20VC：Mercor CEO 谈为何应用层公司缺乏护城河、模型即产品 | 五年内代币支出将超过人力支出 | 与 Brendan Foody 探讨当前硅谷招聘 AI 研究人员的真实成本

你做的那些事，说实话，让我有点害怕。

you did, which slightly terrified me, to be quite honest.

你说大概有百分之三十到四十是这样的？

So you said like 30 to 40% is kind of how

我们通常这么看待来自这方面的收入吗？

we think about like our revenues from that?

一般来说，没错。

Generally, yeah.

明白了。

Okay.

如果我们看看我们融过的几轮资金，哪一轮你觉得最离谱？

So if we take the the rounds that we've raised, which round felt most uncomfortably high?

好问题。

Good question.

嗯，所以

Well, so

我来逐个聊聊每一轮的估值和收入情况。

I'll talk through the valuations of each, of the revenue of each.

我们的种子轮，Phantom 没有用直升机接你吗？

So our seed round Did Phantom not fly you in the chopper?

那真是够吓人的。

That was serious ice.

我们的种子轮是在2023年9月，当时年化收入大约一百万美元，差一点就到这个数了。我本来不想融资，想靠自己把公司做起来，但阿达什安·苏里亚坚持说，如果我要退学，就必须融资，所以我们见了General Catalyst的八号。

So our seed round was in September 2023, we were at called a million in revenue run rate, or just shy of that, and I initially didn't wanna raise because I wanted to bootstrap the company, but Adarshan Suriya's condition on dropping out was that we needed to raise money, and so we met General Catalyst eight a.

嗯。

M.

那个星期天早上。

On a Sunday morning.

他们三十六小时内就给了我们一份条款书，230万美元，投后估值2300万美元。

They gave us a term sheet within thirty six hours for $2,300,000 at a $23,000,000 post money valuation.

挺高的。

Pretty

毛收入。

gross.

所以那个价格还挺合理的，当时是马克斯和尼科。

So that was pretty reasonable insofar as price And this was was Max

还有他吗？

and him on?

当时是马克斯和尼科。

This was Max and Nico.

到了A轮融资时，业务从种子轮到A轮增长不多，但我们发现市场才是关键差异点。

And then at our Series A, the business didn't grow that much from the seed to the Series A, but we found the market was a key differentiation.

2024年5月，当我们营收达到一百五十万美元左右时，见到了维克托。

And we met Victor when we were at, call it, 1,500,000 in revenue run rate in May 2024.

维克托特别兴奋。

And Victor got super excited.

我本来拒绝再见他第二次，但他问：‘你坐过直升机吗？’

Initially, I refused to take a second meeting, but then he said, Oh, have you ever been in a helicopter?

于是彼得带我们坐了直升机，Benchmark特别想和我们合作。

And so Peter took us on the helicopter flight and Benchmark really wanted to work with us.

等到他们给我们投资条款清单时，我们的年收入已经到了大约250万美元，他们给我们开出了2.5亿美元的投后估值。

And so by the time that they gave us a term sheet, were at, call it, 2,500,000 in revenue and they gave us $250,000,000 post money valuation.

这让你觉得不舒服吗？

Did that feel uncomfortable?

因为从2.23亿跳到2.5亿，这跨度可不小。

Because that's a big jump, $2.23 post to $2.50.

你要明白，当时我们年收入才250万，听起来很疯狂，但我预计年底能到5000万，明年年底能冲到5亿。

So keep in mind, the time, this sounds crazy because we were at 2,500,000 in revenue, but I was projecting 50,000,000 in revenue run rate by the end of the year and 500,000,000 by the end of next year.

所以那时候感觉……

And so it felt

简直太划算了。

like a bargain.

老兄，你知道所有创始人都会这么想吧，

Dude, you do know all founders project that day,

我们从没打算做这个项目。

we in the never the project.

当然。

Sure.

这种事可不常见。

It just doesn't happen often.

对，没错，确实如此。

Yeah, yeah, yeah.

四个月后，我们认识了Felicis，压根没做任何演示文稿，也没去见投资人。

And then four months later, we met Felicis, and we never would make a slide deck or take investor meetings.

于是Lisa给我们发了封邮件，说：‘我们知道你们的联合创始人Surya特别喜欢法拉利。’

And so Lisa sent us an email saying, Hey, we know your co founder Surya really likes Ferrari.

所以，要不要跟我们一起去开法拉利赛车？

So do you want to go racing Ferrari's with us?

我回了她一句：‘你们这话可真抓住我了。’

Then I replied and I said, You caught my eye.

再给我讲讲。

Tell me more.

他们说会在海沃德机场碰头，然后坐艾登的私人飞机去拉斯维加斯，在F1赛道上开法拉利比赛。

And they said, We'll meet at the airport in Hayward and go on Aiden's private jet to Las Vegas to race Ferraris around F1 track.

我就回他们说，我们三周后的周日有空。

And so I was like, We're available in three weeks on a Sunday.

于是我们就这么干了。

And so we do this.

我们开法拉利比赛。

We race Ferraris.

我们现在的年收入已经两千万美元了。

We're at $20,000,000 in revenue.

他们问我们觉得什么样的估值比较合理。

They ask us what valuation do we think makes most sense?

我说十亿到二十亿美元之间，结果他们直接给我们开出了二十亿美元的估值条款。

And I say $1 to $2,000,000,000 So they give us a term sheet at a $2,000,000,000 valuation.

那时候，估值是收入的100倍。

And at the time, that's 100 times revenue.

大家都觉得这个估值太高了。

And everyone thinks that that's a high valuation.

但实际上，这是一笔超棒的投资。

Meanwhile, was an incredible investment.

说实话，那时候我正在面试阿达什。

So I'm gonna be honest, this is when I interviewed Adash at that time.

面试结束时，我就跟他说：兄弟，我真想投你。

And at the end, I was like, dude, I would love to invest.

求你了，让我投一笔吧。

Please let me invest.

你很慷慨，允许我投了一小笔。

And you very kindly let me put a small check-in.

之后我跟世界上好几位顶级投资人聊了这事。

And I then spoke to several of the biggest investors in the world.

当然不是说他们瞧不起我，但他们一听就笑了，说：哥们，那价格也太高了吧。

And no offense, they chuckled at me like, Dude, that was such a high price.

真的高得离谱。

It's such a high price.

但事实是这样。

Well, so here's

关键是，之前六个月我们每个月都增长50%，而他们根本没意识到，这种势头还会再持续一整年多。

the thing is, we'd been growing 50% month over month for the prior six months, and I think what none of them really realized was that it would continue for the subsequent twelve plus months.

到了2025年9月，或者说是10月，我们的年化收入已经冲到了四亿美元左右。

And so then that compounded more and more by September 2025, or say October, were at call it 400,000,000 in revenue run rate.

这时候Felicis就来找我们了，说想追加投资。

And then Felicis was like, We want to invest more.

于是他们开出了100亿美元的估值条款。

And so they gave us a term sheet at a $10,000,000,000 valuation.

我们其实没打算花太多时间在融资上，因为业务每个月都在涨50%，我们忙得根本顾不上这些事。

We didn't really want to spend much time on a financing because the business was growing 50% month over month, and so we were very preoccupied.

那时候估值大概是25倍，之后业务规模又涨了近四倍。

And so that was about 25 times, and the business has almost 4x since then.

所以回顾一下，哪一轮让你感觉最不舒服？

So review, which one felt most uncomfortable?

如果非得选一个的话，我觉得B轮定价最超前，比A轮更脱离了实际增长。

If I had to choose any, the Series B priced in the most, the furthest ahead of our growth, or that or the Series A.

应该是B轮吧，虽然A轮和B轮都是估值100倍收入，但250万美元收入时做到100倍，和2000万美元收入时做到100倍，完全是两回事。

Think it was probably the series B, because both were a 100 times revenue, but it's very different to be a 100 times revenue when you're at 2,500,000 in revenue versus 20,000,000 in revenue.

所以B轮可能是最夸张的一轮，不过回头来看，这两轮投资都挺成功的。

So that was probably the largest one, but obviously both were great investments in hindsight.

那下一轮会是什么情况？

What is the next round then?

到时候再说吧。

We'll see.

估计估值会高得多。

Probably a much higher valuation.

我们现在收到了不少报价，估值都比之前高出一大截。不过公司已经相当盈利了，所以我们打算慢慢看，挑个最合适的投资人。

We're getting a lot of offers at meaningfully higher valuations, but the company is fairly profitable, and so we're taking our time to see who the right is.

我们这也算是换着花样享受不同的出行方式吧？

We're also just going through modes of transport, aren't we?

我们坐过直升机，也开过法拉利，现在又是……

We had the Chopper, we had the Ferraris, we've had That's

这个观察挺到位的。

a good observation.

应该能拿到

Should get

一艘战舰，也就是D轮融资。

a warship now to get the Series D.

对。

Yeah.

我完全同意。

I totally agree.

我从来没上过军舰，但听起来特别有意思。

I've never been on a warship before, but that sounds a lot of fun.

这不就对了。

There you go.

红杉资本该安排一艘军舰了。

Sequoia line up the warship.

未来十二个月，位于Anthropic和OpenAI上游的基础设施公司，会比下游的应用层公司表现好得多。

The next twelve months will be dramatically better for infrastructure companies upstream of Anthropic and OpenAI than for application layer companies downstream of them.

这条推文是你发的。

This was your tweet.

你为什么这么认为？

Why do you believe that?

我觉得是因为应用层公司的业务和基础模型公司的业务其实很接近。

The reason I believe that is that the application layer companies businesses are not far removed from the foundation model companies businesses.

比如Claude公司，要转型并不难。

Like it is not a far leap for Claude Co.

努力拓展在医疗和法律领域的功能。

Work to add capabilities across medical and legal.

他们已经在软件工程上做到了，自然也能在金融领域复制同样的模式。

Obviously, they did it with software engineering and can do that across finance.

我觉得在模型之上构建软件层的护城河会极其困难。

And so I feel like building defensibility in the software layer on top of the models is going to be incredibly difficult.

但反过来，在基础设施这一端，明显正在形成真正有厚度的壁垒。

Whereas on the other side of things, in the infrastructure side, it feels like there are meaningful moats that are getting built.

我们在业务中不断积累巨大的网络效应，同时随着为客户扩充资源库，数据壁垒也在稳步建立。

Like we're compounding enormous network effects in the business and a pretty significant data moat as we build out the inventory for our customers.

计算公司当然能通过漫长的研发周期建立起自己的护城河。

Compute companies obviously are able to build moats through these very long R and D cycles.

所以我认为，基础设施层有望实现高利润率，打造出可持续盈利的业务，而这一点在应用层还远没那么明确。

And so I think that there are going to be high margins that get achieved at the infrastructure layer and sort of sustainable, profitable businesses in a way that it's less immediately clear at the application layer.

你看Nebius，我不知道你有没有留意，他们把支持服务的价格上调了30%，但需求完全没受影响。

I mean, you saw Nebius, I don't know if you saw us, but they increased their pricing by 30% across support, which will have absolutely no impact on demand.

这难道不简直疯狂吗？

Isn't that absolutely nuts?

你涨了30%的价，需求却一点没受影响。

So you increased price by 30%, zero impact on demand.

说实话，我们的情况大概也差不多。

It's probably the same for us honestly.

我们的需求一夜之间就能翻倍，只是没那个能力接得住。

We have the demand to double overnight, we just don't have the capacity.

所以关键在于，我们能多快地调动人手，把这些环境快速搭建起来。

And so it's mainly a question of how effectively can we scale to mobilize people to build out these environments much more quickly.

你们会做价格弹性测试吗？

Do you do pricing elasticity tests?

因为如果你能把价格翻倍，业务也能翻倍的话。

Because if you can double price and double the business.

我们可能没法把价格翻倍。

We maybe can't double prices.

我们可以把产能翻倍。

We could double capacity.

我们大概能把价格上调30%，对需求影响不大，但你还要考虑到，定价不只是为了未来六个月的利润最大化。

We could probably increase prices by 30% without much of an impact, but the other thing you need to consider is that pricing is not merely a question of optimizing for the next six months.

而是为了构建一个在未来十年里能赢得市场的体系。

It's optimizing for a structure that wins the market over the next decade.

正因如此，我们特别关注：怎样才能真正为客户提供最好的服务？

For that reason, we're very focused on how do we do what's best for customers?

怎样才能为专家们创造最好的条件？

How do we do what's best for experts?

同时，我们还要打造一个可持续的商业模式，但也不能把市场留太多利润空间，因为高利润会引来竞争。

And how do we build a sustainable business while we're doing it, but make sure that we're not leaving oxygen in the market because high margins invite competition.

好的，我投资了几家下游的应用层公司，比如你刚才提到的Legora。

Okay, I am an investor in several application layer companies downstream, like a Legora, which you mentioned there.

我们一直在观察Legora和Harvey之间的竞争。

We see the Legora versus Harvey battle.

我觉得大家其实都慢慢意识到，他们不该互相内耗。

I think everyone actually is coming around to the fact that they shouldn't be fighting each other.

正如你所说，他们更该提防Anthropic。

They should be wary of Anthropic to your point.

没错。

Totally.

但我一看就觉得，他们的护城河实在太深了。

But then I look at it and go, there is incredible defensibility.

这个产品非常深入，专门针对律师的工作流程设计。

It's a very deep product, specifically suited to the workflows of lawyers.

Anthropic要想追上来，得专门组建全新的产品团队和部门。

Anthropic would have to build out whole separate product teams, divisions to come after them.

还得搭起市场推广、客户成功和用户推广的全套团队。

They'd have to build out GTM teams, customer success teams, adoption teams.

这根本就是两家完全不同的公司。

It's a different freaking company.

这种护城河是实实在在存在的。

The defensibility is there.

你倒是反驳看看。

Argue back.

我可能想说两点。

Maybe I would say two things.

首先，我觉得过去两年里，大家都越来越明白，模型本身就是产品。

First is that I think over the last two years, everyone has increasingly realized that the model is the product.

以前大家总想着用各种方式把一堆 API 调用拼起来，搞一堆零零碎碎的逻辑，那时候还有人热衷于拖拽式智能体构建工具。

That we can build so many of these different abstractions of trying to stitch together API calls and having all this patchwork logic where people used to have all these drag and drop agent builders.

但后来大家发现，只要给模型一个明确的目标，再训练它去实现这个目标，它在几乎所有场景下都比其他方案表现得更好。

And then they just realized that if we give the model the end goal and we train it to accomplish that end goal, it has outperformed every other solution in almost every case that we go after.

这对那些从头到尾训练模型的人来说，是个非常有利的信号。

That bodes incredibly well for those that are training models end to end.

其次，现在的软件层很容易被快速复制出来。

The second thing to consider is that software layers are able to get recreated very quickly now.

我们正在构建一个评估集，用来衡量代理能否端到端地开发SaaS应用。2025年关注的是模型怎么提交代码合并请求，而2026年要看它能不能完整复刻出Slack。

Like we're building out an eval set that measures how effectively agents can build end to end SaaS applications, where 2025 was the year of how do you get a model to make a PR and a code base, and 2026 is the year of how do you get the model to clone Slack end to end.

这些能力在未来十二个月内就会出现在模型里。

Those capabilities are going to exist in the models in the next twelve months.

这对那些依赖软件护城河来维持业务的公司来说，意义重大。

That means very significant things for companies that are betting on software moats sustaining their businesses.

如果我们

If we

再往远了想，我们能多大程度上用代理完全内部搭建出Slack？

take that extrapolated further, how effectively can we build Slack internally agent led entirely?

这正好印证了‘SaaS已死’的说法——如果你是大公司，只需要一点定制或集成，比如房地产公司要对接特定的价格系统，干脆自己搭一个更划算。

That would very much concur with the SaaS is dead, because if you're a large company needing maybe small customizations, integrations, say you're a real estate company and you need very specific integrations to pricing providers, you'd build your own.

我基本同意。

I generally agree.

不过有个例外：当这些公司拥有网络效应时，它们的护城河可能比市场意识到的要深得多。

I think that the caveat is when those companies have network effects, there's probably a significant moat that isn't being priced in fully.

比如 Salesforce，有大量公司基于它的平台开发集成工具，形成了一个类似市场和网络效应的生态；Slack 也有 Slack Connect，对吧？

For example, Salesforce has tons of companies that are building integrations on top of their platform that creates this almost marketplace and network effect around it, or Slack has Slack Connect, right?

我觉得 Carta 也是个很好的例子，它的用户群体希望在所有公司内部都使用同一个平台，这就形成了强大的网络效应。

And I think even Carta is another great example of this whole network effect of the people that use it and want to use the same platform across all of their companies.

那些拥有网络效应的公司，某种程度上能创造更多价值，因为它们可以借助这些效应，以十倍的速度迭代产品，为用户带来更大价值，从而做出更值钱的产品，提高定价，增加收入。

I think that the companies that have network effects will be able to, in some ways, generate more value because they can iterate 10 times faster while leveraging those network effects to create more value for their customers, and therefore build more valuable products, charge more money, etcetera, and increase revenue.

没有网络效应的公司会非常吃力，因为它们所开发的纯软件产品，根本没什么能守住的护城河。

The companies that don't have network effects are going to struggle very significantly, because then there's not really a defensible moat in the pure software associated with the products that they build.

所以在我看来，这才是真正的试金石——能决定一家公司是彻底贬值，还是凭借把产品迭代速度提升十倍而获得巨大价值。

And so to me, that is the litmus test that determines whether this company is going to become worthless or whether this company is going to gain dramatic value from their ability to 10x product velocity.

你说我们越来越发现，

You said we're learning more and

模型本身才是产品。

more that the models are the product.

如果我反过来问，营销和市场策略才是产品呢？

What if I push back and say the go to market is the product?

你卖的对象是律所，关键是要能走进那些顶级律所——比如柯史莫、高伟绅、伟凯、 Clifford Chance——跟客户建立关系，再推动客户成功和产品使用。

You're selling to law firms, it's about being in the room with, you name your biggest law firms, your Coolies, your Goodwins, your Wyden case, your Clifford Chance, building the relationship with the buyer, and then the CS and the adoption.

这其实靠的是市场推广，而不是产品本身。

And it's actually in the go to market, not in the product.

我部分同意，但我想补充一点：与其说是市场推广，不如说是售后的深度介入更关键。

So I agree with this in part, but the caveat I would give is I think it's arguably more the forward deployed motion rather than the go to market.

售后介入指的是客户上线后的支持，而市场推广是上线前的销售。举个例子，就算你销售能力超强，卖的是SaaS产品，客户每年花一百万美元，但后来发现随便让Claude复制一下就能得到一模一样的功能，这时候你很难继续维持高价，哪怕你销售天下第一。

And forward deployed motion being the post sales, go to market being the presales, because ultimately, say you're just really good at sales, and then you provide a SaaS product, and you have a savvy customer who's spending a million dollars a year on the SaaS product, and they realize they could just like tell Claude to copy it and they'll get the same exact thing, it feels very difficult to maintain your pricing power, even if you're the best in the world at sales.

但如果你能深度介入客户，把公司内部那些隐性知识都教给系统，训练出真正懂业务的智能体，那这种能力就非常独特，别人根本复制不了。

Whereas on the other hand, if you have a great forward deployed motion, where you're going deep with a customer, you're training the agents based on all of this tacit knowledge within the company so that it understands how to perform effectively, that feels incredibly differentiated and hard to recreate.

这也是为什么OpenAI和Anthropic这些实验室都在大力投入这种售后深度服务。

And that's also the reason that we see obviously the labs at OpenAI Anthropic investing so much in this for deployed motion.

所以我觉得，Sequoia那篇‘服务才是新软件’的文章特别有道理——传统软件正在被一点点蚕食，真正能构建护城河的，是把服务叠加在软件之上，贴着客户的需求走完最后一公里。

And so I think that the Sequoia article that services are the new software resonated a lot in that these software boats are whittling away, and it's the ability to layer services on top of software to meet customer where they're at and go the last mile that is creating stronger defensibility.

你认同这个新潮的说法吗？

Do you buy this new sexy category?

我的意思是，风投人士确实很棒，但这种由AI驱动的服务新潮流，未来简直就是金矿。

I mean, venture investors are wonderful people, but this new sexy category that AI enabled services is like the future gold mine.

很大程度上，我同意。

I think in a large way I do.

你真这么觉得？

You do?

关键在于，你得确保他们真的在用AI，而不是装装样子。

I think the key thing is that you need to make sure that they're actually going to leverage AI.

现在有很多公司只是在做服务，却没从AI里获得真正的竞争优势，这点要特别小心，但我觉得这个思路非常合理。

I think there are a lot of companies that are just building services and not gaining a significant competitive advantage from AI and using that to that's the thing you've got to be careful about, but I think it's very rational.

拿Mercor举个例子：在这个把人才网络中的人力时间转化为高度定制化工作环境的过程中，涉及大量人为协调——怎么回答客户问题、怎么追踪项目KPI、怎么为每个项目开发专属工具。我们交付团队有大约一百到一百五十人，专门负责这些落地工作，帮客户走完最后一公里。

I'll give an example in the context of Mercor, which is that within this process of turning human time from the talent network into building these super rich environments that mirror everything that people could do in their jobs, there is a lot of human coordination of how do we answer people's questions, how do we track the KPIs of the project and manage it effectively, how do we build the bespoke tooling for that project, and we have about 100 people, or call it 150 people, in our delivery organization that do that for deployed work of helping to go the last mile for the customer.

但现在，我们的AI项目经理已经成功完成了第一个项目，端到端地管理了整个流程：它能招聘专家、回答他们的问题，还能用平台里的编程工具自己搭建标注工具，最终产出所需的数据格式。

But now we have an AI project manager that just completed its first project managing that entire thing end to end, where it's able to hire the experts, it's able to answer their questions, it's able to build the annotation tool using its coding tools within our platform and produce the end data type.

所有参与的专家都反馈说，给这个AI项目经理干活，体验非常好。

And the experts all had a really good experience on the project reporting to the AI project manager that was running it.

所以我认为，我们现在正亲眼目睹服务被自动化，这将带来经济上一场非凡的变革。

And so I think we're seeing in real time that services are getting automated and that that is going to be this extraordinary transformation in the economy.

驱动我们所用智能体的一个关键因素，就是支撑它们的令牌数量。

One thing that powers obviously the agents that we use is the tokens that power them.

我主要想说的是，我们的令牌使用效率提升了，单位令牌的成本也在下降。

And I the whole point was that we have increased token efficiency and token costs come down.

但对所有人来说，令牌成本其实都在上升。

Token costs are rising for everyone.

对的。

Mhmm.

你能不能说说，你认为未来六到十八个月令牌成本会怎么变化？为什么？

Help me understand how you see token costs changing in

接下来六个月到十八个月，令牌成本会怎么变化？为什么？

the next six to eighteen months and why that is?

这其实是个很有趣的案例，印证了西奥班悖论，就像我们之前讨论的——让人类更高效，反而创造了更多工作机会，对吧？

Well, it's a fascinating case study in Siobhan's paradox, similar to what we were talking about in the context of making humans more efficient leading to more jobs, right?

每年模型性能提升十倍，导致整体模型消耗量持续攀升，因为单位性能的成本在不断下降。

When we make models improve by 10x year over year, that has just been causing the total consumption of the models to go up and up and up as the cost per performance go down.

我觉得这个趋势还会继续大幅发展，企业里的令牌消耗量要到很久以后才会开始趋于平稳。

I think insofar as how it's going to develop is that this trend is going to continue very, very significantly before we start seeing any leveling off of token consumption within the enterprise.

现在我们花在内部智能体上的令牌费用，已经超过了员工薪资总额。

Like right now, we're spending more on tokens for our internal agents than we are on employee headcount.

我认为未来大多数企业都会变成这样。

And I think most businesses are gonna look like that in

你们花在智能体上的令牌费用，比员工工资还高？

But you're spending more on tokens for agents than you are in headcount.

没错。

Exactly.

所以你们在智能体上的令牌支出，已经超过工资了。

So your token spend on agents is more than salaries.

正是如此。

That's correct.

太惊人了。

Incredible.

所以我们是通过公司里一系列关键工作流程来管理的，就像我之前说的，每个流程都有一个AI项目经理负责运营。

And so the way we manage it is that we have a variety of these key workflows throughout the company, where we have an AI project manager, as I was describing, that manages operations.

我们有个面试问题机器人，已经完成了超过五百万次面试，问遍了所有面试问题。

We have our interview question agent where we've done over 5,000,000 interviews and asked all the questions in the interviews.

我们还有面试评分和候选人综合评估系统，帮助我们分析所有候选人，决定该录用谁。

We have our interview ranking, or the broader candidate ranking, where it helps to assess all of the candidates and figure out who we should be hiring.

我们也有用于财务自动化的小程序。

We have agents for accounting automation.

还有用于欺诈检测的小程序，等等。

We have agents for fraud detection, etcetera.

每个小程序都配有一个评估系统，告诉我们针对这个具体场景该用哪个模型，以及在价格和性能之间最优的平衡点在哪里。

And corresponding to each of these agents, we have an eval that tells us which model is best to use for this given use case, and what is the Pareto frontier of price performance for that specific use case.

这个评估系统帮我们决定该把推理预算投到哪里，该选哪家服务商，等等。

And that eval allows us to make the decisions around where should we be allocating our inference spend, what providers should we be using, etcetera.

我相信，随着时间推移，这种模式会逐渐在所有财富500强企业中变得高度相似——它们都需要一套统一的系统，用来评估和规范每个业务流程中的智能代理行为。

And I believe that over time, this is going to develop to look very similar across every Fortune 500, where they'll need to have this system of record for evaluating and specifying agent behavior across every workflow in their business.

它们会借助这套系统，把模型层变成标准化商品，因为它们希望实现模型之间的完全竞争，而且切换成本为零。

And they're going to use that to commoditize the model layer because they want to enable perfect competition for the models having zero switching costs.

因此，我们在企业市场增长得非常快，帮助它们建立这套系统记录，并为它们每个业务场景开发相应的评估体系。

And so we've been growing extremely quickly with the enterprise and helping them to populate the system of record and building out those evals for each of

也就是它们在公司内部各种实际应用中的具体场景。

the use cases that they have throughout their business.

你认为模型层真能实现这种商品化吗？就是企业能高效地把自身流程标准化，从而让模型层变成可互换的商品。

Do you think you will see that commoditization at the model layer, whereby enterprise clients are able to really efficiently package the workflows that they do, so it does commoditize the model layer.

因为现在还远远没达到这个程度。

Because right now it's not commoditized quite.

是的，我觉得关键区别在于，API层会先实现商品化。

Yeah, so I think the key distinction is that I think the API layer will get commoditized.

你当然可以围绕这些API构建出用户粘性强的流程和系统。

You can definitely build stickiness and workflows that people people have on top of those APIs.

比如说，我现在在云代码里跑了一堆自动化流程，我觉得要我把这些全都迁过去，估计挺麻烦的，至少我不会花这个时间。我在ChatGPT里也有不少类似的东西在跑。

Like for example, I have all of these routines running in Cloud Code, and I feel like it would probably be difficult, or I at least wouldn't put in the time to move those routines over, and I have a bunch of similar things running in ChatGPT.

所以我觉得人们会用各种方式建立黏性，但对于纯粹基于API的产品来说，如果我们每年在某个特定流程上花一千万美元，那肯定得有个评估体系来管它。

So I think that there's going to be various ways that people can build stickiness, but for pure API based products, where it's like, if we are just spending $10,000,000 a year on a specific workflow, obviously we're gonna have an eval for that.

每次有新模型出来，我们都会做基准测试，搞清楚该怎么无缝切换模型，怎么蒸馏模型。

And every time a new model comes out, we're gonna benchmark that and understand exactly how we should be hot swapping between models and distilling models.

为什么API层会变成大宗商品？

Why does the API layer get commoditized?

因为切换成本是零。

Because the switching costs are zero.

一旦切换成本是零，每两个月就会冒出一个全新的前沿模型。

Like when the switching costs are zero, there's a new frontier model every two months.

这意味着我们很快就会把旧模型换掉，对吧？

That means that we very quickly are going to swap them out, right?

归根结底，我们做决定的依据，就是跟这个流程对应的评估分数。

And ultimately the decisions that we make boil down to the score on the eval corresponding to that workflow.

所以模型之间可以一对一地直接对比，无缝替换，这几乎就是商品的定义。

And so it's very easy to compare model to model one for one in a perfectly hot swappable way, which is almost the definition of a commodity.

我还在震惊于你花在智能体上的token费用竟然超过了人力成本。

I'm still reeling from your token spend with agents more than headcount.

马克·贝尼奥夫前几天说，他们在Anthropic上花了三亿美元，听起来真是笔巨款。

Because actually, Mark Benioff said the other day that they spent 300,000,000 on Anthropic, which seemed like a lot of money.

但仔细算下来，这笔开销只占开发者薪资的3.8%，其实比想象中低得多。

But actually, when you baked it down, it worked out to be about 3.8% of developer salaries is being spent on Anthropic, which actually is much less than one would think.

是的。

Yeah.

你觉得二十四个月后，这个数字会变成什么样？

What do you think that is in twenty four months' time?

对Salesforce来说？

For Salesforce?

对。

Yeah.

我

I

我不知道五年后会怎样，但我敢打赌，五年后，普通企业花在计算资源上的钱会超过人力成本。

don't know about twenty four months' time, but I would bet that in five years, the average enterprise spends more on compute than headcount.

原因在于，模型的能力越来越强，用模型来完成每年十万美金的工作，回报率高得惊人，而且这种优势还会以指数级持续放大，而人类的智力做不到这一点。

The reason for that is that the models are just becoming so capable that it seems like there is just enormous ROI to being able to have models do something for 100 ks a year, that is going to continue compounding at an exponential rate in a way that human intelligence is not going to.

所以人类依然会在模型做不到的事情上发挥重要作用，但我预计，推理和计算的成本最终会超过人力成本。

And so humans will still play an important role at the things models can't do, but I expect that cost of inference, cost of compute will exceed that.

这让我特别感兴趣的是，针对你具体的业务流程做评估——比如以Salesforce为例，评估某个模型在代码生成方面的表现，往往能带来十倍的性价比提升。

The reason that that's so interesting to me is that having an eval for your specific workflow, like say we take the case of Salesforce, having an eval for how good is a specific model at code generation and their use case is often a 10x lever on the price performance of that model.

因为他们可以压缩模型，用开源模型达到甚至超过商业模型的性能，成本却低得多。

Because they can distill the model, they can have an open source model that is performing as well, if not better, for a dramatically lower cost.

随着企业中每个工作流程都在大规模转向计算和推理支出，他们需要评估体系作为权威标准，来判断这些流程是否执行到位，以及是否选对了合适的模型。

And so as we see this enormous shift towards compute and significant inference spend across every workflow in the enterprise, they're going to need to have evals that act as a source of truth for whether those workflows are being done correctly and whether they're using the right models to accomplish that.

恕我直言，现在的评估体系是不是还相当不靠谱？

With the greatest of respect, are evals today not relatively unhelpful?

这就像是驾照考试时，专门练怎么拐弯，可现实中根本不是这么开车的，实际用处不大。

It's like, how good are you at driving around the corner for the driving test in a very specific way, but that's not how it works in the real world, and it's actually not very practical.

这正是问题所在，对吧？

That's exactly the problem, right?

过去我们总依赖一堆学术评测，跟企业真正关心的结果完全脱节，大家拼命搞GPQA这种博士级推理、IMO奥数题、人文类期末考，全是些没人真正在意的冷门题目。

Is that we used to have this paradigm of all of the academic benchmarks that were totally disconnected from the outcomes that enterprises actually care about, where people were building everything ranging from GPQA for PhD level reasoning to IMO for Olympiad math to humanities last exam for this long tail of academic problems no one really cares about.

现在大家关注的是，怎么让模型端到端地完成实际工作，比如跟多个同事协作做财务模型，或者像我们之前说的做演示文稿。

And now they're focused on how do we get the model to do this end to end workflow coordinating with multiple colleagues for a financial model or a slide deck like we were discussing.

怎么让模型从头到尾自己搭建一个完整的SaaS应用？

How do we get the model to build an entire SaaS application end to end?

正因如此，评估体系才成为当前模型研发最关键的前沿课题，正在被大力推动。

That's why there's this enormous build out in pushing the frontier of evaluation as a critical research problem for the next frontier of model development.

好的，模型发展的下一个前沿。

Okay, next frontier of model development.

听你刚才说的这些，我总结出两点。

If I listen to everything that you just said, I would draw two conclusions.

第一，天哪，我们干脆把所有钱都投给 OpenAI 和 Anthropic 算了。

One, shit, we should just invest all of our money into OpenAI and Anthropic.

后来我突然意识到，大多数初创公司——你可以反驳我，尽管反驳——尤其是西海岸的那些，其实是先用前沿模型去试探边界，看看能走多远、能推到什么程度。

And then the realization dawned on me that the majority of startups, and you can shoot me down, again, shoot me down, is the majority of startups, say, especially on the West Coast, use frontier models to see where they can go and how far they can push them.

然后再用开源模型，经常是中国的模型，以更低成本尽可能接近那个水平。在这种情况下，OpenAI 和 Anthropic 自然会受到这种更具成本效益的开源模型的挑战，对吧？不管对错？

And then they use open source, often Chinese models, to get as close to that as possible at a much better cost basis, in which case OpenAI and Anthropic are inherently challenged by that much more cost efficient open source model, right or wrong?

我觉得这两点都有道理。

I think both are true.

五年后的需求会比现在高出好几个数量级，可能是四到五个数量级。

There's going to be many orders of magnitude more demand in five years than there are today, maybe four or five orders of magnitude more demand.

但竞争也会加剧，因为会有更多人通过蒸馏和微调开源模型来完成他们的工作流程。

But there's also going to be increased competition with people just distilling and having fine tuned open source models that accomplish their workflows.

在我看来，OpenAI 和 Anthropic 都是极具投资价值的公司，而且现在大家似乎也开始达成共识，这在几年前可不是这样。

Ultimately, I think OpenAI and Anthropic are incredible investments, and it seems like they're starting to be consensus around that in a way that there wasn't just a couple of years ago.

不过同时我也认为，五年后绝大多数的推理任务将使用开源、定制微调或蒸馏后的模型，而不是前沿模型。

But at the same time, I think that majority of inference in five years is going to be using a open source or custom fine tuned or distilled model, not using a frontier model.

嗯，有意思。

Okay, interesting.

你说他们显然是绝佳的投资。

You said they're obviously incredible investments.

五年后他们会变成什么样？

Where will they be in five years time?

从估值来看，这跟收入有关。

Valuation wise, That's revenue

我是说估值方面。

on valuation wise.

估值要是到万亿级别，真是惊人。

Valuation wise, wow.

要是把两家都估值到一万亿美元左右。

If we put them both at a trillion stake, give or take.

对。

Yeah.

这实在很难想象。

This is hard to imagine.

这个我五年后放给你听，到时候咱们一起回看。

This is one I'll play back to you in five years' time, and we'll both look back

然后会说，要么我们当初眼光太准，要么就是完全看错了。

and go, oh, either we were very prescient or just completely wrong.

我完全能想象，其中一家公司市值会达到一万亿美元，甚至更高。

I could definitely see one of them being a $10,000,000,000,000 company, maybe even significantly higher.

作为前沿模型，背后的机会实在太大了。

It feels like the opportunity associated with being the frontier model is so large.

它会吞噬经济中其他大部分需求，因为一旦你掌握了前沿模型，就能用它当老师，去提炼出更小但更优秀的模型等等。

It will just eat up so much of the other demand within the economy, because that also means when you have the frontier model, you can use that as a teacher model to distill your own models, to have the best small models, etcetera.

所以我猜，至少有一家的估值会超过一万亿美元。

So I would guess at least one of them is worth more than $10,000,000,000,000

我接下来的疑问是，你说算力投入会远超人力成本，那为什么我们不干脆把所有钱都投进英伟达呢？

My next assumption was when you talk about orders of magnitude more when you talk about spending more on compute than you will on salaries, why didn't we just put all of our money in NVIDIA?

我知道。

I know it

听起来有点傲慢又轻率。

sounds supercilious and glib.

我觉得这想法并不离谱。

I think it's not a crazy idea.

英伟达显然是个极其出色的公司，未来还会继续表现优异。

NVIDIA is obviously a phenomenal business that will continue to execute super well.

唯一的变数是，我们似乎正朝着多芯片格局发展，比如Cerebras就做得不错。

The only caveat is that it feels like we're starting to move towards a multi chip future, where obviously Cerebras is executing well.

我和Etch团队的人关系很好。

I'm good friends with the etch guys.

大多数实验室都在自研芯片。

Most of the labs are building in house chips.

所以我猜五年后，英伟达可能不再拥有绝对垄断地位，但这没关系——哪怕它只占全球最大市场的三成到四成，也依然是全世界最有价值的公司。

And so I would guess that in five years, it doesn't feel like NVIDIA has quite the same monopoly, but that's okay, because even if they only have 30% or 40% market share in the largest market in the world by far, that is the world's most valuable company.

说到全球最有价值的公司，你有没有发现，如今财富越来越集中在这前八大企业身上，前所未有地明显。

Speaking of the world's most valuable company, you're seeing this concentration of value towards the top eight names more than ever before.

今年整个市场的上涨，有84%都是由前十名公司推动的。

84% of the year's state rally was driven by the top 10 names.

你有没有担心过，财富这么高度集中在少数几家公司手里？

Do you worry about the concentration of value to such a small number of players?

某种程度上我确实担心，我们该怎么让社会大众更公平地分享这些成果。

Maybe to some extent, I definitely worry about how do we smooth out the benefits to society.

比如，怎么确保每个企业、每个人都能真正享受到人工智能带来的全部好处，而不只是旧金山那少数几个人？

Like how do we ensure that every enterprise and every individual is able to reap the full benefits of AI rather than just like a handful of people in San Francisco?

但说到底，我也觉得资本分配本身有个自然的逻辑——把算力给Anthropic这样需求迫切、能立刻用起来的公司，比给那些可能无法最大化利用它的公司更有效率。

But ultimately, I also think that there is some natural dynamic associated with capital allocation where it is going to be more valuable to give the compute to an anthropic where they have the marginal demand and can use that right away versus a less successful company that might not be able to create the most value with that.

所以从资本配置和效率的角度看，这可能是好事，只要我们能妥善应对贫富差距扩大的社会影响。

So I think that it's probably good from a capital allocation and efficiency standpoint, so long as we are able to manage the societal implications of increasing inequality.

说到贫富差距，你写过一篇关于这个的文章，我记得你推特上提过，应该取消美国底层一半人口的所得税。

Speaking of increasing inequality, you wrote an essay about, and this is taken from your Twitter, how we should eliminate income tax for the bottom half of Americans.

你详细说说这个。

Talk to me about that.

我对此深信不疑。

Well, believe this very strongly.

其实我上大学一年级时就写过这篇论文。

I actually wrote this essay as a research paper when I was a freshman in college.

那是我大学期间为数不多真正有成效的事之一。

It was one of the few productive things I did in college.

核心观点是，经济中最大的正外部性其实是就业。

And essentially, the thesis of this was that the largest positive externality in the economy is jobs, right?

人们总在谈各种经济理论，说碳排放、吸烟这些负面外部性该征税。

People talk about all of these economic theory of how we have negative externalities like carbon, or like smoking, or whatever it is, we should tax those.

但另一方面，就业才是最大的正外部性。

But on one hand, the largest positive externality is jobs.

可现实是，大多数经济体的税收结构反而在抑制就业——既对个人征所得税，又对企业征社保税。

Yet on the other hand, the way that most economies structurally collect income is by disincentivizing jobs, both on the income tax side by taxing the individuals, as well as on the payroll tax side of taxing the companies.

随着社会越来越多人失业，尤其是美国底层一半人口的工作前景变得越来越不确定，我觉得这个问题会变得极其严重。

And as we move towards a world where there's increased job displacement, increased uncertainty around how many jobs are there gonna be, especially for the bottom half of Americans, I think that this is going to become extremely problematic.

所以我建议我们转向一种新的模式，把税收重点放在那些不会严重打击经济激励的事物上。

And so I would suggest that we move towards a paradigm where we instead focus on taxes of things that aren't necessarily going to have a negative impact on incentives in the economy.

比如资本利得税，不管怎样我都会把钱投到资产里。

Like one great example is capital gains, where I'm going to invest money in assets regardless.

就算资本利得税提高了，我也不会因此就不投资了，对吧？

And so if there's higher capital gains tax, it's not like I'm just gonna not invest, right?

所以我认为，尤其是对短期资本利得征税，比我

And so I think that taxing capital gains, especially short term capital gains, which I

觉得长期资本利得对经济更有利，比起直接征收入税，这种方式在结构上要合理得多。

think is probably not as beneficial for the economy as long term capital gains, would probably be structurally much better off than taxing income.

恕我直言，如果你提高资本利得税，投资者冒险的意愿就会被打压。

With the greatest of respects, if you increase the tax on capital gains, you will disincentivize those investors to take risk.

我凭什么要多交钱？

Why the fuck should I pay more?

我本来就在冒风险。

I'm already taking a risk.

别人都不敢投，银行也不肯给钱，所有数据都告诉我别干，可我还是在投资创新。

I'm already investing in innovation when other people won't, when banks won't, when all the data tells me not.

现在你反而想对我这种行为多收税，对我承担这些风险多收税——

Now you wanna tax me more for doing that, for taking

问题是，当投资者承担极高风险时，通常都是通过投资组合整体来操作的，所以你完全可以对整个投资组合的收益征税。我知道你不喜欢听到资本利得税这个词，哈里，但——

The thing is when investors are taking very high risks, it's generally in an aggregated way in a portfolio, and so you would tax the gains on the portfolio overall, and so even if you have a portfolio of like, and I know that you don't like to hear the capital gains tax, Harry, but

不，不，不，我这么说完全是出于尊重，但你这想法完全不对，因为你只是把问题推走了。

No, no, no, I think, and I say this with the nicest respect, it's just wrong, because you just move.

我同意，最关键的是得小心别让投资者转移到其他地区，这确实会带来麻烦。

I agree that the main thing you need to be careful about is if people would move to other geographies, because obviously that creates problems.

但我认为资本利得税还是一个可行的选项。

But I think capital gains is one option.

不过我真不好意思

Think But I'm so sorry to

会上瘾，到时候你可以说退出。

be addicted, and you can say withdraw.

这会带来问题。

That creates problems.

对，这正是关键所在。

Yeah, that's kind of the whole point.

你跑到那些不征资本利得税的地方去，结果政府的税收就全没了。

You fuck off to somewhere that doesn't have capital gains, and then you lose all the tax revenue completely.

抱歉，原谅我，我们英国这里有绿党，有这种理想主义的潮流，结果一走，你就什么都没了。

Sorry, forgive me, we live in The UK where there's the Green Party, there's this idealist movement, Well, then we leave, and then you have nothing.

我同意。

I agree.

我认为有必要做些敏感性分析，看看加税会不会让人纷纷离开，最终反而让政府收入减少。

I think that there needs to be sensitivity analyses associated with how does the increased amount of taxation cause people to just leave and reduce overall government revenue.

但我认为另一个办法是，对那些可能并不太好的消费品征税。

But I think that another way of going about it is also taxing consumption of items that probably aren't the best.

我觉得太荒谬了，我们不征碳税，反而去征底层美国人的税。

Like, it's crazy to me that instead of taxing carbon, we tax the bottom half of Americans.

那为什么不直接征碳税呢？

But why don't we tax carbon, right?

碳排放明显是经济中的负面外部性，至少在美国，这居然没被征税。

That's like a very clear negative externality in the economy, at least in The US, that's not taxed.

我觉得还有很多容易下手的税源，比如那些不会扭曲激励、不会逼人移民的，比征底层人群的税要明智得多。

And so I feel like there is a lot of low hanging fruit with respect to things that we could tax without damaging incentives in a perverse way or causing people to flee the country that would be far better than taxing the bottom half of Americans.

而且，碳税只占政府收入的3%。

And the other thing is that it's only 3% of government revenue.

这么小的占比，政策制定者在权衡对民众的影响时，其实很容易做出决定。

The fact that it's only 3% feels like it's a very easy decision for policymakers to make in the grand scheme of the impact that it would have on people.

你会对预测市场征税吗？

Would you tax prediction marketplaces?

这本质上就是赌博。

It's gambling.

我大概会征税。

I probably would.

预测市场其实有它的价值，能帮人们更好地预判未来，为生活和投资组合做对冲，但征税应该也没太大问题。

There's probably some value of having good prediction marketplaces for allowing people to have effective predictions of the future and hedge things within their lives and investment portfolios, but it's likely okay to tax.

说到对底层50%的人征税这件事，杰夫·贝佐斯转推了我的观点，我简直高兴坏了。

The thing on that point around taxing the bottom 50% is Jeff Bezos retweeted me, which I was ecstatic about.

这真的很棒。

It's pretty cool.

确实挺让人兴奋的。

It was pretty great, yeah.

你见过最厉害的人是谁？

Who's the coolest person you've met?

我真的很

I really

喜欢詹森，也很欣赏萨提亚。

like Jensen, and I really like Satya.

我的意思是，有太多了不起的人了。

I mean, so many incredible people.

当然，达里奥和萨姆都非常出色。

Obviously, Dario and Sam are incredible.

但如果非要选一个，我觉得詹森真的太酷了。

But if I had to choose one person I mean, Jensen's so cool.

对吧？

Right?

他的夹克、风格，永远都那么到位。

Like the jacket, his style, he's always on point.

所以我会说，詹森可能是最让人着迷的一个。我其实很想问你一个问题，你不用回答，但很多人都问过我。

So I would say Jensen is probably one of the The fascinating one I would love to ask you, and you shouldn't give the answer to this, but I think this and people have asked

之前就有人问过我，这真的很难回答。

it for me before, it's very hard

你觉得谁原本以为会特别厉害，结果却让人大失所望？

to answer, is who did you think would be amazing who was surprisingly underwhelming?

别回答这个问题，这太难了。

And don't answer that one, it's a really

确实是个好问题。

good one.

这问题挺有意思的。

It is an interesting question.

我确实见过一些人，

And I have met a

那种让你觉得，哇，他们能做到，那我也行。

couple where you're like, Wow, that gives me confidence that I can do that too.

我倒想说一件事，当年我去乔治城大学的时候，没考上哈佛，那时候我就想，哈佛的人肯定比我聪明多了。

Actually, will say this one thing, which is that I remember when I went to Georgetown, I didn't get into Harvard, and I was like, wow, the people at Harvard are probably dramatically smarter than me.

后来我去了一家叫Prod的非营利机构，那里一堆哈佛和MIT的孩子在搞创业。

And I went to this nonprofit called Prod, where it was a bunch of kids from Harvard and MIT that were all building startups.

他们确实很聪明，这我承认，但我觉得大多数人其实都有种很平等的感觉。

And they're very smart, don't get me wrong, but I do think that most of us have this very equalizing feeling.

大多数做出非凡成就的人，你和他们多接触一段时间后就会发现，他们本质上就是普通人——不是全部，但大部分确实如此。

That majority of people that accomplish extraordinary things, when you spend more time with them, you realize that they're sort of just a normal person to a significant not all of them, but most of them to a significant extent.

我觉得这会让你觉得，比如我看到伊森·索恩在19岁的时候融了七千万美元，我就想，等等，伊森这个人其实挺随和的，人也很好，也许有一天我也能做成类似的事。

And I think that makes you feel like like when I saw Ethan Thornton from Mock raising $70,000,000 as a 19 year old, I'm like, Wait, Ethan is like a chill guy and a good friend, and maybe I could do something like that one day.

这会让你觉得自己其实也能做到更多。

It just gives you the sense of being able to accomplish so much more.

你这么说真有意思，朋友取得成就确实会带来一种扩散效应。

It's so interesting you said that, that kind of dispersion effect from seeing your friends achieve.

我觉得这在很多方面拖了欧洲的后腿。

And I think it's one thing that's held Europe back in many ways.

你跟世界上一些最大的模型提供商合作，

You work with some of the largest model providers in the world.

那你怎么看欧洲在提供顶尖模型方面竞争力不足的问题？

How do you feel about Europe's inability to competeprovide leading models to the world?

你看那些评测指标，说实话，

Like when you look at the benchmarks, mean, Mr.

Aal 可能排在第72名，就像欧洲歌唱大赛垫底那样。

Aal might make an entry at 72, it's like the Eurovision Song Contest kind of at the bottom.

我很喜欢 Off，也很喜欢 Mr.

I love Off, I love Mr.

Ahl 我确实为它在欧洲歌唱大赛上的表现感到骄傲，但说实话，我们在模型这块没做好。

Ahl, I'm very proud of it as a Eurovision, but shit, we haven't delivered on the model side.

欧洲能改善这种情况吗？

Does Europe improve that?

这重要吗？

Does that matter?

我觉得要改变很难，因为人才周围的网络效应太强了。

I think that it's gonna be difficult to change because there's just so many strong network effects around talent, right?

我们拥有最顶尖的人才，就连我认识的很多优秀的法国研究者，也都去 OpenAI、Dropbox 和 DeepMind 工作了。

When we have the best talent, even I knew so many brilliant French researchers that go to work at OpenAI and Dropk and DeepMind, right?

因为当这些实验室聚集了最顶尖的人才，人才就会不断向那里集中，进而带来更多的资金、算力和影响力，形成良性循环。

Because when we have the best talent at those labs, that's where they all aggregate, and then that compounds to them having more capital, more compute, more impact, etcetera.

所以我预计这一趋势将成为美国最大的优势之一，不仅是经济上的，更是地缘政治上的。

And so I expect that trend and to be one of the largest, not only economic, but geopolitical advantages that The US has.

那如果你是现在的欧洲，会不会干脆认了：模型竞赛我们输了，但至少我们还能当个主要的能源供应方，就像我所在的挪威那样？

So if you're Europe today, do you just go, You know what, sort it, we've lost that model race, but we can still be a dominant energy provider if you're Norway where I'm from.

实际上，挪威在能源供应方面表现还不错。

Actually, we do pretty well on Norway providing energy.

我们就这么认了吗？

Do we just accept that?

我会接受这个现实，是的。

I would accept that, yeah.

我觉得保留一些模型训练后的能力可能值得，毕竟在基础模型建成之后，知识蒸馏和应用开发还是会带来价值，但我不确定是否该全力去跟Anthropic正面硬刚。

I think that maybe it's worth having some post training capabilities, because there is gonna be value to distillation in some of the work that happens after foundation models are built, and there's definitely gonna be some value in applications, but I don't know if I would lean aggressively into how do we compete head to head with Anthropic.

你认同那种说法吗——我们需要自主的模型，因为不想让数据流向美国、中国，或者别的地方？

Do you buy the sovereignty argument of we need sovereign models because we don't want our data going to US or China or wherever that is?

在某些情况下，也许吧。

Maybe in some cases.

本地化确实有价值，我举个例子，经常有实验室来找我们，说他们的模型不能只懂美国法律，还得懂英国法律、法国法律，或者任何其他地区的法律。

There is value in localization, and I'll give an example, which oftentimes labs will come to us and say, they need their models, not just to be good at American law, but also to be good at British law or good at French law or whatever the jurisdiction is in the world.

我觉得，这将是让模型在不同司法管辖区真正派上用场的关键一步。

I think that that is going to be an important last mile in making the models useful in whatever jurisdiction that they're operating in.

但话说回来，这些实验室直接在法国雇上一万人，专门教模型怎么更好地理解法国法律。

That said, the labs are just going to hire 10,000 people in France to teach the models how to be better at French law.

我觉得别人很难阻止这种趋势，因为他们从其他领域积累的迁移学习能力实在太强了。

I don't think that there's so much that others are gonna be able to do to stop that because the transfer learning capabilities from all of the other domains that they're focusing on are just so powerful.

而且当

And when

你说雇一万人的时候，最让我震惊的是那股资金洪流。

you say about hiring 10,000 people, the thing that's just astonishing me is the wave of cash.

对我来说，我相信OpenAI也一样，但我亲眼看到Anthropic就是这么干的。

For me, I'm sure OpenAI is the same, but I've seen it specifically with Anthropic.

薪酬水平高得离谱。

I mean, insane levels of comp.

没错。

Totally.

你该怎么跟这个竞争呢？

How do you compete against that?

这确实是目前最让我们头疼的问题之一，尤其是现在创业热潮这么高，我们已经有三位员工创办了估值超过一亿美元的公司。

It's definitely one of the things that's most top of mind, in particular because markets for people founding companies are so hot, where we've had three employees that have founded companies worth in excess of 100,000,000

我看到你发的那些推文，就是那个‘McCormaffia’。

I saw your Tweets where you do the McCormaffia

推文？对，就是那个。

Tweets, Yeah, exactly.

我们公司还很年轻，对吧？

And we're a very young company, right?

我觉得这在很多方面都很难。

And I think that it's difficult for a variety of reasons.

很多人可能根本不清楚创业有多难，Harry，你很清楚，成功概率有多低，我们一路走来有多幸运。

A lot of people probably don't have a full understanding of just how hard it is to build a company, as you know well, Harry, and how low the probability of success is and how fortunate we were and how lucky we got along the way.

所以我觉得这确实是其中一个主要的挑战。

And so I think that that's definitely one of the large challenges.

前几天我还在招人，对方拿到了TBD公司每年两千万美元的现金offer。

Even like there was someone I was hiring the other day and he had an offer for $20,000,000 in cash per year from TBD.

这种事我们几乎天天遇到。

And like that's the kind of stuff we run into on a regular basis.

TBD？

TBD?

就是Meta的超级智能团队。

Meta's super intelligence group.

两千万元现金？

20,000,000 in cash?

每年。

Per year.

或者给的是股票，但可以随时套现。

Or it's in stock, but liquid.

这真不好竞争。

That's hard to compete against.

确实很难。

It's hard

很难竞争，没错。

to compete against, yeah.

这种情况会变吗？

Does that change?

会不会继续不断攀升？

Does that just continue to escalate?

可能对一小部分人来说，还会继续上涨。

It'll probably continue to escalate for a smaller group of people.

但我猜，随着越来越多的人了解这些实验室的运作方式，以及如何训练前沿模型，市场上具备这类技能的人会越来越多，薪资水平也会慢慢回归理性。

But I also suspect that as more people gain knowledge of how these labs operate and what the capabilities of how to train a frontier model, that means that there's going to be more supply in the market for people that have that skill set and thus a little bit more reasonable pricing.

所以我觉得这种疯狂还会持续一阵子，但希望至少市场中前1%的超高薪资能自己回落下来。

And so I expect there to be some craziness that continues, but hopefully sort of the ninety ninth percentile at least within the market will bounce itself out.

现在最难招聘的是什么岗位？

What is the hardest role to hire for today?

研究人员。

Researchers.

是因为人手不够吗？

Just because of supply?

供需关系不平衡。

Because of supply and demand.

现在这个市场，需求是供给的十倍，所以特别难招。

It's just this market where there's 10 times more demand than there is supply, and that makes it very difficult.

我们一直在组建一支非常强大的研究团队，比如Laura论文的第一作者爱德华·胡，他以前在OpenAI，现在和我们在一起，还有其他不少顶尖研究者，但市场确实越来越热了。

We've been building out an incredibly strong research team, like Edward Hu, the first author on Laura, who was previously at OpenAI, is working with us and a bunch of other top researchers, but the market is definitely getting very hot.

招聘一位高质量的AI研究员要花多少钱？

How much does it cost to hire a high quality AI researcher?

顶尖人才年薪的股票激励常常要上千万，确实是这样。

Oftentimes it would be in the tens of millions of stock per year, For the really good people, yeah.

我记得以前研究人员的薪水并不高。

I remember when researchers weren't paid very much.

大概是十年前、十五年前吧，他们是社会上一群被低估但非常聪明的人。

This was like ten, fifteen years ago, were like the underpaid but brilliant people in society.

是的。

Yeah.

现在感觉情况已经不一样了。

Now I feel like that's relatively changed.

对。

Yeah.

现在经营公司比以往更难了吗？

Is it harder than ever to run the company?

我觉得没有。

I don't think so.

举个例子，去年年初我们公司只有40人，年收入却达到了五千万。

Like, to give a frame of reference, we were 40 people and 50,000,000 in revenue run rate last year, at the start of last year.

从那以后，我们的团队规模扩大了七到八倍，业务整体规模也增长了二十五到三十倍。

Since then, we've seven or 8x headcount and we've increased the broader scale of the business by 25x, 30x.

一路走来，要跟上这种增长确实压力很大。

It's definitely been very stressful to keep up with the growth along the way.

但现在我们有了支持部门，比如财务和法务，人力资源也在逐步建立起来。

But I think that now we have the supporting functions, like we have HR, or not really HR, but we have finance and we have legal and we're building out HR.

这样一来，我就不用再亲自处理那么多琐碎的紧急事务了。

And that brings some sense of stability where I don't have to deal with all of these little escalations.

我可以专心做产品、搞研究，还有和客户交流。

And I'm able to just spend my time focusing on building great products, research, and time with customers.

我觉得，这让公司运营变得轻松多了，轻松不少。

And that, I think, has made it easier, significantly easier to run the business.

我现在每天忙得不可开交。

I get in a lot

如今我说的每句话都会引来一堆批评，这其实挺好的。

of shit for everything I say these days, which is wonderful.

我团队一看到我就说：‘天啊，哈利。’

My team just go, Oh, no, Harry.

问题是，我从不故意惹怒别人，可就是有人特别讨厌我。

The trouble is I don't deliberately rage bait, but people just hate me.

这真是最糟糕的事。

Which is the worst thing.

是的。

Yeah.

不过说到HR，我之前在AppLovin和亚当做完节目后发了条推文，说没见过哪个出色的CEO是喜欢HR的。

But HR, I tweeted after a show with Adam at AppLovin, no great CEO that I've met.

这话说得没错。

And it's true.

他们确实喜欢HR。

Loves HR.

HR只会拖慢你的进度。

They slow you down.

他们搞一堆政策和流程，纯粹是添堵。

They implement policy and procedure, and it's just pain.

你同意我的说法吗？

Do you agree with me?

我要补充的是，我觉得这真的特别重要。

The caveat I'll give is that I think it's really important.

我们从40人扩张到400人的时候，文化传承确实遇到了不少问题。

We definitely had challenges in scaling culture when we went from 40 people to 400 people.

具体表现在哪些方面？

How does that show up?

方方面面都有，比如保持顶尖的人才标准，让每个人都认同公司的使命，甚至细到让主管及时跟团队沟通绩效反馈，别让人突然被考核结果吓一跳。

Well, it's so many things, ranging from making sure that we keep a really high talent bar, to making sure that people are bought into the mission of the company, to even the tactical things of making sure that managers are communicating to their team about their performance review and how they're doing so that they're never surprised by a performance review.

而且我们团队里年轻经理多，很多人是第一次带人，不习惯给反馈，也很难持续贯彻团队的价值观和使命。

And when we have a young team with a lot of first time managers, that just creates culture challenges of people that aren't used to giving feedback and maintaining all of the values and commitment to the mission of the team.

所以某种程度上，我同意你的观点。

And so I think that to some extent, I agree.

我觉得有些大型科技公司可能在赋予HR权力方面做得过头了，但我同时也认为这很重要。过去大约一年半的时间里，我们学到的一个重大教训是：在扩大团队规模时，必须把这些基础工作落实到位，否则就会引发问题。

And I think that some of the big tech companies probably go too far on empowering HR, but I also think that it's important, and one of the large lessons we've had over the last eighteen months or so is that it's critical to really get these foundations in place as you scale headcount, otherwise it creates problems.

文化方面的挑战。

Culture challenges.

节目录制前，我们提到过节目结束后关于dosh的讨论。有几个人以为Mercor的管理方式就像九月那样，是那种打卡上班、打卡下班的模式。

Before the show, we said that after the show with the dosh, a couple of people thought that like September was the way that like Mercor is run, and it's like clock in, clock out.

事实并非如此，你是怎么看待这个问题的？

Why is that not true, and how do you think about that?

之所以不成立，是因为公司从未强制规定工作时间。

So the reason it's not true is that we've never mandated hours at the company.

而且我确实工作非常努力。

And obviously I work extremely hard.

Adarsh 也极其努力。

Adarsh works extremely hard.

我们基本上是从醒着开始工作到睡着为止，就算在健身的时候，脑子里想的也还是工作。

We work from when we wake up until we sleep pretty much all the time, aside from maybe working out, but I'm still thinking about work during that time.

我们整个领导团队大多数也都是这样，但同时，我的团队里很多人有孩子，我们希望他们能回家陪家人。

Most of our leadership team of course does as well, but at the same time, majority of my leadership team has kids, and we want them to be able to go home and see their families and all of that.

所以我觉得，这中间需要平衡：打造一家卓越公司确实需要对事业有极大的投入，但也要确保这是一个能让全球顶尖人才长期安心工作的环境。

And so I think that it's some combination of knowing that building a legendary company requires immense dedication to the mission of the business, while also recognizing we need to ensure that it's a sustainable environment for the best people in the world to do their life's work.

准备好了吗？来个快问快答？

Are you ready for a quick fire round?

当然。

Of course.

你打算上市吗？

Would you like to go public?

肯定要上市。

Definitely.

什么时候？

When?

未来几年内吧，我觉得所有卓越的公司最终都会上市。

In the next few years, I think that all legendary companies eventually go public.

所以这确实是成长过程中重要的一环，尤其是当我们公司规模比现在大得多的时候，但今年或明年我们并不着急上市，毕竟我们俩辍学还不到三年，公司还很年轻，我们想先把企业端的所有工作都扎实做好，再考虑上市。

And so it's an important part of the journey and maturing and having a much larger company than we have today, but it's not something we're rushing to do this year or next year, in part because we dropped out of college less than three years ago at this point, and it's still a very young business where we want to make sure that we properly actualize everything that we're working on, on the enterprise side especially, before going public.

别笑。

Don't laugh.

你晚上躺床上的时候，会不会突然想：天啊，这真够疯狂的？

Do you ever lie in bed at night and just go like, Wow, pretty wild.

我每天都觉得自己像在做梦，特别感激我的团队，还有阿达什和苏里亚，是他们让这一切成为可能。我本来能想到一百种不同的发展路径，结果可能完全不一样。

I'm always pinching myself, and I feel extremely grateful for the team and Adarsh and Surya and how all of them made it possible because I could have imagined a 100 things that would have gone differently, and we'd be in a totally different circumstance.

过去一年里，你对什么事情改变了看法？

What have you changed your mind on in the last twelve months?

我改变了对什么的看法？

What have I changed my mind on?

以前我一直在想，基础模型实验室会不会成为全球最大的企业，毕竟就像你刚才问的，这些模型在竞争激烈的环境下，真能维持住定价权吗？

I used to have some questions around whether the Foundation Model Labs would be the largest businesses in the world because of the exact things you asked about in the context of how much those models are gonna be able to maintain pricing power amidst a competitive environment.

但看到这些业务的收入增长速度后，我现在非常确信，它们会成为世界上最有价值的公司。

But I think that as we've seen the sheer revenue ramp of these businesses, I've gained immense conviction that they will be the most valuable companies in the world.

你可以投资OpenAI或者Anthropic。

You can invest in OpenAI or Anthropic.

选哪个？

Which one?

哦，我

Oh, I

我不能回答这个问题。

can't respond to that.

我会

I would

选那个。

choose that.

到目前为止，还有哪些投资人你特别希望拉进来？

Who do you not have as an investor in the company yet that you would most like to have?

我非常钦佩杰夫·贝佐斯。

I really admire Jeff Bezos.

我觉得他对亚马逊的企业文化特别有原则。

I think he's so disciplined about the culture of Amazon.

这一点一直让我印象深刻。

That's one of the things that's always stuck with me.

那里每个人都清楚公司的价值观，而且都朝着同一个战略方向努力。

Everyone there just understands the values and is steering in the same direction as such a strategic business leader.

我从来没见过他，

I've never met him,

但我一直很想见见他。

but I've always wanted to.

你最敬佩哪个竞争对手？为什么？

Which competitor do you most respect and why?

好问题。

Good question.

我很佩服Surge的Edwin，他一直紧贴前沿研究，这一点我们自己也一直在做。

I admire that Edwin from Surge has done a really good job in staying super close to research, and it's something that we've obviously been doing a lot of as well.

但我觉得，这可能是我们和Surge之间最大的区别之一——我们有能力训练模型，还能招到世界上顶尖的研究人员，我对他们在这一方面的执行力非常钦佩。

But I think that's probably one of largest things that differentiates both us and Surge is our ability to train models, to hire some of the best researchers in the world, and I admire them for execution on that front.

有多少比例的数据提供商只是单纯的、尊重彼此的交易型人才平台？

What percent of data providers are just respectfully transactional talent marketplaces?

你是说在数量上，竞争对手有多少？

In terms of volume or number of competitors?

竞争对手的数量。

Number of competitors.

大概一半。

About half.

一半？

Half?

是的。

Yeah.

我真希望也能那样。

I'd love that.

你今天最想改变自己工作中的哪一部分？

What would you most like to change about your role today?

有不少人事事务都会 escalated 到我这儿来。

I would say that there's a decent amount of HR things that get escalated to me.

所以我们正在找一位非常出色的人力资源负责人，能接手这些事。

And so we're looking for a really strong head of people that is able to handle a

很多这样的工作。

lot of this.

最后一个问题是，老兄。

Final one for you, dude.

别人对你做过最暖心的事是什么？

What's the kindest thing that anyone's ever done for you?

有一件事让我印象特别深，我想这得归功于整个 prod 社区，尤其是像 Rob Walkin、Ben Spector 和 Richard DeHaan 这几位。

One that really stuck with me is I remember, and I'll probably attribute this to the entire prod community, namely especially a couple of people like Rob Walkin, Ben Spector, and Richard DeHaan.

Prob 是个起源于 MIT 和哈佛的非营利组织，我当时算是个外人，因为没考上那两所学校，去了乔治城大学。

But Prob was this nonprofit that got started at MIT and Harvard, and I was sort of a blow in because I didn't get into those schools, but I went to Georgetown.