Jev：为生产环境打造System One模型，而非上帝——与TypeSafe AI首席执行官Diogo Almeida对谈

虽然是合成的，但我们有一群品味很好、很用心的人在仔细看这些数据，指出哪里有问题，然后回去重新生成。

Synthetic, but we have a people with a lot of taste and a lot of care looking at looking at these, articulating what's wrong, going back, regenerating.

这就是现在优秀数据人才的样子吗？

Is is that what a good data person is these days?

让我

Let

试着想想怎么解释，这非常复杂，而且我实际上是用一个比这期播客还要长的演讲来培训数据人员的。

me try to figure figure out how to like, it's it's super complicated, and, like, I literally onboard the data people with a talk that I assume is longer than this podcast will end up being.

所以我尽量讲个大概。

So I will try to say, like, the high level of it.

首先，我们做的不是那种人们常说的合成数据，好吧，其实第一条是零。

So number one, we don't do the kind of synthetic data that people kind well, I'll do a actually, number is zero.

数据和合成数据取决于你的任务。

Data and synthetic data depends on your task.

比如，你的数据形态、任务形态都会改变数据。

Like, the shape of your data, the shape of your task changes the data.

比如，RLVR的数据有点像环境。

Like, RLVR's data is kind of environments.

对。

Yes.

RLHF的数据就是人类反馈，你懂吧？

RLHF's is the human feedback, you know?

每个任务都有自己独特的数据，我们当然也有自己独特的数据。

Each task has its own unique kind of data, and we, of course, have our own unique kind of data.

所以第一点，我们有这个。

So number one, we have that.

第二点，我们不想用用户的数据来训练，即使我们可以，对吧，比如我们现在可以随便要求条款，但我不确定会不会有影响。

Number two, the thing I the reason why we don't want to train on our users' data even if we could, right, like we could probably ask for any terms right now and it'll we I don't know if it would make a difference.

我们真的不想要那样，因为不管怎样，真实世界的数据有太多偏见。

We truly don't want that, because no matter what, the real world data has so much bias.

就像有个幂律分布，人们总是问同样的问题，最后你会过度拟合，然后分裂，诸如此类。

There's, like, a power law of, like, people, like, asking the same things where you'll end up, like, overfitting to it and, like, fracturing to it and all of that.

第二点，我们的目标是实现一个完整的科幻未来，多年以后，这些模型将成为通用的基础设施，一层一层又一层，深入到人们无法想象的底层。

And number two, we are, like, aiming for, like, a complete sci fi future years from now where, like, these models are going to be, like, the general infrastructure layers and layers and layers and deep down the stack to to, like, things people can't even imagine.

我想把我们的模型想象成，就像你知道的，UDP之于LLM，TCP之于我们的模型。

Like, I would like to think of our model, like, kind of, like, you'd like, you know, UDP as LLMs and TCP as our models.

在那之上可以构建各种各样的东西，我们需要能够搞定那些未来感十足的应用场景，这样软件开发者才能真正构建出那些未来的东西。

All sorts of stuff can be built on top of that, and we need to be able to nail those futuristic use cases such that software developers can actually build that futuristic stuff.

要做到这一点，即使我们拥有现在所有的数据，我们也只会过度拟合于现在，然后就不起作用了。

And the way to do that is even if we had all of the data of the present, we would just overfit to the present and then it wouldn't work.

我们需要的，感觉就像他们是艺术家。

What we need is to like it almost feels like they're they're artists.

你知道吗？

You know?

他们研究这个认知核心。

They study this cognitive core.

我们的认知核心比任何其他人都要平滑得多。

Our cognitive core is, like, way less jagged than anyone else's.

然后他们找出那些不规整的地方，再用精准的方式去处理，虽然你永远不可能做到完美。

And then they find the jaggednesses, and then they address them surgically in a way that and you can never perfectly do this.

但他们处理的方式，是覆盖所有可能的维度，也就是通用的情况。

But they do it in such a way that it addresses it in every single possible, like, dimension The general case

而不是针对某个特定情况。

rather than the specific case.

而且每次这么做都需要很高的智能。

And, like, that requires a lot of intelligence every time.

好的。

Okay.

我们之前提到过，你有点批评我的想法受RLVR影响，这确实很公平。

So we mentioned a little bit you you you sort of criticized my thinking as RLVR influence, which is, like, very fair.

那我们来说说RLCD吧。哦。

Let let us let us actually mention RLCD Oh.

你在这方面显然有一些独门秘方。

Which obviously you have some secret sauces.

据我所知，你其实从来没有发表过相关的论文或者类似的东西。

To my knowledge, you've never actually published a paper or anything like that on it.

没有。

No.

没有。

No.

还没有。

Not yet.

但是，大家应该从这里面得到些什么呢？

But, like, what what should people get from this?

就是说，你怎么让大家相信你不是在编造一些术语来显得自己很酷？

Like, what, you give people some confidence that you're just not just making up jargon for the sake of sounding cool?

对我来说，有一点就是校准。

Like, one thing for me is, like, calibration.

我觉得这个我理解得比较清楚，因为我们在播客里已经聊过了。

I do think is, to me, like, well understood because we've covered it in on the podcast.

但你说RLCD的时候，我不太明白你指的是什么，跟人们熟悉的东西有什么区别。

But I I don't know what you mean when you say RLCD versus what people are familiar with.

好。

Good.

这是个很好的问题。

It's a great question.

是的。

Yes.

实际上，我要提一个相关的问题。

And actually, I will give a related question.

好。

K.

什么是RLHF？

What is RLHF?

好的。

Okay.

其实，RLHF这个词有好几种不同的意思。

And actually, RLHF means multiple different things.

比如，最初的RLHF，我记得好像是保罗·克里斯蒂亚诺教机器人做后空翻之类的。

Like, there's the RLHF of the original I think it was like Paul Cristiano teaching a robot to backflip or something like that.

是不是有个什么？就是那个吗？

Wasn't there something Was that it?

那是最初的，我……

That that was the original I

不是PPO那篇论文，但我不太确定。

don't the PPO paper, but I don't I don't know.

所以，如果我没记错的话，PPO不一定来自你的反馈。

And so PPO is not necessarily from you and feedback if I recall.

好的。

Okay.

确实。

That's true.

但我记得好像是OpenAI的对齐工作，能教一些很难明确指定的输出，比如后空翻。

But I I believe it was like an OpenAI alignment work that could teach a hard to specify outputs like a backflip.

我不太确定。

I'm not a 100% sure.

然后还有那个学习总结的工作。

And then there was actually learning to summarize.

你知道，这是由很多参与Instruct和共同撰写指令跟随论文的团队成员做的，他们是在语言模型上做PPO。

You know, this was work, by a bunch of the the team that helped with, instruct and and coauthored, the instruction following paper, which was teaching, doing PPO on language models.

这是，抱歉。

This is the, sorry.

我正试着，对。

I'm trying to trying Yeah.

试着操作这个东西。

Trying to manipulate this this thing.

这是2017年。

This is 2017.

我不太确定，但看起来差不多就是这样。

I'm not a 100% sure, but, like, that looks quite right.

如果里面有机器人做后空翻之类的，那可能就是它了。

If it has, like, a robot doing backflips or something like that, that might be it.

是的。

Yes.

好的。

Okay.

不错。

Cool.

看来我猜对了。

I guess I I got it right.

太棒了。

Hell yeah.

这就对了。

There you go.

就是那个。

That's the one.

所以这个想法有点像是，你能不能用它做一些定义不太明确的事情？

So the idea was kinda can you do, like, ill specified things with it?

所以那是第一版。

So that's, like, version one.

第二版是OpenAI做的那个学习总结的工作，实际上就是在语言模型上用PPO来做一些定义不太明确的事情。

Version two was the learning to summarize work that, like, OpenAI did, which is actually, like, PPO on language models to do something somewhat ill specified.

这是人们常说的RLHF的另一个例子，那个我没有参与合著。

This is, like, another thing that people refer to as RLHF, which I did not coauthor.

哦，Diogo在那儿。

Oh, Diogo's there.

好。

Cool.

太好了。

Hell, yeah.

还有Refford。

And Refford.

给Alec和Ryan点个赞。

Shout outs to Alec and Ryan.

我很喜欢他们。

Love them.

但我说的RLHF，其实是那个，天哪。

But, the thing that I refer to RLHF is the, oh, man.

你对那个有意见吗？

Of I'll You have comments on that.

我对那篇论文有意见，但我们现在已经跑题太远了。

I have comments on paper, but, like, we're so many tangents deep.

所以我真正说的RLHF，是指指令跟随这个任务。

So the thing that really got, to to meet the thing that I'm calling to RLHF is the task of instruction following.

重点不是PPO。

It's not about the PPO.

那部分不重要。

That part doesn't matter.

关键是设定一个北极星，指明这是有价值的方向。

It's about, like, setting a north star of this is a valuable direction.

这有点像“最苦涩的教训”那个北极星。

It's kind of like the bitterest lesson north star.

对我们来说，RLCD 就是这项新任务，我不觉得它是行话。

And for us, RLCD is this new task, and it is not I I don't see it as jargon.

我尽量精确地表达。

Like, I I try to communicate with precision.

就是说，嘿，这里还有另一个北极星，就像 DPO 及其所有变体一样，虽然没用那篇论文里的算法，但本质上也是在搞 RLHF。

It's just that, hey, here's another North Star, just like DPO and all of its, like, you know, descendants also do RLHF despite not using the algorithm in that paper.

所以明确地指出这个北极星，就是让 AI 可编程，这是我特别认同的一个词，把人类从循环中移除，因为 RLHF 就是为了这个而调优的，这样你就能自动化一切。

And so clear clearly stating the North Star is, being program programmable AI is is is one, word that I really catch on to, removing the human in the loop from, because RLHF is tuning for this so that you can automate everything.

是的。

Yes.

那关于这个北极星的核心论点，我还有什么遗漏的吗？

And did I miss anything else in the in the thesis of, like, what the North Star is?

对，你说得没错。

There is that is is right.

我说话确实有点太讲究细节了。

I'm overly nuanced in my communication.

但有一个细节就是，我们得务实一点。

The one nuance is that we need to be practical.

我们要清楚语言模型真正擅长什么。

We need to be aware of what language models can do really well.

你知道，就是AI能做什么。

You know, like what AI can do.

比如，有些程序化的东西可能特别酷，但如果技术还没准备好，那它没出现在世界上也不算悲剧。

Like, there could be programmatic types that are like sick AF, But if you if the technology is not ready for it, it to it's not a tragedy if that's not out in the world.

但对我来说，Jev之前的世界才是悲剧，虽然这么说听起来有点自大。

But to me, like, the pre Jev world was a tragedy because it it sounds arrogant.

不是。

No.

不是。

No.

不是。

No.

我非常相信你。

I I strongly believe you.

好。

Cool.

听起来有点自大，但我早在开公司之前就有这种感觉了。

It sounds arrogant, but like I felt this way since long before I even had a company.

我可以作证，是的。

I can vouch that Yes.

我谈这个已经很久了。

I've talking about this for long.

大概三年了吧。

For like three years.

这件事我已经说了很久了，我一直这么说，是因为我原本以为会更容易。

I've been talking about this for so long, and I've been saying it because I thought it would have been easier.

他们说他们做事情不是因为容易，而是因为他们以为容易。

They say they do not do things because they it they're easy dates because they thought it was easy.

对。

So Yeah.

没错。

Exactly.

没错。

Exactly.

没错。

Exactly.

我原本以为整个项目一周就能搞定，结果我错得离谱。

I I thought it this whole project would take a week, and I was unbelievably wrong.

所以，我要向OpenAI的每个人道歉，我当时觉得自己马上就要解决这个问题了。

So I am so sorry to everyone at OpenAI that I thought I was like, man, I'm solving this right now.

但是，我觉得悲剧的是，嗯，我觉得过度承诺、交付不足也很悲剧，而AI在这个方面特别极端。

But, like, I think that the tragic thing is when well, I think overpromise, underdeliver is tragic too, and, like, AI is super extreme on that axis.

我觉得RLVR是主要的，嗯，RLVR和RLHF都是这个问题的极端例子。

And I think RLVR is, the main well, both RLVR and RLHF are extreme perpetrators of this.

但对我来说，就是，那里有太多潜力了。

But like, it to me, it's like, it's just there's just so much potential there.

比如，AI显然非常聪明。

Like, AI is clearly so smart.

我聪明。

I smart.

我喜欢在我的演讲中这么说。

I love this in my talks.

你知道，当我问人们，AI怎么能聪明得这么离谱？

You know, when I ask people, like, how can AI be so unbelievably smart?

我们怎么能解决数学上的千年难题，却连最基本的活儿都自动化不了呢？

How can we, like, solve millennium prize problems in math, but still not automate even the most basics of works?

就是那种特别基础的重复性工作，根本不需要特别聪明的人来做。

Like like, really basic rote stuff that, like, you know, it it doesn't take, like, extremely smart people to do this.

这活儿干着也没啥意思。

It's not a satisfying job.

这些人本来可以做别的事情，但就是因为我们现在还自动化不了这些特别基础、特别没劲的工作，所以还得靠他们来做。我们明明有这么一个超级强大的自动化引擎，但就是没法把它插到这些有经济价值的工作上去。

Like, there's other things these people could be doing, but yet we need them to do, like, this bait like, super basic, none unsatisfying stuff because, you know, like, we can't automate it yet, but we have this, like, supercharged engine of automation that just does not have, like, the right plugs and stuff to plug into all of this economically valuable work.

而且，如果整个TypeSafe公司消失了，可能得花一两年时间，大家才能真正反应过来。

And, you know, like, if the whole company of TypeSafe disappears, like, maybe it'll take, like, a year or two for people to, like, truly catch up.

其实我也不确定要多久。

I actually don't know how long it'll take.

如果模型质量真的很重要的话，那我们在很长一段时间内都会处于一个非常有利的位置。

If if model quality matters, then we are gonna be in a very good position for a long time.

但这事儿，就这么定了。

But it it like, it it it's done.

这已经改变了技术历史的走向。

Like, there like, this has changed the path of, like, technological history.

而且，我们整个领域都会去探索这个空间。

And, like, we will be exploring that space as a field.

我完全同意这一点。

I I think I I definitely agree with that.

你创造了新的可能性。

You've created possibilities.

所以我觉得，如果让我用通俗的话来说，让大家也能理解，你不应该把TypeSafe和Jev的成功仅仅看作是一种新的模型类型。

So I think, you know, if I can paraphrase so that people can also understand, you should not take the success of TypeSafe and Jev as just like, well, you know, that is a new new model type.

现在事情结束了。

Now we're done.

我们回到正轨。

We go back to business.

不是这样的。

Like, no.

其实，还有五种其他模型类型值得探索，让百花齐放。

Like, actually, there's there are like five other model types that you should be exploring and like let a thousand flowers bloom.

没错。

Absolutely.

比如

Like

其中一些你可能会看到。

And some of which you will probably also see.

是的。

Yes.

早期互联网的活力。

Early Internet energy.

我觉得又回到了技术乌托邦。

I think it's back to tech utopia.

你懂吗？

You know?

不再是那种‘哎呀’的感觉了。

It's no longer like, oh, man.

比如，有时候我的编程助手能干活，但所有最好的都被内部藏起来了。

Like, sometimes my coding agents work, but the all of the best ones are hoarded internally.

感觉创造又回来了，你知道吗？

It's like creation is back on the menu, you know?

虽然这世界会变得很疯狂，但，你懂的，系好安全带吧。

Though it's gonna be a wild ass world, and, you know, buckle up.

我对此超级兴奋。

It's I'm so so jazzed about that.

我的意思是，现在你有资金和动力去做你设想的一切，我觉得看到你终于把这些东西展示给世界，真的很让人欣慰，毕竟你说了这么久。

I mean and now you have the funding and the momentum to do whatever you you envision there, which I which I think is, like, very gratifying to see you have after, you know, so long of, of saying these things when I actually actually show the world.

我知道。

I know.

我觉得一直吊着大家的胃口真是件有趣的事。

I just such a such an interesting thing to be a tease the whole time.

我的演讲感觉就像是个悬念，因为我没有说自动化会怎么实现。

Like my talk, like felt like it was a cliffhanger because I didn't say how the automation would occur.

肖恩看了我们的宣言，他说这些部分有点模糊。

Sean reviewed our manifesto and he's like, it's a little bit vague in these parts.

你知道，第一步是什么？

And you know, like, what's step one?

智能体是什么？

What is the intelligence I for

你当时说，对，模型快来了。

see for model and you were like, yeah, model coming.

其实，我主要反对的是“可组合”这个词，但比尔·普拉达纳格特很棒。

And like, well, just, I mainly objected to the word composable, but Bill Pradnagh got is fantastic.

谢谢。

Thank you.

我们真的围绕这个团结起来了。

I I we we've really rallied around that.

我希望我们不像有些公司那样完全是个邪教，但我们确实对自己的工作充满热情，而且我的品牌就是务实，我们所有人都超级无敌务实。

I'd like to think we're not entirely a cult like some companies are, but, like, we are, like, jazzed about what what are doing and, like, we are like, my brand is being practical and, like, we are all, like, so super duper practical.

这真的很棒。

It's really great.

所以，顺便说一下，这就是秘密总计划。

So here and by the way, here here's the, the step the the the secret master plan.

这应该是机器原生可组合AI的形态。

Should the shape of machine native composable AI

这是你的主意，

It was your idea to

制定一个

make a

秘密总计划。

secret master plan.

这是埃隆的那种做法。

It's a it's a Elon thing.

他当初创办特斯拉的时候，就说，咱们就这么干。

When he started Tesla, he was like, here's here's what we'll do.

我正式把功劳记在你头上。

I I'm giving official credit to you.

谢谢。

Thank you.

谢谢。

Thank you.

但是，你知道吗，你早该告诉我你还要搞这个模型发布会。

But, like, you know, you you should have told me you're you're you're also gonna do this model launch.

因为你只跟我说了一半，另一半呢，你当时连Doom演示都没有。

Because you like, you told me you told me half of the story, and then the other half, you didn't have the Doom demo at the time.

对。

Yep.

你也没给我任何数据。

You didn't have any numbers to give me.

对。

Yep.

对。

Yep.

我当时就想，哦。

I was like, oh.

问题是，我不太相信基准测试。

Well, the problem is I don't believe in benchmarking.

没错。

Exactly.

没错。

Exactly.

所以，这东西你得亲自感受，我觉得这才是建立长期信任的方式，虽然这让我们吃了不少亏。

So, like, it is a thing that you need to feel, and, like, I think that this is the way to build long term trust, even though it, like, hurt it hurt us a lot.

你懂吧？

You know?

比如去年我们融资的时候，没人相信我们。

Like like last year when we did fundraise, no one believed us.

你知道吗？

You know?

他们只想要基准测试之类的东西，但我们说我们不会那么做。

Like like and they wanted just benchmarks and stuff, and we're like, we're not gonna do that.

我们是有原则的。

We are principled.

我们会坚持自己的立场。

We're gonna stand by our guns.

那样会奖励那些做坏事的人。

That rewards bad actors.

我根本不在乎，你知道吗，你到底想要什么？

I don't give a shit, you know, like what do you want?

这就是我们，我们会坚持这一点。

Like, this is who we are and we are standing by that.

非常抱歉。

So sorry.

不是，不，

It's not No,

对，从某些方面来说，我觉得选择一条艰难的路，但最终你会打造出自己真正想工作的公司。

yeah, Well, in some ways I think like choosing the hard path, but you end up making the company that you wanna work in.

是的。

Yep.

否则，如果你妥协了，那你就像在OpenAI工作一样，只不过身边都是自己人。

Otherwise, if you sell out, then you're just working in like OpenAI, but with my people.

差不多就是这样。

Which is like Yeah.

我的意思是，我对此显然没有太多遗憾。

I mean, like, I'm, I don't have too many regrets on that, obviously.

结果好得令人难以置信。

Like it worked out so unbelievably well.

而且，你知道，确实是这样。

And, you know, like I, it did.

昨晚我谈到离开OpenAI的原因时，情绪很激动。

I was emotional last night when I was talking about, like, the reasons I left OpenAI.

因为实际上，发布之后我不得不修改我的措辞。

And because, like, it actually had to change my wording after the launch.

我的原话是，如果AI寒冬真的来了，而我没有尽一切可能去避免它，我会觉得自己负有个人责任，既因为RLHF方向——我觉得这确实加剧了过度承诺和实际交付之间的差距——也因为我没有全力投入这件事，因为我认为这才是价值将要被创造出来的地方。

My, my my phrasing was if an AI winter did happen and I did not do every fucking possible thing I could to, like, avert that, I would see myself as personally responsible both for, you know, the the RLHF direction, which I think really widened overpromise versus underdeliver, and also not going all in on this because I think this is this is where value is going to just be like printed.

所以，这真的很酷，因为我感觉我担心的AI寒冬已经被避免了。

So and it was really cool because I feel like the the AI winter I'm worrying about is averted.

你知道，AI会很有用，会被用于自动化，才不到一周，数据就已经无可否认地表明它正在被用于实际工作，而且这就像狂野西部一样。

You know, like AI will be useful, it'll be used for automation, it's been less than a week and like the numbers are already undeniable that it's like being used for real work and like there's it's the wild west.

你能不能，如果你脑子里有数的话，说说你看到的数据是多少？

Can you just just if you have top of your head, what numbers are you seeing?

比如，注册量什么的，你能分享什么就分享什么？

Like, what's what's, like, sign ups, like, what whatever you can share?

其实我对所有事情并不是特别了解。

I'm actually not super on top of everything.

都是团队在告诉我这些。

Like, the team is the ones who are telling me all

这些每天都在变。

of these changing every day.

确实有点疯狂。

It's it's kinda nuts.

如果有一个里程碑让你觉得，嗯，是的。

If if there's a milestone that you're like, well, yep.

那是我们期待的一件事。

That's one thing we're hoping for.

我们达到了。

We reached it.

我要说的是，我们跨过的一个里程碑是每天的令牌数。

I will say a milestone that we've passed is tokens per day.

不错。

Nice.

而且这不是那种短暂的每日令牌量。

And this is not like fleeting tokens per day.

即使是晚上也在持续运转，所以你知道是机器在调用，而不只是人们在试用。

This is like even at night, like it's constantly churning, so you know machines are calling it and not just people trying things out.

所以这真的很酷。

So that is that is so cool.

每天这么多令牌量是很大的。

Tokens a day is a lot.

能超过这个数字真是太棒了。

So surpassing that is awesome.

对我来说，注册量其实并不重要。

Sign ups to me don't really matter.

说实话，这其实是我们犯的一个小错误。

And actually, this is like a bit of a mistake we made if I'm like totally honest.

推特上的人都说我们是营销天才什么的。

People on Twitter were calling us like marketing geniuses and all of that.

但那其实就只是我们。

And that was just us.

我们也没有专门雇营销人员。

We don't have a marketer also hiring.

我们就是做自己，有点傻乎乎的，也不太正经。

And we were just being our genuine goofy, like irreverent selves.

我们当时就是特别狠地把人从候补名单里放进来。

And we were we were just, like, offboarding people off the waitlist so hard.

我们的平台团队真的强得离谱。

Our platform team is so unbelievably cracked.

我觉得我们的正常运行时间比Anthropic还多，而且我们还经历了史上最空前的发布。

I think we have more nines of uptime than Anthropic while having the most unprecedented launch ever.

这真的挺疯狂的。

Like, that is kind of nuts.

所以，真的要给他们点赞。

So, like, props to them.

而且，我们当时没意识到的一点是，第一，对于开发者平台来说，候补名单的注册人数其实不重要，我觉得。

And, the thing we didn't realize so number one, waitlist waitlist signups don't matter for, a developer platform, in my opinion.

你懂吗？

You know?

我猜其中很多人甚至都不是开发者。

I would guess that a large number of them are not even developers.

他们进去，试几个查询，但很多人搞不明白，因为他们不会编程。

So they go in, they try some queries, and a lot of people don't get it because they are not programming.

他们就会想，这是什么鬼？

Like, they're just like, what?

这不是聊天机器人。

This is not a chatbot.

我的ChatGPT2在哪？

Where where's my chat GPT2?

但是，我还没有精确计算过这个。

But if, I I haven't exactly calculated this.

我的感觉是，如果全世界每个人只是随便写几个查询，那跟一个重度用户用循环创造的价值相比，简直是小巫见大巫。

My sense is that if every single human being in the world, like, just wrote a couple of queries, that would be a rounding error compared to, like, one power user's for loop that is just, like, creating value.

我们之前没意识到的是，关于候补名单，我们只是把所有人都从候补名单上移除了。

And the thing we didn't realize with the waitlist is like we just went off off board anyone off the waitlist.

这其实无所谓。

It doesn't matter.

真正让人担心的是速率限制。

The scary part is rate limits.

一旦人们从中获得价值，他们就会想要越来越多的速率限制，因为软件就是这样运作的。

And then once people start getting value from that, then they just want tons and tons of rate limits because this is what software is.

就像你前期投入精力去定义你的任务，然后这个任务创造的价值会超过你投入的成本。

Like you spend effort upfront to specify your road task, and then this road task creates more value than it takes to put in.

现在你有了它，就忘了这回事吧。

And then now that you have it forget.

对。

Exactly.

你在后台运行它。

You run it in the background.

你把它作为依赖项，或者关联到其他东西上。

You make it a dependency or to, like, other things.

你可以做更高级的东西，然后你就在世界上创造了巨大的价值。

You can make, like, higher level stuff, and, like, you just create so much value in the world.

你知道，早期互联网的人可能没想到，2000年代初互联网的奇妙之处，虽然那还不算早期互联网，但正是通过组合性，没有冒犯的意思，不，

You know, early Internet people probably did not imagine, like, the wonder of early two thousands Internet, which is still not early Internet, but, like, it's it's through no offense, composability No.

所有疯狂的事情都是这样发生的。

That all of the crazy stuff happens.

我真的很想在宣言中强调这一点。

I just really wanted to emphasize that in our manifesto.

我们追求的是涌现。

We are going for emergence.

我们想要成为那个催化剂。

We are going for like being the catalyst.

我们想赋能他人，为此我们会尽一切努力，不管是开Discord还是开市政厅会议，哪怕我穿着垃圾袋也行。

We're wanting to empower people and we are going to do whatever we can for that, be it like Discords in our town hall with me wearing a garbage bag or not.

还有播客，我们想要长格式的内容，对吧？

And podcasts and getting off because of the I want the long form, right?

是的，我们会先跳过一些表面的东西，然后深入探讨，这样人们就会真正信任并理解你的使命，那些有共鸣的人最终会加入你或者购买你的东西。

It is like, yes, we'll get past some of the superficial things and then we'll go deep and people will really trust and understand your mission and like, you know, the people that will resonate that will end up joining you or buying you.

不，不，抱歉。

No, no, sorry.

作为……哦，播客。

As a Oh, podcast.

那真有趣。

That was funny.

对不起。

I'm sorry.

抱歉。

Sorry.

我不是故意那么说的。不过，有个很夸张的说法是，你的发布视频有3600万次观看。

I didn't mean to say But, no, any one one one version one very flattering version of this, like, 36,000,000 views of your launch video.

不错。

Cool.

现在已经涨到3800万了。

Up to 38 now.

四舍五入的误差。

Rounding error.

你知道，纳维-斯托克斯方程那个视频有7400万次观看。

You know, Navier Stokes got 74.

《神鬼寓言5》有5700万次。

Fable five got 57.

所以，我觉得，至于最初的Cache BT，我没有统计过，因为当时没有视频。

So I like, as far as I didn't do the stats for like original Cache BT, which there was no video.

所以就像上面那样，对吧？

So like up there, right?

如果说你要在2026年推出一个Neolab，我觉得你现在就是第一名，这挺疯狂的。

As far as like, if you were to launch a Neolab in 2026, I think you're like number one right now, is which like pretty crazy.

其实我宁愿这样，我确实有那件T恤，上面写着“你最喜欢的NeoLab，最喜欢的NeoLab，最喜欢的NeoLab”。

Well, I actually would rather I do have the shirt like your favorite NeoLab, favorite NeoLab's favorite NeoLab.

我根本不在乎成为NeoLab。

I don't give a shit about being a NeoLab.

我觉得成为NeoLab，其实我们有很多周边，都是对NeoLab的恶搞。

I think being a NeoLab actually, we have a lot of swag that's being a parody of a NeoLab.

其中有一件是，NeoLab带产品，但实际上那并不是NeoLab。

One of them one of them I have is, like, NeoLab with product, which actually is not a NeoLab.

我真的不在乎那些。

Like, I don't care about that, really.

我在乎的是成为一个可靠的开发平台。

What I care about is being a reliable dev platform.

是的。

So Yes.

感谢你的比较，但希望我们能超越他们，回到那种，你知道的，对开发者来说的革命性时刻，以及那种人们可以依赖和信任的稳定状态。

Appreciate the comparison, but like, hopefully we transcend past them and we go back into like a thing you know, like, you know, a revolutionary moment for developers and like this stable thing that people can rely on and trust.

对。

Yes.

我的意思是，就这一点来说，你们真的让我印象深刻的是，你们确实在谈论可靠性。

I mean, to that end, I mean, I think that's one thing that really impressed me about you guys is that, yes, you do talk about reliability.

我原以为主要是关于校准，比如我们说的RLCD，但实际上也关乎正常运行时间和可扩展性等等。

I thought it was mostly about calibration, which like, you know, we talk about RLCD, But actually it's also about just like uptime and scalability and all those things.

这些都是，而且……

They're all sort of And of

九个九。

nines.

在我看来，那就是正常运行时间。

Which is uptime in my opinion.

但那只是其中一部分。

But that's part of it.

但可靠性还体现在，比如这个系统有多智能，它能不能稳定地按你的要求去做事。

But like there's reliability in, like how intelligent the thing is, like how consistently does it do the thing that you want.

我觉得那些大型推理模型确实很聪明。

And I think that like the the big reasoning models are very smart.

但在我看来，它们还是缺乏可靠性。

In my opinion, they still lack reliability.

我觉得有很多场景，看起来它们应该足够聪明，能自动完成工作。

I think there's many use cases where you they look like they should be smart enough to automate their work.

自动化这些工作有经济上的动力，但它们还是不够可靠，连实习生都不如，因为它们的优化目标不同。

There is economic incentive to automate that work, yet still they're not reliable enough as at an intern because they're optimized for different things.

所以我觉得，可靠性还包括能不能信任它们的输出。

And so like, think that there's the reliability of being able to like trust the outputs.

而且，还有一些我们还没达到的可靠性维度，这让我特别兴奋。

Also we are like their dimensions of reliability that we are not yet at that I'm like so excited by.

你知道吗，我想先把简单的工作自动化，再处理复杂的，你懂吧？

You know, like I want to automate the easy work before the hard work, You know?

我觉得这是很基本的常识。

Like, I think that that's just a common sense thing to do.

但对我来说，这样就足够了。

But to me, we will be sufficient.

我不知道有没有所谓的“足够可靠”这回事，但我想做到特别好，好到人们甚至不用试模型就知道它能行。

I don't know if there's a such thing as sufficiently reliable, but I wanna get so good that people don't even need to try the model to know that it'll work.

这就像编程中的心流状态一样。

It's like that's like what flow state is in programming.

就像我在写查询，因为这里需要智能，而且你知道，对于复杂的分支逻辑。

Like, I'm just like writing queries because I need intelligence in here and, you know, like, went for nontrivial branching.

我可以用一个类型安全的系统写一个查询，然后直接得到结果，它就能准确地进行分支。

I could just write it in in like a like like a TypeSafe system one query and then get the results out and it just branches accurately.

那简直太好了。

Like, that would be so, so good.

这就是梦想啊。

Like that is the dream.

而且这将会是一个漫长而艰难的过程。

And that is like going to be like a long, long slog.

我们待会儿会聊聊你的API设计，给大家举些例子，比如那些没走过的路之类的。

We're gonna go into your API design in a little bit, just to give people examples and like maybe path not taken, that kind of stuff.

在前面有个问题我挺好奇的，关于可靠性，我注意到没有设置种子。

One thing up the front that I do wonder about in terms of reliability is I noticed that there's no seed.

没有种子的话，那同样的输入，是不是总能得到相同的输出？

There's no And so basically same input, do I always get same output?

如果不能，那又是为什么？

And if not, why not?

哦，好问题。

Oh, great question.

这其实是我们经常遇到的一个问题，可靠性其实是个很宽泛的概念。

So this is actually like a common question we have between so reliability is actually a catchall.

比如，只要AI没法自动化某件事，那都是因为某种可靠性问题。

Like, whenever AI can't automate something, it's due to some form of reliability.

可能是类型安全的问题。

It could be like type safety.

也可能是确定性的问题，或者就是结果不稳定。

It could be determinism, it just could be like it's jagged.

所以可靠性是个总称。

So reliability is a catchall.

我觉得它也是我们最终目标的一个总称。

I just think that it's also a catchall for like what the North Star is.

确定性就是同样的输入，同样的输出。

Determinism is like same inputs, same outputs.

我觉得这对单元测试来说有点意思，但我认为这不是正确的最终目标。

I do believe that this is like slightly interesting for unit tests, but I believe that to be the wrong north star.

我认为稳健性才是人们真正想要的——我不想告诉人们他们真正想要什么，因为那样有点自大。

I believe robustness is what people I don't want to tell people what they really want because that would be a little arrogant of me.

我觉得那才是更重要的特性。

I believe that that is like the more important property.

你希望的是，输入相似，输出也相似。

You want, given similar inputs, get similar outputs.

而且语言模型有多不可靠，真是挺让人惊讶的。

And it's kind of wild how unreliable LMs are.

我们测试的方式就是往提示里放一些UUID，你知道的，就是那些小东西，我记得叫nonce，然后你希望所有这些都能给出相似的输出，因为从语义上讲，它们确实是同一个问题。

Like, way that we test this is you put like UUIDs in, you know, like, little I think they're called nonces in in the prompt, and what you want is similar outputs from all of those because it's truly semantically the same question.

而正是在这个方面，你特别需要那种稳健性，因为人们在让AI做决策时，往往就是在这里栽跟头。

And that is the part where you really want like like like, that robustness is where, like, people get, like, burnt with AI making decisions.

所以我觉得这是一个超级重要的特性。

So I think that is the a super duper important property.

我们也可以追求确定性。

We could also have determinism.

那也是一种可以实现的特性。

That is that is a thing that can be available.

就我作为程序员的理解来说，在某些用例中它可能很有价值。

As far as I can, like, mentally model for programmers, like, it It could be valuable for some use cases.

所以，请在评论里指正我，或者你也可以。

So like, please educate me in comments or you.

但总的来说，这很容易。

But my in in general, it's easy.

确定性是你可以用来换取更好成本的东西。

Determinism is something you can like trade off for better cost.

你知道，我们一直在追求每美元智能的前沿。

You know, like we are we are constantly wanting to be on the intelligence per dollar frontier.

我们为了达到那个目标，做了很多非常疯狂的事情。

We are doing like absolutely disgusting things to be there.

你知道，这我不该说，但没人拦着我。

You know, like this is, I shouldn't say this, but no one's here to stop me.

你知道，你自己批准自己的代码合并。

You know, you sign off on your own PR.

那

That

在我们公司可不是这样的。

is not how it works at this company.

我觉得这周，我的幕僚长K是科技界最有权力的人。

I believe for this week, my chief of staff, K is the most powerful person in tech.

还要感谢K组织了这一切。

And shout out to K for organizing this.

天哪。

Holy.

天哪，太厉害了。

Holy shit.

她简直太能干、太有权力了。

She is so fucking competent and powerful.

她太了不起了。

She's incredible.

我是说，她其实挺糟糕的。

I mean, she sucks.

别挖她。

Don't poach her.

但我也试着说得委婉一点，不过大家都跟我说，别把它叫做“弗兰肯斯坦的怪物”模型，因为那听起来有点负面。

But so I tried to be a bit more filtered, but, like, people are telling me, don't call it a Frankenstein's monster of models, but because that has, like, negative implications.

我觉得弗兰肯斯坦的怪物在这整个故事里其实是好人，我是说，它是无辜的。

I think Frankenstein's monster was, like, the good guy in this whole I mean, it was innocent.

我没读过那本书。

I didn't read it.

好吧。

Okay.

我承认。

I'll confess.

好吧。

Okay.

就是那个面部表情，我……我的牌没（跟上/对上）

That that one facial expression I my cards didn't

如果你想看的话，推荐一部还不错的雅各布·维拉尔迪的电影。你可以看看标注内容。

Decent Jacob Villardy movie if you wanna see I You the annotation.

总之嘛。

Anyway.

你有

You have

我现在根本不知道自己能挤出多少时间。

no idea how little time I have right now.

我现在的头等大事就是睡觉。

My priorities are sleep.

你懂的吧？

You know?

还有那些开发者们。

Developers.

开发者。

Developers.

开发者。

Developers.

开发者。

Developers.

对。

Yes.

开发者。

Developers.

开发者。

Developers.

开发者。

Developers.

但是，是的，我们确实会做一些非常疯狂的事情，就为了在智能和成本之间达到帕累托最优，而且我们会继续这么做。

But, yes, we we do like absolutely disgusting things to be on the Pareto curve of intelligence per dollar, and we are going to keep doing that.

我们会做一些疯狂的事情，我觉得大家真的需要跳出框框来思考。

We're gonna be doing crazy ass stuff, and I think people really need to think outside of the box.

比如说，之所以让人惊讶，是因为人们都困在框框里，而我们继续在框框外做事。

Like, like, part of the reason why it's surprising is, like, people are thought inside the box, we continue to do that.

目前来看，我们显然是最强的，我们也想继续保持这个领域的领先地位。

As of right now, we are obviously the best at this, and we want to continue being the best at that whole thing.

等等。

So, wait.

我刚才说到哪儿了？

Where did where did my tangent from?

所以，我问你的是，你们会有种子和确定性吗？

So so I I asked you about, will you have seeds and determinism?

是的。

Yes.

然后你基本上定义了可靠性和鲁棒性。

Then you basically define reliability and, like And robustness.

是的。

Yes.

但是，嗯，怎么说呢，

But, like, determine well, like

我有个关于稳健性的例子，很快就能说完。

I have a robustness example that's that's, real quick.

可以给你看看。

Can show you.

太好了。

Love that.

我就说一点。

I'll just say one thing.

我们可以做一个确定性的模型。

We can make a deterministic model.

没错。

Exactly.

就是说，如果有人能说服我们这是有价值的事情，我们也很乐意去做。对。

Like like, we're hap if people can convince us that that is a valuable thing to do Yeah.

而且我们也没有严重的GPU短缺问题。对。

And we don't have a gigantic GPU shortage Yeah.

我们可以很轻松地把这些模型都做出来。

We can happily make all of these models.

我们就是想让别人满意，然后不断改进、不断革新。

We live to please, and rev and and revolt revolute.

你们会推翻一切，但方式很温和。

You will throw over everything except you do it in a nice way.

而且很有趣。

And fun.

所以呢，确定性模型也是有可能的。

So so like, yeah, determinism could be on the card.

只是每花一块钱能买到的智能会少一些。

It just gets you less intelligence per dollar.

嗯，看看OpenAI和Endtopic的发展轨迹，你现在就相信我吧，你会被同伴压力推着去做的。

Well, just having seen the trajectory of OpenAI and Endtopic, you will just trust me now that you will be peer pressured into doing it.

所以，就像，即使你告诉人们他们不需要，他们还是会想要。

So, like, just people won't want it even if they if you tell them they don't need it.

他们还是会想要。

They'll still want it.

所以，对，这就是大概的意思。

So, like, yeah, that that's the TLDR of Okay.

好的。

Okay.

我很期待，也许有一天我们会看到这一切发生。

I will love to maybe one day we will see how that happens.

有人告诉我，他们说我们品牌的一部分就是坚定不移，他们说这只是固执的好听说法。

I've been told I'm they they they say that part of our brand is being unshakable, and they say that that's just the nice way of saying stubborn.

固执。

Stubborn.

没错。

Exactly.

而且我是个非常固执的人。

And I'm a very stubborn person.

我觉得我们不可能

I don't think we could have

做到那一点。

done that.

不。

No.

但是，我，好吧。

But so, I I okay.

但我之前跟你争论过。

But I I, like, have argued with you before.

而且你每次都在写关于开发者的事。

And And you've been writing about developers every time.

那好吧。

So, okay.

我认输了。

I give up.

你赢了。

You win.

你赢了。

You win.

我承认我之前跟你争论过。

I I'm sold that I've argued with you before.

不，不，我只是说，我觉得你可以坚持自己的立场，但如果我给你正确的证据，你也能放下之前的想法，说，对，这确实说得通。

No, no, I'm just saying like, I think that you can hold your ground while also like, if I give you the right evidence, you can, not you can sort throw away your priors and be like, yep, like that actually makes sense to me.

所以，你知道的，就相信你自己的直觉吧。

And so like, you know, just just trust your own gut on this.

对。

Yep.

对。

Yep.

对。

Yep.

我来给大家讲个话题。

I'll bring us a talk.

不过我怀疑，我们会在很长很长一段时间内受限于GPU。

I suspect though that we will be GPU constrained for a very, very long time.

任何单位智能成本更低的东西，意味着同样的智能需要消耗更多的GPU，这你知道的，目标不是让公司入驻。

And anything that has less intelligence per dollar means it consumes more GPUs for the same intelligence, which is, you know, like, goal is not to onboard companies.

就是说，这有价值，但我们的目标是让人们去尝试，去做些奇怪的事情。

Like, it's it's valuable, but, like, our goal is to have people, like, experiment and do weird shit.

我们需要尽可能多地让更多人接触到它，就像加州淘金热那样开始。

And we need, like we need to, like, get it to as many hands as possible and, like, starting, like, the California Gold Rush for that.

我觉得现在就有这种情况。

I think there is right now.

提醒一句。

Just a word of caution.

我是说，我之所以这么说，是因为现在有人在考虑这件事。

I mean, I I would just say it because somebody is thinking about it right

嗯，对。

now Mhmm.

就是当你说出像“我们不会承诺确定性模型”这样的话时。

Which is when you say things like, we will not commit to deterministic models.

我们会为了每美元智能度不惜一切代价，而且我们正面临GPU限制。

We we will, we'll do whatever it takes for intelligence per dollar, and we are we are facing GPU constraint.

人们会想，你们可能会对模型进行量化。

People are thinking you may quantize your models.

比如，无论你们发布时是什么样，你们可能会降低量化精度，以牺牲质量来释放内存或带宽之类的。

Like like, whatever you had at launch, you may quantize down in to reduce the quality in order to to free up memory or bandwidth or whatever.

所以你们最好做出某种承诺，现在不必承诺，但比如“我们会保持发布时的模型质量”。

And so you should probably have some kind of promise, which you don't have to make now about like, we will uphold model quality at launch.

就像，比如当人们，当我们，我是说，我们在OpenAI的时候，你推出所有这些API，还有Claude也是。

People like, so like when people, when we, I mean, were at OpenAI when you launched all these APIs and even Claude as well.

就像他们最初推出模型的时候，模型字符串并不是一直保持同一个模型的。

Like when they first launched the models, the model strings did not stay the same model at all times.

你们的模型是有版本管理的。

You have versioning in your models.

那很好。

That's great.

但是，

But like

你们应该公开承诺，比如，一旦某个东西发布了，我们就不去改变它。

you should publicly commit to some kind of like, once the thing is launched, we don't change it.

我们部署模型后不会去改变它们。

We will not change our models when we deploy them.

那太疯狂了。

That is insane.

我们很关心开发者。

We care about developers.

就是说，如果你再这么做的话，那就有问题了，这就是第一方产品和API之间的区别。

Like like like, it makes sense if you're so doing something like that again, this is the problem with a four first party product and an API.

在第一方产品里，你想怎么做都行。

It makes you can do whatever you want in a first party product.

随便他们，只要能带来好的体验，那没问题。

Like like, more power to them, whatever gets that experience, that is fine.

但API就不行了，显然不能那样。

With an API, you obviously can't do that.

不过我要说的是，我们计划比大家习惯的模型提供商动作快得多。

But I will say that we plan to move a lot faster than many people are used to model providers doing things.

所以我们会比大家想象的更快推出新模型。

So we will be launching new models a lot faster than people think.

而且我们不会承诺对模型的长期支持，因为我们认为还有很多改进空间。

And we are not promising long term support for the models because we think that there's lots of improvements to have.

所以有可能我们会暂时对现在的Jev 1.13.0提供长期支持。

So there is a world that we might temporarily LTS what is right now Jev one point thirteen point zero.

我们可能会这么做，因为用的人太多了。

We might do that because so many people are using it.

而且我知道开发者最讨厌依赖关系被破坏。

And I know developers hate breaking dependencies.

另一种选择就是让我们的模型版本四分五裂，这对大家来说都不是好事。

The alternative is fracturing our fleet, and that is a very bad vibe for everyone.

对，你确实会有

Yeah, you do have

上百个不同的模型版本。

like a 100 different versions of the model.

没错，如果我们迭代得很快，这些版本也会很多。

Exactly, and if we're iterating very fast, there would be a lot of those versions as well.

所以我们最终不只是想要一个长期支持的版本，而是长期支持。

So, we we do want to have not just a LTS supported thing eventually, long term support.

我们想要一个特别棒的方式来做这件事。

We want a really sick way of doing that.

我们在这方面有一些研究正在推进，我觉得那会是史上最受开发者欢迎的东西。

We have research stuff cooking in that direction, and I think it's going to be the most pro developer thing ever.

但这还不是我们当前的模型，而且我不能保证我们能保持完全相同的模型。

But it is not yet our current models, and I'm not promising that we will be able to keep the exact same models.

它们每次肯定会变得更聪明。

They will get smarter every time for sure.

我的感觉是，即使在我们模型的迭代中，它已经很聪明了，在模型版本之间，变化往往比字符串模型调用两次还要小。

And my sense is that even our model iterations where it already is smart, it's between model versions, the the changes tend to be even smaller than the string models calling them twice.

但是当我们从，比如，锯齿状变成，哇，那才是大差异所在。

But when when we go from, like, you know, jagged to, like, wow, that is where the big deltas are.

LTS相同模型的一个好处是，实际上你也可以把它们移植到其他芯片上。

One thing that's beautiful about LTS same models is that actually you can also port them to other silicon.

我不知道你有没有想过这个。

I don't know if you've thought about this.

无可奉告。

No comment.

好的。

Okay.

所以我关心的是每美元能买到多少智能。

So I care about intelligence per dollar.

是的。

Yes.

但速度呢。

But speed.

什么？

What?

速度也很重要。

Speed as well.

走着瞧吧。

We'll see.