Latent Space
OpenRouter: from Seed to Stripe — with OpenRouter’s Alex Atallah & AMP’s Anjney Midha
00:00
00:00
Okay, we are here in Anjus House, which is where all great startups in San Francisco start.
好，我们现在在 Anjus House，这里就是 San Francisco 所有伟大创业公司起步的地方。
00:08
Howdy.
你好啊。
00:10
And congrats on Cursor, Mr. All.
也恭喜 Cursor 啊，Mr. All。
00:13
I don't know what else you got so much stuff going on.
我不知道你还有啥，你手头事情也太多了。
00:17
There's a lot going on.
确实有很多事情在忙。
00:18
Well, open router is probably the most, I would say, one I'm excited about, Chris.
嗯，open router 大概是，要我说，我最兴奋的一个，Chris。
00:24
Yeah, yeah.
对对。
00:25
And we have Alex, first time on the pod, but you've been there every year, a few times.
还有 Alex，第一次上这个播客，但你每年都来，来过好几次了。
00:29
I appreciate every time you've shown up for the community.
很感谢你每一次都为社区站出来。
00:31
Congrats.
恭喜。
00:32
I just, like, what a journey.
我就觉得，哇，真是一段旅程。
00:34
When I was looking back at your past post, one of the earliest principles that I saw
我回头看你的旧帖时，看到你作为一个产品人最早写下的原则之一，就是把 PubSub 作为一种产品原则。
00:38
you write as a sort of product person, is PubSub as a product principle.
我想让你，嗯，解释一下，你是怎么思考这个世界上应该存在什么的。
00:42
And I wanted to, you to maybe explain how you think about what should exist in the
是啊，PubSub 那部分，是 2023 年初的，直到我们聊了之后我才想到它。
00:48
world.
世界。
00:49
Yeah, the PubSub piece, which was early 2023, I didn't think about it until we talked
是啊，PubSub 那块，是 2023 年初的事，直到我们聊过之后我才开始想它
00:55
like 10 minutes ago, is about how there is like a way of thinking about products as an
就大概10分钟前吧，讲的是有这么一种思考产品的方式，把产品看成是
01:01
intersection between subscribing to data and publishing data.
subscribing to data 和 publishing data 之间的交集。
01:06
And marketplaces are an easy, easy, easy example of this.
而 marketplaces 就是一个非常、非常、非常简单的例子。
01:10
You have suppliers that are publishing some kind of products to a skew.
你有 suppliers 会把某种产品 publishing 到一个 skew。
01:16
And the skew is kind of like a PubSub topic that a consumer is subscribing to and just
而这个 skew 就有点像是一个 PubSub topic，consumer 订阅它，然后
01:22
going to consume whatever they want.
就去 consume 他们想要的任何东西。
01:24
And humans consume in a very like a discrete ad hoc way.
而人类 consume 的方式非常像一种 discrete、ad hoc 的方式。
01:29
It's not very scalable.
这不太 scalable。
01:31
All their attention is on the topic when they're buying the thing and their attention is
他们买这个东西的时候，全部注意力都在这个话题上，而一旦发生购买，注意力就不在别处了。
01:35
nowhere else when that happens.
Agents 和 inference 的消费者不会这样。
01:39
Agents and consumers of inference don't act like that.
他们一直在持续消费，而且一直在改变自己消费所依据的 skews。
01:43
They're consuming continuously and they're changing the skews that they consume from all
所以 OpenRouter 是普通 API 体验和 marketplace 的混合体，我们会在这里创建 model slug。
01:48
the time.
我们有 auto router。
01:49
So open router is a blend between a normal API experience and a marketplace where we create
所以 open router 是普通 API 体验和 marketplace 的混合体，我们会在里面创建
01:56
model slug.
model slug。
01:57
We have the auto router.
我们有 auto router。
01:58
We have all kinds of product skews that you can subscribe to and anything like continuously
我们有各种各样的 product skews，你可以订阅，而且差不多可以持续
02:04
at a derived value and make decisions based on those consumers.
按某个 derived value 来，并基于那些 consumers 做决策。
02:12
This is something that was more consensus now, but not consensus when you guys started,
这件事现在算更有共识了，但你们刚开始的时候还不是共识，
02:17
which is that there is such a demand for swapping models and changing things out and that people
也就是，大家对 swapping models 和换东西有这么大的需求，以至于人们
02:23
would not use the native SDKs.
不会去用 native SDKs。
02:25
I guess for each of you, what was your sort of realization moment that this would be
我猜，对你们每个人来说，你们那种意识到这会是
02:30
it?
‘就是它了’的时刻是什么？
02:31
You've given a talk at the IE about Alpaca as one of your inspiring moments.
你曾在 IE 做过一个关于 Alpaca 的演讲，把它作为你受到启发的时刻之一。
02:35
Alpaca, I can like rehash the Alpaca moment for a sec like the very beginning at the end
Alpaca，我可以稍微把 Alpaca 时刻再讲一遍，就像 2022 年底最开始那会儿，OpenAI 是当时唯一的玩家。
02:41
of 2022, opening I was the only game in town.
当时有像 OpenAI、Cohere，然后还有一些零星的早期 open weight models 尝试。
02:44
There was like opening I co here and then a smattering of early attempts at open weight
当 Wama 在 2023 年 1 月出来的时候，就感觉是，哇，真的很令人兴奋。
02:53
models.
这真的很重大。
02:55
When Wama came out in January of 2023, it was like, wow, really exciting.
它在一两个 benchmarks 上超过了 GPT-3，但你不能跟它聊天。
02:59
This is really big.
它其实并不是一个很吸引人的 model，但看起来只是需要有人来修一下。
03:00
It outperforms GPT-3 on one or two benchmarks, but you can't chat with it.
它在一两个 benchmark 上比 GPT-3 表现更好，但你没法跟它聊天。
03:06
It wasn't actually an engaging model, but it seemed like someone just needed to fix
它其实并不是一个很吸引人的 model，但看起来只是需要有人去修一下
03:11
a couple of things and do some RLHF on it to get it all the way there.
做几件事，然后对它做一些 RLHF，把它一路推到那个水平。
03:15
Alpaca was the first model that I saw that did that.
Alpaca 是我看到的第一个做到这一点的模型。
03:20
It only took $600 to do a team at Stanford, generated a bunch of synthetic data, fine-tuned
Stanford 的一个团队只花了 $600，生成了一堆 synthetic data，fine-tuned
03:27
Wama, and made Alpaca, 7 billion parameter model, maybe it was 13 billion parameters.
Wama，然后做出了 Alpaca，7 billion parameter 模型，也许是 13 billion parameters。
03:33
It was so good.
它真的太好了。
03:35
I was just on an airplane using it.
我当时正好在飞机上用它。
03:38
In many cases, it could not discern a chat GPT versus an Alpaca result, and I figured
很多情况下，它分辨不出 chat GPT 和 Alpaca 的结果，我就想
03:45
it was this easy to make a model.
做一个模型居然这么容易。
03:49
One, we have a whole new way of monetizing data for the first time.
第一，我们第一次有了一种全新的把 data 变现的方式。
03:54
You can just take really valuable data and turn it into a service in $600, and that
你完全可以把真正有价值的 data，用 $600 就做成一项 service，而且这个
04:00
cost will probably go down over time.
成本很可能会随着时间下降。
04:03
When you say monetizing your data as what eventually will become an MCPM point, or as a training
你说的把 data 变现，是指它最终会变成一个 MCPM point，还是指作为 model 的 training
04:10
data for a model.
data。
04:11
Yeah, training data for a model, an abstract way of saying, hey, I have this data.
对，model 的 training data，一种抽象的说法就是，嘿，我有这些 data。
04:15
I press it into a model.
我把它压进一个 model。
04:17
It makes sense for me in my product, but I could re-package it in the form of a model
它在我的产品里说得通，但我也可以把它重新打包成 model 的形式
04:23
and sell it, and so it's just a whole new business model for the economy.
然后把它卖掉，所以这对整个经济来说就是一个全新的 business model。
04:28
It also, of course, provides a way of following what frontier labs are doing, but in a way
当然，它也提供了一种方式，让你能跟进 frontier labs 在做什么，但是以一种
04:38
that a single developer or a small team of developers can rule on their own.
单个 developer 或一个小型 developer 团队可以自己说了算的方式。
04:45
Whenever you have an example of that, a breakout app that's doing really well, and then some
每当你有一个这样的例子，一个表现特别好的 breakout app，然后还有某种
04:51
kind of framework for imitating it in your own flavor, you have an immediate ecosystem
用来按你自己的风格模仿它的 framework，你就立刻有了一个 ecosystem
04:58
of like an intermediate ecosystem should arise because there's just a huge gap between
——就好像应该会出现一个 intermediate ecosystem，因为这两者之间有一个巨大的 gap：
05:04
the decisions that the single company is making, and all of the variations in those decisions
单个公司正在做的决策，以及这些决策的所有变体，
05:10
that a wider ecosystem can create themselves, and so then you need a marketplace to discover
而这些变体是一个更广泛的 ecosystem 可以自己创造出来的，所以这时候你就需要一个 marketplace 来发现
05:16
all of those services and all of those products.
所有这些服务，以及所有这些产品。
05:20
There wasn't any place on the internet that was a home base for LLMs in terms of seeing
互联网上没有任何一个地方能作为 LLMs 的大本营，用来查看
05:25
how much they were being used, and seeing who was using them and why.
它们被用了多少，以及是谁在用它们、为什么用。
05:29
The closest would be hugging face, they just started a few years ago before that.
最接近的会是 hugging face，他们在那之前几年才刚刚起步。
05:34
Yeah, and hugging face also didn't have the close source models, and you couldn't use
对，而且 hugging face 也没有 closed source models，而且你也没法使用
05:40
the models at the time, and there wasn't data about who was using that.
那些模型，而且当时也没有关于谁在使用它们的数据。
05:45
There were like a bunch of differences between open router and hugging face, and those differences
open router 和 hugging face 之间好像有一大堆差异，而那些差异
05:51
felt really critical to me, especially when I was just trying to learn about LLMs, and
对我来说都特别关键，尤其是当我只是想了解 LLMs 的时候，而且
05:57
why people are choosing these different little ones that are emerging over time.
为什么人们会选择这些随着时间不断涌现出来的、各不相同的小家伙呢？
06:03
Got it.
明白了。
06:04
And an unsh, no stranger to wanting more model diversity at the time, you know, a couple
而 Anush，当时对想要更多 model diversity 并不陌生，你知道，在你 Anthropic 旅程走到几年的时候，这在之前的 podcast 里你也聊过。
06:10
of years into your anthropogenic journey, which you have covered in the previous podcast
你最初是怎么认识 Alex 的？
06:13
as well.
嗯，那次认识，我想是在那之前的 13 年，但 OpenRouter 的那次 handshake 其实就发生在那边，如果你还记得的话，也就是 Alex 和我初次见面的地方。
06:14
What was your introduction to Alex?
你当初是怎么认识 Alex 的？
06:16
Well, the introduction was, I think, 13 years before that, but the open router had
嗯，那次介绍，我想是在那之前的 13 年，不过 open router 的
06:21
handshake actually happened right over there if you remember, which was, Alex and I met,
handshake 其实就发生在那边，如果你还记得的话，那也就是 Alex 和我认识，
06:27
I believe it was sophomore's, now I remember, at a Stanford review, eating for the first
我想那是在大二的时候，现在我想起来了，是在 Stanford review，第一次吃饭，我觉得。
06:33
time thinks.
对，对。
06:34
Yeah, yeah.
所以 Stanford review 是 Stanford 校园里的 Libertarian 报纸，是 Peter Tiel 当年创办的，而不知为什么，Alex 和我都去参加了一次他们的会议。
06:35
So Stanford review was the Libertarian newspaper on campus at Stanford that Peter Tiel
我记得当时的主编是我们俩共同的朋友 Lisa，她真的是一个非常非常棒的主编。
06:39
started back in the day, and for whatever reason, in Alex and I both showed up to one
从那会儿开始的，而且不知为什么，Alex 和我都来参加了其中一次
06:46
of the meetings.
会议。
06:47
And I remember the editor-in-chief was a mutual friend of ours, Lisa, was really a really
我记得当时的主编是我们俩的共同朋友 Lisa，她真的是一个非常非常
06:53
great editor-in-chief.
棒的主编。
06:54
You know, part of an editor-in-chief's job is to assign responsibilities to people and
你知道，主编的一部分工作就是给人分配职责，并且
06:58
make sure the work gets done.
确保工作能完成。
07:01
And I, I may be misremming the details, but I remember wanting to, it was kind of surprising
而且我，我可能记错细节了，但我记得我当时想，让我有点意外的是
07:07
to me that at the time there was no dedicated technology section in the newspaper, you
那时候报纸里居然没有专门的科技版块，你
07:12
know.
知道吗。
07:13
Because it's political, right?
因为这是政治性的，对吧？
07:14
It's like a political, yes, yeah, states, correct things, yeah.
这就像是政治性的，对，是的，州，正确的事情，对。
07:18
But to take us back in time, you may remember this, but there was, there was this technology,
但让我们回到过去，你可能还记得这个，但当时有，有这么一个技术，
07:25
kind of legislation that was being debated called the Net Neutrality Act, and Net Neutrality
当时正在被讨论的一种立法，叫做 Net Neutrality Act，而 Net Neutrality
07:31
is like inherently this political concept, right?
本质上就像是一个政治概念，对吧？
07:34
It's about the regulation of Internet broadband access.
它涉及的是对 Internet broadband access 的监管。
07:38
And so there was a community of us who were kind of technologists, but also debating the politics
所以当时我们有一群人，算是技术专家，但也在辩论其中的政治
07:43
of the technology.
也就是这项技术的政治。
07:44
And I thought the review would be a great place to, like, write about that.
而且我觉得 review 会是一个很棒的地方，可以写这些东西。
07:47
And I was working on, I think, a Net Neutrality article.
而且我当时在写，我想，一篇 Net Neutrality 的文章。
07:50
And I remember proposing, well, maybe you should start a technology kind of section.
而且我记得我提议说，嗯，也许你们应该开一个科技类的版块。
07:54
And Alex was on the only people who said, yes, that would be cool.
Alex 是少数几个说“对，那会很酷”的人之一。
07:58
And I forget whether we end up writing stuff together, but that's when we first met
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的时候
08:03
was 2011 or 2012.
是 2011 还是 2012 年。
08:06
I forget which year it was, was one of those.
我忘了是哪一年了，反正是那两年之一。
08:09
Yeah.
是啊。
08:10
Old Union, if I remember correctly, that's where we used to meet.
如果我没记错的话，是 Old Union，我们以前常在那儿见面。
08:13
But along the way, Alex and I've had a chance to hang out often.
但这一路下来，Alex 和我经常有机会一起聚。
08:18
And probably the, the time when we had the most professional overlap was when I was running
而且可能我们职业上交集最多的时候，就是我当时在经营
08:25
the platform at Discord, and it had become this explosive kind of platform for crypto.
Discord 上的这个 platform，它已经变成了 crypto 领域里那种爆发式增长的 platform。
08:32
Yeah.
对。
08:33
And NFTs in the middle of the pandemic, which also, by the way, you were in charge of safety
还有疫情期间的 NFTs，顺便说一句，你当时还负责 safety
08:37
and security as well, right?
和 security，对吧？
08:38
I was the head of platform, which meant all of the crypto-dow and NFT launch security debugging
我当时是 platform 负责人，也就是说所有 crypto-dow 和 NFT launch 的 security debugging
08:45
fell on me.
都落到了我头上。
08:46
And they're switching.
而且他们还在切换。
08:47
And the phishing, the social engineering attacks that Katana DDoS that we were getting hit
还有 phishing、social engineering attacks，还有我们当时一直被攻击的 Katana DDoS
08:52
by.
by.
08:53
It's around the time that we started teaching security, it's okay, it's Stanford, CS153.
差不多就是我们开始教 security 的时候，没事，是 Stanford，CS153。
08:57
And Alex was in the open sea at the time, and I was trying to figure out how we could defend
而且 Alex 当时在 OpenSea，我正试着搞清楚我们要怎么防御
09:03
against all these attacks that we were at Pika, if you remember how much NFT volume was
所有这些攻击，我们当时在 Pika，如果你还记得有多少 NFT volume
09:09
running through Discord, but it was a meaningful amount of like, it's like several billion
在 Discord 上流转，但那可是个很可观的量，就像，几十亿
09:13
dollars in NFT volume of GMV, so to speak, we're running through the platform, and it was
美元的 NFT volume，也就是 GMV，可以说，都在平台上流转，而且它
09:17
all coming from open sea.
全都来自 OpenSea。
09:18
It was these like buy, sell, trade, servers, the D in dollars Discord.
就是那些像 buy、sell、trade 的 servers，还有 the D in dollars Discord。
09:23
Yeah.
嗯，对。
09:24
So that's when I think we had hung out professionally, but a year after that, OpenAI gave Discord
所以我觉得，那大概就是我们已经在工作上有过交集的时候，但一年之后，OpenAI 给了 Discord
09:31
early access to GPT.
GPT 的 early access。
09:33
So GPT-3.
所以是 GPT-3。
09:34
No, it was GPT-3.5 actually, which is the RL version of GPT-3.
不，其实是 GPT-3.5，也就是 GPT-3 的 RL 版本。
09:38
And that's around the time we made a Discord bot with OpenAI for internal deployment, and
差不多也是在那个时候，我们和 OpenAI 一起做了一个 Discord bot，用于 internal deployment，然后
09:44
that's when I realized we would need, like since I was part of the deployment team, what
就是那时候我意识到我们需要，嗯，因为我是 deployment team 的一员，那个
09:50
was the use case?
use case 到底是什么？
09:51
There were two that were, and there's actually a post now called Discord is your place
当时有两个，而且现在其实有一篇帖子叫 Discord is your place
09:56
for AI with friends that somebody sent me recently that I wrote and published in 2023,
for AI with friends，是最近有人发给我的，是我在 2023 年写并发布的，
10:01
but there were two use cases.
但当时有两个 use cases。
10:04
One was Clyde, which was like a first party friend inside of Discord that could help
一个是 Clyde，它像是 Discord 里面的 first party friend，可以帮你
10:11
you set up your Discord server and talk to you about onboarding and get your friends
设置你的 Discord server，跟你聊 onboarding，还能让你的朋友们
10:15
to hang out more.
多来一起玩。
10:17
And then there was content moderation.
然后还有 content moderation。
10:20
And one of the realizations we had with content moderation was, it would refuse to
而且我们在 content moderation 上的一个发现是，它会拒绝
10:26
moderate.
审核。
10:27
It refused our prompts because the RL, the post training was, we were very early in
它拒绝了我们的 prompts，因为 RL、post training 那会儿，我们还处在非常早期的
10:31
the post training era.
post training 时代。
10:33
And our prompts would trigger it, it's like guardrails.
而且我们的 prompts 会触发它，就像 guardrails 一样。
10:38
And we told OpenAI, hey guys, we need access to the way it's because if we're going to
然后我们跟 OpenAI 说，嘿各位，我们需要能访问它当前的状态，因为如果我们要
10:41
be doing content moderation at scale, we had 250 million monthly active users, we need
要做 content moderation at scale，我们有 2.5 亿 monthly active users，我们需要
10:45
more reliability that the model will do what we needed to.
更高的可靠性，确保 model 会做我们需要它做的事。
10:48
And they said, well, sorry guys, that's not how this works.
然后他们说，呃，抱歉各位，事情不是这么运作的。
10:50
We're a closed source company.
我们是一家 closed source 公司。
10:52
And so that was my first realization that we needed open models.
所以那是我第一次意识到我们需要 open models。
10:55
The enterprises would need more control or over capabilities.
企业会需要更多的控制权，或者对能力有更多掌控。
11:00
And then ultimately would need some kind of control play and a management system to orchestrate
然后最终会需要某种 control play，以及一个 management system 来 orchestrate
11:04
these open models.
这些 open models。
11:05
But I wasn't no good up, they were no good open alternatives until maybe six months later
但当时我并不看好，市场上也没有好的 open alternatives，直到大概六个月后
11:11
when Lama came out.
Lama 出来的时候。
11:12
And six months after that, I led the Susie into Mistral, which was started by Guillaume and
然后在那之后六个月，我领投了 Mistral 的 seed，Mistral 是 Guillaume 创办的，而且
11:16
the Lama team.
Lama 团队。
11:17
And that around that time is when I remember hearing what Alex launching OpenRouter and
而大概就在那个时候，我记得听说 Alex 在启动 OpenRouter，然后
11:24
going, these worlds are going to collide.
我心想，这些世界要撞上了。
11:27
And I don't know when it'll make sense to team up, but Alex was so early and I think he
我不知道什么时候联手才有意义，但 Alex 当时实在太早了，而且我觉得他
11:32
was totally right about this ecosystem starting with Lama that then needed like an easy
完全说对了，这个生态从 Lama 开始，然后需要一个简单的
11:37
layer to manage for it, especially for I was approaching it from the enterprise perspective
layer 来管理它，尤其是我当时是从企业视角来看的
11:41
because I did not like the, that's the VP platform or discord is my job to ensure that when
因为我不喜欢那种，那是 VP platform 或 Discord，我的工作就是确保当
11:46
we deployed models to like 250 million users, they did what we wanted them to.
我们把 models 部署给大约 2.5 亿用户时，它们会做我们想让它们做的事。
11:52
And that was very hard because a few outsourced it to the labs and they controlled the guardrails
这非常难，因为有几方把它外包给了那些实验室，而他们控制着 guardrails
11:58
and their guardrails, their safety policies forbid the model from responding to your prompts.
而他们的 guardrails、他们的安全政策禁止 model 回应你的 prompts。
12:03
That was quite catastrophic.
那相当灾难性。
12:05
Yeah, but you know, a moderation is the thing that they want to support and obviously beyond
嗯，但你知道，moderation 是他们想要支持的东西，而且显然，除此之外
12:09
that, they would work, opening up woodwork with you, presumably to give you a moderation
他们会跟你合作，一起打开 woodwork，大概是为了给你一个 moderation
12:14
endpoint, which they offer for free.
endpoint，而这是他们免费提供的。
12:16
It was an interesting use case that they, so they did give us a moderation endpoint.
那是个挺有意思的 use case，他们，所以他们的确给了我们一个 moderation endpoint。
12:21
However, as you guys know, every discord server is like a mini deployment of itself.
不过，正如你们所知，每个 discord server 都像是它自己的一个 mini deployment。
12:26
And so the use case was instead of having human moderators that have to interpret the norms
所以这个 use case 是，不再需要人类 moderator 去解读
12:30
of the community, you just give the, they often like every, you know, subreddit, discord
社区规范，你只要给他们，他们通常，像每个，你知道的，subreddit、discord
12:36
server's public ones have their own rules.
服务器，公开的那些都有自己的规则。
12:39
That the user, the user's creed randomly is the squad in, yeah.
就是说用户，用户的 creed 随机就是 squad in，对。
12:43
And then humans used to read those norms and then enforce it every day manually, like observing
然后以前是人类去读这些规范，然后每天手动执行，比如观察
12:50
each message in these communities.
这些社区里的每条消息。
12:51
And these communities are like million of users.
而且这些社区都有上百万用户。
12:53
So we had a 5,000 plus person team globally on the discord content moderation team.
所以我们当时在 discord content moderation 团队里，全球有 5,000 多人。
12:58
These outsource contractors are a really tough job.
这些外包承包商真的是个很难搞的活儿。
13:01
And so the idea was instead, if you could give the norms of that server to the LLM, then
所以当时的想法是，换个做法，如果你能把那个 server 的规范给到 LLM，那么
13:07
the LLM would do custom moderation for that server.
LLM 就会为那个 server 做 custom moderation。
13:10
It's almost like a, like in context moderation for that server.
这几乎就像，像是为那个 server 做 in context moderation。
13:14
And many of those servers norms just violated open AI's rules.
而且那些 server 的很多规范本身就违反了 OpenAI 的规则。
13:19
And so that it was like, we had our own custom e-vails, so instead of each server had its
所以那就像是，我们有自己的 custom evals，所以不是让每个 server 都有它
13:23
own custom e-vail, but this at the time, open AI's e-vails, we were also primitive in
自己的 custom eval，但当时，OpenAI 的 evals，我们在这一点上也还很原始，
13:28
our thinking about how to deploy these LLMs, that often the post-training prompts were super
关于怎么部署这些 LLM 的思考上，以至于 post-training prompts 往往都超级
13:34
heavy handed.
太一刀切了。
13:35
It said, oh, anything about Harry Potter or anything that has trademarked content, you
它说的是，哦，任何关于 Harry Potter 的东西，或者任何受商标保护的内容，你
13:40
know, don't refuse.
知道吧，别拒绝。
13:42
And it was a fan, Harry Potter fan community, this is a real use case that had a content
而且那是一个 Harry Potter 粉丝社区，这是一个真实的 use case，涉及 content
13:46
moderation.
moderation。
13:47
The LLM would just refuse, and that was just not precise enough.
LLM 就是会直接拒绝，而那根本不够精确。
13:51
Another one that we heard was like, if someone was trying to write like a detective story,
我们听到的另一个例子是，如果有人想写个侦探小说，
13:58
and there's one chapter with a lot of violence, like maybe someone killed someone, the LLMs
而且其中有一章有很多暴力内容，比如可能有人杀了人，那些 LLM
14:03
would just refuse to help with that part of the story.
就会直接拒绝帮忙处理故事里的那部分。
14:08
And then these would be like, okay, this is not structurally inherent to LLMs, there
然后他们就会说，好吧，这不是 LLM 结构上固有的，
14:13
must be some choice out there so that I can switch to another model when I'm getting
肯定有某种选择，让我能切换到另一个 model，当我从
14:19
like a refusal or a bad result from the main one that I have.
主用的那个 model 那里得到拒绝或者糟糕结果时。
14:24
And that tension also drove me for a marketplace.
而那种张力也促使我去做一个 marketplace。
14:28
Yeah.
对。
14:29
I think that is well accepted now.
我觉得现在这一点已经被大家普遍接受了。
14:32
What was it like back then when you were raising, or starting this, did people get it?
当年你在融资、或者刚开始做这个的时候，那是什么感觉？大家能理解吗？
14:37
What was some of the struggles, basically, I like getting stories out of him about how
基本上，当时都遇到过哪些困难？我喜欢从他那儿挖故事，讲的是
14:41
other VCs don't get it.
其他 VC 怎么就是不懂。
14:44
So anything you want to talk about now that early journey of a provider is done, right?
所以，既然一个 provider 的早期旅程已经走完了，你现在有什么想聊的吗，对吧？
14:51
You can obviously talk about some of the early day stuff.
你当然可以聊聊早期那些事。
14:54
Well, I was going to say that like, but the biggest objection we got is big model win,
嗯，我本来想说类似那样，但我们遇到的最大反对意见就是 big model win，
15:01
which is scaling loss.
也就是 scaling loss。
15:05
Yeah, scaling loss, and natural network effects are just going to kind of accrue to
对，scaling loss，而且 natural network effects 差不多就是会逐渐累积到
15:11
one company, which will be like, it'll be a Google style monopoly.
一家公司身上，那就会像是，会变成 Google 式的垄断。
15:15
Just like how Google won the search market by a large, large margin.
就像 Google 当年以巨大、巨大的优势赢得了搜索市场一样。
15:21
And you'll just be fighting for its graphs at the end, basically.
而到最后，你基本上就只是在争抢它的 graphs。
15:24
That was probably the biggest objection we got.
那大概是我们收到的最大的反对意见。
15:27
And it is interesting.
而这很有意思。
15:28
They Google won the search engine race with such a huge margin.
Google 以如此巨大的优势赢得了搜索引擎这场竞赛。
15:33
I think like, had there been more interesting benchmarks or
我觉得吧，如果当时有更有意思的 benchmarks，或者
15:39
have search engines been, have people seen them a little bit more like LLMs
如果搜索引擎被，如果人们把它们看得更像 LLMs 一点
15:44
where there are services that you can build companies on top of that might not have been the case.
也就是有一些服务，你可以基于它们建立公司，而那种情况可能就不会发生了。
15:50
But LLMs don't merely have a user interface.
但 LLMs 可不只是有一个 user interface。
15:53
They're also ways of building entirely new businesses.
它们也是打造全新业务的方式。
15:56
And a Google level monopoly would be like the Dutch East India company
而一个 Google 级别的垄断，就会像 Dutch East India company
16:01
times quadrillion in magnitude.
规模上乘以千万亿倍。
16:04
Because the whole economy and depending on the one monopoly as well.
因为整个经济也依赖那一个垄断。
16:09
So it didn't seem like, would be a really crazy outcome if that happened.
所以如果那真的发生了，看起来也不会是一个特别疯狂的结果。
16:15
And it's also less likely because the economics of creating good competitors
而且这也没那么容易发生，因为打造优秀竞争者的经济逻辑
16:21
are much more decentralized.
要 decentralized 得多。
16:25
Everything Alex said is true.
Alex 说的一切都是真的。
16:27
And I came at it from a completely different perspective, which is this is why we're here.
而我是从一个完全不同的角度切入的，那就是：这正是我们在这里的原因。
16:33
The scaling laws were never like, in my mind, we're always a feature,
scaling laws 在我心里从来都不是……它们一直是 feature，
16:36
not a bug for why OpenRadar would be very valuable.
而不是 bug，这正是 OpenRadar 会非常有价值的原因。
16:38
Because I was one of the first investors in Anthropic.
因为我是 Anthropic 最早的投资人之一。
16:41
And it was obvious to me that other researchers in our friends group,
而且对我来说很明显，我们朋友圈里的其他研究员，
16:45
and I went to a grad school for machine learning.
而且我读的是 machine learning 方向的研究生。
16:47
And I just had a lot of friends in the ML community who,
而我在 ML 圈子里就是有很多朋友，他们，
16:50
it was very obvious to us that the bidder less than holds.
对我们来说非常明显，bitter lesson 确实成立。
16:53
And so I was like, oh, fantastic.
所以我当时就想，哦，太棒了。
16:55
Now we have at least two proof points that compute scaling works.
现在我们至少有两个 proof points，证明 compute scaling 是有效的。
16:59
It was the opening eye and Anthropic.
就是 OpenAI 和 Anthropic。
17:02
And by the time I think we decided to team up on OpenRadar,
而我想，等到我们决定在 OpenRadar 上联手的时候，
17:05
I had already invested in Mistral and Black Forest Labs and Luma.
我已经投资了 Mistral、Black Forest Labs 和 Luma。
17:09
So there was multiple model companies and teams that I was working with.
所以当时我合作的有好几家模型公司和团队。
17:14
But you did other modalities, whereas this is different modalities.
但你做的是其他 modalities，而这是不同的 modalities。
17:17
Yes.
是的。
17:18
Exactly.
没错。
17:19
And it was so obvious to me that an ecosystem of different kinds of models
而且对我来说特别明显的是，一个由不同种类的 models 组成的生态
17:22
were being created.
正在形成。
17:24
And that this whole narrative of like only one company will dominate.
而且那种“只有一家公司会主导”的整套说法，
17:29
Like Google was maybe true.
比如 Google 当年也许确实是这样。
17:32
But one, I don't believe that, been two.
但第一，我不相信这一点；第二，
17:35
There was so much extraordinary innovation happening
当时正在发生太多非凡的创新。
17:37
across several different research teams.
在好几个不同的研究团队里。
17:39
But the shared problem I was noticing across all of them
但我注意到，在所有这些团队身上，共同的问题是
17:43
was often, you know, the research teams were fantastic
这些研究团队往往，你知道，特别擅长
17:45
at figuring out how to reason about new capabilities.
搞清楚该怎么对新的 capabilities 进行推理。
17:48
They think in terms of capabilities.
他们是从 capabilities 的角度来思考的。
17:50
But never are not developer mindset oriented.
但他们从来都不是以 developer mindset 为导向的。
17:53
Like, what happens after the training is done
比如说，training 做完之后会发生什么
17:55
and the checkpoint comes out, you'd be shocked how similar
然后 checkpoint 出来了，你会震惊地发现它们有多像
17:59
they are early pre-training teams at OpenAI and Tropic,
他们是 OpenAI 和 Tropic 早期的 pre-training 团队，
18:02
BFL, Mistral were in their default approach
BFL、Mistral 采用的都是他们默认的做法
18:08
to taking their research out of the lab and kind of scaling their impact,
就是把研究从实验室里带出来，然后某种程度上 scaling 他们的影响力，
18:13
which is often, oh, the checkpoint is done,
这通常就是，哦，checkpoint 做完了，
18:15
put it out as an API done.
把它作为 API 放出去，搞定。
18:18
And then they'd be crickets.
然后就没动静了。
18:20
In the case of Cloud, the first Cloud checkpoint
Cloud 的情况是，第一个 Cloud checkpoint
18:23
was actually done a year before they released it internally.
其实是在他们内部发布的一年前就已经做完了。
18:27
And then ChatriPT came out and we decided, OK, yes,
然后 ChatriPT 出来了，我们决定，好吧，是的，
18:29
it's a good idea to release a Cloud version externally.
把它以 Cloud 版本对外发布是个好主意。
18:34
And they had no plan.
可他们根本没有计划。
18:35
Like, no plan for how to get developers to actually try it out.
就是，完全没有计划去让开发者真正上手试试。
18:38
And so if you go to the Cloud 1 blog post,
所以如果你去看 Cloud 1 的那篇 blog post，
18:42
you'll notice they like three kind of developer
你会注意到他们大概有给 API 用户准备的三种开发者
18:44
examples for users of the API.
示例。
18:46
And one is a discord bot.
其中一个是个 discord bot。
18:48
And the second is Vivian, my wife's startup called
第二个是 Vivian，我妻子的创业公司，叫
18:50
Juni Learning.
Juni Learning。
18:50
And then there was like notion, because these were all friends
然后还有 Notion，因为这些人都是
18:54
of like the anthropic team, because that's
Anthropic 团队的朋友，因为这就是
18:56
how like last minute the planning was around,
当时计划有多临时，差不多就是，
18:58
hey, once the model's done training,
嘿，一旦 model 的 training 完成，
19:00
how do you get it out to the world?
你怎么把它推向世界？
19:01
There was no distribution platform
当时根本没有 distribution platform
19:03
that understood what developers needed.
一个理解开发者需要什么的东西。
19:04
All the key management provisioning,
所有关键的管理和 provisioning，
19:08
like simple endpoint management, versioning control,
比如简单的 endpoint management、versioning control，
19:12
like all these things that the scientists and researchers go,
就像那些科学家和研究员会说的所有这些事，
19:14
I mean, that's plumbing.
我是说，那不就是 plumbing 嘛。
19:15
I don't really implement the Cloud.
我其实并不真的去实现 Cloud。
19:16
And instead, Alex came at it from that perspective.
而 Alex 则是从那个角度切入的。
19:19
And so it was so obvious to me that like every single lab I was
所以对我来说，这太明显了，就像我所在的每一个实验室都
19:23
funding would spend literally sometimes billions of dollars
资金有时候真的会花掉几十亿美元
19:27
into training.
投到 training 里。
19:28
And then a checkpoint would be done.
然后就会完成一个 checkpoint。
19:30
And it'd be crickets during early access
结果 early access 期间却鸦雀无声
19:32
because they're like, oh, that's right.
因为他们会说，哦，对哦。
19:35
It's hard to use a checkpoint to make anything.
光靠一个 checkpoint 很难做出任何东西。
19:37
You're actually need a whole bunch of plumbing around it
你实际上需要在它周围搭一大堆 plumbing
19:39
to make it usable by a developer.
才能让开发者用得上。
19:41
And so by the time I think we think
所以等到那个时候，我觉得我们会觉得
19:43
you're so obvious to me that a distribution platform,
你对我来说已经再明显不过了，以至于一个 distribution platform，
19:45
like OpenRouter, was critical to have
比如 OpenRouter，对于拥有一个
19:47
an ecosystem if we wanted there to be competition to Google.
ecosystem 来说是至关重要的，如果我们想要有能跟 Google 竞争的东西。
19:50
Like, with Google, DeepBind is done training a new checkpoint.
就像 Google 那边，DeepBind 已经 training 完一个新的 checkpoint。
19:55
And then they push a button and it gets blasted out
然后他们按一下按钮，它就会被推出去
19:57
across all their surfaces from Google Docs to everywhere,
推送到他们所有的 surfaces 上，从 Google Docs 到所有地方，
20:01
even if I don't want to.
哪怕我根本不想这样。
20:02
Everywhere, you want to know about Android.
不管在哪儿，你都想了解 Android。
20:04
Like overnight, they can deploy a new checkpoint
就像一夜之间，他们就能部署一个新的 checkpoint
20:06
to like a billion devices, right?
到差不多十亿台设备上，对吧？
20:09
And that invisible, in for advantage, distribution advantage.
还有那种看不见的、in for advantage、分发优势。
20:12
Most people don't realize, but until OpenRouter showed up,
大多数人没意识到，但在 OpenRouter 出现之前，
20:14
you had to think about all of that yourself as a model lab.
作为一家 model lab，所有这些你都得自己考虑。
20:17
And it was very daunting.
而且这真的让人望而却步。
20:19
And on the topic, I think it took more than 12 months
而且说到这个话题，我觉得这花了超过 12 个月。
20:22
to get to our first 10 million in revenue.
达到我们第一个 1000 万美元的营收。
20:24
And contrast with Black Forest Labs,
再对比一下 Black Forest Labs，
20:27
I remember the early days you guys had a conversation
我记得早期那会儿，你们有过一次对话
20:31
with the BFL team.
是跟 BFL 团队。
20:33
And it was so simple for OpenRouter to say,
而 OpenRouter 当时说起来特别轻松，
20:37
oh, no problem, like the day you launched,
哦，没问题，你们上线那天，
20:39
we can send a million developers to you.
我们就能给你送过去一百万个开发者。
20:41
You know, that was crazy.
你知道吧，那太疯狂了。
20:43
That was like a step function change in like, power.
那就像是，在 power 上来了个 step function 式的变化。
20:46
Is that a real number of million?
那真的是百万这个数字吗？
20:47
I think today it's like a million.
我觉得今天大概就是一百万吧。
20:50
How many developers are on OpenRouter today?
今天 OpenRouter 上有多少开发者？
20:52
We know over 10, but yeah.
我们知道超过 10，但，嗯。
20:54
Over 10 million, but like, it's hard to do.
超过一千万，但，就，这挺难做的。
20:59
I don't know how to, yeah.
我不知道该怎么，嗯。
21:00
We do a lot of like, you know, account deduping work,
我们会做很多，就，你知道的，account deduping 的工作，
21:03
but you know, if you could get a thousand developers,
但你知道，如果你能找来一千个开发者，
21:06
just to put in context, if you get a thousand developers
这么说吧，如果你有一千个开发者，
21:08
who actually try the model on day one after you release it,
他们会在你发布后的第一天就真的去试这个 model，
21:12
and just like do inference and give you feedback,
然后就直接跑 inference，给你反馈，
21:15
that's a thousand more developers
那你就多了一千个开发者
21:16
than they knew how to get to on their own.
比他们自己知道怎么触达的还多。
21:19
Well, you know, BFL had a reputation.
嗯，你知道，BFL 是有名声的。
21:20
But yes.
但没错。
21:21
They had a stable diffusion.
他们当时有 stable diffusion。
21:23
And with Mistral, I don't know if you guys remember,
然后 Mistral 那边，我不知道你们还记不记得，
21:26
but the first checkpoint they released was like torrents.
但他们放出的第一个 checkpoint 就跟 torrents 一样。
21:30
There was like torrent weights.
就是那种 torrent weights。
21:31
Yeah, they just put up a magnet link.
对，他们就直接放了个 magnet link。
21:33
There was no API, because they weren't in for people.
没有 API，因为他们当时不是面向大众的。
21:36
You know, like, saying, OK, download these weights
你知道，就像在说，OK，把这些 weights 下载下来，
21:39
and you guys go ahead and have a story on his side, yeah.
然后你们就自己去他那边搞出个故事来吧，对吧。
21:41
Yeah, I mean, in addition to the building,
对，我是说，除了这栋楼本身，
21:44
a really good developer experience around it,
还有围绕它打造得非常好的 developer experience，
21:46
the marketing that we do for different models
我们给不同 model 做的营销
21:51
is totally different and perceived totally differently
是完全不一样的，用户的感知也完全不同
21:54
from the marketing that a model lab does for itself.
跟 model lab 给自己做的营销相比。
21:56
Yes, we are like a neutral layer looking at this market.
是的，我们就像是观察这个市场的一个 neutral layer。
22:01
Like it's a big dark room with all the corners,
就像一个大黑屋，所有边边角角
22:04
completely obscure to users and users walking into the room
对用户来说都完全看不清，而用户正走进这个房间
22:07
and like feeling around and trying to figure out
而且就像是在四处摸索，想弄清楚
22:10
what objects to grab off the tables
该从桌子上抓哪些东西下来
22:12
and like build into their companies.
然后像是要把这些东西搭进自己的公司里。
22:16
And it's an insane way of working.
而且这种工作方式简直疯狂。
22:18
Models are not products where you can just
Models 并不是那种你可以直接
22:21
enumerate all their features onto a web page.
把所有 features 都罗列到一个 web page 上的产品。
22:24
They're all black boxes, including the open weight ones.
它们全都是 black boxes，包括那些 open weight 的。
22:27
So you need to shine lights on all corners of this room
所以你得拿灯照亮这个房间的每一个角落
22:32
so that people can see what makes this model good.
这样人们就能看到这个 model 好在哪里。
22:34
And you need the company shining that light
而且你需要公司把那束光照出来。
22:37
to be a neutral third party, which is what we specialized in.
成为中立的第三方，这正是我们专攻的方向。
22:41
So in addition to developer experience,
所以除了 developer experience 之外，
22:45
there's also a very important marketing
还有一个非常重要的 marketing
22:48
and product packaging component
以及 product packaging 这个部分
22:50
and a way of routing and discovering models
以及 routing 和发现 models 的方式
22:55
becomes critical to your go-to-market as a provider
对你作为 provider 的 go-to-market 来说会变得至关重要
22:59
or a model lab or a server tool and more in the future.
或者是一家 model lab，或者是一个 server tool，未来还会有更多。
23:03
And this value to your earlier point
这也正好印证了你之前说的那个点，
23:05
about how many VCs, one of my biggest frustrations
关于有多少 VC，我最大的挫败感之一
23:11
is that venture capitalists, many of them
就是很多风险投资人，他们
23:13
just don't have any operating experience in the field.
在这个领域里根本没有任何实操经验。
23:16
So unlike a traditional investor
所以不像传统投资人，
23:18
who's just maybe come up through the ranks
他们可能只是一路按部就班升上来的，
23:19
as like associate working on financial modeling
比如做 financial modeling 的 associate。
23:23
or maybe hasn't been a real operator in the field
或者说，可能已经十多年没在这个领域里
23:25
for like more than 10 years, which is a big part
做过真正的 operator 了，而这现在占了行业里
23:27
of the industry now, I had just arrived at A16Z
很大的一部分，我当时刚到 A16Z
23:31
like a year after running the platform.
大概是在运营那个平台一年之后。
23:33
And so I knew what the challenges were
所以我很清楚挑战是什么
23:35
of building a real great developer experience
要打造一个真正出色的 developer experience
23:39
and actually being able to create a working piece
而且真的能做出一个能跑起来的
23:42
of software with a model.
用 model 做出来的 software。
23:44
And there were a few, I won name names,
而且有几个，我就不点名了，
23:46
but they were investors who were looking at OpenRouter
但他们是投资人，当时正在关注 OpenRouter
23:50
and felt at the time when I would compare notes
而且在我跟别人交换看法的时候，他们觉得
23:54
with people that it was just a marketplace.
它只是个 marketplace。
23:59
Yeah, just a thin layer, just a proxy, just a wrapper
对，就是薄薄一层，就是个 proxy，就是个 wrapper
24:01
or whatever on other people's APIs.
或者随便什么，套在别人的 APIs 上。
24:03
And I was like, you have no idea how strategic the value
我就说，你们根本不知道这价值有多具战略意义
24:07
that OpenRouter has created by being able to orchestrate
OpenRouter 因为能够 orchestrate 而创造出来的
24:10
even three APIs in production.
甚至有三个 API 已经在 production 里跑了。
24:13
The amount of both engineering work and community design
工程工作和社区设计这两方面投入的工作量，
24:17
that goes into getting that actually live
为了让这些真正上线，
24:19
and running in production at the scale
并在 production 里运行，达到这样的规模，
24:21
the OpenRouter team had started,
也就是 OpenRouter 团队当初起步时的规模，
24:22
just doesn't happen by default.
并不是默认就会发生的。
24:24
And that was one of the things that stood out to me
而这是让我印象特别深刻的其中一点，
24:26
about Alex from the earlier days.
关于早期的 Alex。
24:28
Like he just understood like these,
就像他刚刚搞懂了这些，
24:30
from a systems perspective,
从 systems 的角度来看，
24:31
like how do you get these fly wheels going?
就像，你怎么让这些 fly wheels 转起来？
24:34
Like that stood out to me with OpenC
就像，在 OpenC 这件事上，这一点对我来说特别突出，
24:35
when we were working together on the NFD integration
当我们一起做 NFD integration 的时候，
24:38
and discord, like Alex had a level of community
还有 Discord，就像 Alex 有一种社区层面的
24:40
like systems thinking on how you get these fly wheels going.
像 systems thinking，关于你怎么让这些 fly wheels 转起来。
24:44
The most scientists and machine learning people
最顶尖的科学家和 machine learning 的人
24:47
just don't think, think of it.
就别想，想想它。
24:49
Like we often think in terms of pre-training,
就像我们经常从 pre-training 的角度去想，
24:51
maturing, post-training.
成熟、post-training。
24:52
It's a linear stage.
这是一个线性的阶段。
24:53
It's a linear pipeline.
这是一个线性的 pipeline。
24:54
It's a small loop, yeah.
就是一个小 loop，对。
24:54
Yeah, it wasn't much too much later
对，没过多久，
24:56
that the modern context feedback loop cycle
现代的 context feedback loop 周期
24:58
really got standardized in the industry,
真的在行业里标准化了，
24:59
but at the time, if you remember machine learning was like,
但那时候，如果你还记得，machine learning 就像是，
25:02
it mostly, we did a lot of ML,
它主要是，我们做了很多 ML，
25:04
like when I was in grad school on a laptop.
就像我读研的时候，在笔记本电脑上做的那样。
25:06
So you just like download a data set,
所以你只要下载一个 data set，
25:08
grant some oblations,
跑一些 ablations，
25:10
and you looked at the loss curves
然后你看看 loss curves
25:11
and you're like, great, I made AI.
然后你说，太好了，我做出 AI 了。
25:13
And the idea that you have to like,
而且那种想法，就是你得，怎么说呢，
25:15
deploy those capabilities, collect feedback,
把这些 capabilities 部署出去，收集 feedback，
25:18
trajectories, then like put those into a continuous loop,
trajectories，然后就像是把这些放进一个 continuous loop 里，
25:21
might came much, much, much later.
可能要晚得多、多、多得多才出现。
25:23
And it's very counterintuitive to the sign,
而且这对 sign 来说非常反直觉，
25:24
like the traditionally I mindset.
就像传统的那种 I 思维模式。
25:26
I do remember doing the investment phase for OpenRouter,
我确实记得当时在做 OpenRouter 的投资阶段，
25:32
I just didn't try and re-educate a bunch of other VCs
我只是没想着去重新教育一帮其他 VCs。
25:36
on why it was not just a marketplace.
关于为什么它不只是一个 marketplace。
25:39
I was like, you know what?
我当时就想，你知道吗？
25:40
I'm just gonna invest.
我就直接投了。
25:41
And I'm going to like take the opportunity
然后我想抓住这个机会
25:43
to partner with Alex.
去和 Alex 合作。
25:44
And if not no other VCs get it, that's totally fine.
如果其他 VCs 都没看懂，那也完全没关系。
25:47
Because at the time, it was not obvious,
因为当时，这一点并不明显，
25:50
I think to several of the investors,
我觉得对好几个投资人来说都是如此，
25:51
that like OpenRouter was not more than just a wrapper
就是说，OpenRouter 不过就是个 wrapper，
25:53
around APS, and that infuriated,
包在 APS 外面的，而那把我气坏了，
25:54
I was like, you know, I don't have time to debate you.
我当时就想，你知道吧，我没时间跟你争。
25:57
I'm just, we're gonna invest.
我就说，行了，我们要投。
25:59
And then I think like a month later,
然后大概过了一个月吧，
26:01
that Murphy marked it up by 10X.
Murphy 就把它 mark up 了 10X。
26:03
Like I think, I forget what the exact post money was
我当时想，我忘了确切的 post money 到底是多少了
26:06
and so on, but, you know, to his credit,
等等之类的，但你知道，说句公道话，
26:08
Manlo Ventures realized, okay,
Manlo Ventures 意识到，好吧，
26:10
there's actually much more strategic value here as well.
其实这里还有更多的战略价值。
26:12
Maybe you didn't hear all these conversations
也许你没听到所有这些对话
26:13
behind the scenes, but that frustrated me a lot.
幕后的情况，但那让我很恼火。
26:17
You know, there's a lot of this like opining about wrappers.
你知道，有很多这种关于 wrappers 的瞎议论。
26:20
And if you're like, oh, an app is just a wrapper on no model,
如果你会说，哦，一个 app 就只是套在 no model 上的 wrapper，
26:23
then like, and OpenRouter is like this wrapper
然后就像，而 OpenRouter 就像这种 wrapper
26:27
on top of other APS, and this is the most stupid,
叠在其他 APS 上面，而这简直是最蠢的，
26:30
reductive framework.
reductive framework。
26:31
So it's clearly somebody who has no experience
所以这明显是个没有经验的
26:33
deploying products.
deploy 产品的人。
26:34
It's the thing you dismiss other things with.
这就是你用来否定其他东西的那套说辞。
26:36
Everyone's a wrapper on everything, right?
每个人都是所有东西上的一层 wrapper，对吧？
26:38
Like if there's this some point,
就像，如果真有这么个点，
26:39
some wrappers are value.
有些 wrapper 本身就是价值。
26:40
I mean, investors are wrappers in LBs, right?
我是说，投资者就是 LBs 里的 wrapper，对吧？
26:42
Like Mexican rappers.
就像墨西哥说唱歌手那样。
26:43
So, I mean, yeah, it's all wrappers done
所以，我的意思是，对，全都是 wrappers 做出来的，
26:45
all down to bare metal, I guess.
一路做到 bare metal，我猜。
26:46
And that's okay.
这也没关系。
26:47
When I started the whole e-engineer,
当我开始搞整个 e-engineer 的时候，
26:50
I guess the coining in 333,
我猜是在 333 里造出这个词，
26:52
like that was the number one pushback,
那最大的反对意见就是，
26:54
is that this is no value.
这东西没有价值。
26:55
You should actually just trim models, right?
其实你就应该直接 trim models，对吧？
26:57
And yeah, I mean, obviously this is like,
而且，对，我是说，显然这就像是，
26:59
you guys are one of the testaments to the fact that
你们就是活生生的例证之一，证明了
27:01
that you can actually build very valuable wrappers,
你确实可以做出非常有价值的 wrappers，
27:03
but also very valuable model companies.
但也能做出非常有价值的 model 公司。
27:06
It's so hard to be like the day a model launches,
真的很难做到像 model 发布当天那样，
27:12
the fact that you have an OpenRouter endpoint
你就有那个 model 的 OpenRouter endpoint，
27:15
for that model frequently at the top of Hacker News
而且经常出现在 Hacker News 顶部
27:18
on day one, people don't realize the amount of work
第一天，人们根本意识不到这需要多少工作量
27:22
that goes into accomplishing that,
才能把这件事做成，
27:23
an OpenRouter used, like that would happen over and over
OpenRouter 一被用起来，就像那种情况会一遍又一遍
27:26
again, and I remember going,
发生，我记得我当时就想，
27:28
people have no idea how hard that is.
大家根本不知道那有多难。
27:30
That's not...
那不是……
27:31
Yeah, we've covered some of the inference engineering
对，我们已经聊过一些 inference engineering
27:32
that goes behind some of the,
那些在某些……背后的
27:35
we're based on in all those.
我们那时候都是基于这些的。
27:36
Well, today you have all those like cool code name things
今天你有那些很酷的代号之类的东西，
27:39
that people guess what oxy alpha is and all those things,
大家会猜 oxy alpha 是什么以及所有这些，
27:42
but I guess one of the things that you're teasing
但我猜你在暗示的一件事之一
27:44
is how do you get that initial flywheel going, right?
就是你怎么让那个初始 flywheel 转起来，对吧？
27:47
Because today you have your scale
因为今天你有你的 scale
27:49
and your reputation and all these things,
和你的声誉以及所有这些，
27:50
so obviously you can't be driving immense distribution,
所以显然你没法驱动巨大的 distribution，
27:53
but when you're early on, when it's...
但当你还在早期的时候，当它...
27:55
The boot strap.
那个 boot strap。
27:56
Yeah, boot strap, to bring it back to early Discord days,
对，boot strap，说回早期 Discord 的日子，
28:01
I think we like initially connected with,
我觉得我们最初联系上，好像，
28:04
this is an OpenC story, technically,
严格来说，这是一个 OpenC 的故事，
28:06
but we initially connected when you were at Discord
但我们最初建立联系，是在你还在 Discord 的时候
28:10
and we talked about like Axi and Finity Sturver.
我们还聊到了像 Axi 和 Finity Sturver 这样的东西。
28:13
Oh, yes, yes.
哦，对，对。
28:14
This server was like the biggest server at the time,
这个 server 在当时算是最大的 server 了，
28:17
at Discord.
在 Discord 上。
28:18
That's right.
没错。
28:19
And you were kind of like constantly bumping up the window.
而且你有点像是一直在把 window 往上调。
28:22
The limits on the server, oh my God,
那个 server 的限制，我的天，
28:24
for no signal, like 10% of Philippines was...
No Signal 的话，像 10% 的 Philippines 都...
28:28
Axi was on that server.
Axi 就在那个 server 上。
28:29
That's it.
就这样。
28:30
It was like a meaningful controller to the GDP of the country.
它有点像是这个国家 GDP 的一个挺重要的控制因素。
28:33
And that's the crypto game.
而这就是 crypto 游戏。
28:34
But it was like a good one.
但它算是挺不错的一款。
28:35
It was good on breeding thing.
它在繁殖这块做得不错。
28:36
Yeah, similar, yeah.
对，差不多，对。
28:37
There was battling, there was breeding,
有对战，有繁殖，
28:40
and then there was like a marketplace for trading.
然后还有一个像交易市场一样的东西。
28:43
To earn as well.
还能赚钱。
28:44
Yeah, to earn.
对，就是为了赚钱。
28:46
And like the graphics were really cute and fun,
而且画面真的很可爱、很有趣，
28:49
and you kind of like, you know,
然后你会有点，呃，你懂的，
28:51
you get kind of emotional about your Axi that you make.
你会对自己做出来的 Axi 有点感情。
28:54
So to like start a community like that,
所以，要像那样去启动一个社区，
28:57
which we had to do many times at OpenSea
而这种事儿在 OpenSea 我们得做过很多次，
29:00
with basically every early project
基本上每一个早期项目都是这样，
29:03
for us to create a marketplace for it,
我们要为它创建一个 marketplace，
29:05
we need to make sure that like the community actually wants it.
我们需要确保社区真的想要它。
29:09
And it's kind of like building something that people want
而这有点像去做人们想要的东西，
29:12
and going and telling them about it.
然后去告诉大家这件事。
29:14
Like you can do that on a one-on-one basis.
你可以一对一地做这件事。
29:16
But as a way higher leverage to do that in a community,
但在社区里做这件事，杠杆会高得多，
29:20
where everyone can talk to you at the same time.
因为大家可以同时跟你交流。
29:22
So we spent a lot of time like building things
所以我们花了很多时间去做东西，
29:26
that the community really wanted.
做社区真正想要的那些东西。
29:27
We did the same thing for OpenRouter.
我们为 OpenRouter 也做了同样的事。
29:29
And you know, like the Axi community
而且你知道，像 Axi 社区
29:31
was one of like a zillion communities we did that with.
就是我们做过这件事的无数社区之一。
29:35
And on just like saw us doing it.
而且在那上面，他们就像看到我们在做这件事一样。
29:37
And because you could just see people
而且因为你能直接看到人们
29:39
sharing OpenSea links constantly in that discord.
在那个 Discord 里不停地分享 OpenSea 链接。
29:42
Like user sharing links is really clear indicator
就像用户分享链接是一个非常清楚的信号
29:45
that like something important is going on.
说明有重要的事情正在发生。
29:48
So we spent, you know, a lot of time like
所以我们，你知道，花了很多时间，就是
29:52
first figuring out what the gap is in the technology
先搞清楚技术里的 gap 到底是什么
29:56
that people care about.
是人们真正在意的。
29:58
Like what was the actual problem that needs to be solved?
就是，真正需要解决的问题到底是什么？
30:00
You know, in early LLM days,
你知道，在早期 LLM 的时候，
30:03
it was, you know, OpenAI refusing to finish the prompt
就是，你知道，OpenAI 不愿意把 prompt 写完
30:09
or like to like complete the task.
或者，就是，去完成 task。
30:12
It was also, you know, inability to customize models.
还有，你知道，就是没法 customize models。
30:16
And so there are communities that like
所以有些社群，就是那种
30:19
are just completely blocked on that issue.
完全卡在那个问题上的。
30:22
And those are the communities that are most useful
而那些社群其实最有用，
30:24
to sort of learn about and diving to and explore.
可以去了解、深入进去、还有探索。
30:28
Something that really struck me at that time,
当时有件事真的让我印象很深，
30:30
as I was just hearing your talk,
就在我刚听你的演讲时，
30:31
I remember noting how you may not remember this.
我记得我注意到，你可能都不记得这件事了。
30:35
But we were, we had these like working Zoom calls
但我们当时，我们会开那种有点像工作讨论的 Zoom 通话。
30:38
that we were doing a sprint around
我们当时在围绕
30:40
for like this OpenSea integration with discord.
这个 OpenSea 跟 discord 的 integration 做一个 sprint。
30:44
And you know, we'd get, it was myself,
而且你知道，我们当时会聚在一起，就是我、
30:47
my engineering team, I think you were there.
我的工程团队，我记得你也在。
30:49
And I remember, you know, Alex in the middle
然后我记得，你知道，Alex 在其中一次
30:52
of one of those calls, just like, there was like silence.
电话会中间，就，突然像是一阵沉默。
30:57
You know, we were, we were all like,
你知道，我们当时，我们都心想，
30:58
oh yeah, this totally makes sense.
哦对，这完全说得通。
30:59
Let's do this and then there's some everybody aligned.
咱们就这么做，然后大家就都 aligned 了。
31:02
And Alex was like, no, this makes no sense to me.
然后 Alex 就说，不行，这对我来说完全说不通。
31:05
And everyone's like, I remember going,
然后大家都说，我记得我当时就想，
31:08
what, like it works.
什么，它居然能用。
31:09
Like you click on a link and then it bounces you out
就像你点一个 link，然后它就把你弹出去
31:13
to like OpenSea and he was like,
到 OpenSea 这种地方，然后他就说，
31:15
it's not a good user experience.
这不是一个好的 user experience。
31:16
Yeah, we should not do this.
对，我们不应该这么做。
31:18
And I remember going, you know, he was the only person
我记得我当时说，你知道，他是唯一一个
31:22
out of all of us to actually raise his hand and go,
在我们所有人当中真正举手说，
31:25
yes, it made sense from a technical implementation perspective
对，从 technical implementation 的角度来看，这是说得通的
31:28
like we were bouncing the user out into OpenSea.
就像我们当时是把用户跳转到 OpenSea。
31:31
And so it kind of checked the box
所以这算是满足了
31:33
of the product manager's requirements on both sides.
两边 product manager 的要求。
31:36
But Alex went one step further and was like,
但 Alex 更进一步，他说，
31:40
you know what would be better guys,
你们知道吗，更好的做法是，
31:41
if we just embedded the experience right here
如果我们直接把这个体验 embedded 在这儿
31:42
inside of Discord.
就放在 Discord 里面。
31:43
So the link opened up as an embedded iFrame
所以这个链接就以 embedded iFrame 的形式打开了
31:45
and you can just check out right there.
然后你直接在那儿就能查看。
31:47
And not one person on the corner, like seven of us
而且角落里不止一个人，像我们七个
31:50
who had met like, you know, we got to read it.
我们之前见过面，然后，你懂的，我们有机会读到它。
31:52
And it's the guy who doesn't work for Discord.
而且就是那个不在 Discord 工作的人。
31:53
And it's the guy who doesn't work for Discord.
而且就是那个不在 Discord 工作的人。
31:55
Like technically you benefit if they bounce.
就像，严格来说，如果他们 bounce，你反而会受益。
31:57
Exactly.
没错。
31:58
And that was like adversarial to keep the user
而那就像是 adversarial——把用户留在
32:01
inside of Discord would be adversarial to OpenSea.
Discord 里，对 OpenSea 来说会是 adversarial。
32:03
And yet Alex put that user experience first.
然而 Alex 却把 user experience 放在第一位。
32:06
And I was like, that's special.
而我就想，这很特别。
32:08
Wow.
哇。
32:09
Because it's very hard to have somebody who's technical
因为要有一个懂技术的人，真的很难
32:10
like Alex and understands the developer flow.
像 Alex 那样，也懂 developer flow。
32:13
But also understands the best user experience
但也理解最好的 user experience
32:14
and wants to prioritize it.
并且想优先考虑它。
32:15
And that's two sides of the fly
而这就是 fly 的两面
32:17
we let you can get spinning.
我们让你能把它转起来。
32:18
Like it's often hard to stop.
就像它经常很难停下来。
32:20
And you just reminded me that that one was one of those moments
你刚刚提醒了我，那一次就是那种时刻之一
32:23
where I go, I really, I got to be better
我会想，我真的，我得变得更好
32:25
at user experience because I should have been the one
在 user experience 上，因为我本该是那个
32:27
who came up with that and I didn't.
想出那个点子的人，但我没有。
32:29
And I learned from you.
而且我是从你身上学到的。
32:29
And I think that went into one of our case studies
而且我觉得，这被写进了我们的一个 case study 里
32:33
for the PM training program at Discord.
用在 Discord 的 PM training program 上。
32:35
I won't be there.
我不会到场。
32:36
I don't know if it's there as they want.
我不知道它是不是他们想要的样子。
32:38
You need an Alex, this is a conclusion.
你需要一个 Alex，这就是结论。
32:40
Yeah, yeah, you need an Alex.
是啊，是啊，你需要一个 Alex。
32:42
And this is why nobody should be surprised
这就是为什么没人该感到意外，
32:44
why Stripe decided like they had to buy OpenRouter
为什么 Stripe 决定他们必须收购 OpenRouter，
32:47
because it's a really rare combination of people
因为这是一种非常罕见的人才组合，
32:50
who understand the machine learning community,
他们懂 machine learning 社区、
32:53
the developer experience and the end user experience.
developer experience 和 end user experience。
32:55
And putting all that together as a result
而把这一切整合在一起，结果就是
32:58
in this extraordinary scale that very few other
达到这种非凡的规模，很少有其他
33:00
marketplaces have been able to achieve
marketplaces 能够实现
33:02
over the last five years.
在过去五年里。
33:04
Yeah, well, we should talk about the other reasons
对，嗯，我们应该聊聊其他原因
33:06
or the other reasons, which you've written about.
或者说你写过的那些其他原因。
33:08
I wanted to sort of proceed somewhat chronologically as well.
我也想大致按时间顺序来讲。
33:11
So there is a point that one of the questions
所以有一点，其中一个问题
33:14
that Dave from HF0 sent in was,
是来自 HF0 的 Dave 发来的，
33:17
when did you know it really started to work?
你是什么时候知道它真正开始奏效的？
33:19
And you brought up a mixture.
而且你提到了 mixture。
33:21
I don't know if you want to bring up that sort of thing.
我不知道你想不想提那种事。
33:22
Oh, yeah.
哦，对。
33:23
Which obviously you overlap with, so.
而这显然跟你有重叠，所以。
33:26
Yeah, the MOE was, I don't know when,
是啊，MOE 那回事，我也不知道是什么时候，
33:29
I mean, there's no like one moment where I was like,
我是说，并没有某一个瞬间让我觉得，
33:33
oh, this is officially starting to work.
哦，这正式算是开始起效了。
33:36
It was like moments of increasing connection.
更像是连接不断增强的一个个瞬间。
33:38
They really super early on.
他们真的超级早就入场了。
33:39
Oh, yeah.
哦，对。
33:40
So before OpenRouter, I wanted to like explore
所以在 OpenRouter 之前，我就想探索一下
33:44
a bring-your-own-model experiment.
一个 bring-your-own-model 的实验。
33:46
And anyone who's familiar with crypto
而且任何熟悉 crypto 的人
33:48
is like, yeah, a fan to them in all these things.
都会说，是啊，在这些事情上都是他们的粉丝。
33:50
Yeah, so it felt like doing a meta-mask analogy
对，所以感觉像是做一个 meta-mask 的类比
33:54
for AI would be kind of a fun way of exploring that.
对 AI 来说，会是一种挺有趣的探索方式。
33:59
And at the time, there were no AI apps.
当时还没有 AI apps。
34:02
There were probably as many AI apps
那时候，可能有多少 AI apps，
34:04
that were like hitting an LLM via an API call
那种通过 API call 调用 LLM 的，
34:10
as there were like games just doing it in JavaScript.
就有多少用 JavaScript 就能做出来的游戏。
34:15
Basically, there was a moment in time
基本上，曾经有那么一个时刻，
34:18
where it could have been the case that web apps
本来可能是 web apps
34:22
call LLM through the browser,
通过 browser 调用 LLM，
34:25
like through some kind of desktop-managed app
就像通过某种 desktop-managed app 那样。
34:29
that is controlled by the user.
那是由用户控制的。
34:31
And of course, there are like, I think many reasons
当然，我觉得，有很多原因
34:33
that that did not happen.
导致那件事没有发生。
34:35
But back when the days were that primordial,
但回到那个还很原始的时候，
34:39
I built a Chrome extension called Window AI.
我开发了一个叫 Window AI 的 Chrome extension。
34:43
And with Plasma, which I had come across early on
还有 Plasma，我很早就接触到了
34:46
and I was like, who's going to actually use this?
我当时就想，谁会真的用这个啊？
34:48
You did.
你用了。
34:49
Plasma had a couple, like, I think Phantom was using it.
Plasma 有几个，呃，我觉得 Phantom 在用。
34:54
There were some other like real companies.
还有一些其他，像是，真正的公司。
34:56
There was like a Shem.
有一个叫 Shem 的。
34:57
So it basically reacts for Chrome extension.
所以它基本上就是给 Chrome extension 用的 React。
34:58
It compiles to all these.
它会编译成所有这些。
35:00
Kind of like Next.js for Chrome extension.
有点像 Chrome extension 版的 Next.js。
35:02
Next.js, OK.
Next.js，好。
35:03
And yeah, built Window AI on top of it,
然后，对，就在它上面构建了 Window AI，
35:06
the creator of Plasma started contributing code
Plasma 的创建者开始贡献代码
35:09
to Window AI in GitHub.
在 GitHub 上给 Window AI。
35:12
And that turned out to be Lewis Vichy.
结果发现那人就是 Lewis Vichy。
35:15
Oh, you're the co-founder of OpenRouter.
哦，你是 OpenRouter 的联合创始人。
35:17
That's right.
没错。
35:18
You have told me how you met Lewis.
你跟我说过你是怎么认识 Lewis 的。
35:19
Yes, OK.
对，OK。
35:20
So that allowed users to kind of like configure
所以这让用户可以有点像去配置
35:24
which model they wanted to use for a web page in their browser.
他们想在自己的 browser 里，为一个 web page 用哪个 model。
35:27
And then the app would just call out
然后 app 就会直接调用
35:29
to that model when it needed to do things.
那个 model，在它需要做事的时候。
35:32
Not the right form factor for LLMs.
对 LLMs 来说，这不是合适的 form factor。
35:34
But it's like fun experiment.
但这就像个有趣的实验。
35:36
You learn a lot.
你会学到很多。
35:37
And I open-sourced it.
然后我把它 open-source 了。
35:40
And the main learning is like, OK, this has to be an API.
而主要的收获就是，OK，这必须得是一个 API。
35:45
And it has to look a little bit like there
而且它得看起来有点像是，
35:48
has to be more of a developer experience here
这里得更像是一种 developer experience，
35:50
and more of a discovery experience as well.
而且也要更像是一种 discovery experience。
35:53
I don't know where to use these models.
我不知道该在哪儿用这些模型。
35:55
And a little Chrome extension is not going to help me discover.
而且一个小小的 Chrome extension 根本帮不了我去发现。
35:58
It's not enough real estate.
这点地方不够。
35:59
I need more space.
我需要更多空间。
36:00
I need visuals.
我需要视觉呈现。
36:01
I need graphs.
我需要图表。
36:02
I need examples.
我需要例子。
36:04
I need images.
我需要图片。
36:06
I need to be able to explore both as a human and as an agent.
我需要既能作为人类、也能作为 agent 去探索。
36:10
So that's kind of how OpenRouter came to be.
所以 OpenRouter 差不多就是这么来的。
36:13
You know, a meta point that I think is underappreciated.
你知道，有一个 meta point，我觉得它被低估了。
36:17
But Alex is reminding me, is that we were quite lucky
但 Alex 在提醒我，我们当时真的很幸运
36:20
that we were so adjacent to the crypto community
能和 crypto 社区离得那么近。
36:24
in those days because in hindsight, crypto
当年，因为事后回看，crypto
36:27
ended up being kind of like a dress rehearsal for generative models.
最终有点像 generative models 的一次彩排。
36:31
If you think about the Axi experience,
如果你想想 Axi 那段经历，
36:36
Alex is totally right.
Alex 说得完全对。
36:37
They were not that many AI apps at the time.
当时并没有那么多 AI apps。
36:39
And while I was dealing, my job was
而在我做这些的时候，我的工作是
36:41
to be the head of platform at Discord,
在 Discord 当 head of platform，
36:43
which meant to be a general purpose place for communities
也就是要把它做成一个面向社区的通用场所。
36:45
and friends to create, for developers,
以及朋友们来创造，为开发者们，
36:48
to create apps and bots and other services
去创建 apps 和 bots 以及其他服务
36:52
that could be deployed across Discord.
这些可以 deploy 到 Discord 的各个地方。
36:54
And while 80% of the attention of the time
而虽然当时 80% 的注意力
36:56
was being spent on crypto,
都花在了 crypto 上，
36:57
because that's where all the NFT volume was,
因为所有的 NFT 交易量都在那里，
36:59
there was like 20% of my time I was spending
我大概有 20% 的时间是花在
37:02
with a friend who would get hotpot with me
和一个会跟我一起吃火锅的朋友待在一起
37:05
and ask me for, we'd play Magic the Gathering on weekends.
还会来问我，我们周末就打 Magic the Gathering。
37:08
And he was working on a little Discord bot
他当时正在做一个小 Discord bot，
37:10
that could take a text input and turn it into an image.
它能把 text input 变成图像。
37:13
And it was called mid-journey.
它叫 mid-journey。
37:14
You know, there was David Holt,
你知道，当时有 David Holt，
37:16
so that was a good friend.
他是我一个好朋友。
37:17
And David and I both then sort of failed.
然后 David 和我后来都算是失败了。
37:18
They are VR founders, you know, in the last before that.
他们是 VR 创始人，你知道，就在那之前的上一波。
37:22
And I remember this, you know, mid-journey
而且我记得，你知道吗，Midjourney
37:26
was one of the fastest growing communities
是我们当时增长最快的社区之一
37:28
we had after-axe infinities started to beater off.
那是在 Axie Infinity 开始逐渐退潮之后
37:31
And many of the, like the abstractions
而且很多，比如那些 abstractions
37:35
and the infrastructure decisions we made to scale-axe
以及我们为了 scale Axie 所做的 infrastructure 决策
37:37
happened just in time,
都来得刚刚好
37:39
because they inaxe did this and then fell off a cliff.
因为后来 Axie 这么做了，然后就断崖式下跌了
37:41
And then as mid-journey was taking off,
然后，随着 Midjourney 开始起飞，
37:43
we like explicitly decided to help David make the server,
我们其实明确决定帮 David 做这个 server，
37:46
the mid-journey server is the primary place
mid-journey 的 server 是主要的地方
37:49
for interaction with the model
用来和 model 互动
37:50
because it was very hard for people
因为对大家来说真的很难
37:53
to understand how to use the model
搞懂该怎么用这个 model
37:55
if they couldn't see other people using it and copy them.
如果他们看不到别人怎么用、然后模仿他们的话。
37:58
And so the single-player mid-journey web app
所以那个 single-player 的 mid-journey web app
38:00
on its own, like mid-journey.com,
单独来看，就像 mid-journey.com，
38:02
had like terrible retention.
retention 特别差。
38:04
Because people would show up,
因为人们会来，
38:05
they'd see this empty field.
就会看到这个空白的输入框。
38:06
It's kind of like Dolly too.
这也有点像 Dolly。
38:08
And they would type in like cat or dog.
然后他们会输入像猫或狗这样的词。
38:11
And it was like paralyzing for them to have this blank canvas
而面对这块空白画布，他们简直不知所措，
38:13
that they had to fill,
还得自己把它填满，
38:14
because they'd never used an AI model before.
因为他们以前从来没用过 AI model。
38:16
But instead in a Discord server,
但相反，在一个 Discord server 里，
38:17
you could see other people using it
你能看到其他人在用它，
38:18
and riff off of their prompts.
然后顺着他们的 prompts 即兴发挥。
38:19
And the engagement was off the charts.
而且 engagement 简直爆表。
38:21
And so scaling mid-journey from zero
所以把 Midjourney 从零开始 scaling
38:23
to like 10 million monthly actors
到差不多 1000 万 monthly actors
38:25
was a much smoother approach, both to Axi and Finity.
对 Axi 和 Finity 来说，都是顺畅得多的做法。
38:28
And so, like,
所以，就，
38:29
don't forget the best of four pictures, any of these.
别忘了四张图里最好的那张，这些里面随便哪张都行。
38:31
The best, yeah.
最好的那个，对。
38:32
Which is the feedback loop.
也就是那个 feedback loop。
38:33
The Arla HR feedback loop,
Arla HR 的 feedback loop，
38:35
which by the way, separately,
顺便说一句，另外，
38:36
Ekton Brown, David and I used to play Magic Gathering
Ekton Brown、David 和我以前周末会玩 Magic Gathering。
38:39
on weekends.
那是一群会一起出去玩的朋友。
38:40
It was one group of friends who would hang out
那是一群经常一起玩的朋友
38:42
and these concepts were all being discussed all the time.
而且这些概念一直都在被讨论。
38:45
But I think there were few of us
但我觉得，像我们这样的人很少
38:49
who bridged both the crypto worlds and the AI worlds
能同时把 crypto 世界和 AI 世界连接起来
38:52
and compared to crypto where the question was always,
而跟 crypto 比起来，那边的问题永远是：
38:54
what's the use case for this technology?
这项技术的 use case 是什么？
38:57
There was never any need to ask that for you.
但对你来说，从来都不需要问这个。
38:59
Because the use case was so visceral.
因为它的 use case 太切身了。
39:01
I can create now anything I can imagine.
我现在能创造任何我能想象到的东西。
39:05
I can write novels, I can code.
我能写小说，也能写代码。
39:07
And the infrastructure that those of us
而那些基础设施，我们这些
39:09
who believed in the distributed systems,
相信 distributed systems 的人，
39:12
like value of crypto,
比如 crypto 的价值，
39:13
like the censorship resistance part,
比如 censorship resistance 这部分，
39:14
found this use case that was explosive.
找到了这个爆发式的 use case。
39:16
And I think between mid-journey,
而且我觉得，在 mid-journey 之间，
39:19
the plot was a Discord bot pre-launch
整个剧情就是一次 Discord bot pre-launch
39:22
that we were using internally as an LEM.
我们当时在内部把它当作 LEM 用。
39:24
11 labs had a TTS model that we had on Discord as well.
11 labs 有一个 TTS 模型，我们同样也放在 Discord 上用了。
39:28
Discord became this P3 dish for early apps to innovate.
Discord 成了早期 app 用来创新的 P3 dish。
39:32
And I don't think it's a coincidence
而且我不觉得这是巧合：
39:33
that they found a home there before OpenRouter
它们在那里找到了家，那时 OpenRouter
39:36
gave the world a public home store,
还没给世界一个公共的 home store，
39:39
or a storefront.
或者说一个 storefront。
39:42
Discord was this almost kind of P3 dish storefront
Discord 几乎就像是这种 P3 dish 式的 storefront
39:46
that was kind of piggybacked on the infrared
那有点算是搭在 infrared 上，
39:49
we'd built for crypto communities.
那个 infrared 是我们为 crypto 社区搭建的。
39:51
And then I think Alex was one of the first people
然后我觉得 Alex 是最早的一批人之一，
39:52
to realize, wait a minute,
意识到，等一下，
39:54
like these apps need their own home on the internet.
这些 apps 需要在 internet 上有自己的家。
39:58
And then OpenRouter, to me, was a continuation of that,
然后 OpenRouter，对我来说，是那件事的延续，
40:01
of that community's needs.
是那个社区需求的延续。
40:02
And of course, there was the crazy distribution
当然，还有那疯狂的 distribution
40:05
that you enabled for a lot of these developers.
你为很多这类开发者提供了支持。
40:07
So then my question is,
那我的问题是，
40:08
how come you were,
为什么你会，
40:09
my recession is OpenRouter is not that Discord centric, right?
我的印象是，OpenRouter 并没有那么以 Discord 为中心，对吧？
40:12
You have a Discord.
你们有 Discord。
40:13
Yeah.
对。
40:14
It's not to engage your community,
它不是为了跟你们的社区互动，
40:16
but it's not like mid-journey where like,
但它不像 mid-journey 那样，就是……
40:18
no, that is like the primary way
不，那基本上就是最主要的方式
40:19
of people experience OpenRouter.
人们体验 OpenRouter 的方式。
40:21
Yeah, mid-journey, like,
对，mid-journey，就，
40:22
it really helps us see visually, really quickly
它真的能帮我们很快、很直观地看到
40:26
how people are using the model and how to prompt it.
人们是怎么用这个 model 的，以及该怎么 prompt 它。
40:29
And I think that is partly why the server was so critical.
而且我觉得，这也是这个 server 之所以这么关键的部分原因。
40:33
It's like, it is the user experience.
就像，它就是用户体验。
40:35
It actually adds a ton.
它其实加了很多东西。
40:36
Yes.
对。
40:38
And you can go the whole mile
而且你完全可以一路走到底
40:39
with just like prompting via mid-journey,
就只是像在 mid-journey 里做 prompting 这样，
40:42
like via the mid-journey Discord server,
比如通过 mid-journey 的 Discord server，
40:45
getting your images,
拿到你的图片，
40:46
and then sharing them and having fun.
然后把它们分享出去，玩得开心。
40:48
For OpenRouter, for LLMs,
对于 OpenRouter，对于 LLMs，
40:50
like, you need a lot of user experience around LLMs,
就是，你需要很多围绕 LLMs 的 user experience，
40:53
make them like really use the tools.
让他们真的去用这些 tools。
40:54
Yes, exactly.
对，就是这样。
40:55
And yeah,
然后，嗯，
40:57
like seeing the examples of other people
比如说看别人的例子
40:59
is also not as useful,
其实也没那么有用，
41:00
because it's a lot of stuff to read.
因为要读的东西太多了。
41:01
It takes a long, long time.
要花很长很长时间。
41:03
Yeah.
是啊。
41:04
You need like code-based integration,
你需要那种 code-based integration，
41:06
not possible to do in a Discord server.
在 Discord server 里做不到。
41:08
You need, or technically, it's possible.
你需要，或者说，严格来讲，其实也能做到。
41:11
I shouldn't say that.
我不该这么说。
41:12
It's just not a great, great developer experience.
只是它真的不是一个特别特别好的 developer experience。
41:15
You need governance
你需要 governance
41:17
for, at the point where you've got code-based integration,
因为，到了你有 code-based integration 的时候，
41:20
now you need governance
现在你就需要 governance
41:21
for managing the LLMs that have access to it,
用来管理那些能访问它的 LLMs，
41:23
the data policies, which teams,
data policies、哪些团队，
41:25
all that stuff needs a lot more
所有这些东西需要的，远不止
41:27
than a Discord server can provide.
一个 Discord server 能提供的。
41:29
So it's just like it's not the right.
所以这就好像，它不是合适的那个。
41:32
Well, in addition, you're not wrong,
嗯，另外，你也没说错，
41:33
but also there's the very important distinction
但另外还有一个非常重要的区别，
41:36
that mid-journey was an end-user application.
那就是 mid-journey 是一个 end-user application。
41:40
And that's why Discord has 250-minute monthly
这也是为什么 Discord 有每月 250 分钟使用时长的
41:45
end-consumers, it made sense for Discord
终端消费者，所以 Discord 成为
41:48
to be a host for that application experience.
这种应用体验的宿主是合理的。
41:54
What I knew was going to happen
我当时就知道接下来会发生什么，
41:56
soon after mid-journey found explosive product market fit,
就在 mid-journey 找到爆发式的 product market fit 后不久，
41:59
because I think when mid-journey launched,
因为我觉得 mid-journey 刚推出时，
42:01
from launch to 100 million revenue run rate
从推出到 1 亿 revenue run rate
42:04
it was less than eight months.
用了不到八个月。
42:06
And shortly thereafter, stable diffusion launched.
紧接着没多久，stable diffusion 就发布了。
42:09
And all of us used to hang out in the Discord server.
我们所有人以前都泡在那个 Discord 服务器里。
42:11
I think it was the...
我记得是那个……
42:13
Distability Discord?
Distability Discord？
42:14
It was the Lyon...
是 Lyon……
42:16
Yeah, Lyon, the Lyon, the image community,
对，Lyon，Lyon，那个图像社区，
42:18
that's far more stable diffusion.
那里 stable diffusion 要多得多。
42:20
And so when stable diffusion came out, I realized,
所以当 stable diffusion 出来的时候，我意识到，
42:23
oh, now other people can build their own mid-journey.
哦，现在其他人也可以自己搭建mid-journey了。
42:27
Because until then, mid-journey did not have an API.
因为在那之前，mid-journey还没有API。
42:30
So they were a full-stack company, right?
所以他们当时是一家full-stack公司，对吧？
42:31
They were training their own models
他们在训练自己的models，
42:32
and they were deploying them as an application,
然后把它作为application部署出来，
42:34
but if you want to build your own mid-journey
但如果你想要搭建自己的mid-journey，
42:36
there was no API of that quality.
当时并没有那种质量的API。
42:39
And I think Dolly too was still quite primitive,
而且我觉得Dolly当时也还相当primitive，
42:41
like mid-journey actually had great quality.
像 mid-journey，其实质量已经非常好了。
42:42
And then when stable diffusion came out,
然后等 stable diffusion 出来之后，
42:44
suddenly there was this new person
突然出现了这么一种新角色，
42:46
who could...
这个角色可以……
42:47
There was this new capability in the world,
世界上出现了这样一种新能力，
42:48
which is a developer could create their own mid-journey.
也就是开发者可以创建属于自己的 mid-journey。
42:51
And that I think created the need
而我觉得，这就催生了对
42:53
for something like OpenRouter.
类似 OpenRouter 这样的东西的需求。
42:54
Because then you need an API to...
因为那样的话，你就需要一个 API 来……
42:55
If you had the kind of creativity of David Holes
如果你有 David Holes 那种创造力
42:58
and you had stable diffusion as the model
而且你有 stable diffusion 作为 model
43:01
and you wanted to put these things together,
然后你想把这些东西组合到一起，
43:02
how could you do that
你怎么才能做到
43:04
without having to figure out how to host the weights?
而不用去搞清楚怎么 host weights？
43:07
And what OpenRouter...
那 OpenRouter 呢……
43:08
The shape of OpenRouter enabled is that, right?
OpenRouter 所开启的形态就是这个，对吧？
43:12
When you have OpenModel's alternatives
当你有 OpenModel 的替代方案
43:14
to closed sort of applications,
去替代那种封闭式应用时，
43:16
OpenRouter's value in the world becomes extraordinary
OpenRouter 在这个世界上的价值就变得非同寻常
43:18
because I need developer can just show up
因为我需要开发者能直接出现
43:20
and use the API of the university.
并使用大学的 API。
43:21
Did you just save the shape of OpenRouter?
你刚才是不是把 OpenRouter 的 shape 保存下来了？
43:23
Oh, no.
哦，不。
43:24
I didn't know that was the real idea.
我之前不知道那才是真正的想法。
43:26
I've been...
我一直……
43:27
I didn't know that.
我之前不知道。
43:28
I'm misaligned now, I've been over-trained.
我现在 misaligned 了，我被 over-trained 了。
43:31
I've been using CloudRapial much, haven't I?
我一直在用 CloudRapial，用得挺多的，对吧？
43:34
Slotish is what people have said.
Slotish 是大家所说的。
43:35
I've got to show up.
我得去露个面。
43:36
I've got to train myself.
我得训练自己。
43:38
Okay, and I just want to cap off the Mistral side.
好，然后我只想把 Mistral 这边收个尾。
43:40
My TLDR is, there was a Mistral price war,
我的 TLDR 是，出现了一场 Mistral 价格战，
43:43
is what I called it, right?
我是这么叫它的，对吧？
43:44
Like, run about in Europe since 2023 or...
就像，从 2023 年起就在欧洲那边跑起来了，还是……
43:46
Yes.
对。
43:47
December.
十二月。
43:48
They launched Mistral by A by seven B.
他们发布了 Mistral AI 7B。
43:51
And like the price went down like 80%.
然后价格大概降了 80%。
43:54
To me, that's very positive
对我来说，这是非常积极的。
43:55
because it's like the first real competition
因为这就像是第一次真正的竞争
43:58
to host Mistral.
来 host Mistral。
44:00
Is there more?
还有更多吗？
44:01
Yeah, that was...
是啊，那是……
44:03
I'm like trying to remember it.
我正试着回想呢。
44:05
All the things that happened.
所有发生过的事情。
44:06
Like, we saw that model come out
就像，我们看到那个 model 出来
44:09
and immediately saw people say
然后立刻就看到有人说
44:12
that it was the best model in the world.
说它是世界上最好的模型。
44:15
Yes.
是的。
44:16
This was, to my knowledge, the first time
据我所知，这是第一次
44:18
an open weights model was called that
一个 open weights model 被这样称呼
44:21
in real seriousness.
而且是认真的。
44:22
It's hype, right?
这是炒作，对吧？
44:23
Is it?
是吗？
44:24
You know?
你知道吗？
44:25
It was hype, it was hype.
就是炒作，就是炒作。
44:27
It was also hype from AI influencers at the time.
当时也是 AI 网红们在炒作。
44:31
And there were many examples
而且有很多例子
44:33
where it was like outperforming GPT4.
看起来就像是能超过 GPT4。
44:37
So people really wanted to try it out
所以大家真的很想试试看
44:40
and see, oh, it's this can be true for me too.
看看，哦，这事对我来说是不是也能成立。
44:43
And if so, at what price?
如果是这样，那要付出什么代价？
44:45
And the like inference landscape was really messy.
而且当时 inference 这块的格局真的很乱。
44:50
Yes.
是的。
44:50
We cleaned it up.
我们把它整理好了。
44:52
It allowed providers to compete on price.
这让 providers 可以在价格上竞争。
44:55
So we could give users the best price in one spot.
这样我们就能在一个地方给用户最好的价格。
44:58
And so it was, I think the first clear example
所以，我觉得这是第一个清晰的例子，
45:01
of like a provider marketplace working
就是一个 provider marketplace 在运作，
45:05
in a way that adds value to developers.
而且是以能为开发者增加价值的方式。
45:08
Sean, you may not remember this,
Sean，你可能不记得这件事了，
45:09
but I think we met for the first time
但我觉得我们第一次见面
45:12
a few days after McStrawl came out at Nureps,
是在 McStrawl 在 Nureps 发布几天后，
45:14
at a luncheon.
在一个午餐会上。
45:15
Yeah, that's where I also met BFF as well.
对，我也是在那儿认识 BFF 的。
45:17
And Guillaume was there.
Guillaume 也在那儿。
45:19
Yeah, you were there too.
对，你当时也在。
45:20
And we had just announced the Mistral investment.
而且我们当时刚宣布了对 Mistral 的投资。
45:24
And Guillaume was over there.
而 Guillaume 就在那边。
45:26
And I remember turning to Guillaume and asking him,
我记得我转向 Guillaume，问他，
45:30
is it how are you feeling after the launch of McStrawl
就是，McStrawl 发布之后你感觉怎么样
45:33
in 7B?
7B 的？
45:34
And his typical French fashion was like,
而他一贯的法国式风格就是，
45:37
I mean, it's an OK model.
我是说，这是个还行的 model。
45:39
It's not that great.
也没那么好。
45:41
It was so unconscious.
这也太下意识了。
45:42
But I remember him also saying that part of the reason
但我记得他还说，部分原因是
45:46
he felt a lot of people thought that it was better
他觉得很多人认为它更好
45:49
than GPT-4 was because of the speed.
比 GPT-4 好，是因为速度。
45:52
It was an MOE model that they had absolutely
它是个 MOE 模型，他们绝对
45:55
kind of figured out how to make super efficient.
算是搞明白了怎么让它超级高效。
45:57
It was on the period of frontier.
它当时处在 frontier 的那个阶段。
45:59
And this is an important thing about LLM, right?
这是关于 LLM 的一件重要的事，对吧？
46:01
Sometimes when they're faster, you think they're smarter.
有时候它们更快，你就会觉得它们更聪明。
46:04
Even though if you did any of these common evils
即使你做了这些常见的坏事中的任何一件
46:08
that are seven, you do seven tries.
有七次，你就做七次尝试。
46:11
And I don't actually remember.
而且我其实不记得了。
46:13
I think we should go back and figure out
我觉得我们应该回去弄清楚
46:15
what the data says.
data 到底说明了什么。
46:16
But I wouldn't be surprised if it turns out
但我不会惊讶，如果结果发现
46:18
on an end of seven attempts, GPT-4 was smarter on evils.
在七次尝试结束时，GPT-4 在 evils 上更聪明。
46:22
But the perception of on correctness
但对 correctness 的感知
46:26
would be smarter or more accurate.
会更聪明，或者更准确。
46:28
But people from a human preference perspective
但从 human preference 的角度来看，人们
46:32
felt that it was faster because it was smarter
觉得它更快，因为它更聪明
46:35
because it was so fast.
因为它就是这么快。
46:37
And actually, most queries do not take that level.
而且实际上，大多数 query 都不会用到那个级别。
46:39
Don't take that.
不会。
46:40
So this is the start of humans as router,
所以这就是人类作为 router 的开端，
46:42
which then eventually becomes open router as router
然后最终会变成 open router 作为 router
46:45
of the auto model.
来路由 auto model。
46:46
That's interesting, right?
这挺有意思的，对吧？
46:47
I mean, because humans are the routing mechanism.
我是说，因为人类就是 routing mechanism。
46:49
I will ask the fast model first.
我会先问一下 fast model。
46:50
And then if not good enough, I'm going to upgrade manually.
然后如果不够好，我就手动升级。
46:52
Yes, yes.
对，对。
46:53
But then he's going to auto it.
但然后他会把它自动化。
46:54
I didn't auto it that way, but that makes sense.
我没有那样把它自动化，但那样说得通。
46:57
Which then there's a lot more techniques.
那接下来还有很多 techniques。
46:59
Like fusion is the thing that we should talk about.
就像 fusion 才是我们应该聊的那件事。
47:01
Before I move on to those things,
在我接着讲那些事情之前，
47:03
I just want to close off the early years.
我只想把早期那几年收个尾。
47:06
One thing that I observe, which you are also
我观察到的一点是，你也是
47:09
an investor in arena, and we talked about mid-journey
arena 的投资者，而且我们聊过 mid-journey
47:11
having that feedback group of A, B, C, D
有那种 A、B、C、D 的反馈组
47:14
and choosing that being very important.
而做出那个选择非常重要。
47:16
And you understand the flywheel.
而且你理解 flywheel。
47:19
So how come you didn't build arena
那你怎么没去做 arena 呢？
47:21
and how come arena didn't build open router?
那 arena 怎么没去做 open router 呢？
47:23
Well, arena started before open router, right?
嗯，arena 比 open router 先开始的，对吧？
47:28
They had the school project in hell and then
他们当时在 hell 有个学校项目，然后
47:29
it became a completely arena.
它就成了一个彻头彻尾的 arena。
47:32
So I know you had some arena experiences
所以我知道你有过一些 arena 的经历
47:34
like the heads-up comparison type things.
比如那种 heads-up 对比之类的。
47:37
But you never really went as hard as arena did.
但你从来没有像 arena 那样真的那么拼。
47:40
And doing heads-up experience.
而且还要做 heads-up 体验。
47:42
And all of them actually did have a router project
而且他们所有人其实都有一个 router 项目
47:44
based on Elmerina Elos, which they never commercialized.
基于 Elmerina Elos，但他们从没把它商业化。
47:48
It's hard to do a company that does both,
很难做一家两者都做的公司，
47:50
because one company is taking data and selling it
因为一家公司是拿数据去卖，
47:54
and the other company really can't.
而另一家公司真的做不到。
47:57
But by default.
但默认情况下。
47:58
So I think there is a branding reason
所以我觉得这里面有个品牌方面的原因
48:03
that there are two companies here.
就是说，这里有两家公司。
48:07
When you set up open router, there's no training.
当你设置 OpenRouter 的时候，没有任何 training。
48:10
There are no prompts, aside from what your provider
除了你的 provider
48:13
policy sets.
policy 设定的之外，没有 prompts。
48:15
Open router can't see your prompts or completions.
OpenRouter 看不到你的 prompts 或 completions。
48:17
If you want to see that as an org,
如果你想以 org 的身份看到这些，
48:19
you have to opt into it and enable it.
你就得主动 opt in，并把它 enable 起来。
48:22
And so we're pretty conservative and careful
所以我们相当保守，也很谨慎。
48:24
about data policy and security and privacy.
关于 data policy、安全以及隐私。
48:29
And Elmerina is like their business model
而且 Elmerina 有点像，他们的 business model
48:31
is oriented around the labs.
是围绕 labs 展开的。
48:33
And you don't get a fee to give a fee.
而且你不会为了给出一笔费用而拿到一笔费用。
48:37
Yeah.
对。
48:38
We do have free endpoints too, but those free endpoints.
我们确实也有免费的 endpoints，但那些免费的 endpoints。
48:42
I think we're not collecting any prompts.
我觉得我们不会收集任何 prompts。
48:44
We're not monetizing the data unless you opt into it
除非你 opt into it，我们不会把 data 变现。
48:47
for some reason.
不知道为什么。
48:48
This comparison, you're not the first person to ask me this.
这个比较，你不是第一个问我这个的人。
48:50
And Alex knows this, but I was the first CEO of Arena
Alex 知道这一点，但我是 Arena 的第一任 CEO
48:58
for the first five months when we were helping
在最初的五个月里，当时我们在帮忙
48:59
on the status in Wailin kind of spin out of Berkeley.
处理 Wailin 那边的情况，算是从 Berkeley 剥离出来。
49:03
And I did invest in that before open router,
而且我在 open router 之前确实投资了那个，
49:07
but it was very strange to me the comparisons
但对我来说，那些比较非常奇怪
49:11
that outside folks would make between the two projects
外面的人会拿这两个项目做比较
49:13
because the missions were completely different.
因为使命完全不同。
49:15
The founding entity for Arena,
Arena 的创始实体，
49:17
we called it the AI Reliability Institute
我们叫它 AI Reliability Institute
49:19
because it was actually there as an EVAL service.
因为它实际上是以一个 EVAL 服务的形式存在的。
49:22
Like the data, so to speak, that they were originally
可以说，他们最初，就像数据一样，
49:27
kind of offering the labs was how do you make
有点像提供给 labs 的是，你怎么让
49:30
the evaluation of models more reliable
对模型的 evaluation 更可靠
49:33
than kind of like the state of the art at the time,
比当时的那种 state of the art 更可靠，
49:36
which is like really just finger in the wind.
这真的就只是像用手指测风向一样。
49:39
That's kind of what Anastasia and Wailin's PhD work was
这大概就是 Anastasia 和 Wailin 的 PhD 工作
49:42
as scientists at Berkeley was on statistical methodologies
作为 Berkeley 的科学家，研究方向是 statistical methodologies
49:46
for sort of correcting EVAL estimates
算是用来校正 EVAL estimates 的
49:50
based on like intrinsic biases
基于像 intrinsic biases 这样的因素
49:52
and how you collected the data.
以及你是如何收集 data 的。
49:54
And so control and stuff like that.
还有 control 之类的东西。
49:56
And which is very much like a hey,
而这非常像是在说一声“嘿，”
49:58
how if you're a scientist and you're trying to kind of
比如说，如果你是个科学家，然后你试图某种程度上
50:03
the highest expectation customer for Arena was always
对 Arena 期望最高的客户一直
50:05
like a post training and like a researcher at a lab,
像是做 post training 的，像是实验室里的研究员，
50:10
whereas the highest expectation customer
而期望最高的客户
50:12
from my perspective that Alex really understood
在我看来，Alex 真正理解了的
50:18
and the mission was to serve was like a developer
而且这个使命要服务的，是像 developer 这样的人
50:22
who then takes the result of the research
然后他拿研究的结果
50:24
and then produces an application that's deployed
然后做出一个被 deployed 的 application
50:26
to the world.
向世界。
50:27
It was actually a completely different problem
其实那完全是一个不同的问题
50:28
and person that these two teams were focused on.
以及这两支团队所关注的人。
50:31
And so from the outside in, actually I don't know
所以从外部来看，其实我不知道
50:33
if you remember this, but I have a distinct memory
你是否还记得这件事，但我有一个很清晰的记忆
50:35
of a few weeks before we did the term sheet together
就在我们一起做 term sheet 的几周前
50:39
for Open Router, I'd given you a call
为 Open Router，我给你打过一个电话
50:42
because we were trying to get a pooled data set together
因为我们当时正试着一起弄一个 pooled data set
50:45
from Open Router and from Arena
从 Open Router 和 Arena
50:47
to create like an open source repository of prompts.
来搞一个开源的 prompt 仓库之类的。
50:51
I mean, these projects was so kind of different
我是说，这些项目真的挺不一样的
50:54
in their goals that it was totally normal to me.
在目标上差这么多，对我来说完全正常。
50:57
We'd be like, oh yeah, let's call Alex
我们就会说，哦对，给 Alex 打个电话吧
50:58
and see if he'd want to team up on pooling data
看看他愿不愿意一起组队做 data pooling
51:00
because they're so different.
因为他们太不一样了。
51:01
We need, we actually don't have that kind of data at all.
我们需要，可我们其实根本没有那种 data。
51:04
We didn't have API prompts.
我们没有 API prompts。
51:06
We didn't have like what developers want to do
我们没有那种开发者想用
51:08
with the models, which is very different
这些 models 来做什么的信息，
51:10
from what researchers inside a model lab want to do
而这和 model lab 里的研究人员
51:13
before releasing the model.
在发布 model 之前想做的事非常不同。
51:15
Does that make sense?
这说得通吗？
51:17
And so to this day, I think you see that there's difference
所以直到今天，我觉得你能看到是有区别的，
51:21
even though at a 30,000 foot level,
尽管从三万英尺的高度来看，
51:22
I guess you could kind of conclude
我觉得你大概可以得出这么个结论
51:25
that Arena and Open Router are adjacent
Arena 和 Open Router 算是相邻的
51:28
but the road maps, the missions and so on
但 roadmap、使命之类的
51:31
at the time, at least, were like in very different
至少在那个时候，感觉像是走在非常不同的
51:35
sort of directions.
那种方向上。
51:36
That ideal customer I get, I totally get that.
那个理想客户，我懂，我完全懂。
51:39
As a founder, I want to own everything, right?
作为创始人，我就什么都想自己掌控，对吧？
51:41
So, like this is clearly the adjacency
所以，这明显就是那个相邻机会。
51:43
and I'm like, I'm going to explore that.
然后我就想，我要去探索一下这个。
51:45
Oh, and everything meaning like, you don't know what to do yet.
哦，而且这一切的意思就像是，你还不知道要做什么。
51:48
So you want to like make sure you catch PM there.
所以你想确保在那里能抓住 PM。
51:51
No, I think what he says is you want to own
不，我觉得他说的是，你想要拥有
51:54
the entire infrastructure space
整个 infrastructure 领域
51:56
and so you kind of expand to whatever demand you can capture.
于是你就某种程度上扩展到你能抓住的任何需求。
51:59
Yeah, I think that's, that's hard.
是啊，我觉得那，那很难。
52:01
In reality, because serving multiple customers is
实际上，因为服务多个客户是
52:05
fairly, you know, this is the one that's the focus, right?
平心而论，你知道，这个才是重点，对吧？
52:08
Yeah, I still think even in the age of AI,
是啊，我还是觉得，即便在 AI 时代，
52:11
like focuses is underrated and critical,
就，专注这件事被低估了，而且很关键，
52:15
not just because you end up with a better product
不只是因为你最终能做出更好的产品
52:17
by focusing your humans on it,
通过让你的人专注在这上面，
52:19
but also because the world knows what your focus is.
还因为外界知道你的专注点是什么。
52:22
One thousand.
一千。
52:23
The world can map like, oh, I have this issue,
外界就能对应起来，像，哦，我有这个问题，
52:26
which brand out there is going to help me with that issue.
市面上到底哪个品牌能帮我解决这个问题。
52:29
This is the brand that's known for that focus.
这就是那个以专注这一点而闻名的品牌。
52:32
So like, if I want real attention on this issue,
所以说，如果我想要这个问题真正得到关注，
52:35
like this really matters to me,
这对我来说真的很重要，
52:36
I should go to the brand that cares the most about it.
我就应该去找那个最在乎这个问题的品牌。
52:39
To underscore Alex's point,
为了强调一下 Alex 的观点，
52:41
but how important focus is in the early days of Anthropic,
但专注在 Anthropic 早期有多重要，
52:43
it was not easy to, like people think that
这并不容易，就像人们以为的那样
52:48
the early days of Anthropic were like super easy
Anthropic 早期那会儿感觉就是超级轻松
52:49
because they're on their GPT-2 guys who left,
因为他们就是那批离开的 GPT-2 人，
52:52
but it was actually very competitive.
但其实竞争非常激烈。
52:54
The company was starting $10 billion
这家公司一开始就差了 100 亿美元，
52:56
behind OpenAI, right?
落后 OpenAI，对吧？
52:58
And so to get to the frontier,
所以，要想走到前沿，
52:59
like the big question was, what do we want to be known for?
最大的问题就是，我们想以什么闻名？
53:02
What's the mission?
使命是什么？
53:02
And the mission was at AGI, we have bear programming.
而使命就是 AGI，我们做的是 bear programming。
53:04
And so to the exclusion of all kinds of other things
所以当时，我们把各种其他东西都排除在外了，
53:08
that were really shiny at the time,
那些东西在当时真的很亮眼，
53:10
like image models and video models
比如 image models 和 video models，
53:12
that were getting lots of momentum,
它们当时正获得很多势头，
53:15
the Anthropic team was like,
Anthropic 团队当时就想，
53:16
we just got a focus on coding.
我们就专注在 coding 上。
53:18
Like that is the core capability that we're focused.
因为那才是我们聚焦的核心能力。
53:20
And today you can see the results, right?
而今天你就能看到结果了，对吧？
53:21
It's a trillion dollar company within five years.
它在五年内就成了一家市值万亿美元的公司。
53:23
And that focus, I think, like the focus
而那种专注，我觉得，就像那种专注，
53:26
on who your highest expectation customer is
专注于谁是对你期望最高的客户，
53:28
and how you exceed the expectation,
以及你如何超越这个期望，
53:30
because exceeding anyone's expectations is hard
因为超越任何人的期望都很难，
53:32
and doing it for multiple customers
而要为一群客户都做到这点，
53:34
is so even more difficult.
那就更是难上加难。
53:36
It's part of the reason why OpenRider succeeded
这也是 OpenRider 成功的原因之一
53:38
and Anthropic as well.
Anthropic 也是如此。
53:39
Was the focus on coding that early, though,
不过，真的那么早就专注在 coding 上了吗，
53:41
or did it come later?
还是后来才开始的？
53:42
Literally from day one, it was AI, bear programming
毫不夸张地说，从第一天起，就是 AI，bear programming
53:45
is responsibly commercializing AI, bear programmer
就是负责任地将 AI 商业化，bear programmer
53:49
was the seed memo.
就是 seed memo。
53:50
That was when I invested, right?
那就是我投资的时候，对吧？
53:52
We actually got to refine that memo a lot.
我们其实把那份备忘录改了很多。
53:54
Well, you got to ask Darryl and Tom for permission on that.
嗯，这个你得先得到 Darryl 和 Tom 的许可。
53:57
But it's an extraordinary piece of writing
但那真是一篇非常了不起的文章。
53:59
that they had put together
是他们一起写出来的。
54:00
and AI, you know, commercializing it,
还有 AI，你知道，把它商业化，
54:02
and responsibly commercializing AI, bear programmer
并且负责任地把 AI 商业化，bear programmer
54:04
was the mission, you know, from day one.
你知道，从第一天起，这就是使命。
54:07
And I would say there were maybe like a couple moments
而且我会说，大概有那么几个时刻。
54:10
in the company's history where like they did experiments
在公司历史上，他们做过一些实验
54:13
to kind of see if like little detours made sense
多少是想看看稍微绕点路是不是也说得通
54:15
like a general chatbot, like to our AI when chat GPT
比如一个通用 chatbot，就像在 chat GPT
54:17
was really taking off.
真正起飞的时候，对我们的 AI 来说
54:18
But in the end of the day, especially once they got there
但说到底，尤其是当他们把那
54:23
really significant pre-training computer online,
非常可观的 pre-training compute 上线之后
54:25
I think, like all the main e-veils
我觉得，公司里所有主要的 evals
54:28
at the company, for example, have always been coding e-veils
比如，一直都是在做 coding evals
54:31
long horizon, agentic programming,
long horizon，agentic programming，
54:33
from day one, that was always like that.
从第一天起，就一直是这样。
54:34
There's one like quad, instant came out,
有一个像 quad 的，instant 出来了，
54:36
and quad two came out.
然后 quad two 也出来了。
54:38
Yes.
对。
54:39
I remember the marketing mostly being focused on pros.
我记得当时的营销主要都聚焦在专业用户身上。
54:41
Like this model, good writer, better, and long context.
比如这个模型，擅长写作、更好，还有 long context。
54:46
You watched the first of long context.
你见证了 long context 的开端。
54:49
This directly affected me because it was something like that.
这直接影响到我了，因为差不多就是那样。
54:51
What did you make?
你做了什么？
54:52
Small developer, which was my devin before dev.
小 developer，那是我在 dev 之前的 devin。
54:54
Oh yeah, yes, that's small.
哦对，是的，那确实很小。
54:56
Yes.
是的。
54:57
And you know, so I think there's all that really
而且你知道，所以我觉得那些真的都
55:00
like good focus, is there anything that is a question
挺聚焦的，有没有什么问题是
55:03
that people do want to ask?
大家真的想问的？
55:04
You know, you could have built any other things like,
你知道，你本来可以做任何其他东西，比如，
55:06
and obviously OpenRoto was working, working, working.
而且显然 OpenRoto 一直在运转、运转、运转。
55:09
Were there other ideas that you wanted to pursue
有没有其他你本来想追求的想法
55:12
that you turned down?
但你放弃了？
55:13
You know, just the past road's not taken.
你知道，就是那条没走的路。
55:16
We made a couple prototypes for things that we didn't launch.
我们为一些没发布的东西做了几个原型。
55:18
One was a fine-tuning model as a service.
其中一个就是 fine-tuning model as a service。
55:23
Yeah, lots of that open pipe and all those things.
对，很多那种 OpenPipe 之类的东西。
55:25
But it was, it kind of was at a very
但它其实是，它有点像是一种非常
55:27
consummary form factor where you would give us a YouTube video
consummary 的 form factor，就是你给我们一个 YouTube 视频
55:31
or two or three.
或者两三个。
55:33
We would then extract all the transcripts from it
然后我们会从里面提取出所有的 transcripts
55:36
and try to fine-tune a model to talk like the person
并试着 fine-tune 一个 model，让它说话像那个人
55:39
in the YouTube video or the people in the videos that you sent.
也就是 YouTube 视频里的那个人，或者你发来的那些视频里的那些人。
55:42
So like a really, really easy way of creating a fine-tuned model
所以就是一种真的真的很简单的方式，来创建一个 fine-tuned model
55:46
based on like some kind of videos that you like.
基于一些你喜欢的视频之类的。
55:48
That would be so useful.
那会特别有用。
55:50
We made it too.
我们也做了。
55:52
It was like kind of a, it was, it was, it was, it was,
它有点像是一种，它，它，它，它，
55:57
we didn't actually like test it with that many people
我们其实没有真的在那么多人身上测试过它
55:59
because the model marketplace was our main focus
因为 model marketplace 是我们的主要重点
56:03
and it was like growing
而且它当时在增长
56:05
and we were building more conviction in it over time.
而且随着时间推移，我们对它越来越有信心。
56:09
Just just, yeah, I was, as a creator.
只是只是，对，我当时，作为一个创作者。
56:11
Yes, yes, I'm not a creator of it.
对，对，我不是它的创作者。
56:12
I've been ditched many like,
我被人这样甩过很多次，
56:14
I have 500 hours of recorded voice of myself,
我有 500 小时我自己录下来的声音，
56:17
make a thing of you, you know, a charge access to it.
把你做成一个东西，你知道的，访问它要收费。
56:20
It works for only fans, doesn't work for us as regular people.
这在 OnlyFans 上行得通，但对我们普通人不行。
56:24
I think this is mostly, it's just a glorified ragbot,
我觉得这基本上，就是个被美化的 ragbot，
56:28
whether it's in the weights or outside the weights,
不管是在 weights 里还是 weights 外，
56:30
doesn't really matter, you're just doing rag on the videos
其实都不重要，你只是在视频上做 rag。
56:32
and people ultimately always just want to find the source video
而人们最终总是只想找到那个源视频
56:35
that directly answers it.
能直接回答这个问题的。
56:36
My use case was mostly to practice with myself
我的 use case 主要就是自己跟自己练
56:40
because I often like to see what, like the way I practice
因为我经常喜欢看看，比如说我是怎么练习的
56:43
for a job interview, I'm hiring a candidate
比如准备面试、我在招候选人的时候
56:45
or public speaking or whatever.
或者公开演讲之类的。
56:46
I wish there was like a good mini-me
我希望有个不错的 mini-me
56:48
that I could like critique
能让我像这样点评一下
56:50
because it's kind of hard to pull yourself out.
因为要让自己抽离出来还挺难的。
56:51
I would never get, I would never offer it
我永远不会得到，我永远不会提供它
56:53
to other people in the service.
给这个服务里的其他人。
56:54
Like pick your top five mentors
比如挑出你排名前五的导师
56:56
that didn't talk to them instead of talk to themselves.
那些没跟他们聊，而是跟自己聊的。
56:57
That'd be cool too, yeah.
那也挺酷的，是啊。
56:58
That's the character in the eye, that's a good question.
那就是 AI 里的那个角色，这是个好问题。
57:01
And that was the use case we were aiming at.
而这正是我们当时瞄准的 use case。
57:02
I see.
我懂了。
57:03
It was like, you want to create an experience.
就像，你想创造一种体验。
57:06
Like, AIC jobs and AIC jobs was the initial use case.
就像，AIC jobs 和 AIC jobs 是最初的 use case。
57:10
That's like, you know, it's not allowed.
那就像，你知道的，这是不被允许的。
57:13
That's like, that's like an entrototype, yeah.
那就像，那就像个 entrototype，对。
57:15
It's on a lot of agencies.
很多机构都有这个。
57:16
Fine tuning as a service, as part of the router service,
Fine tuning as a service，作为 router service 的一部分，
57:19
is something that I would typically think about as well, right?
也是我通常会考虑的事情，对吧？
57:22
Like, why don't you do that?
就是说，你为什么不这么做呢？
57:23
Because if people are running already,
因为如果人们已经在跑了，
57:24
their inference through you,
他们的 inference 都走你这边，
57:25
store everything, log with thing,
把所有东西都存下来，用东西 log 一下，
57:27
fine tune to a smaller model
fine tune 到一个更小的 model
57:28
that is cheaper faster, all these things
那个更便宜更快，所有这些
57:29
that's within your control, right?
这些都在你的控制范围内，对吧？
57:31
You didn't do that,
你没这么做，
57:32
but like other people would have pitched that
但像其他人可能就会去 pitch 那个
57:34
in the general state of the infrastructure startup.
在 infrastructure startup 的总体状态里。
57:37
Yeah.
是啊。
57:37
I think you were just maybe a little bit early
我觉得你可能只是稍微早了一点
57:39
because today that's an extraordinarily fast growing segment.
因为今天那是一个增长极其迅速的赛道。
57:42
Like, you know, from a straw,
就像，你知道，从 straw 那边，
57:43
where they do a lot of enterprise deployment,
他们那里做很多 enterprise deployment，
57:44
it's often fine tuning is a custom model for ASMR
经常就是 fine-tuning，就是给 ASMR 做一个 custom model
57:47
or whatever.
或者随便什么吧。
57:48
But not as a router.
但不是拿它当 router。
57:50
They're just like, I come to you because I like your
他们只是会说，我来找你，是因为我喜欢你们的
57:51
industrial models.
industrial models。
57:52
I want custom industrial model, right?
我想要 custom industrial model，对吧？
57:54
It is not I want to run all my open AI prompts,
并不是说我想跑我所有的 OpenAI prompts，
57:58
get to store all my results
能把我所有的结果都存起来
57:59
and then just move off of open AI.
然后就直接从 OpenAI 迁走。
58:01
Right?
对吧？
58:02
They're not doing that.
他们没这么做。
58:03
As like a way to export off of dependency
作为一种摆脱对
58:05
on a frontier lab, I've not seen that yet,
frontier lab 依赖的方式，我还没见过这种情况，
58:07
which was your kind of position to do.
而这算是你那种定位会去做的事。
58:10
I mean, we decided, really we leaned into our focus
我是说，我们决定了，真的就是更聚焦于我们自己的重点，
58:15
and figured that like we just saw
然后想着，我们就是看着
58:18
the ecosystem develop over time.
这个 ecosystem 随着时间慢慢发展起来。
58:20
All these inference providers that do want to help
所有这些确实想帮忙的 inference providers，
58:23
companies do that.
帮公司做到这一点。
58:25
Like, it makes sense for us to partner with them
就是，我们跟他们合作很合理，
58:28
and to like give users lots of choice
然后就是给用户很多选择，
58:30
and to like figure out what makes them,
然后就是搞清楚他们靠什么、
58:34
what gives them competitive advantages.
是什么给了他们竞争优势。
58:36
It's a whole new business basically.
基本上，这是一个全新的业务。
58:39
And there's value in being a neutral marketplace
而做一个中立的 marketplace 本身就有价值。
58:41
that just kind of like works with those companies.
那种方式就是有点能跟那些公司配合得上。
58:45
Could you share a little bit to Sean's point,
能不能就 Sean 刚说的那点，稍微多分享一点，
58:47
like how you prioritized,
比如你是怎么 prioritize 的，
58:49
what are some ways you prioritize features?
你都有哪些方式去 prioritize features？
58:52
Because you've always done it so elegantly,
因为你一直做得都特别优雅，
58:54
that never just happens.
这种事从来不会凭空发生。
58:56
And you make all the right decisions
而且你总能做出所有正确的决定，
58:57
that always have product market fit
总是有 product market fit
58:58
from the outside looking in.
从局外人的角度来看。
58:59
But consistently, you seem to have prioritized,
但一直以来，你似乎都在优先考虑，
59:02
you know, a lot of hit features that worked
你知道，很多成了爆款、也确实有效的 feature
59:03
and maybe I have samples that bias or whatever.
也许我手上的 samples 有 bias，或者什么的。
59:05
But Sean,
不过 Sean，
59:06
well, it's what you think hit features worked.
嗯，那也只是你认为哪些 hit feature 成功了。
59:08
Well, like the leaderboards, like leaderboards, okay?
嗯，比如 leaderboards，比如 leaderboards，行吧？
59:11
Yeah, you know, like from day one
是啊，你知道，就像从第一天开始
59:13
to think of how they're like, the feedback charting,
想想它们是怎么样的，像，那个反馈图表化，
59:15
BYU, okay.
BYU，好。
59:15
But like, he had like plugins,
但就像，他有一些插件，
59:17
you know, he had like,
你知道，他就像，
59:19
and I think there was a whole thing
而且我觉得有一整套东西
59:20
I want to get into about like completions versus check,
我想深入聊聊，比如 completions 和 check 的对比，
59:23
check features versus completions.
check features 和 completions 的对比。
59:24
And then also it's called it like the rise of the reasoning
然后还有，它被叫做，像是，reasoning 的崛起
59:27
models and how you deal with those,
models 以及你怎么处理这些，
59:29
multi-modality, all those things.
multi-modality，所有这些。
59:31
BYU, okay.
BYU，好。
59:32
That was a huge one.
那是个大事。
59:34
There was one I think it was in early 2024.
我记得有一个，应该是在 2024 年初。
59:40
Very early 2024,
2024 年非常早的时候，
59:42
we thought it might be interesting to use the results
我们觉得，把结果拿来用可能会挺有意思，
59:45
of multiple models together.
也就是多个 models 放在一起的结果。
59:47
And we launched a prototype called Mom,
然后我们发布了一个叫 Mom 的 prototype，
59:51
mixture of models that let you like pick a couple models
一个 mixture of models，让你有点像可以挑几个 models
59:55
and we'd pick them for you
然后我们会帮你挑
59:57
and then it would fuse the results together at the end
然后在最后把结果融合到一起
60:00
and it would show you all the intermediate results
而且它会给你看所有 intermediate results
60:02
in this like big con-bon board-looking product.
就在这个像很大的 con-bon board 一样的产品里。
60:05
What does diffusion at the end, another model?
最后 diffusion 是做什么的，另一个 model？
60:07
Another model.
另一个 model。
60:08
The smartest of the set of the three of the set.
这组三个里面最聪明的那个。
60:12
So this is a good console idea.
所以这是个不错的 console 想法。
60:14
It was like a very early LOM council.
那就像是一个非常早期的 LOM council。
60:16
This is a multi-agent swarm as like they would call it
这就是他们所说的 multi-agent swarm
60:19
at one of the frontier labs.
在某个 frontier labs 里。
60:20
There it is, you know?
就是这样，你懂的？
60:23
Yeah, like some of those ideas are like
是啊，有些想法确实
60:24
going the right direction but the devil's in the details
方向是对的，但魔鬼在细节里。
60:27
there's a lot of like product refined that needed
就有很多像产品打磨之类的事需要做
60:29
to make them really work.
才能让它们真正跑起来。
60:32
They take your focus away from like, you know,
它们会把你的注意力从，就，你知道的，
60:34
whatever else you have going on
你手头其他所有事情上拉走
60:37
and there's a lot of like community building
而且还有很多像社区建设
60:39
and learning that you need to do
和学习之类的事需要你去做
60:40
and the technology might be too early.
而且技术可能还太早了。
60:43
Like all kinds of reasons they might go wrong
就各种各样的原因吧，它们都可能出问题。
60:45
and in our case the technology was a little too early.
而在我们这种情况下，这项技术有点太超前了。
60:48
In other words, the fused result was a little bit worse
换句话说，fuse 之后的结果稍微差了一点
60:53
sometimes the same as the best model
有时候跟最好的模型一样
60:56
that was being used to fuse
也就是当时被拿来 fuse 的那个
60:57
because the best model was so far ahead
因为那个最好的模型当时领先太多了
61:00
of options two and three at the time, you know,
比当时的第二和第三选择领先太多了，你知道，
61:04
over time the top three or four LOMs
随着时间推移，前三或前四的 LOMs
61:08
have gotten closer together, still neurodivergent
已经越来越接近了，但还是 neurodivergent
61:11
but like all capable of inserting like pretty interesting ideas
但就像，全都能塞进一些挺有意思的想法
61:14
like RL has basically like expanded the surface area
就像 RL 基本上，就像，扩大了
61:18
of creativity for machine learning researchers
machine learning 研究者的创造力发挥空间
61:21
within each lab and so they can, you know,
在每个实验室里，这样他们就能，你知道，
61:23
diversify the reasoning power of different models
让不同 models 的 reasoning 能力更加多样化
61:26
more effectively.
而且更有效一些。
61:27
At least that's my theory for why fusion is like works better
至少这就是我的理论，解释为什么 fusion 就像，效果更好
61:32
than it used to but early 2024.
比起以前，但那是 2024 年初。
61:34
And so the technology was a little bit too primitive.
所以技术还有点太原始了。
61:38
The form factor was not right
form factor 也不对
61:40
and so we would have had to go through
所以我们就得经历
61:41
like a couple more iterations
大概还得再来几轮 iterations
61:43
and so we decided to just delete all the count.
所以我们就决定干脆把所有 count 都删掉。
61:45
And then years later,
然后几年后，
61:49
middle of 2026 or early 2026
2026 年中期或者 2026 年初
61:52
were like, let's bring it back.
我们又想，把它弄回来吧。
61:54
Like the research is looking kind of promising for fusion.
就感觉 fusion 的研究看起来还挺有希望的。
61:58
The models now have like two, three, four
现在这些 models 大概有，像，两、三、四
62:01
top frontier models that are all really good
个顶级 frontier models，而且全都特别强
62:04
and like I'm frequently trying to like
而且我经常会，就，试着去
62:07
consult multiple models to get the best results.
咨询多个 models，来拿到最好的结果。
62:10
And then I ran a little personal experiment
然后我还做了个小小的个人实验
62:13
where I was like, I'm gonna like do a
当时我想，我要，弄一个
62:16
an architecture plan for a code change.
针对 code change 的 architecture plan。
62:20
I'm gonna give it to all the models.
我会把它发给所有 models。
62:21
I'm gonna fuse the result.
我会把结果融合一下。
62:22
And then I'm gonna ask all the models
然后我会问所有 models
62:24
if the fused result is better
融合后的结果是不是更好
62:25
than the individual result each model came up with.
比每个 model 自己给出的单独结果更好。
62:28
And they all said yes, that the fused result was better.
它们都说对，融合后的结果更好。
62:31
And this happened a couple times
这种情况发生了好几次
62:33
and I was like, okay, spot check, pretty good.
我就想，好吧，抽查一下，挺不错。
62:36
We should like benchmark this
我们应该 benchmark 一下这个
62:38
and that's how we built fusion.
然后我们就是这样做出 fusion 的
62:40
Yeah, it came on your fable.
对，它是在你的 fable 上出现的
62:41
So you were like, this is fable level.
所以你就说，这就是 fable 级别的
62:43
Yeah, yeah.
对，对
62:44
Let's start leading up to this year.
咱们从今年之前开始讲起吧
62:47
Should we have been gone to this year?
我们是不是应该已经讲到今年了
62:50
Can you mark out the main milestones in the journey?
你能标出这段历程中的主要里程碑吗
62:52
I think it seems like your promise was,
我觉得，好像你的承诺是，
62:56
routing, you decided the business model
routing，你决定了 business model，
62:58
very early, you take a cut.
很早就定了，然后你抽成。
63:00
And like, what are the major milestones
而且，像，有哪些重要的里程碑
63:03
that inflect the growth, right?
会让增长出现拐点，对吧？
63:05
Like you're growing like 9% week and week now,
就像你现在周环比增长大概 9%，
63:08
is this the official number?
这是官方数字吗？
63:10
Interpreter token volume.
Interpreter 的 token volume。
63:11
I think that sounds about right, yeah.
我觉得这听起来差不多，对。
63:13
So just like, can you mark out
所以就是，你能不能大概梳理一下
63:15
like the sort of brief history of open router
就是 open router 的那种简要历史
63:17
up to the acquisition?
一直到被收购？
63:20
Let's call it, we're just talking about,
就这么说吧，我们只是在聊，
63:24
people have sort of your birth moment
大家有点就是在聊你诞生的那一刻
63:27
with the missile stuff
在 missile 那些事上
63:29
where people are really competing.
那时候大家真的在竞争。
63:30
You have your state of the eye thing, where it's very cute.
你有你那个 state of the eye 的东西，挺可爱的。
63:33
You have 100 trillion tokens, haha, haha.
你有 100 万亿 tokens，哈哈，哈哈。
63:35
Because now you're doing 10 a week, you know.
因为现在你一周做 10 个，你知道吧。
63:39
We're doing 10 a day.
我们现在一天做 10 个。
63:41
10 a day now?
现在一天 10 个？
63:42
Yeah, so yeah, you do this in 10 days.
对，所以对，你 10 天就能做完这个。
63:45
What are the major points there?
那主要的点是什么？
63:47
I just want to, like, this is move curve,
我就是想，那个，这是 move curve，
63:49
but like, you feel the inflections.
但就像，你能感受到那些语气起伏。
63:51
A lot of this is kind of oriented around model launches.
很多这些其实都有点围绕 model launches 展开。
63:55
We had a huge focus on pros all the way up through May of 2024,
一直到 2024 年 5 月，我们都特别关注专业用户，
64:07
because coding was just not there,
因为 coding 当时根本还没跟上，
64:09
and no apps were able to build much on top of that.
而且没有 apps 能在这之上构建出多少东西。
64:11
So a diversity in models, but not a wide diversity
所以 models 是有多样性的，但算不上很广泛的多样性，
64:17
and not a wide diversity in use cases.
在 use cases 上也没有很广泛的多样性。
64:20
Dream Tavern was one of our top apps.
Dream Tavern 是我们排名最高的 apps 之一。
64:22
At the time, the creative Dream Tavern
当时，Dream Tavern 的创意者
64:24
now runs product at Cognition, Devon.
现在在 Cognition 负责产品，Devon。
64:30
Then we, in the middle of 2024, we saw Claude Saun at 3.5.
然后，在 2024 年年中，我们看到了 Claude Saun 3.5。
64:36
That came out incredible leap forward in coding.
那在 coding 上是一次不可思议的飞跃。
64:40
And we saw the dynamics of apps building on top of us change.
我们也看到，构建在我们之上的 apps 的动态发生了变化。
64:45
We saw a huge surge in volume in users using OpenRouter.
我们看到使用 OpenRouter 的用户量出现了巨大激增。
64:51
And this is when I think people started to look
而我觉得，也就是在这个时候，人们开始关注
64:54
at like, money that they were spending and get a little bit like,
比如说，他们花的那些钱，然后开始有点，像是，
64:57
whoa, what's going on?
哇，这是怎么回事？
64:59
I might need to think about like, more cost efficient,
我可能得考虑一下，像是，更 cost efficient 的，
65:01
but equivalent models.
但等价的 models。
65:03
And shortly after that, I think it was after Saun at 3.5,
然后没过多久，我想是在 Saun at 3.5 之后，
65:06
mixed role, 8X7B came out.
mixed role，8X7B 出来了。
65:09
And everyone was like, what, this is the model?
然后大家都说，什么，这就是那个 model？
65:12
Like, the open weights community delivered.
就像，open weights 社区真的交付了。
65:14
And so it was really good timing from the strong.
所以从 the strong 来说，时机真的非常好。
65:17
And all of on just board closer, just helping you.
而这一切，归根结底，就是在帮你。
65:19
It takes an ecosystem to go in OpenRouter.
要在 OpenRouter 里做起来，需要一个 ecosystem。
65:24
Yeah, that was the, yeah, it was like, this early ecosystem,
对，那就是，对，就像是，这个早期的 ecosystem，
65:29
it was like a swing action, where like model apps would come up
它就像一种 swing action，model apps 会冒出来
65:34
with some sort of frontier innovation.
带着某种 frontier innovation。
65:37
Like usage would surge.
然后 usage 就会暴涨。
65:39
Then users, you know, look at their invoices 30 days later,
然后用户呢，你懂的，30 天后看自己的账单，
65:42
and they're like, whoa, what's going on here?
然后他们就会说，哇，这是怎么回事？
65:44
And then open weight models would deliver like cost effective options
然后 open weight models 会带来那种高性价比的选择
65:49
to three months later.
大概三个月之后吧。
65:51
We saw that happen several times.
这种情况我们见过好几次。
65:54
One thing for the coding agents was that you broke out,
coding agents 这块有一点是，你冲出来了，
65:57
which are the top coding agents, and they love that.
也就是那些顶级 coding agents，他们特别喜欢这个。
66:00
They love that leaderboard.
他们超爱那个 leaderboard。
66:01
The client versus the root code versus the what have you.
client 对 root code，对那个什么来着。
66:04
Yeah, like, client was like the top of our leaderboard
对，就像，client 差不多是我们 leaderboard 的第一名。
66:08
at the time.
当时。
66:09
We then, at the end of, and I'll skip for it a little bit,
然后，到了……末尾的时候，这里我会稍微跳过一点，
66:15
at the end of 2025, there were quite a few coding apps
到了 2025 年底，已经有相当多的 coding apps
66:18
on the leaderboard.
在 leaderboard 上。
66:19
They were all IDs or terminal-based agents.
它们全都是 IDs 或者 terminal-based agents。
66:25
And at the end of 2025, we saw OpenClaw appear.
然后到了 2025 年底，我们看到 OpenClaw 出现了。
66:29
And OpenClaw was particularly interesting
而且 OpenClaw 特别有意思
66:32
because one, it was like a new form factor
因为第一，它像是一种新的 form factor
66:35
that brought in a new type of user,
这带来了新一类用户，
66:38
not just a developer, but like a productivity
不只是 developer，而是像 productivity
66:41
or sort of an internet creator came to AI
或者某种 internet creator，来接触 AI
66:46
for the first time.
第一次。
66:48
And it also had an interesting architecture
而且它还有个挺有意思的 architecture
66:52
where it was like calling your chosen model
就是会去调用你选好的 model
66:54
for these heartbeats to see if it was still alive
发这些 heartbeats，看看它还活着没
66:57
in addition to actually using the model for real task.
除了真的用这个 model 来做实际任务之外。
67:00
And the heartbeats are like, they're kind of,
而且 heartbeats 就是那种，有点像，
67:02
you don't want to pay every 30 minutes heartbeats.
你不会想每 30 分钟就为 heartbeats 付费。
67:05
So the auto router that we provided
所以我们提供的那个 auto router
67:09
was really, really useful to this like wide range
真的真的对这么广泛的一类
67:12
of users all of a sudden.
用户来说，突然之间就特别有用。
67:13
And so we just saw it rocket exponentially.
然后我们就看到它指数级飙升。
67:17
And then we saw like OpenClaw just blow up
接着我们又看到 OpenClaw 直接爆了
67:20
and a couple other apps leaned into that new paradigm
还有另外几个 app 也拥抱了这个 new paradigm
67:25
and do something similar.
然后做类似的事情。
67:27
Hermes came out and really leaned into things
Hermes 出来之后，真的把重心放在了这些东西上
67:30
like the auto router and built like a really good community
比如 auto router，还打造了一个特别好的社区
67:34
and leaned into like basically skill management
而且基本上就是主攻 skill management
67:39
and making it really easy and effective
让它变得特别简单又高效
67:40
for people like set their memory in the agent
让人们可以在 agent 里设置自己的 memory
67:43
and build really good skills.
并构建出特别好的 skills。
67:46
Another thing you never did memory skills sandboxes,
另外一件你从来没做过的事：memory、skills、sandboxes，
67:49
always like adjacent things, you could have done.
总是，像，相邻的那些东西，你本来可以做的。
67:52
Could have but like it's, I think like it's hard to bit.
本来可以，但就像，我觉得就像很难 bit。
67:55
They're also very,
它们也非常，
67:57
there are things that developer that really matter
有些东西，对 developer 来说真的很重要
68:00
for like the developer use cases
对于像 developer use cases 这样的
68:01
that were coming out at the time.
当时正在出现的那些。
68:03
Like developers wanted to architect those things.
就像 developers 想架构那些东西。
68:05
Those are kind of critical to building a good user experience.
那些对构建好的 user experience 来说算是挺关键的。
68:08
It's really, it's been hard for companies to find abstractions
说真的，公司一直很难找到 abstractions
68:12
that work for all developers on the memory layer.
那种能适用于 memory layer 上所有开发者的。
68:15
It is, yeah, there are some like Mastera
是啊，确实，有些像 Mastera
68:19
has done a pretty good job, for example,
比如说，就做得相当不错，
68:21
but like developers have like lots of very preferences for them.
但是呢，开发者对它们就是有很多偏好。
68:26
And then we saw, you know,
然后我们看到，你知道，
68:29
the way our leaderboard has changed over time
我们的 leaderboard 一路是怎么变的
68:33
is kind of like a movie of how the AI space has changed over time.
有点像一部电影，讲的是 AI 领域一路是怎么变的。
68:37
If you just sort of like go the way back machine
如果你就，怎么说呢，去 way back machine 翻一翻
68:39
and look at the rankings leaderboard
然后看看 rankings leaderboard
68:42
and the apps leaderboard over time,
再看看 apps leaderboard 随着时间的变化，
68:44
it sort of shows you like what's happened
它差不多就让你看到，像是发生了什么
68:46
in the AI over the last couple of years.
在 AI 里，过去这几年。
68:48
To me, the coming of each moment was,
对我来说，每个时刻的到来就是，
68:50
Andre Karpathy was like, I no longer read Locolama
Andre Karpathy 当时就说，我不再读 Locolama 了
68:52
because like I just go to Open Cloud,
因为，怎么说呢，我就直接去 Open Cloud，
68:54
Open Router's leaderboard.
Open Router 的 leaderboard。
68:55
Which I remember that, anyway.
反正这个我记得。
68:58
I think you probably like to say, like,
我觉得你大概会想说，就是，
68:59
sorry guys, I'm gonna send a bunch of traffic to you.
抱歉各位，我要给你们送一大堆 traffic 过去。
69:03
So I was gonna bring it into the strike thing.
所以我本来打算把它扯进 strike 那件事里。
69:06
How does that kind of conversation start?
这种对话是怎么开始的？
69:09
We had this long standing relationship with Stripe though
不过我们跟 Stripe 有很长久的合作关系，
69:13
from you know, like many different projects
是从，你知道，像很多不同的项目来的。
69:17
that we had worked on with them.
那个我们之前和他们一起做的。
69:19
We invest, you know, a lot of effort
我们投入了，你知道，很多精力
69:22
in countering abuse and token fraud.
在打击滥用和 token fraud 上。
69:26
Can you give some numbers just to so people understand?
你能不能给点数字，好让大家明白？
69:29
I think I like, I posted about this.
我想我，呃，我发过关于这个的帖子。
69:31
We blocked 10X as much dollar volume last month
上个月我们拦截的美元交易额是 10X
69:37
as the month before.
那么多，相比前一个月。
69:39
And the types of token fraud are diversifying
而且 token fraud 的类型正在多样化。
69:42
quite a bit.
还挺多的。
69:43
You know, there are like fraudsters
你知道，有那种骗子
69:46
going after typical stolen credit cards,
专门盯着常见的被盗信用卡，
69:48
but they're also, you know, people trying to resell traffic
但也有，你知道，想转卖 traffic 的人
69:52
against the terms of service.
违反 terms of service 的。
69:54
There's like hacked accounts.
还有那种被 hacked 的账号。
69:55
There's people who just lose act, you know,
还有人就是会失去访问权限，你知道，
69:57
like their whole company is compromised
就像他们整个公司都被 compromised 了
70:00
and they don't even realize it.
而且他们自己都完全没意识到。
70:01
And we help them like regain control and detect it.
而我们会帮他们重新掌控局面，并把这种情况检测出来。
70:06
There's, there are accounts that are like
有些，有些账号会
70:08
reselling inference on the side.
顺便转卖 inference。
70:09
There's, there are accounts that are dealing with,
还有些，有些账号是在处理，
70:14
you know, like an accidental runaway agent
你知道，就像是一个意外失控的 agent
70:18
and they don't realize it.
而他们自己还没意识到。
70:19
Not a hack, but it's something that blows up
不是被 hack，而是那种会直接爆掉的事情。
70:22
and the company doesn't want it.
而公司并不想要它。
70:24
And so our trust and safety team
所以我们的 trust and safety team
70:26
like works a lot on all of these like categories
就是会花很多精力处理所有这些，像是各种类别
70:30
of problems and helps block it and detect it.
的问题，帮忙拦截和检测它。
70:34
And so we've built these, you know,
所以我们建了这些，你知道，
70:36
we have models around them.
我们有围绕它们的 models。
70:39
We worked closely with Stripe for a while on this.
我们有一阵子在这件事上和 Stripe 密切合作。
70:42
And I think it's going to become a huge problem
而且我觉得这会变成一个巨大的问题
70:44
in the ecosystem.
在这个 ecosystem 里。
70:45
Like we're already seeing a lot of companies
就像我们已经看到很多公司
70:48
start to see these fraudsters like spread
开始看到这些骗子在蔓延
70:52
and look for other ways, other, you know,
然后去找其他方式，其他，你懂的，
70:55
other than open router to other fraud vectors.
除了 open router 之外，转向其他 fraud vectors。
70:59
And if you're making a gateway
而如果你在做 gateway
71:02
or selling like generalized inference,
或者卖类似 generalized inference 的东西，
71:04
you are a target for fraud.
你就会成为欺诈的目标。
71:06
If you're selling very discrete like intelligence products,
如果你卖的是那种非常 discrete 的智能产品，
71:11
intelligence products that are like doing something
智能产品，就是那种做某件事
71:12
pretty specific but not like, you know,
相当具体，但不是那种，你知道，
71:14
just reselling inference with some added capability,
只是把 inference 转售出去、再加点能力，
71:18
then you're way less likely to get these fraudsters.
那你碰到这些骗子的可能性就小得多。
71:20
So it, I think we'll see companies also move away
所以，我觉得我们会看到公司也开始不再
71:23
from just reselling inference
只是转售 inference
71:25
with some sort of like added capability
再加上某种类似附加能力
71:28
and move towards sort of like discrete tasks
然后转向有点像离散任务那种方向
71:32
and charging for those tasks
然后对这些任务收费
71:34
and charging for those enhancements
然后为这些增强功能收费
71:35
and letting people bring their own inference
然后让人们自带 inference
71:38
like in a third party way.
有点像第三方的方式。
71:39
Whoa, okay.
哇，好吧。
71:41
And yeah, obviously you would power that.
而且是啊，显然你会为它提供算力。
71:44
But people will pay for outcomes or per task.
但人们会为结果付费，或者按任务付费。
71:48
I think people will pay, you know,
我觉得人们会愿意付钱的，你知道，
71:51
I think like the data dog pricing page
我觉得就像 DataDog 的 pricing page
71:55
is a good look at like the future to come.
能很好地看出未来会是什么样。
71:58
It's like companies, like infrastructure companies
就是像那些公司，像基础设施公司
72:00
will like charge for different types of events
会按不同类型的 events 来收费
72:05
that they're providing
也就是他们提供的那些
72:06
and there'll be lots of like continuous pricing models
而且会有很多那种 continuous pricing models
72:09
that look like that.
看起来就是那样。
72:10
And of course there will be like,
当然也会有那种，
72:11
if you go down towards consumer apps, you know,
如果你往 consumer apps 那边走，你知道，
72:15
simpler pricing, more subscriptions,
更简单的 pricing，更多 subscriptions，
72:19
you know, fewer events to worry about
你知道，要操心的 events 更少，
72:23
and ones that like are not focused
而且那些也不是那种只专注于
72:25
on just adding a markup on top of inference.
在 inference 之上加一层 markup。
72:28
It's not just because fraud is hard,
这不仅仅是因为 fraud 很难，
72:31
but also because the pressure from the labs
也是因为来自 labs 的压力
72:35
and from like good inference providers
而且从那种好的 inference providers
72:39
to like do a commit
到比如去做个 commit
72:41
and then bring your inference elsewhere
然后再把你的 inference 带到别处
72:43
is gonna be very high.
会非常高。
72:44
Any comments?
有什么评论吗？
72:45
Two, one, you know, I think Alex has done
二，一，你知道的，我觉得 Alex 已经
72:48
a very eloquent job of describing something,
非常雄辩地描述了某件事，
72:51
you know, counterintuitively I knew
你知道，反直觉的是，我知道
72:53
would be a thing at scale like four years ago
大约四年前就会大规模地存在了
72:55
because of discord and the particular experience
因为 discord 和那种特别的体验
72:57
that taught me this was, you know,
教会了我这是，你知道，
73:00
as we started scaling mid-journey,
当我们开始 scaling mid-journey 的时候，
73:02
you know, the one of the primary ways
你知道，其中一个主要的方式
73:04
that we used to give away or like get people
我们过去用来赠送或者让人们
73:07
to try mid-journey early on to get the first 10 generations
早期尝试 mid-journey 来获得前 10 个 generations
73:10
because you know, 10 generations of,
因为你知道，10 个 generations 的，
73:12
10 images generated was roughly
生成 10 张图片差不多就是
73:14
the magic moment activation point we found.
我们找到的那个 magic moment activation point。
73:16
Like once you've done 10, you were like,
就像你一旦做完 10 张，你就会觉得，
73:18
this is extraordinary.
这太不可思议了。
73:20
But for that we, so we had a free trial with mid-journey.
但为了那个，我们，所以我们当时有 mid-journey 的免费试用。
73:23
And one day I woke up because I had a platform
然后有一天我醒来，因为我有个平台
73:25
and had to monitor all these dashboards.
还得盯着所有这些 dashboards。
73:28
You know, I had like three missed calls from David
你知道吧，David 给我打了差不多三个未接来电
73:30
and it turns out like there had been
结果发现，好像之前有
73:32
this flood of new users overnight.
一夜之间涌来的一大波新用户。
73:35
And we were like, this is great.
我们就说，这太棒了。
73:38
And he was like, no, actually we shot down
他却说，不，其实我们关掉了
73:39
the free trial and I was like, why is that?
免费试用，我就问，为什么？
73:42
And he said, I don't look at the geolocation IP addresses.
他说，我不看 geolocation IP addresses。
73:44
And basically somebody in China had started
基本上就是中国有人开始
73:47
to resell mid-journey free, you know, subscriptions
转卖 mid-journey 的免费订阅，你懂的
73:49
with the free trial as a way to like basically, you know,
就是把免费试用当成一种方式，基本上，你知道，
73:52
it was fraud abuse, right?
那就是欺诈滥用，对吧？
73:54
And was it a specialized model like mid-journey?
而且那是一个像 mid-journey 那样的 specialized model 吗？
73:56
Yeah. And that was an action application.
对。而且那是一个 action application。
73:58
So this idea that I think the big picture
所以这个想法，我觉得，那个整体图景
74:01
of realization I had back then was,
我当时意识到的就是，
74:03
hey, there's a new type of unit of value
嘿，有一种新的价值单位
74:07
that's being streamed across the internet called a token.
正在互联网上被流式传输，叫做 token。
74:11
And over the next 10 years, the entire internet value chain
而在接下来的 10 年里，整个互联网价值链
74:16
was going to have to deal with the fact that like
都将不得不面对这样一个事实：
74:18
the more valuable tokens got, the more bad actors
tokens 越有价值，就会有越多坏人
74:23
were going to go to try to get their hands on those tokens.
想方设法去染指这些 tokens。
74:26
And anytime you scale something and the payload
而任何时候，当你把某个东西规模化，并且载荷
74:29
gets more and more valuable, more bad thing people
变得越来越有价值时，就会有更多坏人
74:32
try to get access to that value.
试图获取那份价值。
74:34
And so it was very obvious to me back then.
所以当时这对我来说非常明显。
74:38
And so look, to this day, I don't think there's a free turn.
所以你看，直到今天，我都不觉得有免费机会。
74:40
Like I don't think mid-journey's ever actually
就像我觉得 Midjourney 其实从来没有
74:42
turned on the free trial since then
从那以后开启过免费试用
74:44
because it was really not an easy problem
因为那真的不是一个容易的问题，
74:47
to solve interpretive trust and safety.
也就是要解决 interpretive trust and safety。
74:49
And that's why I started teaching the class security at scale.
所以我才开始教 Security at Scale 这门课。
74:52
Like one of that and the anthropic learnings
就像其中一点，以及 Anthropic 的那些经验教训
74:55
to me was clear that the need for security at scale
让我清楚看到，对 security at scale 的需求
74:58
was going to be enormous a few years from then.
从那时起再过几年，这会变得极其庞大。
75:00
Because if you just do the math, think about it.
因为只要你算一下，想想看。
75:03
If where online payments started roughly in the 80s and 90s,
如果在线支付大概是在80年代和90年代起步的，
75:10
and grew to over $1 trillion over the next 10 years,
并在接下来的10年里增长到超过1万亿美元，
75:14
and we needed to build entirely new payment solutions
而且我们需要构建全新的支付解决方案
75:16
to deal with online fraud, where we are today
来应对在线欺诈，而今天我们所处的位置
75:19
is roughly there on tokens.
在 tokens 上差不多就在那里。
75:22
But over the next even five years,
但在接下来哪怕只有五年内，
75:25
we're expecting the token economy to get to roughly $5 trillion.
我们预计 token economy 会达到大约 5 万亿美元。
75:29
And over the next 10 years, I'd be shocked
而在未来 10 年里，我会非常震惊
75:30
if we went to $10 trillion of token flow.
如果我们达到 10 万亿美元的 token flow。
75:33
And so if we were starting to see such aggressive abuse
所以，如果我们开始看到如此猖獗的滥用
75:37
and fraud at sub-scale mid-journey,
以及欺诈，在 mid-journey 还处于 sub-scale 时，
75:39
remember mid-journey at this point
记住，mid-journey 当时
75:40
was less than 300 million revenue run rate a year.
的年 revenue run rate 还不到 3 亿。
75:44
I just realized we were going to need entirely new systems
我这才意识到，我们将需要全新的系统。
75:49
to deal with the fraud that was going to happen
为了处理即将发生的欺诈
75:51
for trying to get into the token flow.
因为想进入 token flow。
75:53
And so I forget the board meeting it was
所以我忘了是哪次董事会会议了
75:57
when you brought up that Stripe wanted to partner up.
当时你提到 Stripe 想合作。
76:00
And it made so much sense to me because Stripe rate
而这对我来说太有道理了，因为 Stripe，对吧？
76:02
are when I was a client or 10 years ago,
而且在我还是客户的时候，或者说10年前，
76:04
we invested in Stripe in the whole pitch
我们投资了 Stripe，整个 pitch
76:05
that Patrick and John communicated so eloquently
Patrick 和 John 把它讲得如此有说服力
76:08
was like, hey, unlike traditional payment tools,
就像在说，嘿，跟传统的支付工具不一样，
76:11
like Braintree, that do a seven day verification
比如 Braintree，它们要做七天的验证，
76:14
like KYC and AML to get the fraud out of the way,
比如 KYC 和 AML，先把欺诈问题处理掉，
76:18
we actually just bite the fraud cost up front
我们其实是先自己把欺诈成本扛下来，
76:20
as a customer acquisition cost and give tell a developer,
把它当作 customer acquisition cost，然后告诉开发者，
76:23
like just use five lines of code
就像，你只要用五行代码，
76:24
and we start accepting your payments in five minutes.
我们就能在五分钟内开始接收你的付款。
76:27
And what'll happen is over time we collect all this data
接下来会发生的是，随着时间推移，我们会收集到所有这些数据
76:30
on the developer's...
在开发者的……
76:31
Cloudflip model.
Cloudflip model。
76:32
...is the Cloudflip model, right?
……就是 Cloudflip model，对吧？
76:33
And they did.
而他们确实做到了。
76:34
Five years later they launched Stripe Rader
五年后，他们推出了 Stripe Rader
76:35
and Stripe really today is a security company.
而 Stripe 如今真是一家安全公司。
76:37
That's the real people think it's a payments company?
这就是人们真的以为它是一家支付公司的原因？
76:39
No, there's lots of other payments providers today
不，如今有很多其他支付服务商。
76:42
that give you cheaper payments transmission.
能给你更便宜的 payments transmission。
76:44
But the reason Stripe keeps being the dominant one
但 Stripe 之所以一直是那个主导者
76:47
here in Audien and Europe is because they have extraordinary
在 Audien 和 Europe，是因为他们有非凡的
76:50
fraud detection that they've built over the years.
fraud detection，这是他们多年来建立起来的。
76:52
It's the same story with Elon and Max Levchin and...
Elon 和 Max Levchin 也是同样的故事，还有……
76:55
And a firm.
还有一家公司。
76:56
Yeah.
是啊。
76:56
So I think the story shows up over and over again
所以我觉得这个故事会一次又一次地出现。
76:59
where every time you have value streamed across the world
每当有价值在世界各地流动的时候
77:03
in large amounts, you need new protection
而且是大量流动，你就需要新的防护
77:05
and security infrastructure to keep the bad guys out
以及安全基础设施，把坏人挡在外面
77:08
and allow the good people to have their transactions
让好人能够让自己的交易
77:10
happen really fast.
非常快速地发生。
77:11
And so I think this is why, from my perspective,
所以我觉得，从我的角度来看，这就是为什么，
77:15
the Stripe and Open Router story is a security story
Stripe 和 Open Router 的故事是一个安全故事
77:18
for the Internet ecosystem, for the Frontier AI ecosystem
对于 Internet 生态，对于 Frontier AI 生态
77:21
without a partnership like that.
没有那样的合作关系。
77:22
It becomes very hard to defend the quality of experience
要守住体验质量就变得非常难
77:25
and the speed and all the good stuff
还有速度，以及所有那些好的东西
77:27
without letting the bad guys get in the way.
同时又不让坏人挡道。
77:30
The second is that there's this underappreciated thing
第二点是，有这么一个没被充分重视的点
77:34
about the fact that you need to...
就是关于你其实需要……这个事实
77:37
All the bad things that Alex described
Alex描述的所有坏事
77:41
as being perpetuated by humans right now
现在正被人类持续做着
77:43
is going to be perpetuated by AI agents
会被 AI agents 一直延续下去
77:45
over the next 10 years.
在未来 10 年里。
77:46
Oh, right.
哦，对。
77:47
So think about the recursive scale
所以想想我们即将看到的
77:49
we're about to see of bad actors.
bad actors 的 recursive scale。
77:51
It's not just bad human beings.
这不只是人类里的坏人。
77:52
It's all the bad agents that are going to be attacking
而是所有那些即将攻击
77:55
the token flow.
token flow 的 bad agents。
77:57
And it's very hard if you're a researcher
而且，如果你是个研究人员
78:01
and an AI lab to reason about that problem
又在一家 AI lab，要推理这个问题会非常难
78:04
because the only data you have is how agents
因为你唯一能拿到的数据，就是 agents
78:06
you're training are going rogue.
你训练它们的时候会怎么失控。
78:09
But that's just a fraction of all the bad behavior
但那只是所有不良行为的一小部分
78:11
on the Internet that we're going to see.
是我们之后会在 Internet 上看到的。
78:12
And so what you need is defenders,
所以你需要的是防御者，
78:14
new sheriffs and town, which is in cowboy hats,
新的治安官来到镇上，戴着牛仔帽，
78:17
that can see all the bad behavior from AI agents
能看见 AI agents 所有坏行为的
78:21
across the ecosystem from different model labs
在整个 ecosystem 里，来自不同 model labs
78:24
and different post-trained deployments
以及不同的 post-trained deployments
78:25
and different developers and take all of that data
还有不同的开发者，然后把所有这些数据都拿到
78:28
and say we're going to build a shield
然后说，我们要建一面盾牌
78:29
for the entire token economy
为了整个 token economy
78:31
because without that, the amount of fraud
因为要是没有这个，那欺诈的规模
78:33
we're going to see of the $10 trillion in GMV
我们会在 10 万亿美元的 GMV 里看到的
78:36
and global GDP growth is like a huge percentage of that
而且全球 GDP 增长，很大一部分就来自那个
78:39
I think is going to be fraud abuse.
我觉得这会是 fraud abuse。
78:41
And we might never get there
而我们可能永远都到不了那一步
78:43
if people just don't trust tokens, right?
如果人们就是不相信 tokens，对吧？
78:46
And I don't think this infrastructure exists.
而且我觉得这套 infrastructure 根本不存在。
78:47
So you have your work cut out for you with a stripe
所以你在 Stripe 这边，任务可不轻松
78:49
but I don't think people have realized the scale
但我觉得人们还没意识到这个规模
78:51
at which agents, agent-agent fraud,
也就是 agents、agent-agent fraud 会达到的规模
78:54
like bad behavior perpetuated by AI agents
就像 AI agents 延续的那些不良行为
78:56
is about to hit us like a tsunami.
马上就要像海啸一样向我们袭来。
78:58
Yeah, I mean, there's a lot to dig into there.
是啊，我是说，这里面有很多可以深挖的东西。
79:00
I want to give you the last word.
我想把最后的话留给你。
79:02
We do have to wrap.
我们确实得收尾了。
79:04
What can people expect from open-roader interest?
大家能期待 open-roader interest 带来什么？
79:07
I mean, I think this is a really good way
我是说，我觉得这是一个非常好的方式
79:09
for us to accelerate go-to-market
来让我们加速 go-to-market
79:11
and to go up market more quickly.
并且更快地进军高端市场。
79:14
It's also, as I'm eloquently described,
而且，正如我被精彩地描述的那样，
79:18
this is a really clear, better together story here
这里有一个非常清晰的、强强联合的故事
79:22
when it comes to improving trust and safety
在提升 trust and safety 方面
79:25
and making it really easy to accept tokens
并且让接受 tokens 变得非常容易
79:28
and let people bring their own inference to your app.
并让人们可以把自己的 inference 带到你的 app 里。
79:31
And to help developers just build on top of inference
而且为了帮助开发者直接基于 inference 来构建
79:35
going forward, we have a really strong brand
接下来，我们有一个非常强大的品牌
79:40
with open router and we're keeping the brand.
有了 open router，而且我们会保留这个品牌。
79:43
So like open router as a product and the roadmap
所以，open router 作为一个产品，还有 roadmap
79:47
and the name and the brand is saying the same.
以及名字和品牌，表达的都是同一个意思。
79:50
And so what you should expect in the next six months
所以接下来六个月，你应该期待的是
79:55
is that most things will be like,
大多数事情都会像，
79:56
what we would have done had we been independent
我们如果一直独立的话会做的那样，
79:58
except everything will be moving faster.
只不过一切都会推进得更快。
80:01
And that's kind of like our near-term goal,
而这差不多就是我们的近期目标，
80:05
longer-term, hopefully I can comment on it.
更长远来说，希望我能聊聊这个。
80:09
So yeah, I'm glad I can.
所以，是的，我很高兴我能。
80:10
Now.
现在。
80:11
OK, well, we'll hopefully do a follow-up at some point.
好吧，希望我们之后某个时候能做一期后续。
80:14
But thank you for being so generous of your time
但谢谢你这么慷慨地抽出时间，
80:16
and congrats on the partnership.
也恭喜这次合作。
80:17
I mean, this is one of the most beautiful bromances
我是说，这是最美好的兄弟情之一，
80:20
I've seen in the AI.
我在 AI 领域见过的。
80:22
Starting from Stanford to here.
从 Stanford 开始，一路走到这里。
80:25
Lots more to do.
还有很多事要做。
80:26
Lots of sheriff, policing to do of the token economy.
token economy 还有大量执法和监管工作要做。
80:31
We need new sheriffs, for sure.
我们肯定需要新的执法者。
80:34
Awesome.
太棒了。
80:35
Thank you.
谢谢。
80:35
Thank you.
谢谢。