生。你们能够抓住并真正把握住多年来 AI 和科技领域增长最快的新品类之一,如果不是最快的话。同时,你们有非常大的竞争对手被 Meta 这样的公司收购。还有其他竞争对手每隔一个月就冒出来,不知怎么地就达到了十亿美元的收入。然而,你们克服了数据泄露之类的挑战,但仍然是市场第一。那么,你们的文化和创始团队有什么特质,让你们能够继续坚持并获胜?

Early 20s, right? You're dropouts of college. You've been able to grasp and really hold on to one of the fastest, if not the fastest growing new category in AI and in technology in years. And meanwhile like you have very large competitors get bought out by like Meta. You have other ones that are popping up like every other month and somehow reaching like a billion in revenue. And yet like and you guys like you know you overcome challenges like the data leak and that kind of thing, but still you're like number one in the market. So, what is it about your culture and your founding team that you guys are able to continue to persevere and win?

Brendan

我想说,我们可能有三个价值观,它们是公司里明确的价值观,但真正被付诸实践,产生了巨大影响。第一个是高标准,就是有非常高的招聘门槛,确保我们不妥协,这相对于其他公司是一个巨大的差异化因素,团队人员的质量是无与伦比的。第二个是我们称之为“时间旅行者”的价值观,就是非常注重长期导向,活在未来,想象五年后世界会是什么样子,然后我们如何从那里倒推?这就是我们比其他所有人才网络更早发现市场的方式,我们一直非常了解研究的前沿,以及前沿正在发生的事情是 AI 原生公司和应用层公司发生的事情的领先指标,是全球经济中发生的事情的领先指标,非常深思熟虑地思考事物如何系统地流经这些采用曲线,嗯,这非常重要。第三个是客户痴迷。我们是一家非常勤奋、痴迷于客户的公司,我们为客户移山,因为他们建造地球上最重要的技术,我们嗯非常致力于帮助他们实现这一目标的使命。我想说,这三件事在我们发展市场领导地位的过程中是最具形成性的。

I would say there is probably three values that we have that are explicit values in the company, but they're really operationalized that have been super impactful. The first one is high standards, just having like a really high hiring bar and making sure that we don't compromise that has been a huge differentiator relative to everyone else in the company and the quality of the people on the team is unparalleled. The second thing is a value we call time travelers, which is about being super like long-term oriented of just like living in the future of what is the world going to look like in five years and how do we work backwards from there? And that was how we sort of found the market before all of these other uh talent networks and we just have always understood the frontier of the research super well and how what's happening at the frontier is a leading indicator of what happens at the AI native companies and application layer companies is a leading indicator of what happens in the global economy and just being very thoughtful about how things systematically flow through these adoption curves um has been really important. And then the third thing is customer obsession. Like we are an incredibly hardworking customer obsessed company where we move mountains for our customers because they build the most important technology on earth and we um are extremely committed to the mission of how we help them to do that. And I would say those three things have been uh the most formative as we've developed market leadership.

动力与竞争 Motivation and Competition

Host

你会说你用的是清洁燃料,还是现在有一点救赎燃料?

Would you say you have clean fuel or at this point do you have a little bit of redemption fuel?

Brendan

第一个是什么?清洁燃料?

What was the first one? Clean fuel?

Host

清洁燃料。

Clean fuel.

Brendan

你如何定义清洁燃料?

How do you define clean fuel?

Host

就像,好吧,Dylan Field 在他们被收购之前跟我解释过,我忘了,是 Adobe 被取消了吗?就像他当时只是靠纯粹的清洁燃料运行。没有什么让他心怀不满。然后那件事发生了,他有了很大的不满,他说,现在我要赢了。然后带着,你知道,一些进取心去做。你是靠清洁燃料运行,还是你有什么不满?

Like okay, Dylan Field explained this to me before their acquisition by I forget, was it Adobe got scrapped? Like he was just running on pure clean fuel. Like nothing put a chip on his shoulder. Then that happened, got a major chip on his shoulder and he was like, now I'm going to win. And like do it with, you know, some gumption. Do you run off of clean fuel or what do you have any chips on your shoulder?

Brendan

我想说,我总是有各种不满。我们一直有一种疯狂的紧迫感,公司非常紧张。如果人们没有执行,我们保持非常高的标准,我们行动非常快,所以我们总是想思考竞争对手在做什么,我们如何能更快,但我要提醒的是,我认为许多公司过于陷入关注竞争对手的陷阱,这变得模仿性。这真的是成为市场第二的好方法,对吧?我们对成为市场第二没有兴趣,对吧?我们有兴趣定义重要的品类,我认为这需要更多这种第一性原理思维,嗯,更多来自时间旅行者的动力,从 5 到 10 年后世界会是什么样子倒推,而不是关注竞争对手。嗯,所以也许这与你问的我的动机有关,但我认为我所有的联合创始人和我们的早期团队真的只是兴奋于我们如何建立一家传奇公司,解决未来十年人类如何融入经济的问题?

I would say I always have like varying chips on my shoulder. I like we always have had a maniacal sense of urgency where um the company is super intense. Like if um people aren't executing um we keep a really high bar and we move super quickly um and so we always want to be thoughtful about what are competitors doing and how can we move faster, but the thing I would caveat is that I think many companies fall too much in the into the trap of looking at what their competitors do, which becomes mimetic. Like it's a really good way to be second place in the market, right? And we have no interest in being second place in the market, right? Like we have an interest in defining important categories and I think that just requires more of this like first principles thinking, uh more motivation from being a time traveler of working backwards from what will the world look like in 5 to 10 years, less so focus on competitors. Um so maybe that's related to to your question of what motivates me, but I I think all of my co-founders and our early team are really just excited about how do we build a legendary company that solves this problem of how do humans fit into the economy over the next decade?

挑战与韧性 Challenges and Resilience

Host

你必须克服的最大挑战是什么?

What were the biggest challenges that you had to overcome?

Brendan

有很多。我的意思是,我认为很多没有真正创办过公司的人不知道那种被一次又一次打脸的感觉。你知道,我们有很多情况,早期失去了种子阶段的客户,因为最初大约一年半的业务都是种子阶段公司。所以,你知道,80% 的公司在 12 个月内倒闭。那是与这种持续紧张作斗争。嗯,有些变成了大公司,比如 Cognition。嗯,但那肯定总是有点痛苦。嗯,有过像安全事件这样的情况。有过我们失去早期团队成员的情况。就像有很多事情你只是被打了脸。我认为创始人会认识到混乱是常态。你只需要习惯咀嚼玻璃和凝视深渊,就像埃隆说的,因为最终是那 5 到 10 年旅程另一边的影響激励着你,让这一切都值得。

There's been so many. I mean I think that a lot of people who haven't actually founded companies don't know the feeling of just getting punched in the face over and over again. And you you know, we we've had so many cases where we've lost our you know, our seed stage customers early on cuz the the initial business for the first roughly year and a half was all seed stage companies. And so, you know, 80% of them go out of business within 12 months. And it was uh sort of fighting that constant tension. Uh some of them became large companies like Cognition. Um but that was definitely always a little bit painful. Um there have been uh cases like the security incident. There's been cases where uh we lose a a early member of the team. Like there's so many things where you just get punched in the face. And I think founders gain an appreciation that like chaos is the norm. And you just have to get used to uh chewing glass and staring into the abyss as Elon says because ultimately it's the impact on the other side of that 5 to 10 year journey that is motivating and makes it all worth it.

领导力与团队管理 Leadership and Team Management

Host

你如何为你的团队管理这一点?真的很有趣。就像我们,我认为我们,所以,我不久前在 Upfront Summit 上看到你,我在台上采访 Alfred Lin,他多次提到,甚至通过我们的采访,他们喜欢成为创始人和公司的减震器。那么,你喜欢成为员工的减震器吗?你如何帮助指导他们并让他们知道发生了什么?

How do you manage that for your team? It's really interesting. Like we I I think we So, I saw you at the Upfront Summit not too long ago and I was interviewing Alfred Lin on stage and something that he's talked about a couple times even through our interviews is they like to be shock absorbers for their founders and their companies. So, do you like to be a shock absorber for your employees? How do you help direct them and let them know what's going on?

Brendan

我同意这一点。我认为 Alfred 说的非常准确,当每个人都感觉站在世界之巅时,那正是你需要严厉打击所有出问题的事情的时候。而当事情不顺利,你知道,是困难时期之一时,那正是你想成为啦啦队长,告诉每个人他们做得有多好,并帮助团结公司的时候。我认为嗯那当然非常重要。

I agree with this. I think what Alfred was saying is super accurate that when everyone feels like they're on top of the world, that's when you need to be like cracking down on what are all the things going wrong. And when things are not going well and you know, it's one of the rough patches, that's when you want to be the cheerleader that's telling everyone how good a job they're doing and and helping to rally the company. I think that um that is certainly super important.

长期导向 Long-term orientation

Brendan

与此相关的另一点是推动长期导向。即使在我们高管团队,有人放了一张每周营收图表,我就说不行,我不想让公司盯着这个看。公司应该关注的是客户有多满意、专家有多满意、专家的利用率有多高——所有这些真正驱动创造世界价值的东西,以及它们如何复利。因为我认为,正如我学到并从那些早期在众多传奇公司工作过的人那里有幸获得的洞见,这种长期导向是成就非凡事业的人和公司最重要的特质之一。

The other thing that relates to this is just driving the long-term orientation. Even on our executive team, someone put up a weekly revenue graph, and I was like, no, I don't want the company looking at this. The company should be looking at how happy are our customers, how happy are our experts, how well utilized our experts are—all these different things that actually drive creating value in the world and how that compounds. Because I think that is one of—as I've learned and been fortunate to get insights from people who've been at a lot of the legendary companies early on—this long-term orientation is one of the most important characteristics of the people and the companies that accomplish extraordinary things.

Host

你敬仰的人有哪些?

Who are some people that you look up to?

Brendan

嗯,我确实认为从结果看埃隆是最伟大的企业家。我希望他别那么深入政治,但我觉得——

Well, I do think Elon is the greatest entrepreneur by results. I wish he didn't get into politics so much, but I think—

Host

幸运的是他已经脱身了。

Luckily he's out of it.

Brendan

幸运的是他已经脱身了。是的,我听人说他说过政府是唯一无法修复的机构。不管这告诉我们什么。但我钦佩他的一点——不仅是反复的成功记录,这很明显,而是那种使命导向,真正忍受如此多的痛苦,因为他在意建立这些塑造一个非常重要市场轨迹并定义该品类的公司。所以我认为他非常杰出。杰夫·贝索斯是我们谈到过的一个人,他几乎比任何人都更能体现这种长期思维,并且在亚马逊创立之初就非常有原则地确定哪些事情重要,以及这些如何映射到几十年后的产品,这比看起来要困难得多。我对此感到惊叹,他创造创业文化的能力。现在也有很多创始人显然非常出色。比如 Sam 和 Dario 都很了不起。我常常感到难过,他们俩总是会招来很多仇恨。对吧?总是有这个循环:你站在世界之巅,然后你成了美国的头号敌人,然后突然你又回到圈内。我对此感到难过。但这些绝对是最激励我的一些企业家。

Luckily he's out of it. Yeah, I heard from someone that he said the government is the one institution that wasn't fixable. So whatever that tells us. But I think one thing I admire about him—not only the repeated track record, which is obvious, but just the mission orientation of really enduring so much pain because he cares about building these companies that shape the trajectory of a really important market and define that category. So I think he's been exceptional. Jeff Bezos is one person we talked about where he really embodies this long-term thinking more so than almost anyone, and being so principled about what are the things that matter at the very beginning of Amazon and how those map on to the product decades later is a lot more challenging than it looks. And I've been amazed by that, the ability of him to create an entrepreneurial culture. There are also a lot of founders now that are clearly extraordinary. Like Sam and Dario are both spectacular. I often feel bad that both of them will always get so much hate. Right? It's like there's always this constant cycle of like you're on top of the world and then you're America's worst enemy and then all of a sudden you're back inside. I feel bad about that. But those are definitely some of the entrepreneurs that inspire me the most.

Host

我想谈谈杰夫·贝索斯。你写了一篇论文——哪个先发生?

I want to talk about Jeff Bezos. So you wrote a paper that—what came first?

Brendan

所以我写了——这不是发表的论文,但我在 Twitter 上开的玩笑是,我记得大一的时候上这门课,我必须写政府如何通过政策更好地对齐经济中的激励。我的研究论文的论点是,经济中最大的正外部性是创造就业。而政府征收税收的最大方式是对就业征税。对于没学过经济学的人,如果你对员工征税,这在功能上等同于对公司征税,对吧?这就像对创造就业的直接抑制。也许对于年收入十亿美元的人来说这有点道理——比如我不在乎,或者我挣不到那么多,但如果税单更大我也不太在乎。但我们对年收入低于 7 万美元的后半部分美国人这样做,就非常不理性。后半部分美国人只贡献了联邦政府税收的 3%。然而这对他们是一个巨大的负担,经济上以及所有相关的时间。所以论点是免除后半部分美国人的税收,因为这会激励更多就业创造,尤其是在低端。然后我看到我的英雄杰夫·贝索斯在 X 上发布了这个。所以我转推了他,然后他转推了我,这让我感到站在世界之巅。因为你知道,这感觉很奇怪,我从小到大听他的播客,如此钦佩他。终于至少得到了一些互动。我实际上从未见过他本人,所以没有。正在努力,是的。

So I wrote—it wasn't a published paper, but the joke I made on Twitter was that I remember I was in this class my freshman year of college and I had to write about how can the government pass policies that better align incentives in the economy. And the thesis of my research paper was that the largest positive externality in the economy is creating jobs. And the largest way that the government collects tax revenue is by taxing jobs. And for people that haven't studied economics, if you tax the employee, that is functionally the same as just taxing the company, right? It's like a direct disincentive to job creation. And maybe this sort of makes sense for people that are making a billion dollars a year—like I don't care, or I don't make that, but I don't care that much if I have a bigger tax bill. But the fact that we do this for the bottom half of Americans that are making under 70k a year is like so deeply irrational. The bottom half of Americans constitute 3% of federal government tax revenue. And yet it's a giant burden on them, economic as well as all of the time associated with doing this. And so the thesis was just eliminate taxes for the bottom half of Americans because that will incentivize more job creation, especially at the low end. And then I saw my hero, Jeff Bezos, posting about this on X. And so I retweeted him, and then he reposted me, which was—I was on top of the world about. And so because you know, it's like this weird feeling when I grew up listening to all of his podcasts and admiring him so much. And to finally at least get some engagement. I've never actually met him in person, so no, I haven't. Working on it, yeah.

Host

好吧,我们会让那件事成真的。

Well, we'll manifest that one.

Brendan

是的,是的,是的。我们得让它发生。

Yeah, yeah, yeah. We'll have to make it happen.

Host

这很令人兴奋。AI 需要的不仅仅是芯片。它需要电力、土地和基础设施。Zone 开发下一代数据中心园区,与 AI 公司、站点开发商和技术领袖合作,以更快、更大规模地将算力上线。Zone 正在构建 AI 前沿的基础。访问 zonefrontier.com 了解更多。那就是 zonefrontier.com 了解更多。Fundrise 旗下的 VCX,私人科技的公开股票代码,让各种规模的投资者都能投资风险资本。在 getvcx.com 查看投资组合。那就是 getvcx.com。Public 是一个多资产投资平台,你可以用 AI 智能体自动化投资。用简单的提示创建你自己的自定义智能体。智能体可以监控市场、执行交易策略,并管理你的现金,涵盖股票、债券、期权、加密货币等。想在市场收盘时买入并在第二天开盘时卖出?告诉你的智能体就行。在 public.com 自动化你的投资。创始人在 Deal 上更快交付。几分钟内为任何国家设置薪资,在任何地方雇佣任何人,快速处理签证,然后回去建设。访问 deel.com/sorcery。那就是 deel.com/s o c e r y。

That's exciting. AI needs more than chips. It needs power, land, and infrastructure. Zone develops next generation data center campuses, partnering with AI companies, site developers, and technology leaders to bring compute online faster and at scale. Zone is building the foundation of the AI frontier. Visit zonefrontier.com to learn more. That's zonefrontier.com to learn more. VCX by Fundrise, the public ticker for private tech, allowing investors of all sizes to invest in venture capital. View the portfolio at getvcx.com. That's getvcx.com. Public is a multi-asset investing platform where you can automate your investing with AI agents. Create your own custom agent with a simple prompt. Agents can monitor markets, execute trading strategies, and manage your cash across stocks, bonds, options, crypto, and more. Want to buy at market close and sell the next day it's open? Just tell your agent. Automate your investing at public.com. Founders ship faster on Deal. Set up payroll for any country in minutes, hire anyone anywhere, get visas handled fast, and get back to building. Visit deel.com/sorcery. That's deel.com/s o c e r y.

Host

那么,我们来谈谈今天的业务状况。你最近有一些大新闻。你收购了 Deep Tune。

So, let's talk about the state of the business today. So, you recently had some big news. You acquired Deep Tune.

Brendan

是的。

Yeah.

Host

你为什么这么做?

Why did you do that?

Brendan

所以这确实回到我们需要自动化经济中一切所需的强化学习环境的关键要素是什么。我提到的三个部分是:第一是应用,第二是世界,第三是任务。而我们最深的能力在于所有世界和所有任务,因为那是很多专家所做的,我们有智能体帮助增强并使其高效。我们最大的瓶颈是如何获得应用——Salesforce、Adobe 和 Google Workspace 的克隆——让所有专家填充相关内容,因为自然那不是我们开始做的。那不是我们比较优势所在。所以我们合作的一个客户主动建议,Deep Tune 是他们顶级的应用合作伙伴,他们应该直接给我们应用,让我们在构建这些环境时填充。所以当客户主动发起合作时有点讽刺。这是一个好信号,说明有东西。所以我们开始与他们合作。我更了解创始人 Tim。他非常出色。这显然是任何这类事情的一半,对吧?就像你想在未来十年与谁合作。这明显是 1+1=3 的那种情况,我们正在设立纽约办公室。

So, it really ties back to what are the critical ingredients of the RL environments we need to automate everything in the economy. Where the three parts I mentioned are first of all the apps, second of all the worlds, and third of all the tasks. And the thing that we have the deepest competency in is all of the worlds and all of the tasks because that's what a lot of the experts do and we have agents that help to augment and make that efficient. And our largest bottleneck was how do we get the apps—that clones of Salesforce and Adobe and Google Workspace—that we can have all of the experts populate with the relevant content because naturally that is not what we started doing. It wasn't as much of our comparative advantage. And so one of the customers that we work with actually suggested proactively that Deep Tune was their top apps partner and they should just give us apps that we could populate as we were building out these environments. And so it's sort of ironic when a customer initiates a partnership. It's a good signal that there's something there. And so we started working with them. I got to know Tim, the founder, better. And he was spectacular. And that's obviously half of any of these things, right? Just like who are the people that you want to work with for the next decade. And it was clearly just a 1 + 1 = 3 kind of thing where we're setting up our New York office.

收购策略 Acquisition Strategy

Host

这是你们的第一次收购吗?

Was this your first acquisition?

Brendan

这是我们的第二次收购,但规模比第一次大得多。我们还收购了一家叫 Seople 的公司,创始人是 Robbie 和 Cat,他们也非常出色。但 Tim 在 Deep Tune 上取得了巨大进展,而且刚从 A16Z 融了 A 轮,所以走得更远一些。

This was our second, but it was much larger than our first. We also acquired a company called Seople, founded by Robbie and Cat who are also spectacular. But Tim had made a giant amount of progress with Deep Tune and had just raised his Series A from A16Z so it was a little bit further along.

Host

你们未来有并购计划吗?或者你从战略上怎么考虑这件事?

Do you have an M&A plan going forward or how do you think about this strategically?

Brendan

我认为并购非常重要。这在一定程度上取决于公司的类型,但至少它是非常高杠杆的事情,对吧?你可能会在大规模上犯错,但也可能在大规模上产生巨大的积极影响。所以这绝对是我们会考虑的事情。我们正在探索几个我们认为可能有趣的类别。

I think M&A is really important. It depends a little bit on the type of company, but it's at least something that is super high leverage, right? Where you can make mistakes at large scale, but you could also achieve high positive impact at large scale. And so it's definitely something we think about. And we're exploring a few different categories that we think could be interesting.

挖角与激励 Poaching and Motivation

Host

你被挖角过吗?

Have you gotten poached?

Brendan

我被挖角过吗?

Have I gotten poached?

Host

是有人在试图收购你吗?我肯定你遇到过不少。

Are people trying to acquire you? I'm sure you've gotten a handful.

Brendan

确实有过这种情况。但这真的关系到我和我的联合创始人的动力。Adarsh 和我曾有机会拿着大量现金离开,再也不用工作,得到生活中想要的一切,但那不是我们的动力。我们宁愿每周工作 100 小时来解决世界上真正重要的问题。我们有幸在一个不可思议的市场中与一支出色的团队一起做到这一点。很难知道我们是否还会再有这样的机会。所以我认为这才是真正框定它的东西,以及我们如何思考我们的优先事项和想要实现的目标。

There have been cases of that. But it really ties to what motivates me and my co-founders. Adarsh and I have had opportunities where we could walk away with a lot of cash and never work again and get whatever we want in life, but that's not what motivates us. We'd rather work 100 hours a week to solve a really important problem in the world. And we have the privilege of doing that in an incredible market with a spectacular team. And it's hard to know if we'll have that privilege again. And so I think that's really the thing that frames it and how we think about our priorities and what we want to accomplish.

规模收购与Meta Scale Acquisition and Meta

Host

那会怎么运作?比如你们会被谁收购?就像 Scale 被收购那样?因为他们仍然有外部客户?

How would that work? Like who would you get acquired by? Like what happened with the scale acquisition? Because they still have outside customers?

Brendan

是的,Scale 的收购当然有点奇怪,因为并购的事情或并购限制,但 Alex 和核心团队过去了,然后很多其他团队成员留在了 Scale,Scale 将大部分业务转向只专注于 Meta。我认为我们永远不会做那样的事情。

Yeah, the scale acquisition is a little bit funky of course because of the M&A stuff or the M&A restrictions, but it's Alex and the core team went over and then a lot of the rest of the team stayed at scale and Scale pivoted most of their business to just focus on Meta. I don't think we would ever do something like that.

Host

但 Meta 还在用其他公司吗?

But does Meta still use other people?

Brendan

Meta 还在用其他供应商,或者我,是的,我不应该过多谈论他们做什么,但是的。

Meta still uses other vendor or I yeah, I shouldn't speak too much to what they do, but yeah.

Host

嗯,Alex 非常有才华。

Well, Alex is very talented.

Brendan

他非常有才华,我只见过 Alex 一次,但我有幸与很多和他共事过很长时间的人合作,我认为他是过去十年中最具一代人代表性的企业家之一,所以看到他在 Meta 取得的成就也很棒。

He's incredibly talented and I've only met Alex once, but I've been fortunate to work with a lot of the people that spent a lot of time with him and I think he was one of the most generational entrepreneurs of the last decade and so it's been great to see what he's been able to accomplish at Meta as well.

Scott Wu与独立性 Scott Wu and Independence

Host

有趣的是我们刚请了 Scott Wu 来。我们刚发布了这期关于保持独立的节目。

It's funny cuz we just had Scott Wu on. We just put out this episode about being independent.

Brendan

听 Scott 说话很有趣。

Scott is so fun to listen to.

Host

太好了。我是说,我们剪掉了很多,因为有很多笑声。

Great. I mean, we cut out a lot cuz there was a lot of laughing going on.

Brendan

是的,是的。

Yeah, yeah.

Host

但他很棒。他们的团队非常厉害。我觉得现在没人用这个词了,但 Cognition 确实很厉害。你知道,我们谈到了因为 Coding Agents 领域的大多数公司都被收购了或者在做类似的事情。所以我问他,你为什么要保持独立?而他想建立一家大公司。

But he's awesome. Their team is cracked. I don't think people use that term anymore, but cognition is cracked. And you know, we talked about because most of the companies in Coding Agents have gotten acquired or they're doing something like that. And so I asked him like why are you going to be independent? And he wants to build a massive company.

Brendan

是的。我认为这很合理,对吧?就像所有取得了非凡成就的创始人,动机从来不是经济上的。为了钱去做说不通。它总是一种情感上的牵引,去做世界上真正重要的事情。所以即使当我在考虑我们要收购哪些公司、收购哪些创始人、与哪些工程师和高管合作时,也是那些有情感牵引去做非凡事情的人,往往最有效率,能完成最令人印象深刻的事情。

Yeah. I think it makes sense, right? It's like for all of the founders that have accomplished something extraordinary, the motivation is never economic. It wouldn't make sense to do it for money. It's always an emotional pull to doing something really important in the world. And so even when I'm indexing on what are the companies we want to acquire, who are the founders we want to acquire, who are the engineers and executives that we want to work with, it's people that have that emotional pull of doing something extraordinary that tend to be the most effective and accomplish the most impressive things.

Harvey与法律基准 Harvey and Legal Benchmarks

Host

我们刚去的另一家公司是 Harvey。

Another company we were just out was Harvey.

Brendan

嗯。

Mhm.

Host

所以我们和

And so we talked with

Brendan

我也喜欢那些人。

I love those guys too.

Host

是的。他们办公室很漂亮。

Yeah. They're off their office is gorgeous.

Brendan

我知道。我前几天在那里,是的,非常漂亮。

I know. I was there the other day and yeah, it's very beautiful.

Host

我觉得我没见过比那更漂亮的办公室。

I don't think I've seen a prettier office than that.

Brendan

嗯,你很快得来看看我们的办公室,是的。我们的也不错。可能不是一个级别,但是的。

Well, you'll have to check out our office soon, yeah. Ours is nice. It might not be the same caliber, but yeah.

Host

真的非常不错。总之,我们请了 Winston、Gabe 和 Nico 来,我们谈到了 Lab 和他们最近发布的基准测试。你被提到是帮助完成那件事的合作伙伴之一。所以你能讲讲你是怎么帮助的,以及你在 Lab 中的角色吗?

is like really really nice. Anyway, so we had Winston on, we had Gabe on, and we had Nico on, and we talked about lab and the recent benchmarks that they put out. And you were mentioned as one of the partners that helped with that. So can you walk through how you helped and like what your I guess your role in lab was?

Brendan

嗯。这和我描述的过程类似,我们与他们合作的工作对应于从他们关心的所有细分领域的优秀专家开始,然后最终帮助策划这些专家构建的管理流程和具体数据形态,以转化为高质量的数据集。我们发布了很多公开的法律基准测试,比如 APEX 和 APEX agents。所以他们发现其中一些工作非常有趣,当然他们在看待这个问题时有独特的视角,因为他们了解律师事务所关心的所有分类法,以及如何覆盖实验室不擅长的分布。所以我对他们在后训练模型方面所做的工作感到非常兴奋。我认为这就像我对 AI 市场将如何发展的心智模型:前沿实验室是 AI 原生公司(如 Apple 公司)将做什么的领先指标,而后者又是企业的领先指标。而将要生存下来的 Apple 公司需要在其特定领域构建前沿模型,比如 Harvey 在法律领域构建模型,因为最终模型就是产品,你需要具备前沿能力才能保持竞争力并相对于实验室建立护城河。所以我们非常兴奋能与他们合作,很快会有更多进展。

Mhm. Well, so it's similar to the process I described where the work that we do with them corresponds to starting with great experts across all the granular domains that they care about and then ultimately helping to curate what are the management processes and specific data shapes that those experts build to translate to a really high quality data set. And we've put out a lot of public legal benchmarks such as APEX and APEX agents. And so they found some of that work really interesting and of course have a unique lens in looking at this problem because of all of understanding all the taxonomy of what law firms care about and how to cover the distribution that the labs aren't effective at. And so I'm really excited about the work that they're doing in post training models. I think that that is like my mental model for how the AI market is going to develop is that the frontier labs are leading indicator on what the AI native companies like Apple companies are going to do, which is a leading indicator on the enterprises. And the Apple companies that are going to survive need to be building the frontier model in their specific domain, like Harvey building the model in law because ultimately the model is the product and you need to be able to have frontier capabilities to stay competitive and build a moat relative to the labs. And so we're super excited to partner with them on that and there's a lot more coming soon.

AI应用层公司 AI App Layer Companies

Host

这真的很有趣。我想你还说过 AI 应用层公司将会消亡。

It's really interesting. I think you also said that AI app layer companies are going to die.

Brendan

不完全是。我说的是,除了数据公司之外,有三类主要公司具有有意义的营收规模。第一类是前沿实验室。第二类是推理提供商。所以第一类是 Anthropic、OpenAI,第二类是 Fireworks、Baseten,第三类是像 Harvey 这样的应用层公司。

Not exactly. Well, what I said is that there's sort of three main categories of companies that have meaningful revenue scale outside of data companies. The first one is frontier labs. The second one is the inference providers. So first one being, you know, Anthropic, OpenAI, second one being Fireworks, Baseten, and the third one is the app layer companies like Harvey.

AI栈中的护城河 Moats in the AI stack

Brendan

我的论点是,这些业务中每一层都有很大一部分会被商品化。如果所有人都在推理层买 GPU,那 GPU 就被商品化了。如果应用层公司在搭建软件层,那随着造软件越来越便宜,这些也会被商品化。但我看到的、公司能够建立且持久的护城河有两样:第一是品牌,Harvey 拥有世界级的品牌——在法律这个类别里肯定是顶尖的,也许在整个应用层创业公司里也是。第二是数据护城河,这一点我们可以切实帮上忙:他们如何理解这个飞轮——不是拿客户数据去训练,而是学习客户关心什么的分布,并以此启发他们要建哪些基准测试、需要整理哪些训练数据集,才能在这些方面做到世界最好。我认为这将是应用层公司要寻找的两大主要护城河,只有具备这些的公司才能建立起持久、甚至传奇级的公司。

And my thesis is that there are large portions of each of those businesses that get commoditized. If everyone's buying GPUs in the inference layer, then the GPUs got commoditized. If the app layer companies are building software layers, those will get commoditized as it becomes cheaper and cheaper to build software. But the two things where companies are able to build moats that I've seen that have been durable is, first of all, their brand, of which Harvey has a world-class brand — certainly top in the law category and maybe more generally across application layer startups. And then second of all, their data moats, which we can help with meaningfully: how do they understand this flywheel from not training on customer data, but learning the distribution of what their customers care about and using that to inspire what benchmarks they want to build, and what training data sets they need to curate to be best in the world at those. And I think those are really going to be the two primary moats to look for in the application layer companies that are able to build durable and potentially legendary companies.

Host

那最顶层那几层的护城河是什么?

What are the moats for the top top layers?

Brendan

嗯,对于前沿实验室来说,肯定有很强的规模经济,就是单纯在算力和数据上花更多钱——如果你有最大的分发和从模型上能产生的收入,你就能投入更多算力和更多强化学习环境,而实验室们正在指数级地加大这些投入。我觉得还有——所以我会说主要就是规模经济,然后还有一些飞轮效应,即一个给定模型帮助改进它后续的模型,也就是递归自我改进,这不会是垂直起飞,但会起作用,让事情随着时间稍微快一点。然后品牌护城河,我会说是基础模型的主要护城河。

Well, with the Frontier Labs, there's certainly a lot of economy of scale associated with just spending more on compute and more on data, where if you have the most distribution and revenue that you can generate from your models, you can invest more on having more compute and more RL environments, which the labs are scaling up their investments in exponentially. I think there's also — so I would say it's probably mainly economies of scale, and then there's some flywheels associated with a given model helping them to improve their subsequent model, i.e. recursive self-improvement, which won't be a vertical takeoff but it will contribute and make things slightly faster over time. And then brand moats, I would say, are the main ones for foundation models.

Brendan

对于推理公司来说,我认为它们最需要做对的一件事——Fireworks 在这方面执行得非常好——就是构建定制模型。这正好是 Harvey 所思考问题的另一面,因为转售 GPU 的转换成本是零。不管人们多想讲一个利润率很高的美好故事——当然,你买一堆 GPU,它们升值 50%,你在一段时间内会有不错的利润率——但真正的转换成本和持久利润来自你如何帮客户定制一个转换成本很高、无法迁移到其他基础设施的模型。所以这就是我如何看待前沿模型实验室正在形成的护城河,以及推理层正在形成的护城河,还有这些公司如何建立持久的业务。

For the inference companies, the primary one that I think they need to get right — and Fireworks is executing really well on this — is building custom models. It's just the other side of the coin of what Harvey is thinking about, because the switching costs of reselling GPUs are zero. As much as people try to tell this nice story of how there are great margins — of course when you buy a bunch of GPUs and those appreciate like 50%, you'll have good margins for a period of time — the real switching costs and durable margins come from how you help a customer customize a model that has high switching costs and can't be ported over to other infrastructure. And so that's how I think about the moats that are developing at the frontier model labs, and then also the moats that are developing in the inference layer, and how those companies will build durable businesses.

Host

哪一层最脆弱?

Which is the most vulnerable?

Brendan

我觉得 API 的转换成本太低了,所以我认为投资者高估了一些 API 业务的收入质量,因为如果你把它简化来看,这些很大程度上就归结为转换成本。我们自己就按天在 API 之间切换,取决于哪个模型在我们评估业务运营方式的评测上表现最好。我认为每家公司成熟、推理账单变大之后都会这么做。所以其中一些公司可能规模很大,但我认为收入质量有时被高估了。

I think that it feels like the switching costs for APIs are so low that I think investors overrate some of the revenue quality of API businesses, because so much of these just boil down to switching costs if you make it that simple. And we switch between APIs on a daily cadence based on whatever model's performing the best on our evals for how we run the business. I think every company will do that as they mature and have scaled inference bills. And so some of these companies may be really big, but I think the revenue quality is sometimes overrated.

数据为何定价错误 Why data is mispriced

Host

好。既然我们聊到宏观层面,你还发过另一条推文。顺便说一句,那是条推文,我们会把那条推文标记出来。但你发过另一条推文,说 AI 进步需要三样东西:算力、算法和数据。你为什么认为在这三者之中,数据的定价是错的?

Okay. Well, since we're talking macro, there was another tweet you put out. By the way, that was a tweet. So we'll tag that tweet. But there was another tweet that you put out that AI progress requires three things: compute, algorithms, and data. Why do you think data is mispriced between that stack?

Brendan

我认为数据仍然是逆共识的。这是最突出的一点:出于某种原因,投资者一直有一种看法,认为我们三年内就会实现超级智能,AI 会在所有事情上都比人类强,所以我们就不需要人类来创建评测和训练模型了。而投资者过去三年一直在这么说。我记得甚至在我们的种子轮,硅谷大多数顶级机构——我们一直很幸运能和喜欢的投资人合作,但硅谷很多顶级机构就是认为这个市场会在几年内消失。我认为大家一直忽略的是杰文斯悖论,或者实际上就是劳动总量谬误:当我们让专家更高效,当专家能把模型改进三倍、五倍、十倍,这反而大幅增加了对改进模型的人力的需求。归根结底,只要我们没有超级智能,只要经济中还有人类能做而 AI 做不到的事,我们就需要人类来衡量那个前沿。所以我认为这就是最大的脱节。说实话,我在很多方面对此心怀感激,因为我们一直有充足的资金渠道,但逆共识减少了市场上的竞争。所以我很高兴我们的市场远没有基础模型业务那么竞争激烈。但我认为它正处在成为共识的边缘,尤其是当人们看到所有那些开始评估自己部署的智能体和后训练模型的企业时。

I think data is still contrarian. It's the top thing where for some reason there has been this perception by investors that we will get superintelligence in three years and AI will be better than humans at absolutely everything, and so we won't need humans to then create evals and train models. And the investors have been saying this for the last three years. I remember even in our seed round — most of the top firms in Silicon Valley, we've always been very fortunate to work with investors we love, but a lot of the top firms in Silicon Valley just thought that the market would dissipate within a couple of years. And I think the thing that everyone has always missed is the case of Jevons Paradox, or functionally the lump of labor fallacy, which is just like as we make the experts more efficient, as the experts are able to improve models three, five, 10 times more, that has dramatically increased the amount of demand for people to improve the models. And ultimately, as long as we don't have superintelligence, as long as there are things that humans can do that AI can't do in the economy, we need humans to measure that frontier. And so I think that has been the largest disconnect. And I honestly feel sort of grateful for it in a lot of ways because we've always had lots of access to capital, but it being contrarian reduces the amount of competition in the market. So I'm glad our market is not nearly as competitive as the foundation model business. But I think that it is on the verge of becoming consensus, especially as people see all of the enterprises that are starting to evaluate agents that they're deploying and post-rate models.

未来十年 The next 10 years

Host

那未来 10 年会怎么发展?

So how does this play out over the next 10 years?

Brendan

我认为关键的发展是,每一家《财富》500 强公司都会变得像一家新实验室:它们有自己的核心数据集,有自己关心的、用来服务客户的核心分发。它们需要调动全职员工和承包商的某种组合,去整理真正高质量的任务,用于评估和训练它们正在构建的模型,因为归根结底,目前由人完成的许多流程都将变成智能体式的。而这是把整个公司的人类知识蒸馏进模型的最高效、最可扩展的方式。所以这将是一场巨大的转型,而我们才刚刚打完第一局。

I think that the key development is that every Fortune 500 is going to look like a neolab, where they have their core data sets, their core distribution that they care about to serve their customers. And they need to mobilize some combination of their full-time employees and contractors to curate really high quality tasks for evaluation and training of the models that they're building, because ultimately so many of these processes that are currently human are going to become agentic. And this is the most efficient and scalable way to distill all of that human knowledge throughout their company into the models. And so that will be a giant transformation that we're just starting the first inning of.

AI数据主权 AI data sovereignty

Host

天哪,我正想问你这个。其实我提前问了 Winston 一些问题,他想让我问问 AI 数据主权。

Damn, I was going to ask you that. So I actually asked Winston for some questions ahead of time and he wanted me to ask about AI data sovereignty.

Brendan

这会非常关键。我是说,Jensen、Alex Karp 和 Satya 都在谈这个,是有原因的。我认为 Anthropic 和 OpenAI 会拥有非凡的业务。我认为它们两家都很可能成为市值 10 万亿美元以上的公司。但归根结底,我们不希望它们吞并全球经济,成为你获取前沿智能的唯一去处。

It's going to be so critical. I mean, there's a reason that Jensen and Alex Karp and Satya are all talking about this. I think that Anthropic and OpenAI will have extraordinary businesses. I think both of them could very well be 10 trillion plus dollar companies. But ultimately, we don't want them to subsume the global economy and be the only place that you can go to access frontier intelligence.

企业AI采用与推理支出 Enterprise AI Adoption and Inference Spending

Brendan

而且,为了让价值惠及财富 500 强中的其他所有公司以及更广泛的美国人,我们需要确保所有这些公司都能训练自己的模型,并构建自己的智能,从而将他们的竞争优势编码化。我认为,某些实验室在融资时会有这样一种叙事:突然之间,他们就可以开启对财富 500 强所有企业数据的使用数据飞轮进行训练,而这些企业会对此无所谓,这就是他们收集整个经济中所有环境的方式。但显然事实并非如此,对吧?显然,像高盛这样的公司之所以存在,是因为他们拥有一系列与业务中的人员、流程、文档等相关的竞争优势。而允许前沿实验室基于这些数据进行训练,就相当于与竞争对手分享所有这些数据。所以他们不会这么做。我相信,相反,他们会构建自己的模型,部署自己的智能体,将所有这些知识编码化。而真正要关注的启发式方法是,企业内这一转型的催化剂是什么,就是他们在推理上花费了多少。因为我们基本上发现,当一家公司在推理上的花费低于 1000 万美元时,他们通常不太关心评估或如何优化之类的。他们还在摸索。一旦他们的推理花费达到 1000 万到 1 亿美元的范围,他们通常就会想要花费……

And for value to accrue to every other company in the Fortune 500 and Americans more generally, we need to make sure that all of these companies are able to train their own models and build their own intelligence that codifies their competitive advantage. Like I think there's sometimes this narrative from certain labs when they're fundraising that all of a sudden they're just going to turn on training on the usage data flywheel of all of the enterprise data at the Fortune 500s are going to be fine with this and this is the way that they're going to like collect all the environments throughout the economy. When like clearly that's not the case, right? Clearly the reason that a company like Goldman Sachs exists is that they have a set of competitive advantages associated with the people, the processes, the documents, etc. in their business. And turning on allowing a frontier lab to train on that data would be analogous to just sharing all of it with a competitor. And so they're not going to do that. And I believe that instead they're going to be building their own models, deploying their own agents that codify all of this knowledge. And really the heuristic to look out for what is the catalyst of this transition and happening within the enterprise is how much they're spending on inference. Cuz what we basically find is that when a company is spending under $10 million on inference, they don't generally care that much about evals or like how to optimize it or whatever. They're still figuring things out. Once they go from the 10 to $100 million on inference range, then they generally want to spend...

Host

这么大的差距。

Such a wide gap.

Brendan

是的,差距很大,

Yeah, it's a wide gap,

Host

就像,我的意思是,我们每天在推理上花费近一百万美元。所以,有些公司扩张得很快。

Like, I mean, we're spending almost like a million dollars a day on inference. So, like, some companies scale fast.

Brendan

而且,一旦他们的推理花费达到 1000 万到 1 亿美元,那么大约 10% 的推理预算应该用于评估,即如何构建数据集,告诉他们应该将请求路由到哪些模型。一旦他们的推理花费超过 1 亿美元,比如 Harvey 或 Casetext 或所有这些公司,那么他们就需要真正认真地思考如何后训练模型,因为经济账开始算得过来了。如果你每年可以花费数千万美元,有时,你知道,像 Casetext 那样花费数亿或数十亿美元来训练自己的模型,因为你可以将固定成本分摊到所有的推理上。所以我预计企业的推理支出将大幅增加,但他们也会经历这些采用曲线:在 1000 万到 1 亿美元范围内推出评估,然后在 1 亿美元以上开始后训练,因为他们要在整个业务中部署前沿智能。

And um once they're in the 10 to 100, then it's roughly 10% of their inference budget should be going to eval of how do they build out the data sets that tell them like what are all the models they should route this to. Once they're past $100 million in inference, like a Harvey or a Casetext or all of these companies, then they need to be really seriously thinking about how do they post train models because the economics start to check out. If you can spend tens of millions of dollars a year, sometimes, you know, hundreds or billions in the case of Casetext, in training your own models because you can amortize that fixed cost across all of the inference that you have. And so I expect the inference spend of enterprises will ramp materially, but they'll also go through these curves of adoption where they roll out evals in that 10 to 100 range and then start post training in the 100 plus range as they deploy Frontier Intelligence throughout their businesses.

企业构建主权AI建议 Advice for Enterprises Building Sovereign AI

Host

因为你拥有如此独特且有利的视角,我的意思是,你 literally 看到每家公司,你看到他们的数据如何被使用等等,考虑到你刚才提到的关于推理的一切,你对这些企业在继续构建其主权 AI 栈时有什么建议?

Because you have such a unique and advantageous perspective, I mean, you see literally every company and you see how their data's being used and all this kind of stuff, what advice do you have taking into account everything that you just mentioned on inference, what advice do you have to these enterprises as they continue to build up their sovereign AI stack?

Brendan

我想说,企业最大的误解是前沿模型实际上是如何学习的,现在训练中绝大部分算力不再只是读取公司所有文档或从输入输出和 SFT 中学习。现在大部分算力都用于强化学习,模型会多次尝试问题,然后根据这些评分标准获得分数。而在绝大多数情况下,强化学习需要主题专家来创建这些评分标准和任务,以评估和训练模型。所以对于任何训练自己模型的公司,他们需要构建能产生清晰评估分数的任务。然后一旦他们对评估有信心,就应该扩大他们关心的分布,以爬坡那个评估。所以也许这两条建议是:首先,构建任务,而不是仅仅假装模型可以读取一切并学习。其次,他们应该从评估开始,因为这是每个 AI 实验室和成功应用层理解他们做得如何以及拥有改进他们正在训练的模型的基线的基础。

I would say that the largest thing that enterprises misinterpret is how Frontier models actually learn, where the vast majority of where the compute gets spent on in training is no longer just like read all the documents in the company or learn from inputs and outputs and SFT. Like most of the compute is now going towards RL where the model will attempt the problem many times and then get a score on of these rubrics. And RL in the vast majority of cases requires subject matter experts that are able to create these rubrics and tasks to evaluate and train models. And so for any company that is training their own model, they need to build tasks that produce clear eval scores. And then once they feel confident in the eval, then they should scale up the distribution of what they care about to hill climb that eval. So maybe those are the two pieces of advice is first of all build tasks rather than just like pretending the model can read everything and learn. And second of all that they should start with the eval because that is the foundation for how every AI lab and successful application layer understands how well they're doing and has a baseline in improving their models that they're training.

未来计划与日常 Future Plans and Daily Routine

Host

很好,不错。

Good and nice.

Brendan

谢谢。是的。

Thank you. Yeah.

Host

好的,那么还有两个问题。未来 12 个月你最期待什么?我们可以从 Merquar 期待什么?你们会上市吗?

Okay, so two more questions. What are you most looking forward to in the next 12 months? What can we expect from Merquar? Are you going to go public?

Brendan

我不认为我们会在未来 12 个月内上市,但我们是。我们,这很有趣。我想在 2024 年,我发推说年底前 IPO,我们,我以为这明显是个玩笑,因为我们大约 12 个月前才从大学辍学。人们却说,哇,不可能吧。所以我们不会在未来 12 个月内上市,但这肯定在我们的雷达上。我最期待的是企业方面的东西,我认为随着推理账单的成熟,它开始如此迅速地发生,真正经典的公司,财富 50 强中的非科技公司开始理解这有多重要。其中一些公司正面临来自前沿模型的存在性威胁,因为这些公司正在侵入他们的领域。这将成为转变智能在经济中使用和部署方式的巨大催化剂。所以我认为这将令人难以置信。

I don't think we'll go public in the next 12 months, but we are. We It's funny. I like think in 2024 I tweeted IPO by end of year and And we were I thought it was like clearly a joke cuz we dropped out of college like 12 months prior. And people were like, whoa, like no way. And so we're we're not going to go public in the next 12 months, but it's certainly on our on our radar. The thing I'm most looking forward to is really the stuff on enterprise where I think that it's starting to happen so quickly as the as the inference bills mature that the really classic companies, the Fortune 50 non-technology companies are starting to understand how important this is. Some of those companies are facing existential threat from the frontier models as those companies creep into uh their sectors. And that's just going to be a huge catalyst for transitioning the way that intelligence gets used and deployed in the economy. And so I I think it'll be incredible to see.

Host

酷。嗯,我知道这是一个非常奇怪的结束问题,但你在我们开始录制前问过我这个问题,你说,你如何度过你的一天?所以我想问你,布兰登,你如何度过你的一天?你还在翻甜甜圈吗?

Cool. Well, I know this is a very strange closing question, but you asked me this before we started recording and you were like, how do you spend your days? So I want to ask you, Brandon, how do you spend your days? Are you flipping donuts?

Brendan

不再翻甜甜圈了。呃,我想想。我通常会在 7:45 左右醒来,然后如果我,我大多数日子会在 7:50 到 8:30 之间醒来,然后 9:00 到办公室,然后我通常每天有 4 小时的会议。有时接近 8 小时,我会在其他地方处理大量消息、邮件和文档。而且我的时间大概三分之一用于招聘,三分之一用于客户,三分之一用于研究和产品。呃,然后我在办公室待到晚上 10 点,每周六天重复。然后周日我在家工作。

No longer flipping donuts. Uh I let's see. So I generally will wake up around 7:45 and then if I'm I wake up most days around 7:50 until 8:30 and then I'm in the office by 9:00 and then I would usually have 4 hours of meetings a day. Sometimes it's closer to 8:00 to and I would knock out lots of messages, emails, and docs elsewhere. And that And my time is probably roughly a third on hiring, a third time with customers, and a third time on research and product. Um and then I'm in the office until 10:00 p.m. and I rinse and repeat 6 days a week. And then Sundays I work from home.

Host

不错。当我和温斯顿交谈时很有趣,很多人听了那期节目后联系我,因为他们说,天哪,听到有人经历那种超高速增长,以及他们如何管理自己的一天和花时间在什么上,真的很棒。有一部分他谈到修复人们的日历和删除会议。你是否试图不断消除干扰?你如何管理专注力?

Nice. It was interesting when I talked to Winston, I got so so many people reached out to me listening to that episode because they're like, oh my god, like it's really great to hear someone who's going through that hyper growth of how they're managing their day and what they spend their time on. And there was one part where he was talking about like fixing people's calendars and like deleting meetings. Are you trying to like constantly try to just like remove distractions? How do you manage focus?

Brendan

完全正确。

Totally.

领导力扩展与授权 Scaling leadership and delegation

Brendan

这很有挑战性,因为我觉得大家很容易就想把事都上报给我,而组织要扩张,唯一的办法不只是靠我,还要靠一线的人对组织里的领导者和高管有高度信任。所以确保我们有一支非常强的领导团队,大概是最重要的事。我感觉我可以花一整天甚至更多时间,光回复消息和邮件。所以到了某个节点,你也必须不再什么都回。我每天还是会做到收件箱清零,因为我想确保每件事都过一遍,但要非常审慎地考虑这一点,以及在哪里借助智能体来提升效率。

It's challenging because I think it's easy for everyone to want to escalate things to me, and the only way the org scales is not only me, but also the people on the ground having high trust in the leaders and the executives in the organization. So making sure that we have a really strong leadership team is probably the most important thing for that. I feel like I could spend all day and more just responding to my messages and emails. And so at some point you also have to just stop responding to everything. I still inbox zero every day because I want to make sure I get through everything, but being very thoughtful about that and where to leverage agents to make things more efficient.

Host

不错。你们在招人吗?

Cool. Are you guys hiring?

Brendan

我们在大量招人。这挺讽刺的,因为大家都说,哦,我们要取代所有工作,但所有 AI 公司,包括我们,都在尽可能快地招人。而且我会——是的,我们很乐意和你合作。去看看 mercor.com。我们不仅在招专家,还在招很多全职员工加入我们的工程团队、研究团队,以及运营团队。

We are hiring a ton. And so it's ironic because everyone says, oh, like we're going to displace all the jobs, but all of the AI companies, us included, are hiring as quickly as possible. And I would—yeah, we'd love to work with you. Check out mercor.com. And not only are we hiring for the experts, but lots of full-time employees to join our engineering team, our research team, as well as our operations team.

Host

太棒了。这是个很好的收尾。那么,非常感谢你,Brendan。

Amazing. Great place to end it. Well, thank you so much, Brendan.

Brendan

谢谢邀请我。

Thanks for having me.

互动版:逐字朗读 + 针对本期提问 →