AI 代理与争夺客户数据的较量

我们会有一个HR团队，专门负责新AI的入职流程。

Like, we'll actually have an HR, and that HR team will onboard AIs as they come.

他们会培训这些AI。

They will train them.

还会指导它们访问所需的文档。

They will show them the access to the documents that you need.

它们会加入不同的团队。

They'll be part of teams.

也会加入Slack，就跟真人一样。

They'll join the Slack, you know, just like humans do.

在这个世界里——我们之前稍微提过一点——这些其实不是软件。

And in that world, and we touched on this a little bit before, but in that world, these aren't software.

这更像是增加了软件的使用席位和消耗量。

That's actually more seats, more consumption of software.

所以，你觉得对企业级AI来说，我们是不是该把它们当人一样来管理？还是觉得这想得太远了？

And so do you think that for enterprise agents, we're moving more to these, you treat them like humans or do you think that that's too far?

我觉得这是一种过渡形态。

I think it is a intermediate form.

我觉得这之所以有效，是因为它能无缝融入现有的工作流程，而不用彻底重构一切。

I think the reason this works well is because you can slot it in to the existing workflows without having to refactor the whole universe.

没错。

Yes.

比如我那个管理美国网球协会球队的助手，它可以给球员发邮件。

So in my example of my agent that manages my USCA tennis team, It can email with the players.

它能登录USTA官网，查看赛程、排阵容、核对人员可用性。

It can go to the USTA website, check the schedule, make lineups, check availability.

它拥有独立身份这一点特别有用，因为这样就能自然接入所有为人类设计的现有流程。

And it works well having its own identity because it can slot into all of these existing workflows that were designed for humans.

在Fivetran，我们有个AI助手专门处理客服工单，会去查看日志、分析代码。

At Fivetran, we have an AI agent that helps respond to support tickets that goes and inspects the logs, inspects the code.

它能调用Fivetran里我们集中管理的所有上下文信息。

It uses all the context that we centralized with Fivetran Yeah.

为了弄清楚这位客户的情况。

To find out what is going on with this customer.

这个问题可能的解决方案是什么？

What might be the solution to this problem?

它还会草拟回复内容。

And it drafts responses.

目前，它在系统中的运作方式就跟真人一样。

And right now, it slots into the system a lot like a person would.

但我们正在努力把它改成一个完全封闭的纯AI系统，只保留一个身份。

But we are working on making the whole thing just a closed loop pure AI system where it will only have one identity.

到时候就只有一个连接故障排查的‘博格集合意识’了。

There will just be the the connector troubleshooter Borg hive mind.

也许确实可以只有一个，但让我来谈谈为什么我用Mac mini来运行我的代理。

And and there may and there may be, but I mean, me give the argument for the the multiple so so why why do I use a Mac mini for my my agent?

对吧？

Right?

其实有几个原因。

I mean, there's a couple reasons.

其中一个是为了能访问iMessage，因为根本没法通过编程方式接入iMessage，所以它得有个桌面环境。

One of them is so it can access iMessage because there's no programmatic way to access iMessage, and so, like, you know, it has a desktop.

另外，现在大家都往无头化走。

Another one is like everybody's moving headless.

比如Salesforce在搞无头架构，还有无头浏览器这些。

Like, Salesforce is doing their headless thing, and there's headless browsers.

但咱们还是先聊聊浏览器吧。

But it it turns out let's just talk about browsers.

你用无头浏览器，反爬虫系统立马就察觉了，功能立马打折扣。

Like, you have a headless browser, all of the anti scraping software kicks in, and then, like, it's not as functional.

所以还不如直接给它一个完整的Safari浏览器，效果好得多。

So it's actually much better to just give it a fully functional version of Safari.

过去三十年里，那些为应对不懂技术的用户而发展出来的界面，其实就是我们现在用的这些。与其费劲重写所有这些东西，不如直接让这些基于人类数据训练的代理，去访问完整的端到端系统，这才是最省事的办法。

So you could argue that the interfaces that have evolved over the last thirty years to deal with unpredictable users that know how to use computers is the UIs we have today, and actually the simplest thing to do rather than try to rewrite all of that stuff is just to give these agents that have been trained on human data access to full end to end systems.

我不确定。

I don't know.

那些系统呢，

The systems that

我做过的项目里，从来不需要用浏览器。

I've worked on, browser use has never been necessary.

用浏览器代价很高，速度特别慢。

Browser browser use has a big cost, which is very slow.

对。

Mhmm.

而且会消耗大量令牌。

And it consumes a lot of tokens.

没错。

Yes.

我发现在网球这个例子中，美国农业部的网站根本没有反爬虫机制，好吧。

And I have found that in my tennis example, the USDA website does not have any anti scraping provisions and Okay.

所以你并不是想从 LinkedIn 或 Zillow 上抓数据吧。

So you're not trying to you're not trying to like read from LinkedIn or Zillow.

要是你真想试，那当然可以。

Try that if Well, you yes.

我只是在用 Python 做浏览器自动化。

And so I'm just using I'm using just, like, Python browser automation.

没错。

For sure.

是的。

Yeah.

我实际上用的是 Selenium。

I'm actually using Selenium.

是的。

Yeah.

而且我还发现有个工具能直接输出你需要的信息，这样代理就不必一直去解析 HTML、浪费大量令牌了。

And then I'm actually there's actually a skill that emits just just exactly what you wanna know so that so that the agent is not reading HTML and consuming all those tokens all the time.

在Fivetran，我们现在正在开发一个Salesforce管理代理，用来实现对Salesforce的持续集成，处理那些小改动，因为现在这部分工作特别耗人力。

And then at Fivetran, for example, we are working right now on a Salesforce administration agent to to integrate to basically do continuous integration of small changes into our Salesforce, which is very labor intensive right now.

而且我们没用浏览器自动化，因为发现Salesforce CLI已经能完成我们所有需要的操作。

And it also does not use browser automation because the Salesforce CLI, we found does everything we need to do.

如果真有必要，我们也会用浏览器自动化。

We'll use browser automation if we need to.

要是遇到CLI做不到的事，我们再考虑，不过得说句公道话，Salesforce CLI的功能已经非常全面了。

If we come to something that can't be done with the Salesforce CLI, but to their credit, the Salesforce CLI is quite comprehensive.

基本上你在界面里能做的所有操作，CLI命令都能实现。

Pretty much anything you can do in the UI, you can do with a CLI command.

而且这些代理似乎早就知道怎么用它了。

And the agents seem to already know how to use it.

那我能问一下吗

So can I do a

要不我们打个一美元的赌？等我们这档播客做满五期的时候，你猜会怎样？

$1 bet with you for when we're on this podcast for five Sure?

我觉得五年后，智能体的主要应用场景还是会沿用人类现在用的那些界面，因为大量已有的集成问题其实已经解决了，权限控制、数据共享这些也都配套到位了。

So I think the majority of use for agents in five years is gonna be the same interfaces that humans are using just because it's a long tail of integration that's already been solved, and like all the protections, and all the sharing, everything And I and would you say the majority would be through APIs that are kind of more of a more of a traditional computer software?

是的。

Yeah.

它们直接调用 API 就行了。

They'll just hit the APIs.

对。

Yeah.

那你对像 MCP 这类新出现的技术怎么看？它们就是想解决这类问题的。

What do you think about things like, you know, these these technologies that mediate that like MCP, which have emerged to kinda try and solve that problem.

你觉得这类技术还有未来吗？

Do think there's a future for those?

还是说它们最终会被更纯粹的工具型 API 取代？

Or do you think that those just give way to like strict tool API usage?

理论上讲，这类中间层其实有点多余，因为现在的智能体本来就擅长调用 API 和命令行工具。

You know, in theory, it seems like they're an unnecessary layer because these agents are great at calling APIs and calling command line tools.

那为什么还要多这一层呢？

So why bother having this other layer?

但实际动手做系统的时候，MCP 确实解决了不少关键问题，尤其是身份验证。

In practice, when you sit down and actually try to build systems, MCPs do solve important problems, particularly authentication.

是的。

Yeah.

还有就是。

And Yeah.

还有就是能清楚看到有哪些功能可用。

Just like discoverability of what's available.

虽然从理论上讲，这层东西可能根本没必要存在，但真要开发一个需要访问上下文的系统时，你几乎都会忍不住加一个 MCP 服务进去。

So even though you can say kind of from first principles that maybe they shouldn't exist, when you actually sit down and write a a real system that's accessing context, you almost always end up sticking an MCP server into it.

这涉及到授权、认证和功能发现，这些功能确实很重要。

And this is authorization authentication authorization and discoverability, like the fact that these things exist.

是的。

Yeah.

而且围绕MCP构建的AI工具还提供了很多小细节，比如用户授权特定任务，这些操作都是在MCP层面完成的。

And there's just also a lot of little affordances in the AI tools that are built around MCP, like user user granting authorization for specific tasks are done at the tool level in MCPs.

是的。

Yeah.

这已经成为这些框架中默认的规则，在很多场景下都运行得很好。

And that's just a rule that's been baked into the harnesses that ends up working well for a lot of situations.

但在某些情况下，这种做法也会出问题。

It also works badly in some situations.

不过，使用方的工具已经开始围绕MCP发展起来了。

But the the the tools on the consuming side have started to grow around MCP.

对。

Yeah.

所以即使理论上你可能不需要它，但在实际应用中，它正在逐渐站稳脚跟。

And thus even if maybe theoretically you don't need it, I think in practice, it's it's taking it's taking hold.

我觉得特别奇怪的是，就连工具使用本身，你也可以认为，随着模型越来越智能，它们完全可以自己构建更好的工具。

And the thing that I find so strange about this is like even the tool use itself, like, you could argue that, like, as smarter models come out, they could build better tools anyways.

所以有时候我在想，代理是不是应该做到极致简单。

And so sometimes I wonder if, like, like, an agent should just be the most minimal thing ever.

它只需要管理持久化状态。

Like, it manages, like, durable state.

它只需要管理计算资源。

It manages compute.

然后你直接调用 Anthropic 或 OpenAI 的 SDK，告诉它自己去构建所需的工具。

And then, you know, you run, like, whatever, like, the Anthropic SDK or the OpenAI SDK, and then you just tell it to build its own tools.

比如，自己去建立和这个系统的连接。

Like, you know, build your connection to this.

这正是 Nanoclave 的工作方式。

Well, that's how that's how Nanoclave works.

不对。

No.

现在实际就是这么运作的。

It's exactly how it works right now.

但你可以说，现在有这么多钱涌向这些基础模型。

But you could argue that, like, there's so much money being poured into these foundation models.

几十亿甚至可能几百亿美元吧。

Like tens of billions of dollars, maybe come hundreds of billions.

所以，我手上拥有的，是目前全球最智能的系统之一。

So I have the most intelligent thing at any point on the planet as one of these models.

既然它们能自己造出更好的工具，干嘛还要用旧的呢？

So why would you use an old tool if they could build a better tool?

是的。

Yeah.

所以，对不了解的人说一下，NanoClaw 是一个个人AI代理，有点像 OpenClaw，但它的运作方式是，你启动时直接 fork 代码库。

So I mean, NanoClaw, for those who don't know, is a personal AI agent, sort of like OpenClaw, except the way it works is you fork the repo when you start up.

是的。

Yeah.

然后你就随心所欲地修改代码，让它变成你想要的样子。

And then you just sort of vibe code it into whatever you want.

是的

Yeah.

随你便吧。

Whatever you want.

没错。

Exactly.

我遇到的问题是，它有点跑偏了。

And the problem I encountered is it sort of went awry.

我真不想去一一排查所有问题。

And I didn't really wanna go and troubleshoot I all of the

调试嘛，我确实调试过Nano。

debug I debug Nano Exactly.

我我

I I

它就是一个光标。

I It's a It's cursor.

我那时候就在里面，按理说该这么做

I'm in there like I should

在 NanoClaw 上花了不少时间调试，最后我就想，我需要一个更清晰分离职责的系统。

go to the some point, spent enough time debugging NanoClaw that I was like, I want something that has more separation of concerns.

比如这边有个代理，提供一套 API，按它自己的方式运行。

Where there's sort of an agent over here that has an API and it works the way it works.

然后我就在上面写各种功能和模块。

And then I just write like skills and and stuff on top of that.

讽刺的是，就算我这么做了，最后还是得给 NanoBot 加功能，因为它分不清 WhatsApp 的群消息和私信。

Ironically, even when I did that, I ended up having to add things to nanobot because like it couldn't differentiate WhatsApp group messages from DMs.

我估计还为此给 NanoBot 提交过一个合并请求。

And I think I actually have a PR against nanobot because of that.

所以某种程度上，我又回到了原点。

So I sort of ended up back in the same place a little bit.

不过我明白你的意思。

But I take your point.

如果它们足够聪明，可能会根据需要自行构建中间层抽象。

Like, if they get smart enough, they may, you know, they may just build their own intermediate abstractions as needed.

对。

Yeah.

没错。

Yeah.

确实如此。

That's right.

随着发展吧。

As you go.

咱们先把这个话题收一收。

Just to gonna maybe wrap this this bit up.

萨提亚说过，SaaS 会崩溃，而 Databricks 已经在尝试用他们的平台重建大量 SaaS 应用，这事他们公开讲过很多次。你觉得真会迎来 SaaS 的末日吗？会不会出现一场巨变——未来全是智能代理，运行在全新的基础设施上？还是说，其实 SaaS 根本没事，市场反应过度了，萨提亚其实说错了？

I mean, Satya has said that there's gonna be the collapse of SaaS, and know Databricks has tried to rebuild a lot of SaaS on top of their platforms, I mean, they've been very public about that, so do you think we're gonna see a massive, like do you think the SaaS pocalypse is a thing, and we're gonna see a massive shift, or and like, you know, it's gonna be agents and it's gonna be on new types of infrastructure that they run, or do you think that, you know, SaaS is fine, the markets are overblown, Satya is, you know, wrong.

要说公允一点，我觉得资本市场看得很准，至少方向没错，至于幅度对不对，我不好说。

Well, to give credit to the public markets, I I think that they are accurately I mean, whether the magnitude is right, cannot say, but the direction is right.

这些SaaS公司里头，不确定性比一年前多太多了

There's a lot more uncertainty embedded in all these SaaS companies

对

Yeah.

包括我自己的公司，也比以前更不确定了

Including my own, than than there was a year ago.

这反映在股价的持续下跌上

And that's reflected in the changing in the in the decline prices.

但我并不觉得是因为这个原因

But I I don't think it's for this.

我不太认同这种说法，对

I don't really buy into this reason Yeah.

说所有SaaS类别都会消失，被几行代码写的软件取代

That all these SaaS categories are gonna disappear and be replaced with five coded software.

我觉得可能会有一些，但更大的威胁其实是新公司不断冒出来

I think there will be some, but I think the bigger threat is simply new companies coming along.

现在写软件容易多了，更原生一些，是的。

It is just so much easier to write software More native, yeah.

AI原生公司会迅速追赶并超越那些老牌企业，有些方面甚至可能做得更好。

Now that AI native companies will just zoom and catch up to to the established incumbents and maybe be better in some ways.

你有没有觉得特别有意思的一点？

You know what's so interesting?

你刚才提到Fivetran，但如果我们看看我们合作的那些传统公司——当然，我不把你算在内，毕竟你们的实验室才是你们的主要客户。

I mean, you just you just mentioned that about about Fivetran, but if you look across the companies we work with that are traditional companies, and I wouldn't I actually would not include you in that, I mean, listen, you know, like both of the labs are are your primary customers.

但如果你看那些传统公司，数据其实显示它们并没有放缓。

But but if you look at across the traditional companies, like, you don't actually see in the data that they're slowing down.

Fivetran的业务做得非常好。

Like, Fivetran is doing great.

它的增长反而在加速，那你怎么解释你刚才说的高风险担忧和实际业务表现之间的矛盾？

It's actually the company's accelerating, and so how do you reconcile this high risk concern that you just voiced with the actual the business?

是不是只是因为……

Is it more just kind of

说实话，这很难说。

Well, I it's it's hard.

不是的。

No.

这就像是X风险。

It's like x risk.

你明白吧？

You know?

突然出现一种全新的做事方式，效果好得多，总会有一些公司被取代。

Some some some new way of doing what you do comes along that's dramatically better, and this will happen to some companies.

当然会了。

Of course, it will.

对吧？

Right?

但再过几年

But And and more years

到目前为止，已经进入这个阶段了。

into this at this point.

是的。

Yeah.

你明白吧？

You know?

我不确定。

I don't know.

我觉得，就我所见，这些情况是没法从数据里推导出来的，也许将来会出现吧。

I just feel like you do you you you can't derive this from the data as far as I can tell, and, like, maybe it's coming.

或者对。

Or Yeah.

或者这根本就反映不到上市公司的数据里。

Or maybe It's not even really in the public company data.

没错。

No.

不是的。

It's not.

对。

No.

是的。

Yeah.

没错。

Yeah.

普遍都是这样。

Across the board.

当你开始看到净收入留存率低于1的时候，那就说明情况不对了。

It's when you start to see, like, less than one net dollar retention, then, you know, it's it's Yeah.

对。

Yeah.

已经出现了。

It's come.

是的

Yeah.

对

Yeah.

没错

Yeah.

当然

For sure.

是的

Yeah.

也许会吧

And maybe it will.

也许将来会，但我们一直都在聊这个话题。

Maybe it's in the future, but like, we've been having this conversation.

我记得清楚，你上次打电话来的时候，大概一年前，你说马丁，这AI真不是闹着玩的，能用它写连接器，效果不错，而且趋势已经很明显了。

I I remember I actually remember when I was when you put in the call, and it was a year ago or so, and you're like, Martin, this AI stuff is like really real, and, you know, like, you can kinda code connectors with it pretty good, and it's pretty good, and it's coming.

但你打那个电话的时候，其实已经是很久以前的事了，自那以后公司发展得特别好。

But that was actually quite a while ago when you put in that call, and the company's done fantastic since then.

所以很可能的情况是，与其有人想重新改造一个已经摸清长期积累的公司，不如去 tackling 更合适的新问题。

So it could be the case that, like, rather than someone trying to redo an existing company that, you know, has figured out, like, a long tail of stuff, they'll go work on different problems that are more suitable.

就我们自身而言，这些年一直尝试用人工智能来构建数据复制连接器，这正是我们业务的核心。

So we have, you know, in our particular case, we have been trying ourselves to use AIs to build data replication connectors, which is the core of what we do.

是的。

Yeah.

对。

Yeah.

从 GPT-3 那会儿就开始了。

For years, since gbt three.

没错。

Yeah.

它们在输出能力上一直在持续提升。

And they they continue to improve in terms of what they can put out.

他们仍然没意识到这种复杂的长尾问题。

They still do not discover this long tail of complexity.

当然。

Of course.

很多人没想到，要准确复制一个系统并保持同步居然这么难。

It is it is it really surprises people how difficult it is just to make an accurate copy of a system and keep it up to date.

现在，Fivetran 开始展现出一些新能力，进一步提升质量，尤其是在数据源的覆盖完整性和复制准确性方面。

And so and and and now we are actually starting to see new capabilities inside Fivetran to push the bounds of quality, particularly quality even further, like like completeness of coverage of the sources and the correctness of replication.

你可以想象，用人工智能来做这件事，远比靠人力全面得多。

You can imagine how you can use AIs to do that more comprehensively than you ever could with human beings.

是的。

Yeah.

我觉得，除了人工智能的威胁越来越近，但它离我们目前的工作还有不小的距离。

And so I I think in addition to the sort of, you know, the AI threat is getting closer, but it's still it's still, I think, a ways away from what we do.

我们反而正被这些新机会推着向前走。

We're actually starting to see the opportunity pull us forward.

所以我们正通过内部使用人工智能，在一些非常不显眼的方式上，不断提升自己的核心业务，我觉得别人还没发现这些方法。

So we're we're starting to get better at our own core business by leveraging AI internally in extremely non obvious ways that I don't think anyone else has has discovered yet.

你能

Can you

就是说

just So

你能说说这些吗？还是说这是个重大秘密？

Can you talk to those, or is is that a great secret?

说实话，最终来说，Fivetran 内部其实就是一场永不停歇的、疯狂的排错工作。

Well, it's it's you know, at the end of the day, what's going on inside Fivetran is just this crazy mass troubleshooting effort that never ends.

等等。

Wait.

这听上去跟每个初创公司都一样。

That sounds like every startup ever.

我们确实就是这样。

Well, we have it is.

但对我们来说，这个问题的范围要大得多，因为我们有750个连接不同数据系统的连接器。

But it it's it's the the breadth of it is much larger for us because we have 750 connectors to different systems of record.

从Oracle到SAP，再到Qualtrics，只要是主流系统，我们都有。

Everything from Oracle to SAP to Qualtrics to you name it.

每个系统都有自己的独特问题。

They all have different idiosyncrasies.

这些毛病只有真实客户遇到时才会暴露出来。

And you only discover these idiosyncrasies when real customers bump into them.

是的。

Yeah.

它们通常表现为性能问题、数据错误或者直接失败。

And they show up as performance problems, correctness problems, and failures.

我们一直以来的解决方式，我总说，其实没什么秘诀。

And, you know, the way we have have always solved it is I I always like to say the trick is there's no trick.

就是背后下了大量功夫，说白了，这是一种经济上的取巧。

It's just a lot of effort behind the scenes and it's kind of an economic trick.

我们只需要修复每个bug一次，之后所有使用这个连接器的客户都能从中受益。

We only have to fix every bug once and then every customer who uses that connector benefits from it.

你可以想象，AI编程助手就像是取之不尽的初级工程师资源。

But you can imagine how you can use AI coding agents, are basically an infinite supply of junior engineers.

对于这种问题来说，这确实是个特别有用的工具。

That is a particularly valuable tool for this kind of problem.

是的。

Yeah.

但要把这套方案真正落地，细节上其实相当棘手。

And the the the details of putting that into practice turn out to be quite tricky.

当然。

Sure.

不过最近几个月，我们真的开始看到效果了，规模化的改进也逐渐显现。

But we've really, especially the last couple months, started to see it work and started to see, you know, like like improvements at scale.

无数微小的改进累积起来，我们终于看到问题开始得到缓解。

Many many many small improvements start the we've seen the flood start to come.

没错。

Right.

我认为，正因为如此，今年Fivetran的质量和稳定性还会再上一个台阶。

And I think you'll see the quality and reliability of Fivetran take yet another leap this year because of that.

你知道，你处在非常独特的视角，因为你的客户包括像OpenAI和Anthropic这样的大实验室。

You know, you're in this very unique vantage point because you have the big labs as customers, like OpenAI and Anthropocure customers.

他们是原生的AI公司，站在最前沿。

You know, they are AI native, They're at the forefront.

他们使用Fivetran的方式，跟传统企业有什么不同吗？

Do they use Fivetran differently than traditional enterprises or anything that the enterprise can run

从这方面来看？

from that?

没有。

No.

他们的使用场景其实很典型。

Their use cases are very typical.

他们用 Fivetran 把来自众多不同数据源的信息同步到一个集中的数据湖里。

They use Fivetran to replicate data from lots of different systems of record into into a centralized data lake.

然后用这些数据做分析。

And they do analytics with that.

再把这些数据作为输入，接入他们自己的内部 AI 工作流。

They feed that as contacts into their own internal AI workflows.

所以他们的数据基础架构，其实和很多其他公司的非常相似。

So they they have built data foundations that look very much like the data foundations of many other companies.

Anthropic 的数据平台，就是由一位顾问搭建的，这位顾问之前也为多家公司设置过 Fivetran 和 dbt。

The data the the systems at Anthropic, one of the people who helped set them up was a consultant who had set up Fivetran and dbt and many other companies.

是的。

Yeah.

所以他们的数据平台看起来非常典型。

So their their their datas their data platforms look very typical.

我觉得这一点非常重要。

And I think this is a very important message.

如果你正在考虑为人工智能搭建数据基础，千万别误以为必须打造一个全新的、特别复杂的系统来作为AI的数据底座。

If you are thinking about data foundations for AI, do not make the mistake of thinking you need to build some exotic new system as a data foundation for AI.

最适合AI的数据基础，很可能就是你现在已经有的那个。

The right data foundation for AI is probably the one you already have.

如果你用的是比较现代的数据平台，比如Snowflake、Databricks、BigQuery，或者已经迁移到了Iceberg数据湖，上面跑着相应的计算引擎，那这些都完全可以作为你AI应用的绝佳基础。

If you have a reasonably modern data platform, something like Snowflake, Databricks, or BigQuery, or maybe even you have transitioned to an iceberg data lake with those compute systems running on top, that is a great foundation for your context for AI as well.

你知道吗，以前总有一种说法，我们之前聊Fivetran的时候也提到过，更广泛地说，大家觉得AI会把整个基础设施都变得廉价化，对吧？

You know, there there used to be this idea, again, we've touched on in in the context of Fivetran, Fivetran, but more broadly, there's this idea that AI commoditizes infrastructure broadly, right?

所以当时的想法是，既然AI什么都能写，那基础设施就没那么重要了。

And so the idea was like, well, it can write anything.

但实际情况恰恰相反。

The opposite seems to be true.

现在写出来的软件比以往任何时候都多，而且这些软件问题不少，反而更需要底下有稳定可靠的基础设施支撑，所以大多数基础设施公司因此都受益了。

More software's being written than ever before, the software's actually pretty buggy, it needs kind of stable infrastructure below it, you know, and so most infrastructure companies have seen a lift as a result of this.

从这个角度看，这种现象是暂时的吗？未来的AI消费者级基础设施终会到来，还是说——我简单说下我的看法：能长期稳定运行的高质量软件，恰恰不是AI最擅长的，你不如把精力放在别的地方。不过，我的观点会不会太保守了？毕竟AI未来的能力会越来越强。

You know, in sense, is this transitory, like the eventual AI consumer's infrastructure's coming, or do you act let me just give you my quick view on this, which is building rock solid software that you can operate for long periods of time is just not what AI is best at, and you're probably better spent focusing it on other things, But is my view blinker than, like, how powerful it's gonna get over time?

没人知道它未来会强大到什么地步。

No one knows how powerful it will get over time.

不过话说回来，如果它真变得足够强大，这些问题就都不重要了，因为我们直接进入一个物资极度丰富的世界了。

I mean, the nice thing about that is that if it gets sufficiently powerful, all these questions become sort of moot because we'll just be living in a post scarcity world.

但我觉得从当下来看，AI确实主要是在推高对基础设施的需求，而不是让它变得廉价和通用。

But I think if we look at at the present day, I think it is mostly true that AIs are are just creating more demand for infrastructure and not commoditizing it at all.

没错。

Yeah.

你可以把基础设施看成多层结构，最底层是数据中心。

You can think of infrastructure as having layers and at the bottom are like data centers.

再往上是像AWS这样的云服务商，然后还有像Convex这样的无服务器平台，它们让使用云服务变得更简单。

And then, you know, you go to cloud vendors like AWS, and then you have systems like Convex sort of serverless platforms that try to make the cloud vendors easier.

再往上，还有一些更高层次的抽象系统，比如Databricks这样的平台。

And then you even have, you know, systems that set up at a higher level of abstraction than that, which could include, you know, Databricks.

它们很多业务其实都是在托管笔记本环境。

A lot of their businesses hosting notebooks.

对吧？

Right?

我觉得最后一层才是最受AI威胁的。

I think that last layer is the one that is is threatened by AI.

是的。

Yeah.

那种感觉。

The perception there.

没错。

Yeah.

确实。

Yeah.

毫无疑问。

For sure.

AI在处理这类事情上确实很在行。

AI is is quite good at navigating Yeah.

基础设施会稍微复杂一点。

Slightly more complicated infrastructure.

所以如果你有个AI代理，可能就不需要最用户友好的那一层了。

So if you have an AI agent, maybe you don't really need that very most user friendly layer.

你可以直接用下一层就可以了。

You can drop down to the next one and use that.

对。

Yeah.

我的意思是，每当消费层变得不确定时，互联网当年也是这样。

I mean, you could argue that whenever the consumption layer is up for grabs, which also happened with the Internet.

对吧？

Right?

那时候你得去不同的地方做不同的事。

Like, you kind of went to different places to go do things.

是啊。

Like Yeah.

它改变了用户界面。

It changed the UI.

它改变了很多东西，但核心基础设施还是保持不变的。

Like, it changes a bunch of stuff, but, like, the core infrastructure stays in place.

对吧？

Right?

操作系统还在，芯片还在，数据库也还在，它们是慢慢演进，而不是被彻底取代？

Like, you still have operating systems, you still have chips, you still have databases, and, like, you they kind of evolve over time rather than they get replaced?

是的。

Yeah.

可能只去掉一层，也可能去掉三层。

Maybe you peel one layer, or maybe you peel three.

没错。

Exactly.

你不可能把所有层都彻底掀掉。

You're not gonna peel it all the way Yeah.

再往下说

Back down to

没错。

That's right.

我得先查个时间，因为跟你聊天实在太愉快了，简直能聊到天荒地老。

Let's talk let me actually just do a quick time check here just because I just enjoy talking to you so much, we can just talk forever.

好的。

Okay.

咱们换个话题，聊聊 dbt 的并购吧。

Let's change topics a bit to the dbt merger.

对。

Yeah.

你们在2025年收购了Census，接着又收了Tobico Data和SQL Mesh，后来还跟dbt Labs达成了合作。这个行业一向并购频繁，但要说新兴公司里最积极的，Fivetran绝对是头一个。你能说说背后的策略和规划吗？还是说这些收购都是临时起意？

So you acquired Census in 2025, and then Tobico data and SQL mesh, and then you signed with dbt labs, and so I mean this has always been a space that's been relatively acquisitive, but I would say for the new style companies, Fivetran has been the most acquisitive, so maybe can you talk through the strategy and the plan, Or is this was this ad hoc?

有没有一个长远的蓝图？

Is there some grand strategy?

我父母都是投资银行家出身，

Well, I I am the child of investment bankers,

所以也许我不过是实现了自己的宿命。

so maybe I'm just realizing my my destiny.

不只是投资银行家，

Not not not just investment bankers.

是私募股权，

PE.

对吧？

Right?

我哥哥做私募股权，我表亲还有家里好些人也都干这行。

Well, my brother did PE and my cousin and a bunch of other people in my family.

这不就对上了。

There you go.

但我的父母没干过这行。

But not not my parents.

我妈妈是商业银行的，我爸爸做并购投资。

My mother was a commercial banker, and my dad was a m and a investment.

这就对上了。

Oh, there you go.

明白了。

Okay.

并购，威尔斯那家。

M and a, Wells.

是的。

Yeah.

对。

Yeah.

不过是在石油行业里做的。

But but in the in the oil industry.

哦，原来如此。

Oh, okay.

我明白了。

I see.

我明白了。

I see.

不过话说回来，我一直在听。

But, anyways, I I listen.

看着你们做初创公司，亲眼见证你们执行这个策略，真的让人印象深刻。

For for a startup and having watched you do this, it's actually been very impressive to watch you run this strategy.

但说真的，这根本不是我们一开始打算做的。

Well, but but seriously, it it was not something we set out to do.

Fivetran 没有企业并购部门。

I Fivetran does not have a corp dev function.

我觉得每一次

I felt with every single

我从来没想过这一点。

I never really thought about that.

确实如此。

That's true.

对。

Yeah.

对。

Yeah.

我一直觉得，任何收购或合并，第一次大的就是HBR。

I I've always felt that any acquisition or merger, and the first big one was really HBR.

是的。

Yeah.

我记得很清楚。

Remember very well.

这种决定应该基于独特的原因，而且感觉像是你最后一次这么做。

It should feel like it's for these unique reasons, and it feels like it's the last one you're ever gonna do.

但你以后肯定还会做，只是做决定的理由必须非常充分。

And it's not gonna be the last one you ever do, but the reasons to do it should be really really strong.

是的

Yeah.

你不该主动去寻找这样的机会。

You shouldn't go looking for this.

不过，我们确实多次找到了充分的理由。

Nonetheless, we have found these strong reasons several times.

我觉得 dbt 这个很合适。

I I think the dbt one is a great fit.

这两个产品一直以来几乎总是被一起使用。

These are two products that have historically almost always been used together.

它们俩天生就搭配得非常好。

And they they they kinda go together.

Fivetran 是把所有数据集中到一处的工具。

A Fivetran is the tool that gets all your data in one place.

dbt 则是用来整理数据，把它转化成能精准反映你业务特点的模型的工具。

Dbt is the tool that you use to organize it and turn it into a model that reflects the particular details of your business.

然后，这些数据就会供给所有的数据使用者。

And then that is what feeds into all of the data consumers.

你知道吗，你曾经公开说过，dbt 将会是代码智能代理最大的受益者之一。

You know, you you've actually said publicly that dbt is gonna be one of the the biggest beneficiary of coding agents.

你能稍微展开说说吗？

Can you kinda pencil that out a bit?

是的。

Yeah.

对。

Yeah.

所以现在有个很好的情况是，

So there's this great

我其实不太明白那是什么意思。

I didn't really know what that means.

我觉得，dbt 的使用量会大幅增加。

There well, I think I think there's gonna be way more usage of dbt.

我觉得编程代理会写出大量的 dbt 模型。

I think coding agents are gonna write tons of dbt models.

我跟你说过这事。

I told you that.

我们其实已经看到这种情况了。

We're actually already seeing that.

我的意思是，确实如此。

I mean, that's Yeah.

而且它也会成为最大的受益者之一。

And and and it's gonna be a great beneficiary.

我觉得迪杰斯特拉说过一句很精彩的话，就是那个迪杰斯特拉算法，对吧？

There there is this great quote from Dijkstra, I think, which is Dijkstra's algorithm, Dijkstra?

没错。

That yeah.

计算机代码本质上应该是人与人之间沟通的工具，而不仅仅是给计算机执行的指令。

That computer code should be seen as a means of communication between humans and only incidentally as an execution format for computers.

这一点在dbt项目中的SQL查询里体现得最为明显。

And nowhere is that more true than in SQL queries in dbt projects.

这是一种很好的方式，能清楚表达我们公司数据的规则。

It is a great way to express these are the rules of data at my company.

是的。

Yeah.

即使这些代码是AI写的，你仍然需要保留这份文档，它真实记录了你业务的运作方式。

And even if it's being written by AIs, you still wanna have that artifact that that is is a is a executable documentation of how your business works.

明白了。

Alright.

你在AI浪潮中担任一家规模不小的公司的CEO，这真是一种独特的体验。

So you have the pleasure of being the CEO of a relatively large company during the AI wave.

你亲身参与了编程工作，这是我们很多人都没能做到的。

You've gotten to factor writing coding, which a lot of us have.

你做了大量实践，比如你提到的NanoClaw实验。

You're running a lot of experience, like you mentioned, experiments, your your NanoClaw experiment.

不过，你现在在搞的是什么？

Though, you're now running what?

纳米机器人。

Nanobot.

纳米机器人。

Nanobot.

是的。

Yeah.

所以，这部分有多少是来自乔治——那个科学家、技术专家——的视角，又有多少是你作为一家大公司的CEO，觉得它真的实用？

So how much of this is, you know, George, the the the scientist, the techie versus like you actually view this like pragmatically useful for the the CEO of a large company.

对。

Yeah.

我不确定这是不是个好主意。

I don't know if it's a good idea.

但我就是忍不住。

I just can't resist.

是的。

Yeah.

对于想在业余时间写代码的CEO来说，编码代理特别方便，因为它们可以异步运行。

And coding agents are great for CEOs who wanna write code on the side because they work sort of asynchronously.

这样你就能让很多任务在后台同时运行。

So you can you can have a lot of things spinning in the background.

我现在手头有不少项目在进行。

I I have I have a lot of projects going right now.

你能说说都是哪些吗？

Can you can you name them?

我的意思是

I mean

比如说，我有些项目纯粹是当爱好在做，像我那个管理网球团队的系统。

I I mean, I have things I am just doing as hobbies like like the system for managing my tennis team.

我还在开发一个网球统计的机器视觉应用。

I'm working on a little on a on a tennis statistics machine vision app.

但我在 Fivetran 还有好多项目。

But then I have many things at Fivetran.

这些都是实验性的原型，我会跟别人分享，一起讨论。

They're all experimental proof of concepts that I share with people and we talk about.

我正在搞一个叫‘纳米数据湖目录’的东西，就是想解决用数据湖时必须额外搭个目录服务的问题。

There's a potential like nano data lake catalog that I have going that attempts to if you when you use data lakes, have to adopt this additional service called the catalog.

核心问题是：我们能不能让这个目录彻底隐形？

And it's the answer to the question, could we make the catalog invisible?

这就是其中一个例子。

So that's an example.

我纯粹是闲着没事，从零开始写一个经典的 OLTP SQL 数据库。

I'm working on just for the hell of it, a from scratch classic OLTP SQL database.

什么？

What?

我觉得这听起来太疯狂了，简直不可思议。

I I think That sounds really there is an well, it's crazy.

这太疯狂了。

It's crazy.

对吧？

Right?

但这个项目的核心恰恰就是不要有。

But it it the whole point of the project is no.

也就是说，它根本不需要被创建出来。

That being that being created.

是的。

Yeah.

它它

It it

它的目标是做一个像 SQLite 那样的东西，只不过底层用 S3 来存储。

it it's it's what it attempts to do is to be like SQLite except s three is the backing store.

哦，这挺酷的。

Oh, that's cool.

这主意真不错。

That's a great idea.

就是为了这个。

For it.

因为当你构建AI工作流时，会需要成千上万个小数据库。

Because when you build AI workflows, you have this need for like zillions of tiny Yeah.

数据库。

Databases.

这只是一个概念验证。

And it's a proof of concept.

它主要是想探索一下：靠足够的AI编程，能不能搞定一件根本不可能的事？

It's mostly an exploration of like, could you with sufficient AI coding just take on something absolutely ridiculous?

如果你在听这个，又想参与这个项目，那你一定是数据库专家，去GitHub上找我们吧。

If you're hearing this and you wanna work on this you are an expert in data bases GitHub.

我可不是。

I don't.

我不行。

I don't.

但如果你真想参与这个项目，来找我聊聊吧。

But if you wanna work on this, come to come come talk to me.

这太好了。

That's great.

也许你可以到 Fivetran 来做这个。

Maybe you can come do this at Fivetran.

你根本不用非得用我的原型。

You don't even have to use my proof of concept.

我觉得现在正是一个真正的机遇。

I think there's actually a real opportunity right at this moment.

我真的很遗憾，我们好像不得不一直用 Postgres。

I I really lament that we're sort of stuck with Postgres forever.

Postgres 实际上，跟很多人想的不一样，是很老的技术了。

Postgres, contrary to popular belief, is very old technology.

它根本不是一个好的数据库。

It is not a good database.

本科生写课程项目写的数据库都比Postgres强。

Undergraduates writing class projects write better databases than Postgres.

不是因为开发Postgres的人不够聪明，而是因为它太老了，积累了大量技术债务。

Not because the people who built Postgres were not smart, but simply because it was written a long time ago, it has a lot of technical debt.

我真的觉得世界应该开发一个新的生产级数据库，而不是没完没了地重新包装Postgres。

And I really think the world should create a new operational database rather than just endlessly repackaging Postgres.

但是

But

这个观点太棒了。

That's an amazing take.

对。

Yeah.

这又是我一个反主流的看法。

That's another one of my contrary.

其实，Postgres 很差。

Postgres is bad, actually.

我还得说，不行。

I I should also say No.

不行。

No.

不行。

No.

更好了。

Even better.

比如说，上数据库课的本科生写的数据库比

Like, an undergrad in a database course is writing better databases than

在很多方面，本科生在数据库课程里写的存储引擎，比 Postgres 的堆存储引擎还要强。

In in many respects, the storage engine that you would write as an undergraduate in a database course is better than Postgres' heap storage engine.

明白了。

Gotcha.

Postgres 的存储引擎，我觉得，今天那些推广它的人自己都不会承认这设计有多好。

Postgres' storage engine, I don't think you the creators of the the people who who promote it today would admit it is not a good design.

它一直都是东修西补过来的。

It's been patched up in a

很多地方。

lot of

但 Postgres 的存储引擎，

ways, but the the postcresist the postcresist storage engine

是的。

is Yeah.

我不会在这上面扯太远。

I'm not I'm not gonna digress on this.

现在问题多得数不清。

There's a million questions now.

但就像这场对话，你有没有觉得，这反而成了种干扰？

But but but I I will say just like this conversation, do you ever find this becomes a bit of a distraction?

这些事情真的让人眼前一亮，特别有趣，做起来特别带劲，你就会想，哎呀，我本该去干那个一对一的活儿，结果却坐在这儿办公室里。

Like, it is so dazzling and so fun and so interesting to work on these things, and you're like, oh, you know, I should be like doing that one on one, but instead I'm here in my office.

不会。

No.

我经常想这个问题。

I think about that a lot.

我得时刻提醒自己，得把这种冲动压一压。

I make sure, you know, you gotta you gotta like keep it at bay.

没错。

Yeah.

因为确实有风险，就是所谓的‘克劳德精神病’——这词儿我特别喜欢，但你真的很容易陷进去。

Because it it it is like, you know, there there is a danger of Claude psychosis, so which is a term that I I love, but you can just get sucked in.

没错。

Yeah.

不过呢，我时间还挺多的。

But, you know, I have a lot of time.

我没孩子，所以时间挺多的

I don't have any kids, so I have a lot

真的很多。

of time.

有孩子的人常说，养孩子就像多了一个自己。

I mean, people with children will will say that it's like it's like having a whole another self.

是的。

Yeah.

没错。

Yes.

如果你没有的话。

If you don't.

相信我。

Trust me.

我确实有体会。

And I do.

是的。

Yeah.

对。

Yeah.

我说啊，作为一位高管、首席执行官、创始人，我特别欣赏你的一点是你真的很善于反思。

I mean, one thing I do I do really appreciate about you as an executive, a CEO, as a founder, is you're actually quite reflective.

而且从我认识你以来，你一直都在做这样一个心理练习：假设董事会新派来一位CEO来整顿Fivetran，你会不会立刻就大刀阔斧地改？

And one of the mental exercises that you've been doing as long as I've known you was like pretending if you were a new CEO brought in by the board to fix Fivetran, like, would you immediately unwind?

说起来，新来的CEO上任后几乎都会这么干，我知道你经常这么想。

Which by the way, always happens when you bring in a CEO, And I know you kind of do this mental exercise.

那最近有没有哪件事让你觉得，要是真有外人来接手，我肯定得改一改？

So is there anything recently that you've thought, like, boy, if I was brought in to, like, run things, I would change this?

太多了。

Many things.

这确实是我经常在做的一个练习。

I that's an exercise I I do regularly.

我正想找个最近的例子。

I'm trying to think of a recent example.

说起来，不算太近的例子是，我确实做过一些简化定价的尝试。

I mean, a not so recent example is I did do some things to try to simplify pricing.

我觉得这些举措还挺成功的。

I I think those have been successful.

虽然对公司来说挺痛苦的，因为主要降低了小客户的定价。

They were painful for the company because they mostly cut prices for small customers.

是的。

Yeah.

我另一个常做的练习是问自己：其他CEO该怎么做？

Another exercise I do is I ask myself, what should other CEOs do?

这招挺难的，得逼自己去做那些重大又让人害怕的事。

It's hard it's a it's a trick to get yourself to to do things that are big and scary.

是的。

Yeah.

所以在dbt合并这件事上，我当时在考虑要不要认真对待，我常用的一个方法是问自己：如果换成Certar，他会怎么做？

So in the dbt merger, you know, when when I was reflecting on that, on whether that was a thing we should seriously consider, One of the tricks I use is I ask myself, what should what should Certar do?

Certar是Snowflake的首席执行官，哦，对。

Certar is the CEO of Snowflake Oh, yeah.

是的。

Yeah.

我们和他合作很久了。

Who we've worked with for a long time.

这真是个不错的思维训练。

It's a good mental exercise.

哦，对，想起来了。

Oh, That's right.

然后我就照着去做。

And then I just go do that.

是的。

Yeah.

那一刻我心里就有了明确的答案：和dbt合并，绝对应该。

And and that was like a clear answer in my mind was like merge with dbt absolutely

对。

Yeah.

没错。

Yeah.

就是靠这个思路来判断的。

By that by that framework.

虽然看起来规模很大、让人害怕，但只要你想象一下——

Even though it seems very big and scary, when you imagine Yeah.

如果是别人，这么做是个好主意吗？

Is this a good idea for someone else?

是的。

Yeah.

这样就能想明白了。

Then you can kinda get there.

是的。

Yeah.

也许这只是一个很轻松、老生常谈的最后问题。

May maybe just kind of a softball cliche final question.

但你知道，你在这种过渡期管理着一位CEO，拥有独特的视角。

But, you know, you you do have like the the unique perspective of managing a CEO during this transition.

所以首先，你最担心的是什么？

So first, what are you most kind of existentially worried about?

其次，你最期待的是什么？

And second, what are you most excited about?

我觉得我最担心的是，有一天，编程代理会变得太厉害，能把连接器写得特别好，人们就干脆自己动手了。

Well, I think the thing I most worry about is just, you know, that the at some point, the coding agents will get so good at writing connectors that people will just shift to DIY.

我觉得这对Fivetran是个真正的威胁。

I think that's a real threat to Fivetran.

有些业务可能更受或更少受这种影响，毕竟客户自己写代码也不是不可能。

Think some different businesses are more and less threatened by like maybe the customers will just vibe code it themselves.

这确实是我对Fivetran的担忧。

That is a thing I worry about with Fivetran.

如果真到了那一步，我们也会找到办法在这个环境中继续发展。

And we we will we will find a way to thrive in that world if we get there.

即使短期内会对我们自己造成损失，我们也会提供相应的工具，供人们去实现这些功能。

And we we we will provide the tools that you use to do that if that indeed becomes possible, even if it comes at a short term cost to ourselves.

但我真的非常担心这一点。

But I do worry a lot about that.

而最大的机会在于，AI为数据处理带来了全新的可能性。

And then the the biggest opportunity, I think, is that AI is just a whole new set of things to do with data.

现在，把所有数据集中起来并加以整理的需求，比以往任何时候都更迫切。

The need for getting all your data in one place, organizing it, is so much greater now than ever before.

我认为，在这个数据平台的下游，人们会需要一整套新工具，而我们，尤其是我们和dbt，正好最适合提供这些工具。

I think that there's a whole set of tools that people are gonna need on the other side of that data platform, and I think we, and especially we and dbt, are perfectly positioned to provide them.

太棒了。

Amazing.

乔治，每次请你来做客都特别开心。

Well, it's always a pleasure to have you, George.

谢谢你抽空过来。

Thanks for coming.

马丁，见到你真好。

Good to see you, Martin.

感谢您收听这期的a16z播客。

Thanks for listening to this episode of the a 16 z podcast.

如果喜欢这期节目，别忘了点赞、评论、订阅，给我们打分或写评价，也推荐给你的朋友和家人。

If you like this episode, be sure to like, comment, subscribe, leave us a rating or a review, and share it with your friends and family.

想收听更多节目，请前往YouTube、Apple Podcast和Spotify。

For more episodes, go to YouTube, Apple Podcast, and Spotify.

在X上关注我们@a16z，或在Substack订阅我们的频道：a16z.substack.com。

Follow us on x at a sixteen z, and subscribe to our Substack at a sixteen z dot Substack dot com.

再次感谢收听，我们下期再见。

Thanks again for listening, and I'll see you in the next episode.