本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
我正在英伟达年度GTC大会现场，将采访四位杰出的AI首席执行官。
I'm here at Nvidia's annual GTC conference, and I'm gonna interview four amazing AI CEOs.
本节目由纽约证券交易所赞助。
Stick with sponsored by the New York Stock Exchange.
你是否希望改变世界并筹集资金？
Are you looking to change the world and raise capital?
那就去纽约证券交易所吧。
Do it at the NYSE.
纽约证券交易所是一个现代化的市场，一个为规模化和长期影响力而打造的庞大平台。
The NYSE is a modern marketplace and a massive platform built for scale and long term impact.
因此，如果你正在为未来而构建，纽约证券交易所正是这一切发生的地方。
So if you're building for the future, the NYSE is where it happens.
人工智能时代的一家杰出公司，当然是CoreWeave。
One of the great companies of the AI era is, of course, CoreWeave.
他们正在为这些超大规模企业构建庞大的基础设施。
They're building massive infrastructure for these hyperscalers.
从某种意义上说，迈克尔·因特拉托，欢迎来到本节目。
And in some ways, Michael Intrator, welcome to the program.
你是最初的超大规模云服务商。
You're the original hyperscaler.
你们很早就入局了，搞到了我不知道是哪些型号的GPU。
You guys got in very early and secured your I don't know which GPUs you end up getting.
不。
No.
你在这个趋势中确实非常早入局。
You were very early to this trend.
你是怎么这么早就进入这个领域的？又是如何建立起当时所谓的Neo Cloud的呢？
How did you get to it so early and how did you build out this, you know, first, I guess at the time, Neo Cloud?
是的。
Yeah.
其实我们最初并没有把它当作Neo Cloud来启动。
So we didn't we didn't really start it as a Neo Cloud.
我当时运营一家专注于天然气的算法对冲基金。
I I I was running an algorithmic hedge fund focused on natural gas.
当你构建一个算法对冲基金时，一旦算法搭建完成，你主要就是监控它、测试不同的假设，以及做各种相关工作。
And when when you build an algorithmic hedge fund, once the algorithms are built, you're really just monitoring it and testing different theses and doing all that.
但同时也有很多空闲时间，我们对加密货币产生了浓厚兴趣。
But there's also a lot of downtime and we got super interested in crypto.
而且，你知道，我们挺技术宅的。
And, you know, we're pretty nerdy.
我们喜欢深入探究底层，于是开始对安全层产生兴趣。
We kind of dig under the hood and we started to get interested in the security layer.
我们研究了比特币和比特币挖矿，但并不喜欢它。
We looked at Bitcoin and the mining for Bitcoin, and we didn't like it.
我们觉得，肯定有某个天才工程师设计了ASIC芯片，他们自己运行起来肯定比我们更在行。
We just thought that, like, there's some brilliant engineer that built the ASIC, and they're probably gonna be better at running it than we are.
所以我们开始重点关注GPU，主要是因为GPU不仅能用来挖以太坊，还能做很多其他事情。
So we really began to focus on the GPUs, mostly because the GPUs were you can mine Ethereum with them, but you could also do all these other things.
因此，从一开始，我们就将计算能力视为一种可以部署到不同应用场景的资源。
And really, so right from the start, we looked at the compute as an option to be able to deploy our computing power to different use cases.
所以，我们在2017年创立了这家公司，前三年主要做加密货币挖矿，经历了几次加密货币寒冬。
And so, you know, began the company in 2017, you know, spent the first kind of three years mining crypto, went through a couple of crypto winters.
因为我们来自对冲基金。
Because we had come from a hedge fund.
我们确实擅长风险管理，懂得如何思考资本、风险敞口和配置等问题。
Our you know, we we have real chops in risk management and how we think about capital and risk exposure and allocation and all of that.
因此，从一开始我们就对此非常谨慎。
And so we were really careful around that right from the start.
所以我们很好地度过了加密货币寒冬，并开始扩大公司规模，同时立即寻找其他可以利用这些计算能力的场景，因为加密货币波动性很大。
So we weathered crypto winter really well and began to scale the company and immediately started to look for other use cases that you could use this compute for because crypto was pretty volatile.
是的。
Yeah.
当时加密货币还是个未知数。
And crypto was a question mark at that time.
当然。
Absolutely.
对。
Yeah.
我的意思是，比特币具有投机性，还有很多其他投机性项目。
I mean, Bitcoin was speculative, and there were many other speculative projects.
当时使用这种硬件的其他人都是一些量化交易员，是的。
The only other people using this type of hardware, quants Yeah.
医学研究人员。
Medical researchers.
所以，可以这样理解我们逐步开始涉足的产品方向。
So a good way to think about it is, like, the progression of products that we kind of started to work on.
首先是加密货币，但我们很快就从加密货币转向了计算机图形渲染，并开发了帮助人们进行动画和图像渲染的项目，就是那种让电影变得酷炫的东西。
You know, first was crypto, but we immediately moved from crypto to CGI rendering, and we built projects that would allow folks that were trying to animate and render images, you know, kinda what makes the movies cool.
对吧？
Right?
是的。
Yeah.
我们开始着手这个领域，随后转向批量计算，研究如何利用计算能力推动医学研究和其他科学应用。
And and we started to work on that, and then we moved to batch computing and started to look at medical research and different ways of using the compute to be able to drive science.
我们不断向上游推进，探索GPU在更复杂场景中的应用。
And we just kind of kept moving up the stack in terms of complexity on how GPUs could be used.
最终，在大约2020年到2021年期间，我们开始认真探索如何将GPU用于神经网络，而当时我们并不知道该如何做到这一点。
And ultimately, in like, call it like 2020, 2021, we started to really try to figure out how you can go ahead and use GPUs for neural networks, and that was not something that we knew how to do.
于是我们实际购买了一批A100，并捐赠给一个正在研究Luther AI的团队。
And so we actually went out and bought a bunch of a one hundreds and donated them to a a group that was working on Luther AI.
他们正在开展一个开源项目，我们捐赠GPU计算资源，就是希望这些团队能充分利用它。
They were working on an open source project with the thought that these guys are taking the GPU compute because we're donating it.
如果我们一开始做得不够好，他们也无法责怪我们。
They can't really get at us if we're not very good at it initially.
这取得了非常好的效果，因为
And that worked out really well because
他们不能抱怨服务等级协议。
They can't complain about the SLA.
他们一直告诉我们，我们需要更多这样的东西。
They they kept telling us like, we need more of this.
你得专注于这个。
You gotta work on this.
这让我们真正理解了运行大规模并行计算所需要的条件。
And that began to really give us an understanding of what was necessary to run scale parallelized computing.
而且，你知道，我们确实经历了这一切。
And, you know, that that we went through it.
我觉得购买那些最初的GPU就像是我们支付的学费，是的。
I I I kinda feel like buying those initial GPUs was the tuition we paid Yeah.
来学习如何经营这项业务。
To learn how to run this business.
然后有一件有趣的事是，所有那些人后来都回到了他们的本职工作，因为他们都是志愿参与这个项目的。
And then one of the interesting things is all of those guys went back to their day jobs because they were all volunteers working on this.
他们是有相同理念的科学家。
They were like minded scientists.
当他们回到日常工作时，都表示：我想要这样的基础设施。
And when they got to their day jobs, they were all like, I want that infrastructure.
它被正确地构建了。
It's built the right way.
研究人员会希望以这种方式使用它，这推动了我们的业务发展。
That's the way that researchers are gonna wanna use it, and that launched our our business.
这是一个了不起的故事。
It was an amazing story.
所以你从加密货币转向了这些学术界和深度研究领域的研究人员。
And So you went from crypto to these researchers into academia and deep research.
扑克牌局中接下来要翻的牌是什么？
What's the next card to turn over in the poker game?
是的。
Yeah.
因此，早在2020年到2021年、在ChatGPT出现之前，我们就非常清楚地意识到，规模定律将起到决定性作用。
So so what became very clear to us very, very early on was that the scaling laws were going to drive and remember, this is really back in the 2020, 2021, before ChatGPT moment occurred.
我们开始明白，计算能力在规模化后会变得不再商品化，对吧？
And we began to understand that computing decommoditizes at scale, right?
任何人都可以运行GPU，但你能运行一个足够大的集群来训练一个改变世界的大模型吗？
Anybody can run a GPU, but can you run a cluster that's large enough to train a model that can change the world?
这是个完全不同的问题。
And that's a different question.
因此，我们开始认真思考：如何为客户——尤其是规模越来越大的客户——提升计算能力的交付能力？
And so we really began to think about, like, how do you go about scaling up your delivery of this computing to clients, larger and larger clients?
接下来要翻的牌，就是从这样一个角度出发：这其中有一部分取决于我们获取资本的能力，以便将我们的解决方案提供给最广泛的受众，以及最精通计算资源的高端用户。
And that was the next card to turn is to think about it from a, okay, there's a component of this that is going to lean into our ability to access the capital to be able to deliver our solution to the broadest possible audience, to the most sophisticated consumers of this compute.
是的。
Yeah.
而接下来要思考的，就是把它当作一项业务，而不仅仅是一个工程项目的延伸，这样才能交付基础设施、软件，以及所有介于你所做的事情之间的环节——当我们思考我们所做的工作时，我们实际上处于Nvidia GPU之上、模型之下。
And and that was really the next card is thinking about it as a business rather than as a engineering project to be able to deliver the the the infrastructure and the software and really everything between you know, when you when you're thinking about what we do, we kind of live above the Nvidia GPUs but below the models.
是的
Yeah.
所有这些内容，包括软件、软件与运维的集成、可观测性，以及构建专为这一特定用例设计的云所需的一切。
And everything in there, all the software, the integration of software and operations and observability and all the things that you need to be able to build a cloud that's purpose built for this one specific use case.
对吧？
Right?
所以我们并不做所有事情。
So we don't we don't do everything.
我们专注于一个特定用例，那就是网页服务器。
Really focus on one use case, which allows us web servers.
你跟AWS不一样。
Different you got AWS.
你知道吗？
You know what?
他们做得很好。
They do a great job.
这简直是一个绝佳的解决方案。
It's like it's a it's a great solution.
这是一个解决难题的 brilliant 方案。
It was a brilliant solution to solve a problem.
我们只是看了看，发现了一个新问题，于是决定去研究这个问题，尝试找到一个解决方案，以提供能够解决该问题的计算能力。
We just looked at it and said, there's a new problem, and let's go about let's go about looking at this problem and try and come up with a solution to deliver compute that solves that problem.
那么，语言模型是什么时候开始打电话给你，要求扩容的呢？
And when did the language model start dialing and calling you for, you know, capacity?
是的。
Yeah.
所以，我们的第一个语言模型实际上是 Luther。
So our our our first well, our our first language model was really a Luther.
没错。
Yes.
但我们的第一个大规模商业应用是 Inflection。
But our our first, like, large commercial was inflection.
啊。
Ah.
所以，你知道，我们与穆斯塔法和Inflection合作，之后我们进一步扩展到超大规模云服务商、OpenAI，覆盖了各种基础模型，并持续不断扩展，因为我们坚信，计算的去商品化、提供解决方案的能力，以及构建能够改变世界的超级计算机，正是我们开始专注的方向。
And so, you know, we work with Mustafa and and and inflection, then we we really diversified from there into the hyperscalers, into open AI across the the the model, the foundation models across you know, and and just kept scaling and scaling with the belief that, you know, once again, the the the decommoditization of compute, the ability to to deliver a solution, and the solution is building supercomputers that can change the world, And that's really what we began to focus on.
这为训练铺平了道路，而如今世界已经经历了这样一个时刻：我们从研究阶段迈入了产品的商业化阶段。
That was the lead into training and now the world has gone through, you know, this this moment where we've moved from research into the productization of this.
它正逐渐从组织的边缘渗透到核心业务中。
It's it's it's beginning to work its way in from the fringe of organizations into the core of what they do.
你每天都能看到，通过我们基础设施层运行的推理计算量巨大无比，简直就像是……
And you can see that every day in the amount of inference compute that is being driven through our infrastructure layer, which is just massive, which is just like one of the
人们消费它的渠道，不只是在构建模型，而是在部署并使用它们。
channels people are consuming it, not just building models, but they're deploying them and and utilizing them.
我总是把推理看作是人工智能投资的变现，是的。
I always think of inference as the monetization Yeah.
对人工智能投资的回报。
Of the investment in artificial intelligence.
所以当我们看到我们的计算资源被用于支撑每天海量的推理请求时——你知道，推理就是人们向模型提问，模型给出回答的时候。
So when when when we when we see our compute being used to stand up the massive scale of inference that's hitting our compute every day and, like, you know, inference is when people ask the model a question, it comes back with an answer.
那就是推理。
That's an inference.
或者当你向模型提问，然后它去执行某个操作，那也是推理，对吧？
Or when you ask the model a question and then to go do something, that's inference, right?
而这才是真正有机会将价值从模型本身延伸到现实世界的地方。
And that's actually where you're you're you have the opportunity to really drive value outside of the model itself, but into the real world.
这对我们来说非常令人兴奋。
And that's really exciting for us.
这正是我们喜欢关注的。
That's what we like to watch.
这正是我用来衡量发展状况所关注的。
That's what I like to watch in terms of gauging the health.
那些是哪些芯片？
What chips are those?
所以，说实话，我们是率先将英伟达的新架构推向市场的先锋。
So really, you know, we are we are the tip of the spear in bringing the new architecture out of Nvidia Yeah.
实现大规模商业化生产。
Into into commercial production at scale.
是的。
Yeah.
当我们率先大规模部署H100时。
And so when when you know, we were the first ones to bring the H100s at scale.
我们是第一个大规模部署H200的。
We were the first ones to bring the H200s at scale.
第一个部署GB200，现在你们看到的是GB300。
First ones with the GB200s, and now you've got the GB300s.
对我们来说，一件令人惊叹且非常有趣的事情是，人们在新架构推出后，立即用最前沿的GPU来训练模型，然后把这些GPU转移到不同的实验中。
And one of the things that's that's that's amazing and really fascinating for us is is, you know, people are using the bleeding edge GPUs to train models as the new architectures come out, and then they take those GPUs and they move them into different experiments.
随着时间推移，他们再把这些GPU用于推理任务。
And then over time, they move them into inference.
而且它们会在推理任务中继续使用很长时间。
And they continue to use them in inference for a very, very long time.
现在一个100的使用寿命是多久？
What is the shelf life of a 100 right now?
是的。
Yeah.
这一直是个热门话题，我认为对你们公司、微软，还有迈克尔·默里——你当量化交易员时肯定认识他——来说都是如此。
That's been a big debate is, I think, for your company, for Microsoft, and I guess Michael Murray, you're you know, who you must have known when you were a quant Yeah.
你知道，你会说，天哪。
You know, saying, oh my god.
整个行业都要完蛋了。
The whole industry is the sky is falling.
但我们行业里都知道，人们并不会直接把这种硬件扔掉，而是会找到其他用途。
And then we all know in the industry that people don't just throw this hardware away, that they find uses for it.
街头总会找到技术的新用途。
The street finds its own use for technology.
那么这些设备的实际使用寿命是怎样的？
So what's the reality of the lifespan of these things?
关于GPU折旧这个说法，我的看法是。
So so my my take on the the the GPU depreciation bait Yeah.
这完全是无稽之谈。
Is that it's nonsense.
对吧？
Right?
这个争论是由一些持有股票空头头寸的交易员推动的，他们试图打压股价。
It's a debate that is being brought to the forefront by some traders that have a short position in the stock and they're trying to talk down.
听好了。
Look.
我们来看看事实。
Here's what we know.
对吧？
Right?
当我们购买基础设施时，我们是一家以成功为导向的公司，对吧？
When when we buy infrastructure, infrastructure, we're a success based company, right?
相对于我们所竞争的那些巨型公司，我们是一家相对较小的公司。
We're a small company on a relative basis compared to the enormous companies that we're competing with.
因此，我们的客户来找我们，购买五到六年的计算资源。
And so they come our clients come in to us and they buy compute for five years, for six years.
我们的平均合同期限是五年。
Our average contract is five years.
所以，无论是行业内还是行业外的任何人，只要说这些东西在十六个月后就会过时，或者散布其他类似的胡言乱语，都完全不符合实际情况。
So any commentary by anyone either inside or outside of the industry that this stuff becomes obsolete in sixteen months or whatever nonsense they're spewing, it's it doesn't it doesn't in any way match up with the facts on the ground.
实际情况是，客户是按五年周期购买的。
The facts on the ground is they're buying it for five years.
对吧。
Right.
对吗？
Right?
如果人们愿意为它付费，那它就仍然有价值——这一直是我的观点。
If if and my approach to this has always been, if people are willing to pay me for it, it still has value.
没错。
Correct.
这种看待问题的方式很简单。
Pretty simple way of of approaching it.
我们采用六年折旧。
We use a six year depreciation.
我们相信这些GPU的使用寿命会超过六年，但我们认为，面对如此快速发展的技术周期，六年是一个公平合理的选择。
We believe that the GPUs will last in excess of six years, but we felt like that was a fair and reasonable approach to a technology cycle that's moving at this velocity.
A100、安培系列，今年它们的价格在整个年度里都有所上涨。
The a one hundreds, the Amperes, this year, the price has appreciated through the year.
那为什么会这样？
And why is that?
我认为这是因为，随着更多已部署的算力投入使用，新的公司开始涌现，它们有新的应用场景、不同规模的模型，正试图建立新的商业项目，而这些项目之前可能被A100所排除。
I I think it's because one of the things that happens is as more installed capacity becomes available, you have new companies that come into existence that have new use cases, that have different size models, that are trying to build new commercial ventures that maybe have been blocked out of the h one hundreds
是的
Yeah.
而且从未有机会运行在
And never had an opportunity to run on
那个。
that.
我的意思是，给观众举个简单的例子：当你在三四年之后换掉你的iPhone时，你会想，谁还会用iPhone 12？
I mean, to make a very simple example for the audience, like, when you trade in your iPhone after three or four years, you're like, who's gonna use an iPhone 12?
而在南美或非洲，你去商店里买的可能是iPhone 12或者Pixel 7，只要50美元，但它们依然有很长的使用寿命。
And it's like, have you been to South America or Africa where you go to the store and you buy an iPhone 12 or you buy the Pixel seven and it costs $50, that's still got great life left in it.
没错。
Absolutely.
你知道的？
You you know?
而且，你看，我们发现了许多惊人的应用场景——新成立的公司，或者已经将新模型整合到工作流程中的现有公司，它们都能够使用Ampere芯片。
Well and so, look, you know, we we we find these amazing use cases, new companies that have come into existence or existing companies that have integrated new models into their workflow that are able to use the Amperes.
所以他们一直在购买我们所有可用的显卡。
And so they keep buying any GPUs that we have available.
而且again，说一块显卡在十六、十八个月，或者两年后就不再相关或商业上可行，这根本说不通。
And once again, you know, the the concept that a GPU is no longer relevant or commercially viable after sixteen more, eighteen months, or two years Yeah.
就是这样。
That's It
这根本说不通。
just it just doesn't make it sense.
荒谬至极。
Farcical.
是的。
Yeah.
我觉得有时候人们会被摩尔定律或我们行业增长的速度所迷惑，觉得有太多利益牵涉其中，大公司都要求使用最新产品。
I think sometimes people get caught up in Moore's Law or in just how fast our industry is growing Yeah.
因此，大公司都在要求最新型的产品。
And that there's so much at stake that big companies are demanding the most recent products.
这并不意味着使用寿命变短了。
That doesn't mean that the lifespan has gotten shorter.
这意味着机会以及机会的范围变得大得多。
It means the opportunity and the surface area of the opportunity has gotten much larger.
是的。
Yeah.
其中一个原因是，这个行业因为涌入的前所未有的巨额资本而受到了极大关注。
One of the things is is, like, you know, the the the industry has gotten so much attention for the unprecedented scale of capital that is coming to bear on this.
是的。
Yeah.
正因为如此，人们往往极度关注那些基于最先进芯片构建的公司。
And because of that, there tends to be a incredible focus on the companies that are building on these most advanced chipsets.
事实上，即使在这些公司内部，它们的芯片仍有很长的使用寿命，可以用于提供推理算力、开展其他实验、处理一些不那么前沿但仍需完成的工作。
And the truth of the matter is is, you know, even within those companies, they have a long tail of useful life to provide inference horsepower, to work on other experiments, to do less bleeding edge activity, but still needs to be done.
而且，是的，我想到了渲染也是如此。
And, yeah, I mean, rendering comes to mind as well.
或者是的。
Or yeah.
我们正在Nano Banana上生成图像。
We're making images on Nano Banana.
比如，它会派上用场的。
Like, there there there will be a use for it.
有一个时间节点，那时计算与功耗的比率可能不再合理。
There is a moment in time where maybe the compute to power ratio doesn't make sense.
我的预期是，过时将由这样一个时刻定义：数据中心的电力可以被重新用于比现有基础设施提供更高利润率的用途。
My my expectation is is obsolescence will be defined by the moment in time where the power in the data center for me will be able to be repurposed for a higher margin than the existing infrastructure provides.
而且，正如我所说，我完全预计这些基础设施的使用寿命会超过六年，但这一领域普遍采用的标准——除了亚马逊之外——实际上一直是一年，而亚马逊确实是六年。
And, you know, like I said, I I fully expect this infrastructure to last in excess of six years, but the the the standard in the in in in the space has really been used with one exception, which is Amazon, which is, yeah, it's six years.
这似乎是合理的周期。
That's that seems like the right schedule.
我不是瞎说的。
I'm not making it up.
这就是大家都在用的。
That's what everybody's using.
对。
Yeah.
能源成本就是机会，因为嘿。
And the energy cost is the opportunity because, hey.
我们需要的就是这个空间。
It's just this we need that space.
这里的回报更好，这些硬件可能会被转售给其他需要的人，比如爱好者之类的。
There's a better reward here, and that might get resold that hardware to somebody else who wants it, a hobbyist or something.
是的。
It's Yeah.
我的意思是，或者它可以被送到其他有更多容量的地方，在那里可以重新利用。
I mean or it could be sent someplace else where they have more capacity when they can repurpose it there.
但我总觉得，嗯，我们等到那时再处理这部分业务吧。
But I I I I kind of feel like, you know, we'll we'll deal with that part of the business when we get there.
我现在知道的是，这极其盈利。
What I know right now is it is extraordinarily profitable.
继续保留那些已经运行、并签有长期合同的基础设施，对我的公司非常有利。
It's very accretive to my company to continue to keep the infrastructure that's been up and running, that's been on these long term contracts.
当这些设备使用满五年后逐渐到期，我依然能以比一年前更高的价格出售它们。
And as it rolls off, as it's been in use for five years, you know, as it becomes available, I am still able to sell it at a higher price than it was at a year ago.
现在竞争很激烈。
There's competition now.
当你当年从詹森那里购买这些设备时，是的，你下单后应该能在三十天内收到货。
When you were buying these from Jensen back in the day, yeah, you could buy them and have them shipped, I would assume, within thirty days or less.
如今，即使是像你这样的老客户，等待时间有多长？
Nowadays, what's the wait like even for you, a loyal old customer?
这中间是不是有点竞争？
And is there a bit of a battle?
服务器的分配会不会涉及一些政治因素？
Is there politics to who gets the servers?
比如，我看到一些非常知名的人物在说他们必须拿到配额。
Like, I've you see some, like, very big names talking about they gotta get an allocation.
这还是有点疯狂吗？
Is it still a little bit crazy?
身处这种必须购买人人都想要的东西的境地，感觉如何？
What's it like to So be in that category having to buy something everybody wants?
听好了。
Look.
你知道，我
You know, I
我觉得这证明了我们所从事的业务是成功的。
I I I think of it as an affirmation of the business that we're in.
对吧？
Right?
比如，我们吸引到竞争对手，恰恰说明这个业务是健康的，有很多人试图提供这项服务，因为对这种基础设施的需求很大，需要将基础设施集成到软件层中，以向人工智能提供支持——无论是模型层面、推理层面、应用层面，还是杰ensen所关注的五层蛋糕中的任何一层。
Like, the fact that we are attracting competitors, the the means that the business is healthy and there's a lot of people trying to deliver this service because the need for this infrastructure, the need to integrate the infrastructure, you know, into the software layers to deliver it to artificial intelligence either at the model level or the inference level or the application level or whatever, you know, level of the five layer cake that Jensen's, you know, focused on.
有更多人进入这个领域，这并不会让我气馁。
The the the fact that there are more people coming into this, it doesn't discourage me.
是的。
Yeah.
至于获取GPU，我们和其他人一样，带着采购订单前来，准备好付款。
As far as getting access to the GPUs, we show up like everybody else with a you know, we we'd like to buy here's a PO and we're ready to pay.
那一个是
The one what's
等待时间是怎样的？是竞争非常激烈吗？
the wait time like, and is it just really competitive or not?
因为我跟詹森聊过这件事。
Because I talked to Jensen about it.
我说，你是怎么管理这么多大人物、大公司争相购买的？
He said I said, how do you manage all these, like, big egos and names and companies trying to buy stuff?
他说，他们下单，我们就按顺序发货，是的。
And he said, well, they order it, and we give it to them in the order Yeah.
他们就是按顺序订购的。
In which they order it.
真的就是这样吗？
That's Is it really like that?
真的就是这样。
It really is.
对吧？
Right?
是的。
Yeah.
你知道的，他不想陷入偏袒任何人的境地，那样对客户来说显然不是个好做法。
Like, you know, he doesn't wanna be in the position of playing favorites or ally like, that that just seems like a bad place to be with your clients.
或者拿去拍卖。
Or auctioning them off.
是的。
Yeah.
所以你能想象吗？
So Can you imagine?
那会那会
That would that that
那太疯狂了。
That'd be crazy.
是的。
Yeah.
我不确定这对长期业务会有好处。
Don't I'm not sure that would be good for the long term business.
不。
No.
是的。
Yeah.
所以我们的做法是
So so our our our approach is
你可能会遇到一些主权基金过来表示，我愿意出双倍的价格。
You might get some sovereigns coming in and saying, I'll pay double.
是的。
Yeah.
他们对法拉利也偶尔会这么做。
They do that with Ferraris too sometimes.
这些就是计算领域的法拉利。
These are the Ferraris of computing.
对吧？
Right?
某种程度上，确实如此。
In a way, they are.
是的。
Yeah.
还有布加迪。
I Bugattis.
我们的做法是与整个领域的客户合作，寻找真正有趣且符合我们收购要求的公司，以便能够设计出所需债务结构，从而在这个规模上建设基础设施。
Our our approach is to work with clients across the entire space to find opportunities that are really interesting companies that can fit into our contraction requirements, the way we're gonna be able to go out and structure the debt that we require in order to go out and and build infrastructure at this scale.
而且
And
这些债务是如何运作的？
How does all that debt work?
这正是你们擅长的领域。
I that is something that you guys specialize in.
企业债务。
Corporate debt.
我从事风险投资业务。
I'm in the venture business.
人们常说：既然企业债务回报这么好，我为什么还要做风险投资？
People are like, why should I be in venture when corporate debt pays so well?
企业债券市场规模巨大。
Corporate paper is so huge.
我想知道这如何与之契合，比如，人们为数十亿美元的基础设施融资要支付多少利率？
I'm curious how this fits in and, like, what interest rate people are paying on, you know, a billion dollars in infrastructure.
他们为此支付多少利息？
What do they pay on that?
是的。
Yeah.
CoreWeave 在推动这一领域诸多融资模式方面一直是创新者。
So so CoreWeave has really been the innovator around a lot of the financing engines that have come to bear on this.
我们做了首批基于GPU的贷款，我认为这一点很重要，或者我会试着用一种让人能理解的方式解释一下。
We did the first GPU based loans and like, I I think it's important or I'm gonna try to explain this in a way people can understand.
我们的做法是走出去，寻找客户。
So what we do is we go out and we find a client.
我们就以微软为例。
Let's use Microsoft.
你之前提到过他们。
You brought them up before.
对吧？
Right?
微软来找我们，说他们想向你们购买一些计算资源。
And Microsoft comes to us and says, we'd like to buy some compute for you.
我们说，好的。
And we say, okay.
太好了。
Great.
我们要签订一份合同。
We're gonna sign a contract.
一旦我手里有了合同，我就会创建一个东西。
Once I have a contract in hand, then what I do is I create something.
这个名字并不特别有创意。
It's not a particularly creative name.
它叫作‘盒子’。
It's called the box.
是的
Yeah.
对吧？
Right?
我用这个盒子的方式是，把我和微软的合同放进盒子里。
And what I do with the box is I take my contract with Microsoft and I put it in the box.
我去见詹森，购买GPU。
I go to Jensen and I buy the GPUs.
我把它们放进盒子里。
I put it in the box.
我拿我的数据中心合同。
I take my data center contract.
我把合同放进盒子里。
I put it in the box.
现在这个盒子管理现金流，它有一个现金流的分配机制，资金流入和流出都经过它。
And now the box governs cash flow, and it has a waterfall of cash flow that comes into it and goes out of it.
所以它的运作方式是，我先构建计算资源，然后将计算资源交付给微软，由他们向这个资金池付款。
And so the way it works is then I build the compute and then I deliver the compute to Microsoft and they pay the box.
他们不会直接付给我。
They don't pay me.
对。
Right.
资金进入资金池后，第一件事是支付数据中心的费用。
It goes into the box and the first thing it does is it pays the data center.
支付电费。
It pays the power bill.
支付利息和本金。
It pays the interest and the principal.
然后剩下的部分会流回给我们。
And then whatever's left flows back to us.
对吧？
Right?
因此，这是一种结构极为完善、经过时间检验和压力测试的机制，能够以客户票据及交易中的其他抵押品为担保进行融资，这正是为什么CoreWeave这家许多人都没听说过的公司，能够在十八个月内筹集到350亿美元用于大规模建设基础设施。
And so it is an incredibly well structured, time tested, pressure tested vehicle to be able to borrow money against client paper and all of the other collateral around the deal, which is why CoreWeave, is a company that many people haven't ever heard of, was able to go out and raise $35,000,000,000 in eighteen months to build infrastructure at scale.
但重要的是要理解，这个系统内的经济模型使得在五年协议的前两年半内，我们就已经偿清了所有费用。
But what's important to understand is the economics in this box are such that within two and a half years of a five year deal, we have paid for everything.
本金已经还清。
The principal's been paid off.
这个项目，
The well,
本金已经还清。
the principal's been paid off.
利息也已经付清。
The interest has been paid off.
进入这个系统的回报足以让我们在设备层面实现公司收益，这让全球最精明的贷款方——无论是银行、私募股权基金，还是其他任何机构——都相信他们能够实现借贷的唯一铁律：把我的钱还给我。
The return into the box is such that we are able to generate returns to our company at the box level, which gives the most sophisticated lenders in the world, whether it's banks or private equity funds or, you know, whoever, confidence that they're going to be able to achieve the one rule of lending, which is give me my money back.
是的。
Yes.
所以当这种情况发生时，效果会更好。
So Works better when that happens.
所以他们看到这个结构时，会说：哇。
So they look at this box and they're like, wow.
我们非常有信心能拿回我们的钱。
We're really confident we're gonna get our money back.
也许他们想要十个这样的结构。
And maybe they want 10 boxes.
没错。
That's correct.
如果任何一个结构出现亏损，你都可以应对，而且不会那么严重。
And and if any one box goes upside down, you can deal with it and it's not as acute.
没错。
That's correct.
而且它们之间不会相互影响。
And they don't cross pollinate.
它们不会在各个箱子之间引发传染。
They don't cause a contagion across the boxes.
它们都是独立且互不干扰的。
They're all independent and discreet, one.
第二，当你这样做，并向贷款方展示这种融资工具和融资机制如何运作时，他们会持续以越来越低的利率向你放贷。
And number two is as you do this and as you show the lenders how this financing tool and how this financing mechanism works, what they do is they continue to lend you money at progressively lower rates.
因此，当你回顾过去两年我们的资本成本时，我们已经将资本成本降低了600个基点。
And so when you think about our cost of capital over the last two years, we have dropped our cost of capital by 600 basis points.
哇哦。
Wow.
这太惊人了。
It is enormous.
对吧？
Right?
因此，你看到的是一家正在将资本成本逐步降低至超大规模云服务商借贷水平的公司，这将使我们未来能够与它们竞争。我们一直非常严格和细致地维护、滋养和照顾这些箱子，以确保我们能持续获得资本市场支持，从而推动业务发展。
And so you're seeing a company that is driving its cost of capital down towards where the hyperscalers borrow, which will enable us to be able to be competitive with them over time and we have been extremely militant and diligent about feeding, watering, and caring for those boxes so that we continue to have access to the capital markets in a way that allows us to build and drive our business.
这意味着你必须拒绝一些想进入这个盒子的人吗？
It means you have to say no you have to say no to maybe some people who wanna be in the box?
是的。
Yeah.
所以有些客户。
So Some customers.
我们会看一些交易，然后觉得，你知道，他们只想租用GPU一年。
We we look at some deals and we're just like, you know, they wanna buy GPUs for a year.
我一看就认为，这个交易我不能接，因为时间太短，无法摊销成本。
And I look at it and say, I I that's not a deal that I can do because it's too short for me to amortize Yes.
这些支出。
The expenses.
所以我就不会做这种交易。
Or and so I won't do that.
对吧？
Right?
比如，一旦
Like, once
他们可以去找其他供应商，那些供应商可能愿意承担这种风险，并且有额外的产能。
And they can go to another provider who maybe wants to take that risk on who has extra capacity.
完全正确。
Absolutely.
但我们的业务核心其实是风险管理，即如何实现规模化，因为在我看来，在这段供需失衡的时期——全球GPU数量不足以满足人工智能所有应用场景的计算需求——对我和我的公司而言，最重要的是实现超大规模，从而降低我们的资本成本，让我们能够从市场的各个角落获取信息流，包括大语言模型、高频交易、搜索等所有领域。
But our business is really built about around the risk management of being able to get to scale because in my mind, during this period of disequilibrium, during this period where there are not enough GPUs in the world to provide the compute for all of the different use cases in artificial intelligence, the part that's important for me and for my company is to get enormously large so we can drive down our cost of capital so that we have information flow coming in from all different parts of the market, the large language models, high speed trading, search, all of these things.
这些领域都在向我们反馈信息，帮助我们了解下一个需要开发的产品是什么，或者客户在扩展时需要什么帮助，以及他们需要哪种类型的计算资源。
They're feeding they're feeding information back into us that is letting us know what the next product we need to build is or where, you know, they need help scaling or what type of compute they need.
所有这些信息流对我们来说都极其宝贵。
And all of that information flow is incredibly valuable to us.
关于需求，你能告诉我们些什么？
What what can you tell us about demand?
有报道称，说……
There's been reports of, hey.
也许是甲骨文与OpenAI的Starbase项目被缩减了，也许并没有。
Maybe the Oracle Starbase thing with OpenAI's been downsized or maybe not.
然后，你知道的，其他公司比如微软和谷歌都在大举投入。
And then, you know, other folks, Microsoft is going big, and Google's going big.
Meta也在大举投入，这些公司显然拥有巨额现金流。
Meta's going big, and those people obviously have massive cash flow.
苹果似乎缺席了。
Apple seems to be MIA.
他们似乎不想参与。
They don't seem to wanna play.
你提到了很多拥有庞大资产负债表、有能力推动巨大需求的大公司。
You you you've you've you've named a lot of really big companies with really big balance sheets that have capacity to drive a lot of demand.
听好了，多年来我一直坚定不移地坚持这一点。
Look, I I have been truly steadfast in this for years now.
四年来，我们所提供服务的需求一直持续不断，远远超过了全球提供足够算力以满足所有人工智能需求的能力。
For for for four years, the depth of the demand for the service we provide has been relentless and overwhelms the global capacity of the world to deliver enough compute to enable all of the demand for artificial intelligence to be stated.
展开剩余字幕（还有 480 条）
我们一直对此毫不松懈。
And that has been we have been relentless about that.
听起来像是帕特里克·尤因时代抢不到票。
Have Sounds like snick tickets during the Patrick Ewing era.
是的。
Yeah.
他们排队人数一度达到了五万人。
Like, they they got up to 50,000 people on the wait list.
所以，如果奇迹发生，排队消失了，如果限制被解除，我们拥有了大量的GPU、充足的能源和充足的数据中心，系统中会突然释放出多少产能？我应该说，会部署多少产能？
So if magically, wait list went away if the if the constraint went away and we just had an a large amount of GPUs available, a lot of energy available, a lot of data center available, how much capacity would just all of a sudden come out of the system or would be deployed, I should say?
所以，记得我们是如何通过这个五年周期来构建业务的。
So remember how we build
我们的业务是通过这个五年周期构建的。
our business through this box, and it's a five year box.
所以，如果我们遇到空档，如果需求因为技术突破或战争而突然消失。
So if we had an air pocket, if if demand were suddenly to disappear because of a technology breakthrough, because of a War.
战争，或者其他任何事情，对吧？
A war, anything, right?
从风险管理的角度来看，原因并不重要。
Like like the why from a risk management perspective does not matter.
你必须为公司做好准备，应对‘如果发生’的情况。
You have to prepare your company for the what happens if it happens.
是的。
Yeah.
因此，通过签订这些长期合同，与资产负债表雄厚的对手方合作，我们正在保护自己和我们的贷款人。
And so by entering into these long term contracts, into entering into contracts with counterparties that have large balance sheets, you are or we are protecting ourselves and our lenders
对。
Yeah.
因此我们有信心，他们也有信心，因为你可以看到，他们向我们收取的利率持续下降，这表明他们最终一定能收回资金，而这正是借贷的唯一准则。
So that we are confident, and they are confident because you can see how confident they are by the rate that they're charging us continuing to decline, that they're ultimately going to get their money back, and that is the one rule of lending.
是的。
Yeah.
所以，你知道，如果只是
And so, you know, if But just
从产能角度来看，如果我们没有限制，而英伟达的黄仁勋说：
in terms of the capacity, if you if we're unconstrained and Nvidia Jensen says, hey.
想订多少就订多少。
Order as many as you want.
那会发生什么？
What would happen?
所以，也很重要的是要明白，限制因素不仅仅是GPU。
So the the it's also important to understand the constraints aren't just GPUs.
对。
Right.
对。
Right.
电力。
Electricity.
是电力，是的。
It's it's power Yeah.
是内存。
It's memory.
是存储。
It's storage.
是网络。
It's it's networking.
是光器件。
It's optics.
所有这些方面。
All of the things.
而且存在各种限制
And there's there's various throttles
内存现在就是一个限制因素。
that will limit the Memory is a throttle right now.
对吧？
Right?
哦，是的。
Oh, yeah.
确实是。
It is.
哦，是的。
Oh, yeah.
确实是。
It is.
为什么内存会成为瓶颈？
Why how did memory become the throttle?
如果内存，而且它历来是一个周期性行业。
If memory and it has historically been a cyclical business.
对吧？
Right?
我们已经看到需求浪潮推高了内存成本，然后崩溃，接着又再次上涨。
We have seen these waves of demand driving up the cost for memory and then it collapses and then it drives up.
这是一个典型的繁荣与萧条交替的行业。
It's a very boom and bust business.
这种周期性源于制造工厂资本密集的特性：人们投资建厂，大量扩充产能，一旦出现这种情况，就会过度建设，而我们已经多次见证过这种循环。
It's cyclical in its nature because the fabs are so capital intensive that people invest in the fabs, build a ton of capacity, and then overbuild if And there's any type of that we've seen that cycle again and again.
目前发生的是两股力量的交汇。
What's happening right now is the confluence of two things.
对吧？
Right?
一是人工智能需求激增，随之而来的是对计算能力以及GPU相关配套服务的庞大需求，
One is is with all the demand for artificial intelligence and the corresponding demand for compute and the ancillary services around the GPU,
的
the
需求飙升到极点。
demand is through the roof.
这是第一点。
That's number one.
第二点是，2023年可能已经到了需要进行投资周期的时候了，明白了。
Number two is is that there was probably an investment cycle that needed to happen back in 2023 Got it.
这将带来必要的晶圆厂产能来满足需求。
That would have brought on the necessary fab capacity to be able
去服务？不可能预测刚刚发生的事。
to serve Impossible to predict what just happened.
就能源而言，根本无法预测刚刚发生的事。
Just with energy, it's impossible to predict what just happened.
现在人们正在追逐能源。
And now people are chasing energy.
数据中心正在向能源丰富的地方迁移。
The data centers are going where the energy is.
这并不是基于房地产。
It's not based on real estate.
这取决于是否有风？
It's based on It's and and is there some wind?
当你有非常资本密集型的时候，不行。
And anytime you you have a very cap no.
不是每次都是这样。
Not every anytime.
但很多时候，当你从事像建造晶圆厂这样资本密集型的业务时，就会出现类似能源行业的繁荣与萧条周期。
But many times when you have a a capital intensive business like, you know, building fabs, you will get this boom and bust cycle just like in energy.
他们过度建设，是的。
They overbuild Yeah.
然后，你知道，光纤。
And then, you know Fiber.
是的。
Yeah.
我的意思是，这种情况有很多例子。
I mean, there's there's there's a lot of examples of that.
我们的方法
Our approach
从某种角度看，这种繁荣与萧条的周期正是资本主义的美妙之处，我们能够应对它。
Some ways when you look at that, it's a beautiful aspect of capitalism that we're able to have a boom bust cycle, that we're able to weather it.
对吧？
Right?
如果你从第一性原理来看资本主义，像光纤过剩这种情况就会发生，这为谷歌或下一家公司收购所有资产创造了机会。
If you think just like capitalism from first principles, something like that happens and we have too much fiber, it creates an opportunity for Google to buy it all up or the next person.
听我说。
Listen.
你知道，繁荣与萧条的周期确实带来了很多影响。
The the the you know, it it does it does a lot of things having a boom bust cycle.
它清理了那些低效的参与者。
It clears out the underbrush.
是的。
Yeah.
最强的公司能够生存下来并利用这一机会，从而为未来的商业奠定基础。
The strongest companies will be able to survive and take advantage of that, and it sows the seeds of future business.
它做的另一件事是将基础设施铺设到位。
The other thing that it does is you put that infrastructure into the ground.
你把光纤埋入地下，这成为了我们日常看电影、沟通、使用Zoom以及应对新冠疫情等所有这些活动所依赖的基础架构。
You put the fiber into the ground, which became the backbone of how, you know, we watch movies every day and how we, you know, communicate and how we hop on a Zoom and, you know, COVID and all of these things were based on that infrastructure that was available to be consumed.
是的。
Yeah.
人们并没有意识到这个事实。
People don't recognize this fact.
关于YouTube的创始理念，我认识的查德·赫利和他的合伙人，他们意识到，在这个阶段，存储成本正迅速下降，因此我们可以提供免费且无限的上传服务。
If you the the premise of YouTube from the founders who I knew, Chad Hurley and his other partner, they basically had the realization, at this curve, storage is coming down so quickly, we could offer free unlimited uploads.
带宽成本也在下降，所以我们不必向用户收取在线分享视频的费用。
And bandwidth is coming down, so I guess we don't have to charge people for sharing a video online.
在此之前，如果你的视频走红，人们会震惊不已，但你的服务器会因此宕机，并提示：‘这个人需要支付账单了。’
Before that, if your video went viral, people are gonna have their minds blown, but your server would turn off and it would say, this person, you know, needs to pay their bill Yes.
因为他们每传出一兆字节的数据就要被收费。
Because they were getting charged for carriage by the megabit going out.
是的。
Yes.
我的意思是，你看。
I mean, it look.
而且，你知道，这些商业模式确实发生了变化，是的。
And and, you know, these these the business models change Yeah.
并且在演进。
And evolve.
而且，正如你所说，摩尔定律，当然詹森也会谈到，在加速计算领域发生的一切远远超过了——是的。
And, you know, like you said, Moore's Law and and and certainly Jensen will talk about the fact that, like, what what is going on within the the the accelerated compute space dwarfs Yeah.
摩尔定律。
Moore's law.
对吧？
Right?
所有这些都将带来更多的机会，催生更多像YouTube那样彻底改变世界的公司。
And all of that is going to lead to more opportunity to build more companies that are going to do things like YouTube did, which has really changed the world.
是的。
Yeah.
是的。
Yeah.
我的意思是，我不确定是不是每小时或每分钟都有百万小时的视频上传，但后来苏珊·伍吉奇（愿她安息）告诉我，每分钟上传的视频量有多大，这听起来完全不合逻辑，直到你意识到，世界上有30亿人使用这项服务，其中1%甚至0.1%的人上传内容就够了。
I mean, the the concept that I I don't know if it was, like, a million hours being uploaded every hour or minute, but at some point, Susan Wojciech, rest in peace, said told me just, like, how much was being uploaded every minute, and it made no logical sense until you realized, well, there's 3,000,000,000 people out two or 3,000,000,000 people in the service and 1% upload or point one 10 bips upload.
哦，原来是这样。
It's like, okay.
每千个人里有一个人上传。
One in a thousand people upload.
这是一个很大的分母。
It's a big it's a big denominator.
就像
Like
我当时和OpenAI的首席财务官莎拉·弗莱尔一起参加了一个小组讨论
I I was sitting on a a panel with Sarah Fryer, CFO
是的
Yes.
她时不时会分享一些非常有趣的信息
OpenAI, and she every every once in a while, she she really puts out, like, interesting information.
她谈到ChatGPT-3刚发布时，一百万令牌的成本是32美元零钱
And so she was talking about the cost of a million tokens when ChatGP three came out, and it was $32 in change.
而现在，一百万令牌的成本只有9美分
And now a million tokens cost 9¢.
是的
Yeah.
对吧？
Right?
所以你就能看到，资本市场和资本主义如何推动工程和技术的发展
And so you you just see, like like, the incredible power of how the capital markets, how capitalism is fueling engineering and fueling
现在这已经变成一种递归了。
And it's become recursive now too.
我的意思是，如果你对这些模型说：嘿，
I mean, these models if you say to the model, hey.
让自己更高效，少花钱，降低令牌成本，它们就会说：好的，船长。
Make yourself more efficient, spend less money, and lower the cost of tokens, you'd be like, okay, captain.
是的。
Yeah.
我不知道你有没有看到卡帕西的递归
I don't know if you saw Carpathi's recursive
对。
Yeah.
上周末那件事，现在连从未接触过语言模型或计算机科学的普通人也开始说：这个周末我要试试做点递归的东西。
It was thing last weekend, but it's like now civilians who've never worked in a language model or doing computer science are like, I'm gonna try to do something recursive this weekend.
你知道，这是我和其他创始人聊天时提到的一件事，当你思考人工智能所做的一些事情时，它确实降低了运营的门槛。
You you know, it's one of the things that I that I talked to, you know, the other founders about, you know, and it's like when you think about some of the things that AI does, right, it's lowering the barrier to operations.
所以，如果你有一个好点子或绝佳的创意，你可以打开你的模型，告诉它，你可以用直觉编程，做各种各样的事情，创造出前所未有的东西。
So if you have a good idea or a great idea, you can open up your model and you can tell your model, you can vibe code it, you can do all kinds of different things and create things that never existed before.
这太棒了。
That's amazing.
对吧？
Right?
这正在打破一个长期限制人类创造力的壁垒，现在突然间，整个全新的可能性领域被打开了，比如医学研究，或者棒球卡收藏，或者你感兴趣的任何领域。
Like that's bringing down this incredible barrier that kept human creativity contained and now all of a sudden this whole new vector of, you know, medical research or different approaches to, you know, baseball cards or whatever you want.
如果你有一个伟大的想法，一个新颖的创意，这现在就是那个宝贵的核心，能让你构建和创造新事物。
If you've got a great idea, if you've got a new creative idea, that's the valuable kernel right now that allows you to to build new things and to create new things.
我只是觉得这令人无比兴奋。
And I just think that's incredibly exciting.
是的。
Yeah.
这就像是给了全球80亿人的头脑一个工具，让他们能够突破那些曾经看似不可逾越的障碍。
It's like you're bringing the minds of 8,000,000,000 people a tool that allows them to overcome what was insurmountable for forever.
为了人类。
For humanity.
是的。
Yeah.
这是一个光明的新未来，迈克尔。
It's a bright new future, Michael.
感谢你与我们分享这些信息和愿景。
Appreciate you sharing the information with us and the vision.
我非常高兴能邀请阿拉文德·斯里尼瓦斯来到节目。
I am really delighted to have Aravind Srinivas on the program.
谢谢你邀请我，杰森。
Thank you for having me here, Jason.
太棒了。
It's so great.
我想谈谈我爱上你们产品的三个阶段。
I wanna go through three stages in which I fell in love with your product.
第一阶段是，我可以自由选择我的语言模型，无论是想用OpenAI、Quad，还是其他任何模型。
The first phase was I could go and pick my language model if I wanna choose OpenAI, if I wanna choose Quad, whatever it was.
这对我来说真是一个巨大的突破。
That was, like, a real unlock for me.
在侧边栏里，我注意到你做了类似雅虎早期做的内容，比如金融、体育。
And on the sidebar sidebar, I noticed you had done essentially, like, what Yahoo did in the early days, finance, sports.
当我打开我的尼克游戏时，它给我展示了实时版本。
And when I pulled my Nick game up, it gave me a live version of that.
当我查看我的股票时，它实时总结了新闻，我当时就想，哇。
When I pulled my stocks up, it summarized the news in real time, and I was like, wow.
这个执行效果太棒了。
This execution is great.
我差不多把你当成了我接入不同模型的入口，让我更容易查看和使用。
And I I kinda made you my front door to different models, and it made it easier for me to check it.
然后你们推出了评论浏览器，我当时就想，天哪。
Then you came out with the comment browser, and I was like, holy cow.
我可以给它一系列指令。
I can give this a series of instructions.
去我的LinkedIn。
Go to my LinkedIn.
找出这家公司里的所有人。
Find everybody from this company.
把他们放进一个Google表格，搞定。
Put them into a Google Sheet, and boom.
你是第一个推出这个功能的。
You are the first out of the gate with that.
就在过去几周，我一直在用OpenClaw，但你们推出了Computer。
And then just the last couple of weeks, I had been claw pilled and using OpenClaw, but you came out with computer.
于是我开始用Computer，天啊，真不错。
And I started using computer, and, boy, it's good.
它起步非常出色，让我能轻松完成重复性任务，某种程度上类似于Claude的Co-Work，或者工程师/开发者在使用它。
It's a really strong start, allowing me to do repetitive tasks, very similar in some ways to co work from Claude or basically an engineer or developer using it.
所以这些是公司的发展历程吗？
So are these the evolution of the company?
我应该从这个角度来思考。
And I should think about it that way.
但你现在怎么看Perplexity？
But how do you look at Perplexity now?
你们拥有非常忠实的粉丝群体。
You have a very loyal fan base.
你们赚了很多钱。
You're making a lot of money.
我不知道你们是否公开过，但我认为是数亿到数十亿美元的规模。
I don't know if you disclose it, but I think it's hundreds of millions to billions.
你可以告诉我们。
You can tell us.
但当Claude表现强劲、OpenAI依然强势、Grok表现优异、Gemini迅速崛起时，Perplexity该如何应对呢？
But what is Perplexity in the face of, wow, Claude's having a great run, OpenAI still doing strong, Grok doing very well, Gemini coming on strong.
你们一共有六七个人，而你恰好是我目前最顶尖的两位之一。
There's, like, six or seven of you, and you just happen to be one of my top twos right now.
谢谢。
Thank you.
所以告诉我吧。
So tell me.
首先，谢谢你。
First of all, thank you.
非常感谢。
Thank you so much.
Perplexity 一直为那些追求额外优势、充满好奇心的人打造，因此你成为我们的核心用户非常自然。
Perplexity has always been built for people who are always looking for the extra edge, the curious people, so it's very natural that you are one of our power users.
在过去三年半里，对我们来说一个共同的主题是准确性。
One common theme for us for the last three and a half years is accuracy.
Perplexity 希望成为打造最准确人工智能的公司。
Perplexity wants to be the company that's building the most accurate AI.
所以当你想给人答案时，准确性对于建立信任至关重要，因为只有这样，用户才会继续提出下一组问题。
So when you wanna give somebody answers, accuracy is very essential for building trust because only then the user is gonna ask the next set of questions.
事实证明，让AI访问互联网以确保准确性是个绝佳的主意。
It turns out it was a great idea to give AI access to the Internet to be accurate.
这就是Perplexity Ask产品。
So that's the Perplexity Ask product.
事实证明，让AI完全访问浏览器也是一个极好的主意，这样当您委托它完成您自己在浏览器上会做的操作时，它就能保持准确。
It turns out it's a great idea for AI to have full access to a browser so that it can be accurate when you task it to go do something that you would do yourself on a browser.
代理式浏览。
Agentic browsing.
Comet。
Comet.
最后一个阶段是，事实证明，让AI获得对计算机的完全访问权限是个绝佳的主意，这样它就能像您一样独立完成计算机上的所有操作，本质上成为计算机本身——一个汇集了今天AI所有能力的交响乐团，无论是GPT、Claude、Gemini还是其他任何模型，所有这些能力融为一体，这就是Perplexity Computer。
Now the last phase is it turns out it's a great idea for AI to give it be given a full access to a computer so that it can do whatever you do on a computer on its own, essentially becoming the computer itself, an orchestra of everything AI can do today, every single capability each individual AI model has, be it GPT or Claude or Gemini or anything else, an orchestra of all those capabilities, that that's what Perplexity Computer is.
而在计算机内部运行的所有这些子代理，就是乐手。
And all these sub agents that are running inside computer are the musicians.
这些模型本质上就是乐器，目前有数百个模型，每个都有自己的专长。
The models are essentially the instruments, and there are, like, hundreds of models out there, each having their own specialization.
有些擅长编程。
Some are good at coding.
有些擅长写作。
Some are good at writing.
有些擅长多模态、视觉合成、图像生成、视频生成和音频处理。
Some are good at multimodal, visual synthesis, image generation, video generation, audio.
但重要的是最终的输出，也就是你演奏出的音乐。
But what matters is the end output, the music you play.
这就是AI为你完成的工作，也是Perplexity Computer的核心。
That's the work AI gets done for you, and that's what Perplexity Computers.
AI本身现在就是计算机。
The AI is itself is the computer now.
仍然运行在浏览器中。
Still lives inside of a browser.
你有没有考虑过给它桌面的根权限？
Have you considered giving it desktop root access?
这似乎是接下来的发展方向，但这会带来很多安全问题和信任问题。
That feels like the next place this is going, but that comes with a lot of security issues, a lot of trust issues.
正如你所说，信任是至关重要的。
As you mentioned, trust is paramount.
获得正确的答案能建立信任，但同样重要的是不能被黑客攻击，也不能让它删除你的文件。
Getting the right answer is what builds it, but also not getting hacked and not having it delete your files.
是的。
So Yes.
你怎么看待对我的 Windows 机器授予根权限？
How do you think about root access to my Windows machine?
对。
Correct.
显然，iOS 不会允许你这样做。
Obviously, iOS, they won't let you.
但安卓手机是允许你这样做的。
But with an Android phone, it would let you.
是的。
Yes.
所以你们正在开发这个功能吗？
So do you have that in the works?
对。
Yes.
所以我们发布了一款名为 Perplexity 个人电脑的产品。
So we announced something called personal computer, Perplexity personal computer.
这本质上会将 Perplexity 电脑的所有信任机制、可靠性以及服务器端执行能力，与你的本地电脑同步，让你可以通过手机使用它。
That's essentially gonna take all the trust and reliability and the server side execution of Perplexity computer, but synchronize it with your local computer so that you can use it from your phone.
我们将会通过 Mac mini 实现这一点，让你的电脑与 Mac mini 同步。
And we're gonna do this with the Mac mini where you synchronize your computer with the Mac mini.
这样，它就成为了你的本地服务器。
So that becomes your local server.
所有与你的本地私有数据相关的代理编排都将运行在该本地编排循环和Mac mini的运行时环境中。
All the agent orchestration that has to do with your local private data will run on that local orchestration loop, that runtime with the Mac mini.
不是在你们的服务器上，也不是在Anthropix上。
Not on your servers, not on Anthropix.
没错。
Exactly.
是的。
Yeah.
它仍然可以在获得你许可的情况下调用前沿模型，但所有编排工作都会在你的本地硬件上进行。
It could still ping frontier models if it needs to with your permission, but it will be orchestrating everything on your local hardware.
是的。
Yeah.
如果需要在服务器端硬件上运行，而你又不希望太复杂、运行时间过长的统计任务在你的本地硬件上执行，是的。
And if it needs to run on the server side hardware, if you don't want very complicated, long running stats to be running on your local hardware Yeah.
你可以将其委托给你的服务器端计算机运行，而这台计算机同样仅对你本人开放。
You can delegate it to run on your server side computer, which is, again, only accessible to you and you alone.
这样，我们将实现一种完美的混合模式，在本地和服务器端之间取得平衡。
So that way, we're gonna bring this perfect hybrid of trustworthycom hybrid between local and server side.
而且你会让它变得很容易操作。
And you And you'll make it easy to do it.
只需抽象化即可。
Just be abstracted.
你安装一个可执行文件。
You install one executable.
搞定。
Boom.
就完成了。
It's done.
这就像给新手用的OpenCloud。
It's it's like OpenCloud for dummies.
没人需要学习怎么使用它。
Nobody needs to learn how to use it.
没有人需要管理API密钥。
Nobody needs to manage API keys.
没有人需要管理跨越100多个服务的独立计费，也不用纠结哪些功能可以开放访问、哪些不能。
Nobody needs to manage separate billing across, like, 100 different services, figure out what you can give access to and not access to.
这些我们都帮您处理好了。
We take care of that.
这就像是史蒂夫·乔布斯的方式，端到端的整合。
So it's a Steve Jobs way of doing it, you know, end to end integration.
那你怎么看待本地模型呢？
And and how do you think about local models?
我已经在Mac Studio上运行Kimi 2.5了。
I have started running Kimi 2.5 on a Mac Studio.
它没有Claude、Gemini或Grok那么好，但你大概能免费获得80%的性能。
It's not as good as Claude or Gemini or Grok, but you can probably do about 80% there for free
是的。
Yeah.
本质上是这样。
Essentially.
对。
Yeah.
考虑到我其他的账单，比如Claude之类的费用越来越高，这确实很有吸引力。
And so that's quite compelling considering some of my other bills, Claude, and and stuff were getting expensive.
那你有吗？
So do you have one of those?
你开始在你的本地Mac Studio上测试了。
You started testing on your local Mac Studio.
我猜你有一台Mac Studio，并且你自己在做这个。
I assume you have a Mac Studio, and you're doing this yourself.
是的。
Yeah.
或者
Or
现在，我不知道你有没有看到戴尔和英伟达发布了一款大型工作站。
now, I don't know if you saw Dell and Nvidia announced a giant workstation.
是的。
Yeah.
是3800美元吗？
Was it a 3,800?
差不多吧。
Something like that.
差不多吧。
Something like that.
是的。
Yeah.
它有750吉字节的内存。
It's 750 gigs of RAM.
那么，你怎么看待桌面电脑回归工作站/服务器方向？
So what do you think about the desktop going back to workstation slash server
是的。
Yeah.
状态如何？
Status?
我觉得这非常有前景。
I think it's very promising.
我的预测是，它最初会作为一个子代理启动。
My my my prediction is it initially start off as a sub agent.
无论你需要处理什么——比如报税、个人照片、邮件、日历、各种本地应用、私人笔记——这些非常私密的内容，如果你注重隐私，完全可以确保访问这些数据的模型运行在你的本地硬件上。
So whatever you need to go, like your tax returns, your personal photos, your emails, your your calendar, all that stuff, those local apps, your personal notes, very personal notes, you could make sure that the models that access those tokens will be running on your local hardware if you want to, if you're that privacy conscious.
而更复杂、需要访问服务器端已有数据的操作，比如你的谷歌日历，是的。
And more complicated stuff that accesses your data that's already on the server side, example, your Google Calendar Yeah.
你的Gmail。
Your Gmail.
这仍然是个人数据，但AI运行时可以通过你的连接器访问这些数据，比如你的谷歌日历连接器、谷歌工作区连接器。
This is personal data still, but AI runtime can access that through your connector, your Google Calendar connector, your Google Workspace connector.
这可以在服务器端运行，因为数据本来就在服务器上。
And that could run on the server side because, anyway, the data is on the server.
它甚至根本不会存在于你的设备上。
It's not even lying on your device.
所以这种混合编排正是我们未来的发展方向。
So that sort of hybrid orchestration is where we are headed to.
我不认为这完全是本地与完全服务器端之间的二选一。
I don't think it's a dichotomy between fully local versus fully server.
关键在于选择。
It's all about choice.
而且，不管你用手机时，其实你并不关心这个计算任务在哪个服务器上运行，因为根本不可能在手机上运行。
And, anyway, when you're on your phone, you wanna you don't care actually which server that workload's running from because it's not gonna be able to run on your phone anyway.
芯片必须存在于Mac Studio、Mac Mini或服务器上。
The chips need to exist on a Mac Studio or a Mac Mini and or on the server.
或者即将到来的这款新戴尔电脑。
Or this new Dell that's coming out.
我真的觉得，花一万美元购买一台高性能台式机这个想法会吸引很多人，因为它能降低他们的
And I I really think the idea of spending $10,000 on a powerful desktop will appeal to people if it lowers their
对。
Right.
每月500美元的
$500 a month
是的。
Yes.
欺诈账单。
Fraud bill.
是的。
Yes.
这是一笔巨大的节省，而且你还能获得隐私保护的好处，是的。
This is an incredible savings, plus you get the benefit Yes.
无需用你的个人数据来训练语言模型。
Of privacy and not educating the language models on your personal data.
是的。
Yes.
而且这将会像你购买冰箱或互联网调制解调器一样。
And it's gonna be it's gonna be like you're buying a refrigerator, your your your your Internet modem.
这些设备的成本最终会下降。
Like, the cost for these will eventually go down.
对。
Yeah.
但你不会觉得是在浪费钱。
But it's not gonna feel like you're wasting your money.
每个家庭都有很多其他传感器，对。
Every a every home has a lot of other sensors Yeah.
它们控制着你的家居，也将成为这种协同运作循环的一部分。
That runs your home that'll also be part of this orchestration loop.
对。
Yeah.
所以，这正是它变得令人兴奋的地方，因为现在你只需对手机下达指令，就能控制整个家。
So so that's where it gets exciting because now you can just dictate something to your phone, and that can control your entire home.
这就是每个人的梦想，所有这些协调循环都可以在你的本地硬件上运行。
So that's the dream that everybody has, and all that orchestration loop can run on your local hardware.
没问题。
No problem.
我很好奇，你对操作系统有什么看法？
And I'm curious what you think of the operating system.
最终，这个工作站的操作系统会是什么？
What's eventually going to be the operating system of this workstation?
人工智能就是操作系统。
AI is the operating system.
在传统的操作系统中，你之前是通过编程来执行任务的。
Like, earlier in the traditional operating system, you execute programmatically.
现在，你从目标出发，而不是具体的指令。
Now you start with objectives, not specific instructions.
对。
Right.
你提出一个高层次的目标。
You come up with a high level objective.
去帮我建一个网站，把所有播客的 transcripts 都收集起来，跟踪播客前后股票价格的变化。
Go build this website for me that, you know, takes all the transcripts of all in podcast and tracks the stock price just before the podcast and after Yeah.
并为前七天绘制图表。
And chart it for the max seven.
是的。
Yeah.
并且随着时间推移绘制图表。
And and chart it over time.
我们可以做到，这就是目标。
We can so that's the objective.
但单独来看，它在运行文件系统、代码沙箱，并访问互联网。
But individually, it's running a file system, a code sandbox, access to the Internet.
它有自己的HTML工具之类的，是的。
It's having, like, its own HTML tools and, like Yeah.
所以我认为，这基本上就是模型、系统、文件和连接器融合的地方。
So I think that's basically where, you know, models, systems, and files, and connectors are all coming together.
你会把它看作一个操作系统，但你所操作的抽象层级更高，你思考的是目标。
You would think of that as an OS, except you're operating at an abstraction about that where you are thinking in terms of objectives.
是的。
Yeah.
在你看来，它最终是否需要成为一个独立的操作系统？
And does it need to eventually become its own operating system in your mind?
有可能。
It could be.
比如，人们可能会觉得，是的。
Like, people could think about it as just like, yeah.
我有一台Perplexity电脑一直在运行。
I have my per Perplexity computer running all the time.
现在它基本上运行在Linux机器上。
Whether it essentially, it runs on Linux machines right now.
所有的服务器端计算机都是Linux机器。
Every server side computer is a Linux machine.
是的。
Yeah.
所以我认为马克在我们发布后不久发了一条推文，说Linux计算机确实是正确的选择。
So I think Mark in recent tweeted this right after our release that turns out Linux computers was the right idea.
桌面版的Linux计算机终于要行得通了。
Desktop desktop Linux computers are finally gonna work.
是的。
Yeah.
我的意思是，它们已经很稳定了。
I mean, they're stable.
是的。
Yeah.
它们是可以自定义的。
They're customizable.
没错。
Exactly.
而且你不会受苹果公司想要控制体验的限制
And you're not at the mercy of Apple's desire to contain the experience
没错。
Exactly.
或者微软Surface对黑客的限制。
Or Microsoft Surface area as for hackers.
没错。
Exactly.
你打造一个坚如磐石的东西，确实感觉Linux可能会真正成为
You build something rock solid, and it does feel like Linux might actually become the
正确。
Correct.
最终的胜者。
The eventual winner.
它可能不需要有前端界面。
It may not need to have a front end.
这就是关键。
That's the thing.
你可以通过手机访问Linux机器。
You could you could access the Linux machine on your phone.
对。
Right.
可以运行iOS或Android系统。
Could be running iOS or Android.
无所谓。
Doesn't matter.
对。
Right.
实际有价值的运行时环境是在服务器上的Linux上运行的。
The actual valuable runtime is running on Linux on the server.
作为一家消费公司，你做得非常出色。
You've done great as a consumer company.
那里充满了热情。
Lot of love there.
现在我开始看到一些企业级公司也开始参与进来。
Now I'm starting to see corporations with computer starting engaging.
事实上，你会很高兴听到这个消息。
And in fact, you'll be happy to know this.
上周，我让后办公室的两个人停下了OpenClaw的工作。
Last week, I took two people in my back office, and I said, stop working on OpenClaw.
你们的任务是仅使用Perplexity为我们的风投公司实现后办公室自动化。
Your job is to do the back office automation at our venture firm only using Perplexity.
他们说：Perplexity电脑。
And they were like Perplexity computer.
他们就说，哦，好吧。
And they were like, oh, okay.
它在 Slack 里对话效果不好。
It doesn't talk well in Slack.
它在 Slack 里没有智能代理。
It doesn't have an agent in Slack.
我说，它会有的。
I was like, it will.
我打算去见一下阿拉文德。
I'm gonna see Aravind.
我会跟他谈谈这件事。
I'll talk to him about that.
所以我们需要一个非常强大的 Slack 连接器。
So we need a really strong Slack connector.
它已经发布了。
It's already out.
是的。
It is.
好的。
Okay.
太好了。
Great.
目前，Computer 作为一个 Slack 机器人存在。
Computer exists as a Slack bot right now
好的。
Okay.
你可以将其添加到企业版 Slack 工作区，我们整个公司都是这样运作的。
That you can add to your Slack workspace on enterprise plan, and our entire company works like that.
人们在 Slack 上与 Computer 的交流比与其他人的交流还多。
People are talking more to computer on Slack than other than other people.
在我们的第一个 Volley 中，我们发送报告，但不具备交互性。
In our first Volley, we were sending reports in, but it wasn't interactive.
这太完美了。
That's perfect.
所以你现在让公司朝着两个不同的方向发展。
So now you've got your company going in two different directions.
你们这个出色的面向消费者的产品，每个月有多少人使用？
This incredible consumer run you have, how many people are using the product every month?
几千万人。
Several tens of millions.
所以是几千万人。
So tens of millions of people.
这与谷歌和雅虎的消费者业务发展轨迹非常相似。
That's very much similar to the trajectory of the Google and Yahoo consumer business.
现在你们有了企业市场。
Now you've got corporate.
你们在企业市场这边做得怎么样？
How are you doing on the corporate side?
数千家公司。
Thousands of companies.
对我们来说增长最快的业务。
Fastest growing business for us.
它的增长速度超过了消费者业务和收入。
It's growing faster than the consumer and revenue.
像电脑解锁这样的功能带来了全新的可能性。
And things like computer unlock entirely new possibilities.
例如，我们为处于企业最高层级的企业客户节省了超过一亿美元。
For example, we've saved more than a $100,000,000 for our enterprise max customers who are on the highest tier of enterprise.
解释一下这是什么。
Explain what that is.
这需要多少钱？
What does it cost?
每人每月200美元？
200 a month per person?
所以有两个层级。
So there are two tiers.
一个是企业专业版，每月40美元，还有一个是企业至尊版，每月400美元。
One is the enterprise pro, which is $40 a month, and there's the enterprise max, which is $400 a month.
在电脑上，当你用完积分后，就需要为令牌付费。
And that that and and and and on a computer, after you run out of your credits, you would pay for the tokens.
你为使用量付费。
You pay for the usage.
你从每月400美元、每年5000美元的这个套餐中赚钱吗？还是说现在人们都疯了？
Are you making money on the $400 a month, $5,000 a year one, or at this point in time, are people going so crazy?
Perplexity的一个特点是，我们所有的收入，与其他一些包装公司不同，都具有正的毛利。
Our one thing that Perplexity has is every revenue we make, unlike certain other wrapper companies, every revenue Perplexity makes has positive gross margins.
明白了。
Got it.
因为我们不只是在卖令牌。
Because we're not just selling tokens.
对的。
Right.
我们大部分收入是经常性的，因为人们在支付订阅费。
Most of our revenue is recurring because people are paying a subscription fee.
由于我们通过多个不同的模型进行路由，因此在令牌使用上非常高效。
And because we route through multiple different models, we are very efficient in terms of how we spend on the tokens.
因为我们拥有RAG、编排和搜索的所有优势，实际上并不需要扩大模型的上下文窗口。
Because we have all this advantage with Rag and orchestration and search, we don't actually need to blow up the context window of the models.
是的。
Yeah.
因此，我们所有的收入都实现了正毛利。
As a result of that, we have positive gross margins on all the revenue.
意思是，我们赚的每一分钱都实现了盈利。
Mean, every single penny we make, we make profits on that.
但整体来看，公司目前尚未盈利，但我们正在朝着这个目标努力。
But the overall, the company is still yet to be profitable, but we're working towards that.
你曾经有机会退出。
You've had the opportunity to exit.
有很多传言。
A lot of rumors.
苹果公司，还有其他人，都说过。
Apple, other people were like, hey.
这是一个很棒的团队。
This is a great team.
现在团队有多少人？
How many people on the team now?
大约400人。
About 400.
是的。
Yeah.
你拥有一支非常抢手的团队。
You've you've got a very coveted team.
你显然理解消费者。
You obviously understand consumer.
你显然理解商业。
You obviously understand business.
这是一家产品驱动的公司。
It's a product driven organization.
报告显示你已经退缩了，但这个世界正变得极度竞争。
Reports are you declined, but the world's getting hypercompetitive here.
当你们只有400人的团队时，如何跟上节奏？看看萨姆·阿尔特曼在那里筹集了1000亿美元，再看看埃隆把数据中心放到太空，还把SpaceX和推特合并。
How do you keep up as a 400 person organization when you got Sam Altman over here raising a $100,000,000,000, you know, and then you have Elon putting data centers in space and merging with SpaceX and Twitter.
你有谷歌，资源无限；亚马逊也入局了；还有Gemini，产品非常强大，而谷歌在消费端确实很出色。
You have Google with unlimited resources, Amazon getting in the game, and, obviously, Gemini, very strong product, and Google, really good at consumer.
我想我们都同意，Facebook和Meta至今还没搞明白，除了可能更精准地给我们推送广告，他们还没找到真正的消费者场景。
I think we'd all agree Facebook and Meta haven't figured it out yet except maybe for serving us better ads, but they they haven't figured out the consumer case yet.
但他们一定会模仿。
But they'll copy it.
他们总是会这么做。
They always do.
你怎么看待这个竞争格局？
How do you look at the playing field?
因为这里的难度不是在下跳棋，而是像同时与世界前十的国际象棋大师对弈。
Because the degree of difficulty this isn't playing checkers or this is like playing against the 10 best chess players in the world.
你每天都必须面对这样的挑战。
That's what you have to do every day.
是的。
Yeah.
那你如何看待这个问题？
So how do you think about it?
从长远来看，作为一个独立公司，你觉得你最终需要联手合作吗？
Long term and independent company, do you think you'll need to join forces at some point?
那么，你当初为什么没有接受那个交易？
Well And why didn't you take the deal?
这些交易简直太棒了，
These deals were incredible that
你当时得到的报价。
you got offered.
我们拥有一个所有你提到的公司都不具备的优势，那就是多模态协调能力。
So one advantage we have that all these companies you mentioned don't have is the multimodal orchestration.
我们就像瑞士一样。
We're like Switzerland.
我们不必押注于某一个模型。
We don't have to have one horse in the race.
无论GPT胜出、Gemini胜出、Quadrant胜出、Llama胜出，对我们来说都没关系，甚至开源模型胜出也无所谓。
If GPT wins, Gemini wins, quad wins, llama wins, it doesn't matter to us, or even open source models can win.
没问题。
No problem.
而且你们把这些都作为服务提供。
And you have them on the service.
我们有DeepSeek和Kimi。
We have DeepSeek and Kimi.
我们有Kimi。
We have Kimi.
我们有Nematron，还有大量使用Quinn，阿里巴巴的Quinn，是的。
We have Nematron, and we have lot of usage of Quinn, Alibaba Quinn Yeah.
在后台默默运行。
Silently under the hood.
对我们来说，优势就在于能够整合每个模型的最佳能力，为用户提供它们所有功能的交响乐。
So for us, like, that advantage of being able to take the best in each model and give the user the orchestra of everything they can do.
我认为你们提到的任何公司都无法做到这一点。
I don't think any of the companies you mentioned can do that.
没错。
Right.
它们也不会这么做。
Nor would they.
他们也不会。
Nor would they.
这对他们来说毫无意义。
It it makes no sense for them.
这等于承认，尽管他们投入了大量数据中心和资本支出，仍无法打造出最好的模型。
It would be an admission that all the data centers and CAPEX they've built out mean still couldn't produce them the best model.
Anthropic的首席执行官Dario最近在一次采访中表示，模型正在走向专业化。
And Dario, CEO of Anthropics, said recently in an interview that models are specializing.
去年年初，人们认为模型会趋于同质化。
Towards the beginning of last year, people thought models are gonna commoditize.
但到了去年年底，模型开始走向专业化。
But towards the end of last year, people models started specializing.
即使在编程领域，Quad Code和Codex的能力也大不相同。
Even within coding, quad code and codex have very different capabilities.
我们的iOS工程师非常喜欢使用Codex。
Our iOS engineers love using codex.
我们的后端工程师喜欢使用 Quad Code。
Our back end engineers love using quad code.
是的。
Yeah.
因此，即使在编程这样的专业化领域内，模型也有各自独特的专长，而在编程之外的许多其他应用场景中，不同的模型擅长不同的任务，这意味着那个无需将所有希望寄托于单一模型的‘指挥家’，可以通过为客户提供每个你提到的优秀模型都无法提供的独特价值和服务而胜出。
So even within a specialization like coding, models have their own unique specialties, and there are many other use cases outside coding where different models are good at different things, which means the orchestra conductor that has no one model to the horse in the race can win by providing a very unique value and service to the customer that each of these amazing names that you mentioned cannot.
所以你是从他们那里批量购买令牌，然后向客户收费，还是认为所有这些都由你来处理？
And so you're buying tokens wholesale from them, and then you'll charge customers to do it, or do you think it's all You're
你会负责所有这些编排工作。
gonna take care of all their orchestration.
对。
Yeah.
因此你无需在不同模型之间管理令牌。
So you don't have to manage tokens across different models.
因为我认证了我几个不同的账户，我的专业账户，Interperplexity。
Because I authenticate a couple of my different accounts, my pro accounts, Interperplexity.
但你是否在抽象化这一点？我没有足够的知识来判断，用户是否可以跨这些模型进行搜索，而这是否是他们Perplexity订阅的一部分？
But does it I I I don't have enough knowledge to know if you're abstracting that and people can just search across them and it's part of their Perplexity subscription?
不。
No.
我们不会把其他AI的订阅捆绑在一起。
We're not bundling subscriptions from into other AIs.
是的。
Yeah.
我们只是直接调用这些模型。
We just ping the models directly.
明白了。
Got it.
你从我们这里获得的是Perplexity或编排功能。
What you get in us is the Perplexity or Orchestration.
明白了。
Got it.
框架。
The harness.
对。
Right.
当模型逐渐走向专业化时，懂得构建优秀框架的人价值更大。
So the when when when when models are kinda specializing, the there's a bigger value in the one who knows how to build a great harness.
对。
Right.
可以整合
That can take the
每个模型的最佳优势。
best in each model.
现在是自动路由吗，还是仍需下拉菜单，让人手动选择？
Does it auto route today, or do you still have the drop down somebody's gotta pick?
它确实会为每个提示自动选择最佳模型，但我们也会让用户自由选择他们想要的模型。
It it definitely auto routes the best model for each prompt, but we also give users the flexibility to pick whatever model they want.
我觉得我见过不少初创公司把这种做法拼凑起来，对多个模型执行相同的查询。
What do you think of I've seen a bunch of startups hack this together, but doing the same query across multiple
我们开发了一个叫模型委员会的东西。
We built this thing called model council.
模型委员会。
Model council.
是的。
Yeah.
这是Perplexity中的一种模式，我看到Jensen在一次采访中说，他会把同一个提示输入五个不同的AI，然后看看每个AI的回答。
That's one of the one of the modes and Perplexity where I saw Jensen say in one of his interviews that he he puts the same prompt in five different AIs and sees what each of them says.
对。
Yes.
每个人
Everybody
都这么做。
does that.
是的。
Yeah.
但你仍然需要运用生物计算能力来阅读每一个答案，是的。
But then you still have to apply a biological compute to read every answer Yes.
然后找出它们之间的差异。
And then figure out where they differ.
向五位律师咨询关于
To five lawyers about
你的信任或五位不同的医生。
your trust or your Five different doctors.
五位不同的医生，试图弄清楚。
Five different doctors and trying to figure it out.
没错。
Exactly.
这太傻了。
It's dumb.
所以，模型委员会是我们开发的一个功能，它不仅会提供每个模型的答案，还会明确指出它们在哪些地方一致、哪些地方不一致，以及细微的差别在哪里。
So the model council is a feature we built where it will not just give you the answers of each model, but it'll tell you exactly where they agree, where they disagree, and where the nuances are.
这个功能在界面里吗？
And that's in the interface?
是的。
Yeah.
最终建议？
Final counsel?
我不知道它在这里。
I didn't know it was there.
它在
It's
那里。
there.
我的意思是，你们发布产品的节奏相当快。
I mean, you you release product at a pretty great cadence.
怎么说 yes。
How do Yes.
你在哪里学到的，
Where did you learn that,
你们的产品发布理念是什么？
and what's your philosophy of shipping product?
我们的理念是，速度就是我们的核心。
Our philosophy is, like, speed is our mode.
你知道的，大公司做不到像我们这么快地行动、以这样的速度服务客户，同时保持质量、速度和信任是非常困难的。
Like, you know, again, one of the things that big companies cannot do is move at the speed we do, serve customers at the speed and qual it's it's very hard to maintain quality, speed, and trust at the same time.
是的。
Yeah.
比如，苹果公司推出任何产品都要花很长时间。
Like, Apple takes a long time to ship anything.
对。
Right.
因为他们非常担心人们不信任他们。
Because they're very worried about people not trusting them.
是的。
Yep.
所以一些公司官僚主义严重，推出产品总是要花很长时间。
And so some companies are bureaucratic, and they just take forever to ship something.
他们不维护自己发布的产品。
They don't maintain what they ship.
他们可能为某个活动大肆宣传，但根本没人知道怎么使用那个功能。
They may make a big deal about an event, but nobody even knows how to go and use that feature.
对。
Yeah.
它们会被遗弃。
They get abandoned.
没错。
Exactly.
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。