OpenAI如何为8亿周活跃用户构建服务：模型专业化与微调

我觉得很多听众可能还没意识到这个数字有多庞大，但这真的...太疯狂了。

don't think many people listening to this don't understand how big that is, but that is I mean It's crazy.

是啊。

Yeah.

按照达到8亿所用的时间来看，这简直可以载入史册了。

That's gonna be, like, actually historic for the time it's taken to get to 800,000,000.

确实是历史性的。

It's historic.

而且想想我们投入的时间和需要扩展的规模...

Also It's just like, yeah, the amount of time and just like how much we've had to scale up.

这相当于全球人口的十分之一。

That's a tenth of the globe.

对吧？

Right?

没错。

Yeah.

是啊。

Yeah.

全球10%的人口每周都在使用它

10% of the globe uses it every

每周。

Every week.

没错。

Yeah.

这太棒了。

That's great.

而且还在增长。

And it's growing.

而且还在增长。

And it's growing.

所以，在某个时刻，你知道，它会达到甚至更高的水平。

So, like, at some point, know, it'll hit like, you know, it'll go even higher than that.

所以，是的，显然，那里的覆盖范围是无与伦比的。

And so so, yeah, like, obviously, the reach there is is unmatched.

但除此之外，能够拥有一个平台让我们触及更广泛的受众。

But then also just, like, being able to have a platform where we can reach even more than just that.

比如，我们内部有时会讨论，我们的API最终能覆盖多少终端用户？

Like, one thing we we talk about internally sometimes is, like, what does our end user reach from the API?

实际上，它的覆盖范围确实非常非常广泛。

Like, it's actually it was, really, really it is really broad.

可能甚至——这很难说，因为ChatGPT增长太快了——但在某些时候，它肯定比ChatGPT还要大。

It might might even it's hard because ChatGPT is growing so quickly, but, like, at some points, it was definitely larger than than ChatGPT.

我们能够利用所有这些资源并获得我们想要的覆盖范围，我认为这真的很好。

And the fact that we're able to get tap in all of this and and and get the reach that we want, I think, is really good.

是的。

Yeah.

不过，确实，有时候也存在一些紧张关系。

But, yeah, I mean, there's definitely some tension sometimes.

我认为这个问题已经在几个地方出现了。

I think the I think it's come up in a couple of places.

我认为其中之一是在产品方面。

I think one of them is is on the product side.

就像你提到的，有时候会有竞争对手在我们的平台上进行开发。

So as you mentioned, you know, sometimes there are competitors kinda like building on our Yeah.

在我们的平台上。

On our Yeah.

这些平台上的开发者可能不太高兴如果ChatGPT推出与他们竞争的产品。

Platform who, you know, might not be happy if ChatGPT launches something that competes with them.

是的。

Yeah.

我的意思是，这就像云操作系统一样古老的故事。

I mean, that you know, that's the tale of as the old as the cloud operating systems or whatever.

所以，更像是ChatGPT是否担心竞争对手的问题。

So, like, that's, you know I think it's more like, does ChatGPT worry about the competitor Yeah.

你知道的，这类事情。

You know, type thing.

就像，你知道的，你在扶持一个竞争对手。

Like, you know, you enabling a competitor.

是的。

Yeah.

是的。

Yeah.

所以，我的意思是，有趣的是，我觉得并不特别担心。

So, I mean, the interesting thing is, like, I would say not particularly.

主要只是因为我们增长得太快了。

Mostly just because we've been growing so quickly.

这就像

This is like

我明白了。

that I get.

你知道，它现在就像一股强大的力量。

It's such a, you know, force right now.

是啊。

Yeah.

没错。

Yeah.

增长能解决太多太多不同的问题。

Growth solves so many so many different things.

而且，我们另一种思考方式是，大家都在围绕AGI进行建设，朝着AGI方向努力。

And, like and and the other way we think about it is, like, everyone's kind of building building around AGI, building towards AGI.

当然，这里难免会有一些重叠。

Of course, there's gonna be some overlap here.

但我想说，至少从我的立场看，我更多感受到来自客户端的压力，比如API用户本身。

But I would say, at least in my position, I feel more of this tension from the customer, like the API customers themselves.

就像在说，天啊，你们是不是要开发我正在做的这个东西？

It's like, my gosh, are you gonna build this thing that I'm working on?

这个故事和系统一样古老。

That story is as old as the system.

从来没有一个计算机平台不具备这种特性。好吧，我在这点上有些摇摆不定，想听听你的看法：历史上提供核心服务和API的问题在于可能会被绕过，对吧？

There's never not been a computer platform that have that So okay, so I kinda go back and forth on this one, wanna try one out on you, which is the problem historically with offering a core services and APIs, can get disintermediated, Right?

所以我可以在它上面构建，但用户并不知道，诸如此类。

And so I can build on top of it, but then, you know, the user doesn't know, like, whatever.

在云上构建，我就能绕过云服务商，然后可以切换到其他云平台等等。

Build on top of the cloud, I disintermediate from the cloud, and then I can switch to another cloud or whatever.

我突然意识到，用这些模型很难做到这一点，因为它们实在太难以抽象化了。

And it occurs to me that that's kind of hard to do with these models because the models are so hard to abstract away.

就像，它们就是难以驾驭。

Like, they're just they're just unruly.

对吧？

Right?

如果你试图用传统软件来驱动它们，它们就是不太容易管理好。

If you try to, like, have traditional software drive them, they just don't kind of manage very well.

所以我部分认为这几乎就像一种反去中介化技术，你必须直接将其暴露给用户。

So part of me thinks that it's almost like this, like, anti disintermediation technology that you kind of have to expose it to the to the user directly.

是的。

Yep.

这样讲得通吗？

Does that make sense?

所以我在想，即使我认为ChatGPT实际上只是试图将模型暴露给用户，API也只是试图将模型暴露给用户。

And so I'm wondering if, like so even if I think ChatGPT is really just trying to expose the model to the user, the API's just trying to expose the model to the user.

因此我认为几乎存在这样一种论点：如果真正的价值在于模型，那么你如何将其传递给用户其实并不重要，因为要让某人以计算机科学的经典方式将其抽象化——比如用户不知道自己正在使用该模型——将会非常困难。

So I think there's almost this argument that's like, if the real value is in the models, it doesn't really matter how you get it to them, because it's gonna be very tough for someone who's gonna to abstract it away in in in classic sense of computer science of, like, they don't know that they're using the model.

比如，你总是知道自己正在使用GPT-5。

Like, you always know you're using g p d five.

是的。

Yeah.

而且有趣的是，我认为整个行业也慢慢改变了对此的看法。

And and the interesting thing is I think, like, the entire industry kind of has slowly changed their mind around this too.

想想看，一开始我们大概觉得，哦，这些都可以互相替换。

Think, like, in the beginning, we kinda thought, oh, these are all gonna be interchangeable.

就像软件一样。

It's just like software.

是的。

Yeah.

没错。

Yeah.

正是如此。

Exactly.

所以很容易就能替换掉。

So the easy thing to swap out.

对。

Yeah.

我想我们在产品端也正在认识到这一点，比如GPT五的发布，以及四零版本，还有那么多人喜欢零三和四零版本等等。

I think we're learning this on the product side with, like, know, the GPT five launch and, like, four o and, like, how how so many people like o three and four o and and all of that.

我感觉到它变化的时候。

I felt that when it changed.

我就觉得，你对我没那么好了。

I'm like I'm like, you're not as nice to me.

是啊。

Yeah.

比如，我喜欢

Like, I like the

验证。

validation.

验证。

Validation.

对。

Yeah.

没错。

Yeah.

我也这么想。

Think think so.

所以。

So.

其实挺有趣的，因为我真的很喜欢GPT五号的个性，但我觉得我使用ChatGPT的方式非常功利性。

It's It's actually fun because I I really loved GPT five's personality, but I think it's like the way I used, you know, ChatGPT was very utilitarian.

哦，它主要是用于工作或获取信息之类的。

Oh, it's it's like, you know, mostly for work or just like information.

是啊。

Yeah.

我确实逐渐接受了，但说实话，当它改变时我感受到了那种不协调。

I've definitely come around just, you know, but, like, I actually felt the dissonance when it changed.

就像有种情感上的波动，但这几乎像是一种消除不协调的调解技术。

It's like there's this emotional thing that goes on, but it's almost like it's an anti dissonant mediation technology.

你某种程度上得向用户展示这一点。

You kinda have to show this to the user.

是的。

Yeah.

没错。

Yeah.

然后你会看到很多更成功的产品如Cursor直接这样做，尤其是那些用户想要更多控制权的编程产品。

And then you see a lot of more successful products like Cursor do this directly, especially the coding products where users want more control.

我们甚至看到一些更通用的消费产品也在这样做。

We've even seen some, like, you know, like, more general consumer products do this.

所以在消费者这一侧确实如此。

And so it's definitely been true on the on the consumer side.

有趣的是，我认为在API这一侧也同样如此。

The interesting thing is I think it's also been true on the API side.

这也是我认为不...

And that's also something that I think No.

正是。

Exactly.

乔纳森，这正是我想说的。

Jonathan, that's exactly what I'm saying.

所以，就像，

So, like,

有人可能会辩称，我可以通过API绕过你们。

the argument could be that I could use the API to disintermediate you.

但实际上，这种情况并未发生，因为在模型与人之间插入一层软件实在太难了。

But, like, you don't see that happening because it's so hard to put a layer of software between a model and a and a person.

你几乎必须直接暴露模型本身。

You almost have to expose the model.

是的。

Yes.

是的。

Yes.

而且

And

我认为，如果说有什么变化的话，那就是这些模型似乎在功能上越来越分化，各自擅长不同的领域和特定用例。

I think, if anything, I think the models are, like, almost like diverging in terms of, like, their what they're good at and, like, their specific use case.

而且我觉得这种情况会越来越多。

And I think there's gonna be more and more of this.

但基本上，令人意外的是，基于我们API开发的用户留存率出奇地高，尤其是在人们原以为可以随意切换的情况下。

But, yeah, basically, it's it's been surprisingly hard for or, like, the the retention of people building on our API is, like, surprisingly high, especially when people thought you could just kind of swap things around.

要知道，现在甚至有些工具能帮你实现这种切换。

You might have, you know, like, even tools that help you swap things around.

是的。

Yeah.

但模型本身的用户粘性确实令人惊讶。

But, yeah, the the stickiness of the of the of the model itself has been has been surprising.

你觉得这是因为用户与模型之间建立了某种关系，还是更多出于技术原因——比如我的评估体系是针对OpenAI设计的，且能保持准确性？

And do you think that is because of a relationship between the user and the model, or do you think it's more of a technical thing, which is like, my evals work for OpenAI and this, you know, and the correctness maintains?

没错。

Yeah.

是的。

Yeah.

我认为两者都有。

I think it's both.

所以我认为这里确实存在终端用户层面的因素，这是我们部分客户反馈的情况。

So I think there's there's definitely an end user piece here, which is what we've heard from from some of our customers.

比如，他们只是对模型本身越来越熟悉。

Like, they just get familiar with with the model itself.

但我也认为存在技术层面的因素，作为开发者——尤其是初创公司——你会深入研究这些模型，不断迭代优化，试图在你的特定应用框架内让它表现得更出色。

But I also think there's a technical piece, which is like the also, as a developer, especially with startups, you're like really going deep with these models and like really like iterating on it, trying to get get it really good within your particular harness.

你同时在迭代优化自己的应用框架。

You're iterating on your harness itself.

你会时不时为它添加不同的工具。

You're giving it different tools here and there.

所以你最终确实是在围绕这个模型打造产品。

And so you really do end up, like, building a product around the model.

因此，这里存在一个技术层面，当你持续围绕某个产品（比如GPT-5）进行开发时，实际上你是在围绕它构建更多功能，从而使你的产品与该模型形成独特的适配优势。

And so there is a technical piece where, you know, as you kind of keep building with a particular product like GPT five, you're actually, like, building more around it so that your product works uniquely well with that with that model.

所以我使用Cursor，主要用于各种事务，比如写博客——你知道我们是投资者——有时也用于编程，令人惊讶的是我在Cursor里会用到这么多模型。

So so I I use I use Cursor, and a lot of just for, like, a lot of stuff, like, writing blogs and, like, you know, know, we're investors, and I use it for sometimes for coding, and it's remarkable how many models I use in Cursor.

比如，我最常用的模型确实是GPT-5。

So, like, literally my go to model is g p d five.

我太爱GPT-5了。

I love g d five.

我觉得这太棒了，你懂吗？

I think it's a phenomenal, like, you know?

然后我用GPT-5的max模式来做规划，但你知道，我在Cursor里用的是tab补全模型，他们刚发布的新模型是用于一些基础的...一些功能。

And then, like, I use, like, max mode with g p d five for planning, and then but, you know, like, I mean, I the tab complete model that's in Cursor, and, like, you know, the the new model they just dropped is for, like, some basic you know, some stuff Yeah.

Composer不错。

Composer good.

是啊。

Yeah.

所以，你知道，我觉得

And so, like, you know And I think that

某种程度上也反映了这一点，因为就像是，每个特定用途都有对应的特定模型。

kinda reflects this too, because it's like, there's a particular model for each particular use Yeah.

对。

Yeah.

没错。

Yeah.

比如，我和很多用过新Composer模型的人聊过，它确实在快速处理上表现优异，速度超快。

Like, I've talked to a bunch of people who use the new Composer model, and it's just really good for fast It's super fast.

第一遍草稿确实如此。

First pass Exactly.

让你保持心流状态，之后如果需要更深入的思考，可以再切换到其他模型

To keep you in flow kinda thing, and then you kinda bubble out to another model if you want deeper thinking

我确实会直接使用GPT-5来帮我规划事项，它在这方面非常出色。

about I literally sit down on this GPT-five to help me plan something out, and is really good at that.

然后当我编码时，我会用快速聊天功能，接着切换到Composer，如果遇到什么棘手的问题或奇怪的bug之类的。

And then when I'm coding, I'm doing the quick chat thing, then I'll use Composer, and if there's whatever, there's some crazy bug or something like that.

还记得早期大家都认为只会有一个模型吗？连投资者都说我们永远不会投资模型公司，因为最终只会有一个AGI模型。但现实却是模型数量爆炸式增长。

So do you remember in the early days of all of this, where there's gonna be one model, even investors, we will never invest in a model company because there will only be one model and it's gonna be AGI, but the reality it feels like there's this massive proliferation of models.

就像你之前说的，它们承担着多种功能。

Like you said before, they're doing many things.

所以可能有两个问题，也许太直接或粗鲁，但第一个是：这对AGI意味着什么？

And so maybe two questions, maybe too blunt or too crass, but the first one is what does that mean to a for AGI?

第二个问题是：这对OpenAI意味着什么？

And the second was what does that mean for OpenAI?

这是否意味着最终会形成一个模型组合？

Like, does that mean that, like, you end up with a model portfolio?

你们会选择其中一部分吗？

Do you select a subset?

你认为未来这一切会被某个终极模型取代吗？

Do you think this all gets superseded by some god model in the future?

那么，这会如何发展呢？

Like, how does that play out?

因为这和大多数人的想法相悖。

Because it's it's against what most people thought.

大多数人曾认为最终会有一个全能的大型模型。

Most people thought this is all going towards one large model that does everything.

是啊。

Yeah.

我觉得这一切最疯狂的地方在于，大家的想法是如何随时间改变的。

I think the the crazy thing about all this is just, like, how everyone's thinking has just changed over time.

完全同意。

Totally.

我对此记忆犹新，最疯狂的是这并不久远。

Like the I I distinctly remember this, and and and the crazy thing is not that long ago.

大概就在两三年前吧。

It's just like three like two or three years ago.

是的。

Yeah.

我记得，即使是OpenAI，当时的想法也是会有一个统治一切的模型。

I remember, like, even with an open AI, the the thinking was that there would be like one model that rules them all.

对。

Yep.

这就好比，你为什么会...我是说，这某种程度上指向了微调API产品。

And it's like, why would you I like, mean, this kinda goes to the fine tuning API product.

就像是，你为什么要开发一个微调产品呢？

It's like, why would you even have a fine tuning product?

你为什么会想要在上面进行迭代呢？

Why would you even want to, like, iterate on it?

没错。

Yep.

确实。

Yep.

将会有一个模型能够涵盖一切。

There's gonna be this one model that just subsumes everything.

是啊。

Yeah.

没错。

Yeah.

这也是对通用人工智能形态最简化的看法。

And that was also kind of the that that is also like the most simplistic, like, view of what the what the AGI will look like.

而且，是的，自那以后情况已经完全改变了。

And and, yeah, it's like definitely completely changed since then.

但另一方面要记住的是，可能还会继续变化，甚至从我们现在所处的位置开始。

Think one and and but then the other thing to keep in mind is like, might continue to change, like, even from where we are today.

对。

Yeah.

但越来越清楚的是，我认为未来会有许多专业模型的存在空间。

But it's like becoming increasingly clear, I think, that there will be room for a bunch of specialized models.

很可能会出现其他类型的模型大量涌现的情况。

There will likely be a proliferation of other types of models.

我是说，你可以看到我们正在用Codex模型做这类事情

I mean, you see us do this with, like, the codex model

是啊。

Yeah.

完全同意。

Totally.

本身。

Itself.

我们有GPT-4.1、4.0、5.0等等这些版本。

We have, like, you know, have, GPT four one and, like, four o and, like, five and and and all of this.

所以我不认为有空间容纳所有这些。

And so I I don't think there's there's room for all for all this.

不过我觉得这未必是坏事。

I I don't think that's bad for what it's worth.

我认为，就像我们试图迈向通用人工智能一样，事情总是出人意料，市场也在不断演变，产品组合也因此而发展。

Like, anything, I think, you know, as we've tried to move towards AGI, things have just been very unexpected, and I think the market just evolves, and the product portfolio evolves because of that.

所以我完全不认为这是件坏事。

So I I don't think it's bad thing at all.

而我真正认为的是

What I do think it

你可以轻易地辩称这对OpenAI非常有利，对模型公司也非常有利。

means You could easily argue it's very good for OpenAI and very good for, like, the model companies to, like Yeah.

因为没有那种赢家通吃的集中化动态。

Because not have, like, you know, winner take all consolidated dynamics.

对吧？

Right?

我是说，你会拥有一个更健康的生态系统，能提供更多解决方案。

I mean, you just have a healthier ecosystem, lot more solutions you can provide a lot.

是的。

Yeah.

你知道吗？

You know?

是的。

Yeah.

随着生态系统的扩大，这通常是有益的。

And as as the ecosystem grows, it generally is helpful.

这是我们经常思考的一个问题：随着整个AI生态系统的壮大，OpenAI将从中受益匪浅。

Like, this is one thing we actually think about a lot too is is as the general, like, AI ecosystem grows, like, OpenAI just stands to benefit a lot from this.

这也是为什么我们的一些产品甚至开始向其他模型开放。

And and this is also why we've like, some of our products, we even started opening up to other models.

对吧？

Right?

比如，我们的产品现在允许你接入其他模型。

Like, our product now allows you to bring in other models.

是的。

Yeah.

我们认为这一切就像水涨船高，总体上对我们有利。

So all of this, we think it's like any any rising tide generally helps us here.

但随着我们进入一个将出现更多模型的世界，这就是我们投资于模型定制产品的原因，包括微调API、强化微调，并开放这些功能。

But, yeah, I think as we move into a world where there'll be a bunch more models, this is why we've kind of invested in our model customization product with fine tuning API, with the reinforcement fine tuning, opening that up as well.

这也是我们开源GPT OSS的部分原因，因为我们希望能够促进发展

It's also why part of why we open sourced GPT OSS as well, because we wanna be able to, you know, facilitate

我想超级

I wanna super.

我稍后想详细谈谈这个，因为开源确实非常有趣。

I wanna talk about that in in just a bit, because the open source is actually very interesting.

我 我

I I

实际上，我认为开源模型非常棒。

mean, actually, I thought the open source model was great.

是的。

Yeah.

但这显然是企业必须谨慎对待的事情。

But it's clearly something that a company has to be careful with.

是的。

Yep.

但在那之前，我想先聊聊微调API。

But before that, I wanna talk a little bit about the fine tuning API.

是这样的，我注意到你们正在向更复杂的应用方向发展，比如微调技术。某种程度上，这可以被解读为一种让步，承认存在产品特定数据和用例，通用模型无法满足这些需求。

So so so I I've I've noticed that you are moving towards kinda more sophisticated use of things like, you know, like fine tuning, which, you know, in a way, can read that as a bit of a capitulation that, like, you know, there is product specific data and there's product specific use cases that a general model won't do to your point.

对吧？

Right?

所以与模型扩散相反，你们选择了这条路。

So, like, as opposed to proliferation model, you do that.

看起来这些数据实际上非常有价值，对吧？

It seems like a lot of that data is actually very very valuable, right?

那么，在多大程度上存在一种互惠机制？比如你们开放产品数据接入微调的能力，同时也能从这些。

And so, you know, to what extent is there like interest in almost a tit for tat where you can like expose, you know, the ability to get product data into fine tuning, and then you also benefit from that data because the the vendors provide it to you versus, like, this is a 100%, you know, like, they keep their own data and there's kind of no interest in that.

因为在我看来这像是下一阶段的规模化发展，我们目前就处于这个阶段。

Because it feels to me like the next level of scaling, this is kind of where we're at.

所以我有点好奇如何

And so I'm just kinda curious how

是的。

Yeah.

我的意思是，或许退一步说，我们最初投资微调API的主要原因，一是人们对于能够更多定制模型有着巨大需求。

So I mean, maybe even like taking a step back, the the main reason why we even invested in a fine tuning API in the very beginning is, one, there's been huge demand from people to be able to customize the models a bit more.

这涉及到提示工程，而且我认为行业对此的看法也发生了变化。

It kinda goes into, like, prompt engineering and also, like, I think the industry's changed their mind on that as well.

就像，它已经进化了。

Like, it's evolved.

但第二点正是你所说的，就是企业拥有大量他们囤积的数据宝藏，他们希望以某种方式在这次AI浪潮中加以利用。

But the second thing is exactly what you said, which is the companies just have giant treasure troves of data that they are sitting on that they would like to utilize in some fashion in this AI wave.

你知道，最简单的方法就是把它放进某种向量里，比如用它做检索增强生成之类的。

And you can know, the simple thing is to put it in, you know, some like vector like, do rag with it or something.

是的。

Yeah.

但你也知道，如果他们有一个更专业的技术团队，他们会想看看如何利用这些数据来定制模型。

But there's also you know, if they have a more technical team, they do wanna see how they can use it to customize the models.

因此，这实际上是我们投资于此的主要原因。

And and so that is actually the main reason why we've invested in in this.

有趣的是，早在2223年左右，我们的微调服务功能过于有限，导致人们很难真正利用这些数据。

The the interesting thing was way back, kinda back in like 2223, our fine tuning offering was, I'd I'd say like too limited, so that it was very difficult for people to to tap into and use this data.

当时它就像是一个监督式微调（SFT），没错。

So it was just like an SF like a supervised fine tuning Yep.

API。

PI.

我们当时觉得，'你可以勉强用用'，但实际上它只适用于——说实话——基本上就是指令跟随的加强版。

Like, we're like, oh, you can kind of use it, but in practice, it really is only useful for, like like it's it's honestly just like instruction following plus plus.

比如稍微调整语气，本质上还是在指导它。

Like, kind of change the tone, you're just really like instructing it.

是的。

Yeah.

但我认为最近的关键突破在于强化微调模型，因为通过这种设置，我们现在允许您实际运行RL（强化学习），这虽然更复杂且难度更高，需要投入更多资源。

But I think the the big unlock that has happened recently is with the reinforcement fine tuning model, because with that setup, we're now letting you actual run actually run RL, which is more finicky and it's like harder, and and, you know, like, you need to invest more in it.

但它能让您更充分地利用数据。

But it allows you to leverage your data way more.

顺便说一句，这对我而言是个天真的问题——根据我自己投资组合的理解，似乎存在两种使用模式。

By the way, this is just a naive question for me, which is it feels from just my understanding from my own portfolio, it feels like there's two modalities of use.

一种是长期拥有大量数据宝藏，基于这些数据离线创建模型，然后部署使用。

One of them is I've got a treasure trove of data that I've had for a long time, I create my model on that treasure trove of data, and all that happens offline, and then I deploy that.

对。

Yep.

另一种则是产品需要实时使用的情况。

There's another one which is like, actually have the product being used in real time.

我有

I've got

一群用户。

a bunch of users.

是的。

Yeah.

而且实际上，我可以更贴近用户。

And like, I can actually get much closer to the user.

我可以进行A/B测试并决定使用哪些数据，这更像是一种近乎实时的操作。

I can kind of a b test and decide which data, and, like, it's kind of more of a near real time thing.

这是更侧重于产品方面还是... 是的。

Is is it like, is this focus on, like, more product stuff or more Yeah.

宝藏数据？

Treasure trove?

所以我们的梦想是

So the dream with the

微调API的目标是能够同时处理这两种情况。

fine tuning API was that we should be able to handle both.

对吧？

Right?

就像我们实际上已经有了这个数据流，然后我们还有完整的LoRa微调推理设置，理论上我们应该能够扩展到数百万个这样的微调模型，这通常是在线学习场景下会出现的情况

It's like it's like we actually had this stream, then we have this whole, like, LoRa setup with the fine tuning inference where we should just be able to scale to, like, millions and millions of of these fine tuned models, which would is usually what would happen if you have, like, this online learning

事情。

thing.

完全正确。

Exactly.

是的。

Yeah.

实际上，主要还是那种形式。

In practice, it's mostly been the the form.

对吧？

Right?

实际上，主要还是他们已创建的离线数据，或是他们与专家合作生成的数据，或是通过使用他们产品所产生的数据，这些数据能够在这里被利用。

In practice, it's mostly been, like, the offline data that they've, like, already created or they are creating with experts or or something and, like, using their product that they're that they're able to use here.

但关于强化微调API，我主要想说的是，它改变了原有范式，不再只是像SFT那样做小的渐进式改进，比如语气调整。

But the main thing I I was trying to say around the reinforcement fine tuning APIs, it kind of changes the paradigm away from just, like, small incremental improve like, tone improvements, which is what SFT did Yeah.

而是真正提升模型，使其在特定用例上可能达到苏打水准，这是你所了解的。

To actually improving the model to potentially soda level on a particular use case that you you know about.

对。

Yeah.

没错。

Yeah.

这正是人们真正开始使用强化微调API的地方，也是它获得越来越多采用的原因。

Like, that's where people have really started using the reinforcement fine tuning API, and that's why it's it's it's gotten more more more uptake.

因为如果讨论的内容更像是'我能让这个模型不以某种特定方式说话'，那就没那么吸引人了。

Because if if the discussion is less like, hey, I can make this model, you know, not, like, speak in a certain way better, it's less compelling.

但如果像是'在医疗保险编码或编码规划、代理规划等方面，能打造出世界上最优秀的...'

But if it's like, hey, for, like, you know, medical insurance coding or for, like, coding planning, agentic planning or something, can create the world's best Yeah.

利用你的数据集通过RFT训练的模型，那价值就大得多了

Model using your dataset with RFT, then it becomes a lot more

那么你们会不会，或者说你们有没有可能

And will you will you ever, like or maybe do you?

你们会不会想办法获取那些数据？

Will you ever, like, find ways to get access to that data?

比如说你

Like, you

知道，是的

know Yeah.

听着

So the Listen.

如果我手上有数据又想要便宜的GPU，我会跟你交换的。

If I if I had the data and I wanted cheap GPUs, I'd trade you for it.

这个嘛，我不确定。

Like, I don't know.

嗯，是的。

Like Yeah.

我是说，我们我们之前讨论过这个，我们实际上也在这里试行了一些定价方案，因为这些数据确实很有帮助，而且获取起来有点困难。

We I mean, we we we've talked about this, and we've actually been piloting some pricing here too where it's like because this data is, like, really helpful and and it's kinda hard to get.

如果你真的使用强化微调API进行构建，你实际上可以获得折扣推理，如果你愿意分享数据，甚至可能获得免费训练。

And if you actually build with a reinforcement fine tuning API, you can actually get discounted inference and potentially free training too if you're willing to share the data.

这总是，你知道的，取决于客户的选择。

It's always kind of, you know, it's up to the customer there.

但如果他们这样做，对我们有帮助，对客户也会有好处。

But if they do, it is helpful for us, and there'll be benefits for the customer as well.

那太棒了。

That's awesome.

好的。

Okay.

你说过对提示工程的看法已经改变了。

You said that views on prompt engineering have changed.

是的。

Yeah.

实际上，我之前并不知道这一点。

I wasn't actually, I wasn't aware of that.

其他所有事情我都不知道，这个我也不知道。

All the other things I wasn't aware of, this one I wasn't.

是的。

Yeah.

我是说，我认为2022年时主流观点是这样的。

I mean, I think the prevailing view this is back in 2022.

我记得当时和很多人交流过，他们基本上认为——这类似于单一模型AGI的观点——提示工程将不再重要，未来你甚至不需要考虑在上下文窗口中输入什么内容。

I remember I was talking to so many people, and they're basically I mean, this is similar to, like, the single model AGI view as well, which is like like, prompt engineering is just not gonna be a thing, and you're just not gonna have to think about what you're putting in the in the context window in the future.

就像，模型会足够智能，它自己就能明白。

Like, the model would just be good enough, and it'll just, like, know.

是的。

Yeah.

没错。

Yeah.

它会知道你需要做什么。

It'll know what what what you need to do.

是啊。

Yeah.

那不是

That's not

那根本不存在。

that's not a thing.

对。

Yeah.

嗯，就像，那个...我也不确定。

Well, like, that that, like I don't know.

也许人们忘了，但这确实是当时普遍的看法。

Maybe people forget it, but, like, that was, like, very common release.

我觉得这可能和扩展法则之类的有关。

I think it's, like, scaling laws or whatever.

在扩展法则中有某种机制，你会与模型心灵相通。

Something in scaling laws, and, like, you'll just mind meld with the model.

而且，就像，你只需要提示和遵循指令，效果就会好到你根本不需要刻意去做。

And, like, you just, like like, prompting and, like, instruction following will just will be so good that you won't really need to do it.

是的。

Yeah.

没错。

Yeah.

如果非要说的话，显然它一直是错的。

And if anything, like, yeah, it's, clearly been wrong.

而且，是的。

And Yeah.

对。

Yeah.

嗯。

Yeah.

但这很有趣，因为我认为我们现在所处的世界略有不同，模型在执行指令方面已经变得非常、非常出色，相比之前的GB三·五之类的模型。

But it is interesting because it's I think it's a slightly different world that we're in now where the models have gotten really, really good at instruction following relative to the, you know, like, GB three five or something.

是的。

Yeah.

但我认为现在的关键已不再是两年前我们所想的提示工程。

But I think the name of the game now is is less on, like, prompt engineering as we had thought about it two years ago.

更多的是上下文工程方面，比如你给它提供什么工具？

It's more of, like, it's like the context engineering side where it's like, what are the tools you give it?

它获取的是什么样的数据？

What is like the data that it pulls in?

它何时会获取正确的数据？

When does it pull in the right data?

嗯，这确实非常有趣。

Well, it's just very interesting.

我是说，它已经简化到几乎荒谬的简单程度了。

I mean, it's reduced it to like an almost absurdly simplistic level.

比如，关于RAG的奇怪之处在于，经典用法就像是用余弦相似度来选择要输入给超级智能的内容。

Like, the weird thing about rag, for example, the classic use of rag is like you're using like cosine similarity to choose something that you're gonna feed into a super intelligence.

是啊。

Yeah.

你知道吗，这简直像在侮辱人。

Know, you're like, I'm a rare It's like insulting.

我居然要根据该死的嵌入空间随机抓取内容。

I'm gonna like randomly grab this thing based on fucking embedding space.

这其实...当你想让超级智能做决策时，却要把智能推给检索环节...没错。

It doesn't really you know, I'm like and then I'm you know, when you want the super intelligence to decide the thing to do, and so it's like pushing intelligence in that retrieval Yep.

显然这种做法

Clearly is something that makes

很有道理。

a lot of sense.

这几乎就像是在推动

It's almost like pushing

将智能外推

the intelligence out in

某种程度上

a way.

没错

Exactly.

而且公平地说，我认为RAG的引入是在模型还处于前推理阶段的时候

And and and to be fair, I think, like, RAG was kind of introduced when the models were like it was like pre reasoning models.

所以当时的情况是，你只有一次机会来做这件事，而且它并不那么聪明。

So it was like, you only had to kinda like one shot to like do this, and it wasn't that smart.

但现在我们有了推理模型，我的意思是，如果你喜欢的话，我最喜欢的模型实际上是o three，因为它是最勤奋的模型之一。

But now that we do have the reasoning models, now that we have I mean, if you like, one of my favorite models is actually o three because it was like one of the most diligent models.

它有点

It kinda

比如o three。

like o three.

它它会执行所有这些工具调用，本质上就是智能体本身在尝试完成工具调用、RAG或其他类似操作，或者编写执行代码。

It It would just like do all these tool calls, and it's like really the the intelligence itself trying to like do the, you know, tool calls or reg or or anything like that or write the code to execute.

对。

Yep.

所以范式已经转变了，但正因如此，上下文工程、提示工程——你提供给模型的内容变得格外重要。

And so the the the paradigm has shifted there, but, yeah, because of that, think, like, context engineering, prompt engineering, what you put what you give the model is, like, extra important.

没错。

Yep.

是的。

Yep.

对。

Yeah.

好的。

Okay.

所以你们有横向的API，还有像ChatGPT这样的垂直产品。

So you have API so you have the API, which is horizontal, you've got ChatGPT, other products which are vertical.

我们甚至还没讨论到像素层面。

We haven't even talked about pixels.

这些都还只是语言层面的东西。

This is all just this just language.

智能体是一种新的模态吗？

Are agents a new modality?

还是说它是别的什么？

Is that something else?

比如，你知道的，像是代码本或者你这里说的模态具体指什么？

Like, you know, like a codex or What do you mean by modality here?

在我看来它们似乎兼具垂直和水平特性。

Like and they feel both vertical and horizontal to me in a way.

对我来说，ChatGPT就是一个产品。

Like, to me, ChatGPT is a product.

对吧？

Right?

它就像一个产品，连我妈妈都在用。

It's like it's a product and, like, my mom uses it.

对吧？

Right?

是的。

Yep.

而API是开发者的工具。

And an API is a dev thing.

你把它交给开发者，而CLI在我看来介于两者之间。

You kinda give it to a developer, and, like, a CLI is kinda somewhere in between to me.

这就像是，它是个产品吗？

It's like, is it a product?

它是横向的吗？

Is it, like, it is horizontal?

嗯，是的。

Like Yeah.

内部是如何处理它的？

How how is it handled internally?

是一个完全独立的团队负责

Is it a totally separate team that

代理相关的工作吗？

does agents?

或者不。

Or No.

是的，这很有趣，因为我觉得你刚才的描述方式几乎把代理看作是一个单一的概念，好像它可能有自己特定的...也许

So it's yeah, it's interesting because like, I I think the way that I I the way that you frame it just now almost seemed like agents was like this, like, singular concept that like, you know, might or like might have its own particular Maybe

更好的问题是你眼中的代理是什么？

a better question is what is an agent to you?

是的。

Yeah.

对。

Yeah.

对。

Yeah.

对。

Yeah.

对。

Yeah.

对。

Yeah.

确立共同语言对这次对话很重要。

Getting a language is, like, important for this conversation.

对。

Yeah.

所以我...我不知道。

So I I I don't know.

其实我都不确定分享这个是否有帮助，但我对智能体的基本看法是——它是一种能代表你采取行动、并能长期运作的人工智能。

I actually don't even know if it'd be helpful for me to share, but my my general take on agents is it's it's a it's an it's an AI that will take actions on your behalf that can work over long time horizons.

好的。

Okay.

我认为这是更为普遍的预监管观点。

And I think that's the that's the more pretty general Pregulatory.

是的。

Yeah.

是的。

Yeah.

确实如此。

Definitely.

不过，如果你这样想的话，确实。

But, like, if you think about it that way yeah.

我的意思是，也许这就是你所说的模态，但这只是使用AI的一种方式。

I mean, maybe this is what you mean by modality, but it is just a, like, way of, like, using AI.

我想它可以被视为一种模态，但我们并不认为它是独立于AI之外的单独事物。

And it is a I guess it could be viewed as a modality, but we don't view it as, like, a separate thing, separate from AI.

让我来说

Let let me let let

让我试着给你解释一下这个问题是从何而来的。

let me just try and kinda, you know, give you a sense of where this question's coming from.

就像，我知道如何打造一个产品，我们也知道如何为产品开拓市场，我们了解将它们转化为平台意味着什么，我们做这行已经很久了，对吧？

Like, I know how to build a product, like, and we know how to do go to market for products, we know how to do, like, you know, we know the implications of turning them into platforms, like, it's just we've been doing this for a very long time, right?

我们也知道如何为API做同样的事情，对吧？

We know how to do the same thing for APIs, right?

我们了解如何计费，明白人们基于它构建时产生的张力等等。

We know how to do billing, we know, like, the tension of, people build on top of it, and all of that stuff.

而我一直在尝试的——这可能只是个人探索。

And like, what I've been trying to and this is just maybe a personal inquiry.

对我来说还不明确的是，对于一个代理来说，它究竟属于这两个阵营中的哪一个，是更像产品阵营？

It's just not clear for me for an agent if you if it if it sits in one of those two camps, is it more like the product camp?

还是更像...因为它有点两者兼有。

Is it more like the because it's kinda both.

比如，我完全可以给你提供编码。

Like, I could like literally give you coding.

对。

Yeah.

对。

Yeah.

然后作为用户，你只需要和它对话，或者我可以把它嵌入到我的应用里。

And like, as a user, and then you just talk to it, or I could like build in a way kind of embed it in, like, my app.

所以这对你来说意味着，比如定价策略和生态系统影响的问题。

And so, like but then that means something to you as far as, like, you know, how do you price it and what does it mean for ecosystem?

就像，是的。

Like Yeah.

举个例子，如果我成立一家公司完全围绕Codecs来开发，你觉得可行吗？

Like, for example, like, would you be fine if I started a company and just, like, built it around Codecs?

这能成立吗？

Is that a thing?

创办一家公司并围绕它发展

Starting a company and building it around

Codex？

Codex?

是的。

Yeah.

实际上我认为那会很棒。

I actually think that would be great.

就像我们发布了Codex SDK，我们希望人们能够基于它进行开发和创新。

Like, it's a we release the Codex SDK, we want people to be able build it and hack it.

没错。

Yeah.

其实，我想你可能指的就是这个——这也是OpenAI的一个独特之处，反映了它的运营方式：归根结底，OpenAI是一家AGI（通用人工智能）公司。

Actually, I think this might be what you're getting at, which is and this is like a kind of a unique thing about OpenAI, and it kinda reflects on how how it's run, which is at the end, like, at the end of the day, OpenAI is like an AGI company.

它本质上是一家智能公司。

It's like an intelligence company.

是的。

Yeah.

当然。

For sure.

因此，智能体只是这种智能得以体现的一种方式。

And so agents are just like one way in which this intelligence kind of be manifested.

没错。

Yeah.

所以我认为我们内部的实际思考方式是，我们所有的不同产品线——Sora、Codex、API、ChatGPT——都只是部署的不同界面和不同方式

And so the way that I'd say we actually think about internally is all of our different product lines, Sora, Codex, API, ChatGPT, are just different interfaces and different ways So of deploying

你们并没有真正

you don't really

所以并不存在像这样专门考虑智能体的独立团队。

So there's no, like, single teams like this is, you know, like, thinking about agents.

我认为它更多是以这种方式体现：每个产品领域都在思考如何将这种智能转化为实际可呈现的形式，使智能体行为更具可行性。

I would say the way that it it it manifests itself more is, like, each product area thinks about, like, what is, you know this intelligence is actually turning into a form where, like, you can actually agentic behavior is more possible.

我明白了。

I see.

在ChatGPT这样的第一方产品中，这会是什么样子？

What would that look like in a first party product like ChatGPT?

是的。

Yeah.

那会是什么样子？

What would that look like?

这正是Codex最终成为独立产品的原因。

That's that's actually why Codex ended up becoming its own product.

在编程类产品中，它会是什么样子？

Like, what would it look like in a coding style product?

是的。

Yeah.

我们探索过，ChatGPT在那里某种程度上可行，但实际上CLI界面更有意义。

Like, we explored it, and ChatGPT, like, kind of worked there, but, like, actually, the CLI interface actually makes a lot more sense.

这是部署它的另一种界面。

That's another interface to deploy it.

是的。

Yeah.

我明白了。

I see.

然后如果你看API本身，这又是部署它的另一种方式。

And then if you look about the API itself, it's like, this is another interface to deploy it.

它是以一种略微不同的方式思考，因为采用的是开发者优先的思维模式。

It's it's thinking about it in a slightly different way because it's a developer first mindset.

我们正在帮助其他人构建它。

We're helping other people build it.

定价略有不同。

The pricing is slightly different.

但这些都是这种核心智能的不同表现形式，也就是亚洲行为。

But it's all these, like, different manifestations of this core, like, intelligence that is the the the Asian behavior.

是啊。

Yeah.

整个经济体系中有如此大比例本质上只是代币洗钱，这太惊人了。

It is so remarkable how much of this entire economy is basically just token laundering.

字面意义上

Literally

就像我能做的任何事都是为了把英语或自然语言输入进去，然后让智能输出结果。

like anything I can do to get like like English in or like a natural language in and then like, you know, the intelligence out.

对。

Yeah.

我是说，这这这是因为这些系统对分层处理有很强的抵抗力。

And I mean, and and and it's because these are so resistant to layering.

要把语言分层处理实在太困难了。

It's so hard to layer language out.

就像，你知道的，没错。

Like, you know, like Yep.

是的。

Yep.

我甚至可以用编解码器轻松实现，相当容易。

I could even do it easily pretty easily with, like, codecs.

我可以直接把它当作程序的一个组件使用，基本上就是将智能进行'洗白'处理

I could just, like, use it, you know, as a as a component of a of a program and just, you know, basically launder intelligence to

它。

it.

当然，你知道，这样做是要收费的。

I mean, of course, you know, I'd be charged to do that.

所以

So

对。

Yes.

实际上，根据我观察过这么多不同产品的发布——包括智能体发布、你们定义的产品、各种API以及基于这些的产品——我的看法是，它们与我们习惯的东西其实大不相同。

It I I actually my my view of this and having seen now so many kind of launches of different products, I've seen agent launches and the definition that you have, I've definitely seen APIs, and I've seen products on these is like, they're actually quite different than, like, what we're used to.

比如，成本结构不同，防御性也不同。

Like, the COGS is different, the defensibility is different.

是的。

Yep.

哦，所以我们某种程度上是在重写它。

Like, oh, so we're kind of rewriting it.

所以这有点像，你知道的，你来自定价背景。

And so it's kind of like, you know, you came from a kind of pricing background.

我是说，你正在研究定价的演示模型。

I mean, you're working on demo model for pricing.

现在你有了API。

Now you have the API.

所以我真的很想听听你的想法，我是说，你是如何逐步调整思路的，以及你如何为这些智能访问定价，毕竟你无法预知会有多少人使用它。

So I just love your thoughts on like I mean, how how how have you evolved your thinking and how do you price these, you know, access to intelligence where, you know, you don't know how many people gonna use it.

几乎可以肯定的是，这会是基于使用量的计费方式，而非其他形式。

It's almost certainly usage based billing, not something else.

你能稍微谈谈关于这些东西的定价理念吗？

Like, can you can you talk just a bit about, like, philosophy around pricing on these things?

产品和API的定价方式是否不同？

Is it different for product versus API?

比如

Like

是的。

Yeah.

我认为这里的实际情况是，它也随着时间的推移而演变。

I think the the the honest truth here is, like, it's evolved over time as well.

而且，说实话，我们在API上采用基于使用量的定价，最根本的原因是这最接近实际使用情况。

And and, like, I actually think the simplest like, the reason why we've done usage based pricing on the API honestly is because it's been the likes it it's closest to how it's actually being used.

所以我们最初就是这样开始的。

And so that's kinda how we how we started.

实际上我认为API的基于使用量的定价出人意料地一直很稳固。

I actually think usage based pricing on the API has has has like surprisingly held strong.

实际上，我认为这可能是我们会长期坚持的做法，主要是因为

And like, I actually think this might be something that we'll keep doing for for quite a long time, mostly because

在我不明白你们为何不采用按使用量计费的背景下

In the context of how I don't know how you don't do usage based.

是的。

Yeah.

是的。

Yeah.

是的。

Yeah.

是的。

Yeah.

就是不知道该怎么... 是的。

Just don't know how that Yeah.

然后...然后...然后还有我们如何定价的策略。

And then and then and then there's also the strategy of, like, how we price it.

在内部，我们做的一件事就是始终确保我们的按使用量定价实际上是从成本加成的角度来制定的。

And and internally, one thing we do is is we always make sure that we actually price our our our usage based pricing from a, like, cost plus perspective.

就像，我们实际上只是试图确保我们在...完全合理。

Like, we're we're actually just, like, trying to make sure that we're being responsible from a Totally makes sense.

从利润率的角度来看。

From a margin perspective.

顺便说一下，这是

By the way, this is

这是整个行业的一个巨大转变，因为，我记得从本地部署到订阅制的转变。

a huge shift in the industry in general just because, like, I remember the shift from on prem to to recurring.

是的。

Yeah.

那是个非常非常大的变化。

That was a big, big deal.

就像，这催生了Zoro。

Like, that created Zoro.

它催生了整个公司。

Like, it, created whole companies.

这就像里面整本书的内容。

It's like whole books on in there.

比如一大堆关于如何进行这种变革的顾问，对吧。

Like, a bunch of consultants on how you do this to change, like Yeah.

你知道吗？

You know?

而且我认为向按使用量付费的转变更加重大。

And, like, I think the shift to to usage is as bigger, bigger.

这还是个非常棘手的技术难题。

And it's also even a really hard technical problem.

是啊。

Yeah.

我甚至无法想象8亿，哇，你们是怎么搭建的？

I can't even imagine 800,000,000 wow, like, do you build?

是啊。

Yeah.

8亿用户的情况会简单些，因为是订阅制而非按用户计费。

800,000,000 wow is a little easier because it's not user based pricing, it's subscription.

所以听起来容易多了。

So it's like, that sounds like way way easier.

确实容易很多。不过API上还是有大量用户，我们得处理所有账单相关的事务。

That's way way But I mean, there's still like a like a lot of users on the API that we need to, like, you know, manage all the billing side.

有些超额使用的情况或者

There's some, like, overages or stuff you've

需要处理的问题？

gotta deal with on that?

你说的超额是指什么？

Or What do you mean by overages?

我也不太清楚。

Like, I don't know.