Stripe 的约翰·科利森谈代理式商业将如何重塑互联网

另外，别忘了，过去很多网站都认为机器人是坏事。

And also, don't forget, lot of websites have historically bots used to be a bad thing.

它们一直对机器人抱有偏见。

They used to discriminate against bots.

还专门设置各种防机器人机制之类的。

They would have bot protections and things like that.

但现在情况变了，我们反而希望智能代理能顺利完成这些操作。

Whereas now, it's like, Okay, no, we actually want agents to be able to do this.

所以我们正在逐步取消那些防机器人措施，为合作的智能代理开辟类似‘快速通道’的机制，安全地传输一次性支付凭证之类的东西。

And so you're unwinding some of the bot protections and having a little TSA pre for the bots that we work with, a way to kind of securely transfer one time use payment credential and things like this.

背后其实有大量复杂的系统在运作。

There's a lot of wiring up beneath the scenes

这其实正好可以引出代理式商业中的B2B层面，因为这让我想到另一个问题。

This that businesses have to is actually maybe like a good spot to slightly seg over into the B2B element of agentic commerce, because this gets to another thing that I think about.

你知道吗，其中一种很有价值的训练数据，就是直接分析用户的鼠标行为。

You know, I know one of the things that a valuable training data is literally just looking at user mouse behavior.

对吧？

Right?

这样，理论上AI模型就能像真人一样去网站上点击按钮了。

Because then, theoretically, the AI model can go to a website and click the buttons just the way a human wants.

但我觉得这有点荒谬。

But it all it seems silly to me.

既然我们做网站本来就是给人用的，结果又要让机器人去爬取，那为什么不一开始就直接为机器人设计呢？

Like, ultimately, if we're building a human if we're building websites for human use and then a bot's gonna actually scrape them, why not just actually build it for bots in the first place?

这背后到底在发生什么？

What are we seeing there?

你看到零售业的现状了吗？根本体验正在被重塑，所谓的用户体验，那个‘用户’不再是人，而是机器人。

What are you seeing from retailers where the fundamental experience is it gonna be reshaped such that what we call the UX, the user experience, is not really the you is not the person, but the you is the bot.

这会不会彻底改变互联网的面貌？

And is that changing how the internet is just going to look?

你说到点上了，这正是当前AI领域最核心的争论之一，关于AI在实际应用中究竟该往哪儿走。

So you're getting to a very core debate that's happening in AI right now, like some of these kind of questions at the frontier of AI as to where things go for practical purposes.

但其中一个观点是，当考虑到AI在现实任务中的应用时，实际上正在上演一场竞赛：要么我们改造世界，让AI更容易使用，比如让网站更适合AI浏览等等；

But one of them is when you think about AI adoption for real world tasks, there's almost a race between maybe we make the world consumable by AI, and we make the websites friendly for AIs to navigate and things like that.

要么就是AI本身变得足够强大，能应对非结构化的电脑操作——你给它一个桌面、一个鼠标，哪怕系统没做任何兼容性设计，它也能自己摸索出怎么用；

Or maybe just the AIs get good enough at unstructured computer use that you give them a desktop and you give them a mouse and a And there's no backwards compatibility that we built, but they just end up figuring it out anyway.

我觉得在未来一两年，甚至三年内，这个问题就会有答案了；

And it feels like over the next year or two or maybe three, this question will be resolved.

现在，包括我们在内的很多人，都在努力让Stripe的控制面板、商家网站这类平台能被AI轻松理解；

Right now, a lot of people, including us, are working on making the Stripe dashboard, merchant website, things like that consumable by the AIs.

这本质上意味着，我们要把那些原本靠点击网页完成的操作，转成AI擅长处理的文本形式；

And what that means basically is text where the AIs are very good at working with text.

所以，很多以前需要在网站上点来点去的事情，现在都需要提供文本接口；

And so you need a textual interface for a bunch of things that previously might have been clicking around on a website.

抱歉，我插一句，

Sorry, just to

说清楚一点。

be clear.

这让我想起当年关于App Store里的应用该不该用拟物化设计的争论。

This reminds me a little bit about the debate whether apps in the app store should be skeuomorphic or not.

不过说清楚一点，Stripe其实是押注在另一个方向上——那些看起来专门为人类设计的网站，可能并不是未来该下注的地方。

But just to be clear, Stripe is kind of taking a bet on the other side that actually the websites that look like they're designed for humans will not be the way to place your where to place your chips?

这确实是一个选择。

Well, it's an option.

对吧？

Right?

因为如果AI在操作电脑方面进步得很快，那这方案就站得住脚了。

Because the AIs get really good at computer use quickly, then it's good.

你明白吗？

Know?

那样的话，答案就出来了。

Then we have that answer.

你其实已经能看到一些苗头了。

And you're seeing this a little bit.

当人们用Open Claw去买东西，它在自家的Mac mini上运行时，商家可能会说：哎，我没做智能代理界面啊。

When people get their Open Claw to go buy something and it's running on their Mac mini at home, the merchant might say, oh, but I don't have an agentic interface.

但问题是，我们已经买了啊。

It's like, well, we bought it anyway.

那个爪子自己跑过去完成了操作，还用了无头浏览器。

The claw just went and did it and used a headless browser.

所以你根本不需要特意去适配这个计算机自动操作的世界，因为它的核心就是完全兼容现有的系统。

And so you don't need to do anything to enable the computer use world because the whole point is that it's fully backwards compatible with the current world.

但与此同时，我觉得人们还是想要这种超强的智能代理体验。

But in the meantime, I think people want to have these really powerful agentic experiences.

这其实是一场竞赛，对吧？

And it's a bit of a race, right?

我觉得企业们很清楚，谁先抓住这波需求，谁就能占得先机。

I think businesses correctly want to be the first ones to capture all this demand.

因为我想，如果你在AI应用里看到结果，它说：喏，这里有好几个选项……

Because I think it'd be the case if you start getting presented results in an AI app, and it's like, Okay, here are a bunch of options.

这个直接点一下‘立即购买’，两天就到货了。

This one, you can just click Buy Now, and it'll arrive in two days.

或者这个，你得自己去网站上到处点来点去，烦死了。我觉得消费者会越来越倾向于选那种一键搞定的，就像在亚马逊上，第三方卖家要是标了‘亚马逊Prime’，大家就更愿意买。

Or this one, you have to go to the website and click around and blah, blah, I think you'll start seeing some consumer preference for the ones that click one, kind of like on Amazon when there is a seller that has Amazon Prime in the third party sellers.

你选它的可能性大得多。

You're way more likely to pick that.

亚马逊确实做过这方面的研究。

And Amazon has done studies on this.

你可以想象，在AI应用里也会出现类似的情况：那些早早支持智能购物的商家，会比不支持的跑得快得多。

You can imagine an analogous world within the AI apps where the businesses that support agentic commerce sooner outperform the ones that don't.

可以

Can

你能说说，智能购物会怎么影响价格吗？

you talk to us about how agentic commerce could impact prices?

我们都清楚，现在人们买东西前做研究，经常会输入类似这样的指令：帮我找一双最便宜的运动鞋，还得符合这些标准。

And we all know that already when people are researching for products, you might put in a prompt saying, like, find me the cheapest pair of, I don't know, sneakers that fit the following standard.

你可以想象，会有越来越多的消费者倾向于选择最便宜、最能满足需求的选项。

And so you could imagine a world where lots and lots of consumers are being driven to the cheapest possible option for their needs.

同时，你也可能看到商家开始推出更复杂的定价策略，比如动态调价之类的，这让我忍不住想，未来是不是就会变成机器人之间没完没了地讨价还价。

And you could also imagine a world where maybe shops start to respond with more sophisticated price pack architecture, maybe dynamic pricing or something like that, which makes me wonder if the future is just going to be bots negotiating with each other over price endlessly.

我觉得这

I think it's a

确实有点难说，毕竟大家都能猜一猜。

bit I mean, we can all speculate.

但现在下结论还太早，因为这还属于新兴现象。

I think it's a bit too early to have realized data there because one, it's an early phenomenon.

其次，目前AI应用的前沿用户群体收入普遍较高，未必能代表未来长期稳定的市场格局。

Obviously, secondly, the leading edge of AI apps is quite high income, which is not necessarily representative of what the durable equilibrium will be.

我们经常讨论的一点是，微交易现在很可能已经变得可行了。

The one that we think about a lot is microtransactions are probably gonna be now possible.

对。

Yeah.

所以，可能会出现一大块以前根本不存在的商业形态。

And so you probably have this whole segment of commerce that previously wasn't double.

因为微交易过去从来行不通，主要是做决定时的心理负担太大了。

Because microtransactions, they never work or never have worked historically because of the the kind of mental load of deciding.

卖每篇彭博文章0.4美元根本没意义，等你纠结完要不要买，还不如直接办个日卡或者订阅服务呢。

There's no point at selling Bloomberg articles for $0.40 each because by the time you've decided, you might as well just have bought a day pass or a subscription or something like that.

不过要是彭博能提供一个正规的API就好了，尤其是搞金融相关的东西，写代码的时候能直接调用彭博数据，那真的太方便了——毕竟终端界面根本不是给AI用的。

However, if Bloomberg, and actually, I would love if Bloomberg had a proper API for their data, because if you want do anything involving finance, when you're vibe coding, it'd be really handy to have the Bloomberg data and the terminal is not exactly the right interface for AI.

要是能按SIP计费卖彭博数据就好了，比如你查一下霍尔木兹海峡的情况。

But selling Bloomberg data by the SIP, so it's like, okay, we're doing a query about the Strait Of Hormuz.

然后系统自动调取相关的彭博数据。

Let's pull in the Bloomberg data for that.

每次查询就付一小笔费用。

And you pay whatever small amount for that one little data request.

这种微交易，才是真正能落地的。

That's a microtransaction that could actually happen.

所以我们觉得，现在时机终于成熟了，因为过去人类在做决定时的心理负担实在太大。

So that's the big effect that we think the time is finally there because the mental load has traditionally been too high for humans.

但人工智能就喜欢过度思考。

But AIs love overthinking things.

它们乐于反复推敲每一个小决定，那我们干脆就利用这一点。

They love scrutinizing every little decision, and so we might as well do that.

这一点我也一直在考虑。

This is something I've been thinking about as well.

如果微支付变得毫无障碍的话。

If microtransactions become frictionless.

对吧？

Right?

那人们会不会开始把越来越多的内容都拿来收费，以至于你基本上每看一点都要付钱？

Do people try to monetize more and more content such that basically, like, you're kind of paying a

这就像交税。

tax Yeah.

用来浏览互联网？

To to browse the Internet?

对了，约翰，我想告诉你一件事。

By the way, John, I just wanna let you know.

我不知道你知不知道，但彭博社其实把一些重要的战略问题交给了我和特蕾西。

I don't know if you know this, but Bloomberg has actually delegated some of these big strategic questions to me and Tracy.

所以他们明确说了，不行。

So we could they've actually said, no.

应该由播客主来决定，这样我们才能好好讨论这个话题。

The podcasters should get to decide so that we could actually talk about this.

如果你们能跟迈克和团队说一声，我就希望我的编程助手能直接跟彭博社对接，处理数据，那就太好了。

If you guys could put word in with Mike and the crew that just, I would like my coding agent to be able to talk to Bloomberg and to do data process, that'd

我们现在正跟他频繁沟通呢。

be We're really messaging him right now.

不过，这又正好引出了我下一个问题。

But no, this actually, once again, perfectly segs to my next question.

我这些年听软件开发者朋友常说一个说法，就是‘免费’有两种：一种是免费如啤酒，一种是自由如言论。

One of the things for I hear from I've heard for years from my software developer friends, they use this binary free as in beer or free as in speech.

对吧？

Right?

比如有些软件你可以免费下载，这叫‘免费如啤酒’；但‘免费’也可以指你可以随意使用、修改，这就是‘自由如言论’。

So there's, like, software that you can download for free, and that's, you know, that's free as in beer, but then free can also mean sort of, like, you can do anything you want with it, free as in speech.

我在想，是不是该再加一种——‘自由漫游’式的免费？

I'm wondering on this point whether we need a third sort of one, free as in roaming, free roaming.

因为每一条数据，理论上对模型开发者都有价值。

Because every piece of data is theoretically valuable to a model maker.

而你和数据的每一次互动，可能叫元数据，这种行为本身也非常宝贵，等等。

And every interaction with a piece of data, which maybe is called metadata, that interaction activity is also very valuable, etcetera.

既然每一次互动、每一个版本都这么有价值，我们最终会不会走向一个不再有‘自由漫游’的互联网？还是说，我们只能不断支付那些十亿分之一美分的微交易？随着上网、浏览网页，这些信息都在被不断暴露，理论上这些小额付费迟早变得可行又合理。

And I'm curious, given that every interaction and any version of that is valuable, are we gonna be at the end of this sort of free roaming Internet, or we were we always just be paying these billionth of a cent micro transactions, which theoretically become doable and reasonable as we just sort of go along the Internet, surf the web, because this is all very valuable information that's theoretically being exposed.

我觉得不

I don't think

人工智能正在终结免费浏览互联网的时代，因为一直以来都有很多专有内容，比如报纸的付费墙。

AI is causing the end of the free roaming Internet because there's always been a lot of proprietary stuff, newspaper paywalls.

我不记得曾经能免费浏览彭博社的内容，哪怕我现在用着AI，也得花钱。

I don't I don't remember a free roaming Bloomberg that I could read without paying even I'm a using the AI.

你说得对。

You're right.

办个订阅吧，老兄。

Get the subscription, man.

你可以导入数据。

You can import.

直接用终端就行。

Just get the get the terminal.

不行。

No.

不好意思。

I'm sorry.

对。

Yes.

我觉得

I feel

这么说吧，我觉得你真的很想要我们的终端。

like half of this, I would say, I know you really want our terminal.

不过，抱歉，你继续说。

Anyway, sorry, keep going.

对。

Yes.

一直以来，消费端都有很多付费内容，都是藏在付费墙后面的。

So there's always been a of proprietary stuff on the kind of consumer side where this content is only available behind a paywall.

再举个例子，像彭博、FactSet、FMP，还有其他好多家数据提供商。

And then again, to the data point, there's Bloomberg and FactSet and FMP and all these different providers.

所以我觉得，所有这些提供商都在想，怎么才能让AI用上这些内容，又不至于免费开放。

And so I think what all the providers are thinking about is, how do I make this stuff available to the AI without just necessarily making it available for free?

所以对数据所有者来说，这通常表现为内容授权协议，对吧？

And so what that looks like with the data owners is content licensing deals, right?

Reddit 现在靠出售 Reddit 数据赚了不少钱，毕竟上面有大量特别有意思的人类讨论。

Reddit now makes a lot of money from selling the Reddit data because you have all this super interesting human discussion.

各种媒体出版商要么在起诉 AI 公司，要么和它们合作，甚至两者都在做。

You have the various media publishers either kind of suing or doing partnerships with or both the AI companies.

我们对数据提供商整体上非常感兴趣。

We are very interested in data providers generally.

我们曾经做过一个演示，展示如何用编码代理进行金融研究项目——它会自行购买所需信息，最终生成一份 PDF 报告。

Like we did a demo of sessions as to how you could do a financial research project with your coding agent buying the information it needs to put together the PDF answer.

不过我们在这个方向上还处于非常早期的阶段。

But we're quite early in that journey.

但说实话，长期以来一直都有大量专有数据存在。

But no, I think there's been lots of proprietary data for a long time.

所以现在关键在于，怎么把这些数据开放给 AI 代理使用。

And so it's more about how do we open that up to AI agents.

我不觉得免费互联网上的内容会因此大幅减少。

I'm not sure you'd see that much of a constriction on what's available on the free internet.

不过我要稍微反驳一下，互联网上以前也有很多免费内容。

But just to push back on that a little bit, there's also been a lot of free stuff on the Internet.

其中一部分原因是靠流量和广告支撑的。

And one of the reasons it's been free is because of traffic and advertising.

如果现在流量和广告减少了，因为大家都通过平台访问，不再直接跳转链接，那内容方就有动机开始对内容收费来弥补收入损失。

And so if that's not gonna get as much advertising or traffic anymore because everything is going through platform and people just aren't linking back directly, the temptation would be to try to make up that revenue by charging for the actual content.

是的，但这前提是AI想浏览的内容和人类想浏览的内容是一样的。

Yeah, but that assumes that what the AIs want to browse and what the humans want to browse will be the same thing.

我不确定情况一定会这样，比如那些面向人类用户的媒体平台，可能还是老样子。

And I'm not sure that'll necessarily be the case, where it'll still be the case, let's say, for media properties that is aimed at human consumption.

它们的广告收入应该还能维持得住。

And so they'll do fine with the advertising revenue they had.

也许它们还能从AI相关的业务里多赚一点额外收入。

Maybe they have some incremental revenue that they can get from the AI stuff.

还有啊，你说得也许对，那些特别差劲的聚合页面，比如什么‘十大榜单’，你一搜东西就跳出个SEO页面，满屏都是广告，点一下弹窗还占满整个屏幕，这种网站确实挺烦人的。

And then, I mean, maybe you're right that the really crummy aggregator pages where it's like top 10 lists, you know, you Google something and you get some SEO page and there's like a million tabula ads on us that you're trying to click through and the popover takes over your screen.

这些网站可能日子会更难熬，但我估计没人会真为它们感到惋惜。

Maybe those sites have a harder time, but I don't think anyone's going to really be too sad about that.

但你自己不也提过吗？我觉得你最近的一次演讲里就谈到过这个。

But you have talk you've talked about this yourself, I think, in a recent speech you gave.

这其实让我挺沮丧的，因为现在有了这些Vibe类的编程工具，我本来有很多想法，比如基于推特数据做点什么应用。

It's actually depressing to me because now thanks to the Vibe coding apps, there's a bunch of stuff I would have liked to build, say, around Twitter data.

对吧？

Right?

以前确实有个阶段，很多社交平台都开放API，挺开放的。

And there was an actual era in which a lot of these social platforms are very open with their APIs, etcetera.

可现在呢？全给关了。除非你愿意花几千美元，否则根本别想做个能操作别人推特账号的应用之类的东西。

And now everything's sort of shut down, so you can't even really you know, unless I want to spend thousands of dollars, I couldn't you couldn't really build, like, an app that does something with someone's Twitter account, etcetera.

我觉得你最近那场演讲里也提到过这一点。

And I think you talked about this in your recent speech.

这难道不是进一步促使所有人收紧限制，只允许人们自由浏览网站并大量抓取数据吗？

Doesn't this further accelerate the move of everyone to just sort of clamp down and allow people to roam their site and slurp up data?

是的。

Yes.

但澄清一下，人类正常浏览仍然可以免费使用，比如你依然可以免费浏览 Reddit。

But just to clarify, the human roaming has you know, you can still browse Reddit for free.

你也可以免费浏览 Twitter。

You can still browse Twitter for free.

他们限制的其实是某些情况下的数据抓取，没错。

What they're clamping down on in some cases Yeah.

主要是针对 AI 训练用的爬取行为，他们想对这种行为收费。

Is AI training scraping, where they want to put a toll on that.

也许他们原本没打算让这些信息被 AI 应用程序随意免费使用。

And maybe they didn't kind of intend for all that information to be kind of freely available to the AI apps.

至于这些数据是否应该免费提供给 OpenAI、谷歌这样的公司，或者他们是否应该付费获取，这本身是可以讨论的。

And there, you can argue whether that should be freely available to OpenAI and Google and everyone like that, or they should have to pay for it.

但像Reddit、X以及我们常去的那些地方，这些对我们人类来说至关重要的公共信息空间，并不会就此消失。

But it's not the case that our very important intellectual commonses of Reddit and X and all the other asylums that we like to frequent are going away for us as humans.

你好。

Hi.

我是P.J. Voat。

I'm PJ Voat.

我的播客搜索引擎推出了一档两部分的新系列，讲的是：在所有新兴的人工智能技术中，最具变革性的可能是无人驾驶汽车。

My podcast search engine has a new two part series for Of all the new technologies coming out of AI, the most transformative one might be driverless cars.

目前它们已经在美国十个城市的道路上运行，并且正在迅速扩展到更多地方。

They're already on the road in 10 American cities, and they're quickly coming to more.

我们将讲述我们是如何走到今天的。

We tell the story of how we got here.

揭秘谷歌内部一支花了十五年时间打造可能是路上最安全车辆的团队，同时也会探讨那些蓝领城市里，工会和政客们如何努力阻止这些汽车上路的纷争。

The secret team at Google that spent fifteen years building what might be the safest vehicle on the road, and we cover the fights brewing in blue cities where unions and politicians are working to keep those cars off the streets.

请在您收听播客的平台搜索‘Search Engine’收听。

Listen to Search Engine wherever you get your podcasts.

你知道吗，说到微交易，这话题人们已经聊了几十年了。

You know, you talk about micro transactions, and people have been talking about micro transactions for literally decades probably.

你举的例子，比如为什么不能花五毛钱买一篇文章呢？

And the example you gave, oh, why can't we just buy an article for 50¢?

我觉得这事儿一直没起来，原因有很多。

I think there's various reasons why that never took off.

其中一个就是交易成本的问题。

But one aspect is just the cost of a transaction.

理论上，尤其是有了稳定币之后，交易成本几乎可以忽略不计。

Theoretically, particularly with stablecoins, you could have almost free transactions.

那你有没有看到稳定币在企业间交易中被实际使用？毕竟成本这么低。

Are you seeing actual use of stablecoins in sort of B2B transactions because the sheer cost is so low?

稳定币在这个场景里到底扮演什么角色？

Where do stablecoins come into this?

你说得对，如果是按次付费的模式，从经济角度来说，不可能每次使用都走一次银行卡交易。

Yeah, you're right that when you have this pay by the SIP model, you obviously just can't, economically speaking, have an individual card transaction per usage.

顺便说一下，这正是人工智能经济的商业模式，因为每个令牌的边际成本都很低。

And by the way, this is the economic model of the AI economy because you have marginal costs for each token.

所以根本没人提供无限使用的服务套餐。

And so no one really offers an all you can eat plan.

你可以订阅ChatGPT、Clawd或者类似的服务，选个套餐，但迟早会用完配额。

You can subscribe to ChatGPT or Clawd or something like that, and you'll have a plan and you'll eventually hit usage limits.

或者你也可以按API调用付费，先充值一些积分，然后慢慢消耗掉这些令牌。

Or you can pay on an API basis and you'll top up with some amount of credits and then burn down the tokens.

但归根结底，所有人都在按使用量收费。

But everyone charges per usage at some level.

那接下来该怎么办呢？

And so then what do you do?

我们大量采用的常见模式就是这种按使用量计费的方式。

And the common mode which we power a lot of is this usage based billing.

我们其实收购了一家叫Metronome的公司，专门用来追踪积分使用情况。

We actually bought a company called Metronome to power that where you track credits.

所以这种电信计费方式，你把卡片信息留着，系统就会自动扣减积分。

And so that telco billing, you have your card on file, and then you track credits.

我们对你说的这种用稳定币来做支付的方式非常感兴趣。

We are quite excited about what you're referencing, which is stablecoins for this.

你是在问现在这种做法的使用量有多大吧？

You're asking how much are we seeing of it right now?

目前还不多，但我们正在全力搭建相关的基础设施。

Not that much yet, but we're building out all the infrastructure for us.

我觉得稳定币真正有潜力的地方在于，你不用再一个个去跟这些服务提供商建立关系。

The place where I think stablecoins could be really exciting is you don't have to then go have a relationship with all these services.

如果你只是想随便做个网站，觉得‘行吧’。

And so if you want to vibe code a little website and you're like, okay.

明白。

Yeah.

你得去Cloudflare买域名，再用Vercel托管，诸如此类，你肯定也经历过，得一个个注册账号、完成各种设置。

I'm gonna buy the domain from Cloudflare and host on Vercel and all those things, you know, you've had the experience of you have to go and sign up for all these services and do these things.

我们觉得，注册这么多服务的体验根本不该存在。

And we think that experience of signing up for all these services should go away.

所以我们正在打造一种无需再做这些繁琐注册的方式。

And so we are making it where you don't need to do that.

通过稳定币支付，可以按每次网页请求或极小的金额来付费。

Paying via Stablecoins could be one way where you pay per web request or per individual very small amounts.

你就不必一个个去注册这么多不同的服务了。

And you don't have to go sign up for a whole bunch of different services.

我明白，电商服务其实是一整条光谱，从简单填个表单，到给代理一笔预算，让它全权替你打理生活、自由支配这笔钱，都算在内。

So I realize that ejected commerce is a spectrum of services, everything from just simply filling out a form to giving an agent a certain amount of wallet spend and saying, go off and figure out my life for me and spend the following on whatever.

大家也总爱聊机器人失控、跑偏的事儿。

Everyone also likes to talk about bots going wild, right, and going rogue.

所以我想，如果你给代理一笔钱，让它按某种方式花出去，指令的详细程度肯定有高有低。

And so I imagine if you are giving an agent a certain amount of money and telling it to spend in a certain way, instructions, specificity can vary.

但万一出问题了，或者你根本不喜欢代理做的决定，该怎么办？

But what happens if things go bad or what happens if you're simply dissatisfied with the choice that the agent makes?

这该怪谁呢？

Whose fault is that?

也就是说，责任到底该由谁来承担？

Like, where does the liability actually lie?

我觉得这种担忧被夸大了，因为首先，代理在有人监督的情况下依然能为你完成大量有用的工作。

I think this concern is quite overstated because, one, the agents can do a huge amount of useful work for you while there's still a human in the loop.

比如，当云代码问你：‘我花这笔钱可以吗？’

And so imagine if Cloud Code asks you, are you okay if I spend this amount?

你只需要按个回车键就行。

And you just have to hit enter.

这比注册服务、在网页上到处点来点去的体验好太多了。

That's much better than the experience of signing up for a service and clicking around the web face and things like that.

即使你还要审核每一笔交易，也已经省下大量时间了。

So you've still saved a huge amount of time, even if you're reviewing the transaction.

同样，对于消费者场景，我们建议大家这样理解代理式商业：你始终在场，亲自决定买什么、是否愿意花这笔钱。

Similarly, again, the consumer use case, the way we suggest people visualize agentic commerce, is you still have a human in the loop deciding, making the final decision of both what thing I buy and am I Okay spending this much money?

但即便最终由你本人来确认这笔交易，AI在你的购物过程中依然能带来巨大帮助。

But the AIs can still be hugely beneficial as part of your shopping journey, even if you're as a human signing off on that transaction.

然后你就进入了授权委托的层面。

And then you get a bit into delegated authority.

就像我们在Stripe这家大公司，每年花掉几十亿美元，里面有很多授权委托的机制。

It's kind of like we run at Stripe a large company, and we spend billions of dollars a year, and there's lots of delegated authority.

你可能会问，要是有个别疯狂的人做出了错误的消费决定怎么办？

You might say, what if you have one crazy person who makes a bad spending decision?

这种情况我们偶尔也会遇到。

And we occasionally have that.

那你怎么解决呢？

And, you

说实话，我也不确定。

know, how do you solve it?

你知道的，这事儿怎么处理？

It's like, I don't know.

给每个人设个消费上限，要是谁干了特别离谱的事，就敲打一下，用合理的限额、授权机制和审批流程，问题自然就解决了。

Give people spending limits on how much they can spend and you, you know, whack them on the knuckles if they, do something really crazy, and you you just solve it with kind of sensible limits and delegated authority and approval process.

人和AI代理之间，道理也差不多。

And it's kind of the same with a human with an agent.

这让我想到一个问题，我想问问在AI时代管理一家大公司是什么感觉。

Well, this gets me to a question I wanted to ask about running a big company in the age of AI.

现在变化快得惊人。

Things are changing insanely fastly.

快。

Fast.

快地？

Fastly?

快。

Fast.

事情在变化，抱歉。

Things are change I'm sorry.

事情

Things

变化快得惊人。

are changing extremely fast.

去年大家特别期待的一些事，并没有像预期那样迅速降温；甚至到了2025年12月，有些东西突然火了起来，成了下一个热门方向。

And there are certain things that people were excited about last year that didn't fizzle out in the same way or certain things that people got really excited about even in December 2025 that became the next hot thing.

对Stripe这样的公司来说，你怎么判断该抓住哪个机会？

For Stripe specifically or for a company, how do you think about saying like, okay, here is a thing that we want to exploit.

你是想趁现在这个时机赶紧投入资源做出来，还是该把资源分配给探索未来24个月可能爆发的方向？

Here is a moment and we want to take advantage of it and we're going to build the thing for this moment versus we want to allocate resources for anticipating what's going to be the thing twenty four months from now, explore where this is going.

从企业角度看，在这么快的变化节奏里，你怎么平衡探索和执行之间的资源分配？

How do you think about the sort of resource allocation between exploring and exploiting in such a fast moving environment from the corporate standpoint?

我觉得你必须非常灵活，就像你说的，如果信息落后三个月、六个月，那就完全过时了。

I would say you need to be very flexible because as you were saying, if your information is three months out of date or six months out of date, it's just woefully old.

所以我们尽量保持灵活。

So we try to be pretty flexible.

我知道人工智能 somehow 激发了人们想要成为思想领袖的欲望。

And I know there's AI somehow brings out a desire for thought leadership in people.

没错，确实如此。

Yes, my word.

每个人都特别想突出自己的品牌。

Everyone's very brand pronounced.

是的。

Yeah.

而且说实话，我们现在的问题比答案还多。

And just, I don't know, we have more questions than answers.

我们只能边走边摸索。

And we're trying to figure it out as we go through.

所以第一点，就是要保持适当的谦逊。

So that's one, having the right level of humility.

第二点，你得有点‘AI妄想症’——剂量决定毒性：你得真心相信模型的能力，但同时也要能清醒地认识到它的局限。

The second thing is you kind of need to have the similarly the right level of AI psychosis, the dose makes the poison, where you need to really believe in the capabilities of the models, but you also have to be willing to, reason about the limits.

举个例子，AI在写代码方面特别厉害，能大幅提升效率。

And so to give you an example, really amazing at writing codes and huge productivity, uplift there.

如果你试着让模型做点财务分析之类的，不知道你们试过没有。

If you try to do anything, I don't know if you guys have tried to do anything like a financial analysis with the models.

它们在数字上简直一塌糊涂。

They're just horrible at numbers.

完全不懂算术。

Innumerate.

但它们能骗过你，因为能做出看起来很漂亮的表格。

And they can kind of fool you because they can put together something that they can build a really nice spreadsheet.

可你一检查就会发现，表格做到一半，数字全乱了。

But you're like, but the numbers just go completely wrong halfway through the spreadsheet.

说得太对了。

Oh, great point.

你完全正确。

You're absolutely right.

所以你需要对这些模型擅长什么、不擅长什么，有个敏锐的直觉。

And so you need to have kind of a bit of a fingertip sense for what the models are good at versus bad at.

我可以把公司分成传统三大块：研发、市场销售，以及行政和财务。

I think of this in you split up the company into the traditional there's R and D, go to market, and G and A.

如果你看看这几个不同领域，它们的应用方式其实很不一样。

And if you think about the different domains, they work quite differently.

在工程团队里，工程师们一直对工具和更好的工具充满热情。

Within engineering, engineers have always been passionate about tools and better tools.

但凡有点自尊心的工程师，都会积极拥抱人工智能。

And any kind of a half self respecting engineer is very eagerly adopting AI.

很少会遇到抗拒的人。

You don't get too many laggards.

因此，我们在Stripe的软件工程团队里看到了非常广泛的应用。

And so we have seen a huge amount of adoption within software engineering at Stripe.

进展特别快。

That's going really quickly.

我觉得我们还没到终点。

I don't think we're at the end of state.

但可以说，这套系统已经很好地嵌入了自我优化的循环，因为那些天天做工程的人，自然会想：怎么才能让这个流程更自动化？

But we're pretty like, that machine, I would say, is pretty well set up to be in this self improving loop because the people who are kind of doing the engineering day in, day out also think about, how can I better automate this process?

怎么用AI让事情做得更快？

How can I use AI make it work faster?

在Stripe，很多bug修复根本没人打开编辑器。

And so a lot of bugs that are fixed with Stripe, no one ever opens an editor for them.

他们直接点一下界面里的‘修复这个bug’按钮，事情就解决了——这才是真正的软件工程。

They just hit the fix this bug button in the interface and that's So done that's software engineering.

在销售和市场推广这边，最大的优势是，这是科技公司乃至所有企业里最可量化的领域。

On sales and kind of go to market, what helps you is that it is the most measurable domain inside of tech companies and companies broadly.

所以你能清楚衡量某个销售的具体产出效率，但对软件工程师、律师这类岗位，任何指标都很容易被Goodhart定律扭曲。

And so you can reason about the specific productivity of a specific salesperson in a way that you can't with a software engineer or a lawyer or something where any kind of metric there is very subject to Goodhart's Law.

而市场推广这一块，数据是实实在在能算出来的。

Whereas with go to market, it's measurable.

所以我们在Stripe至少看到，我们的销售团队对AI特别热衷。

And so we have seen, at Stripe at least, our sellers are super into AI.

他们非常积极地在采用这项技术。

They're adopting it very eagerly.

虽然没法直接证明是AI提升了生产力，但数据确实表现得非常好。

And you can't ascribe causality in the productivity numbers, but the productivity numbers are very good.

我觉得最难推进的地方，是在其他职能部门。

The place where I think it's trickiest is within the other general functions.

比如法务、财务、风控这些领域。

You think about legal and finance and risk and things like that.

这些地方其实已经在大量使用AI了。

And there, there is a lot of AI happening.

但问题在于，首先得做大量基础设施的工作。

But it's just tricky because, one, there's a huge amount of infrastructure work that needs to happen.

说白了，我们得先把财务数据整理成AI能用的格式。

It's like, Okay, we actually have to get this financial data into a form where it's usable by the AIs.

以前我们都是把数据严加保管的。

And we used to keep it under lock and key.

现在却得把它整理成AI能用的格式。

And now we need to load it in a way that the AI can use.

所以我们当时在Stripe投入了大量基础设施建设，既要让数据能被AI使用，又要确保安全可控。

So we had a big infrastructure build at Stripe to make the data usable to the AIs, but also keep it locked down appropriately and things like that.

就像我之前说的，AI在不同领域的能力参差不齐。

And then, like I said, the AIs are just uneven in which domains they are good at.

如果是编程相关的问题，或者能转化成编程问题的，AI表现得非常好。

And so if it is a coding problem or if it could be turned into a coding problem, very good at that.

公开的税法和法律条文，AI特别擅长，因为这些内容它训练时都接触过。

The public tax code and the public legal code, they're very good at that because it's text that they've been trained on.

但要是处理Stripe内部具体的财务数据，或者一些私有的内部信息，就很难取得理想的效果。

But working with, I would say, specific financial data at Stripe or kind of specific internal stuff that we have, that's just kind of harder to get good results out.

这并不是做不到，而且模型未来会进步很多。

It's not impossible, and the models will improve a lot.

但我觉得，要在这上面取得出色成果，是最具挑战的。

But I would say that is kind of the trickiest domain to get great results.

我们回到一开始聊的话题，那时候我们说到，互联网本质上是靠商业驱动的。

So just going back to the beginning of this conversation, we talked about how the internet basically runs on commerce.

如果这种商业的本质在发生变化，那互联网肯定也会受到相应影响。

And if the nature of that commerce is changing, you would expect some impact on the Internet.

我们之前问了你很多相关的问题。

We've been asking you various questions about that.

但我现在想直接问你一句：

But I'm just going to ask you very directly.

在智能代理驱动的商业世界里，你对互联网的总体愿景是什么？

What is your sort of broad vision of the Internet in the world of agentic commerce?

我们对智能的爆发感到非常兴奋。

We're very excited for the intelligence explosion.

在我们所关注的互联网领域里，创业活动正变得越来越活跃。

And so in our pockets of the Internet, we are seeing much more entrepreneurship happening.

这些数据都有体现。

That shows up in the numbers.

举个例子，今年第一季度Stripe上新成立的商家数量同比上涨了71%。

So to give you a sense, new business creation on Stripe in Q1 was up 71% year over year.

但一般来说，这些企业并不会增长得这么快。

That business generally is not growing that quickly.

这说明前沿领域的企业创建活动明显增多。

It's like the leading edge numbers are seeing much more business creation.

其他数据也能看到类似趋势。

You see this in other numbers too.

过去很多年，App Store上的应用上线数量一直很平稳。

So the App Store number of app launches was quite flat for many, many years.

但现在它的增长势头也跟上了。

And now it's seeing similar growth to that.

所以，智能革命带来的成果会非常多。

And so there are going to be many, many fruits of the intelligence explosion.

但一想到有更多人开始创业，我们就特别兴奋。

But we get quite excited about the idea that you have more people creating companies.

这或许会让经济更有活力，对吧？

And maybe you get more dynamism, right?

我们之前聊过一些影响，比如价格竞争可能会加剧。

We talked about some of the effects where you'll maybe see more price competition.

如果你在大公司工作，心里想着：我们这么做也太蠢了吧。

If you work at a big company and you think, hey, the way we are doing this is really silly.

我完全可以做得更好。

I could do it so much better.

现在，你辞职创业、推出更优产品的门槛变低了，因为你手里的资源更多了。

The bar to you being able to quit the company and develop a competing better offering is now lower because you have more power available to you.

你可以调动身边可用的智能工具来帮你。

You can recruit the intelligence that's available to you.

所以我们预测，经济活力会增强——这一点，我们已经在前沿数据里看到了。

And so our prediction is that you get more economic dynamism, which we're already seeing in our leading edge numbers.

你知道吗，我们现在看到初创公司越来越多了。

You know, we're seeing more startup formation.

不过我们之前聊过Stripe的活动，某种程度上，现在小公司更容易彼此协作，所以数量也变多了。

But we talked about this about Stripe sessions, you know, kind of in a sense, maybe you see smaller firms and more of them because it's easier for them to coordinate with each other.

约翰，我还有上百万个问题想问你，但时间确实不早了。

John, I have, like, a million more questions for you, but I'm aware of the time.

那我们就先到这里吧。

So let's just leave it there.

这次谈话真是太棒了。

This was such a treat.

非常开心能有机会和你聊天，下次一定再请你来。

Really enjoyed getting the chance to, talk, and would love to have you back sometime.

当然没问题。

Absolutely.

我非常期待再和你聊。

Would Would love love it.

是啊。

It.

特蕾西，我觉得刚才那番对话特别有意思。

Tracy, I really thought that conversation was super interesting.

我的意思是，光是开头就有很多内容了。

I mean, just to start, there is a lot here.

这事儿可不小。

This is a big deal.

我觉得接下来会往哪儿发展，比如商业直接参与，还有你提到的互联网更广泛的影响，这些问题都相当重大。

I think where it's going to go, the question of whether just commerce directly, and then also your question of the internet more broadly, there's some pretty big questions.

这些事情如何演变，背后的影响可不小。

Pretty big implications from which way some of this stuff unfolds.

对。

Yeah.

我很高兴约翰能体会到其中的乐趣。

And I'm glad that John appreciates the enjoyability

是的。

of Yeah.

对。

Yes.

我确实对智能代理的可发现性有些疑问。

Abs I do have questions about, I guess, discoverability when it comes to agents.

现在大家都熟悉这个例子：你打开Netflix的时候。

So everyone by now knows the example of you go on Netflix.

你第一次看的视频、电影或剧集特别关键，因为算法会围绕这个选择慢慢构建推荐逻辑。

The first video or movie, TV series you ever watch is incredibly important because the algorithm sort of builds itself around that one choice.

如果你一开始看了个无聊的东西，那以后它就一直把你归到这个单一的标签里。

And if you watch something stupid, then forever and ever, it's gonna, like, pinpoint who has this one demographic.

所以算法往往容易陷入一些明显的模式里。

So there is a tendency for algos to sort of cluster among, like, recognizable patterns.

想要发现新的东西、新产品，就变得越来越难了。

It becomes harder to discover new things, new products.

有些算法确实做得比其他的要好。

I will say some algos obviously do it better than others.

我觉得YouTube的算法在推荐新内容方面特别出色。

YouTube's algorithm, I find, like, actually really phenomenal for resurfacing or resurfacing new things.

我觉得大语言模型的内存功能是一把双刃剑，某种程度上，当它能了解一点用户背景时，确实更有用。

Well, LLMs itself, I find memory to be a very double edged sword, which is on some level, I do think it's more useful when it knows a little bit about the context of the user.

但有时候它会硬把六个月前的查询翻出来，硬扯回那个话题。

But there'll be times when it'll shoehorn any you did a query six months ago, and we'll try to bring it back around to that.

这里面大有文章。

I think there's a lot there.

这跟广告的未来也有很大关系。

I think there's a lot there with the future of advertising.

对。

Yeah.

抛开这些不说，大语言模型里的SEO对应物到底是什么？

But just setting aside, what is the equivalent of SEO in the LLM?

我们大概应该做一期节目，从那些必须认真思考如何在平台上展现自己的零售商的角度来看。

We should probably do an episode from the perspective of retailers that are really having to think about how they surface within

确实如此。

Totally.

电子商务就是如此。

So that ecommerce.

如果你是个零售商，现在还得同时针对机器人和人类做推广，这本身就是一个巨大的转变。

That in itself is a huge shift if you're a retailer and now you're trying to target bots as well as humans.

再往大了说，这个世界里互联网会变成什么样？我个人对互联网的整体走向一向很悲观，总觉得每项新技术的出现都会让情况变得更糟。

And then the overall question of what happens to the Internet in this new world, like, I tend to be very cynical about the overall direction of the Internet and just assume that every new advance in technology is gonna make it worse.

但你也可以想象一个未来，微交易变得无缝顺畅，结果人人都开始为自己的内容或API收费。

But, like, you could see a future where the ability to do seamless microtransactions basically means that, like, everyone starts charging for their content for their APIs.

这样一来，你上网其实得付钱了。

And so you actually have to pay a fee to be on the Internet.

这正是我想说的。

This is what I mean.

免费模式的终结，我觉得这是个挺有意思的问题。

Like, the end of free as enrollment, I think is an interesting question.

而且说实话，看到有人也觉得注册域名这件事值得解决，我感觉自己没疯，挺有共鸣的。

I also just personally, I find it very validating that I'm not crazy that actually being able to do something in a buying a domain name is something that people other people thought needed to be solved.

你懂我意思吧？

You know what I'm saying?

最近几周Stripe推出的几项功能，比如域名相关服务，还有直接通过API付费、不用再上网站输信用卡这些，都让我觉得特别有感触。

So some of these things seem very that Stripe announced in the last couple weeks related to both the domain name question, they're related to being able to pay an API directly rather than go to a website and enter your credit card and stuff like that for credits or whatever.

这些事我一直觉得又烦又重要，现在看到有人也在关注，我反而松了口气。

I found these things to be very annoying and big deals, and I'm sort of glad that I feel less crazy.

说白了，这些确实是当下互联网环境中让人头疼的现实问题。

Like, actually, these are annoying things within the new contemporary landscape.

没错。

Absolutely.

说到让人头疼的事，我觉得未来很大一部分场景，就是机器人之间互相谈判。

And speaking of annoying things, I also feel like a big chunk of the future is gonna be bots just, like, negotiating with each other.

在保险行业，我已经听说有人用人工智能批量拒赔了。

And I've already heard in the insurance sphere, you know, you've heard of insurers using AI to, like, bulk deny claims.

对吧？

Right?

我还听说一些新的AI初创公司，或者现有企业，正在用AI来对抗保险公司的拒赔决定，确实如此。

And then I've heard of new AI startups or existing companies using AI to contest the insurers Yeah.

当然。

Of course.

这种反制措施。

Pushback.

感觉以后全都是机器人在打交道了。

And so it just feels like it's just gonna be bots.

唯一能赚钱的，会是韩国的芯片公司，因为大家都得买更强大的芯片，还有旧时代的旧金山房东，毕竟OpenAI和Anthropic的员工总得有地方住。

The only moneymakers are going to be chip companies in Korea because they're everyone's gonna be having to get more powerful chips and old boomer landlords in San Francisco because the employees of OpenAI and Anthropic need a place to live.

还有医疗行业。

And health care.

是的

Yeah.

好吧

Alright.

我们就到这里吧？

Shall we leave it there?

我们就到这里吧。

Let's leave it there.

这是《奇思妙想》播客的又一期节目。

This has been another episode of the Odd Thoughts Podcast.

我是特蕾西·阿拉韦。

I'm Tracy Allaway.

你可以关注特蕾西·阿拉韦。

You can follow me at Tracy Allaway.

我是乔·威森塔尔。

And I'm Joe Wiesenthal.

你也可以关注我，账号是 the stalwart。

You can follow me at the stalwart.

别忘了关注我们的嘉宾约翰·科利森。

Follow our guest John Collison.

他的账号是 collision。

He's at collision.

还有我们的制作团队：卡门·罗德里格斯在 Carmen Erman，达什·奥·贝内特在 Dashbot，凯尔·布鲁克斯在 Kale Brooks，以及凯文·洛扎诺在 Kevin Lloyd Lozano。

Follow our producers, Carmen Rodriguez at Carmen Erman, Dash O'Bennett at Dashbot, Kale Brooks at Kale Brooks, and Kevin Lozano at Kevin Lloyd Lozano.

想了解更多 Odd Lots 的内容，请访问 Bloomberg.com/oddlots，订阅每日简报并收听所有往期节目。

And for more Odd Lots content, go to Bloomberg dot com slash Odd Lots for the daily newsletter and all of our episodes.

你还可以在我们的 Discord 频道 discord.gg/oddlots 里，24 小时随时讨论这些话题。

You can chat about all of these topics twenty four seven in our Discord, discord.gg/oddlots.

如果你喜欢 Odd Lots 这个节目，喜欢我们聊互联网的未来，麻烦在你常用的播客平台给我们一个好评。

And if you enjoy Odd Lots, if you like it when we talk about the future of the Internet, then please leave us a positive review on your favorite podcast platform.

另外，如果你是 Bloomberg 的订阅用户，可以免费收听我们所有的节目，全程无广告。

And remember, if you are a Bloomberg subscriber, you can listen to all of our episodes absolutely ad free.

你只需要在苹果播客里找到彭博频道，然后按那里的指示操作就行。

All you need to do is find the Bloomberg channel on Apple Podcasts and follow the instructions there.

谢谢收听。

Thanks for listening.

我是弗朗辛·拉克鲁瓦，一名获奖记者，现在推出了我的新播客《弗朗辛·拉克鲁瓦的领袖对话》，由彭博播客出品。

I'm Francine Lacroix, an award winning journalist, and I've got a new podcast, Leaders with Francine Lacroix from Bloomberg Podcasts.

我采访过从国家元首到时尚偶像的各种人物，聊他们对当下新闻的看法，但我一直很好奇，这些人在领导岗位上究竟是怎样的人。

I've interviewed everyone from heads of state to fashion icons about the news of the moment, but I've always been curious who are these people as leaders.

我觉得领导力没有唯一正确的做法。

I don't think there's one right way to be a leader.

做决定。

Make decisions.

一个糟糕的决定，也总比不做决定要好。

A poor decision is always better than no decision.

新节目每两周的星期一上线，记得收听。

Listen to new episodes every other Monday.

无论你在哪个平台听播客，都关注《与法兰辛·拉克鲁瓦共话领袖》吧。

Follow leaders with Francine Lacroix wherever you get your podcasts.