20VC：水平代理与垂直代理之争：谁将胜出 | 模型路由：成为大宗商品还是价值提升？ | 推理市场的未来 | 为什么创始人应接受顶级机构的折扣——专访 Harrison Chase（LangChain）

那就是数据、记忆、特定的API和工具。

And that would be data, that would be memory, that would be specific APIs, specific tools that you can use.

控制框架有点难，但你仍然可以把很多领域知识放在里面，比如路由和其他的东西。

Controlling the harness is a little bit harder, but you can still put a lot of your domain knowledge around what should do both routing and things in there.

而模型是最难控制的，但潜力也最大。

And then the model is the hardest to control, but potentially the highest ceiling.

让你能在这三者上迭代的关键是评估和追踪。

The thing that lets you iterate on each of those is evals and traces.

所以这些非常有价值，但只有当你把从中获得的洞察带回上下文、框架或模型时，它们才有价值。

So there is a ton of value in these, but there's only value in that you capture insights that you bring back to the context, or bring back to the harness, or bring back to the model.

这个过程可以各不相同。

And that exact process can look different.

所以，后训练非常适合把追踪中的洞察带回模型。

So post-training, great for bringing these insights from traces back to the model.

要把东西带回上下文或记忆，你可能需要在那些追踪上运行另一个代理，然后让那个代理输出学习内容到记忆里。

For bringing things back to the context or memory, you're probably going to want to run another agent over those traces and then have that agent output learnings for memory.

然后，要把这些从追踪中学到的东西带回给控制系统，你很可能需要在这些追踪上运行一个代理，然后更新控制系统本身的代码。

And then for bringing these learnings in the traces back to the harness, you're probably going to want to run an agent over those traces and update the code of the harness itself.

而且我们确实有东西能帮人们做到这一点。

And we actually have something that does this for people.

它叫Engine。

It's called Engine.

它是LangSmith里的一个代理，负责运行追踪并尝试更新代码。

It's an agent in LangSmith that runs over traces and tries to update code.

然后，当你更新这些东西时，你需要知道它确实变得更好了，这时候评估就派上用场了。

And then when you update these things, you're going to need to know that it's actually better, and this is where evals come in handy.

我能问一下，持续学习是如何改变代理及其工作方式的吗？

Can I ask, how does continual learning change agents and how they work?

持续学习，我不认为它有多大程度上改变代理的工作方式，但我认为它改变了你希望在代理周围运行的系统。

Continual learning, I don't think so much changes how agents work, but I think it changes the systems that you want to have running around the agent.

我的意思是，我认为持续学习代理的核心循环，也许可以说是核心内循环，在很大程度上仍然是一样的。

So what I mean by that is, I think the core loop for a continual learning agent, the core inner loop maybe to say, will still largely be the same.

它会是一个大语言模型在循环中运行，调用工具，与模型和环境互动。

It will be an LLM running in a loop, calling tools, interacting with the model, and an environment.

我认为持续学习增加的是一个外层循环。

I think what continual learning adds is an outer loop.

这个外层循环会查看智能体的输出，然后利用这些输出做些什么。

An outer loop that looks at the exhaust of agents and then does things with it.

同样，具体做什么可能取决于你对持续学习的定义。

And again, what it does could be different depending on what continual learning means for you.

所以我认为持续学习也是一个没有非常精确定义的词，或者可能——

So I think continual learning is also one of those words that doesn't have a super precise definition, or maybe-

你会怎么定义它？

How would you define it?

我会把它定义为一个随着运行而不断改进的系统。

I would define it as a system that gets better over time as it runs.

同样，它可以通过不同的方式随时间改进。

And again, there's different ways it could get better over time.

它可以更新上下文，也可以更新框架的代码，或者更新模型的权重。

It could update the context, it could update the code of the harness, or it could update the weights of the model.

而且外层循环对于所有这些情况来说，看起来都大致相同。

And the outer loop will look roughly the same for all of those.

它看起来就像：获取轨迹，得到经验，修复问题，测量以确保有所改进，然后重复。

It will look like take traces, get learnings, fix something, measure it to make sure it improved, and then repeat.

但具体的变化，再次强调，更新模型权重与更新一个像记忆文件那样的文本文件，是非常不同的。

But that exact changing, again, updating weights of model, very different from updating a text file that is like a memory file.

如果我们看上下文层，我们提到数据是其中的一部分。

If we look at the context layer, we mentioned data being a part of that.

我们看到数据公司以有史以来最快的速度将收入规模扩大到数亿美元。

We're seeing data companies scale to hundreds of millions of revenue in the fastest time ever.

前所未有。

Ever.

而且每个人都在卖数据。

And everyone is selling data.

你觉得这个市场会怎么发展，既然你对它保持中立，不参与其中，但又跟它沾边？

How do you expect this market to play out, given your neutrality on not being involved in it, but being adjacent to it?

未来几年这个市场会不会出现整合？

Will this be a market of consolidation in years to come?

你觉得它会怎么演变？

How do you see it play out?

我其实觉得数据市场在整体格局中还处于早期阶段，因为我们说的是智能体的上下文层。

I actually think the data market is still pretty early in the grand scheme of things, insomuch as we're talking about a context layer for agents.

我的意思是，现在有很多数据库，ClickHouse 可能就是一个很好的例子，最近表现非常出色。

So what I mean by that is there are a bunch of databases that are, ClickHouse is maybe a really good example of one that's doing incredibly well recently.

我觉得他们主要还是专注于自己的数据库，那种传统的老式数据库。

I think they're still largely focused on their database, their good old-fashioned database.

他们擅长存储数据和查询数据，但方式还是比较传统的。

They're good at storing data and querying data, but it's still in pretty traditional ways.

ClickHouse 在人工智能出现之前就存在了。

ClickHouse existed pre-AI.

当我们谈到为智能体提供上下文层时，再说一遍，会有大量来自软件系统的数据产生，而且，我认为现有的数据库，比如ClickHouse、Mongo、Elastic等，非常适合处理这些数据。

When we talk about a context layer for agents, and again, there will be a bunch of data produced from software systems, and great, I think existing databases, ClickHouse, Mongo, Elastic, whatever, are great for that.

当我们谈到为智能体提供上下文层时，我认为关键的一部分是，是的，有一个数据库，但实际上我认为数据库周围还有一个系统。

When we talk about a context layer for agents, I think a key part of what exists there is yes, there's a database, but I actually think there's a system around the database.

我认为为智能体提供的上下文层，看起来更像是一个围绕数据库的智能体，而不仅仅是一个数据库。

I think a context layer for agents will look more similar to an agent around a database than just a database.

原因在于，要把所有这些学习成果吸收到数据库中，你需要某种智能来完成这件事。

And the reason for this is that taking all of these learnings and ingesting them into a database, you need some sort of intelligence to do that.

而且我认为这种智能实际上非常非常难。

And I think that intelligence is actually really, really hard.

记忆中最难的部分根本不是数据库。

The hardest part about memory is actually not the database at all.

当我们与许多公司交流时，他们实现记忆的方式就是把它存储在一个单一的文件中。

When we talk to a number of companies, the way that they do memory is they just store it in a single file.

非常简单。

Very simple.

难点在于把发生的所有观察，无论是用户与软件交互，比如传统的SaaS，还是用户与代理交互，把这些观察提炼成足够有价值的信息，才能进入这个上下文层。

The hard part is taking all of the observations that happen, whether it's from the user interacting with the, let's say it's a traditional SaaS, interacting with the software, or the user interacting with the agent, and distilling that into information that's valuable enough to go into this context layer.

这就是目前记忆方面的难点。

That's the hard part in memory right now.

为什么这么难？

Why is that so hard?

因为要让LLM按照你的意图去做事，这本身就很棘手。

Because it's tricky to get LLMs to do what you want them to do.

而这种元反思，我觉得LLM其实还很不擅长，比如反思提示词并改进提示词，反思代理架构并改进代理架构。

And this type of meta-reflection, I think is something that LLMs actually still struggle at a lot, reflecting on prompts and improving prompts, reflecting on agent architecture, improving agent architecture.

我们应该为此在内部建立一个基准。

We should build a benchmark for this internally.

我们确实有一些这方面的基准，比如Engine，就是我们的代理，它正在运行一些东西。

We do have some benchmarks for this in terms of Engine, which is our agent that's running on things.

但我认为这种元反思对代理来说仍然非常困难。

But I think this type of meta-reflection is still really difficult for agents.

我觉得这对人类来说通常也是一件难事。

I think it's also generally a hard thing for humans to do as well.

我不确定我们是否真的理解，有些人可能理解，所以我不该这么说，但我不确定我们是否理解大脑内部发生的事情。

I don't know if we actually understand, some people probably do, so I shouldn't say this, but I don't know if we understand what goes on inside our brain.

我觉得我们的大脑仍然很复杂，我们还没有完全理解。

I feel like our brains are still complex things that we don't fully understand.

所以如果我们试图在系统中复制这种学习过程，我认为有很多不同的部分。

And so if we're trying to now replicate this kind of learning process in a system, I think there's a bunch of different pieces.

所以我认为模型在反思方面并不擅长，而且我也不确定系统应该是什么样子。

And so I think models aren't great at reflection, and I don't think we know what the system should look like.

输入应该是什么样的？

What should the inputs look be?

输出应该是什么样的？

What should the outputs be?

这个代理的框架应该是什么样的？

What should the harness for this agent be?

不同的工作流程应该是什么样的？

What should the different workflows be?

12个月后，模型在反思能力上会有显著提升吗？

Will the models be meaningfully better at reflection in 12 months?

可能会吧。

Probably.

看起来它们会在各方面都变得更好。

It seems like they'll get better at everything.

你觉得模型改进的速度会继续保持像现在这样的加速趋势吗？

Do you think the speed of model improvement will continue in the same manner of acceleration that it has done?

还是说我们正在接近某种平台期？

Or do you think we are approaching some form of asymptotic plateauing?

我觉得，如果不深入实验室内部，很难准确判断模型改进的速度。

I think the speed of model improvement is something that's probably really tough to say without being inside of a lab.

我想说的是，对于很多任务来说，模型的改进可能会趋于平稳，因为已经足够好了。

I would say, I think there are a lot of tasks for which the model improvements will flatline in the sense that it's already good enough.

我觉得有很多任务，模型已经做得足够好了。

I think there are a bunch of tasks where the models are just good enough.

比如编码。

Which would be like coding.

对，编码已经接近那个水平了。

Yeah, coding is getting there.

像总结、路由任务这些。

Like summarization, routing tasks.

可能还有更难的任务，或者我们甚至没想到要交给模型的任务。

There are probably more difficult tasks or tasks that we don't even know to think about, know to throw a model at.

我可能会说，这种元学习就是模型现在还不擅长的任务之一。

I would probably say this type of meta learning is one of those tasks that the models aren't good at now.

我想这对实验室来说非常重要，关系到破解RSI之类的事情。

I would imagine is very important to the labs in terms of kind of cracking RSI or things like that.

所以我认为模型会在这方面逐渐进步，而且这种进步的速度可能会比之前看起来快得多。

And so I do think the models will get better at that over time, and it's possible that that progress actually looks much faster than previous.

我觉得如果没有在实验室内部工作过，真的很难有有根据的看法。

I think it's really tough to have an educated opinion without being inside of a lab.

有些任务已经饱和了。

There are some tasks where it's saturated.

其他一些任务，我觉得进展可能会比以前更快。

There are other tasks where I think it will probably just increase faster than before.

你觉得现在听起来很疯狂的智能体，在两三年后会变得非常普遍吗？

What about agents today sounds crazy that you think will be incredibly common in two to three years?

我觉得智能体的记忆层仍然是个大问题，远未解决。

I think this memory layer for agents is still massively unsolved.

我认为智能体现在很少见，但未来几年会非常普遍的一个特点就是记忆层。

I think one thing about agents that is rare today but will be incredibly common in the next few years is this memory layer.

我的意思是，如果你看看Instinct或Muse，我觉得他们做得好的地方之一就是这种个性化和学习，让它看起来像是为你量身定制的，无论是语气还是主动跟进，那种主动性。

So what I mean by that is if you look at Instinct or Muse, I think one of the things that they've nailed is this kind of personalization and learning and seeming like it's something for you, whether it's because it's speaking in the tone or because it's just following up, like that proactiveness.

所以我觉得记忆层还处于非常早期的阶段。

And so I think memory is super, super early.

我们两年来一直在跟人聊记忆这个话题，也尝试在开源社区里做些事情。

We've been talking with folks about memory for two years and trying to do things in the open source.

我们推出了一个测试版，后来又砍掉了一个商业化的记忆解决方案。

We launched an alpha and then killed a commercial solution for memory.

记忆这块还太早了，而且非常依赖具体场景。

It's just so early on in memory, and it's so domain-specific.

对Instinct或Muse来说，记忆的样子跟企业或组织工具里的记忆很不一样，跟法律、金融这些领域的记忆也完全不同。

What memory looks like for Instinct or Muse is actually pretty different than what it probably looks like for a company or org harness, and it's pretty different than what it looks like for legal or finance or things like that.

所以我觉得记忆真的是一个能带来巨大差异的东西。

So I think memory is one thing that really makes a difference.

我觉得Instinct和Muse在这方面做得很好，你能看到它在那里带来的不同。

I think Instinct and Muse have done a great job at this, and so you see how it makes a difference there.

但我觉得这还处于非常早期的阶段，我预计几年后所有产品都会带有某种记忆功能。

But I think it's still very early on, and I would expect all products in a few years to have some aspect of memory in it.

恕我直言，这不是很明显吗？

With respect, is that not quite obvious?

我觉得记忆会产生影响这一点是显而易见的，所以过去几年你会看到很多人对它感兴趣，而且一直保持着兴趣。

I think the fact that memory will have an impact is quite obvious, and that's why you see so many people interested in it and having it be interested in it for the past few years.

但我还是觉得答案并不明显。

I still think it's non-obvious what the answer is.

我还是觉得答案极其不明显。

I still think it's incredibly non-obvious what the answer is.

我觉得没人知道记忆问题的答案是什么。

I don't think anyone knows what the answer for memory is.

既然它这么明显，而且几乎所有东西都需要它，那你为什么还要砍掉那个商业解决方案呢？

Why did you kill the commercial solution for it, given the obviousness of it and the clear need for it in almost everything?

我们砍掉它是因为没有找到产品市场契合点。

We killed it because we didn't find product market fit.

为什么没找到呢？

Why didn't we find product market fit?

我觉得我们其实是想做得太泛了。

I think we actually tried to be too general.

我觉得记忆这个领域还太早了，所以很多你想构建的解决方案其实都是非常特定于领域或特定于代理的。

I think memory is so early that I think a lot of the solutions that you want to build are actually very domain or agent-specific.

所以你会构建垂直特定的记忆解决方案。

And so you would build vertically specific memory solutions.

那你会怎么做不同的事情？

What would you do differently?

所以再说回记忆的难点，我认为记忆的难点在于那个从事实中学习该记住什么的过程，以及那个执行记忆的代理。

So again, going back to what the hard part of memory is, I think the hard part of memory is that process for learning from facts about what to remember, and that agent that does that remembering.

就像我说的，我非常看好垂直代理这类东西。

As I said, I'm massively bullish in kind of vertical agents.

我认为让这些代理擅长学习它们领域特有的东西是关键的一部分。

I think having those agents be good at learning things specific to their domain is a key part of that.

所以我的确认为，至少一开始，这类学习代理可能必须是特定于领域的。

And so I do think these kind of learning agents, at least to start, probably have to be domain-specific.

而且我不知道是否存在一个通用的，比如这些学习代理的框架是什么样的？

And I don't know if there's a general, like what does the harness for these learning agents look like?

它可能看起来和我们现有的很多框架差不多。

It probably looks pretty similar to a lot of the harnesses we have now.

它可能看起来挺像的——我们构建的是深度智能体。

It probably looks pretty-- We build deep agents.

深度智能体就是一种智能体框架。

Deep agents is an agent harness.

可能看起来和那个差不多，由大语言模型驱动循环，但带有特定的指令和工具，也许还有一些专门的子智能体，擅长导航现有的记忆结构或提炼事实。

Probably looks similar to that, in LLM running the loop, but it's got specific instructions and specific tools, and maybe there are some specific sub-agents that are really good at navigating existing memory structures or distilling kind of facts.

所以我认为会有专门负责记忆的智能体。

And so I think you'll have agents for remembering things.

我只是觉得它们会是特定领域的。

I just think they'll be domain-specific.

你觉得在消费类应用上会有垂直智能体吗？

Do you think you will have vertical agents for consumer applications?

我理解并同意你关于金融、会计、销售、营销和法律领域的看法。

I understand and I agree with you when you look at finance, accounting, sales, marketing, legal.

你提到了Muse和Instinct，它们会是横向智能体。

You mentioned Muse and Instinct, they would be horizontal agents.

你觉得我们会有纵向的消费者智能体吗？

Do you think we'll have vertical consumer agents?

我觉得在消费领域，我的看法是两者都会有。

I think my take on the consumer space is that you'll have both.

横向和纵向智能体都会有，而且横向智能体会和纵向智能体互动。

You'll have both horizontal and vertical agents, and the horizontal agents will interact with the vertical agents.

在什么情况下适合用纵向智能体，什么情况下不适合？

Where does it make sense to have a vertical agent versus not?

我觉得对大多数应用来说，可能用纵向智能体更合适。

I think for most apps, it probably makes sense to have a vertical agent.

我认为纵向智能体之所以有意义，原因大概有两个。

And I think the reason that having vertical agents makes sense at all is probably twofold.

可能不止两个，但首先想到的就是这两个。

Maybe more, but two that jump to mind.

第一，很多这类应用都了解你，并且能比通用型横向智能体更好地学习你与那个应用的互动方式。

One, a lot of these apps have knowledge about you and can learn things about how you interact with that app better than a general horizontal agent could.

所以回到领域特定性的问题，比如说，你在Spotify上听音乐的方式和你在Strava上的互动方式是非常不同的。

So back to the point about domain specificity, how you interact with music for, say, Spotify is very different than how you might interact with Strava.

因此，这些应用实际上可能最擅长了解你。

And so those apps are actually probably best at learning things about you.

而且我认为，将这些了解到的信息呈现出来的最佳方式可能就是通过一个智能体。

And I think the best way to expose those learnings is probably through an agent.

另一个原因是，很多这类应用都有大量的API。

The other thing is a lot of these apps have a ton of APIs.

所以你可以说，'嘿，Spotify，别构建智能体了，直接开放一堆API吧。'

And so you could say, "Hey, Spotify, don't build an agent, just expose a bunch of APIs."

我实际上认为，如果有一个智能体来封装这些API，并提供一个单一的，或者可能是多个，但很可能是一个单一的自然语言界面，让横向智能体与那个纵向智能体交互，那会是一种更好的体验。

I actually think it's a better experience if there is an agent that kind of wraps those APIs and exposes a single, or maybe multiple, but probably a single natural language interface for a horizontal agent to interact with that vertical agent.

当智能体成为与网站交互的基本界面时，基础设施会如何变化？

How does the infrastructure change when agents are the fundamental interface with sites?

我的意思是，我想买某场演唱会的票，我就说：“每五分钟查一次，有票就买”，等等。

And what I mean by that is, I wanted tickets to a specific concert, and I said, "Check every five minutes and buy any seat that comes available," da da da.

结果它连续三天，每五分钟就查一次。

And for three days, it checked every single five minutes.

查询量是人工操作时的一千倍。

The volume of queries was 1,000 times what it would have been with humans.

在智能体交互而非人类交互的世界里，所需的基础设施会发生怎样的变化？

How does the infrastructure required change in a world of agent interaction, not human interaction?

当智能体进行交互时，请求数量会大幅增加。

When agents are interacting, it makes the number of requests go way up.

我想你看到过一则新闻，好像是Resy封了一个用户的账号，因为那个用户用Instinct去频繁拉取数据，具体频率我不清楚，但确实很频繁。

I think you saw a story about how, I think it was Resy that blocked a user's account because they were using Instinct to pull it, I don't know how frequently, but pretty frequently.

所以我觉得基础设施承受的压力会大得多。

And so I think there's just way more strain on the infrastructure.

这种压力会如何传导到更底层的架构呢？

How does that percolate down to lower layers?

这个我还没有形成看法。

I don't know if I have an opinion on that yet.

我觉得你得构建那种能应对突发流量、扩展性更强的系统。

I think you have to build systems that can be much more bursty, that can scale way broader.

除此之外，我也说不出更具体的了。

I don't know if I have more specifics other than that.

你觉得供应商会允许代理使用吗？

Do you think providers will allow agent usage or not?

我觉得未来供应商会推出自己的代理，让人们要么直接跟那个代理互动，要么通过他们的横向代理跟那个代理互动。

I think the future is providers exposing their own agent and having people either interact with that agent or having people's horizontal agents interact with that agent.

所以我的横向代理，抱歉，我确认一下，会跟Resy的纵向代理互动。

So my horizontal agent, sorry, just so I understand this, would interact with Resy's vertical agent.

我觉得完全正确。

I think that's exactly right.

那为什么这样能行得通？

Why would that work?

因为那还是解决不了很多这类企业的两个核心广告问题。

Because that still doesn't solve the two core problems of advertising for a lot of these businesses.

他们会失去广告模式，然后还有追加销售。

They lose their ads model and then upsell.

他们还会失去那些“你可能也喜欢”的推荐，所以购物篮的规模就变小了。

They lose their, anything else you might like also on the side, so you get reduced basket size.

为什么代理对代理的模式会行得通呢？

Why would agent-to-agent work?

我觉得你可以把其中一些东西包含在代理的回复里。

I think you could include some of those things in the agent response.

所以你可以针对Resy或Etsy等平台上的特定商品做广告，并把它们包含在代理的回复中。

So you could do advertising for specific things on Resy or Etsy or whatever and include that in the agent's response.

比如，“嘿，我找到了这三样东西。

Like, "Hey, I found these three things.

你可能还想考虑一下这个。”

You might also want to consider this other thing."

这现在是一个自然语言回复，会返回给主代理。

That's now a natural language response that goes back to the main agent.

我还想提一点关于系统压力的，如果你有一个垂直代理，它和底层系统交互会比第三方代理更好、更高效。

I think also to the point around the strain on the systems, if you have a vertical agent, it's going to be better and more efficient at interacting with the underlying systems than a third-party agent will.

所以我觉得这样会让它们做这件事更高效。

So I think it will make it a lot more efficient for them to do that.

当我们看代理带来的查询量和数据量变化时，做所有这些工作会消耗很多令牌。

When we look at the amount of queries and the volume that changes with agents, it takes a lot of tokens to do all of this work.

令牌成本和智能成本随时间会怎么变化？

How does the cost of tokens and cost of that intelligence change over time?

因为我们一直以为它会变得更便宜。

Because we always thought it would get cheaper.

但并没有便宜多少。

Hasn't got much cheaper.

我们到底该怎么看待这个问题？

How do we actually think about that?

成本是我们现在跟人聊得最多的一个大问题。

Cost is one of the big things that we talk with people about right now.

我觉得这个问题有几个不同的方面。

I think there's a few different dimensions to this.

第一，前沿模型的价格随着时间推移在逐渐下降。

One, I think the frontier models are relatively getting cheaper over time.

Astra 的性价比，跟一年前的很多产品相比，简直太棒了。

Astra is price per performance, fantastic compared to a lot of things that existed a year ago.

第二，需求量大增。

Two, there's just way more demand.

当你开始谈代理之间互相交流或者多代理系统时，消耗的 token 数量多得惊人，所以成本自然就上去了。

When you start talking about agents talking to agents or multi-agent systems, there are just so many more tokens that are getting burned, and so the costs are going up.

这一点在编程领域尤其明显。

And we see this in coding primarily.

编程成本飙升，我觉得很多公司，包括我们自己在内，现在都在想办法削减这些开支。

Coding costs have skyrocketed, and I think a lot of organizations, including ours, are now thinking about ways to pare those costs back.

那这个问题最后是怎么解决的呢？

How does this end up getting resolved?

如果你看看我们是怎么处理编程代理的，我觉得这里面有一些线索。

If you look at how we're tackling coding agents, I think there's a few hints there.

我们在处理编程代理的时候，基本上分三步走。

So when we're tackling coding agents, we think of it in three steps.

第一步，主要是让成本透明化，让管理员能看到每件事大概要花多少钱，已经花了多少钱。

One, we think of basically just visibility, showing how much things will cost, how much things have cost to admins.

第二步是成本控制，比如你可以说：“嘿，帮我订个位子，但别花超过五美元。”

Two is cost control, so you could say something like, "Hey, go get me a reservation, but don't spend more than $5 doing it."

然后第三步，就是真正去影响整个系统本身。

And then third is actually impacting the harness itself.

所以我觉得，目前很多系统层面的工程工作，主要是为了让代理更高效，因为这是最大的瓶颈。哦，抱歉，其实很多系统层面的工程工作，是为了让代理表现更好、更准确，因为这才是最大的障碍。

So I think primarily a lot of harness engineering has been done around making the agent more efficient because that's the-- Or, sorry, a lot of the harness engineering has been done to make the agent just better, more accurate, because that's the biggest blocker.

我觉得随着成本越来越受关注，我们看到大家开始更重视像提示缓存这样的东西，这对降低成本特别有效。

I think as cost becomes more a concern, we're seeing people care more about things like prompt caching, which are really good for reducing costs.

我们看到模型提供商实际上让这变得更容易了，虽然这有点深入细节，但他们引入了可以在运行中更改代理可用工具的功能。

We're seeing model providers actually make this easier by, and this is getting a little bit in the weeds, but they introduce things where you can change the tools that are available to an agent mid-flight.

以前，这会破坏缓存并导致成本大幅增加。

Previously, this would have busted the cache and caused a massive increase in cost.

现在你可以动态添加这些工具，同时保持成本不变。

Now you can just add those dynamically and keep the cost the same.

所以我认为，在控制成本方面，人们越来越关注模型API和框架。

So I think we're seeing more attention to both the model APIs and harnesses in terms of controlling costs.

当你今天与大型企业交谈时，什么对他们最重要？

When you talk to large enterprises today, what matters most to them?

是准确性吗？

Is it accuracy?

是成本，还是安全性？

Is it cost, or is it security?

顺便说一句，另一个因素是延迟。

Another thing would be latency, by the way.

所以我们每年都会做开发者调查，问他们最关心什么。

So we run developer surveys every year where we ask developers what they care about.

过去两年，排名第一的答案一直是准确性，而且它的票数几乎是第二和第三名加起来的两倍。

For the last two years, the number one response, basically double the next two, has been accuracy.

所以准确性仍然是首要的。

So that's still the primary thing.

接下来是延迟和成本，安全性稍微靠后一些。

The next two are latency and cost, and security is a little bit after that.

其实都挺接近的。

Pretty close.

最近跟大型企业交流时，我觉得成本方面的担忧上升了。

When we talk to large enterprises recently, I'd say cost has risen in terms of concern.

不过我还是觉得它排在准确性后面，但可能比延迟更受关注。至于安全性，对大企业来说，安全性很重要，但可能还是排在准确性和成本之后。

I'd still probably say it's below accuracy, but it's probably higher than latency and then security is, for large enterprises, security is higher, but still probably below accuracy and cost.

你大概把工程薪资的百分之多少花在编程工具上？不用特别精确。

What percent of your engineering salary, give or take, not specifically, but do you spend on coding tools?

我觉得我们大概把工程薪资的30%花在了编码工具上。

I think we spend probably roughly 30% of our engineering salary on coding tools.

实际上，最近两个月这个比例下降了，而我们使用的token数量却上升了。

It's actually gone down the last two months, while tokens, the number of tokens that we've used, has actually gone up.

我们通过三个步骤做到了这一点。

And so we've done this by three steps.

第一步基本上就是提高可见性。

First is basically just visibility.

我们把所有编码代理的使用情况都记录到LangSmith上，这样就能知道谁花了多少，花在了什么地方。

We tracked all of our coding agent usage to LangSmith, so we could tell who was spending how much, what they were spending it on.

我们可以对此进行分析。

We could do analytics over that.

第二步就是基本的成本控制。

Second is just basic cost controls.

我们设定了预算，如果达到预算，你可以和我们的工程副总裁谈谈，他几乎总是会批准增加预算。

We set a budget, and if you hit the budget, you can talk to our VP of engineering, and he'll almost always increase it.

但这只是为了防止代理被错误使用或陷入不良循环。

But it's just a safeguard against agents being used incorrectly or going in a bad loop.

然后第三步是我们控制运行环境。

And then third would be we control the harness.

所以我们做的一件事就是自己开发了一个内部的云端编码代理。

So one of the things we've done is we actually built our own in-house cloud coding agent.

不是IDE，也不是终端代理。

So not an IDE, not a terminal-based agent.

那部分我们让大家随便用自己喜欢的。

We let people use whatever they want for that.

但我们自己建了一个基于云的代理，我们叫它OpenSui。

But we built our own cloud-based agent, OpenSui is what we call it.

其实现在我只用这种方式编码，因为我一直在Slack上，所以我就从Slack给它发消息，它就开始运行。

It's actually the only way I code anymore because I'm in Slack all the time, and so I just message it from Slack, and it kicks off.

但在那个环境里，我们做了很多工作来控制成本。

But in that harness, we've done a bunch of work to control costs.

比如提示缓存，还有路由这些，我们都集成到这个工具里了。

So things like prompt caching, but also things like routing we've built into this harness.

而且我觉得，要控制编程代理的开销，这三步流程是挺通用的。

And I actually think that's a pretty repeatable three-step process for getting coding agent spend under control.

第一步，就是要有可见性。

Just first, visibility.

第二步，基本控制。

Second, basic controls.

第三步，尽量改进工具本身。

Third, try to improve the harness itself.

完全理解。

Totally get that.

我想再深入聊聊这个。

And I do just want to unpack that.

你提到了可见性，就是看谁花了多少钱。

You mentioned that visibility, see who's spending what.

那些花钱最多的人，是不是也是最有效率的人？

Were the people spending the most, the most effective?

有时候花钱最多的人确实最有效率，但有时候也不是。

Sometimes the people spending most are the most effective, and other times they're not.

我的意思是，我们有几个资深工程师，他们花费很多，但我百分之百相信他们做的事情非常值得。

So what I mean by that is we have a few senior engineers who spend a lot, and I 100% trust that what they're doing is well worth it.

这一点毫无疑问。

No questions there.

但也有一些情况，更多是因为一些愚蠢的错误或者设置不当造成的。

We've also just had some cases due more to silly mistakes or improper setup than anything.

我们遇到过使用量的大幅飙升。

We've had massive spikes in usage.

举个例子，有一两个任务中，某个插件基本上会在代理的每一轮操作中，对网页的DOM进行截图或快照，然后发送回来，因为他们当时在做前端工程，想看看页面长什么样，然后把它发回来。

So an example of this, we had one or two tasks where some plugin was being used that basically would screenshot or take a snapshot of the DOM of a webpage and send that back every turn of the agent because they were doing some front-end engineering, and they wanted to take what it looked like and just send it back.

它就会一直运行这个庞大的循环。

It would just run this massive loop.

所以我觉得就那一件事，我们一周就花了四万。

And so I think we spent 40K in a week on that one thing.

真到那种程度的时候，我觉得我们公司算是挺宽松的，也信任大家，但我知道那肯定不值。

When it gets to that, again, I think we're pretty liberal as a company and trust people, but I know that's probably not worth it.

所以这只是一个例子，再说一次，这不是恶意的。

And so that's just an example of, again, it wasn't malicious.

其实也不是故意这么用的。

It wasn't even intended usage, really.

就是不小心犯了个错。

It was kind of just a mistake.

所以我觉得能发现这个问题就很值得。

And so I just think just catching that is well worth it.

所以这要看情况，得看是在哪个极端。

So it depends a bit on the side of the spectrum.

就像你提到的，工资的30%。

30% of salary, as you mentioned.

那是个大概的估计，是的。

That's a guesstimate, yeah.

大概的估计。

A guesstimate.

完全没问题。

Totally fine.

那两年后会是多少？

What is that in two years?

我觉得估算编程代理花费的难点在于，组织变化太快了。

I think the hard part about estimating coding agent spend is that orgs are just changing so rapidly.

我的意思是，未来可能会有组织规模小得多，但编程代理的花费跟人类工资一样，甚至更多，或者高出一个数量级。

So what I mean by that is, I think there's a future where orgs are much smaller and coding agent spend is the same as human salary, or more, or an order of magnitude more.

而且只有少数人在驱动和执行，速度极快，效率极高。

And you have a small number of people that are driving and executing incredibly rapidly and incredibly performantly.

我觉得从思考的角度来看，这是一个很吸引人的愿景。

And I think that's an enticing vision in terms of what you think about.

我觉得很多公司，包括我们，可能都还没到那一步。

I don't think many companies, including us, are probably there.

我觉得我们还是需要很多不同背景、不同专长的人一起协作，而且还有很多沟通、开销和审查的工作。

I think we still see that we need a bunch of different people with a bunch of different backgrounds and a bunch of different expertises collaborating on things, and there's a bunch of communication and overhead and review.

是的，理想情况下，最好是一个人驱动100个代理，然后驱动产品的大部分内容。

And yes, it would be ideal to have one person driving 100 agents and just driving everything that makes up a large part of the product.

但我觉得我们还是看到，这里面仍然有很多人的因素。

But I think we just see that it's still a very human component to it.

所以我花了很多时间思考组织设计。

And so I spend a lot of time thinking about org design.

有没有可能让小型团队，单人团队，双人团队，来驱动代理？

Is it possible to have small teams, one-person teams, two-person teams, that are kind of driving agents?

我觉得这很有吸引力，在这种情况下，我会说代理的花费可能是成本的3倍……

I think it's very tempting, and in that situation, I would say agent spend maybe 3x the cost of...

但另一方面，能够做到这一点的工程师非常宝贵，所以也许实际上花费会保持不变，因为能够做到这一点并驱动所有这些代理的人更少，但你付给他们的工资更高，所以总花费其实是一样的。

But then also, the engineers who can do that are incredibly valuable, and so maybe it actually will just stay the same because the people who can actually do that and who can drive all those agents, there's fewer people, but you're paying them a lot more, and so the spend actually stays the same.

所以我觉得现在组织设计非常有意思，正因为如此，我觉得这个问题其实很难回答。

So I think org design is very interesting right now, and so because of that, I think it's actually really hard to answer that question.

对我来说，这个数字太关键了，因为它是决定Anthropic企业价值的根本因素。

To me, that number is so central because it is the fundamental determinant of Anthropic's enterprise value.

如果这个数字降到5%，那2万亿美元的IPO可能就有点高估了。

If that number goes to 5%, then the $2 trillion IPO might be a little bit overpriced.

如果这个数字真的像McCall的情况那样达到300%，那它就严重低估了，所以我这么关注它。

If that number actually goes to 300% like it is in McCall's case, then it's dramatically underpriced, and so that's why I focus so much on it.

你同意这个数字是Anthropic价值的基础吗？

Do you agree that number underpins Anthropic's value?

我觉得这个数字确实是Anthropic价值的基础，这可能是对的。

I think it's probably right that that number underpins Anthropic's value.

我觉得棘手的是，到目前为止，我们只在软件工程领域看到了智能体的真正影响，而企业里还有很多其他领域我认为还没有受到智能体的影响。

I think the tricky part is that the only place we've seen agents really have effect is in software engineering so far, though, and there are a ton of other places within an enterprise that I would argue have yet to be affected by agents.

所以我觉得这是有可能的。

And so I think it's possible.

再说软件工程，我觉得很难预测，但基本上有两个关键因素。

Again, software engineering, I think very difficult to predict, but I think there's basically two levers here.

一个是软件工程本身会怎么发展。

One is what happens in software engineering.

会和现在一样吗？

Is it the same as today?

会更多吗？

Is it more than?

还是会更少？

Is it less than?

但我认为毫无疑问，在其他领域，肯定会比现在多得多。

But I think undoubtedly, for the other domains, it will be a lot more than it is today.

所以从这个角度看，我认为这对Anthropic、OpenAI以及其他模型提供商来说是非常有利的。

And so I think from that perspective, I think that's very bullish on Anthropic and OpenAI and other model providers.

我觉得很多人还在问另一个问题，就是开源模型的普及会如何改变这个格局。

I think another question that a lot of people are asking is about the proliferation of open models and how that changes the landscape.

你合作过一些非常大的企业。

You work with some very large enterprises.

我的一位好朋友，SaaStr的Jason Lemkin，说我不打算模仿他的口音，因为我会搞砸，而且我以前试过，他有点——他说我学得不太好。

One of my dear friends, Jason Lemkin from SaaStr, says that I'm not going to do his accent because I'll butcher it, and I've done it before, and he got rather-- He said it wasn't very good.

你觉得大型企业会害怕使用中国的开源模型吗，还是说其实他们并不介意？

Do you find that large enterprises are fearful of working with open source Chinese models, or actually, they don't mind?

我们看到有些企业确实害怕使用中国的开源模型。

We see that some enterprises are fearful of working with open source Chinese models.

我们也看到有些企业并不在意。

We see others that don't care.

你觉得这种恐惧有道理吗？

Do you think that fear is justified?

我真的不觉得这种恐惧有道理。

I don't really think that fear is justified.

我觉得有不同的担忧。

I think there's different concerns.

我觉得一个担忧是，它运行的数据会不会以某种方式传回中国的实验室。

I think one concern is that the data that it runs on will somehow go back to the Chinese labs.

再说，如果你用的是Fireworks或者自己托管，我觉得这种担忧完全没道理。

And again, if you're using Fireworks or hosting it yourself, I think that's just ridiculous.

我觉得还有其他担忧，比如它们训练用的数据可能会导致模型以不符合企业价值观的方式回应，诸如此类。

I think there's other fears that the stuff that they've been trained on will cause it to respond in a way that's not consistent with the values of the enterprise or something like that.

我觉得这时候评估就派上用场了，而且这种担忧很大程度上是没有根据的。

I think that's where evals come in handy, and I think that's largely unfounded.

不过当然，在某些领域，可能确实有一些真实的问题。

But sure, maybe for some domains, there's some realness there.

所以我觉得这很大程度上是没有根据的。

So I think it's probably largely unfounded.

但在某些领域，因为那种回应方式，我确实能看到一些细微差别。

But for some domains, because of that kind of response, I could see there being some nuance there.

你觉得未来12个月内，我们会有可媲美的美国开源模型吗？

Do you think in the next 12 months we'll have comparable US open source models?

我认为在接下来的12个月里，我们绝对会有可媲美的美国开源模型。

I think in the next 12 months, we will absolutely have comparable US open source models.

你看看英伟达在生态系统合作方面所做的所有努力。

I think you look at everything that NVIDIA is doing to work with the ecosystem.

我觉得他们非常重视这一点。

I think they care about it tremendously.

你看看Reflection最近发布了一个模型。

You look at Reflection had a model launch recently.

我觉得它在某些基准上还是稍微落后于一些中国模型，但我认为这个领域受到了极大的关注，所以它们绝对会赶上来。

I think it was still a little bit behind some of the Chinese models on some benchmarks, but I think there is a tremendous amount of attention in that domain, and so I think absolutely those will catch up.

当你今天看开源模型的格局时，以你的位置，有什么是别人看不到但你却看得很清楚的吗？

When you look at the state of landscape today in terms of open models, what does the world not see that you see very clearly given your position?

我认为开源模型的使用确实很多，但我认为它们主要用于比人们想象的更狭窄的任务。

I think when it comes to open models, I think there is a lot of usage of them, but I do think they're for more narrow tasks than people think.

所以我觉得我们看到很多人使用开源模型和更多后训练模型，这比我一年或一年半前预期的要多。

So I think we see a lot of people using more open source models and more post-trained models than I would have maybe expected a year or a year and a half ago.

但我还是认为它们主要用于那些比较简单的任务，比如路由、防护、提取或总结。

But I still think they're primarily for these simpler tasks, like routing or guardrails or extraction or summarization.

我觉得它们并没有真正驱动很多主要智能体的核心循环。

I don't think they're really driving the core agentic loops for a lot of these main agents.

所以我觉得开源模型有点两极分化，因为有些人会说：“太好了，它们已经足够好了。”

And so I think open source models are kind of divisive because you have some people on one end who are like, "Great, they're good enough.

“我们就用它们做所有事情吧。”

Let's use them for everything."

但我认为它们还没到那个程度。

And I don't think they're quite there.

然后还有另一些人会说：“前沿模型会不断变得更好、更便宜。”

And then I think you have other people who are like, "The frontier models are just going to keep on getting better and better and cheaper and cheaper."

是的，这没错，但如果你真的清楚自己要做什么，而且有高容量的使用场景，并且其中有些部分的任务足够简单，那么我认为对这些用例进行后训练模型是非常划算的投资。

And yes, that's true, but if you really know what you're going to do, so you have to have a high-volume use case, and if there are a few components of that that are simple enough tasks, I think it's very good ROI to start post-training models for those use cases.

你怎么看待推理市场的走向？

How do you see the inference market playing out?

你刚才很公开地提到了Fireworks。

You mentioned that Fireworks very openly.

我们是Fireworks的大投资者，我很喜欢Lynn和她的团队。

We're a big investor in Fireworks, and I love Lynn and the team.

但每个人都在尝试做这件事。

But everyone's trying to do it.

我们之前请了Cruso的Chase来上节目，他们的业务也包括这部分。

We had Chase on the show from Cruso, who also have that as part of their business.

我和Nebius的那些朋友关系很好，他们也在做这个。

Friends with the Nebius guys, they also do it as part of it.

这个业务会怎么发展？你怎么看待需要做整个技术栈，还是只做推理服务提供商，比如像Fireworks或Base10这样的？

How does this business play out, and how do you think about the need to do all of the stack versus inference-only providers, like a Fireworks or a Base10?

我觉得关于推理市场会怎么发展，有两件事是我从LangChain开始到现在真的低估了的。

I think when it comes to how things will play out with respect to inference, there's two things that I think I really underestimated from the start of LangChain to now.

一个是对推理的需求，另一个是对编码代理的需求。

One is the demand for inference, and two is the demand for coding agents.

我觉得这两个市场都比三、四年前我想象的要大得多。

I think both of those markets are just tremendously bigger than I thought three, four years ago.

我记得Benchmark的Eric Vishria最近在推特或播客上说过，基本上就是说，一切都能运转。

I think actually it's Eric Vishria at Benchmark who said this on either a tweet or podcast recently, where I think he basically said, everything just works.

这个市场太大了。

This market is so big.

Thomas Tunks也提到过，推理市场就是这么大，以至于——我觉得Eric说的是更广泛的东西，但放到推理上，我觉得任何推理服务都能行。

I think Thomas Tunks as well had something around just the market for inference is just so big that-- I think Eric said it about more broad stuff than just inference, but I think applying it to inference, I think any inference will work.

需求太大了，而且我不觉得短期内会减少。

There's just so much demand for it, and I don't know if it's going to get less anytime soon.

所以，对，当然。

And so yeah, sure.

Fireworks会成功。

Fireworks will work.

AWS也会成功。

AWS will work.

Cruso 也能行。

Cruso will work.

我只是觉得需求太大了。

I just think there's so much demand for it.

你今天不会投资哪些领域？

Where would you not be investing today?

我可以给你举几个例子。

So I can give you some examples.

我觉得客户支持这块领域特别没意思。

I think the customer support space is wildly uninteresting.

过去18个月里有20家供应商都融资超过一亿。

20 providers all funded with over 100 million in the last 18 months.

还有 Salesforce 和 Atlassian 这些大公司也在里面。

You got Salesforce and Atlassian and big incumbents there too.

还有像 Intercom 和 Talkdesk 这些老牌公司也在做。

You got kind of more of the old ones like Intercom and Talkdesk doing it too.

好的，不错。

Okay, cool.

或者机器人领域，比如Bold Tape，但现在有太多资金涌入机器人行业了。

Or robotics, like bold tape, but there's so much money going into robotics.

我们离ChatGPT那样的时刻还差得远呢。

We're massively far away from having a ChatGPT moment.

特别是人形机器人，在外形设计上真的很困难。

Humanoids, in particular, really struggle in form factor.

我的回答是，我觉得资金和兴奋程度远远超过了实际能力。

I just think the funding and the excitement exceeds the capabilities significantly, would be my answer.

我倾向于认为自己是个乐观主义者，所以我确实认为这项技术具有变革性，因此会有很多机会。

I like to think I'm somewhat of an optimist, and so I do think this technology is transformative, and so I think there'll be ton of opportunity.

所以我的回答可能是这样，然后通过这个，我们可以反推出我不会投资什么。

And so I might answer that, and then through that, we can back out what I wouldn't invest in.

推理方面我们已经谈过了。

I think inference we already talked about.

我觉得有一个巨大的——

I think there's a tremendous-

好，就这么办。

Yeah, let's do this.

你是我最新的合伙人——

You're my newest partner-

……加入这个基金。

... in the fund.

欢迎。

Welcome.

是的。

Yes.

第一天就有你，真是太好了。

It's so good to have you on your first day.

谢谢。

Thank you.

第一，我们应该有推理能力。

Number one, we should have inference.

我们应该有推理能力。

We should have inference.

好的。

Okay.

将来会有更多的软件被编写，所以我们也应该投资于更底层的、基础设施类的基础构件。

There's going to be way more software written, so we should also invest in lower level, kind of infrastructure primitives.

Modal，这家公司很棒。

Modal, incredible company.

Temporal，这家公司也很棒。

Temporal, incredible company.

Supabase。

Supabase.

Convex。

Convex.

他们不是和Fireworks有竞争关系吗？

Do they not compete with Fireworks?

Modal在某些方面确实有竞争，是的。

Modal does on some things, yeah.

好的。

Okay.

但他们做的更多。

But they also do more.

他们提供沙盒环境。

They do sandboxes.

所以我们有推理层，第一，编码层。

So we've got inference layer, one, the coding layer.

我会说是推理，然后是基础设施。

I would say inference and then infrastructure.

所以，底层基础设施，Temporal、Modal、Supabase、Convex，还有垂直领域的智能体。

So, low-level infrastructure, Temporal, Modal, Supabase, Convex, vertical agents.

我觉得垂直领域的智能体会是一个很好的方向。

I think vertical agents are going to be a good layer.

有没有什么特定的垂直领域？

Any specific verticals?

我不确定有没有哪个垂直领域我会特别看好。

I don't know if there's any kind of verticals that I would overweight so much.

我觉得关键是要有领域内的独特见解，并且深入钻研。

I think the important parts are having domain-specific point of view and really diving deep there.

我记得帕特里克·科里森最近在一个播客里说过，好像是和Replit的Amjad一起聊的，当时他在谈如果自己再创业会做什么。

I think Patrick Collison said this on a recent podcast, I think with Amjad from Replit, when he was talking about what the next startup he would build.

可能不会选一个特别光鲜的领域。

It probably wouldn't be a sexy area.

可能会选一个技术渗透率低，但积累了很多领域知识的行业，然后进去真正改变它。

It would be some area where technology maybe, or at least AI penetration is low, but there's a lot of domain expertise built up there, and kind of going in there and really transforming that.

关键就是要有一些领域专长，这些专长可以体现在数据或者逻辑里，放进系统里，最终可能融入模型权重，但我不会一开始就从模型权重入手。

The key part is basically just having some domain expertise that can either be captured in data or logic that goes in the harness, and eventually maybe the model weights, but I wouldn't start there.

拥有一些可以捕捉到的领域专长，然后把它付诸实践。

Having some domain expertise that can be captured there and then putting it to work.

顺便说一下，我觉得在这些事情中，有三点是我会关注的，而且我认为它们是先决条件。

I think, by the way, there are maybe three things that I would look for in all of these and I think are prereqs.

如果你看看所有成功的团队，第一，我认为他们的团队人才非常强大。

If you look at all of these teams that are succeeding, one, I think the talent that they have is tremendously strong.

比如Modal，他们的团队人才非常强。

Like Modal has an incredibly strong kind of talent team.

Lynn和Fireworks，对吧？

Lynn and Fireworks, right?

他们写了PyTorch。

They wrote PyTorch.

这多酷啊，对吧？

Like how cool is that, right?

我们见到他们的时候，你知道，当你见到创始人时，最后一次简短介绍时，就像说："我写了PyTorch。"

When we met them, you know when you're kind of meeting founders, the last time you do the brief intro, it's like, "I built PyTorch."

哦，那挺好的。

Oh, well, that works.

我们就能打勾了。

We'll get a tick in that box.

所以我觉得要构建这种特别深度的基础设施，你必须技术非常强，而能做到这一点的人真的不多。

So I think to build, especially this really deep infrastructure, you have to be super technically strong, and there's just not that many people that can do that.

第二，你必须非常非常快地行动。

Two, you have to go really, really fast.

能够快速迭代，我觉得极其重要。

Being able to iterate quickly I think is incredibly important.

市场变化太快了。

The market just changes so rapidly.

我觉得速度和执行速度是关键。

I think speed and execution speed.

说到我们在LangChain的优势，我觉得我们迭代得非常非常快，这也是我们相对于其他公司的一个优势。

When I think about our advantages at LangChain, I think we iterate really, really quickly, and I think that's one of the advantages we have over other companies.

最后一点就是品牌。

And then the last thing would be brand.

我觉得在这个AI时代，很多成功的公司基本上都是抓住了机遇，然后乘势而上，充分利用了这一点。

I think a lot of companies that made it in this AI era, basically caught lightning in a bottle and then rode that wave and captured that.

看看LangChain早期的情况。

You look at LangChain early on.

我们很快就火了起来，并且一直在朝着那个方向努力。

We went very viral and have executed against that cursor.

很好的例子。

Great example.

他们也很快就火了，并且抓住了那个机会。

They went very viral, executed against that.

Jev，甚至Claude Code、ChatGPT也是这样。

Jev, even Claude Code, ChatGPT.

看起来这个领域发展得太快了，而且因为竞争对手太多，也许正是因为这样，才发展得这么快。

It seems like the space is moving so fast, and because there are so many competitors, maybe this is even because of it.

因为竞争对手太多了，LangChain 出来后的两个月里，冒出了一百个不同的框架。

Because there are so many competitors, there were 100 different frameworks that came out in the two months after LangChain came out.

但我觉得我们建立起了品牌优势，而且我认为这是我们一直依赖的持久优势。

But I think we had a brand advantage that was built up, and I think that was a durable advantage that we rode.

所以 Instinct 是另一个例子，可能有一堆个人助理。

And so Instinct, another example of this, there's probably a bunch of personal assistants.

我觉得如果有的话，估值可能会高于合理水平，但就一些融资轮次而言，如果有一家脱颖而出的公司，拥有非常强大的技术团队，并且证明了自己能快速执行，那你根本不会去看价格，直接投资就行了。

I think if there is, in terms of it would probably be at maybe higher valuations than is rational, but in terms of some of these funding rounds that are getting done where there's a breakout company with a really strong technical roster that has proven that they can execute quickly, I think you just don't look at the price it, you just invest.

抢先一步是不是比以往任何时候都更重要？

Is being first more important than ever before?

我觉得是的，绝对如此。

I think so, absolutely.

追赶是有可能的。

It's possible to catch up.

我觉得 OpenAI 在追赶 Codex 方面做得非常出色，但 Anthropic 和 Claude Code 抢了先机，哇，那真是巨大的优势。

I think OpenAI's done an incredible job of catching up with Codex, but being first, Anthropic and Claude Code, that wow, what an advantage that was.

但你说得对，LangChain一出来，几个月后，越来越多的类似框架就冒出来了。

But to your point, yeah, LangChain comes out, and then months after, more and more come out.

但Instinct也是一样，一样的道理。

But Instinct, same, same.

但实际上，作为第一个被光环笼罩的品牌，所拥有的持久垄断优势是非常真实的。

But actually, the durable brand monopoly that comes from being the haloed first is very real.

我觉得这确实非常真实。

I think that is super real.

我还认为，在那之后，你必须执行得非常出色。

I also think that you have to execute incredibly well after that.

我刚才说的也有反例，仅仅因为你是第一个，并不意味着你就能保证成功。

Counterexample to what I just said, just because you're first doesn't mean that you're guaranteed success.

AutoGPT，不知道你还记不记得这个项目，它在LangChain之后几个月出现，是GitHub上有史以来增长最快的项目。

AutoGPT, I don't know if you remember the AutoGPT project, but it came out a few months after LangChain, was the fastest-growing GitHub project of all time.

但之后我就没怎么听说过他们了。

I haven't heard much from them since.

这倒不是说他们不好，因为快速执行本身就很难。

And that's not a knock, because it's very hard to execute with speed.

但我觉得，就算你在某个领域是第一个，也不代表你就能存活下来。

But I think just because you're first in an area doesn't mean that you'll survive.

我从来不喜欢批评别人，所以我觉得在某些方面，与其批评，不如表扬他们。

I never like to knock people, so I find it helpful to praise them instead in some respects.

如果你要写一份加入公司的备忘录，说明我们为什么要投资Jev的90亿美元融资，你对Jev的乐观理由会是什么？

If you were to write a memo joining the firm for why we should invest a $9 billion Jev round, what would your bull case be for Jev?

对Jev的乐观理由是，有一种新的推理方式和新的模型类型，而且这个模型类别就是以他们的名字命名的。

Bull case for Jev is that there's a new type of inference and new type of model, and the model category is literally named after them.

所以他们会乘着这股浪潮前进，而且他们是一支非常有才华的团队，能够在这方面执行得很好。

And so they're going to ride that wave, and they're a very talented team that can execute on that.

我对这个团队不太了解。

I don't know the team super well.

我认为，对于这个价值的融资轮次，这必须是假设的前提。

I think that would have to be the assumption for a round at that value.

从我目前看到的情况来看，确实是这样。

From everything that I've seen, that seems to be the case.

但这就是看涨的理由。

But that would be the bull case.

他们基本上开创了一个新的模型类别，并且与之紧密相关，同时拥有一支能够执行这个愿景的优秀团队。

They basically invented a new model category, are associated with it, and have a talented team that can execute on that.

从Benchmark拿钱会改变一家公司的发展轨迹吗？

Does taking money from Benchmark change the trajectory of a company?

百分之百会。

100%.

我觉得Benchmark，当Miles和整个Benchmark合伙团队投资我们的时候，当时只有我和Ankush。

I think Benchmark, when Miles and really the whole Benchmark partnership invested in us, it was me and Ankush.

我们只有两个人。

We were two people.

我们真的把他们当作像是第三个联合创始人一样。

We really thought of them as kind of like a third co-founder.

从两个人到现在大概四百人，回顾整个公司建设的过程，他们绝对是我们一路上最稳定的合作伙伴。

When I think about company building from two people to where we are now today, which is like 400 people, they've been our most consistent partner in that journey, for sure.

我不是做风投的。

I am not a VC.

我从来没做过风投。

I've never done VC.

可能还有其他市场因素，比如Benchmark参与进来本身就是一个信号，而且我们很幸运，在Benchmark那轮之后两三个月，红杉也加入了我们，他们也是非常出色的合作伙伴。

There's probably also other market dynamics where, great, Benchmark's involved, and that signals thing, and we're very lucky to have Sequoia join us after two months, three months after the Benchmark round, and they've been a tremendous partner as well.

红杉那边是谁？

Who was it from Sequoia?

红杉的Sonja，她是对这个领域最了解的人之一。

Sonja from Sequoia, who is one of the most plugged-in people in the spaces.

他们的风格有什么不同？

How do their styles differ?

Benchmark可能更关注公司的核心是什么。

Benchmark is maybe a little bit more, what's the core nucleus of the company?

这就是我们会成为他们的参谋，并在这方面提供帮助的地方。

And that's where we're going to be a sounding board for that, and we're going to help with that.

我的意思是，我觉得，是的，很多关于公司建设、定位方面的事情，我都会去找他们，他们也会花时间在这上面。

So what I mean by that is, I feel like, yeah, a lot of the company-building things, positioning things, I would go to them for, and they would spend time there.

红杉的风格，我觉得，是他们什么都帮忙。

Sequoia's style, I think, is they help with everything.

他们就像一台机器。

They are just a machine.

他们帮助招聘个人贡献者和高管。

So they help with recruiting of ICs, of executives.

他们帮助营销和发布，而且他们会真正地落实到操作层面。

They help with marketing and launches, and they'll really operationalize.

真的有用吗？

Does it really work?

我觉得是的。

I think so, yeah.

我们的很多候选人都是通过红杉的推荐来的。

We've gotten a lot of our candidates from Sequoia references.

真的吗？

Really?

他们有一个人才网络，我们从那里招到了很多非常非常优秀的候选人。

They have a talent network, and we've gotten in a bunch of really, really strong candidates from there.

你会对创始人说，为了红杉，值得接受折扣吗？

Would you say to founders it's worth taking a discount for Sequoia?

我觉得绝对值得为红杉接受折扣，还有Benchmark，可能还有其他一些顶级或一级的风险投资公司，但那些是我们合作的。

I think absolutely it's worth taking a discount for Sequoia, and Benchmark, and probably other of these top one or tier one VC firms, but those are the ones that we work with.

关于融资，你现在知道什么，是你希望当初只有你们两个人的时候就能告诉自己的？

What do you know now about fundraising that you wish you could tell yourself when it was just the two of you?

我觉得融资几乎是一场游戏，而我一开始肯定不知道这一点。

I think there is this game to fundraising almost, and I definitely didn't know that going in.

我是第一次创业。

I'm a first-time founder.

我基本上把所有时间都花在产品上，所以非常专注于此。

Where I spend all my time is on product, basically, and so I'm very focused on that.

我确实觉得融资是有技巧的，你想在不刻意融资的情况下完成融资。

I do think there is this game to fundraising where you want to fundraise without fundraising.

你想先拿到一份条款清单，然后告诉别人你已经有了一份，然后——

You want to get one term sheet and then tell others you have a term sheet and-

在推特上发布收入曲线图，然后就会引来很多主动找上门的，然后——

Post a revenue graph on Twitter, and then it leads to all this inbound and-

它——

It-

你喜欢这样吗？

Do you fucking like this?

我讨厌这样。

I fucking hate it.

也许我现在已经不适合玩这个游戏了。

Maybe I'm too old for this game now.

这不过是摆姿态而已。

It's just posturing.

就像说，"哈里森，你是真的喜欢我，想和我一起创业吗？

It's like, "Harrison, do you really like me and want to build a business with me?

我给你同样的价格和同样的钱，"然后随便吧。

I'll give you the same price and the same money," then whatever.

"咱们就这么定了。

"Let's just do this.

别再搞这些虚的了。"

Stop the fucking posturing."

我一点都不喜欢融资。

I don't like fundraising at all.

我觉得，就像你说的，我想和喜欢的人一起做出伟大的东西。

I think, what you were saying, I want to build something great with people I enjoy working with.

我觉得，是的，尤其是考虑到我们的雄心壮志，我们需要融资才能实现。

I think, yeah, especially for the scale of ambitions that we have, we need to fundraise to get that done.

我觉得Benchmark在向我们融资时也是这么做的。

I think that that was the approach that Benchmark took when we were fundraising from them as well.

就像说，"嘿，我们先看看彼此是否合得来，是否愿意一起合作。"

Like, "Hey, let's figure out if we like each other and want to work together."

顺便说一句，他们这样谈判非常高明，所以我觉得他们也是很好的谈判者。

By the way, that was very good negotiating by them, and so I think they're very good negotiators as well.

但回顾我们最近完成的几轮融资，我们的B轮是从IVP的Tom那里筹集的。我记得从红杉那轮开始，我就和Tom聊了超过两年半。

But when I look at our most recent rounds that we've done, we raised from Tom at IVP, our Series B. I'd been chatting with Tom from, I think right around the Sequoia round, for over two and a half years.

所以那是一段随着时间慢慢建立起来的关系，我逐渐了解了他，他也逐渐了解了我。

And so that was a relationship that we just built up over time, and I got to know him, and he got to know me.

所以到了需要融资的时候，我觉得其中少了很多博弈，我真的很感激这一点。

And so when it came time to fundraising, I feel like there was less games involved there, and I really appreciated that.

我想来个快速问答环节。

I would love to do a quick fire round.

我说一个简短的陈述，你马上给我你的想法。

So I say a short statement, and you give me your immediate thoughts.

这样可以吗？

Does that sound okay?

来吧。

Let's do it.

你最喜欢的一次风投会议是哪次？

What was your single favorite VC meeting?

和Benchmark的第一次路演。

First pitch with Benchmark.

房间里全是传奇人物。

So many legends in the room.

和谁相处最酷？

Who was the coolest to hang with?

比尔·格利。

Bill Gurley.

你当时紧张吗？

Were you intimidated?