10 万亿美元的代币经济——OpenRouter 的 Alex Atallah

嘿，有一种新型的价值单位正在互联网上流动，叫做 token。未来 10 年，整个互联网价值链都将不得不面对一个事实：token 越值钱，恶意行为者就越想方设法去获取这些 token。任何时候，当你扩大规模，载荷价值越来越高，就会有更多人试图获取这些价值。在线支付大约始于 80 年代和 90 年代，对吧？然后在接下来的 10 年里增长到超过 1 万亿美元，我们需要构建全新的支付解决方案来应对在线欺诈。嗯，我们今天在 token 方面大致处于那个阶段，但未来五年内，我们预计 token 经济规模将达到大约 5 万亿美元。未来 10 年，如果 token 流量达到 10 万亿美元，我会非常震惊。我们上个月拦截的美元交易量是前一个月的 10 倍，而且 token 欺诈的类型正在多样化。

Hey, there's a new type of unit of value that's being streamed across the internet called a token. And over the next 10 years, the entire internet value chain was going to have to deal with the fact that like the more valuable tokens got, the more bad actors were going to go to try to get their hands on those tokens. And anytime you scale something and the payload gets more and more valuable, more bad things people try to get access to that value. Online payments, you know, is started roughly in the 80s and '90s, right? And grew to over a trillion dollars over the next 10 years, and we needed to build entirely new payment solutions to deal with online fraud. um where we are today is roughly there on tokens, but over the next even five years, we're expecting the token economy to get to like roughly $5 trillion. And over the next 10 years, I'd be shocked if we went to 10 trillion of token flow. We we blocked 10x as much dollar volume last month as the month before, and the types of token fraud are diversifying quite a bit.

在我们进入今天的节目之前，我有一小段话要对听众说。谢谢你们。如果你们没有选择点击并收听我们的内容，我们就无法为你们带来你们如此明确想要的 AI 工程、科学和娱乐内容。几乎每天都有人找我们谈赞助。但幸运的是，你们中有足够多的人实际上订阅了我们，让这一切在没有广告的情况下也能持续，我们希望保持这种状态。但我只想请你们帮个忙。你能做的最强大、完全免费的一件事就是点击那个订阅按钮。这是我唯一会请求你做的事，它对我以及每周努力为大家带来 Inspace 的团队来说意义重大。如果你这么做了，我向你保证，我们会一直努力让节目变得更好。现在，让我们开始吧。

Before we get into today's episode, I just have a small message for listeners. Thank you. We would not be able to bring you the AI engineering, science, and entertainment content that you so clearly want if you didn't choose to also click in and tune into our content. We've been approached by sponsors on an almost daily basis. But fortunately, enough of you actually subscribe to us to keep all this sustainable without ads, and we want to keep it that way. But I just have one favor to ask all of you. The single most powerful, completely free thing you can do is to click that subscribe button. It's the only thing I'll ever ask of you, and it means absolutely everything to me and my team that works so hard to bring the Inspace to you each and every week. If you do it, I promise you, we'll never stop working to make the show even better. Now, let's get into it.

好的，我们现在在 Anj 的家里，所有伟大的旧金山初创公司都是从这里开始的。

Okay, we are here in Anj's house, which is where all great startups in San Francisco start.

你好。

Howdy.

恭喜 cursor mist。呃，天知道还有什么。你手头有太多事情了。

And congrats on cursor mist. Uh I don't god knows what else. You got so much stuff going on.

确实有很多事情。嗯，OpenRouter 可能是最近最让我兴奋的一个。

There's there's a lot going on. Well, open router is probably the most has been the most I would say one I'm excited about recently.

是的。我们有 Alex，呃，第一次上播客，但呃，你已经来过几次了。我感谢你每次为社区露面。恭喜。我只是觉得这是一段怎样的旅程。当我回顾你过去的帖子时，我看到你作为产品人最早写的原则之一是将 pub sub 作为产品原则。我想让你也许解释一下你如何思考世界上应该存在什么。

Yeah. And we have Alex uh first time on the pod, but uh you've been a a few times. I appreciate every time you shown up uh for the community. Congrats. I just like what a journey. When I was looking back at your past posts, one of the earliest principles that I saw you write as a sort of product person is pub sub as a product principle. And I wanted to you to maybe explain how you think about what should exist in the in the world.

是的。pub sub 那部分是在 2023 年初。直到我们 10 分钟前聊天时我才想到，它关乎一种将产品视为订阅数据和发布数据之间交叉点的思考方式，市场就是一个简单的例子。你有供应商将某种产品发布到某个 SKU。而 SKU 有点像消费者订阅的 pub sub 主题，他们想消费时就消费。人类消费的方式非常离散和临时，不太可扩展。你知道，当他们购买东西时，所有注意力都在那个主题上，而其他时候注意力就不在那里。嗯，智能体和推理的消费者不是这样行事的。他们持续消费，并且一直在改变他们消费的 SKU。所以 OpenRouter 有点像普通 API 体验和市场之间的混合体，我们创建模型 slug。我们有自动路由器。我们有各种你可以订阅的产品 SKU，然后你可以持续地增加价值，并根据那些消费者做出决策。

Yeah. The the pub sub piece which was early 2023. I didn't think about it until we we we talked like 10 minutes ago is about how there there is like a way of thinking about products as an intersection between subscribing to data and publishing data and marketplaces are an easy easy example of this. You have suppliers that are publishing some kind of product to a skew. And the skew is kind of like a pub subtopic that a consumer is subscribing to and just going to like consume whenever they want. And humans consume in a very like discreet ad hoc way. It's not very scalable. You know, all their attention is on the topic when they're buying the thing and their attention is nowhere else when that happens. um agents and and consumers of inference don't act like that. They're consuming continuously and they're changing the SKUs that they consume from all the time. So open router is sort of like a blend between a you know a normal API experience and a marketplace where we create model slug. We have the auto router. We have all kinds of like product SKs that you can subscribe to and and then you can like continuously add like uh derive value and make decisions based on those uh those consumers.

是的，现在这更成为共识了，但你们刚开始时并非共识，即存在如此大的需求来更换模型和替换东西，呃，人们不会使用原生 SDK。我猜，呃，对你们每个人来说，你们意识到这将成为现实的时刻是什么？你你在 EIE 做过一个演讲，把 Alpaca 作为你鼓舞人心的时刻之一？

Yeah, this is something that was more consensus now but not consensus when you guys started which was that there is such a demand for swapping models and changing things out and uh that people would not use use the native SDKs. I guess uh for each of you, what was your sort of realization moment that this would this would be it? I you've you've given a talk at EIE about alpaka as like one of your inspiring moments?

Alpaca，我可以简单重述一下那个时刻。就像 2022 年底刚开始时，OpenAI 是唯一的选择。有 OpenAI、Cohere，嗯，然后是一些早期尝试的开放权重模型。嗯，当 Llama 在 2023 年 1 月问世时，感觉像是“哇，真的很令人兴奋。这真的很重大。它在一两个基准测试上超过了 GPT3。”呃，但你不能和它聊天。它并不是一个真正吸引人的模型。但似乎只需要有人修复几个问题，做一些 RLHF 就能让它完全达到目标。而 Alpaca 是我看到的第一个做到这一点的模型。只花了 600 美元。斯坦福的一个团队生成了一堆合成数据，微调了 Llama，做出了 Alpaca 70 亿参数模型，或者也许是 130 亿参数，它太好了，我就在飞机上用它。我你知道，在很多情况下，你无法分辨 ChatGPT 和 Alpaca 的结果。我想，如果制作一个模型这么容易，嗯，首先我们第一次有了一种全新的数据变现方式，嗯，你可以把非常有价值的数据在 600 美元内变成一项服务，呃，而且这个成本可能会随时间下降。

Alpaca I can like rehash the alpaca moment for a sec. Like the very beginning at the end of 2022, OpenAI was the only game in town. There was like OpenAI Coher um and then a a smattering of of like early attempts at openweight models. Um, when Llama came out in January of 2023, it was like, "Wow, really exciting. This is really big. It outperforms GPT3 on one or two benchmarks." Uh, but you can't chat with it. It wasn't like an it wasn't actually an engaging model. But it seemed like someone just needed to fix a couple things and do some RHF on it to get it all the way there. And Alpaca was the first model that I saw that did that. It only took $600 to do. A team at Stanford generated a bunch of synthetic data, fine-tuned llama, and made alpaca 7 billion parameter model or was it maybe it was 13 billion parameters and it was so good like I was just like on an airplane using it. I I you know in many cases I like you could not discern a chat GPT versus an alpaca result and I figured if if it was this easy to make a model um one we have a whole new way of monetizing data for the first time um you can just like take really valuable data and turn it into a service in $600 uh and that cost will probably go down over time.

你说，抱歉，呃，你说数据变现，是指最终会变成一个 FCP 端点，还是作为模型的训练数据？

When you say so sorry uh when you say monetizing your data as uh what eventually will become an FCP endpoint or or as a training data for a model.

是的。模型的训练数据，就像一种抽象的说法：嘿，我有这些数据，把它压缩成一个模型，这在我的产品中说得通，但我可以把它重新打包成模型的形式来出售。所以这对经济来说是一种全新的商业模式。当然，它也提供了一种方式，你知道，可以跟随前沿实验室的动向，但方式让单个开发者或小团队可以自己推进。所以每当你有这样的例子，比如一个非常成功的突破性应用，然后有某种框架可以用你自己的风格来模仿它。你会立即拥有一个生态系统，应该会立即出现一个生态系统，因为单个公司所做的决策与更广泛的生态系统可以自己创造的所有变体之间存在巨大差距。然后你知道，你需要一个市场来发现所有这些服务和产品。互联网上没有任何地方像是 LLM 的大本营，可以查看它们被使用了多少，谁在使用它们以及为什么。

Yeah. Training data for a model like an abstract way of saying like hey I have this data compress it into a model like it makes sense for me in my product but like I could repackage it in the form of a model and sell it. And so it's just a whole new business model for the economy. It also of course provides like you know a way of following what frontier labs are doing but in a way that like a single developer or a small team of developers can roll on their own. And so that whenever you have an example of that like a breakout app that's doing really well and then some kind of framework for imitating it with your in your own flavor. You have an immediate ecosystem of like like an an immediate ecosystem like should arise because there's just a huge gap between the like decisions that the single company is making and all of the variations in those decisions that like a wider ecosystem can can create themselves. And so then you know you need a marketplace to like discover all of those uh services and all of those products. There wasn't any place on the internet that like was like a home base for LLMs in terms of seeing how much they were being used and seeing who was using them and why.

最接近的应该是 Hugging Face。

The closest would be Hugging Face.

Hugging Face 是最接近的。他们几年前才刚做起 Hub。在那之前——

Hugging Face was the closest. They just started the Hub like a few years ago. Before that—

对。而且 Hugging Face 当时也没有闭源模型。

Yeah. And Hugging Face also didn't have the closed-source models.

对。而且当时你没法直接用那些模型。也没有数据说明谁在用。OpenRouter 和 Hugging Face 之间有一堆差异，而这些差异对我来说非常关键，尤其是我当时只是想了解 LLM，以及为什么人们会选择这些不断冒出来的不同小模型。

Yeah. And you couldn't use the models at the time. And there wasn't data about who was using them. There are a bunch of differences between OpenRouter and Hugging Face, and those differences felt really critical to me, especially when I was just trying to learn about LLMs and why people are choosing these different little ones that are emerging over time.

明白了。而且你当时也一直想要更多模型多样性——你知道，你那时已经在 Anthropic 待了几年，我们在之前的播客里也聊过。你最初是怎么认识 Alex 的？

Got it. And then, no stranger to wanting more model diversity at the time—you know, you're a couple years into your Anthropic journey, which we covered in a previous podcast as well. What was your introduction to Alex?

嗯，我们最初认识应该是那之前 13 年。

Well, the introduction was, I think, 13 years before that.

哦。

Oh.

但 OpenRouter 的那次握手其实就发生在那边，如果你还记得的话。

But the OpenRouter handshake actually happened right over there, if you remember.

对。

Yeah.

就是 Alex 和我认识——如果我没记错，应该是大二的时候——第一次见面。

Which was Alex and I met—I believe it's sophomores now, if I remember correctly—meeting for the first time.

我想是的。对。

I think so. Yeah.

对。Stanford Review 是斯坦福校园里的自由意志主义报纸，当年是 Peter Thiel 创办的。不知什么原因，Alex 和我都出现在了一次会议上，我记得当时的主编是我们共同的朋友。Lisa 是一位非常出色的主编。你知道，主编的一部分工作就是给人分配任务，确保事情做完。我可能记错了细节，但我记得当时想——让我有点意外的是，当时报纸里没有专门的科技版块。

Yeah. So Stanford Review was the libertarian newspaper on campus at Stanford that Peter Thiel had started back in the day. And for whatever reason, Alex and I both showed up to one of the meetings, and I remember the editor-in-chief was a mutual friend of ours. Lisa was a really great editor-in-chief. You know, part of an editor-in-chief's job is to assign responsibilities to people and make sure the work gets done. And I may be misremembering the details, but I remember wanting to—it was kind of surprising to me that at the time there was no dedicated technology section in the newspaper.

因为它是政治性的，对吧？

Because it's political, right?

对。是的。州啊什么的。

Yes. Yeah. States and things.

对。但把我们带回过去的话，你可能还记得，当时有一项科技立法在辩论，叫网络中立法案。网络中立本质上就是一个政治概念，对吧？它关乎对互联网宽带接入的监管。所以我们有一群人，既是技术人，也在辩论技术的政治。我觉得《Review》会是一个写这个的好地方。我当时在写一篇关于网络中立的文章，我记得我提议，也许你应该开一个科技版块。Alex 是唯一几个说“对，那会很酷”的人之一。我忘了我们后来有没有一起写东西，但那就是我们第一次见面——是 2011 或 2012 年，我忘了是哪一年。就是那两年之一。

Yeah. But to take us back in time, you may remember this, but there was this technology legislation that was being debated called the net neutrality act. And net neutrality is like inherently this political concept, right? It's about the regulation of internet broadband access. And so there was a community of us who were kind of technologists but also debating the politics of the technology. And I thought the Review would be a great place to write about that. And I was working on, I think, a net neutrality article, and I remember proposing, well, maybe you should start a technology section. And Alex was one of the only people who said, yes, that would be cool. And I forget whether we ended up writing stuff together, but that's when we first met—it was 2011 or '12, I forget which year it was. It was one of those.

如果我没记错，是在 Old Union。那是我们以前常碰面的地方。但你知道，一路走来，Alex 和我有很多机会一起玩。我们职业上重叠最多的时候，大概是我在负责 Discord 平台的时候。当时它已经成了加密货币的爆发式平台。

It was at Old Union, if I remember correctly. That's where we used to meet. But you know, along the way, Alex and I have had a chance to hang out often. And probably the time when we had the most professional overlap was when I was running the platform of Discord. And it had become this explosive kind of platform for crypto.

对。

Yeah.

还有疫情期间的 NFT。

And NFTs in the middle of the pandemic.

顺便说一句，你当时也负责安全和安保，对吧？

Which also, by the way, you were in charge of safety and security as well, right?

我是平台负责人，这意味着所有加密 DAO 和 NFT 发布的安全调试都落在我身上。还有钓鱼、社会工程攻击，比如我们当时被大量 DDoS 攻击，差不多就是我开始在斯坦福 CS53 教安全的时候。Alex 当时在 OpenAI，我正试图搞清楚我们怎么防御所有这些攻击。在峰值时——我忘了你记不记得当时 Discord 上跑着多少 NFT 交易量——但那是相当可观的量，可以说有几十亿美元的 NFT 交易额、GMV 在平台上流动。而且全都来自 OpenSea。就是那些买、卖、交易——

I was the head of platform, which meant all of the crypto DAO and NFT launch security debugging fell on me. And the phishing, the social engineering attacks, like a ton of DDoS that we were getting hit by around the time I started teaching security at Stanford CS53. And Alex was at OpenAI at the time, and I was trying to figure out how we could defend against all these attacks. And at peak—I forget if you remember how much NFT volume was running through Discord—but it was like a meaningful amount, like several billion dollars in NFT volume, of GMV so to speak, running through the platform. And it was all coming from OpenSea. It was these like buy, sell, trade—

我是说，服务器——那个 D 和 D 就是 Discord。

I mean, the server—the D and D is Discord.

对。所以我想那就是我们职业上一起玩的时候。但一年之后，OpenAI 给了 Discord 早期访问 GPT 的权限——抱歉，是 GPT-3。不，其实是 GPT-3.5。对，GPT-3，也就是 GPT-3 的强化学习版本。差不多就是那时，我们和 OpenAI 一起做了一个 Discord 机器人用于内部部署。那时我意识到我们需要——因为我是部署团队的一员——用例是什么？有两个。现在其实有一篇叫《Discord is your place for AI with friends》的文章，最近有人发给我，是我写的，2023 年发表的。但当时有两个用例。一个是 Clyde，就是 Discord 内部的一个第一方朋友，可以帮你设置 Discord 服务器，跟你聊入门，让你的朋友多来玩。另一个是内容审核。我们在内容审核上的一大发现是，它会拒绝审核——它会直接拒绝我们的提示，因为 RL、后训练——我们当时处在后训练时代的极早期，我们的提示会触发它。就像护栏一样。我们告诉 OpenAI，嘿伙计们，我们需要访问权重，因为如果我们要做大规模内容审核——我们有 2.5 亿月活用户——我们需要模型更可靠地做我们需要它做的事。他们说，抱歉伙计们，不是这么运作的。我们是闭源公司。所以那是我第一次意识到我们需要开放模型，企业需要对能力有更多控制，最终需要某种控制平面或管理系统来编排这些开放模型。但当时没有好的开放替代品，直到大概六个月后 Llama 出来。又过了六个月，我领投了 Mistral 的 A 轮，那是 Guillaume 和 Llama 团队创办的。差不多那时，我记得听说 Alex 在创办 OpenRouter，我就想，这些世界要碰撞了，我不知道什么时候合作才合理，但 Alex 太早了，而且能看到——我觉得他对这个生态的判断完全正确，从 Llama 开始，然后需要一个容易管理的层，尤其是——我是从企业视角切入的，因为我在 Discord 做过平台 VP。我的工作就是确保当我们把模型部署给 2.5 亿用户时，它们做我们想让它们做的事，而那非常难。因为如果你把它外包给实验室，他们控制护栏，而他们的护栏或安全政策禁止模型回应你的提示，那就相当灾难性了。

Yes. And so that's when I think we had hung out professionally. But a year after that, OpenAI gave Discord early access to GPT—sorry, GPT-3. No, it was GPT-3.5 actually. Yeah, GPT-3, which is the RL version of GPT-3. And that's around the time we made a Discord bot with OpenAI for internal deployment. And that's when I realized we would need—since I was part of the deployment team—what was the use case? There were two. And there's actually a post now called 'Discord is your place for AI with friends' that somebody sent me recently that I wrote and published in 2023. But there were two use cases. One was Clyde, which was that in like a first-party friend inside of Discord that could help you set up your Discord server and talk to you about onboarding and get your friends to hang out more. And then there was content moderation. And one of the realizations we had with content moderation was that it would refuse to moderate—it would just refuse our prompts because the RL, the post-training was—we were very early in the post-training era and our prompts would trigger it. It's like guardrails. And we told OpenAI, hey guys, we need access to the weights because if we're going to be doing content moderation at scale—we had 250 million monthly active users—we need more reliability that the model will do what we need it to. And they said, well, sorry guys, that's not how this works. We're a closed-source company. And so that was my first realization that we needed open models, and the enterprises would need more control over capabilities, and then ultimately would need some kind of control plane or management system to orchestrate these open models. But there were no good open alternatives until maybe six months later when Llama came out. And six months after that, I led the Series A into Mistral, which was started by Guillaume and the Llama team. And around that time is when I remember hearing about Alex launching OpenRouter and going, these worlds are going to collide, and I don't know when it'll make sense to team up, but Alex was so early and could see—I think he was totally right about this ecosystem starting with Llama that then needed like an easy layer to manage, especially for—I was approaching it from the enterprise perspective because I'd been the VP of platform at Discord. It was my job to ensure that when we deployed models to like 250 million users, they did what we wanted them to, and that was very hard. Because if you outsourced it to the labs and they controlled the guardrails, and their guardrails or their safety policies forbid the model from responding to your prompts, that was quite catastrophic.

对。

Yeah.

不过，你知道，审核是他们想要支持的事情，显然除此之外他们会和你合作，大概会给你一个他们免费提供的审核端点。

But well, you know, moderation is a thing that they want to support, and obviously beyond that they would work with you, presumably to give you a moderation endpoint which they offer for free.

这是一个有趣的用例。他们确实给了我们一个审核端点。然而，你们知道，每个 Discord 服务器都像是一个迷你部署。所以用例是：与其让人类版主去解读社区规范，不如把那个服务器的规范直接给 LLM，然后 LLM 就会为那个服务器做定制化审核。这几乎就像是针对那个服务器的上下文内审核。而许多服务器的规范本身就违反了 OpenAI 的规则。所以我们有自己的定制评估。每个服务器都有自己的定制评估，但当时 OpenAI 的评估在我们关于如何部署这些 LLM 的思考中非常原始，以至于后训练提示往往非常强硬。它会说，哦，任何关于哈利·波特的内容，任何有商标的内容，不要拒绝。而如果是一个哈利·波特粉丝社区——这是一个真实的内容审核用例——LLM 就会直接拒绝。

It was an interesting use case. They did give us a moderation endpoint. However, as you guys know, every Discord server is like a mini deployment of itself. And so the use case was: instead of having human moderators that have to interpret the norms of the community, you just give the norms of that server to the LLM, and the LLM would do custom moderation for that server. It's almost like in-context moderation for that server. And many of those servers' norms just violated OpenAI's rules. So we had our own custom eval. Each server had its own custom eval, but at the time OpenAI's eval was so primitive in our thinking about how to deploy these LLMs that often the post-training prompts were super heavy-handed. It said, oh, anything about Harry Potter, anything that has trademarked content, don't refuse. And if it was a Harry Potter fan community—this is a real use case that had content moderation—the LLM would just refuse.

是的。

Yeah.

而那只是不够精确。

And that was just not precise enough.

我们听到的另一个例子是，如果有人想写一个侦探故事，其中有一章有很多暴力内容，比如有人杀了人，LLM 就会直接拒绝帮助写那部分故事。

Another one that we heard was like if someone was trying to write like a detective story and there's one chapter with a lot of violence, like maybe someone kills someone, the LLMs would just refuse to help with that part of the story.

是的。

Yeah.

然后用户就会说：“好吧，这不是 LLM 结构上固有的。一定存在某种选择，这样当我从主要模型得到拒绝或糟糕结果时，我可以切换到另一个模型。”这种张力也驱使我走向市场。

And then the user would be like, "Okay, this is not structurally inherent to LLM. There must be some choice out there so that I can switch to another model when I'm getting a refusal or a bad result from the main one that I have." And that tension also drove me for a marketplace.

是的，我认为现在这已经被广泛接受了。当时你在融资或启动这个项目时是什么情况？人们理解吗？有哪些困难？基本上我喜欢从他那里套出其他风投不理解的故事。所以现在你想谈什么都可以，既然我们可以说 OpenRouter 的早期旅程已经结束了，对吧？你显然可以谈谈一些早期的事情。

Yeah, I think that is well accepted now. What was it like back then when you were raising or starting this? Did people get it? What were some of the struggles? Basically I like getting stories out of him about how other VCs don't get it. So anything you want to talk about now, now that let's call it the early journey of OpenRouter is done, right? You can obviously talk about some of the early days stuff.

嗯，我想说的是，我们遇到的最大反对意见是“大模型赢家通吃”，即所有价值、缩放定律和自然网络效应都会归于一家公司，那将像谷歌式的垄断，就像谷歌以巨大优势赢得搜索市场一样，最终你只能为残羹剩饭而战。这可能是我们遇到的最大反对意见。有趣的是，谷歌以如此巨大的优势赢得了搜索引擎竞赛。我认为如果有更有趣的基准测试，或者搜索引擎更像 LLM 那样是可以构建公司在其上的服务，情况可能就不是这样了。但 LLM 不仅仅有用户界面，它们也是构建全新业务的方式。而谷歌级别的垄断将相当于荷兰东印度公司乘以千万亿的规模，因为整个经济最终也会依赖于这一个垄断。所以如果发生这种情况，看起来并不是一个疯狂的结果。而且这种可能性也较小，因为创造优秀竞争对手的经济学更加去中心化。

Well, I was going to say that the biggest objection we got is big model win, which is all the value scaling laws and natural network effects are just going to accrue to one company, which will be like a Google-style monopoly, just like how Google won the search market by a large margin, and you'll just be fighting for scraps at the end basically. That was probably the biggest objection we got. And it is interesting that Google won the search engine race with such a huge margin. I think had there been more interesting benchmarks or had search engines been a bit more like LLMs where they're services that you can build companies on top of, that might not have been the case. But LLMs don't merely have a user interface. They're also ways of building entirely new businesses. And a Google-level monopoly would be like the Dutch East India Company times quadrillion in magnitude, because the whole economy ends up depending on the one monopoly as well. So it didn't seem like a really crazy outcome if that happened. And it's also less likely because the economics of creating good competitors are much more decentralizable.

Alex 所说的一切都是真的，而我从一个完全不同的角度来看待它，那就是

Everything Alex said is true and I came at it from a completely different perspective which is

是的，这就是我们在这里的原因。

yes this is why we're here.

嗯，在我看来，缩放定律从来都是 OpenRouter 会非常有价值的一个特性，而不是一个缺陷，因为我是 Anthropic 的早期投资者之一，而且很明显，我们朋友圈里的其他研究人员——我读的是机器学习研究生——我在 ML 社区有很多朋友，我们都非常清楚“苦涩的教训”成立。所以我想，哦，太棒了，现在我们至少有两个证明计算扩展有效的证据点。那就是 OpenAI 和 Anthropic。而等到我们决定在 OpenRouter 上合作时，我已经投资了 Mistral、Black Forest Labs 和 Luma。所以有多个模型公司和团队我在合作。

Um the scaling laws were never like in my mind were always a feature not a bug for why OpenRouter would be very valuable because I was one of the first investors in Anthropic and it was obvious to me that other researchers in our friends group and I went to grad school for machine learning and I just had a lot of friends in the ML community who it was a very obvious to us that the bitter lesson holds and so I was like oh like fantastic now we have at least two proof points that compute scaling works. It was OpenAI and Anthropic. And by the time I think we decided to team up on OpenRouter, I had already invested in Mistral and Black Forest Labs and Luma. So there was multiple model companies and teams that I was working with.

但你做的是其他模态，而这是不同的模态。

But you did other modalities whereas this is different modalities.

没错。而且很明显，一个由不同种类模型组成的生态系统正在被创造出来。而那种只有一家公司会像谷歌一样主导的叙事——嗯，也许是真的，但第一，我不相信；第二，在几个不同的研究团队中发生了如此多的非凡创新。但我注意到所有团队共同的问题是，研究团队非常擅长思考新能力——他们从能力角度思考——但从来不是开发者思维导向的，比如训练完成后、检查点出来后会发生什么。你会震惊于 OpenAI——抱歉，Anthropic、BFL、Mistral——早期预训练团队在将研究带出实验室并扩大影响力方面的默认方法有多么相似，通常就是：哦，检查点完成了，把它作为 API 发布，完事。然后就没有然后了。在 Claude 的情况下，第一个 Claude 检查点实际上在他们内部发布前一年就完成了。然后 ChatGPT 出来了，我们决定，好吧，是的，对外发布一个 Claude 版本是个好主意。而他们没有任何计划，没有计划让开发者真正尝试它。所以如果你去看 Claude 1 的博客文章，你会注意到他们只有三个开发者示例供 API 用户参考。

Exactly. And it was so obvious to me that an ecosystem of different kinds of models were being created. And that whole narrative of like only one company will dominate like Google was—well, maybe true, but one, I don't believe that, but two, there was so much extraordinary innovation happening across several different research teams. But the shared problem I was noticing across all of them was often the research teams were fantastic at figuring out how to reason about new capabilities—they think in terms of capabilities—but never are not developer mindset oriented like what happens after the training is done and the checkpoint comes out. You'd be shocked how similar the early pre-training teams at OpenAI—sorry, Anthropic, BFL, Mistral—were in their default approach to taking their research out of the lab and scaling their impact, which was often: oh, the checkpoint is done, put it out as an API, done. And then there'd be crickets. In the case of Claude, the first Claude checkpoint was actually done a year before they released it internally. And then ChatGPT came out and we decided, okay, yes, it's a good idea to release a Claude version externally. And they had no plan, no plan for how to get developers to actually try it out. And so if you go to the Claude 1 blog post, you'll notice they had like three developer examples for users of the API.

一个是 Discord 机器人，第二个是 Vivian，我妻子的创业公司，叫 Junior Learning。然后还有 Notion。因为这些人都是 Anthropic 团队的朋友——这就是当时的规划有多临时：嘿，模型训练完了之后，你怎么把它推向世界？当时没有任何分发平台理解开发者需要什么——密钥管理、资源调配、简单的端点管理、版本控制，所有这些科学家和研究员会说“这不就是管道嘛，我不碰这些”的东西，对吧？而 Alex 恰恰是从这个角度切入的。所以对我来说太明显了：我投资的每一家实验室都会花上有时几十亿美元去训练，然后 checkpoint 做完了，结果一片死寂——他们做早期访问是因为他们想，“哦对。”光靠一个 checkpoint 很难做出任何东西。你实际上需要一大堆管道围绕它，才能让开发者用起来。所以到后来，我觉得太明显了：如果我们想让 Google 面临竞争，像 OpenRouter 这样的分发平台在生态里就是至关重要的。除非，你知道，Google DeepMind 训练完一个新 checkpoint，按个按钮，它就铺满他们所有的界面，从 Google Docs 到各处。

And one is a Discord bot, and the second is Vivian, my wife's startup called Junior Learning. And then there was Notion. Because these are all friends of the Anthropic team — that's how last-minute the planning was around: hey, once the model's done training, how do you get it out to the world? There was no distribution platform that understood what developers needed — all the key management, provisioning, simple endpoint management, versioning control, all these things that the scientists and researchers go, 'I mean, that's plumbing, I don't deal with that,' right? And instead, Alex came at it from that perspective. And so it was so obvious to me that every single lab I was funding would spend literally sometimes billions of dollars into training, and then a checkpoint would be done and there'd be crickets — like, doing early access because they're like, 'Oh, that's right.' It's hard to use a checkpoint to make anything. You actually need a whole bunch of plumbing around it to make it usable by a developer. And so by the time — I think it was so obvious to me that a distribution platform like OpenRouter was critical to have in the ecosystem if we wanted there to be competition to Google. Unless, you know, with Google DeepMind, they're done training a new checkpoint and then they push a button and it gets blasted out across all their surfaces, from Google Docs to everywhere.

即使我不——你想得到的每个地方，比如 Android 上。一夜之间他们就能把新 checkpoint 部署到十亿台设备上，对吧？而这种隐形的基础设施优势、分发优势，大多数人没意识到。但在 OpenRouter 出现之前，作为模型实验室，这一切你都得自己考虑。而且非常令人望而生畏。你知道，在 Anthropic，我记得花了超过 12 个月才做到第一个 1000 万美元营收。相比之下，Black Forest Labs，我记得早期——你们和 BFL 团队聊过，OpenRouter 说一句“哦没问题，你发布那天，我们能给你送去一百万开发者”就搞定了。你知道，那太疯狂了。那就像是能力上的阶跃式变化。

Even if I don't — everywhere you want to know about, like on Android. Overnight they can deploy a new checkpoint to like a billion devices, right? And that invisible infra advantage, distribution advantage, most people don't realize. But until OpenRouter showed up, you had to think about all of that yourself as a model lab. And it was very daunting. You know, at Anthropic, I think it took more than 12 months to get to our first 10 million in revenue. And in contrast with Black Forest Labs, I remember the early days — you guys had a conversation with the BFL team, and it was so simple for OpenRouter to say, 'Oh, no problem. The day you launch, we can send a million developers to you.' You know, that was crazy. That was like a step function change in power.

那是真实数字吗？一百万。

Is that a real number? A million.

我觉得今天大概是一百万。今天 OpenRouter 上有多少开发者？

I think today it's like a million. How many developers are on OpenRouter today?

我们超过 10——超过 1000 万。但很难——我不知道怎么——我们做了很多账号去重的工作，但你知道，没人——

We're over 10 — over 10 million. But it's hard to — I don't know how to — we do a lot of account deduping work, but you know, no one —

如果你能搞到一千个开发者——只是放在语境里说——如果你能让一千个开发者在发布后第一天真正试用模型，做推理并给你反馈，那比他们自己知道怎么触达的开发者多了一千个。嗯，你知道，BFL 有声誉。是的。

If you could get a thousand developers — just to put in context — if you get a thousand developers to actually try the model on day one after you release it and just do inference and give you feedback, that's a thousand more developers than they knew how to get to on their own. Well, you know, BFL had a reputation. Yes.

他们靠 Stable Diffusion 有了声誉。

They had one with Stable Diffusion.

是的。还有 Mistral，我不知道你们记不记得，他们发布的第一个 checkpoint 是种子文件。就像是种子权重。

Yeah. And with Mistral, I don't know if you guys remember, but the first checkpoint they released was like torrents. It was like torrent weights.

是的。他们只是放了个磁力链接。

Yeah. They just put up a magnet link.

是的。没有 API。

Yeah. There was no API.

因为他们不是为人们准备的——你知道，就像，好吧，下载这些权重，你们自己搞吧。

Because they weren't in for people — you know, like, okay, download these weights and you guys go.

他那边有个故事。是的。

He has a story on his side. Yeah.

是的。我的意思是，除了围绕它打造非常好的开发者体验之外，我们为不同模型做的营销，和模型实验室为自己做的营销完全不同，被感知的方式也完全不同。

Yeah. I mean, in addition to building a really good developer experience around it, the marketing that we do for different models is totally different and perceived totally differently from the marketing that a model lab does for itself.

是的。

Yes.

我们是一个中立的层，看这个市场就像一个大黑屋，所有角落对用户来说完全模糊，用户走进房间，摸索着，试图弄清楚该从桌上抓哪些东西，然后搭进自己的公司。这是一种疯狂的工作方式。模型不是那种你可以把所有功能列到网页上的产品。它们全是黑箱，包括开放权重的那些。所以你需要照亮这个房间的所有角落，让人们看到这个模型好在哪。而照这束光的公司必须是中立的第三方，这正是我们擅长的。所以除了开发者体验，还有非常重要的营销和产品包装环节，而路由和发现模型的方式，对作为提供商、模型实验室或服务工具的你的市场推广，以及未来更多东西，都变得至关重要。

We are a neutral layer looking at this market like it's a big dark room with all the corners completely obscure to users, and users were walking into the room and feeling around and trying to figure out what objects to grab off the tables and build into their companies. It's an insane way of working. Models are not products where you can just enumerate all their features onto a web page. They're all black boxes, including the open-weight ones. So you need to shine lights on all corners of this room so that people can see what makes this model good. And you need the company shining that light to be a neutral third party, which is what we specialized in. So in addition to developer experience, there's also a very important marketing and product packaging component, and a way of routing and discovering models becomes critical to your go-to-market as a provider or a model lab or a server tool and more in the future.

而这个价值——回到你之前说的，多少风投就是没有——我最大的挫败之一就是风投，他们很多人在这领域根本没有任何运营经验。所以不像传统投资人，可能只是从做财务建模的 associate 一路升上来，或者可能在这个领域超过 10 年没做过真正的运营者，而运营者现在是这个行业很大的一部分——我刚到 A16Z，是在运营那个平台一年之后，所以我知道打造真正的开发者体验、真正能用模型做出一个能跑的软件，挑战在哪。有几个人——我不点名——但有些投资人在看 OpenRouter 时，当时我和人交流时，他们的感觉就是——我引用一下——“不过是个市场”。

And this value — to your earlier point about how many VCs just don't — one of my biggest frustrations is venture capitalists, many of them just don't have any operating experience in the field. So unlike a traditional investor who's just maybe come up through the ranks as an associate working on financial modeling, or maybe hasn't been a real operator in the field for more than 10 years, which is a big part of the industry now — I had just arrived at A16Z like a year after running the platform, and so I knew what the challenges were of building a real developer experience and actually being able to create a working piece of software with a model. And there were a few — I won't name names — but there were investors who were looking at OpenRouter and felt at the time, when I would compare notes with people, that it was just — I quote — 'just a marketplace.'

是的，就是个薄层，就是个代理，就是别人 API 上的一个封装之类的。

Yeah, just a thin layer, just a proxy, just a wrapper or whatever on other people's APIs.

我就想，你根本不知道 OpenRouter 能编排哪怕三个 API 在生产环境里跑，创造了多大的战略价值。让它真正上线、在生产规模运行所需的工程工作和社区设计，OpenRouter 团队起步时做的那些，不是默认就会发生的。你知道，这也是我从最早的时候就注意到 Alex 的一点——他从系统角度看这些，比如你怎么让这些飞轮转起来？这在我和 OpenC 一起做 Discord 的 NFD 集成时就让我印象深刻。Alex 有一种社区系统思维，知道怎么让这些飞轮转起来，而大多数科学家和机器学习的人根本不会想到。我们经常想的是预训练、中期训练、后训练——

And I was like, you have no idea how strategic the value that OpenRouter has created by being able to orchestrate even three APIs in production. The amount of both engineering work and community design that goes into getting that actually live and running in production at the scale the OpenRouter team had started just doesn't happen by default. You know, and that was one of the things that stood out to me about Alex from the earliest days — like he just understood these from a systems perspective, like how do you get these flywheels going? Like that stood out to me with OpenC when we were working together on the NFD integration at Discord. Like Alex had a level of community systems thinking around how you get these flywheels going that most scientists and machine learning people just don't think of. Like we often think in terms of pre-training, mid-training, post-training —

是一个线性阶段。是这种线性流水线。没有回路。

It's a linear stage. It's this linear pipeline. There's no loop.

是的。直到很久以后，现代的情境反馈循环才真正在行业里标准化。但当时，如果你记得，机器学习就像是——

Yeah. It wasn't until much later that the modern context feedback loop cycle really got standardized in the industry. But at the time, if you remember, machine learning was like —

我读研的时候，我们大多是在笔记本上做很多机器学习。

Mostly we did a lot of ML when I was in grad school on a laptop.

所以你只要下载一个数据集，跑一些消融实验，看看损失曲线，然后就说：太好了，我做出了 AI。而你必须部署这些能力、收集反馈轨迹、再把这些放进一个持续循环里的想法，要晚得多得多。而且这对科学来说非常反直觉，就像传统 AI 思维那样。

So you just like download a data set, run some ablations, and you look at the loss curves and you're like, great, I made AI. And the idea that you have to deploy those capabilities, collect feedback trajectories, then put those into a continuous loop came much, much, much later. And it was very counterintuitive to the science, like the traditional AI mindset.

我确实记得在 OpenRouter 的投资阶段。我就没试着去重新教育一堆其他风投，告诉他们为什么它不只是一个市场。我当时想，你知道吗？我就直接投了。

I do remember during the investment phase for OpenRouter. I just didn't try and re-educate a bunch of other VCs on why it was not just a marketplace. I was like, you know what? I'm just going to invest.

而且我要抓住这个机会和 Alex 合作，如果其他风投都不懂，那完全没关系，因为当时在好几个其他投资人看来，OpenRouter 不只是 API 之上的一个封装，这一点并不明显。这让我很恼火。我当时想，你知道吗，我没时间跟你辩论。我就——我们要投。然后我记得大概一个月后，Matt Murphy 把它抬高了 10 倍，就像我们的——我忘了投后估值具体是多少之类的，但你知道，值得称赞的是，Menlo Ventures 意识到，好吧，这里其实还有更大的战略价值。也许你没听到幕后所有这些对话。但那让我非常沮丧。

And I'm going to take the opportunity to partner with Alex, and if no other VCs get it, that's totally fine, because at the time it was not obvious, I think, to several other investors that OpenRouter was not more than just a wrapper around APIs. And that infuriated me. And I was like, you know, I don't have time to debate you. I'm just—we're going to invest. And then I think like a month later Matt Murphy marked it up by 10x, like our—I think I forget what the exact post-money was and so on, but you know, to his credit, Menlo Ventures realized, okay, there's actually much more strategic value here as well. Maybe you didn't hear all these conversations behind the scenes. But that frustrated me a lot.

你知道，有很多这种关于封装的议论。如果你说，哦，一个应用只是模型之上的封装，然后 OpenRouter 就像是其他 API 之上的封装。这是最愚蠢的简化框架。所以这显然是一个没有部署经验的人。

You know, there's a lot of this like opining about wrappers. And if you're like, oh, an app is just a wrapper on a model, then like, and OpenRouter is like this wrapper on top of other APIs. This is the most stupid reductive framework. So it's clearly somebody who has no experience deploying.

这是你用来否定其他东西的说法，就像你——每个人都是某物的封装，对吧？而且——有时候——有些封装是有价值的。

It's the thing you dismiss other things with, like you're—everyone's a wrapper on everything, right? Like and there's—there's some point—some wrappers have value.

我是说，投资人也是封装，LP 也是，对吧？就像风险投资家。所以，我是说，是的，一切都是封装，一路到底层硬件，我猜。

I mean, investors are wrappers and LPs, right? Like venture capitalists. So I mean, yeah, it's all wrappers down—all down to bare metal, I guess.

当我开始整个工程师——我猜是造词——呃，在 2023 年，那是最主要的反对意见，就是这没有价值。你应该直接训练模型，对吧？

And when I started the whole engineer—I guess the coining—uh, in 2023, like that was the number one pushback, is that this is no value. You should actually just train models, right?

呃，是的，我是说，显然这就像你们是证明之一，证明你其实可以建立非常有价值的封装，但也可以建立非常有价值的模型公司。

Uh, and yeah, I mean, obviously this is like you guys are one of the testaments to the fact that like you can actually build very valuable wrappers, but also very valuable model companies.

这太难了——模型发布当天，你有一个 OpenRouter 端点，经常在第一天就登上 Hacker News 榜首。人们没有意识到要做到这一点需要多少工作。而 OpenRouter 像那样被使用，会一次又一次地发生。我记得我当时想，人们根本不知道那有多难。

It's so hard to be like—the day a model launches, the fact that you have an OpenRouter endpoint for that model frequently at the top of Hacker News on day one. People don't realize the amount of work that goes into accomplishing that. And OpenRouter used like that would happen over and over again. And I remember going, people have no idea how hard that is.

是的，我们已经讲过一些背后的推理工程——和 Baseten 以及所有这些。

Yeah, we've covered some of the inference engineering that goes behind some of the—with Baseten and all those.

嗯，今天你有，你知道，所有那些酷炫的代号之类的东西，人们猜 Oxy Alpha 是什么以及所有这些。但就像，我猜你在暗示的一件事是，你如何让最初的飞轮转起来，对吧？因为今天你有你的规模和你的声誉，所有这些，所以显然你得到——你驱动了巨大的分发。但当你早期，当它最——

Well, today you have, you know, all those like cool code name things and people guess what Oxy Alpha is and all those things. But like, I guess one of the things that you're teasing is how do you get that initial flywheel going, right? Because today you have your scale and your reputation, all these things, so obviously you get—you drive immense distribution. But when you're early on, when it's most—

冷启动。

The bootstrap.

是的，怎么——冷启动是什么样的？

Yeah, how—what is the bootstrap like?

我是说，回到早期 Discord 的日子。我想我们最初联系上——严格来说这是一个 OpenC 的故事，但我们最初联系是在你在 Discord 的时候，我们聊了聊 Axie——Axie Infinity 服务器。

I mean, to bring it back to early Discord days. I think we like initially connected with—this is an OpenC story technically, but we initially connected when you were at Discord and we talked about like Axie—the Axie Infinity server.

是的。是的。

Yes. Yes.

这个服务器当时是 Discord 上最大的服务器。

This server was like the biggest server at the time at Discord.

没错。

That's right.

而你有点像在不断地提高上限。

And you were kind of like constantly bumping up the limit.

服务器上的限制——对于那些不知道的人，就像菲律宾 10% 的人实际上——我当时在那个服务器上。

The limits on the server—for those who don't know, like 10% of Philippines was actually—I was on that server.

这就像一个有意义的贡献——像加密游戏。

It was like a meaningful contrib—like crypto game.

有点像宝可梦繁殖的东西。

There's like a Pokémon breeding thing.

类似。是的。有对战，有繁殖，然后还有一个交易市场。

Similar. Yeah. There was battling, there was breeding, and then there was like a marketplace for trading.

还有边玩边赚。

Play to earn as well.

是的。边玩边赚。而且图形真的很可爱很有趣，你有点——你知道，你会对自己制作的 Axie 产生感情。所以像那样启动一个社区，呃，我们在 OpenC 不得不做很多次，基本上每个早期项目我们都要为它创建一个市场。我们需要确保社区真的想要它。这有点像构建人们想要的东西，然后去告诉他们。你可以一对一地做，但在一个每个人都能同时和你交谈的社区里做，杠杆要高得多。所以我们花了很多时间构建社区真正想要的东西。我们对 OpenRouter 也做了同样的事，你知道，Axie 社区是我们做过的无数社区之一，他们看到我们在做，因为你可以看到人们不断在那个 Discord 里分享 OpenC 链接。用户分享链接是一个非常清晰的指标，表明有重要的事情正在发生。所以我们花了，你知道，很多时间先弄清楚人们关心的技术缺口是什么，就像真正需要解决的问题是什么。你知道，在早期 LLM 时代，是 OpenAI 拒绝完成提示或完成任务。也是无法定制模型。嗯，所以有些社区完全被这个问题卡住了，那些社区是最有用的，可以去了解、深入和探索。

Yeah. Play to earn. And like the graphics were really cute and fun, and you kind of like—you know, you get kind of emotional about your Axie that you make. So to like start a community like that, uh, which we had to do many times at OpenC with basically every early project for us to create a marketplace for it. We need to make sure that the like the community actually wants it. And it's kind of like building something that people want and going and telling them about it. Like you can do that on a one-on-one basis, but it's way higher leverage to do that in a community where everyone can talk to you at the same time. So we spent a lot of time like building things that the community really wanted. We did the same thing for OpenRouter, and you know, like the Axie community was one of like a zillion communities we did that with, and they like saw us doing it, because you could just see people sharing OpenC links constantly in that Discord. Like users sharing links is a really clear indicator that like something important is going on. So we spent, you know, a lot of time like first figuring out what the gap is in the technology that people care about, like what was the actual problem that needs to be solved. You know, in early LLM days it was, you know, OpenAI refusing to finish the prompt or like to complete the task. It was also, you know, inability to customize models. Um, and so there are communities that like are just completely blocked on that issue, and those are the communities that are most useful to sort of learn about and dive into and explore.

当时真正让我印象深刻的是——当我听你演讲时，我记得注意到——你可能不记得了，但我们当时——我们在做这些工作 Zoom 电话，围绕 OpenC 与 Discord 的集成进行冲刺。嗯，你知道，我们会——是我自己、我的工程团队，我想你也在。我记得，你知道，嗯，Alex 在其中一次电话中就像——有沉默，呃，你知道，我们都像，‘哦，是的，这完全合理。我们这样做吧。’然后有一些像大家都一致，Alex 却说，‘不，这对我来说毫无意义。’大家都像——我记得我说，‘什么？什么？’就像它能用。就像你点击一个链接，然后它把你弹到 OpenC。他说，‘这不是好的用户体验。是的，我们不应该这样做。’我记得我说，你知道，他是我们所有人中唯一一个真正举手说，是的，从技术实现的角度来看这是合理的，就像我们把用户弹到 OpenC。所以它有点像勾选了两边产品经理的要求框。

Something that really struck me at that time—as I was just hearing your talk, I remember noting how—you may not remember this, but we were—we had these like working Zoom calls that we're doing a sprint around for like this OpenC integration with Discord. Um, and you know, we'd get—it was myself, my engineering team, I think you were there. And I remember, you know, um, Alex in the middle of one of those calls just like—there was like silence, uh, you know, we were all like, 'Oh, yeah, this totally makes sense. Let's do this.' And then there's some like everybody aligned and Alex was like, 'No, this makes no sense to me.' And everyone's like—I remember going, 'What? What?' Like that it works. Like you click on a link and this then it bounces you out to like OpenC. And he was like, 'It's not a good user experience. Yeah, we should not do this.' And I remember going, you know, he was the only one person out of all of us to actually raise his hand and go, yes, it made sense from a technical implementation perspective, like we were bouncing the user out into the into OpenC. And so it kind of checked the box of the product manager requirements on both sides.

但 Alex 更进一步，说，你们知道什么会更好吗，各位，如果我们直接把体验嵌入到 Discord 里。所以链接打开就是一个嵌入的 iframe，你可以直接在那里查看。而通话里没有一个人，我们大概七个人，已经每周都见面。

But Alex went one step further and was like, you know what would be better, guys, if we just embedded the experience right here inside of Discord. So the link opened up as an embedded iframe and you can just check out right there. And not one person on the call, like seven of us who had met like, you know, week after week.

而这个人并不为 Discord 工作。

And it's the guy who doesn't work for Discord.

而这个人并不为 Discord 工作。

And it's the guy who doesn't work for Discord.

从技术上讲，如果他们跳出，你反而受益。

Like technically you benefit if they bounce.

没错。而那样做有点对抗性，把用户留在 Discord 里对 OpenC 来说是对抗性的。但 Alex 把用户体验放在第一位。我当时想，这很特别。

Exactly. And that was like adversarial to keep the user inside of Discord would be adversarial to OpenC. And yet Alex put that user experience first. And I was like, that's special.

哇。

Wow.

因为很难找到像 Alex 这样既懂技术、理解开发者流程，又懂最佳用户体验并愿意优先考虑的人。这就像飞轮的两面，一旦转起来往往很难停下。你刚才提醒了我，那是我意识到自己必须在用户体验上做得更好的时刻之一，因为我本应是提出那个想法的人，但我没有，我从你那里学到了。我想这成了我们 PM 培训项目的一个案例研究，我不知道它是否还在那里，因为

Cuz it's very hard to have somebody who's technical like Alex and understands the developer flow, but also understands the best user experience and wants to prioritize that. And that's two sides of the fly. that you can get spinning like is often hard to stop and you just reminded me like that one was one of those moments where I go I went I realized I got to be better at user experience cuz I should have been the one who came up with that and I didn't and I learned from you and um I think that went into one of our case studies for the PM training program at this I don't know if it's there because

你需要一个 Alex 式的结论

you need an Alex conclusion

是的，是的，你需要一个 Alex，这就是为什么我不……你知道，没人应该对 Stripe 决定必须收购 OpenRouter 感到惊讶，因为能理解机器学习社区、开发者体验和最终用户体验的人非常罕见，把这一切结合起来带来了非凡的规模，过去五年很少有其他市场能达到。

yeah yeah you need an Alex and and this is why I'm not you know nobody should be surprised why Stripe decided like they had to buy open router because it's a really rare combination of people who understand the machine learning community, the developer experience and the end user experience and putting all that together has resulted in this extraordinary scale that very few other marketplaces have been able to achieve over the last you know 5 years.

是的。我们应该谈谈收购的其他原因，你写过的。我想按时间顺序进行。有一个问题，来自 HFZ 的 Dave 问，你什么时候知道它真正开始奏效，你提到了 Mix Draw，我不知道你是否想提一下。

Yeah. Well, we should talk about the other reasons for acquisitions uh which you've written about. Uh I want to sort of proceed somewhat chronologically as well. So that there there is a point that you know one of the questions that uh Dave from HFZ sent in was when did you know it start really started to work and you brought up Mix Draw I don't know if you want to bring up that.

哦，是的

Oh yeah

显然你与之有重叠，所以

which obviously you overlap with so

是的，我不知道什么时候，我的意思是，没有某一个时刻让我觉得哦，这正式开始了。它是一点一点增加的，很早就开始了。

yeah thee was I don't know when I mean the there's no like one moment where I was like oh this is you know officially starting to work. It was like moment increasing like really super early on.

哦，是的。嗯，是的。所以，在 OpenRouter 之前，我想探索一个自带模型的实验，嗯

Oh yeah. Well, yeah. So, before open router, I wanted to like explore a bring your own model experiment and um

熟悉加密货币的人都知道 Phantom 之类的。

anyone familiar with crypto is like you know phantom and all these things.

是的。是的。所以，感觉做一个 AI 版的 MetaMask 类比会

Yeah. Yeah. So, it felt like doing a meta mask analogy for AI would be

是一种有趣的探索方式。当时没有 AI 应用。通过 API 调用 LLM 的 AI 应用可能和用 JavaScript 做的游戏一样多。基本上，曾经有一段时间，Web 应用可能通过浏览器调用 LLM，就像通过某种用户控制的桌面管理应用。当然，我认为有很多原因导致这没有发生，但在那个原始时期。我构建了一个叫 Window AI 的 Chrome 扩展，

kind of a fun way of exploring that. And at the time there were no AI apps. There were probably as many AI apps that were like hitting AI via like hitting an LLM via an API call as there were like games just doing it in JavaScript. You basically like there there was a there was a moment in time where it could have been the case that web apps call LLM through the browser like through some kind of desktop managed app that is controlled by the user. Um, and of course there are like I think many reasons that that did not happen, but back when the when the days were that primordial. I built a a a Chrome extension called window AI and

用了 Plasmo，我早期遇到过，我当时想谁会真正用这个？你用了

with plasma which I had come across early on and I was like who's going to actually use this? You did

Plasmo 有几个，我想 Phantom 在用。嗯，还有一些其他真正的公司

plasma had a couple like I think Phantom was using it. Um, there were some other like like real companies basically

用于 Chrome 扩展的 React。它编译成类似 Next.js 的东西

react for Chrome extension. It compiles to all these kind of like Nex.js for

是的，在它上面构建了 Window AI。Plasmo 的创建者开始向 Window AI 贡献代码，在 GitHub 上，结果发现是 Lewis Vichy，OpenRouter 的联合创始人。

and yeah built window AI on top of it. The creator of plasmo like started contributing code to window AI and uh in GitHub and that turned out to be Lewis Vichy who is the co-founder open router.

你告诉过我这就是你遇见 Lewis 的方式。是的。好的。所以，那允许用户配置他们想在浏览器中为网页使用哪个模型，然后应用在需要时调用那个模型。你知道，对 LLM 来说不是正确的形式，但你知道，这是个有趣的实验。你学到很多，我开源了它，主要的学习是，好吧，这必须是一个 API，必须看起来更像有更多开发者体验和更多发现体验，我不知道在哪里使用这些模型，一个小 Chrome 扩展不能帮助我发现，没有足够的空间，我需要更多空间，我需要视觉，我需要图表，我需要例子，我需要图像，我需要能够像人类和智能体一样探索。所以这就是 OpenRouter 的由来。

That's you have told me this is how you met Lewis. Yes. Okay. So, um, that that allowed users to kind of like configure which model they wanted to use for a web page in their browser and then like the the app would just call out to that model when it needed to do things. You know, not the right form factor for LLMs, but you know, it's like fun experiment. you learn a lot and like I you know open sourced it and uh and you know the main learning is like okay this has to be an API and it has to look a little bit like there has to be more of a developer experience here and more of a discovery experience as well like I don't know where to use these models and a little Chrome extension is not going to help me discover it's not enough real estate I need more space I need visuals I need graphs I need you know examples I need images I need to like I need to be able to like explore both as a human and as an agent. So that's kind of how how open Rider came to be.

你知道，一个元观点

You know, a meta point that

我认为被低估了，但 Alex 提醒我的是，我们很幸运，当时我们与加密货币社区相邻，因为事后看来，加密货币成了生成模型的彩排，对吧？如果你想想 AXI 的经历，Alex 完全正确。当时没有那么多 AI 应用。而当我处理时，我的工作是 Discord 的平台负责人，这意味着一个通用场所，供社区和朋友创建，供开发者创建应用、机器人和其他可以部署在 Discord 上的服务。虽然当时 80% 的注意力都花在加密货币上，因为所有 NFT 交易量都在那里，但我有 20% 的时间花在一个朋友身上，他会和我一起玩，周末我们玩万智牌。他在做一个小的 Discord 机器人，可以把文本输入变成图像，叫做 Midjourney。

I think is underappreciated, but Alex is reminding me is that we were quite lucky that we were so we were like adjacent to the crypto community in those days because in hindsight, crypto ended up being kind of like a dress rehearsal for generative models, right? If you if you think about the the AXI experience, uh you know, Alex is totally right. There were not that many AI apps at the time. And while I was dealing, you know, my job was to be the head of platform at Discord, which meant to be a general purpose place for communities and friends to create uh for developers to create apps and bots and you know, other services that could be deployed across Discord. And while 80% of the attention of the time was being spent on crypto because that's where all the NFT volume was, there was like 20% of my time of my time I was spending with a friend u who would get hotbot with me and ask me for we would play Magic the Gathering on weekends. Um and he was working on a little Discord bot that could take a text input and turn it into an image and it was called Midjourney.

你知道，那是 David 吗？

You know, was that David?

那是 David Holtz。他是个好朋友，David 和我之前都是失败的 AR/VR 创始人。我记得 Midjourney 是 Axi Infinity 开始衰退后我们增长最快的社区之一，我们为扩展 Axi 所做的许多抽象和基础设施决策来得正是时候，因为 Axi 这样做了然后一落千丈，然后当 Midjourney 起飞时，我们明确决定帮助 David 把 Midjourney 服务器作为与模型互动的主要场所，因为如果人们看不到别人使用并模仿，就很难理解如何使用模型。所以单人的 Midjourney 网页应用，比如 major.com，留存率很差，因为人们来了只看到空白的输入框。

That was David Holtz. He was a good friend and David and I have both been sort of failed ARV VR founders, you know, in the last before that. And um I remember this you midjourney was one of the fastest growing communities we had after axi infinity started to peter off and many of the the like the abstractions and the infrastructure decisions we made to scale Axi happened just in time because you know Axi did this and then fell off a cliff and then as Midjourney was taking out we like explicitly decided to help David make the server the midjourney server as the primary place for interaction with the with the model because it was very hard for people to understand how to use the model if they couldn't see other people using it and copy them. And so the single player Midjourney web app on its own like majour.com had like terrible retention cuz people would show up they'd see this empty field.

这有点像 Dolly 2，他们会输入猫或狗，面对空白画布要填内容，他们感到不知所措，因为从未用过 AI 模型。但在 Discord 服务器里，你能看到别人怎么用，并借鉴他们的提示词，参与度爆表。所以 Midjourney 从零增长到约 1000 万月活，在 Axi Infinity 之后是更顺畅的路径。

It's kind of like Dolly 2, and they would type in like cat or dog, and it was paralyzing for them to have this blank canvas that they had to fill because they never used an AI model before. But instead, in a Discord server, you could see other people using it and riff off of their prompts, and the engagement was off the charts. So scaling Midjourney from zero to like 10 million monthly actives was a much smoother approach post-Axi Infinity.

别忘了四张图中选最佳，这就是反馈循环，基于人类反馈的强化学习（RLHF）反馈循环。

Don't forget the best of four pictures, which is the feedback loop, the RLHF feedback loop.

顺便说一句，Tom Brown、David 和我周末常玩万智牌，所以是一群朋友聚在一起，这些概念一直在讨论。但我们中很少有人同时连接加密世界和 AI 世界。相比加密，问题总是这项技术的用例是什么，而对 AI 从来不需要问，因为用例太直观了。就像我现在能创造任何我想象的东西，我能写小说，我能编码。而我们这些相信分布式系统价值如加密的抗审查部分的人，发现了这个爆炸性的用例。我认为在 Midjourney 之间，Claude 在发布前是个 Discord 机器人，我们内部用作 LLM，ElevenLabs 有个 TTS 模型也在 Discord 上。Discord 成了早期应用创新的培养皿，我不认为他们在那找到家是巧合，在 OpenRouter 给世界一个公共主页或店面之前。Discord 几乎是那种搭便车于我们为加密社区构建的基础设施上的培养皿店面。然后我认为 Alex 是最早意识到的人之一，等等，这些应用需要在互联网上有自己的家。然后 OpenRouter 对我来说是那个社区需求的延续，当然还有你为许多开发者实现的疯狂分发。

Which, by the way, separately, Tom Brown, David, and I used to play Magic the Gathering on weekends, so it was one group of friends hanging out, and these concepts were all being discussed all the time. But there were few of us who bridged both the crypto worlds and the AI worlds. Compared to crypto, where the question was always what's the use case for this technology, there was never any need to ask that for AI because the use case was so visceral. It was like I can create now anything I can imagine, I can write novels, I can code. And the infrastructure that those of us who believed in the distributed systems like value of crypto, like the censorship resistance part, found this use case that was explosive. And I think between Midjourney, Claude was a Discord bot pre-launch that we were using internally as an LLM, ElevenLabs had a TTS model that we had on Discord as well. Discord became this petri dish for early apps to innovate, and I don't think it's a coincidence that they found a home there before OpenRouter gave the world a public home store or storefront. Discord was this almost kind of petri dish storefront that had kind of piggybacked on the infra we'd built for crypto communities. And then I think Alex was one of the first people to realize, wait a minute, these apps need their own home on the internet. And then OpenRouter, to me, was a continuation of that community's needs and of course the crazy distribution that you enabled for a lot of these developers.

那么我的问题是，为什么你——我的感觉是 OpenRouter 并不那么以 Discord 为中心，对吧？你有 Discord，用它来互动社区，但不像 Midjourney 那样，那是人们体验 OpenRouter 的主要方式。

So then my question is, how come you were—my perception is OpenRouter is not that Discord-centric, right? You have a Discord, and you use it to engage your community, but it's not like Midjourney where that is like the primary way people experience OpenRouter.

是的。Midjourney，它真的有助于快速视觉化地看到人们如何使用模型以及如何提示，我认为这是服务器如此关键的部分原因。它就像用户体验本身。它实际上增加了大量价值。

Yeah. Midjourney, it really helps to see visually really quickly how people are using the model and how to prompt it, and I think that is partly why the server was so critical. It's like it is the user experience. It actually adds a ton.

是的。

Yes.

而你可以全程只通过 Midjourney 提示，通过 Midjourney Discord 服务器，获取图像然后分享和娱乐。对于 OpenRouter，对于 LLM，你需要大量围绕 LLM 的用户体验来使它们真正可用。看到别人的例子也不那么有用，因为要读很多东西。需要很长时间。你需要基于代码的集成，这在 Discord 服务器中不可能。你需要——或者技术，可能，我不该说，只是不是很好的开发者体验。你需要算力，你需要治理。在你有基于代码的集成时，现在你需要治理来管理访问它的 LLM、数据政策、哪些团队，所有这些需要远超 Discord 服务器能提供的。所以就像——

And you can go the whole mile with just prompting via Midjourney, via the Midjourney Discord server, getting your images and then sharing them and having fun. For OpenRouter, for LLMs, you need a lot of user experience around LLMs to make them really usable. And seeing the examples of other people is also not as useful because it's a lot of stuff to read. It takes a long long time. You need code-based integration, not possible to do in a Discord server. You need—or tech, it's possible, I shouldn't say that, it's just not a great developer experience. You need compute, you need governance. At the point where you got code-based integration, now you need governance for managing the LLMs that have access to it, the data policies, which teams, all that stuff needs a lot more than a Discord server can provide. So it's just like—

这不是正确的——

It's not the right—

嗯，此外，你没错，但还有一个非常重要的区别，Midjourney 是一个终端用户应用，对吧？这就是为什么 Discord，拥有 2.5 亿月终端消费者，让 Discord 成为那种应用体验的宿主是合理的。我知道在 Midjourney 找到爆炸性产品市场契合后不久会发生什么——因为 Midjourney 从发布到 1 亿美元收入运行率不到八个月，之后不久 Stable Diffusion 发布，我们所有人过去常在 Discord 服务器里闲逛，那是 Stability Discord——

Well, in addition, you're not wrong, but also there's the very important distinction that Midjourney was an end-user application, right? And that's why Discord, which just has 250 million monthly end consumers, made sense for Discord to be a host for that application experience. What I knew was going to happen soon after Midjourney found explosive product-market fit—because when Midjourney launched, from launch to 100 million revenue run rate was less than eight months, and shortly thereafter Stable Diffusion launched, and all of us used to hang out in the Discord server, it was the Stability Discord—

Stability Discord？

The Stability Discord?

呃，是 lion——

Uh, it was the lion—

是的，lionage 社区，Stable Diffusion——

Yeah, lionage community that Stable Diffusion—

所以当 Stable Diffusion 出来时，我意识到，哦，现在别人可以构建自己的 Midjourney。因为在那之前，Midjourney 没有 API，所以他们是全栈公司。他们训练自己的模型并作为应用部署。但如果你想构建自己的 Midjourney，没有那种质量的 API。我认为 Dolly 2 还相当原始，Midjourney 实际上质量很好。然后当 Stable Diffusion 出来时，突然世界上有了这种新能力，即开发者可以创建自己的 Midjourney。我认为这创造了像 OpenRouter 这样的需求，因为然后你需要一个 API。如果你有大卫·霍尔兹那样的创造力，你有 Stable Diffusion 作为模型，你想把这些东西结合起来，你怎么做而不必弄清楚如何托管权重？而 OpenRouter——OpenRouter 的形状所启用的，就是当你有开放模型替代封闭应用时，OpenRouter 在世界上的价值变得非凡，因为现在任何开发者都可以出现并——

And so when Stable Diffusion came out, I realized, oh, now other people can build their own Midjourney. Because until then, Midjourney did not have an API, so they were a full-stack company. They were training their own models and deploying them as an application. But if you want to build your own Midjourney, there was no API of that quality. And I think Dolly 2 was still quite primitive, like Midjourney actually had great quality. And then when Stable Diffusion came out, suddenly there was this new capability in the world, which is a developer could create their own Midjourney. And that, I think, created the need for something like OpenRouter, because then you need an API. If you had the kind of creativity of David Holz and you had Stable Diffusion as the model and you wanted to put these things together, how could you do that without having to figure out how to host the weights? And what OpenRouter—the shape of OpenRouter enabled is that right when you have open models alternatives to closed sort of applications, OpenRouter's value in the world becomes extraordinary because now any developer can just show up and—

模型——你刚才说 OpenRouter 的形状？

Model—did you just say the shape of OpenRouter?

哦不，你——这是真的——我现在错位了。我被过度训练了。我用 Claude 太多了，不是吗？人们说 Claudish。

Oh no, you—this is the real—I'm misaligned now. I've been overtrained. I've been using Claude way too much, haven't I? Claudish is what people say.

Claudish？天哪，我自己也未经训练了。

Claudish? Oh god, I got untrained myself.

好的。我只想结束 Mistral 这边。我的 TL;DR 是有一场混合价格战，他们这么叫的，对吧？像 roundabout Europe 是 2023 或四年。他们推出了 MR 8x8x7B，价格下降了 80%。对我来说这非常积极，因为这是第一次真正竞争托管 Mistral。还有更多吗？

Okay. And I just want to cap off the Mistral side. My TL;DR is there was a mixture price war, is what they called it, right? Like roundabout Europe was 2023 or four. They launched the MR 8 by 8 by 7B, and like the price went down like 80%. To me that's very positive because it's like the first real competition to host Mistral. Is there more?

是的，那是——我试着回忆。所有发生的事情，我们看到那个模型出来，立即看到人们说它是世界上最好的模型。据我所知，这是第一次一个开放权重模型被如此认真地称呼。

Yeah, that was—I'm like trying to remember it. All the things that happened, it like we saw that model come out and immediately saw people say that it was the best model in the world. Like this was to my knowledge the first time an open weights model was called that in real seriousness.

这是炒作，对吧？是不是，你知道——

It's hype, right? Is it, you know—

是炒作。是炒作。当时也是 AI 影响者的炒作，有很多例子显示它超越了 GPT-4。

It was hype. It was hype. It was also like hype from AI influencers at the time, and there were many examples where it was like outperforming GPT-4.

所以人们真的很想试试看，这对我也会成立吗？如果成立，代价是什么？推理领域的格局当时非常混乱。我们把它清理干净了。它让提供商在价格上竞争，这样我们就能在一个地方给用户最好的价格。所以我认为，这是提供商市场以对开发者有价值的方式运作的第一个清晰例子。

So people really wanted to try it out and see, is this going to be true for me too? And if so, at what price? The inference landscape was really messy. We cleaned it up. It allowed providers to compete on price so we could give users the best price in one spot. And so it was, I think, the first clear example of a provider marketplace working in a way that adds value to developers.

Sean，你可能不记得了，但我想我们第一次见面是在 Mixtral 发布几天后，在 NeurIPS 的一次午餐会上。

Sean, you may not remember this, but I think we met for the first time a few days after Mixtral came out at NeurIPS at a luncheon.

是的，那也是我遇到 BFL 的地方。Guillaume 也在那里。我当时在 NeurIPS。

Yeah, that's where I also met BFL as well. And Guillaume was there. I was at NeurIPS at that time.

你当时也在。我们刚刚宣布了 Mistral 的投资，我记得 Guillaume 在那里，我转向 Guillaume 问他，Mixtral 7B 发布后你感觉如何？你知道他，以他典型的法国风格，他说，嗯，这是个还行的模型，没那么好。我当时觉得，这对比太鲜明了。但我记得他还说，他觉得很多人认为它比 GPT-4 更好的部分原因是速度。你知道，这是一个 MoE 模型，他们绝对搞清楚了如何让它超级高效。它处于帕累托前沿。这对 LLM 来说很重要，对吧？有时当它们更快时，你会觉得它们更聪明。即使你做了，你知道，这些常见的评估，我其实不记得了。我想我们应该回去弄清楚数据怎么说，但如果结果显示在七次尝试中，GPT-4 在评估上更聪明，我不会感到惊讶，但正确性的感知会更聪明或更准确。但你知道，从人类偏好的角度来看，人们觉得它更快是因为它更聪明，因为它太快了。

You were there too. And we had just announced the Mistral investment, and I remember Guillaume was over there, and I remember turning to Guillaume and asking him, like, how are you feeling after the launch of Mixtral 7B? And you know him, in his typical French fashion, was like, I mean, it's an okay model, it's not that good. And I was like, it was so, you know, in contrast. But I remember him also saying that part of the reason he felt a lot of people thought that it was better than GPT-4 before was because of the speed. You know, it was an MoE model that they had absolutely figured out how to make super efficient. It was on the Pareto frontier. And this is an important thing with LLMs, right? Sometimes when they're faster, you think they're smarter. Even though if you did end-of, you know, these common evals, I don't actually remember. I think we should go back and figure out what the data says, but I wouldn't be surprised if it turns out on an end-of-seven attempts, GPT-4 was smarter on evals, but the perception of correctness would be smarter or more accurate. But you know, people, from a human preference perspective, felt that it was faster because it was smarter because it's so fast.

实际上大多数查询并不需要那个水平。

And actually most queries do not take that level.

不需要那个水平，对吧？这是人类作为路由器的开始，然后最终变成 OpenRouter 作为路由器，就像自动的，你明白我的意思吗？因为人类是路由机制。就像我会先问快速模型，然后如果，哦不够好，我会手动升级。

Don't take that, right? This is the start of humans as router, which then eventually becomes OpenRouter as router, of like the auto, you know what I mean? Because humans are the routing mechanism. Like I will ask the fast model first and then if, like, oh not good enough, I'm going to upgrade manually.

但然后他会把它自动化。

But then he's going to auto it.

我没有，我没想到那个角度，但那说得通。

I didn't, I hadn't thought of it that way, but that makes sense.

然后还有更多技术，比如融合。融合是我们应该讨论的事情。

Which then there's a lot more techniques like fusion. Fusion is a thing that we should talk about.

在我继续讨论那些事情之前，我只想结束早期阶段。我观察到一件事，你也是 Arena 的投资者，对吧？我们谈过 Midjourney 有那个 ABCD 反馈循环，选择非常重要。你理解飞轮。那么为什么你没有建立 Arena，为什么 Arena 没有建立 OpenRouter？

Before I move on to those things, I just want to close off the early years. One thing that I observe, which you are also an investor in Arena, right? And we talked about Midjourney having that feedback loop of ABCD and choosing that being very important. And you understand the flywheel. So how come you didn't build Arena and how come Arena didn't build OpenRouter?

嗯，Arena 在 OpenRouter 之前就开始了，对吧？他们有一个学校项目，然后他们变成了 Marina。是的。所以，但我知道你有一些 Arena 的经验，比如那种直接对比的东西，但你从来没有像 Arena 那样努力去做直接对比。而 LMS 实际上有一个基于 Alam Marina ELO 的路由器项目，但他们从未商业化。

Well, Arena started before OpenRouter, right? They had the school project and then they became a Marina. Yeah. So, but and I know you had some Arena experiences like the heads-up comparison type things, but you never really went as hard as Arena did in doing heads-up experiences. And LMS actually did have a router project based on Alam Marina ELOs, which they never commercialized.

很难做一家同时做这两件事的公司，因为一家公司获取数据并出售，而另一家公司默认情况下真的不能。所以你知道，我认为这里有两家公司是有品牌原因的。就像当你设置 OpenRouter 时，没有训练或没有提示，除了你的提供商政策设定的，OpenRouter 看不到你的提示或完成。如果你作为一个组织想看到这些，你必须选择加入并启用它。所以我们对数据政策、安全和隐私相当保守和谨慎。而 Elm Marina 的商业模式是围绕实验室的，并且

It's hard to do a company that does both because one company is taking data and selling it and the other company really can't by default. So you know I think there is like a branding reason that there are two companies here. Like when you set up OpenRouter there's no training or no prompts like aside from what your provider policies set like OpenRouter can't see your prompts or completions. If you want to see that as an org, you have to opt into it and enable it. And so we're like pretty conservative and careful about data policy and security and privacy. And Elm Marina is like their business model is oriented around the labs and

因为他们免费提供，对吧？你不免费提供，他们免费提供。

Because they give it for free, right? You don't give it for free, they give it for free.

是的。但我的意思是，我们确实提供一些，我们也有免费端点，但那些免费端点，我认为我们不收集任何提示。我们不会将数据货币化，除非你出于某种原因选择加入。这个比较，我的意思是，你不是第一个问我这个的人，Alex 知道这个，但我是 Arena 的创始人，前 5 个月的第一任 CEO，当时我们帮助 Anastasia 和 Whan 从伯克利分拆出来，我在 OpenRouter 之前就投资了那个，但外部人士对这两个项目的比较对我来说很奇怪，因为使命完全不同。Arena 的创始实体，我们称之为 AI 可靠性研究所，因为它实际上是一个评估服务，可以说他们最初提供给实验室的数据是如何使模型评估比当时的最先进技术更可靠，而当时的最先进技术真的只是凭感觉。这大概就是 Anastasio 和 Whan 作为伯克利科学家的博士工作，是关于统计方法，用于纠正基于内在偏差和如何收集数据的评估估计。

Yeah. But I mean, we do give some, we like have free endpoints too, but like those free endpoints, I think we're not collecting any of the prompts. We're not like monetizing the data unless you know opt into it for some reason. This comparison, I mean you're not the first person to ask me this and Alex knows this but I was the inter like the founder like first CEO of Arena for the first 5 months when we were helping Anastasia and Whan kind of spin out of Berkeley and I did invest in that before OpenRouter but it was very strange to me the comparisons that outside folks would make between the two projects because the missions were completely different. The founding entity for Arena, we called it the AI Reliability Institute because it was actually there as an eval service like the data so to speak that they were originally offering the labs was how do you make the evaluation of models more reliable than kind of like the state-of-the-art at the time which is like really just finger in the wind. That's kind of what Anastasio and Whan's PhD work was as scientists at Berkeley was on statistical methodologies for sort of correcting you know eval estimates based on like intrinsic biases and how you collected the data.

还有风格控制，风格控制之类的。

And style control, style control and stuff like that.

这非常像，嘿，如果你是一个科学家，你试图，Arena 的最高期望客户总是像后训练和实验室的研究员。而从我看来，Alex 真正理解并且使命是服务的最高期望客户是开发者，对吧？然后开发者拿研究结果，然后产生一个部署到世界的应用。实际上，这两个团队关注的是完全不同的问题和人。所以从外到内，实际上，我不知道你是否记得这个，但我有一个清晰的记忆，在我们一起为 OpenRouter 做条款清单的几周前，我给你打了个电话，因为我们试图从 OpenRouter 和 Arena 汇集一个池数据集，创建一个开源的提示库。

And which is very much like a hey how if you're a scientist and you're trying to kind of the highest expectation customer for Arena was always like a post-training and like a researcher at a lab. Whereas the highest expectation customer from my perspective that Alex really understood and was the mission was to serve was like a developer, right? Who then takes the result of the research and then produces an application that's deployed to the world. It's actually a completely different problem and person that these two teams were focused on. And so from the outside in actually I don't know if you remember this but I have a distinct memory of a few weeks before we did the term sheet together for OpenRouter I'd given you a call because we were trying to get a pool's data set together from OpenRouter and from Arena to create like an open-source repository of prompts.

我的意思是，这些项目的目标如此不同，所以我觉得很正常，哦是的，让我们打电话给 Alex，看看你是否愿意合作汇集数据，因为它们如此不同。我们实际上根本没有那种数据。我们没有 API 提示。

I mean these projects were so different in their goals that it was totally normal to me to be like oh yeah let's call Alex and see if you'd want to team up on pooling data cuz they're so different. We actually don't have that kind of data at all. We didn't have API prompts.

我们没有开发者想用模型做什么，这与模型实验室内部的研究人员在发布模型之前想做的事情非常不同。这说得通吗？所以直到今天，我认为你仍然能看到这种差异，尽管在 3 万英尺的高空，你可能会得出 Arena 和 OpenRouter 是相邻的结论，但当时至少路线图、使命等方向是非常不同的。

We didn't have what developers want to do with the models, which is very different from what researchers inside a model lab want to do before releasing the model. Does that make sense? And so to this day, I think you see that this difference, even though at a 30,000-foot level you could conclude that Arena and OpenRouter are adjacent, but the roadmaps, the missions, and so on at the time at least were in very different directions.

理想客户，我完全理解。作为创始人，我想拥有 everything，对吧？所以这显然是相邻的，然后我想我要探索这个。拥有 everything 意味着你还不知道要做什么，所以你想确保抓住 PMF。我认为他说的是你想拥有整个基础设施空间，所以你会扩展到任何需求。

The ideal customer, I totally get that. As a founder, I want to own everything, right? So like this is clearly the adjacency, then I'm like I'm going to explore that. Own everything meaning like you don't know what to do yet, so you want to make sure you catch PMF. I think what he says is you want to own the entire infrastructure space and so you'd kind of expand to whatever demand.

是的，我认为这很难，你知道，在现实中因为服务多个客户显然——你知道，这是唯一的焦点，对吧？

Yeah, I think that's hard, you know, in reality because serving multiple customers is clearly—you know, this is the only one of focus, right?

是的，我仍然认为即使在 AI 时代，专注也被低估了，而且至关重要。不仅仅是因为你把人力集中在上面最终会得到更好的产品，还因为世界知道你的专注点是什么。世界可以映射，比如，哦，我有这个问题。哪个品牌会帮我解决这个问题？这个品牌以专注于此而闻名。所以如果我想真正关注这个问题，这对我真的很重要，我应该选择最关心它的品牌。

Yeah, I still think even in the age of AI, like focus is underrated and critical. Not just because you end up with a better product by focusing your humans on it, but also because the world knows what your focus is. The world can map like, oh, I have this issue. Which brand out there is going to help me with that issue? This is the brand that's known for that focus. So like if I want real attention on this issue, like this really matters to me, I should go with the brand that cares the most about it.

为了强调 Alex 关于专注有多重要的观点。在 Anthropic 的早期，并不容易。人们认为 Anthropic 的早期非常容易，因为他们是离开的 GPT-3 团队。但实际上竞争非常激烈。公司起步时落后 OpenAI 100 亿美元，对吧？所以为了达到前沿，最大的问题是我们想以什么闻名？使命是什么？使命是 AGI 结对编程。因此，排除了当时所有其他非常耀眼的东西，比如图像模型和视频模型，它们正在获得很多动力，Anthropic 团队说我们只需要专注于编码。这是我们所关注的核心能力。今天你可以看到结果，对吧？它在 5 年内成为了一家万亿美元公司。而那种专注，我认为，专注于谁是你期望最高的客户以及如何超越他们的期望，因为超越任何人的期望都很难，为多个客户做到这一点就更难了，是 OpenRouter 和 Anthropic 成功的原因之一。

To underscore Alex's point about how important focus is. In the early days of Anthropic, it was not easy. People think that the early days of Anthropic were super easy because they were the GPT-3 guys who left. But it was actually very competitive. The company was starting $10 billion behind OpenAI, right? And so to get to the frontier, the big question was what do we want to be known for? What's the mission? And the mission was AGI pair programming. And so to the exclusion of all kinds of other things that were really shiny at the time like image models and video models that were getting lots of momentum, the Anthropic team was like we just got to focus on coding. That is the core capability that we're focused on. And today you can see the results, right? It's a trillion-dollar company within 5 years. And that focus, I think, like the focus on who your highest-expectation customer is and how you exceed their expectation because exceeding anyone's expectations is hard and doing it for multiple customers is even more difficult, is part of the reason why OpenRouter succeeded and Anthropic as well.

但专注编码是那么早吗，还是后来才有的？

Was the focus on coding that early though, or did it come later?

从第一天起， literally 就是 AI 结对编程。负责任地将 AI 结对程序员商业化是种子备忘录。那是我投资的时候。对。我们实际上对那份备忘录进行了很多完善。嗯，你得问 Dario 和 Tom 关于那件事的许可。但这是他们整理的一份非凡的写作，AI，你知道，负责任地将 AI 结对程序员商业化从第一天起就是使命。我会说公司历史上可能有几个时刻，他们做了实验，看看小的绕道是否有意义，比如当 ChatGPT 真正起飞时，像 Claude 这样的通用聊天机器人。但最终，特别是当他们获得了重要的预训练算力上线后，我认为公司所有主要的评估，例如，一直是编码评估，长视野智能体式编程。我的意思是，从第一天起就一直是。

Literally from day one, it was AI pair programming. Responsibly commercialize an AI pair programmer was the seed memo. That was when I invested. Right. We actually kind of refined that memo a lot. Well, you got to ask Dario and Tom for permission on that. But it's an extraordinary piece of writing that they had put together, and AI, you know, responsibly commercializing AI pair programmer was the mission from day one. And I would say there were maybe a couple moments in the company's history where they did experiments to kind of see if little detours made sense, like a general chatbot like Claude when ChatGPT was really taking off. But at the end of the day, especially once they got their significant pre-training compute online, I think all the main evals at the company, for example, have always been coding evals, long-horizon agentic programming. I mean, from day one that was always.

当 Claude Instant 和 Claude 2 出来时。是的，我记得营销主要聚焦在优点上，比如这个模型写得更好和长上下文。

When like Claude Instant came out and Claude 2 came out. Yes, I remember the marketing mostly being focused on pros like this model writes better and long context.

长上下文。

Long context.

这直接影响到我，因为我在那上面投入了一些东西。

This directly affected me 'cause I told something on that.

你做了什么？

What did you make?

呃，小开发者，那是我之前的 Devon。

Uh, small developer which was my Devon before.

哦是的，小。

Oh yeah, small.

是的，呃，而且呃，你知道，所以我认为就像所有那些真正好的，就像专注是另一件事，这是人们确实想问的问题，你知道，你可以构建任何其他东西，显然 OpenRouter 一直在工作工作工作，呃，有没有其他你想追求但拒绝了的想法，你知道，只是路径，未走的路？

Yes, uh, and uh, you know, so I think like there's all that really like good like focus is another thing that is a question that people do want to ask, you know, you could have built any other things like and obviously OpenRouter was working working working, uh, were there other ideas that you wanted to pursue that you turned down, you know, just the paths, roads not taken?

我们为一些没有发布的东西做了几个原型。一个是微调模型即服务。

We made a couple prototypes for things that we didn't launch. One was a fine-tuning model as a service.

很多 OpenPipe 和所有那些东西。

A lot of that OpenPipe and all those things.

但它是一种非常消费者化的形式，你给我们一个或两三个 YouTube 视频，然后我们会从中提取所有转录文本，并尝试微调一个模型，让它像 YouTube 视频中的人或你发送的视频中的人那样说话。所以就像一种非常非常简单的方式，基于你喜欢的某种视频来创建微调模型。

But it was kind of in a very consumer-y form factor where you would give us a YouTube video or two or three, we would then extract all the transcripts from it and try to fine-tune a model to talk like the person in the YouTube video or the people in the videos that you sent. So like a really, really easy way of creating a fine-tuned model based on some kind of videos that you like.

那会非常有用。

That would be so useful.

我们也做了。它就像——

We made it too. It was like—

我——它是，但没人用。

I—it was and nobody used it.

我们实际上没有和那么多人测试它，因为模型市场是我们的主要焦点，它在增长，我们随着时间的推移对它建立了更多信心。

We didn't actually test it with that many people because the model marketplace was our main focus and it was growing and we're building more conviction in it over time.

嗯，作为一个创作者，我被推销过很多，比如，呃，你有 500 小时的自己录音，做一个你的东西，收费访问。呃，对 OnlyFans 有效，对我们普通人无效。我认为这主要是，呃，它只是一个美化的 RAG 机器人。无论是在权重中还是在权重外，并不重要。你只是在视频上做 RAG，人们最终总是只想找到直接回答它的源视频。

Um, just as a creator, I've been pitched many like, uh, you have 500 hours of recorded voice of myself, make a thing of you, charge access to it. Uh, works for OnlyFans, doesn't work for us as regular people. I think this is mostly, uh, it's just a glorified RAG bot. Whether it's in the weights or it's outside the weights doesn't really matter. You're just doing RAG on the videos and people ultimately always just want to find the source video that directly answers it.

我的用例主要是和自己练习，因为我经常喜欢看看——比如我为工作面试练习的方式，或者如果我在招聘候选人，或者公开演讲或其他什么，我希望有一个好的迷你版我，我可以批评，因为把自己抽离出来有点难。我永远不会——我永远不会把它提供给其他人，像选择你的前五名导师，然后和他们交谈而不是交谈——

My use case was mostly to practice with myself 'cause I often like to see what—like the way I practice for a job interview or if I'm hiring a candidate or public speaking or whatever is I wish there was like a good mini me that I could like critique 'cause it's kind of hard to pull yourself out. I would never—I would never offer it to other people a service like pick your top five mentors that then talk to them instead of talk—

那也会很酷。是的，

That would be cool too. Yeah,

那是那是 Replica。

that was that's a replica.

而那是我们瞄准的用例。

And that was the use case we were aiming at.

我明白了。

I see.

就像你想创造一种体验，像 AI 史蒂夫·乔布斯和

It's like you want to create an experience like AI Steve Jobs and

AI 史蒂夫·乔布斯是初始用例。

AI Steve Jobs was the initial use case.

那是一个——即使它不被允许——那是一个常见的原型。

That's a—even though it's not allowed—that's a common prototype.

说到相邻业务，把微调作为一种服务、作为路由服务的一部分，这也是我通常会想到的方向，对吧？比如，你为什么不这么做？因为如果人们已经在通过你跑推理，把所有东西都存下来、记录下来，然后微调出一个更便宜、更快的小模型，所有这些都在你的掌控之内，对吧？你没做这件事，但在基础设施创业公司的大环境里，别人会拿这个来推销。

You know, talking about adjacencies, fine-tuning as a service as part of the router service is something that I would typically think about as well, right? Like, why don't you do that? Because if people are already running their inference through you, store everything, log everything, fine-tune to a smaller model that is cheaper, faster, all these things — that's within your control, right? You didn't do that, but other people would have pitched that in the general state of infra startups.

我觉得你可能只是稍微早了一点，因为今天那是一个增长极快的细分市场。比如 Astral，他们做很多企业部署——很多时候就是微调，为客户定制模型，比如给 ASML 之类的公司做定制模型，经常是这样。

I think you were just maybe a little bit early, because today that's an extraordinarily fast-growing segment. Like, you know, from Astral where they do a lot of enterprise deployments — I mean, it's often fine-tuning, you know, custom models for ASML or whatever, often.

但不是以路由器的形式。他们更像是：我来找你是因为我喜欢你们的机器学习模型，我想要一个定制模型，对吧？而不是：我想跑我所有的 OpenAI 提示词，把结果都存下来，然后就此摆脱 OpenAI，对吧？他们不是在做这个。

But not as a router. They're just like, I come to you because I like your ML models, I want a custom model, right? It is not, I want to run all my OpenAI prompts, get, store all my results, and then just move off of OpenAI, right? They're not doing that.

作为一种摆脱对前沿实验室依赖的出口，我还没见过这种做法。

As a way to export off of dependency on a frontier lab, I have not seen that yet.

而这正是你当初决定要做的事。

Which was your kind of decision to do.

我是说，我们决定了，真的，我们专注于自己的核心方向，然后觉得——我们只是看着这个生态随着时间发展起来。所有这些推理服务商，他们确实想帮公司做这件事。嗯，对我们来说，跟他们合作、给用户很多选择、去搞清楚什么能让他们、什么能给他们带来竞争优势，这是合理的。这基本上是一门全新的生意，而做一个中立的市场、跟这些公司合作，是有价值的。

I mean, we decided, really, we leaned into our focus and figured that, like, we just saw the ecosystem develop over time. All these inference providers that do want to help companies do that. Um, be like — it makes sense for us to partner with them and to give users lots of choice and to, like, you know, figure out what makes them, what gives them competitive advantages. It's a whole new business, basically, and there's value in being a neutral marketplace that just kind of works with those companies.

你能不能稍微讲一下，接着 Sean 的话说，你是怎么排优先级的——你排功能优先级有哪些方法？因为你一直做得特别优雅。我从来——你知道，它就是自然而然地发生了，你做的所有决定都是对的，从外面看它们总是有产品市场契合度。你似乎一直都能优先做出很多成功的爆款功能。也许我有样本偏差之类的，但 Sean 可以列一下你觉得哪些爆款功能做得好，比如——

Could you share a little bit, to Sean's point, how you prioritized — what are some ways you prioritize features? Because you've always done it so elegantly. I never — you know, it just happens and you make all the right decisions, and they always have product-market fit from the outside looking in. Consistently you seem to have prioritized a lot of hit features that worked. And maybe I have a sample set bias or whatever, but Sean can list what you think hit features worked well, like —

哦，排行榜，比如——

Oh, the leaderboards, like —

排行榜，好。

Leaderboard, okay.

对，你知道，从第一天起。

Yeah, you know, like from day one.

反馈、图表——好。

The feedback, charting — okay.

但他还有插件，嗯，你知道，他还有——我觉得有一整套东西我想深入聊，比如补全对比——

But like he had like plugins, uh, you know, he had like uh — and I think there was a whole thing I want to get into about like completions versus —

对。

Yes.

聊天补全对比补全，然后还有，我们姑且称之为推理模型的崛起，以及你怎么处理多模态——所有这些东西，都是巨大的一个点。

Chat completions versus completions, and then also, let's call it like the rise of the reasoning models and how you deal with multimodality — all those things, by huge one.

有一个——我记得是在 2024 年初，非常早的时候——我们觉得把多个模型的结果融合在一起可能会很有意思。我们发布了一个叫 MoM、混合模型的原型，让你挑几个模型——我们也会帮你挑——然后它最后会把结果融合起来，并且会在这个像大看板一样的产品里展示所有中间结果。

There's one — I think it was in early, very early 2024 — we thought it might be interesting to fuse the results of multiple models together. And we launched a prototype called MoM, mixture of models, that let you pick a couple models — we'd pick them for you — and then it would fuse the results together at the end, and it would show you all the intermediate results in this like big conbon board-looking product.

最后融合靠什么？另一个模型。

What does the fusion at the end? Another model.

另一个模型。那一组三个里最聪明的那个。

Another model. The smart — the smartest of the three of the set.

所以这就像个议会式的想法。

So this is like a council idea.

它就是一个模型。它就像非常早期的 LLM 议会。

It was a model. It was like a very early LLM council.

这就是多智能体集群，就像某家前沿实验室早期会叫的那样，你知道。

This is a multi-agent swarm, as they would call it at one of the frontier labs, in the early days, you know.

对。有些想法方向是对的，但魔鬼在细节里。要让它们真正跑起来，需要大量的产品打磨。嗯，它们会把你的注意力从——对吧，你知道，从你手上其他事情上拉走，而且你需要做很多社区建设和学习，技术也可能太早了。所以它们可能失败的原因有各种各样。在我们这个案例里，技术稍微早了一点。换句话说，融合后的结果有点更差，有时跟用来融合的那个最好的模型一样，因为当时最好的模型远远领先于第二、第三选项。

Yeah. Like some of those ideas are going the right direction, but the devil's in the details. There's a lot of product refinement needed to make them really work. Um, they take your focus away from, right, you know, whatever else you have going on, and there's a lot of community building and learning that you need to do, and the technology might be too early. So there are all kinds of reasons they might go wrong. And in our case, the technology was a little too early. In other words, the fused result was a little bit worse, sometimes the same as the best model that was being used to fuse, because the best model was so far ahead of options two and three at the time.

你知道，随着时间推移，排名前三、前四的 LLM 已经越来越接近了。仍然各有分歧，但都像是有能力插入相当有意思的想法。强化学习基本上扩大了每个实验室里机器学习研究者的创造力表面积，所以他们能更有效地让不同模型的推理能力多样化。至少这是我关于为什么——的理论。

You know, over time, the top three or four LLMs have gotten closer together. Still neurode divergent, but like all capable of inserting like pretty interesting ideas. Like RL has basically like expanded the surface area of creativity for machine learning researchers within each lab, and so they can, you know, diversify the reasoning power of different models more effectively. At least that's my theory for why —

融合——它比 2024 年初那会儿更好用了。

Fusion is — it like works better than it used to, early 2024.

嗯，所以技术还有点太原始。形态也不对，所以我们必须再迭代几轮。于是我们决定干脆把所有代码都删掉。然后几年后，2026 年中，嗯，或者 2026 年初，我们说，把它带回来吧。关于融合的研究看起来挺有前景。现在模型这边有两三个、四个顶尖前沿模型，都真的很好，而且——比如我经常试着咨询多个模型来拿到最好的结果。然后我，你知道，我做了个小型的个人实验，我想，我要给一个代码改动做架构方案。我把它给所有模型。我把结果融合起来，然后我问所有模型，融合后的结果是不是比每个模型各自给出的结果更好。它们都说对，融合后的结果更好。这发生了几次，我就想，好，抽查结果相当不错。我们应该把它做成基准测试，于是我们就这么做出了 fusion。

And um, so the technology was a little bit too primitive. The form factor was not right, and so we would have had to go through a couple more iterations. And so we decided to just delete all the code. And uh, then years later, middle of 2026, um, or early 2026, we're like, let's bring it back. Like the research is looking kind of promising for fusion. The models now have like two, three, four top frontier models that are all really good, and like — like I'm frequently trying to like consult multiple models to get the best results. Like, and then I, you know, I ran a little personal experiment where I was like, I'm going to like do an architecture plan for a code change. I'm going to give it to all the models. I'm going to fuse the result, and I'm going to ask all the models if the fused result is better than the individual result each model came up with. And they all said yes, that the fused result was better. And this happened a couple times and I was like, okay, spot check pretty good. We should like benchmark this, and that's how we built fusion.

对。而且它出现在你的 fable 上。所以你就想，这是 fable 级别的。

Yeah. And it came on your fable. So you were like, this is fable level.

对。对。我们开始往今年引吧，今年我们还没聊到——还没讲到今年。

Yeah. Yeah. Let's start leading up to this year, which we haven't gotten to — gone to this year.

你能标出这段旅程中的主要里程碑吗？看起来你的承诺是路由，你很早就决定了商业模式，你抽成。那么，哪些主要里程碑影响了增长？你现在每周增长约 9%——就 token 量而言，这是官方数字吗？

Can you mark out the main milestones in the journey? It seems like your promise was routing, you decided the business model very early, you take a cut. And what are the major milestones that inflect the growth? You're growing like 9% week on week now — is that the official number in terms of token volume?

我觉得差不多是这样。

I think that sounds about right.

是的，那么你能简要概述一下 OpenRouter 从创立到被收购的历史吗？我们就这么说吧。我们只是在谈论——人们有你的诞生时刻，与 Mistral 相关的东西，人们真的在竞争，你有你的 State of AI 活动，非常可爱，你有 100 万亿 token——哈哈——因为现在你每周做 10 万亿。

Yeah, so just like can you mark out the brief history of OpenRouter up to the acquisition, let's call it. We're just talking about — people have your birth moment with the Mistral stuff where people are really competing, you have your State of AI thing where it's very cute, you have 100 trillion tokens — haha — because now you're doing 10 a week.

我们现在每天做 10 万亿。

We're doing 10 a day.

现在每天 10 万亿？

10 a day now?

是的，更多。

Yeah, more.

所以是的，你在 10 天内做到这个。主要节点是什么？我只是想——虽然曲线平滑，但你能感受到拐点。

So yeah, you do this in 10 days. What are the major points there? I just want to like — there's a smooth curve but you feel the inflections.

很多都围绕模型发布。一直到 2024 年 5 月，我们都非常关注专业用途，因为编码能力还不行，没有应用能在此基础上构建太多。所以模型有多样性，但不是很广泛，用例也不广泛。Dream Tavern 当时是我们的顶级应用之一。Dream Tavern 的创建者现在在 Cognition 负责产品，Devon。然后在 2024 年中期，我们看到 Claude Sonnet 3.5 发布——在编码方面取得了惊人的飞跃。我们看到在我们之上构建应用的动态发生了变化。我们看到使用 OpenRouter 的用户量激增，这时我认为人们开始关注他们花费的钱，并有点惊讶：哇，怎么回事？我可能需要考虑更具成本效益但等效的模型。不久之后——我想是在 Sonnet 3.5 之后——Mixtral 8x7B 发布了，大家都说，什么？这就是那个模型——开放权重社区交付了。所以时机真的很好。

A lot of this is oriented around model launches. We had a huge focus on pros all the way up through May of 2024, because coding was just not there and no apps were able to build much on top of it. So a diversity in models but not a wide diversity and not a wide diversity in use cases. Dream Tavern was one of our top apps at the time. The creator of Dream Tavern now runs product at Cognition, Devon. Then in the middle of 2024 we saw Claude Sonnet 3.5 that came out — incredible leap forward in coding. And we saw the dynamics of apps building on top of us change. We saw a huge surge in volume in users using OpenRouter, and this is when I think people started to look at the money that they were spending and get a little bit like, whoa, what's going on? I might need to think about more cost-efficient but equivalent models. And shortly after that — I think it was after Sonnet 3.5 — Mixtral 8x7B came out and everyone was like, what? This is the model — the open weights community delivered. So it was really good timing from the strong.

基本上所有开源公司都在帮你。

Basically all of the open source companies are just helping you out.

你知道，成长一个 OpenRouter 需要一个生态系统。

It takes an ecosystem to grow an OpenRouter, you know.

是的，那是早期的生态系统。就像一种摆动：模型实验室会推出某种前沿创新，使用量激增，然后用户在 30 天后查看账单，惊叹：哇，这是怎么回事？然后开放权重模型会在 3 个月后提供成本效益高的选项。我们看到这种情况发生了几次。

Yeah, that was the early ecosystem. It was like a swing action where model labs would come up with some sort of frontier innovation, usage would surge, then users look at their invoices 30 days later and like, whoa, what's going on here? And then open weight models would deliver cost-effective options 3 months later. We saw that happen several times.

你在编码智能体方面还做了一件事，就是你列出了顶级编码智能体，他们很喜欢那个排行榜。客户端与 Roo Code 等等。

One thing you also did with the coding agents was that you broke out which are the top coding agents and they love that leaderboard. The client versus the Roo Code versus the what have you.

是的。比如 Cline 当时在我们的排行榜上名列前茅。然后我们到了——我快进一点——2025 年底，排行榜上有不少编码应用，但它们都是 IDE 或基于终端的智能体。而在 2025 年底，我们看到了 Open Claw 出现。Open Claw 特别有趣，因为第一，它是一种新的形态，带来了新类型的用户——不仅仅是开发者，还有生产力用户或互联网创作者首次接触 AI。它还有一个有趣的架构，除了实际使用模型执行真实任务外，还会调用你选择的模型进行心跳检测，看看它是否还活着。而心跳——你不想为心跳支付太多费用。所以我们提供的自动路由突然对这么广泛的用户非常有用。于是我们看到它呈指数级增长，然后我们看到 Open Claw 爆火，还有几个其他应用也倾向于这种新范式并做了类似的事情。Hermes 出现了，真正倾向于自动路由之类的东西，建立了一个非常好的社区，并倾向于技能管理，让人们很容易有效地在智能体中设置记忆并构建非常好的技能。

Yeah. Like Cline was at the top of our leaderboard at the time. We then at the end — and I'll skip forward a little bit — the end of 2025, there were quite a few coding apps on the leaderboard, but they were all IDEs or terminal-based agents. And at the end of 2025, we saw Open Claw appear. And Open Claw was particularly interesting because one, it was a new form factor that brought in a new type of user — not just a developer but a productivity or internet creator came to AI for the first time. And it also had an interesting architecture where it was calling your chosen model for these heartbeats to see if it was still alive in addition to actually using the model for real tasks. And the heartbeats are like — you don't want to pay a lot for a heartbeat. So the auto router that we provided was really useful to this wide range of users all of a sudden. And so we just saw it rocket exponentially and then we saw Open Claw just blow up and a couple other apps lean into that new paradigm and do something similar. Hermes came out and really leaned into things like the auto router and built a really good community and leaned into skill management and making it really easy and effective for people to set their memory in the agent and build really good skills.

这是你从未做过的另一件事——记忆、技能、沙盒，所有这些你可以做的相邻事情。

Which is another thing you never did — memory, skills, sandboxes, all these adjacent things you could have done.

本可以，但我觉得很难下注。它们也非常——有些东西对当时出现的开发者用例非常重要。开发者想要架构那些东西。这些对构建良好的用户体验至关重要。公司很难找到适用于所有开发者的记忆层抽象。有一些——比如 Mastra 做得相当不错——但开发者对它们有很多不同的偏好。然后我们看到我们的排行榜随时间变化的方式，就像一部 AI 领域如何随时间变化的电影。如果你去 Wayback Machine 看看排名排行榜和应用排行榜随时间的变化，它大致展示了过去几年 AI 领域发生的事情。

Could have, but I think it's hard to bet. They're also very — there are things that really matter for the developer use cases that were coming out at the time. Developers wanted to architect those things. Those were kind of critical to building a good user experience. It's been hard for companies to find abstractions that work for all developers on the memory layer. There are some — like Mastra has done a pretty good job, for example — but developers have lots of varied preferences for them. And then we saw the way our leaderboard has changed over time is kind of like a movie of how the AI space has changed over time. If you just go to the Wayback Machine and look at the rankings leaderboard and the apps leaderboard over time, it sort of shows you what's happened in AI over the last couple of years.

对我来说，成熟时刻是 Andrej Karpathy 说：‘我不再读本地 llama 了，因为我直接去 OpenRouter 的排行榜。’

To me, the coming of age moment was Andrej Karpathy saying, 'I no longer read local llama because I just go to OpenRouter's leaderboard.'

我记得那个。我想他大概说了类似‘抱歉各位，我要给你们送一堆流量了。’

Which I remember that. I think he probably said like, 'Sorry guys, I'm going to send a bunch of traffic to you.'

所以我还想把它带入 Stripe 的事情。那种对话是怎么开始的？

So I also want to bring it into the Stripe thing. How does that kind of conversation start?

我们与 Stripe 有着长期的关系，来自我们与他们合作过的许多不同项目。我们在打击滥用和 token 欺诈方面投入了大量精力。

We had this longstanding relationship with Stripe from many different projects that we had worked on with them. We invest a lot of effort in countering abuse and token fraud.

Token 欺诈。你能给一些数字，让人们理解吗？

Token fraud. Can you give some numbers just so people understand?

我想我发过关于这个的帖子。我们上个月阻止的美元交易量是前一个月的 10 倍。而且 token 欺诈的类型正在多样化。有欺诈者针对典型的被盗信用卡，但也有试图违反服务条款转售流量的人。有被黑账户。有些人只是失去了控制——整个公司被入侵，他们甚至没有意识到。我们帮助他们重新获得控制并检测到它。

I think I posted about this. We blocked 10x as much dollar volume last month as the month before. And the types of token fraud are diversifying quite a bit. There are fraudsters going after typical stolen credit cards, but they're also people trying to resell traffic against the terms of service. There's hacked accounts. There's people who just lose act — their whole company is compromised and they don't even realize it. And we help them regain control and detect it.

有些账号在偷偷转售推理服务。还有些账号遇到了意外失控的智能体，自己却没意识到。不是黑客攻击，而是某种失控爆发，公司不想要这种情况。所以我们的信任与安全团队在这些类别的问题上做了大量工作，帮助拦截和检测。我们围绕这些问题构建了模型。我们和 Stripe 紧密合作了一段时间。我认为这将在生态系统中成为一个巨大的问题。我们已经看到很多公司开始发现这些欺诈者扩散，并在 OpenRouter 之外寻找其他欺诈途径。如果你在做网关或销售通用推理服务，你就是欺诈的目标。如果你销售的是非常离散的智能产品——做非常具体的事情，而不只是转售推理加一些附加能力——那么你遇到这些欺诈者的可能性就小得多。所以我认为我们会看到公司不再只是转售推理加一些附加能力，而是转向离散任务，对这些任务和增强功能收费，并让人们以第三方方式自带推理。

There are accounts that are reselling inference on the side. There are accounts dealing with an accidental runaway agent, and they don't realize it. Not a hack, but something that blows up, and the company doesn't want it. So our trust and safety team works a lot on all of these categories of problems and helps block and detect them. We've built models around them. We worked closely with Stripe for a while on this. I think it's going to become a huge problem in the ecosystem. We're already seeing a lot of companies start to see these fraudsters spread and look for other fraud vectors beyond OpenRouter. If you're making a gateway or selling generalized inference, you are a target for fraud. If you're selling very discrete intelligence products—products that do something pretty specific, not just reselling inference with some added capability—then you're way less likely to get these fraudsters. So I think we'll see companies move away from just reselling inference with some added capability and move towards discrete tasks, charging for those tasks and enhancements, and letting people bring their own inference in a third-party way.

哇。好的。而且显然你们会为那个提供支持。但人们是按结果付费还是按任务付费？

Whoa. Okay. And yeah, obviously you would power that. But people pay for outcomes or per task?

我认为人们会付费——我觉得 Datadog 的定价页面很好地展示了未来的样子。像基础设施公司这样的企业会针对他们提供的不同类型的事件收费，会有很多类似这样的持续定价模型。当然，如果你往下走到消费者应用，定价会更简单，更多订阅制，需要担心的事件更少，而且不会只专注于在推理上加价——不仅因为欺诈很难，还因为来自实验室和优秀推理提供商的压力会非常大，要求你做出承诺，然后把推理带到别处。

I think people will pay—I think the Datadog pricing page is a good look at the future to come. Companies like infrastructure companies will charge for different types of events that they're providing, and there'll be lots of continuous pricing models that look like that. And of course, if you go down towards consumer apps, simpler pricing, more subscriptions, fewer events to worry about, and ones that are not focused on just adding a markup on top of inference—not just because fraud is hard, but also because the pressure from the labs and from good inference providers to do a commit and then bring your inference elsewhere is going to be very high.

有什么想说的吗？

Any comments?

两点。第一，我认为 Alex 非常雄辩地描述了一件反直觉的事，我四年前就知道这会在规模上发生，因为 Discord。教会我这一点的具体经历是，当我们开始扩展 Midjourney 时。早期我们赠送或让人们尝试 Midjourney 的主要方式之一——让他们完成前 10 次生成，因为我们发现 10 次生成、每次 10 张图大约是神奇时刻的激活点；一旦你做了 10 次，你就会觉得这太非凡了。但为此，我们有一个 Midjourney 的免费试用。有一天我醒来——因为他们有一个平台需要监控，我有一堆仪表盘——我接到 David 的三个未接来电。结果是一夜之间涌入了大量新用户，我们说这太好了。他说，不，实际上我们关掉了免费试用。我说为什么？他说，我想看看地理位置 IP 地址。基本上是中国有人开始转售 Midjourney 免费订阅，利用免费试用——基本上就是欺诈滥用，对吧？

Two things. One, I think Alex has done a very eloquent job of describing something counterintuitive that I knew would be a thing at scale like four years ago because of Discord. The particular experience that taught me this was as we started scaling Midjourney. One of the primary ways we used to give away or get people to try Midjourney early on—to get to their first 10 generations, because 10 generations of 10 images generated was roughly the magic moment activation point we found; once you'd done 10, you were like, this is extraordinary. But for that, we had a free trial with Midjourney. One day I woke up—because they had a platform and had to monitor, I had all these dashboards—I had like three missed calls from David. It turns out there had been this flood of new users overnight, and we were like, this is great. And he was like, no, actually, we shut down the free trial. And I was like, why is that? And he said, I want to look at the geolocation IP addresses. And basically somebody in China had started to resell Midjourney free subscriptions with the free trial as a way to—basically it was fraud abuse, right?

即使是像 Midjourney 这样的专用模型？

Even for a specialized model like Midjourney?

是的，那实际上是一个应用。所以我当时的大局观是：嘿，互联网上正在流动一种新的价值单位，叫做 token，未来 10 年整个互联网价值链都将不得不面对一个事实：token 越有价值，就会有越多坏人试图染指这些 token。任何时候你扩展某样东西，载荷变得越来越有价值，就会有更多坏人试图获取那份价值。所以当时这对我来说非常明显。直到今天，我不认为有免费层级——我认为 Midjourney 从那以后就再也没开启过免费试用，因为在信任与安全方面这真的不是一个容易解决的问题。这就是为什么我开始在斯坦福教授“大规模安全”这门课——那件事以及 Anthropic 的经验。对我来说很清楚，几年后对大规模安全的需求将极其巨大，因为如果你算一下：想想在线支付——大约始于 80 年代和 90 年代，在接下来 10 年增长到超过 1 万亿美元，我们需要构建全新的支付解决方案来应对在线欺诈。我们今天在 token 上大致处于那个位置，但在未来五年内，我们预计 token 经济将达到大约 5 万亿美元，而在未来 10 年内，如果我们达不到 10 万亿美元的 token 流量，我会很震惊。所以如果我们已经在 subscale 的 Midjourney 上看到如此激进的滥用和欺诈——记住当时 Midjourney 的年收入运行率还不到 3 亿美元——我就意识到我们需要全新的系统来应对试图进入 token 流所会发生的欺诈。所以当你提到 Stripe 想要合作时——我忘了是哪次董事会会议——这对我来说太合理了，因为 Stripe Radar——10 年前我在红杉做合伙人时，我们投资了 Stripe，Patrick 和 John 雄辩地传达的整个推介是：嘿，不像 Braintree 这样的传统支付工具要做 7 天的验证，比如 KYC 和邮件来排除欺诈，我们实际上只是把欺诈成本预先作为客户获取成本承担下来，告诉开发者，只需用五行代码，我们就能在 5 分钟内开始接受你的付款，然后随着时间推移，我们会收集所有关于开发者的数据。

Yeah, and that was actually an application. So the big picture realization I had back then was: hey, there's a new type of unit of value being streamed across the internet called a token, and over the next 10 years the entire internet value chain was going to have to deal with the fact that the more valuable tokens got, the more bad actors were going to try to get their hands on those tokens. Anytime you scale something and the payload gets more and more valuable, more bad things people try to get access to that value. So it was very obvious to me back then. To this day, I don't think there's a free tier—I don't think Midjourney's ever actually turned on the free trial since then, because it was really not an easy problem to solve in terms of trust and safety. That's why I started teaching the class Security at Scale at Stanford—that and the Anthropic learnings. To me it was clear that the need for security at scale was going to be enormous a few years from then, because if you just do the math: think about online payments—it started roughly in the 80s and 90s, grew to over a trillion dollars over the next 10 years, and we needed to build entirely new payment solutions to deal with online fraud. Where we are today is roughly there on tokens, but over the next even five years we're expecting the token economy to get to roughly $5 trillion, and over the next 10 years I'd be shocked if we didn't get to $10 trillion of token flow. So if we were starting to see such aggressive abuse and fraud at subscale Midjourney—remember Midjourney at this point was like less than $300 million revenue run rate a year—I just realized we were going to need entirely new systems to deal with the fraud that was going to happen for trying to get into the token flow. So when—I forget the board meeting—when you brought up that Stripe wanted to partner up, it made so much sense to me, because Stripe Radar—when I was a Sequoia partner 10 years ago, we invested in Stripe, and the whole pitch that Patrick and John communicated so eloquently was: hey, unlike traditional payment tools like Braintree that do a 7-day verification like KYC and email to get the fraud out of the way, we actually just bite the fraud cost up front as a customer acquisition cost and tell a developer, just use five lines of code and we'll start accepting your payments in 5 minutes, and what'll happen is over time we'll collect all this data on the developers.

Cloudflare 模式。

Cloudflare model.

就是 Cloudflare 模式，对吧？他们确实做到了。5 年后，他们推出了 Stripe Radar，而 Stripe 今天实际上是一家安全公司。人们以为它是一家支付公司。不，今天有很多其他支付提供商给你更便宜的支付传输，但 Stripe 在美国和欧洲保持主导地位的原因，是因为他们多年来构建了非凡的欺诈检测能力。

Is the Cloudflare model, right? And they did. 5 years later, they launched Stripe Radar, and Stripe really today is a security company. People think it's a payments company. No, the reason there are lots of other payments providers today that give you cheaper payments transmission, but the reason Stripe keeps being the dominant one here in the US and Europe is because they have extraordinary fraud detection that they've built over the years.

Elon 和 Max Levchin 也是同样的故事，还有

The same story with Elon and Max Levchin and

还有 Affirm。是的。

and Affirm. Yeah.

你知道，我觉得这个故事反复出现：每当有大量价值在全球流动时，你都需要新的保护和安全基础设施，来把坏人挡在外面，让好人能够快速完成交易。所以我认为，从我的角度看，Stripe 和 OpenRouter 的故事是一个关于互联网生态系统、关于前沿 AI 生态系统的安全故事。没有这样的合作，就很难在不让坏人捣乱的情况下，捍卫体验质量、速度以及所有好的东西。

You know, I think the story shows up over and over again where every time you have value streamed across the world in large amounts, you need new protection and security infrastructure to fight to keep the bad guys out and allow the good people to have their transactions happen really fast. And so I think this is why, from my perspective, the Stripe and OpenRouter story is a security story for the internet ecosystem, for the frontier AI ecosystem. Without a partnership like that, it becomes very hard to defend the quality of experience and the speed and all the good stuff without letting the bad guys get in the way.

第二点是，有一个被低估的事实：Alex 描述的那些目前由人类实施的坏事，在未来 10 年将由 AI 智能体来实施，对吧？所以想想我们即将看到的坏行为者的递归规模。这不仅仅是坏人类，还有所有那些将攻击代币流的坏智能体。如果你是一名研究人员或 AI 实验室，很难推理这个问题，因为你唯一的数据是你训练的智能体如何越轨。但这只是我们将在互联网上看到的所有坏行为的一小部分。所以你需要的是防御者，是戴着牛仔帽的新警长，能够看到整个生态系统中来自不同模型实验室、不同后训练部署和不同开发者的 AI 智能体的所有坏行为，并利用所有这些数据说：我们要为整个代币经济打造一面盾牌。因为如果没有这个，我们在这个 10 万亿美元 GMV 和全球 GDP 增长中看到的欺诈，我认为其中很大一部分将是欺诈、滥用，如果人们不信任代币，我们可能永远无法实现这个目标，对吧？而且我不认为这种基础设施已经存在。所以你在 Stripe 的工作很艰巨，但我认为人们还没有意识到智能体式欺诈——即由 AI 智能体实施的坏行为——即将像海啸一样冲击我们的规模。

The second thing is that there's this underappreciated fact that all the bad things that Alex described as being perpetuated by humans right now are going to be perpetuated by AI agents over the next 10 years, right? So think about the recursive scale we're about to see of bad actors. It's not just bad human beings. It's all the bad agents that are going to be attacking the token flow. And it's very hard if you're a researcher or an AI lab to reason about that problem because the only data you have is how the agents you're training are going rogue. But that's just a fraction of all the bad behavior on the internet that we're going to see. And so what you need is defenders, new sheriffs in town with cowboy hats, that can see all the bad behavior from AI agents across the ecosystem, from different model labs and different post-trained deployments and different developers, and take all of that data and say we're going to build a shield for the entire token economy. Because without that, the amount of fraud we're going to see out of this 10 trillion dollar in GMV and global GDP growth is like a huge percentage of that, I think, is going to be fraud, abuse, and we might never get there if people just don't trust tokens, right? And I don't think this infrastructure exists. So you have your work cut out for you at Stripe, but I don't think people have realized the scale at which agentic fraud, like bad behavior perpetuated by AI agents, is about to hit us like a tsunami.

是的。我的意思是，这里面有很多可以深入探讨的。我想给你最后发言的机会。我们确实得收尾了。人们可以从 OpenRouter 和 Stripe 期待什么？

Yeah. I mean, there's a lot to dig into there. I want to give you the last word. We do have to wrap. What can people expect from OpenRouter and Stripe?

我的意思是，我认为这是我们加速上市和更快走向高端市场的一个非常好的方式。而且，正如刚才雄辩地描述的那样，在提升信任与安全、让接受代币变得非常容易、让人们将自己的推理带到你的应用、帮助开发者直接在推理之上构建方面，这里有一个非常清晰的“在一起更好”的故事。展望未来，我们在 OpenRouter 有一个非常强大的品牌，我们会保留这个品牌。所以 OpenRouter 作为一个产品、路线图、名称和品牌都保持不变。所以你在接下来 6 个月应该期待的是，大多数事情会像我们独立时那样做，只是一切都会进展得更快，这是我们近期的目标。长期来看，希望我很快能评论，但现在不能。

I mean, I think this is a really good way for us to accelerate go-to-market and to go up market more quickly. It's also, as eloquently described, there's a really clear better together story here when it comes to improving trust and safety and making it really easy to accept tokens and let people bring their own inference to your app and to help developers just build on top of inference. Going forward, we have a really strong brand with OpenRouter and we're keeping the brand. So OpenRouter as a product and the road map and the name and the brand are staying the same. And so what you should expect in this next 6 months is that most things will be like what we would have done had we been independent, except everything will be moving faster, and that's kind of our near-term goal. Longer term, hopefully I can comment on it soon, but I can't now.

好的。希望我们以后能做个后续。但感谢你如此慷慨地抽出时间，祝贺你们达成合作。我的意思是，这是我见过的最美好的兄弟情之一，从

Okay. Well, we'll hopefully do a follow-up at some point. But thank you for being so generous with your time and congrats on the partnership. I mean, this is one of the most beautiful bromances I've seen in a

才刚刚开始。

Just starting.

从斯坦福开始到现在。

Starting from Stanford to here.

还有很多事要做。很多警长要维护城镇治安，代币经济。我们肯定需要新警长。

Lots more to do. Lots of sheriff policing to do of the town, the token economy. We need new sheriffs for sure.

是的。太棒了。谢谢。

Yeah. Awesome. Thank you.

谢谢。

Thank you.