20VC：认知与工厂之争：Vinod Khosla掀起风暴 | OpenAI接近700亿美元年化收入：Anthropic面临威胁 | ElevenLabs估值翻倍至220亿美元，Salesforce以20亿美元收购Listen Labs

他以后回想起来可能会说：天哪，我真希望当初没做那件事。

He will probably look back and say, "Gosh, I wish I hadn't done that."

这等于给所有竞争对手递了把刀，真希望这事没发生。

It's weapon in the hands of everyone who competes with them, wish it hadn't happened.

这里很多问题在于，法律义务是什么，不成文的规则是什么，以及违反这些不成文规则会有什么后果。

A lot of what's going on here is what are the legal obligations versus what are the implied rules and what are the consequences of breaking the implied rules?

其中一条虽然不是法律但也是不成文的规则是，风险投资公司如果同时投资两家直接竞争的公司，会非常棘手。

And one of the non-legal but nonetheless implied rules is it's really hard for venture firms when they do two directly competitive deals.

唯一的解决办法就是，你得拼命强调你对两家公司一视同仁，都像自己的孩子一样爱。

And the only way you get around it is you make all this argument about you love your children equally.

在公开场合，你对其中一家公司的事绝口不提，对另一家也一样。

You don't say anything in public about one, about the other.

然后偏偏出了这么一件事，把所有的墙都撞破了。

And then you have something like this that crashes through every wall.

这恰恰说明，同时投资两家直接竞争的公司，这种局面非常危险，尤其是当你深度参与这些公司的时候。

And it just shows that those situations of two directly competitive investments are very fraught, especially if you're actively involved in the companies.

这真的很没帮助。

It was very unhelpful.

这对Khosla很没帮助，对两家公司也都很没帮助。

It's very unhelpful for Khosla, very unhelpful for both companies.

感觉就是，恶心。

It's like, ick.

这可不是风投的光彩时刻。

It was not venture's finest hour.

我们继续吧。

Let's move on.

够了。

Enough.

也许他是在当乐观派，本意是好的，是严厉的爱。

Maybe being the Pollyanna, he meant it positively as tough love.

挣扎中的公司——有个理论，对最挣扎的创始人要严厉。

Struggling sec- There's a theory, be harsh on the most struggling founders.

他觉得激励那些苦苦挣扎的二线竞争对手会有效果。

He thought struggling second-tier competitor would be motivating.

哈里，我显然在人员管理方面让位于德夫和CEO，但我隐约记得有句话叫公开表扬，私下批评。

Harry, I yield, obviously, to Dev and the CEO as man manager side here, but I do vaguely remember something about praise in public and chastise in private.

对，这看起来反了，但——

Yeah, it seems backwards, but-

我觉得这就是这里的用意。

I think that was the mission here.

谁知道呢。

You never know.

这确实让人摸不着头脑。

It seems baffling.

过去几周我们聊了很多关于开放模型的话题。

We've been talking a lot about open models over the last few weeks.

我们看到了Reflection Ship Beam。

We saw Reflection Ship Beam.

我觉得那个故事真的很有意思。

I thought that was a really interesting story.

我们还聊到了你的一家公司，Rory，就是在Intercom之前卖给Salesforce的那家。

We also discussed, Rory, one of yours before Intercom, obviously selling to Salesforce.

Salesforce还花了20亿美元收购了Listen Labs。

We had Salesforce buying Listen Labs too for $2 billion.

这两家比起来，你更愿意去哪家？

Of those two, where would you rather go?

我觉得两家都值得花时间聊聊。

I think both are worth some time.

我很想听听你对Reflection的看法，因为你看，他们其实已经说了很久要发布产品，而且从我们之前聊的来看，我觉得一个开源权重、美国出身的模型，现在是一个非常有意思的战略位置。

I'd love to hear your thoughts on Reflection, because look, they've been frankly talking about shipping for a long time, and from what we said earlier, it seems to me that being an open-weight US-sourced model is a super interesting strategic position right now.

所以如果有人能推出接近前沿水平的产品，哪怕晚个六个月、十二个月，就凭市场的动态，那也是很好的位置。

So if someone can ship something that's near frontier quality, even if it's six, 12 months off, just from the dynamics of the market, that's a nice place.

所以你们怎么看，Dev？

So what are your guys' thoughts, Dev?

对，我今天早上其实还发了条推文。

Yeah, I actually tweeted it out this morning.

我其实觉得企业可以两全其美。

I actually thought that enterprises kind of can have their cake and eat it too.

第一，你有一个美国开源的模型，智能水平接近前沿，而且成本效益高三到四倍，也就是更便宜。

One, you get a US open source model that's near frontier intelligence, that's three to four times more cost-effective, i.e., cheaper.

这对任何企业来说都是好消息，因为显然有些企业会对使用中国的开源模型有所顾虑，尤其是那些受监管的行业，或者处理大量敏感信息的行业。

And that's music to the ears of any enterprise, because there's obviously going to be some hesitation about using Chinese open source models, especially if you're in a regulated industry or industry that deals with a lot of sensitive information.

对我来说，这是一个非常值得关注的发布。

To me, this was a very noteworthy announcement.

再次说明，我是红杉资本在Reflection的投资人，但我觉得这对企业来说非常好。

Again, I do want to disclose as Sequoia's investor in Reflection, but I think it's great for the enterprise.

我今年去了Dreamforce，意外地和很多高管进行了不少对话。

I was at Dreamforce this year, and got to actually have a surprisingly large number of conversations with execs.

尽管中国开源模型的影响力势不可挡，但和我交谈的人中，没有一个真正想用中国的模型。

Notwithstanding the incredible force of nature of Chinese-based open-weight models, nobody I talked to really wanted to use a Chinese source model.

不管对不对，公不公平，每个人虽然都觉得是迫于成本压力，但真的没人想用中国的模型。

Right or wrong, fair or not, everybody that was, they felt like it was under duress for cost, but no one really wanted to use one.

再说一遍，不管对不对。

Again, right or wrong.

所以如果他们这次做对了，需求可能会像洪水一样涌来。

So if they nail this right, it could be a torrent of demand.

我同意，而且我觉得这也印证了一点，就是不是每个工作负载都需要最顶尖的智能。

I agree, and I think that this also reinforces the point that you don't need the frontier-level intelligence for every workload.

如果它接近顶尖水平，就像Rory说的，跟最新模型差不到六个月，而且能覆盖你90%的用例，特别是像编程和智能体这类推理任务，我真看不出企业有什么理由不立刻去跟他们谈。

If it's close to the frontier, as Rory said, within six months of the latest models, and addresses 90% of your use cases, especially around reasoning use cases like coding and agentic workloads, I don't see why enterprises would not want to talk to them immediately.

这又回到了第一个话题。

Ties back to the first topic.

从某个层面看，到目前为止的证据表明，并没有出现大规模的迁移，因为人们比较懒，而且账单也没到天文数字，所以他们就一直用顶尖模型，或者降级到Anthropic或OpenAI的便宜一点的模型。

At one level, so far the evidence has been that there hasn't been a massive push because people are lazy and the bills haven't been astronomical, and therefore they've just used the frontier model and maybe downgraded to the Anthropic or OpenAI less expensive model levels.

但我同意，我觉得现在有意思的是，因为数字变得实在太大了。

But I agree, I think the interesting thing now is because the numbers are getting so huge.

如果OpenAI和Anthropic各自年收入达到700亿美元，那就是1400亿美元。

If OpenAI and Anthropic are each making $70 billion in ARR, that's $140 billion.

这比微软在Windows和Office上长期征收的税还要多。

That's more than Microsoft taxed for the longest time across both Windows and Office.

这占美国企业利润的一大块，对吧？

That's a chunk of US corporate profits, right?

到那时，如果你是CFO，你可能会去找你的CIO说：“我不在乎直接用Anthropic更方便，兄弟。

And at that point in time, if you're a CFO, you probably come in to your CIO and say, "I don't care that it's easier to just use Anthropic, dude.

我们需要从这笔token账单上砍掉300万。”

We need to shave 3 million off this token bill."

所以，开始用Jev来处理你20%的决策，用Reflection或其他什么来处理非前沿的任务，然后想办法用Anthropic来处理那些难搞的事情。

So start using Jev for your 20% of your decisions and start using Reflection or something for anything that's not frontier and figure out a way how to use Anthropic for the hard shit.

我猜，Dev，你们很快就会开始进行这些对话，你们的客户也会。

I'd imagine, Dev, you guys will start to have those conversations, and all your customers will.

这就是为什么，再次强调，Anthropic现在的增长率如此引人注目，而在这种被逐渐蚕食的情况下，他们还能保持那种无与伦比的增长吗？

That's why, again, these growth rates are so interesting right now for Anthropic, and can they still maintain that unparalleled growth rate in the context of this kind of chipping away?

就像《老人与海》里那样，老人把鱼拖回来，结果其他鱼在啃食鱼的身体。

It's like, remember "The Old Man and the Sea," the guy's bringing back the fish and then the other fish are gnawing at bits of the body.

他们正在从他们身上夺走大块收入。

They're just taking away slugs of revenue from them.

而且，你还有Jev和其他一堆Jev的竞争对手，也在以不同的方式攻击他们。

Well, and then you've got Jev and you've got a load of the other Jev competitors also, in a different way, attacking them.

是的，而且前进的势头一直很强，所以这一切就像是在说：‘谢谢，Rory，这很有趣，但我有没有提到我们去年和前年都增长了10倍？’

Yeah, and again, the forward momentum has been such that all this is like, "Yeah, thanks, Rory, that's super interesting, but did I mention we 10X'd last year and the year before?"

我觉得还有一点很重要，Jason，我很好奇你在Dreamforce看到了什么，但我认为在企业的实际使用情况和能力之间存在着很大的差距。

The other thing I think is important to understand, and Jason, I'm curious what you saw at Dreamforce, but I think there's a big gap between where the capabilities are and where actual usage is in the enterprise.

我认为企业内部真正有经验的技术人才很匮乏。

I think there's a paucity of real experienced skill sets inside the enterprise.

所以我觉得我们会看到杰文斯悖论发挥作用，基本上随着成本下降，人们会越来越放心地在企业内越来越多的用例中部署它。

And so I think we're going to see Jevons paradox kind of come into play, where basically, as these costs come down, people will get more and more comfortable with deploying it for more and more use cases across the enterprise.

随着他们对这些技术越来越熟悉和熟练，他们会开始将其部署到更多的用例中。

As they get more conversant and skilled around these technologies, they'll start deploying it for more use cases.

我在MongoDB也看到了这种情况。

I see this even at MongoDB.

所以我觉得，坦白说，这是一个巨大的市场，我认为每个人都能成功。

And so I think you're going to see that frankly, this is a massive market, and I think everyone can be successful.

我不评论前沿实验室和开源模型的增长率，但我不认为这是一个零和游戏。

I won't comment on the growth rates of the frontier labs versus the open source models, but I don't see this as being a zero-sum game.

回答这个问题，我在Dreamforce学到的最显而易见的答案是，我遇到的每个人都被内部对token的需求压得喘不过气来。

To answer this, the Captain Obvious answer to what I learned at Dreamforce, everyone I talked to, everyone was just overloaded with the amount of internal demand there is for tokens.

不管需求是什么，都超出了他们的想象。

And whatever it is, it's more than they can imagine.

不管使用是否得当，不管是不是白白烧掉，不管上限是多少，需求可能比他们能想到的还要高一个数量级，对吧？

Whether it's being used properly, whether it's being burned for nothing, whatever the cap is, the demand is probably an order of magnitude more than they've figured out how to surface, right?

这可能会导致使用次优模型。

That may lead to using suboptimal models.

这可能会造成一些浪费，但每个人都得管理这些，再加上对使用中国模型的担忧，不管对错，这都是一个巨大的机会，如果你能抓住的话。

It may lead to a lot of waste in some ways, but everyone's got to manage it, which just that and the fear of using China-based models, again, right or wrong, is just a huge opening if you can nail it.

我只想补充一点，纯属好玩，我们可以继续往下聊，就是，天哪，我个人对公开的评测越来越怀疑了。

The only thing I will just add, just for fun, we can move on, is man, just personally, I'm getting more and more skeptical of published evals.

当然，你可以让一个评测看起来很棒。

Like of course you can make an eval look great.

它到底有多快？

How fast was it?

真正的输出是什么？

What were really the outputs?

它在真实工作流中表现如何？

How well did it perform on a real workflow?

它在关键任务工作流中表现如何，而不是告诉你某件商品有货但实际上没有，这其实也算关键任务工作流。

How well did it perform on a mission-critical workflow versus telling you an item was in stock when it wasn't, which is still kind of mission-critical workflow.

我发现现在每个评测都需要加三个星号和四个剑标。

I found that every single eval needs like three asterisks and four daggers next to it today.

所以我不太相信Beam像他们说的那么好，但我刚才趁我们在聊的时候问了一下，能不能在播客结束前上OpenRouter查一下，但它还没上线。

So I'm skeptical Beam is as great as they say it is, but I just asked while we're on if I could get on OpenRouter and check before we end the podcast, but it's not there yet.

但我会试试。

But I would try.

但不是说我觉得这些都不好，我只是觉得所有的评估都太偏重象棋锦标赛了，所以我有点怀疑。

But it's not that I don't think these are all great, I just think all the evals are so biased and focused on chess championships that I'm slightly skeptical.

我们有Nemotron，有Thinking Machines，还有Reflection。

We have Nemotron, we have Thinking Machines, we have Reflection.

12个月后，开源流程中的token有多少比例会流经美国，又有多少会流经中国？

In 12 months' time, what percentage of token in open source flow will go through US versus Chinese?

嗯，你看，就两件事。

Well, look, just two things.

就像Dev一开始说的，开发者会变，尤其是在这里。

As Dev made the point in the beginning, developers will change, and especially here.

这比换数据库容易多了，对吧？

This is a lot easier than swapping out your database, right?

换数据库也能做，但我最近刚换过，跟你说，那真是费劲。

That could be done, but having recently done it, it's some work, man, I got to tell you.

就算有语言模型也一样。

Even with an LLM.

听我说，这还是有工作量的。

Listen, this is still work.

你得把提示词写好。

You have to qualify your prompt.

你得重新调整工作流程。

You have to redo your workflows.

这可不是网上说的那么轻松。

It's not nothing like the internet says.

但我能想象，12个月后可能会有一半。

But I could see it being half- In 12 months.

12个月时间很长，我觉得主权加美国本土真的很重要。

12 months is a lot of time, and I think sovereignty plus US-based is really important.

我不是说一定会发生，因为我不确定……再说，我怀疑这种对等是不是真的存在，还是只是表面上的。

I'm not saying it will happen, because I don't know that... Again, I'm skeptical the parity is really there versus pretend there.

但如果你比X早三个月，按照Dev的说法，一年内可能就完成一半了。

But if you're even X minus three months, to Dev's point, it could be half in a year.

一年时间很长。

A year is a long time.

但这让我很有信心。

But that gives me a lot of confidence.

我是Fireworks的投资者，红杉也是，我觉得作为Fireworks、Base10或Modal的投资者，最大的担忧是，大型企业其实很在意中国模型，这会阻止他们与Fireworks或Modal这样的合作伙伴合作，然后他们就会转向Frontier。

I'm an investor in Fireworks, as is Sequoia, and I think the biggest concern I would have, as a Fireworks or a Base10 or a Modal investor would do, is that actually big enterprises are concerned about Chinese models, and that would prevent them from engaging with partners like Fireworks or Modal, and they then go to Frontier.

所以你说50%，这让我看到你认为的质量水平足以达到这个比例，这让我很兴奋。

And so the fact that you say 50% shows me that the levels or quality you think is sufficient enough that it would be that rate, that makes me excited.

我觉得企业除了成本之外，还会非常关心安全性，能否信任这些系统，而我们现在才刚刚开始关注这个问题。

I think the other thing that enterprises are going to care a lot about above and beyond cost is safety, being able to trust these systems, and I think we are just getting focused on this.

显然，上周在华盛顿有一个大型会议。

Obviously, there was that big conference last week in Washington.

很明显，会有某种自我监管的方式来处理安全问题。

It's clear that there's going to be some self-regulating approach to figuring out how to address safety.

挑战在于如何定义安全，对吧？

The challenge is how do you define safety, right?

因为如果有人能很好地定义它，我真的很想知道，但安全就像美一样，因人而异。

Because if someone knows how to define it well, I'm really curious, but its beauty is in the eye of the beholder.

但我们需要某种方式或机制，让人们能够像评估基准一样来判断这些模型有多安全，当然，人们可以操纵基准。

But we need some way or some mechanism for people to be able to determine, much like a benchmark, and yes, people can game a benchmark, is how safe are these models?

这些模型越安全，无论采取什么机制来防止人们做坏事，它们就会被越广泛地使用。

The more safe these models are with whatever mechanisms are in place to prevent people from doing bad things, the more widely usable they'll become.

同意，顺便说一句，哈利，我不会猜是50%，但我觉得即使是10%或20%，在令牌数量上也会极其重要。

Agreed, and I wouldn't guess, by the way, 50%, Harry, but I think even 10 or 20% would be extremely significant in terms of tokens.

因为你想一想，OpenRadio上的基础模型，我觉得能获得20%到30%的令牌，却拿走了90%的收入。

Because if you think about it, the foundation models on OpenRadio I think get 20, 30% of tokens and get 90% of the dollars.

如果美国的开放权重模型在令牌数量上能达到那个水平的50%到60%，那已经是很大的进展了，因为那意味着更多以企业为中心、注重安全的客户在采用。

If the US open-weight models are even 50, 60% of that in terms of tokens, that's a lot of traction because it'll be by definition the more enterprise-centric, safety-conscious customers doing it.

我认为，坦率地说，这对Reflection来说将是一个巨大的胜利，而且正如你指出的，对所有推理服务提供商来说，也是真正的风险降低。

I think that would be, frankly, a huge triumph for Reflection, and as you point out, a real de-risking for all the inference guys.

叮，叮，叮。

Ding, ding, ding.

我们开投资委员会了。

We got investment committee.

戴夫，顺便说一句，欢迎加入委员会。

Dave, welcome to the committee, by the way.

我们非常高兴你能成为投资团队的一员。

We're thrilled to have you as part of the investment team.

你现在是奥德里斯科尔、莱姆金、斯特宾斯这家新公司的正式成员了。

You are sartorially on the new firm that is O'Driscoll, Lemkin, Stebbings.

杰森，我们有ElevenLabs，他们的估值翻倍到了220亿美元。

Jason, we have ElevenLabs who doubled their valuation to $22 billion.

我们要投220亿吗？

Would we do 22 billion?

我觉得在220亿这个估值，我们应该考虑投入基金最多10%的资金。

I think at 22 billion we should consider doing up to 10% of the fund.

时间不多了。

Time is short.

我们正处于18个月的投资部署周期。

We're at an 18-month deployment cycle.

我们最尊敬的合伙人之一要回去经营一家上市公司，所以团队里会失去一些人。

One of our most esteemed partners is going back to run a public company, so we're losing some of the folks on the team.

在我看来，ElevenLabs在竞争非常激烈的情况下，做了一件相当困难的事情。

The way I see it, ElevenLabs has done something quite difficult with actually a fair amount of intense competition.

它的领先优势还在不断扩大。

Its lead keeps widening.

它是代理应用最重要的十大基础之一，尤其是语音应用的发展方向。

It is one of the 10 most important underpinnings of agentic applications is the voice applications where they're going.

增长速度，而且令人惊讶的是，利润率也非常可观。

The rate of growth, and surprisingly, the margins are very impressive.

这家公司，等到我们完成朋友和家人轮融资时——我们希望以500亿的估值完成——它的收入可能接近10亿，这样的增长速度，我觉得这是我们必须要做的交易。

This is a company that by the time we close the friend and families round, which we're hoping to do at 50 billion, right, this one could be approaching a billion in revenue, and that's the sort of growth rate that I think it's the deal we've got to do.

而且它的差异化程度非常高。

And it's highly differentiated.

我不会说它是世界上护城河最深的公司，但风险绝对不像通用电气那么高。

I won't say they have the highest moat out there in the world, but this isn't ge-level risk.

ElevenLabs组建了一支非常出色的团队。

ElevenLabs has built an incredible team.

他们的语音响应速度低于毫秒级，这已经成为关键的模式，如今很难找到比这更好的资产了。

They have sub-millisecond response rates for voice, which has become a key modal, and it's tough to find a better asset today.

我是不是希望估值是210亿或180亿？

Do I wish it was at 21 or 18?

当然希望，但我觉得我们应该全力以赴。

For sure, but I say we go all in.

Lamkin先生，我能问您一个问题吗？您之前在节目里说过，今年会是语音模型替代的一年，而成本会推动这个决定。

Can I ask you, Mr. Lamkin, you previously said on the show that this would be the year of substitution for voice models and that cost would drive that decision.

您改变主意了吗？

Have you changed your mind?

不，因为我觉得ElevenLabs，大部分情况下，它总是有风险的。

No, because I think ElevenLabs is, for the most part, it's always a risk.

大部分情况下，它是有保障的。

For the most part, it is covered.

它有多款模型。

It has multiple models.

它覆盖了低端和高端市场。

It has covered the low end and the high end.

而且在这里，质量在很多方面比传统的开放权重模型更重要。

And quality here is in many ways more important than it is in your classic open-weights model.

对于非关键的工作流程，我们可以容忍推理上的一些松散。

We can tolerate a little bit of laxness in the reasoning for non-critical workflows.

但是当那个花店老板接起电话时，它必须得管用。

But when that flower dealer is picking up the phone, it has to work.

它必须在几秒钟内接起电话，而且必须选对花。

It has to answer that phone in seconds, and it has to be the right flowers.

如果它失败了，如果它理解错了，如果我们无法填充数据，那这个产品就不值得了。

If it fails, if it misunderstands it, if we can't populate, this product isn't worth it.

所以价格是有上限的，但ElevenLabs的交易规模让我震惊。

So there is a limit to price, but I'm shocked at the deal size of ElevenLabs.

让我震惊的是，有那么多规模相对较小的公司，每年心甘情愿地支付几十万美元，而且我们的背景调查显示他们非常满意。

I'm shocked at the number of folks that are relatively modest companies paying hundreds and hundreds of thousands of dollars a year happily, and our reference checks said they love it.

他们觉得自己得到了很好的交易。

They feel like they're getting a good deal.

所以我不认为所有模型都是可以互换的。

So I don't think all models are fungible.

尽管我们在基金里讨论过其他一些话题，我确实建议过可能投入5亿到Reflection和其他项目，但今天这个用例并不是可以互换的。

Despite some of the other discussions we've had at the fund, the fact that I did recommend putting maybe 500 million into Reflection and others, this is not a fungible use case today.

12个月后很难预测，但我对团队印象深刻，而且我认为今天没有人能追上他们。

12 months, it's hard to predict, but I'm impressed with the team, and I don't think anybody's going to catch them here today.

这很重要。

It matters.

这就是大语言模型发挥作用的地方。

This is where the LLM matters.

基金的10%投给了ElevenLabs。

10% of the fund goes to ElevenLabs.

哇。

Wow.

我只是不想分两轮来做。

I just don't want to do it over two rounds.

我不想等到估值到50或100亿再看。

I don't need to see it at 50 or 100.

Dev明白我的意思，对吧？

Dev knows what I mean, okay?

我不想再看第二次。

I don't need to see it a second time.

我现在就想把所有的筹码都押上，拿到能拿到的股权，然后继续前进，对吧？

I just want to put all the chips in now, get the ownership I can, and move on, right?

我有个朋友是做首席营收官的，我想让他当董事会顾问。

And I've got a friend of mine that's a CRO I want to put on as a board advisor.

我觉得他能帮上忙，给他们一些见解。

I think he'll be able to help out and give them some insight.

好，从一家红杉公司跳到另一家红杉公司。

Okay, moving from one Sequoia company to another Sequoia company.

抱歉，Dev，这不是故意的。

Sorry, Dev, it's not intentional.

但Rory曾说过一句话，好像是——那句话是什么来着？

But Rory once said a quote, which is like-- What was the quote?

“无论你在玩什么游戏，就做红杉。”

"Whatever game you're playing, just be Sequoia."

你记得我说的是哪句话吗？

Do you remember the one I'm talking about?

那句话太多了。

There's so many quotes.

你看，我干这行已经30年了，我刚入行的时候，红杉就是顶尖的，我得给他们大大的赞，30年过去了，他们依然站在顶端。

Look, I've been doing this 30 years, and when I came in, Sequoia were top, and give them huge credit, 30 years later, they're still there.

所以，是的，我引用一下加里·莱因克尔那句足球名言，如果你还记得的话。

So, yes, I give the Gary Lineker football quote, if you recollect.

所以，令人惊讶的是，红杉在Salesforce以20亿美元收购Listen Labs中赚了数亿美元。

So with great surprise, Sequoia made hundreds and hundreds of million dollars with Salesforce buying Listen Labs for $2 billion.

现在，显然，他们之前有15亿美元的报价。

Now, apparently, they had offers at $1.5 billion.

这对红杉的布莱恩·施雷尔来说是一次重大胜利。

A major win for Bryan Schreier at Sequoia.

投资9600万，回报给投资者约8.5亿美元。

About $850 million back to investors on 96 invested.

红杉做得最好。

Sequoia did the best.

他们领投了种子轮和A轮，赚了大约25倍。Ribbit领投了B轮，八个月内翻了四倍。

They led the seed, they led the A, making about a 25 X. Ribbit led the B, which 4X'd in eight months.

所以，干得好，Mickey。

So well done, Mickey.

各位，你们怎么看这件事？

What did we think of this, guys?

我觉得这是一个很好的结果，因为这是一个非常清晰的领域，很明显——退一步说，听我说。

I think it was a great outcome for a very clear category in the sense of it's pretty obvious that-- Stepping back, listen.

市场调研是一个非常大、分散、几乎还没有技术化的行业。

Market research is a very large, fragmented, almost pre-technical industry.

有一堆咨询类的公司，然后有三家上一代的公司，Qualtrics、Medallia，还有以前的SurveyMonkey，它们的市值都达到了几十亿美元，大概高个位数到低两位数的水平。

There's a bunch of consulting-type companies, and then there's three last-generation companies, Qualtrics, Medallia, and what was SurveyMonkey, all that got to kind of high single digit, low double digit billions of dollars in terms of market cap outcome.

所以这真的是一个很好的领域。

So it's a really good category.

这是一个已有的领域，广告投入很大，而大语言模型在这方面做得更好。

It's an existing category with a big ad spend, and LLMs just do it so much better.

因为你在市场调研中到底在做什么？

Because what are you doing in market research?

你向人们提问，然后汇总和分析他们的回答。

You're asking questions of people and aggregating and assimilating the answers.

而大语言模型擅长的两件事，我们刚才聊了语音。

And the two things LLMs do well, we just talked about voice.

语音作为一种交互方式，对大语言模型来说已经解决了。

Voice as a modality is a solved LLM problem.

还有提问和知道怎么问——不是那种老套的、死板的上一代调查问卷，不管你怎么回答，它都继续问下一个愚蠢的问题。

And then asking questions and knowing how to-- And not just these canned, stupid last-generation surveys where no matter what you say, the survey just asks the next dumb question.

比如问：“您住得满意吗？”

It's like, "Did you like your stay?"

你说：“不满意，我根本没入住。”

"No, I didn't even check in."

然后下一个问题还是：“床舒服吗？”

And then the next question is, "Was the bed comfy?"

这就是上一代的做法。

That's the last generation.

而现在有了大语言模型，你可以根据受访者的回答来调整问题。

Whereas now with LLMs, you can adjust the questions based on what the respondent is saying.

这意味着市场调研可以做得更好，我认为Listen Labs是早期做得好的竞争者之一，执行得非常出色，并且实现了真正的收入。

All that means market research can be done so much better with LLMs, and I think Listen Labs was one of the early contenders that did it well and executed brilliantly and got to real revenue.

结果非常好。

Great outcome.

这算是AI和大语言模型在应用层用例清单中排名前五或前十的用例之一，它就在那里。

It's kind of one of those, the list of use cases for AI and LLMs at the app layer, that comes out like a top five, top 10 use case, and there it is.

但你知道吗，我很想听听Dev对此的看法。

But you know what, I would love to get Dev's thoughts on this.

我非常喜欢Salesforce的很多收购，而且我一直是许多收购的支持者。

I'm pretty much a fan of a lot of Salesforce's acquisitions, and I've been champions of many one.

但这一笔我实在不太理解。

This one I don't really get.

倒不是说我觉得这是个坏主意，但Salesforce的市值已经接近600亿美元了。

It's not that I think it's a bad idea, but Salesforce is coming up on $60 billion.

他们的目标是600亿美元。

Their target's $60 billion.

他们差不多到500亿了。

They're almost at 50.

我不太明白，2000万美元的下一代AI调查收入怎么能改变大局。

I don't know how $20 million of revenue of next generation AI survey moves the needle.

我也不是说它不能。

And I'm not saying it doesn't.

规模要够大，要么就是小打小闹，要么就得有实质影响，对吧？

Mongo scale, it either has to tuck in or move the needle, right?

也许你有不同的看法，但要想有实质影响，规模得很大，Salesforce也不例外。

Maybe you think about it differently, but to move the needle, it's got to be big, and Salesforce even.

所以这个案例，我理解理论上的吸引力，但我在想，几年后它会不会就被遗忘了。

So this is the one where I get the theoretical appeal, but I wonder if this one will just be forgotten in a couple of years.

我想说的是，我们可能会看到更多这样的交易，对吧？

What I will say is that I think we're going to probably see more of these deals, right?

因为我觉得现在开始有区别了，你是在做一个伪装成公司的功能，还是在做一个真正的品牌，对吧？

Because I think there's starting to be a difference between are you building a feature that's masking as a company or a franchise, right?

而且我觉得很多AI应用公司会被收购，因为在一个别人的平台上做一个聪明的产品，说白了就是一个功能，如果他们能利用那个大公司的分销渠道，他们有资金和分销能力，他们就会利用这一点。

And I think a lot of AI application companies will sell because a clever product on top of someone else's platform is frankly a feature, and if they can use the distribution of that larger company, they have the cash and distribution, they're going to take advantage of it.

我觉得那些更持久的公司，是那些本质上在创建数据循环的公司，使用产生数据，而这些数据是别人没有的。

I think the companies that are more durable are the ones that essentially are creating a data loop where the usage creates data that no one else has.

数据让产品更好，产品吸引更多使用，你就创造了那个良性循环。

The data makes the product better, the product attracts more usage, and you kind of create that virtuous cycle.

所以如果有人想模仿你，他们可以模仿功能，但他们没有你所有的交互数据。

And so if someone tries to copy you, they can copy the features, but they don't have all the interactions that you have.

所以我觉得这里的区别是，当你看到这些独立的AI应用公司时，它们是真的在做一个伪装成公司的功能，还是真的在建立一个品牌？

And so I think the difference here is when you look at these standalone AI app companies is are they really a feature masking as a company or are they really building a franchise?

我觉得Listen Lab的创始人认为这个报价对他们来说是有意义的。

And I think the founders of Listen Lab decided that this was an offer that made sense for them.

如果你是Intercom，你肯定会很生气，对吧？

You'd be pretty pissed off if you're Intercom, wouldn't you?

抱歉。

Sorry.

你想想看，你辛辛苦苦打拼了十五年，经历了各种磨难，才把公司做起来，结果ABBA版的贾斯汀·比伯一出来，三年就卖了个差不多的价钱，你心里肯定想："靠。"

You're just like you've gone through 15 years of shit and struggle and building this business and then ABBA's answer to Justin Bieber comes along and three years later sells for a comparable number and you're like, "Fuck .

抱歉。"

Sorry."

所以我知道他们都赚了钱，但你肯定会想："哎呀，真是的，那条路容易多了，不是吗？"

So I know they all made money, but you'd be like, "Oh, bugger, that was an easier route, wasn't it?"

当然，作为创始人你肯定会这么想。

Of course, you think that as a founder.

总有人走的路更容易，你确实会去想这些。

There's always someone that had the easier route, and you do think about it.

当然，你会想——即使在Mongo，我打赌Dev偶尔也会想："天哪，那些家伙——我们在Mongo做得很好，但那些人，他们走的路容易多了。"

Of course, you think-- Even at Mongo, I bet Dev thinks about once in a while, "God, those-- We've done great at Mongo, but those guys , they had an easier way with it."

我常跟我的孩子们说，比较是快乐的小偷。

The lesson I tell my boys is comparison is the thief of joy.

你要是老拿自己和别人比，那你这辈子都会过得很痛苦。

You start comparing yourself to everyone else, and you're going to be very miserable for the rest of your life.

对，罗里。

Yeah, Rory.

嘿，我可没这么说，哈里。

Hey, I didn't do it, Harry.

罗里。

Rory.

对，比较就是快乐的窃贼，明白吗？

Yeah, comparison's the thief of joy, okay?

天哪。

God.

哈里，你这么说可不对，因为最后两家公司都赚了钱，而且就像德夫说的，你回头看看，他们都会说：“是的，我对自己做的事感到满意。”

Good to know, Harry, but that's why you were wrong to mention that because in the end, I think both companies made money, and to Dev's point, you'll step back and say both of them will go, "Yeah, I feel good about what I did."

你说得对。

Now, you're right.

三年内做到比十年或十五年容易多了。

It's a lot easier to do it in 3 years than 10 or 15.

这我们当然知道，但生活本来就不公平，风投的结果也很随机。

Duh, we get that, but life ain't fair and venture outcomes are very random.

更重要的是，这说明了时机和周期有多关键，因为你看，我们最近投了Intercom。

And more than anything, what it says is timing and the cycle is just so interesting because look, we put some money in Intercom recently.

Intercom达到了那个估值。

Intercom achieved that valuation.

如果他们在2018、2019年SaaS热潮时卖掉，Intercom本来也能达到这个估值。

It could have happened for Intercom had they sold in 2018, '19 in the SaaS boom.

而且就像Dev说的，也许五年后回头看，这些应用公司里可能有10%会早早卖掉，拿到30倍营收，剩下的那些会为了6倍营收、3亿估值、3000名员工而苦苦挣扎，头都要炸了，对吧？

And to Dev's point, maybe we will look back in five years from now, maybe 10% of these apps companies will have sold early, they'll have got their 30X revenue, and the rest of them will be grinding it out for six times revenue at 300 million with 3,000 employees and their head exploding, right?

但某种程度上，你无法控制结果。

But to some extent, you can't control the outcomes.

你只能享受过程，对吧？

You just got to like the process, right?

如果你最后做得好，那两者都没问题。

And if you end up doing well, then both of them have done just fine.

我想说一件事，Listen Labs 成立多久了？

I will say one thing from how old was Listen Labs.

它是什么时候成立的？

When was it founded?

三年了。

Three years old.

作为创始人，经历过不同长度的创业旅程，我想说，最初的日子总是最难的，对吧？

I will say as a founder having been through the different lengths of journeys, the earliest days are always the hardest, right?

那些日子也挺有趣的，因为只有你一个人，而且很有创造性，但除非你是那些幸运的年轻人，否则你几乎总是会资金耗尽。

They're sort of fun because it's just you and it's very creative in the way, but unless you're the lucky kids these days, you almost always run out of money.

你没有足够的客户。

You don't have enough customers.

那真的很残酷，对吧？

It's brutal, right?

但我要说的是，如果你能在三到四年内以一个好价格退出，我不是说你在财务上应该这么做，但天哪，那会容易得多。

But I will say if you can exit for a good price in three to four years, I'm not saying you should financially, but man, it's easier .

不管团队经历什么困难，退出都要容易得多。

It's so much easier to exit no matter what struggles you go through as a team, though.

因为我觉得那是五年一个阶段，真的会把人榨干，对吧？

Because I find it's five-year chunks, and they take it out of you, right?

我不是说这是卖公司的理由，但如果你能在头五年内以20亿美元卖掉，我会告诉他们接受。

I'm not saying it's a reason to sell, man, but if you can sell for $2 billion in the first five years, I'll tell them to take it .

因为不是我们不想打造一个像Mongo那样的公司。

Because it's not that we don't all want to build a Mongo.

我们理智上确实想，但天哪，就在第一波浪潮里做吧。

We do intellectually, but man, it's just do it in the first wave.

那会容易得多。

It's so much easier.

我觉得没有人会在电话会议上大喊：“我不敢相信你卖了。”

I don't think anyone is sitting on the call screaming, "I can't believe you sold.

真是的，各位。

Damn it, guys.

为什么只拿二十亿就走呢？

Why would you only take $2 billion?

别——不，不是那样……疯子一样——我们都经历过那种风投会议，公司情况就是这样，桌上没人说这种交易该卖。

Be a- No, that's not true ... maniac and- We've all had VC meetings where a company is like this, and no one around the table says you should sell on a deal like this.

今年我就遇到一次，价格差不多，每个风投都说“别卖”，就在第一波的时候。

I had one this year at about this price, and every single VC said, "Don't sell," right, in that first wave.

因为业务势头很猛。

And it's because there's a lot of momentum in the business.

等等等等。

Da, da, da.

但天哪，为了那个第一次退出……

But man, getting out for that first...

我知道你可以叫它第一次退出，但你其实还能再做一次。

I know you can call it the first exit, but you really can do another.

问题是，你得判断这是不是一代人的机会。

The thing is, you have to decide if it's generational.

你是在打造一个像Mongo那样甚至更好的东西吗？

Are you building a Mongo or better?

那当然，你会说不卖。对吧？

Then, of course, you say no. Right?

如果头五年就能拿到20亿美元退出，而且你做的不是Mongo那种级别的，那我还真不知道该怎么说了。

If you can get out for $2 billion in the first five years and you're not building Mongo or better, I don't know, man.

我会接受，然后在Salesforce享受生活。

I would take it and enjoy my life at Salesforce.

这故事现在都老掉牙了，但不知怎么的，最近又在推特上翻出来了，就是扎克伯格当年想都没想就拒绝了十亿美元的收购，好像是雅虎还是谷歌，我记不清了，反正是很早的时候。

It's such an old story now, but for some reason, it surfaced on Twitter, the whole Zuckerberg turning down a billion from, was it Yahoo or Google, I can't remember right early on, without even thinking about it twice.

但事实就是这样。

But it is.

我觉得真正的重点，Jason，就像你说的，你得有点冷血。

I think the real point, Jason, it comes to your thing is you have to be kind of cold-blooded.

从现在开始，最可能的发展轨迹是什么？

What's the likely trajectory from here?

然后你怎么看待继续经营下去，以及你有多大可能不只是拿到三倍的金额，而是十倍甚至更多，你又是怎么评估的？

And then how do you feel about running the tape, and how likely are you to get not just three times this amount, but 10X this amount and more, and how do you assess that?

再加上个人方面的因素。

Plus the personal side.

我觉得如果你想打造一家世代传承的公司，你就必须对自己诚实。

I think you have to be honest if you're trying to build a generational company.

我不认为那只是一句口号。

I don't think that's a catchphrase.

我觉得这真的很重要，因为否则的话，你有一个终局价值，一个终局状态，如果报价合理，你永远无法超过这些交易的净现值。

I think it's really important, because otherwise, you have a terminal value, a terminal state, and if the offer is reasonable, you'll never beat the net present value of these deals.

打造世代传承的公司并不容易。

This generational stuff isn't easy.

Dev又回到了那个位置。

Dev's back in the seat again.

这不容易。

It ain't easy.

这事儿不容易，打造一家世代传承的公司这事儿。

This stuff ain't easy, this generational company stuff.

所以，我跟你说，最好的公司总是会收到收购邀约，对吧？

So, and I will tell you, the best companies always get offers to sell, right?

所以就拿我自己举例吧，我不说Blade Logic是个特别突出的例子，但我们在成长的每个阶段，都有人想收购我们。

So I'll use my example, and I wouldn't say Blade Logic was a massive outlier, but every stage of our growth, we had people who wanted to buy us.

我记得有一次，我现在可以说这个了，我当时坐在约翰·钱伯斯的办公室里。

I remember sitting, I guess I can share this now, I was in John Chambers' office.

我们真的是面对面坐着，他想收购我们公司，但后来不知怎么的，这事儿就没成。

We were literally sitting face to face where he wanted to buy the company, and then for some reason, it just didn't work out.

然后他们又回来了。

Then they came back again.

EMC也找过我，你就这么一直走下去，最后，我们在上市一年后把公司卖了。

I had EMC at the table, and you just go and go, and then ultimately, you sold a year after we went public.

而且我觉得，就像杰森说的，最终创始人得自己决定，是继续把公司做大做强，还是干脆按下退出键。

And I think to Jason's point, ultimately, the founder has to decide, do I feel good about continuing to build, or should I just hit the exit button?

我对把公司卖给BMC的决定感到满意，但之前有好几次差点就把公司卖掉了，现在回头看，当时的选择是对的。

And I felt good about our decision to sell to BMC, but we came really close a number of times to sell the business earlier, and looking back, it was the right decision.

你卖给BMC卖了多少钱？

How much did you sell for to BMC?

我们卖了大约9亿美元，这在当时是2008年最高的收购价。

We sold for about $900 million, which at that time, was the highest acquisition paid in 2008.

那真是一笔大买卖。

It was a killer deal.

我有一家小公司跟他竞争。

I had a small company that competed with him.

他把我们打得落花流水。

He cleaned our clock.

Blade Logic在那个领域是赢家。

Blade Logic was the winner in that category.

其实，第二名是Aref，他现在在贝恩资本。

Actually, and the number two was Aref, who is now at Bain Capital.

他当时在管CenterView和CenterPoint。

He was running CenterView, CenterPoint.

CenterOne，对。

CenterOne, yeah.

CenterOne。

CenterOne.

你可是把我们打得落花流水啊，Damien。

You've walked all over us, Damien.

你们卖掉的时候营收是多少？

What revenue were you doing when you sold?

我们大概做到一亿，所以是九倍。

We were doing about 100, so 9X.

顺便说一句，那正好是全球金融危机之后，2008年市场已经相当紧张了。

Which by the way, was right on the heels of the GFC, so the market was getting quite skittish in 2008.

我感兴趣的一个是Jason提到的，就是Vercel，年经常性收入达到6亿美元，其中代理业务占新业务的50%，而年初时只有3%。

One that I am interested by is actually one that Jason suggested, which is Vercel at $600 million in ARR, with agents being 50% of the new business, up from 3% at the start of the year.

Jason，每次你聊这些话题，我都觉得特别开阔思路。

Jason, I always find you quite mind-expanding for me when you talk about these topics.

你怎么看这个？到2027年，我们应该给代理分配多少预算？

How do you think about this, how much budget we should give to agents for 2027?

Vercel现在新业务中50%来自代理，而之前只有3%，这对我们有什么启示？

What's the takeaway for us from Vercel now getting 50% of new business from agents, up from 3%?

嗯，我觉得可以从两个不同的角度来看。

Well, look, I would say maybe two different things.

第一，我认为像Vercel这样的公司，在AI和代理出现之前，就已经有了合适的基础设施和产品，对吧？

One, I think that there's a bunch of companies like Vercel that had the right infrastructure and the right product before AI and agents, right?

Vercel是2020年成立的，对吧？

So Vercel founded in 2020, right?

Guillermo在2026年之前肯定没想到，他的大部分业务会来自代理自主创建应用，对吧？

Guillermo did not know that in 2026, the majority of his business would be from agents spinning up applications on their own, right?

这谁也没想到。

Nobody saw that.

有些最惊人的故事，连ElevenLabs也有点类似。

Some of the most amazing stories, even ElevenLabs is a little bit of that.

当然，Replit成立于2016年，是个很大的案例，你只要抓住那个点，它就能运转起来，而且他们确实抓住了老虎的尾巴。

Certainly Replit, founded in 2016, is a big case study where you just latch onto that and it just works, and they do have the tiger by the tail.

竞争对手很多，而且都很强，但智能体选择了Vercel。

There's a lot of competitors, a lot of really strong competitors, but agents pick Vercel.

这非常厉害，而且很有意思。

It's very strong, and it was interesting.

我想，他们是怎么说的来着？

I think, what did they say?

他们在6亿美元的基础上增长了170%，而客户只增加了20%左右。

They're growing 170% at $600 million from 20% increase in customers or something.

扩展收入简直惊人，关键是，当智能体选择你的时候——我相信Dev在数据库方面也在想这个问题——那现在就是一股自然的力量。

The expansion revenue is astronomical, and it's just the thing is when agents pick you, and I'm sure Dev's thinking a lot about this on the database side, when agents pick you, it's a force of nature right now.

说到智能体，我学到的是，你和我一天到晚都在构建21个智能体。

And agents, what I've learned from, well, you have 21 agents we're building all day.

我和Emil之间，我每天真的要花八到十个小时来构建。

Between me and Emil, I'm literally building eight to 10 hours a day.

天哪，我的智能体们都有自己的想法，而且很难反驳它们。

And man, my agents have opinions, and they're really hard to argue with.

如果它们想用某个供应商，那真是费劲。

And if they want to use a vendor, it's a lot of energy.

我知道我是在把它们拟人化，但当智能体做出决定时，那力量很强大，而Vercel就是其中的大受益者。

I know I'm anthropomorphizing it, but it's powerful when an agent decides something, and Vercel is a major beneficiary of this.

即使去年看起来不错，但当时似乎不会像今天这么重要。

Even last year, it looked great, but it never seemed it would be as important as it is today.

哦，那只是我放代码的地方，对吧？

Oh, that's just where I'm going to stick my code, right?

重要，但属于后台基础设施。

Important, but back-office infrastructure.

但说真的，现在很多智能体和应用都在这里运行，我也会把这个推荐给IC。

But man, this is where so many agents and apps are running, and I would take this one to the IC, too.

我觉得Jason提到的这一点很有意思，就是智能体在做选择或做决定，对吧？

I think this is really interesting about the point that Jason made is that agents are picking or making decisions, right?

过去25年，每个营销人员、每家公司都在优化如何出现在谷歌第一页，对吧？

For the last 25 years, every marketing person, every company has optimized how you show up on page one of Google, right?

你是在教育人类买家。

You educate human buyers.

现在你得确保智能体能找到你，理解你卖什么，并且在他们阅读的各个来源上，关于你的信息是最新、准确的。

Now you have to make sure that agents can find you, understand what you sell, have current, accurate information about you across the sources they read.

如果你在AI生成的答案中缺席，就像在谷歌上隐形一样。

And if you're missing from an AI-generated answer, it's like being invisible on Google.

坦白说，红杉有一家叫Profound的公司，专门帮企业解决这些问题。

Full disclosure, Sequoia has a company called Profound that helps companies with those, helping them.

但我告诉你，这是个巨大的问题，而且不只是科技行业面临。

But I will tell you, this is a massive problem, and it's not just in tech.

如果你是电商网站，或者你是想通过电商卖货的零售公司，而现在代理在替消费者做购买决定，那你就得想办法去教育这些代理。

If you're an e-commerce site or you're a retail company trying to sell through e-commerce, and agents are now making the buying decisions, you got to figure out how to educate those agents.

不然的话，你的生意就完了。

Otherwise, your business is going to zero.

我看这个市场的时候，我们也有家公司，Dave，叫Peak，大家都说，"这不过是Semrush的下一代产品。"

When I look at this market, we too have a company, Dave, Peak, and everyone's like, "Well, this is just like a next generation of Semrush."

这真的只是SEO分析工具的下一代，比如你们的Profound和Peak，还是说这其实是一个更大的品类？

Is it just a next generation of SEO analytics providers with your Profounds and your Peaks, or is this actually a much larger category?

我觉得机会不只是分析，告诉你哪里缺失之类的。

I think the opportunity is more than just analytics by telling you where you're missing and all that.

我觉得这些公司的机会不只是告诉你缺什么，而是帮你塑造内容、创建内容，帮你解决那个问题。

I think the opportunity for these companies is not just tell you what you're missing, but to help shape the content and create the content to help you solve that problem.

如果他们能闭环，那就成了一个非常有吸引力的解决方案，而不是只告诉你，"你在这里。

If they can close the loop, that becomes a very compelling solution versus just telling you, "You're here.

你不在那里。"

You're not here.

这就是其他人在做的事情。

This is what the other guys are doing."

我今天就在看，比如规模小一点的，看看Resend。

I'm just looking today, for example, it's smaller, but look at Resend.

比如说，我们第一次通过智能体采购的产品就是Resend的邮件服务。

So for example, the first agentic purchase we made was Resend for email.

好的，去年我试着搭建一堆智能体的时候，一直有人推荐SendGrid，但因为需要CEO或创始人去处理，我就是搞不定那个该死的SendGrid。

Okay, when I was trying to set up a bunch of agents last year, I kept getting recommended SendGrid, but because they need the CEO, the founders to go back, I just couldn't get the damn SendGrid thing to work.

他们好像把免费版给停用了。

They'd kind of deprecated Free.

我搞不清楚那个密钥。

I couldn't figure out the key.

它老是出问题。

It kept breaking.

所以我就问我的智能体该用什么。

And so I asked my agent what I should use.

它说用Resend。

It said use Resend.

Resend从四月份的10.6万次MCP调用增长到九月份的300万次。

Resend April 106,000 MCP calls to September three million.

这就是智能体在推荐产品。

That's agents saying to use products.

我并不是说Profound没在做这件事。

And I'm not saying Profound isn't doing that.

他们可能确实在做。

They may well be.

根据我有限的体验，这次的情况不一样。

My limited experience is this is different.

我想稍微反驳一下你们。

I want to kind of disagree with you all.

你的智能体是怎么决定的？

How did your agent decide?

因为我觉得有一部分是取决于我在答案引擎里最初的表现，还有我在OpenAI上的表现，然后你的代理还会看什么？

Because I assume some part of it is the general how I show up initially in the answer engines, how I show up on OpenAI, and then what else does your agent look at?

是不是只看易用性——

Is it just the ease-

它看的是你的基础设施。

It looks at your infrastructure.

它看你正在构建的应用程序。

It looks at the application you're building.

它看什么是可信的。

It looks what's trusted.

它看它认为什么是最合适的匹配。

It looks what it thinks the right match is.

而且偶尔，它会做出一些非常针对你具体用例的古怪选择。

And once in a while, it'll make wacky choices that are very specific to your use case.

但你必须成为那个特例。

But you got to be that one.

ChatGPT会基于一些事实给你一个通用的答案。

A ChatGPT is going to give you a generic answer based on a number of facts.

我不是说他们用的是同一个大语言模型，对吧？

I'm not saying they're using the same LLMs, right?

所以有一部分是这方面的因素。

So some of that is being into it.

但智能体了解你的整个技术栈。

But an agent knows your whole stack.

它了解你的应用程序。

It knows your application.

它知道你在系统之上构建了什么。

It knows what you've built on top of the system.

所以它会给出非常具体的推荐，告诉你该用什么。

So it will make a very specific recommendation what to use.

而且这个问题对早期创业公司来说更加严重，没有双关的意思，因为你的竞争对手已经存在了10年、15年，有大量的数据积累，而如果你试图颠覆某个竞争对手，智能体会看到关于竞争对手的所有这些数据，而关于你的数据却很少。

And this problem is even more profound, no pun intended, for early-stage companies because there's so much corpus of data for your incumbent who's been around for 10, 15 years, and if you're trying to disrupt some incumbent, the agent's going to see all this data about the incumbent and very little about you.

所以当它做决定的时候，它很容易就会直接选择现有的解决方案。

So when it makes a decision, it could very easy just basically pick the incumbent solution.

所以实际上，初创公司有责任想办法让智能体了解他们做什么，包括文档、API，以及所有包装层的基础设施，这样才能展示出他们确实为智能体做决策提供了一流的体验。

So it's actually incumbent upon the startup to figure out how they basically educate agents on what they do, the documentation, the APIs, and all the infrastructure in the wrapper so that they can then show that they are truly a first-class experience for an agent to make a decision on.

顺便说一句，这是我对我投资或合作的投资组合公司说的。

For what it's worth, here's what I tell my portfolio companies I've invested or work with.

我会说：“找10个你认识的、已经在生产环境中使用智能体做不同事情的人，10个你信任的人，然后每两周问他们：‘在这个类别里你会推荐什么？’

I say, "Find 10 folks you know that have agents in production doing different things, 10 folks you trust, and ask them every two weeks, 'What would you recommend in this category?

“你会推荐什么？”

What would you recommend?'

然后看看智能体推荐什么。”

And see what the agents recommend."

这个产品现在还不存在。

This product doesn't exist today.

我觉得我们四个人应该投资这个。

I think the four of us should fund it.

这和Profound不一样，Profound是在猜测大语言模型会怎么做。

It's not the same as Profound, which is guessing what LLMs do.

这些是在不同应用场景中实际运行的智能体，直接问它们应该用什么，对吧？

These are live agents in the field in different applications, asking the agent what should be used, right?

让我理解一下，那个回答问题的智能体在哪里？

And just to let me understand, where is that agent who will answer the question?

我的意思是，那个智能体大概在做别的工作，然后你就直接去问它——

I mean, that agent presumably is doing some other job, and you'll just go on and say-

这个智能体可以在任何科技公司。

The agent could be at any tech company.

任何有实力的初创公司现在都会有智能体在运行。

Any startup is going to have agents running today that's worth their salt.

是的，你会让你的智能体透露它们内部在用什么。

Yes, and you will allow your agents spew out what they're using internally.

对，比如我现在就在打电话。

Yeah, for example, literally I'm on the phone right now.

我正在问我正在构建的一个应用，SaaS Recruit，它的智能体：“托管应该用什么？”

I'm asking an app I'm building, SaaS Recruit, its agent, "What should I use for hosting?"

它说：“首先，用Render。”

It said, "First, use Render."

说实话，我对Render几乎不了解。

Actually, honestly, I barely know Render.

好的，它说：“用Render，我的选择。”

Okay, it says, "Use Render, my pick."

那个智能体说：“我的选择。”

The agent said, "My pick."

第二，Railway。

Number two, Railway.

我们知道Railway。

We know Railway.

他们做得相当不错。

They're doing pretty well.

第三名，Fly.io。

Number three, Fly.io.

第四名，AIOs。

Number four, AIOs.

第五名，Vercel，不太适合这个。

Number five, Vercel, not suitable for this.

哦，有意思。

Oh, interesting.

我本来会选Vercel的。

I would've picked Vercel for this.

你们可能知道Render。

You guys may know Render.

我从来没听说过，你知道吗？

I never heard of it, and you know what?

我大概会用Render。

I would probably use Render .

它明确写了我的选择。

And it literally says my pick.

所以 Render 暂时领先了。

So Render just got a lead.

从这个 AI 助手那里，它拿到了相当不错的优势。

It just got a pretty good lead from this agent.

你得赢下这场竞争。

You got to win that war.

我们能不能聊聊 Render 的最后一轮？

Can we get into the last round in Render?

Render 那轮融资的估值是多少？

Can we get the Sequoia price at Render?

Render 那轮具体估值多少？

What was the Sequoia price in Render?

有人知道吗？

Anybody know?

我看了，但我不确定。

I did, but I don't know.

我觉得这跟Profound和Peak的市场不太一样。

I don't think it's actually the same market as Profound and Peak.

我其实觉得是一样的。

I actually do.

我认为是一样的，因为我知道这两家公司都给Profound A轮写了投资意向书，而且我们拿到了。

I think it is because I know both companies actually wrote a term sheet on the Profound A, which we got it.

而且我觉得这两家公司，在搞清楚怎么让人类在ChatGPT上找到自己之后，自然会很快转向：如果代理是我需要展示的对象，那么天哪，我得告诉我的客户怎么让代理找到他们。

And I think naturally both those companies, when they're finished figuring out how you show up on ChatGPT for humans, are going to quickly evolve to if agents are who I need to show up for, then by God, I need to tell my customers how to show up for agents.

所以我觉得这是很自然的产品延伸。

So I think it's a natural next product extension.

别太纠结于那些细节，比如“哦，我们只针对人类”。

Don't get kind of caught up in the minutia of, "Oh, we only do it for humans."

从宏观来看，就像Jason说服我们的那样，每个人，无论是人类还是代理，都会用大语言模型和网上的信息来做购买决策，所以作为供应商，你必须确保自己以正确的方式出现，不管需要付出什么代价。

Zooming out, as Jason has convinced us, everyone, humans and agents, will make purchasing decisions using LLMs and using information they get online, and you're going to have to make sure as a vendor that you show up correctly and whatever that takes.

这里的机会就是向那些需要解决这个问题的公司营销人员推销解决方案。

The opportunity there is to sell a solution to the marketing person who's got to solve that problem in any company.

MongoDB 那边肯定有个人，今天下午就得琢磨怎么在智能代理面前展示自己，确保当 Jason 需要数据库的时候，我们排在第一位。

There's some dude at MongoDB who is going to be figuring out this afternoon, check how we show up on agents to make sure when Jason needs to get a database, we show up first.

如果那个人需要一种方式来了解别人怎么说他们，其他智能代理怎么说他们，那 MongoDB 就会买这样的产品。

And if that person needs to have some way of surveying what other people are saying about them, what other agents are saying about them, that's a product that MongoDB is going to buy.

我觉得这就是 Profound 的机会，也是 Peak 的机会。

And I think that's the Profound opportunity, it's the Peak opportunity.

这个领域还有很多其他公司。

It's a bunch of other companies in the space.

我觉得这个领域不错，而且会比 Semrush 那个领域更有深度，Semrush 那个领域最后变得商品化，很无聊，原因有很多。

I think it's a good space, and I think it's going to have more depth than the Semrush space, which ended up commodified and boring for a whole bunch of reasons.

我同意。

I agree.

提醒一下，我们在 HubSpot 做了那个系列，我记得是 C 轮，我在这边说过。

I mean, as a reminder, we did the series, I think C at HubSpot, and I say this over here.

它一开始是一家SEO公司，但他们在上面加了很多其他东西，做得非常出色。

It was an SEO company to start out, but they just did a brilliant job of adding a whole bunch of other stuff on top.

你解决了客户眼前的痛点，尤其是这些营销平台。

You solve the customer's immediate pain point, especially these marketing platforms.

你解决了客户眼前的痛点，然后他们发现还有一个相邻的痛点，你又解决了，接着再解决下一个。

You solve the customer's immediate pain point, and then they realize they have an adjacent pain point, and you solve that, and then you solve the one after that.

我觉得每个营销人员都有一个三年的痛点清单，需要解决如何在新的大语言模型世界中展示自己，他们希望得到尽可能多的帮助。

And I think every marketing person has a three-year list of pain that they have to solve in terms of how they show up in the new world of LLMs, and they want all the help they can get for that.

如果你能提供解决方案，就能在这里赚钱。

If you make solutions, you can make money here.

所以我认为这是一个非常好的市场。

So I think it's a super good market.

兄弟们，我还有什么话题没讨论到，需要补充的吗？

What topic have I missed, boys, from the list that I have to discuss?

好的，有一个话题我很喜欢，因为它可能会改变很多技术方面的事情。

Okay, there was one I liked, but I like the one because it could change a lot of tech stuff.

工程师起诉的那个，是哪个案子来着？

The engineers suing, which one was it?

嗯，是Groq。

Well, Groq.

哪些是许可协议？

Which are the license deals?

我之前没想到这一点，但现在看来，我因为这么做而失去了考虑利益，这似乎是个很明显的问题。

I didn't think about that, but now it seems like an obvious claim that I lost the benefit of my consideration by doing this.

我不知道律师们会怎么看这个案子，但没错，给我们讲讲这个案子的背景吧。

I don't know where the lawyers come out on this one, but yeah, give us the background on this one.

Groq的工程师起诉Nvidia的170亿美元许可和招聘协议。

Groq's engineers sue over Nvidia's $17 billion license and hire deal.

两名前工程师在特拉华州提起诉讼，声称Groq的董事会实际上通过一项110亿美元的许可协议，加上约30亿美元的股票，将其核心技术和前200名员工卖给了Nvidia，让普通股东只剩下一个空壳公司。

Two former engineers sued in Delaware, allege Groq's board effectively sold its core technology and top 200 staff to Nvidia through an 11 billion license, plus about three billion in stock, leaving common holders with a hollowed-out company.

新闻说他们认为可能有法律途径，因为从宏观角度看，公司法试图确保每个人都受到平等对待，而且有200年的先例，你知道事情会怎么发展。

The news is that they perceive that there might be a course of action, because zooming out a million miles, corporate law tries to say everybody gets treated the same, and there's 200 years of precedent, and you know exactly how it happens.

所有普通股东都受到同样的对待。

All common shareholders get treated the same.

董事会成员对所有股东负有注意义务和信托责任。

Board members have a duty of care, a fiduciary duty to all shareholders.

判例法多得数不清，必须遵循。

Case law up the wazoo has to happen.

而这件事把所有的判例法都打乱了，实际上让买家能够重新分配公司的对价。

And this stuff tears up all the case law, and it effectively allows the buyer to reallocate consideration in the company.

“嘿，你们过来，能拿到很多钱。”

"Hey, you guys are coming across and you get a lot of money.

“你们留在后面，只能拿到一点钱。”

You guys are getting stayed behind, and you get a little money."

关键是，这两名员工已经离开了公司。

And crucially, these two employees had already left the company.

我不确定，但我的猜测是，当收购方分配资金时，他们不会说，‘哦，我还想照顾一下前员工’，因为他们会想，‘我为什么要费这个劲？’

And I don't know this for sure, but my guess is when the acquirer is allocating capital, they don't say, "Oh, I also want to take care of the ex-employees," because they're like, "Why would I bother?"

但特拉华州法律规定，所有普通股股东都是一样的，这是这些原则中不可妥协的一条，就是你必须以同样的方式对待持有同一证券的每一个人。

But Delaware law says everyone who owns a common shareholder is exactly the same, and it's one of the non-negotiable principles of this stuff, that you treat everyone in the same security in the same way.

所以我读到那个的时候，我就想，嗯，这挺有意思的。

So when I read that, I thought, "Hmm, that's interesting.

这是个挺巧妙的案子。

That's a clever little case."

而且它还会继续发酵一阵子。

And it'll run a little.

如果它最后和解了，我也不会意外，因为可能这个集体里人不多，我猜所有过来的人都挺满意的。

I wouldn't be surprised to see it settled, because there's probably only a small number of people in the class, because my guess is everyone who came across was happy.

我猜那些留下来拿到股票的人也都挺满意的。

My guess is everyone who stayed and got stock is happy.

所以可能只有少数已经离开的人会有意见。

So it could be a relatively small number of people who'd left already.

但这个案子会很有意思，他们可能确实有一些法律上的论点可以提。

But it's going to be an interesting case where they probably have some legal argument to make.

对，我来把这些点串起来，对吧？

Yeah, I'll connect dots, right?

记得我们一开始讨论的是员工离职，而且是非常非常快地离职。

Remember we started this discussion about people leaving and leaving very, very quickly.

我在这里看到的相似之处是，投资者很久以前，甚至最近，已经学会了在各种情况下保护自己。

And here's the parallel I draw is, investors long ago, and actually even more recently, have learned to protect themselves in situations.

他们会在事前就谈判好这些问题，比如在任何退出情况下会发生什么，包括一些奇怪的情况。

They negotiate these kind of issues up front, what happens in any kind of exit, including a weird one.

员工从来没想到要问这个。

Employees never thought to ask about this.

所以如果你是创始人，你会听到员工说：“等等，如果公司的技术和部分团队成员被收购，但公司本身没有被收购，我的股权怎么办？”

So if you're a founder, you're going to have employees saying, "Hold on, what happens to my equity if the company's tech and some members of the team get bought, but the company doesn't?"

我觉得创始人需要能够回答这个问题，因为随着人们了解这个案例，这个问题会越来越频繁地出现，尤其是如果员工经常跳槽的话。

And I think founders need to be able to answer that question, because that's a question of, as people learn about this, is going to be coming up more and more often, especially if people switch between companies very often.

而且公司法的理念，无论是特拉华州还是像德克萨斯州这样的地方，员工都不应该需要担心这个，因为就像，哥们儿，有普通法在，你们所有人都会得到同样的对待。

And the idea of corporate law, Delaware, or even if you go to a place like Texas, you shouldn't have to think about it as an employee, because it's like, dude, there's common law and you'll all be taken care of the same.

但一旦这种保障被削弱，就像你说的，一旦你知道这事没准了，那就很麻烦了。

But once that gets eroded, as you say there, once that you know it's up for grabs, then it's quite insidious.

你说得对，每个人加入公司时，不光要考虑我的1%值多少钱。

And you're right, everyone has to think, when I join a company, I don't just have to think what my 1% is worth.

我还得想，如果你卖了知识产权，我的1%会怎么样？

I have to think, "Well, what could happen to my 1% if you sell the IP?

如果你卖了知识产权和一半员工，我的1%又会怎么样？

What could happen to my 1% if you sell the IP and half the employees?"

我觉得这可能会让所有事情都变得更难搞定。

And I think it probably makes everything harder to pull off.

所以看看这是否会导致人们抵制这类结构，会很有意思。

So it'll be interesting to see if this leads to some kind of pushback away from those kind of structures.

它们并不理想，人们之所以这么做，只是因为联邦贸易委员会。

They're not ideal, and the only reason people are doing them is because of the FTC.

没人会明说，因为他们不能说，但就是因为联邦贸易委员会，并购在合理时间内根本批不下来，而这些交易几乎一夜之间就能完成，员工第二天就能上班。

No one will ever say this because they can't, but it's because of the FTC and the fact that you can't get M&A through in any appreciable time, whereas you can get these deals done literally overnight and the employees start next day.