推理将成 10 万亿美元市场：Naveen Rao 谈能源、硬件与 AI 基建

Naveen，我太兴奋了，兄弟。我跟你说过我从你那儿偷师了。Matt Bornstein、Matt O'Keefe 等等。非常感谢你今天能来，兄弟。

Naveen, I am so excited for this, dude. I told you I stole the out of you. Matt Bornstein, Matt O'Keefe, many others. Thank you so much for joining me today, man.

当然。是的，我很期待。

Absolutely. Yeah, I was looking forward to it.

是的，我觉得会很有趣，因为我看了你大量的节目。我记下了我认为最有争议的一些话。顺便说一句，有几个很尖锐的。但我其实想从基础设施的基础层开始，也就是能源。你之前说过它将成为瓶颈以及它的重要性。你如何评估当今的能源供应状况？有哪些我们没看到但应该看到的问题？

Yeah, I think it's going to be a lot of fun because I went through a ton of your shows. I wrote down what I thought were the most controversial quotes. There were a few zingers, by the way. But I actually wanted to start on the foundational layer of this infrastructure, which is energy. You've spoken before about how it will be the bottleneck and the importance of it. How do you evaluate the state of energy supply today? And what are we not seeing that we should see?

能源供应一直很难建设，对吧？我觉得人们没有意识到这一点。我的意思是，我见过行业内的数字，建设 1 吉瓦大约需要 500 亿美元。所以，人们把吉瓦这个词随便说，好像很容易。就像，‘伙计们，这不容易。’ 换个角度说，建一个晶圆厂，比如半导体晶圆厂，大约需要 200 或 300 亿美元。所以，1 吉瓦比建一个晶圆厂还贵。如果你从美元和复杂性的角度来看。所以，这很难。我认为过去几年我们一直在吸收过剩的需求。所以我们通常会过度建设，我们不会以 100% 的产出运行基础设施。所以我们为数据中心获取能源的主要方式就是：我们能不能以更高的利用率运行？并从中再挤出几吉瓦。所以，是的，如果你看美国电网，我们平均输出大约是 900 吉瓦，但我认为它有能力峰值达到 1.2、1.3 太瓦。所以，1300 吉瓦。所以，有一些过剩容量，这就是我们主要追求的目标。这是你可以快速获得的。但是，建设基础设施很难，我们需要建设更多。我们落后了。

Energy supply continues to be difficult to build, right? I don't think people appreciate it. I mean, I've seen numbers from general industry is about $50 billion to build one gigawatt. So, people are sort of throwing around the word gigawatt like it's easy. And it's like, 'Guys, this is not easy.' Just to put that in perspective, building a fab, like a semiconductor fab, is on the order of 20 or 30 billion dollars. So, this is 1 gigawatt is more than building a fab. Like, if you look at it in terms of dollars and complexity. So, it's hard. I think what we've seen over the last several years has been soaking up the excess demand that's been around. So, we overbuild generally and we don't run infrastructure at 100% output. So, what we've been doing where we've been getting energy for data centers is primarily been from this. Like, can we run things at higher utilization? And get that next couple of gigawatts out of it. So, yeah, I think if you look at the US power grid, we're on average something like 900 gigawatts of output, but I think it's capable of peaking up to like 1.2, 1.3 terawatts. So, 1,300 gigawatts. So, there is some excess capacity, and that's primarily what we've been going after. That's what you can get access to quickly. But, building infrastructure is hard, and we need to build more. We're behind.

我们落后多少？

How far behind are we?

嗯，我认为过去 20 年，中国在总容量上超过了我们。我认为那发生在 2023 年。但是，他们 20 年来建设速度是我们的 6 倍。他们从远远落后追赶上来，但他们真的又快又狠地做了这件事。我们对此批评很多，因为我说，‘哦，他们用肮脏的方法来做这个。烧煤，这个那个。’ 没错，但他们也在尽可能快地增长，我认为现在他们有点回头清理一些东西。所以，与此同时，我们并没有真正投资能源基础设施，这会伤害我们，因为真正达到目标需要时间。

Well, I think over the last 20 years China surpassed us in terms of total amount of capacity. That happened, I think, in 2023. But, they were building 6x faster than us for 20 years. They were coming back from far behind, but they really went fast and hard at this. And we criticized that a lot because we said, 'Oh, they're using dirty methods to do this. Coal burning, this and that.' True, but they're also trying to grow as fast as possible, and I think now they're sort of going back and cleaning up some of this. So, meanwhile, we didn't really invest in energy infrastructure, and that's going to hurt us because it's going to take time to really get there.

兄弟，中国比我们领先这么多，我们能学到什么，以至于我们应该在政策和监管上更宽松？有什么要点？

Dude, what can we learn from China being so much further ahead of us that we should be like way more loose on policy and regulation? What are the takeaways?

嗯，我认为有些方面很难，因为它触及我们文化和他们文化的核心。我认为中国可以集中规划。我们不能。这两者都有优势，对吧？在这些大规模基础设施推广的事情上，是的，有优势。毫无疑问。然而，我认为在创新方面，最前沿的东西，美国要好得多。所以，AI 诞生在这里，不是那里，你知道，所有这些最新创新都来自这里。现在，他们在中国的某些工作也非常出色。我不想贬低这一点，因为不仅仅是蒸馏。他们真的在做一些真正的创新。然而，我认为真正开辟新领域的最前沿的东西来自这里，而不是那里。所以，我认为我们必须利用我们的优势。归根结底就是这样。

Well, I think some of this is hard because it hits the very core of our culture and their culture. I think China can centrally plan. We don't. And there are advantages to both of these things, right? And in these things where you have these giant infrastructural rollouts, yeah, there's an advantage. No doubt about it. However, I think when it comes to innovation, tip of the sphere sort of things like the US is far and away better. So, AI was born here, not there, you know, like all of these latest innovations have all come from here. Now, they're doing some really great work in China also. I don't want to diminish that because it's not just distillation. They're really doing some true innovation. However, I think that tip of the sphere stuff where you really open up new areas comes from here, not from there. So, I think we have to leverage our advantages. That's what it comes down to.

当你审视当今稀缺的东西时，你认为今天不稀缺但 3 到 5 年后会稀缺的是什么？

When you look at what's scarce today, what do you think is not scarce today that will be in 3 to 5 years?

我认为那种过剩容量将开始变得非常非常稀缺。如果我们看增长率，即使是 AI 更保守的增长率，我们也会在接下来几年内很快耗尽能源。而且，仅仅说‘好吧，我要找更多容量’并不容易。你会用完所有过剩。所以，然后我们就会处于一种收敛到建设速度的状态。我听到有人说，‘哦，我们一年要建 40 吉瓦。’ 不可能。我们从未做到过。我只是不认为这是可能的，至少现在不是。也许几年后我们会想出一些新发明，我们会通过建设来解决。但就目前而言，我们不会一年建 40 吉瓦。也许我们建 10 吉瓦。如果我们建 10 吉瓦，那就太棒了。换个角度说，那是 5000 亿美元。

I think that excess capacity is going to start getting very very scarce. If we look at growth rates, even the much more conservative growth rates of AI, we're going to be out of energy pretty soon in the next few years. And it's not easy just to say, 'Okay, I'm going to find some more capacity.' You will have used up all of that excess. So, then we're in a place where we kind of converge to the rate of building. And I've heard people say, 'Oh, we're going to build 40 gigawatts in a year.' There's no way. We have never done that. I just don't see it as being possible, at least not right now. Maybe in a few years we'll come up with some new inventions and we will build our way out. But as it stands today, we're not going to build 40 gigawatts in a year. Maybe we build 10. If we build 10, that's amazing. That's half a trillion dollars, just to put it in perspective.

所以，当我们看数据周期，已经被谈论为极其波动，我们现在看到它开始级联或在 cool weave 中展开，有很多担忧。它实际上只是开始吗？还是派对在晚上 7 点音乐就停了？

So, when we look at the data cycle that is already being talked about as being incredibly choppy, and we're seeing that now starting to cascade or unpack in cool weave, and a lot of concerns around that. Is it actually just at the start? Or does the music turn off at the party at 7:00 p.m.?

我认为最大的决定因素是对 AI 服务的需求，因此对 GPU 的需求，因此对这个过剩的新能源建设的需求。我的意思是，显然我身处这个行业。我相当相信 AI。我认为有大量需求。我认为我们甚至还没有触及表面。

I think the biggest determinant of that will be is there demand for AI services for therefore for GPUs and therefore for this excess this new energy buildout. I mean, obviously I'm in this industry. I'm pretty AI pilled. I think there's plenty of demand. I don't think we even scratched the surface.

我们在 AI 领域还有很多问题要解决，但随着我们解决这些问题——我们会在未来几年内解决——它只会开辟出我们今天甚至不知道的新需求类型。所以，我的感觉是这只是一个小插曲，而不是某种泡沫破裂。我的意思是，这就像如果你问：“嘿，人们有治愈癌症的需求吗？”当然有，对吧？如果你治愈了癌症，人们会想要那个。如果你建造了一台能做任何事的智能机器，是的，人们会想要那个。

There's a lot of problems we have to solve in AI still, but as we solve those, and we will over the next couple of years, it's just going to open up new types of demand that we don't even really know about today. So, my sense is that this is a blip. It's not some kind of bubble being burst. I mean, it's sort of like if you said, 'Hey, is there demand for people curing cancer?' Of course there is, right? If you cure cancer, people are going to want that. If you build an intelligent machine capable of doing anything, yeah, people are going to want that.

我一直记得维诺德·科斯拉对我说：“我想承担尽可能多的技术风险，尽可能少的采用风险。”

I always remember Vinod Khosla saying to me, 'I want to take as much technology risk and as little adoption risk as possible.'

是的。

Yeah.

正是如此。我有治愈癌症的方法，非常难做到，但采用风险为零。我完全理解。但这是否意味着，如果我们真的面临 18 个月后耗尽的情况，我们今天对过剩供应的定价是错误的？

It's exactly that. I have a cure for cancer, very difficult to do, but there's zero adoption risk. I totally get that. But then does that mean that we're mispricing the excess supply today if we're literally staring down the barrel of 18 months running out?

我的意思是，你可以这么说。我想这取决于能源市场，对吧？它们应该这样定价和上涨。我至少通过轶事看到了证据。我没有具体数字可以引用，因为每一个表后交易都是定制的。人们支付的每千瓦时费用远高于标准电网。所以，我认为我们已经看到两级定价出现，数据中心电力的价值高于一般用途，对吧？所以，现在有很多表后的事情在进行。公司正在建设像 Bloom Energy 这样的新技术。有一个 Bloom Energy 的交易，在亚利桑那州建设几吉瓦。如果成功，那太棒了。那可能是我们需要继续前进的技术进步。但我们知道它还没有真正被证明。而且，顺便说一句，那些经济性与你在电网上看到的非常不同。

I mean, you could argue that. I suppose that's up to the energy markets, right? That's how they should be pricing and moving up. I do see evidence through anecdotes at least. I don't have specific numbers to quote around this because every one of these behind-the-meter deals is bespoke. And people are paying way more per kilowatt hour than they would off the standard grid. So, I think we're already seeing a two-tiered pricing occurring, where the value of electricity for data centers is higher than it is for general use, right? So, there's a lot of behind-the-meter stuff going on right now. Companies are building new technologies like Bloom Energy. There's a Bloom Energy deal for building a couple of gigawatts in Arizona. If that works, amazing. That might be the technological advances we need to keep going. But we know it hasn't really been proven out yet. And those economics, by the way, are very different than what you see on the grid.

怎么不同？

How so?

我的意思是，我认为如果你只是归结为人们愿意为数据中心能源支付多少每千瓦时，它就是更高。你可以容忍更高的价格，因为利润结构要好得多。基本上，如果我有一吉瓦的算力，我想卖掉它，我可以以——让我们称之为五倍的价格出售。我在编数字，但大约是我支付电价的五倍。所以，我有很多回旋余地。因此，我可以为那电力多付 30% 到 40%。

I mean, I think if you just boil it down to how much people are willing to pay per kilowatt hour for energy for data centers, it's just higher. You can tolerate a higher price because the margin structure is just so much better. Basically, if I have a gigawatt of compute and I want to sell that, I can sell it for, let's call it five X the price. I'm making numbers up, but it's on the order of five X the price of what I pay for that electricity. So, I got a lot of wiggle room there. So, I could pay 30% 40% more for that electricity.

当我们从能源——核心基础动力——稍微往下走到使用它的算力时。

When we kind of go slightly further down from energy, which is kind of the core foundational power, to the compute that uses it.

是的。

Yeah.

关于算力，我们现在知道什么你希望两年前就知道的？

What do we know now about compute that you wish you'd known two years ago?

我想我希望知道的一件事是，我们不再真正看到物理效率提升了。所以，你可以把构建计算机的方式分开，让我们称之为三个层次。第一层是物理层。如果我能从一个点发送一个比特到另一个点并更便宜地做到，比如，假设我用这里表示 0，那里表示 1。我的手机有一定重量。所以，在这两个状态之间移动需要能量。如果我做一个更轻的手机，它改变了所需的能量，对吧？所以，那只是一个纯粹的物理事情。如果你通常指的是摩尔定律或登纳德缩放，你可能听说过，让设备更小真的是物理事情。所以，你让物理层更高效、更密集。所以，那些创新基本上停止了。而且，你知道，如果我们看最近几代 GPU 或其他芯片，你不再看到多少效率提升。你看到效率提升来自上一层，即计算架构。所以，我会说过去几年或几代我们看到的大部分效率提升来自更小的数据类型，从 32 位到 16 位到 8 位到 4 位。所以，你获得相当好的效率提升，因为你移动更少的比特，你用更少的比特计算。然后下一层，没有被很好地利用，我们在 Mosaic 开始了一点，是关于我如何开始在模型层面强制局部性和减少数据移动。而且，你知道，五六年前，我们没有这样想，因为模型不是算力的主要驱动力。或者甚至，是的，也许六年前。算力仍然是通用的事情。我运行网站和数据库等等。然而，现在如果我把计算机仅仅看作设计来提供智能的物理设备，它改变了心态。现在，我可以从模型到架构再到物理进行协同设计。这就是我希望我知道的，顶层还有多少 alpha。不仅仅是物理层。我们长期关注物理层。

I think one thing I wish I'd known is the fact that we don't really see physical efficiency gains anymore. So, you can separate out the way we build a computer and let's call it three levels. So, the first level's the physical. So, if I can send a bit from this point to that point and do it cheaper, like, let's say I represent zero as being here and one as being here. My phone has a certain weight. So, there's energy required to move between these two states. If I make a lighter phone, it changes the amount of energy required, right? So, that's just a pure physical thing. If Moore's law is what you generally refer to it or Dennard scaling, you may have heard, where you make devices smaller is really a physical thing. So, you're making the physical layer more efficient and denser. So, those innovations have largely stopped. And, you know, if we look at the last few generations of GPUs or other chips, you don't see efficiency gains much anymore. You see efficiency gains from the next layer up, which is kind of a computing architecture. So, I would say most of the efficiency gains we've seen over the last several years or several generations have come from smaller data types, going from 32-bit to 16-bit to 8-bit to 4-bit. So, you get pretty good efficiency gains cuz you're moving fewer bits, you're computing with fewer bits. Then the next level, which is not really exploited that well, and we started this a little bit Mosaic, is about how do I start enforcing locality and reduction of data movement at the model level. And, you know, 5-6 years ago, we didn't think about it like this because the model was not the primary driver of compute. Or even, yeah, maybe 6 years ago. It was still compute was a general-purpose thing. I run websites and databases and all that kind of stuff. However, now if I look at the computer solely as a physical device designed to deliver intelligence, it changes the mindset of it. Now, I can co-design all the way from the model through the architecture to the physical. And that's what I wish I'd known, how much sort of alpha there is left in that top layer. It's not just the physical. We focus on the physical for a long time.

我能问一下，当从模型前关注转向模型后关注，跨越堆栈时，这种设计思维如何改变？

Can I ask, how does that design mindset change when moving across the stack with that transition from pre-model focused to post-model focus?

是的，我的意思是，这意味着我们不把模型视为神圣不可侵犯，对吧？而且我们的心态是能源优先，意味着我们所做的一切都必须给我们能源优势。这听起来很简单，但过去没有人这样设计这些东西。它总是关于速度。这实际上根源可以追溯到 1940 年代的 ENIAC。因为，你知道，计算机被设计得比替代品更快。一直是这样。顺便说一句，在 ENIAC 的情况下，你知道 ENIAC 为什么被建造吗？它是一个缩写。我认为它代表电子数值积分器和计算器。

Yeah, I mean, it means we don't treat the model as sacrosanct, right? And we're energy first in our mindset, meaning that everything we do has to give us an energy advantage. It sounds like a simple thing, but it's not how anyone has designed these things in the past. It's always been about speed. And this actually has its roots that go all the way back to ENIAC, you know, in the 1940s. Because, you know, computers were designed to be faster than the alternative. That was always the case. And the ENIAC's case, by the way, actually do you know why ENIAC was built? It's an acronym. It's Electronic Numerical Integrator and Calculator, I think is what it stands for.

听起来不错。

Sounds good.

是的。所以，这是一台由美国政府在 1945 年建造的计算机。它是首批建造的大型数字计算机之一，它的目的，就像曾经建造的每台机器一样，都有一个目的。它的目的是做火炮弹道计算，并且比替代品更快更准确地完成。当时的替代品是真正做数学的人。所以，你必须更快。那是关键。下一代出来，它必须比那更快。GPU 出来了，它们必须比 CPU 更快地渲染。每个东西相继试图比前一个更快。这导致了效率，原因有几个。我的意思是，在相同的功率包络下让事情更快是更高效的。

Yeah. So, this was a computer that was built by the US government in 1945. It was one of the first large-scale digital computers built, and its purpose, like every other machine that's ever been built, has a purpose. Its purpose was to do artillery trajectory calculation, and do it faster and more accurately than the alternative. The alternative at that time were humans who literally did the maths. So, you have to be faster. That was the key. Next generation comes out, it's got to be faster than that. GPUs came out, they had to be faster at rendering than a CPU. Each thing successively has tried to be faster than the previous thing. That has led to efficiency for several reasons. I mean, making things faster in the same power envelope is more efficient.

但摩尔定律也一直在推进，我们从真空管走到了晶体管。晶体管越做越小、越来越小。而现在我们处在一个物理层面其实没有太大变化的位置。晶体管并没有变得好很多。不过，我们在模型层面和架构层面还有很大的空间。

But also Moore's law was chugging along, and we went from vacuum tubes to transistors. Transistors kept getting smaller and smaller and smaller. And now we're in a place where that physical side of things is not really changing a whole lot. Transistors haven't gotten much better. However, we have a lot of room left on the model and architectural levels.

这是否意味着我们正在看到模型的商品化，还是说我们实际上正在看到模型质量的分化——那些能够节省能源、获得效率提升的，与那些做不到的之间拉开差距？

Does that mean we're seeing a commoditization of models, or does that mean we're actually seeing a separation of model quality in those that are able to save energy and have efficiency gains versus those that aren't?

我觉得完全正确。我一直比较擅长的一件事，就是判断接下来几年人们会把注意力放在哪里，而我认为这会是一个很大的部分。

I think that's exactly right. One thing I've been pretty good at is picking where people are going to be focusing in the next couple of years, and I think this is going to be a big part of it.

太精彩了，Naveen。我真的很佩服你。你在挑选一个市场的时候，会不会这样推演：我认为世界会往这个方向走，所以我要把自己放在这个位置？

Exceptional, Naveen. I really credit you. Do you walk through, when you're picking a market, and go, this is where I think the world's going to go, and so I'm going to position myself here?

是的，绝对如此。而且这也归结于兴趣本身。对我来说，这是最有趣的问题之一：造一台更好的计算机。

Yeah, absolutely. And it also comes down to just the interest. This is, for me, one of the most interesting problems there is: to build a better computer.

所以你在这里坐下来，在 unconventional 之前，在想法迷宫里，你当时想：我要聚焦的问题集是……什么？那是什么？

So you sat down here, pre-unconventional, in the idea maze, and you were like, the problem set that I'm going to focus on is dot dot dot. What was that?

我大概是这么看的：如果我要解决一个问题，它的影响是什么？它能解锁什么？其实这有点像做科学家。做一个好科学家，关键就在于问对问题——如果你回答了它，就会解锁更多问题。这几乎是一篇好论文的标志。所以对我来说，在这件事上，很清楚就是：好吧，我们将会耗尽能源，这一点很明确。我们现有的范式并没有真正聚焦于此。我认为这会是焦点所在，因为我们需要持续 Scaling，对吧？我们需要以许多新的形态交付智能，并大幅颠覆成本，才能让它无处不在。所以还没人在想这件事。我应该开始做。

Well, I kind of look at it like, if I were to solve a problem, what is the effect? What does it unlock? In fact, it's kind of like being a scientist. Being a good scientist is all about asking the right question that, if you answer it, unlocks many more questions. That's almost the hallmark of a good paper. So for me, in this endeavor, it was clearly like, okay, well, we're going to run out of energy. That's clear. The paradigm we have hasn't really been focused on this. I think this is going to be where the focus will be, because we need to keep scaling, right? We need to deliver intelligence in many new form factors and greatly disrupt the cost in order to make it ubiquitous. So no one's thinking about this yet. I should start it.

我能问一下吗——我们刚才在聊模型效率，以及它在能源方面能带来什么。我们看到今天的模型越来越高效。我们该怎么看自己在这条模型效率曲线上的位置？那里还有很多空间可跑，还是我们已经接近渐近点，只能挤出最后的增量？

Can I ask — we were talking about model efficiency and how it can deliver in terms of energy. We see models getting more and more efficient today. How do we think about where we are on that model efficiency curve? Is there a lot of room to run there, or are we kind of at the asymptotic point where you're eking out the final increments?

不，不，不。我们还有很长的路要走。我认为原因在于，我们还没有在模型和物理层面之间做真正的协同设计。所以你可以想出一个概念，也许仅靠软件，在标准 GPU 上就能获得，比如说，5% 的效率提升。听起来很棒。然而，你能利用某个概念的程度是有限的，因为物理层面的东西是按某种方式造出来的。如果你看我们的大脑，它们是共同演化的。我们让智能层和物理层共同演化，从而能够在物理层中利用智能层的东西。比如稀疏性。稀疏性意味着，比如说，计算单元之间没有连接。在 GPU 上利用稀疏性，我确实能获得一些收益，百分之百可以，但当我用的是标准的片外随机存取存储器时，这些收益的上限就被封住了。我走不了太远，因为那个物理层面的东西未必是为利用这种稀疏性而设计的。所以如果我能造出利用它的东西，我就能获得巨大的收益。这正是大脑所做的。这就是它共同演化的方式。

No, no, no. We have a lot to go. And I think the reason is we've not done this true co-design between the model and the physical side of it. So you can come up with a concept and you may be able to get, say, let's call it 5% efficiency gain just from software on a standard GPU. That sounds great. However, you can only exploit whatever concept so much because the physical thing is built in a certain way. If you look at our brain, they co-evolved. We co-evolved the intelligence layer and the physical layer, and we're able to exploit things in the intelligence layer in the physical layer. So sparsity, for instance. Sparsity means, let's say, just not having connections between computing units. I can get some kind of gain out of using sparsity in a GPU, 100%, but when I have a standard off-chip random access memory, the limit of those gains is capped. I can't go too far because that physical thing is not designed necessarily to exploit this kind of sparsity. So if I can build something that exploits it, I can get massive gains. That's really what the brain did. That's how it co-evolved.

当我们想到那种共同演化，或者说架构几乎针对任务的专门化，这是否意味着我们将坚定不移地走向芯片的专门化——我们看到 Etched，看到 Fractal、Olex，当然还有 Cerebras，还有不少人瞄准特定工作负载？那是算力的未来吗？你怎么看？

When we think about that co-evolution, or that specialization of architecture to the task almost, does that mean that we'll unwaveringly have the specialization of chips we see — Etched, and we see Fractal and Olex and a number Cerebras obviously, but a number of people going after specific workloads? Is that the future of compute? How do you think about that?

百分之百是。而且这其实可以追溯到图灵最初的那篇论文。他想造一台智能机器。那就是他一开始明确说出的目标。所以我们走上了这条路——我们拿了他那台理论构造的机器，造出了它的物理版本。所以我们造的任何东西都不是真正的图灵完备，因为它们内存有限之类的，但我们能造出超级有用的应用。我是说，我们可以靠它们经营企业。我们可以剪辑视频、图片和文字，以及各种很棒的东西。所以过去 60 年生产力获得了巨大提升。但它偏离了最初的目标。而现在我们某种程度上又回来了。所以计算机的目的，我认为，是提供智能。对我来说，它只是智能的物理体现。所以现在我们必须弄清楚如何把它做得更好。有趣的是，如果你造出更好的智能物理体现，你就造出了某种通用的东西。我们的大脑在某些方面是通用的。然而，硬件的工作负载是固定的。它并不像我们在 CPU 里所想的那样具备可编程性。所以 GPU 已经走上了为工作负载构建定制架构这条路。我认为我们只会围绕这一点进行超级优化。这是好事，因为现在我们能造出真正有朝一日能够建造任何东西的智能机器。

100% it is. And this actually goes back to even Turing's original essay. He wanted to build an intelligent machine. That was his stated goal right at the beginning. So we went down this path of — we took his machine, which was a theoretical construct, and we built physical versions of it. So anything we build is not truly Turing complete, because they have finite memory and stuff like this, but we were able to build applications that were super useful. I mean, we could run businesses off of them. We could edit videos and pictures and text and all kinds of great things. So that's a huge productivity gain over the last 60 years. But it deviated from that original goal. And now we've kind of come back. So the purpose of the computer, I think, is to provide intelligence. To me, it's just a physical manifestation of intelligence. And so now we have to figure out how to make that better. And what's interesting is that if you make a better physical manifestation of intelligence, you made something general. Our brains in some ways are general. However, hardware is fixed in its workload. It doesn't do programmability in the same way that we think about it in a CPU, say. So the GPU is already down this path of building customized architectures for workloads. And I think we're just going to hyper-optimize around this. And that's a good thing, because now we can build intelligent machines that are actually capable of building anything at some point.

我派你去——给我一点金融思维。你是一位出色的技术专家。从很多方面说，我是个不务正业的交易员。但我们提到，是的，芯片坚定不移地走向专门化。这在多大程度上会大幅削减一家横向供应商——具体就是 Nvidia——的可服务市场，还是说——老兄，别这么小家子气。现在还很早。我们还有 50 倍的空间可跑。

I sent you — give me my financial mind. You're a wonderful technologist. I am a delinquent trader for a living, in many respects. But we mentioned, yes, unwavering specialization of chips. To what extent does that massively reduce the TAM for a horizontal provider, Nvidia specifically, versus — dude, do not be so small-minded. It is still so early. We have 50x the room to run.

是的。我是说，市场将会非常巨大。你看，我认为 Nvidia 显然在把我们带到今天这一步的路上执行得极其出色。2014 年我创办 Nirvana 时，我们并没有真正理解市场会有多大。我很清楚，我们需要一种围绕神经网络的新计算架构。那是一种描述算力的新方式。对我来说，那就像是计算机能做的最重要的事。我认为 Jensen 也相对较早地意识到了这一点。我是说，他在 2014 年的 GTC 上就开始谈论人工智能了。我记忆犹新，因为很多做图形的人对此很恼火。他们说，你为什么花那么多时间在 AI 上？那什么都不是。

Yeah. I mean, the market is going to be huge. Look, I think Nvidia obviously executed extremely well on the path to getting us here. When I started Nirvana in 2014, we didn't really understand how big the market was going to be. It was clear to me that we needed a new computing architecture around neural networks. That was a new way to describe compute. And to me, it was like the most important thing a computer could do. I think Jensen recognized this relatively early on as well. I mean, he started to talk about artificial intelligence at GTC in 2014. I remember it vividly because a lot of the graphics guys were pissed off about this. They were like, why did you spend so much time on AI? It's nothing.

这不是重要的工作负载。而图形处理才是，对吧？他正确地认识到这将会很重要，在那个方向上投入了大量资源，并且执行得非常好。所以，我们经历了计算架构的演变，从我们使用那些机器的用途到它们最终将成为的样子。但现在我认为很清楚，我不需要关心其他任何事情。我只需要关心非常、非常高效地、非常、非常可扩展地运行神经网络，而这将成为衡量标准。

It's not an important workload. And graphics is, right? And he correctly recognized that this is going to be important, invested a ton in that direction, and actually executed very well. So, we had to go through this evolution of computing architectures from what we use those machines for to what they were eventually going to become. But I think now it's pretty clear to me that I don't need to care about anything else. I only need to care about running a neural network very, very efficiently, very, very scalably, and that's going to be the figure of merit.

你认为我们在多大程度上看到芯片的商品化？你知道，微软昨晚发布了那些芯片。显然，谷歌在 TPU 上取得了惊人的成功，亚马逊也有自己的。Anthropic 也在构建自己的。还有名字很妙的“我爱这个哈拉佩诺”。我想知道下一个版本会是什么。但我们是否看到芯片 inherent 的商品化？

To what extent do you think we see the commoditization of chips? You know, Microsoft last night released that chips. You've obviously got Google with amazing success with TPUs, Amazon with theirs. You've got Anthropic building their own. You've got the wonderfully named I love this jalapeno. Um I wonder what the next version will be. But do we see that just inherent commoditization of chips?

嗯，首先，我认为下一个必须是哈瓦那辣椒，因为我觉得它会更辣。最终你得用鬼椒之类的，对吧？

Well, first off, I think the next one's got to be habanero cuz it's going to be hotter, I would think. You got to go to ghost pepper or something eventually, right?

我觉得你应该注册商标，否则萨姆可能会这么做。

I think you should trademark it, otherwise Sam might do it.

我会问问他。所以，我实际上并不认为芯片会商品化。我有点把它看作与我们在进化中看到的情况平行。所以，我对所有这些事情如何运作有一个非常简单的世界观。所以，对我来说，智能将拥有无限的需求。智能的需求没有上限，因为我们生活的世界、宇宙是对抗性的。所以，起初两个细菌可能试图找到阳光的地方之类的。那个具有感官能力能够移动或 somehow 适应那个环境的细菌获胜并继续增殖。然后那些东西组织成真核细胞和多细胞生物。每次你基本上都在提高能量和赌注的水平，但它仍然是对抗性的。现在我们在人类群体之间看到这一点，对吧？我们谈论美国对中国之类的。会是某个东西对某个东西。你知道，在某个层面上我们开始利用更多的太阳。这就像卡尔达肖夫等级，你知道吗？我们在打某种星际战争。不管那是什么。你需要智能来做到这一点。智能是这些对抗性战斗的燃料、经济。所以，它有无限的需求。为了创造智能，你需要能量。能量到智能的转换就是硬件。当我说硬件时，我指的是整个堆栈。我指的是芯片、系统、云、网络，所有这些东西。所以，对我来说，如果我只从经济角度看，就像我将出售将能量转换为智能的能力。我做得越好，积累的价值就越多。所以，就市场而言，我不认为它会很快商品化。

I'll ask him. So, I actually don't see chips being commoditized. I sort of look at it as it runs in parallel to what we saw in evolution. So, I have a very simple worldview of how all this stuff works. So, intelligence to me will have infinite demand. There is no limit of demand of intelligence because the world that we live in, the universe is adversarial. So, at first two bacteria might be trying to find a spot of sunlight or something like that. The one that has the sensory capabilities to be able to move or somehow adapt itself into that environment wins and keeps proliferating. Then those things organize into eukaryotic cells and multi-cellular organisms. Each time you're basically upping the level of energy and the stakes, but it's still adversarial. And right now we see this between human populations, right? We talk about US versus China or whatever. It'll be something versus something. And you know, at some level we start to harness more of our sun. Like this is like the Kardashev levels, you know? And we're fighting some intergalactic battle. Whatever that is. You need intelligence to do that. Intelligence is the fuel, the economy of these adversarial battles. So, it has infinite demand. And in order to create intelligence you need energy. The conversion of energy to intelligence is hardware. And when I say hardware I mean the whole stack. I mean the chips, the systems, the cloud, the networks, all of those things. So, to me, if I look at it just from an economy like I'm going to be selling the capability to convert energy to intelligence. And the better I do that, the more value I accrue. So, I don't see it commoditizing anytime soon as far as the market goes.

你担心你会贪多嚼不烂，然后降低质量吗？在从能量到智能的过渡中，你在中间能吃掉越多，你赚的钱就越多。是的。但是，我昨天请了 Model 的创始人埃里克上节目。

Do you worry that you bite off more than you can chew and then deprecate quality? In that transition from energy to intelligence, the more that you can eat in the in between, the more money you're going to make. Yeah. But, I had the founder of Model on the show yesterday, Eric.

是的。

Yeah.

我的意思是我喜欢他，但这家伙只想做所有事情。我是说所有事情。我就想，产品广度在什么时候会损害深度？你考虑过这个吗？

And I mean I love him, but the dude just wants to do everything. I mean everything. And I was like, at what point does product breadth damage depth? Do you think about that?

我们一直考虑。专注非常重要。我确实认为世界在变化。现在，有了 AI，我们可以承担更多的广度。我们可以探索的表面积比以前更大。

We do all the time. Focus is super important. I do think the world is changing. Now, with AI, we can take on more breadth. Our surface area that we can explore is greater than it was before.

只是因为开发速度更快了。就像为什么现在每家公司都是所谓的复合型初创公司？

Just because of like speed of dev is much faster. Like why is that now every company is a compound startup, so to speak?

你知道，在我们的第一块芯片中，我们构建了一组耦合振荡器。它们有一个耦合矩阵。它们有一堆相互影响的振荡器。所以，当我们做这个时，我会问——你知道，在旧世界，比如 AI 之前，我可能会有一个问题：“好吧，如果我们不做全连接，噪声如何传播？”意思是，我有一个振荡器的菊花链，我在开始有一些噪声特性。我想看看它在比如四到六跳后如何降低信息质量。在旧世界，我不得不问我的研究团队：“嘿，你能模拟这个吗？加入某种噪声模型。”我只是想要一个零阶的理解，对吧？只是为了引导我自己的想法。那可能需要一周左右，因为他们会说：“啊，我正在做其他事情。我下周再回来处理。”也许写代码和验证需要一周，所以，我花了几周才得到答案。我可以直接向模型问这个问题。它吐出代码。我可以内省代码。我可以来回几次。我在一小时内得到答案。所以，以前需要两周思考的事情现在只需要一小时。所以，因为 AI，我每单位时间可以思考更多事情。现在，我认为我们仍然有一个问题。我实际上昨天刚发了关于这个的帖子，我们有一个 AI 垃圾问题。这些模型产生了太多垃圾，而且很难验证。有些是好的，有些不是，我们需要弄清楚如何大规模地做到这一点，但我可以走得更快。我可以覆盖比以前更多的表面积。

You know, in our first chip, we built a set of coupled oscillators. Okay, they have a coupling matrix. They have a bunch of oscillators that sort of influence each other. And so, when we're doing this, I would ask I you know, in the old world, say before AI, I might have a question like, "Okay, well, how does noise propagate if we don't do all-to-all connectivity?" So, meaning I have like a daisy chain of oscillators, and I have some noise characteristic at the beginning. I want to see how that degrades information over, say, four or six hops. In the old world, I'd have to ask my research team, "Hey, can you go model this? Put some sort of noise model in." And I just want a zero-order understanding here, right? Just to kind of guide my own thoughts. That might take a week or so cuz they're like, "Ah, I'm working on this other stuff. I'll come back to it next week." Maybe it takes a week to write the code and verify it and get So, I got a couple weeks to get an answer. I can literally ask this question to a model. It spits out the code. I can introspect the code. I can go back and forth with it a few times. I get an answer in an hour. So, something that would have taken me 2 weeks to think through has now taken me an hour. So, I can think through more things per unit time because of AI. Now, I think we still have a problem. I actually just posted on this yesterday that we have an AI slop problem. We have too much crap coming out of these models, and it's very hard to verify it. Some of it's good, some of it's not, and we need to figure out how to do that at scale, but I can go faster. I can cover more surface area than I could before.

你为什么说我们有太多垃圾？是不是就像在 X 上看到，你知道，现在覆盖我时间线的东西。无意冒犯美国，但你们政客的脸结构相当有趣，嗯，这是一种英式说法。嗯，我的时间线上全是这些，但有不同 AI 生成的他们。有外星人从国会出来。我是说，这太疯狂了。这就是你说的 AI 垃圾吗？

What makes you say we have too much crap? Is it just like looking on X and seeing like you know, the thing that is just covering my timeline right now. No disrespect to US, but it's your politicians who are rather in interestingly structured faces um would be an English way of saying it. Um that's all over mine, but there's like different AI generations of them. There's aliens and coming out of Congress. I mean, it's nuts. Is that what you mean by AI slob?

不是。嗯，不是。

It's not. Um it's not.

那是娱乐。

That's entertainment.

是的，那更多是娱乐。顺便说一句，我会说英国人垄断了自嘲幽默。这比我们在美国能做的要好得多，所以

Yeah, that's more entertainment. I will say by the way that English have cornered the market on self-deprecating humor. It is so much better than we can ever do in the US, so

老实说，顺便说一句，我们从英国看这个都在想：“哇，我们一直在抱怨自己的政客。”

Honestly, by the way, we're all looking at this from the UK being like, "Wow, we've been moaning about our own politicians."

这些家伙。

these guys.

是的，我知道。不，我们已经把它提升到了新的水平。

Yeah, I know. No, we've taken it to new levels.

毫无疑问。我们，我告诉过你我们是创新者，对吧？所以实际上这是一个完美的例子，你看到一个输出，你说：“好吧，那是否首先符合你的直觉检查？”然后你会说：“嗯，也许不是。”所以，我认为真正有趣的是，今天的老年人看到图像时没有那种直觉检查。

No doubt about it. We I told you we're innovators, all right? And so actually this is a perfect example of you see an output and you say, "Okay, does that kind of does it meet your gut check first?" And you're like, "Huh, maybe not." So, I think what's really interesting is older people today see images and they don't have that gut check.

他们会说：“哇，那是真的。”我的父母曾给我看 AI 生成的特朗普弹吉他的视频，他们说：“哇，他弹吉他。”我说：“妈，他没在弹吉他。那是 AI 生成的东西。”她说：“哦，好吧。”所以对我来说，这很明显是 AI 生成的。我认为我们在公司里也看到了这种验证问题。我在我们公司看到了。我们可以构建各种产物。我们可以构建代码。我们可以构建模型。我们可以做各种事情，但你必须闭环并确保每一件事都是正确的。这实际上是信任结果。而这样做实际上开始成为一个大瓶颈，因为你产生了太多产物。所以，我认为我们已经把问题从生产转移到了验证。

They're like, "Wow, that's real." I mean, I've had my parents say, you know, show me these AI generated videos of Trump playing the guitar and they're like, "Wow, he plays the guitar." I'm like, "Mom, he's not playing that guitar. That's an AI generated thing." She's like, "Oh, okay." So, for me it's quite obvious that it's AI generated. So, I think that verification problem we see in companies. I see it in our company. We can build all kinds of artifacts. We can build code. We can build models. We can do all kinds of things, but you have to close the loop and make sure each one of those things is correct. It's actually trust the result. And doing that actually starts to become a big bottleneck because you're producing so many artifacts. So, I think we've moved the problem from producing to verification.

验证的价值在于，如果你拥有验证层，那么你就拥有自动化层，因为你有可验证性，然后你可以在此基础上推进。是这样吗？

And the value of verification is if you own the verification layer, then you own the automation layer because you have like verifiability that you can then progress there on. Is that it?

嗯，是的，我想是这样。我的意思是，我认为这是使用 AI 开发东西的必要部分，但如果你拥有验证层，你就拥有生产层吗？我不完全确定。我认为你还必须拥有规格。所以，生产有用的输出，比如建造汽车或桥梁之类的，你必须能够制定规格，弄清楚我是否在解决一个有用的问题？然后，产生的每一个产物都必须验证是正确的，然后朝着那个规格前进。最终，你必须说：“是的，我确实按预期实现了规格。”所以，这是一个端到端的过程，不仅仅是验证。验证只是其中的一部分。我不知道人们会如何评估它的价值。我认为我把它作为整体的一部分来评估，比如达到最终结果。

Um yeah, I suppose. I mean, I think it's a necessary part of using AI to develop things, but if you own the verification layer, do you own the production layer? I'm not entirely sure. I think you have to own the spec as well. So, producing a useful output like building a car or a bridge or something, you have to be able to spec it, figure out am I solving a problem that's useful? Then, each one of these artifacts that's produced has to be verified to be correct and then marching towards that spec. And eventually, you have to say, "Yes, I did indeed implement the spec as intended." So, it's an end-to-end process, not just verification. So, verification is a piece of it. I don't know how one would value that. I think I value it as part of the whole, like getting to the end result.

我们节目请来了 Atlassian 的 Mike，很棒的人。他谈到了思维链的重要性，以及理解智能体主导的决策与人类主导的决策在公司不同职能中的对比。我的一部分想法是：“我完全理解，这绝对合理。”另一部分想法是：“这样想感觉有点老套，那意味着什么？我们不像信任人类那样信任智能体的决策？”那也感觉有点奇怪。

We have Mike on the show from Atlassian, awesome guy. And he was talking about kind of the importance of kind of chain of thought and understanding, you know, agent-led decision-making versus human-led decision-making across different functions in companies. And part of me was like, "I totally get that and that makes absolute sense." And part of me was like, "That feels kind of old to think of it like that and like what does that mean? We don't trust agent decisions as much as we trust humans?" That feels kind of weird, too.

我们确实不信任。

We don't.

想想，是的。

Think about Yeah.

嗯，我的意思是，我同意你的看法。我认为这很酷。但在很多情况下，我认为我们还不应该像信任人类那样信任智能体的决策。我们为某些事情训练多年来建立判断力。我的意思是，成为神经外科医生，对吧？那是医学院毕业后 7 年的训练。你知道，这是有原因的。并不是学习如何缝合有多难。你知道，我在研究生院做过。问题不在于成为外科医生的物理机械部分。问题在于要有判断力来决定往哪个方向走。你必须见过足够多的病例，见过输出，见过人们如何思考这些问题，然后才能建立这种判断力。我认为我们现在定义智能的方式是，我们取一个随机的模型，对吧？我们可以多次提示它，我们会得到不同的输出。你所做的是探索可能解决方案的分布。然后我们需要选择哪些解决方案是好的。你知道，你会得到很多不同的输出，其中大多数都不好。你可以说人类也这样做，对吧？我们说：“好的，我如何做决定？”嗯，我可能会思考所有输出，我有某种启发式来决定什么是好的。我们在内部这样做，也许是内心独白或某种内部过程。但模型吐出一个思维链，然后我们有许多思维链，我们决定什么是好的，然后我们走那条路。那个决定元素实际上对于每个任务并没有被很好地理解。所以我认为我们不应该信任这些东西。我认为你建立的决策链越长，它就越可能失控。你可能会得到看起来非常稳健但完全错误的东西。所以这就是我们今天所处的位置。未来可能会改变。我认为我们需要更多突破才能达到那里，但这就是我们今天所处的位置。

Uh I mean, look, I agree with you. I think it's cool. I don't think we should trust agent decisions as much as humans in a lot of cases just yet. We train for certain things for many years to build judgment. I mean, becoming a neurosurgeon, right? It's a 7-year training after med school. You know, and there's a reason for that. It's not that learning how to throw a stitch is that hard. You know, I've done that in grad school. It's not the physical mechanical part of being a surgeon that's the problem. It's having the judgment on which direction to go. You have to have seen enough cases, seen the output, seen how people think through these problems to then build that judgment up. And I think the way we're defining intelligence right now is we take a model that's stochastic, right? We can prompt it several times and we'll get different sorts of outputs. And what you're doing is you're exploring the distribution of possible solutions. And then we need to pick which of those solutions is good. You know, you'll get a lot of different outputs and most of them are not good. And you could argue that a human does this too, right? We say, "Okay, how do I make a decision?" Well, I might think through all the outputs and I have some kind of heuristic of deciding what's good. We do this internally, maybe it's an internal monologue or some sort of internal process. But a model spits out a chain of thought and then we have many chains of thought and we decide what's good and we take that path. That deciding element is actually not well understood for every task. And so I don't think we should trust these things. I think the longer and longer decision upon decision upon decision chain you build, the more off the rails it can go. And you can get things that look very robust, but they're completely wrong. So that's where we are today. That will probably change in the future. I think we need a few more breakthroughs to get there, but that's where we are today.

所以我个人很喜欢这样的问题：“你相信什么今天看起来古怪、奇怪、疯狂，但 5 年后会非常普遍并得到认同？”我记得把信用卡插入 PayPal，我父母责骂我。还记得 OKCupid 吗？当时想：“你在网上找丈夫或妻子？多么荒谬的想法。”当时很疯狂。

So I personally I love a question like, "What do you believe that today looks kooky, weird, nuts that in 5 years time will be very commonplace and agreed?" I remember putting my credit card into PayPal and my parents telling me off. And like remember like OKCupid and being like, "You find your husband or wife online? What a ridiculous idea." Nuts at the time.

是的。

Yeah.

我认为我们会比信任自己更信任我们的智能体。

I think that we will trust our agents more than we trust ourselves.

我认为在某个时候这是真的。是的。

I think at some point that's true. Yeah.

但你认为我们今天离那还很远？

But you think that we are far from that today?

是的，我们今天还没到那一步，专业知识并不会过时。它只是会以不同的形式呈现，我认为未来会看起来像这样：我们开始构建我们称之为模型的东西，但我实际上不认为它是一个模型。我认为它是一整台机器。我认为智能既有物理部分，也有智能层，它们是不可分割的。所以我有一台机器。它几乎像一个有机体，我可以教它成为各种任务的专家，价值将来自建立这种专业知识。就像你知道的，你上哪所学校？你如何学习成为神经外科医生或厨师或其他什么。那个来源很重要，对吧？就像你去某个人手下学习，你学习了他们的方法、他们的思维。价值就来自那里。就像你或任何人类从头到尾都是一台智能机器，但真正让你有价值的是你被教导的方式。我认为我们将在机器中开始看到类似的东西，我们可以复制那种智能等等，但你被教导的方式实际上将是重要的，所以其中一部分实际上是知道如何达到一个真实的、可行的、经过验证的最终结果。

Yeah, we're not there today and expertise doesn't really go out of style. It's just going to take on different forms and what I think the future will look like is that we start building we'll call it models, but I actually don't think it's a model. I think it's a whole machine. I think of intelligence as having a physical and having the intelligence layer as one. They're not separable. So I have a machine. It's almost like an organism that I can teach to be an expert at various tasks and the value will come from building that expertise up. It's like you know, which school did you go to? How did you learn how to be a neurosurgeon or a chef or whatever. That providence matters, right? Like you went and studied under this person and you learned their methodology, their thinking. That's where the value came from. Like you or any human is an intelligent machine top to bottom, but was the way you were taught that actually made you valuable. And I think we're going to start to see something similar in machines and we can replicate that intelligence and all of that, but the way you were taught is actually going to be the important thing and so part of that is actually knowing how to get to an end result that is real, that will work, that is verified.

我没受过多少教育。我在教育上花的时间很少，Naveen，但谢谢你。我希望人类还有一些价值。嗯，不，我完全理解你。当你思考学习中的价值时，很多价值在于一个人随着时间的推移、在多年做某事的过程中积累的数据。

I wasn't really taught much. I spent very little time in education, Naveen, but thank you. I hope there is some value in the human being. Um no, I get you totally. When you think about the value in that learning, a lot of that is in the data that one accrues over time in the years that you're doing something.

是的。

Yeah.

我们今天看到数据公司。我在 McColl，但我们看到 10s 达到天文数字的收入水平。以前是惊人的大。

We see data companies today. I'm in McColl, but we see 10s hit astronomical levels of revenue. Previously were like astonishingly large.

我的意思是，这仍然大得惊人，但发生在如此短的时间内。Naveen，你真是个聪明人，和你聊天真愉快。我就是想问你很多关于世界的问题。这个市场会发生什么？它会随着时间推移而整合吗？还是会越来越解绑，你建立更多的防御性？会发生什么？

I mean that's still astonishingly large, but in such short time. Naveen, you're such a smart dude, it's such a joy talking to you. I just kind of want to ask you lots of questions about the world. What happens to this market? Does it just consolidate over time? Does it unbundle more and more and you build more defensibility? What happens?

老实说，我认为它会专业化。我认为如果你专业化并真正开始解决重要问题，你就会拥有一个可防御的大市场。这又是技能的幂律。你可以说所有人类都能学习许多技能，但某些技能比其他技能更有价值。你构建的数据和课程也会如此，对吧？

I think it specializes, to be honest. And I think if you specialize and actually start to work on important problems, you have a big market that's defensible. It's again a power law for skills. You could argue that all humans are capable of learning many skills, but certain skills are more valuable than others. And the same thing will be true with data and curricula that you build, right?

完全理解。所以我们看到数据公司的专业化。所以你不认为这会商品化？

Totally get that. And so we see the specialization of data companies. So you don't think that commoditizes?

是的，我不认为它是商品。我认为它最初是商品，因为那时只是大量收集观察数据，对吧？但当你开始构建更高质量的数据集时——在 AI 领域我们就是这么说的，比如更高质量的编码数据、更高质量的思维链数据等等——你实际上开始在数据中获得很多价值。模型在某种意义上几乎只是它所学内容的体现。所以我认为数据层在未来会非常重要。我不认为它会商品化。我认为构建模型并知道如何选择数据实际上是现在的关键。大型模型公司——这很大程度上是一个数据问题。我的意思是，他们在模型本身上取得了进展，但大部分能力并非来自那里。它来自使用不同类型的损失函数来整理数据。比如我们之前有预训练，那是一种非常特定的损失，然后我们转向现在所谓的 mid-training，即强化学习、监督微调，所有这些不同的混合。所以那个课程——你如何从预训练过渡到所有这些不同类型的训练来修改模型的行为——实际上是秘诀。

Yeah, I don't think it's commodity. I think it started as a commodity because it was sort of like at volume gathering of observations, right? But as you start to construct higher quality data sets—that's what we call it in the AI world, like higher quality for coding, higher quality for chain of thought or whatever—you actually start to have a lot of value in the data. The model is almost just a manifestation of what it learned in a sense. So the data layer, I think, is going to be pretty important going forward. I don't see that commoditizing. I think building models and knowing how to choose that data is actually the game right now. The big model companies—it's largely a data problem. I mean, they are making advances on the model itself, but that's not been where most of the capabilities come from. It's been from curating data using different sorts of loss functions. Like we had pre-training before, which was a very particular kind of loss, then we moved to what we call mid-training now, which is reinforcement learning, supervised fine-tuning, all these different sorts of mixes. So that curriculum—how you went from pre-training to all these different sorts of trainings to modify the behavior of the model—is actually the secret sauce.

我能问你吗，你提到过那种预训练强化学习，然后显然真正的转变已经转向后训练。

Can I ask you, we've seen you mentioned that kind of pre-training RL and then obviously a real shift has moved to post-training.

是的。

Yeah.

随着重点从预训练转向后训练，你认为这是模型改进效率以及我们如何达到那里的持久永久转变，还是我们会回到强化学习成为主要或预训练成为主要？从重要性随时间变化来看，那会是什么样子？

With that shift of emphasis from pre-training to post-training, do you think that is an enduring permanent shift in the efficiency of model improvements and how we get there, or will we move back to RL being the primary or pre being the primary? How does that look in terms of importance over time?

嗯，我不知道它是否更高效。实际上，我认为强化学习效率更低。如果你从每单位算力或每单位美元模型提升多少来看，它实际上效率更低。但它确实在向前推进，对吧？我认为想法是我们要让模型在所有重要任务上变得更好。所以我认为在未来几年，是的，非常重要。我认为我们会开始在中层构建专业化，甚至可能在预训练层。所以很难确切说未来的范式会是什么，因为这些不是——这些是我们为了下一个问题而设计的非常人为的场景，对吧？就像我在这里，我想去那里，所以我想办法去那里。然后我在这里，我想去这里，所以我想办法去那里。你最终会得到这种随机的点游走。这就是行业演变的方式。我不认为模型公司在性能方面考虑得非常长远。比如，哦，我要在思考 AI 系统的方式上取得重大突破，从而给我优势。我会说大型实验室并没有那么做。

Well, I don't know that it's more efficient. Actually, I think RL is more inefficient. If you look at it in terms of how much better the model gets per unit compute or per unit dollar, it's actually more inefficient. But it does keep moving forward, right? I think the idea is that we want to make the models better at all the tasks that are important. And so I think it's probably for the next several years, yes, very important. I think we'll start to build specialization in those middle layers, maybe even in the pre-training layer. So it's hard to say exactly what will be the paradigm in the future because these aren't—these are very contrived scenarios that we just engineer our way to the next problem, right? It's like I'm here and I want to get to there, so I engineer some way to get there. And then I'm here and I want to get to here, so I engineer some way to get there. And you end up with this kind of random walk of points. And that's how the industry evolves. I don't think the model companies are thinking very long term in terms of performance. Like, oh, I'm going to make some massive breakthrough on the way I think about an AI system that's going to give me an advantage. I would say the big labs are not doing that as much.

你认为在当今生态系统的日复一日的竞争中，他们能承担得起吗？

Do you think they can afford to in the day-by-day race that is this current ecosystem?

这就是问题所在。

That's the problem.

是的。

Yeah.

真的很快。我的意思是，你必须紧跟其上。如果你在模型上晚了 1 个月，那就是灾难，对吧？如果你是 OpenAI 或 Anthropic 或 Grok 或 Google。所以，是的，他们现在还承担不起那些更大的风险。你看到了这一点——当 Elon 创立 xAI 时，他基本上采取了，好吧，这是范式。我们去雇人，然后砰。他很快到达了前沿，因为他有资源。所以这向我指出，并没有大量非常深入的创新。而是把碎片拿出来，与知道如何构建它们的团队一起组装，并非常快地执行。我认为这与试图提出新范式非常不同。你看看 Yann LeCun 的公司和 Fei-Fei Li 的公司。他们试图做更深入的创新。我认为那可能在初创公司的领域，它开始看起来像药物开发，对吧？就像所有大风险由初创公司承担，所有渐进的东西由大公司完成。大公司在初创公司显示出前景时收购它们。

It's really fast. I mean, you've got to be on top of it. If you're 1 month late on a model, that's disaster, right? If you're OpenAI or Anthropic or Grok or Google. So, yeah, they can't afford to take those bigger risks just yet. And you're seeing this—when Elon spun up xAI, he basically took, okay, here's the paradigm. Let's go hire the people and boom. He got to the frontier pretty fast because he had the resources. So what that points out to me was that there wasn't a ton of very deep innovation. It was taking the pieces, putting them together with a team that knows how to build them and executing very fast. I think that's very different than trying to come up with a new paradigm. And you look at like Yann LeCun's company and Fei-Fei Li's company. They were trying to do deeper innovations. And I think that might be in the realm of startups and it starts to look like drug development, right? Like all the big risks are taken by startups and all the incremental stuff is done by big companies. The big companies buy the startups when they show promise.

我有时问简单的问题，但我不知道答案，而你非常聪明。为什么今天每个人都对强化学习如此兴奋？就像你提到的 mid learning，我只是有很多人就像一切都在于强化学习。我想哦我同意。是的。我同意。

I sometimes ask simple questions, but I don't know the answers and you're very smart. Why does everyone get so excited by RL today? Like you mentioned it kind of the mid learning there and I just have so many people just like it's all about RL. I'm like oh I agree. Yeah. I agree.

我的意思是，有几个原因。如果你想想我们学习的方式，它主要是强化学习。如果我要教你一项新技能——学习如何发网球或类似的东西——它实际上是一个强化学习循环。我会给你一个它应该是什么样子的范式。这是你的发球应该看起来的样子，也许演示一下或给你看视频之类的，它会让你建立一个关于什么是好或伟大的内部模型，然后你不断根据那个内部模型评判自己，说 yes，我快到了。我在进步，或者我没有，然后我需要寻求帮助。所以那就是强化学习，我认为它重要的原因是我们无法收集足够的数据来解决世界上的每一个任务。我们根本不能。你不能通过仅仅收集数据来预训练解决这个问题。这太难了。坦率地说，模型不够高效，无法利用它。所以你需要有一个范式，你可以开始说，‘好吧，这些是 10 个最重要的任务。然后我有某种我们称之为可验证奖励的东西。我有某个东西，那个内部模型。比如，一个好的发球是我用力击球并且它进了，对吧？它非常可验证。我可以测量速度。

I mean there's several reasons. If you think about the way we learn it's primarily reinforcement learning. If I were to teach you a new skill—learn how to serve a tennis ball or something—it's really an RL loop. I would give you a paradigm of what it should look like. Here's how your serve should look and maybe demonstrate that or show you a video or something and it will allow you to build an internal model of what good is or great is and then you keep judging yourself against that internal model and say yes, I'm getting there. I'm getting better or I'm not and then I need to ask for help. So that is RL and I think the reason it's important is we cannot gather enough data to solve every task in the world. We simply can't. You can't pre-train your way out of this by just gathering data. It's just too hard. The models, frankly, aren't efficient enough to make use of it. So you need to have a paradigm where you can start to say, 'Okay, these are the 10 most important tasks. And then I have some sort of what we call a verifiable reward. I have something that's that internal model. Like, a good serve is I hit it really hard and it went in, right? It's very verifiable. I can measure the speed.

我可以测量它是否在边界内。当我有这种可验证的奖励时，我就可以开始说：‘好吧，我采取的行动是否导致了那个好结果？’所以现在我有一个结构，可以让我解决任何一组通用任务。这就是为什么强化学习如此令人兴奋。

I can measure whether it was in or out of bounds. And when I have that verifiable reward, then I can start to say, 'Okay, did the actions I took lead to that good outcome or not?' So now I have a structure that allows me to solve any set of general tasks. And that's why RL is so exciting.

你认同 Yann LeCun 的后 Transformer 世界理论吗？那会如何改变一切？

Do you buy Yann LeCun's post-transformer world theory? And how does that change everything?

我确实认为我们会进入后 Transformer 世界。我不认为 Transformer 是终极方案。直到几年前我才真正理解这一点，因为 Transformer 混淆了很多东西。我们关心的是推理，是能够推理一个问题。我不知道是谁说的，但大意是：‘智能是你在没有信息时所做的事，而不是你在有信息时所做的事。’所以能够从某种机制性或第一性原理的理解来推理问题，这就是智能。Transformer 有点那种能力，但它们实际上是通过这种智能体式的流程来实现的：你取多个样本，然后决定哪个好。所以我取一个轨迹，然后与另一个轨迹比较，我说：‘好吧，这个轨迹更好。’所以我把这个称为我的推理轨迹。而我们实现这种推理的方式是通过检索机制。注意力机制基本上有这个上下文，这是一种检索，然后我们可以从中采样，最终得到看起来像推理的东西。我认为这对于我们进行推理的方式来说极其低效。我们不需要这样构建推理。我们的大脑不是这样工作的。我们不会构建一个巨大的轨迹，把它输入到另一个模型的输入中，然后产生另一个轨迹。这不是我们大脑的工作方式。所以有更高效的方法，从算法角度有更高性能的方法。因此我们将超越 Transformer。

I do think we will be in a post-transformer world. I don't think the transformer's the be-all end-all. I don't think I truly understood this until maybe a couple of years ago because transformers conflate a lot of things. What we care about is reasoning, being able to reason through a problem. I don't know who said it, but it's like, 'Intelligence is what you do when you don't have the information, not what you do when you do have the information.' So being able to reason through a problem from some sort of mechanistic or first principles understanding is what intelligence is. So transformers kind of sort of have that, but they really do it through this agentic kind of flow where you're taking multiple samples and you're deciding what's good. So I take one trace and then compare it to another trace and I say, 'Okay, this trace is better.' So I'm going to call that my reasoning trace. And so the way we kind of implement this reasoning thing is through a retrieval mechanism. So attention basically has this context, which is a retrieval thing, and then we can sample off of that and eventually get to something that looks like reasoning. I think this is extremely inefficient for how we do reasoning. We don't need to build reasoning this way. Our brains do not work in this way. We don't build a huge trace, pump it into the input of another model, and then produce another trace. That's not how our brains work. So there are way more efficient ways to do this, and way more high-performance from an algorithmic standpoint ways to do this. And so we will move past the transformer.

你认为成为伟大创始人所需的条件变了吗？

Do you think what it takes to be a great founder has changed or not?

我不认为它变了。我认为基本原则仍然相同。我认为对未来的观点，你的愿景，实际上现在可能比以往更重要。在 AI 之前的世界，你只要能够执行，就可以成为一个相当不错的创始人。你不必拥有很好的愿景。你只需要知道如何执行。如果你只是出去执行一家大公司，这并不难，因为大公司行动缓慢、笨重，我认为你实际上可以建立一家初创公司，然后卖掉它或以此建立一家公司。我认为如今如果你只基于执行，我不认为你会走得很远。

I don't think it's changed. I think that the basic principles are still the same. I think having a viewpoint on the future, your vision, actually may matter more now than it did before. In the pre-AI world, you could be a pretty good founder by just being able to execute. You don't have to have very good vision. You just have to know how to execute. And if you just go out and execute a big company, which isn't that hard because big companies are slow and lumbering, I think you could actually build a startup, and you could sell it or build a company out of it. I think these days if you're just based on execution, I don't think you're going to get very far.

我非常同意。我能问你吗，我们谈到了不同的层次。你说过：‘我不认为数据会商品化。我认为专业化。我认为那里有价值。’哪个层次会随着时间的推移而商品化？

I agree incredibly. Can I ask you, we spoke about the different layers. You said, 'I don't think it's going to get commoditized in data. I think specialization. I think there is value there.' Which layer does get commoditized over time?

是的，这是个好问题。可以说，能源已经是一种商品，对吧？所以这已经商品化了。我认为一定程度的智能确实开始商品化。说实话，我们已经看到了。比如 6 个月大的模型，比如说中国的开源模型，那个水平，它们相当有用，对吧？6 个月前的东西仍然相当不错。所以我认为这是商品化的标志，对吧？前沿不是唯一重要的东西。对于某些问题，它是，但 6 个月后，那个前沿将出现在开源模型中。

Yeah, it's a good question. Arguably, energy's already a commodity, right? So this is already commoditized. I think a certain amount of intelligence does actually start to become commoditized. We're already seeing it, to be honest. Like, models that are 6 months old, let's say the Chinese open models, that level, they're pretty useful, right? Something that's 6 months old is still pretty darn good. So I would argue that's a hallmark of commoditization, right? The frontier is not the only thing that matters. For certain problems, it is, but in 6 months, that frontier will be in the open models.

那么请帮我理解。如果所有基础设施的交付机制或产出都商品化了，会发生什么？

So then help me out. What happens if the delivery mechanism or the output of all the infrastructure is then commoditized?

嗯，我认为商品化不一定意味着不好。它只是意味着利润率大幅下降，但也意味着市场非常大。所以你可以说西红柿是商品，但我们仍然吃它们。我们仍然需要它们，你知道吗？所以我们将有一个巨大的智能市场，那个市场将继续增长，但利润率可能不会很大。我实际上不认为模型公司会是拥有整个堆栈利润的公司。今天甚至都不是这样。比如，英伟达今天拥有堆栈的利润，对吧？我会说他们可能是拿走大部分利润的人。

Well, I think commoditization doesn't necessarily mean it's bad. It just means that margin falls quite a lot, but it also means the market is very big. So you could argue that tomatoes are commoditized, but we still eat them. We still need them, you know? So we will have a huge market for intelligence, and that market will continue to grow, but the margins may not be huge. I actually don't think the model companies will be the ones who own the margin of the whole stack. It's not even true today. Like, Nvidia owns the margin of the stack today, right? I would say they're probably the one who take the lion's share of the margin.

很多人对我说，推理成本实际上会在未来几年大幅下降，我们可能会看到推理市场的商品化。你同意还是不同意？

A lot of people say to me the cost of inference will actually go massively down over the next few years, and we'll see the commoditization potentially of the inference market. Would you agree with that or not agree with that?

好吧，让我们分解一下。推理市场的商品化意味着什么？这意味着运行推理的运营商，我假设，对吧？

Well, let's break it down. What is commoditization of the inference market mean? What that means is that the operator who's running inference, I assume, yeah?

是的。是的，绝对。你的飞轮、你的模型、你的基础时间，以及它们提供的服务。是的。现在完成了。

Yeah. Yeah, absolutely. Your flywheels, your models, your base times, and the services that they provide. Yeah. It's done now.

因为如果你看他们的价值主张，他们所做的并且做得很好的是拿一个现有的模型，比如说一个开源模型，然后让它运行良好，你知道，有正常运行时间保证，并使其在基础设施上运行相对成本高效。所以他们是基础设施玩家，对吧？他们基本上是将模型映射到物理硬件上并做好这一点。现在，我认为未来那里会有价值，但推理的价格，比如人们愿意为 token 支付的价格，会下降。所以他们必须从硬件能力中挤出更多利润来交付那个 token。而那个层次的油水我认为会减少。所以我认为这会导致他们利润率降低。而且我认为模型本身已经商品化了。你现在有几种不同的方式可以获得它。但硬件方面可能不会崩溃。我认为那部分仍然会有优势，因为你仍然可以挤出更多油水，比如‘嘿，我更高效地将能源转化为智能。’与基础设施层相比，我认为那里有更多的油水可挤。所以我实际上认为硬件是未来 10 年该去的地方。

Cuz if you look at their value prop, what they've done and they've done very well is take a model that exists, an open model, say, and then make it work well, you know, with uptime guarantees, and make it relatively cost-efficient to run on infrastructure. So they're infra play, right? They're basically mapping the model onto physical hardware and doing that well. Now, I think there will be value there in the future, but the price of inference, like what people are willing to pay for token, will drop. And so they have to squeeze more margin out of the hardware's capabilities to the delivery of that token. And that amount of juice in that layer I think will decrease. So I think it'll result in them having lower margins. And I think the model itself is commoditized. You have several different ways to get it now. But the hardware side of it probably won't collapse. I think that part of it is still going to have an advantage because you can still squeeze more juice out of like, 'Hey, I convert energy to intelligence more efficiently.' We have a lot more juice to squeeze there versus the infrastructure layer in my view. So I actually think the hardware is the place to be over the next 10 years.

作为投资者，我需要改变我的身份吗？我知道这听起来很奇怪，Naveen，但就像我是一个软件投资者。过去 10 年我投资 SaaS。我做得很好。

Do I need to change who I am as an investor? I know that sounds strange, Naveen, but like I'm a software investor. I've invested in SaaS for the last 10 years. I've done it well.

是的。

Yeah.

老兄，做硬件投资相当难。这是一种非常不同的心态。这一代投资者准备好成为硬件投资者了吗？否则

Dude, it's quite hard doing hardware investing. It's a very different mindset. Is this generation of investors equipped to be hardware investors? Otherwise

我会说，如果你是一个纯粹的软件投资者，你有点……老实说。

I would say if you're a pure software investor, you're kind of Honestly.

我可以很自信地这么说，因为我认为 SaaS 不再是你能获取利润的地方了。从 2011 年到 2020 年这段时间到现在，我们经历了一种虚假繁荣：你可以在 AWS 云服务上快速搭建东西，快速测试市场，尝试各种想法，然后去追逐一个 1 亿或 2 亿美元的市场。SaaS 导致了市场的巨大碎片化，因为交付机制变得非常容易且无缝。现在我们处于这样一个阶段：应用开始更多地关乎智能层。我就是不认为 20 年后软件会是一个巨大的市场。

And I'll say that pretty confidently because SaaS I don't think is going to be where you gain margin anymore. And we had this sugar high from, what, 2011, 2020 timeframe till now, of, you know, I can very quickly spin something up on an AWS cloud service, test the market quickly, try things out, and then go after a $100 million or $200 million market. SaaS was like massive fragmentation of the market because the delivery mechanism became very easy and frictionless. Now we're in this place where application starts to become more about the intelligence layer. And I just don't see software being a huge market in 20 years.

你不觉得——你认为 SaaS 会变成 Legora 那样吗？我的意思是，Legora 是 SaaS 吗？

Do you not think— do you think SaaS becomes Legora? And what I mean by that is, like, is Legora SaaS?

嗯，我的意思是，算是吧，企业级 SaaS，但也不完全是。不。所以，语义不就变了吗？我不认为有谁在创业时会说‘哦，我们是一家 SaaS 公司’。

Well, I mean, kind of, yeah, enterprise SaaS, but kind of not. No. So, does the semantics not just change? I don't think anyone really is starting fresh being like, 'Oh, we're a SaaS company.'

对。是的，我——所以，正是如此。我觉得你说得对。所以，我想的是 SaaS 的经典定义。但我认为，那些通过软件和服务的组合来交付成果的公司，绝对是。我们在几家公司看到了这一点，我在种子阶段非常积极地投资初创公司。我看到一批公司涌现出来，他们接手长期存在的工程任务，这些任务以前以传统方式完成，而且必须正确完成。他们通过使用 AI 加人类，使这些任务变得更快、更便宜。这很棒。那就是——我的意思是，我会说 Harvey 是律师事务所领域的第一个例子，你知道，旧的方式是我有一组律师去看合同、写合同之类的。他们基本上说，‘好吧，我可以用 AI 并用人类验证输出，用更少的人交付同样的东西，以同样的质量或接近同样的质量更快地交付，我将颠覆那个市场。’所以，我认为我们会看到——这就是新型的 SaaS。

Right. Yeah, I— So, exactly. I think you're right. So, I was thinking the classic definition of SaaS, I suppose. But I think companies where you deliver an outcome and you do so through a combination of software and services, absolutely. And we're seeing that in several companies and I invest pretty actively in startups at the seed stage. And I'm seeing a crop of companies come where they're taking engineering tasks that have existed for a long time that were done in a traditional way and have to be done correctly. And they're making those tasks much faster, much cheaper by using AI plus humans. It's great. And that's— I mean, I would say Harvey was a first example of this for law firms where, you know, the old way was I have a set of lawyers who go and look at contracts and write contracts, that kind of thing. And they basically said, 'Well, I can use AI and verify the outputs with humans and deliver the same thing with fewer humans and deliver it faster at the same quality or near the same quality and I will disrupt that market.' So, I think we're going to see it— this is sort of the new SaaS.

我能——我们看到了，你知道，我称之为‘词汇自慰’，就是风投们发明一套关于世界的理论。然后就像，你知道，万亿美元服务那回事。你知道，我觉得红杉资本做了一篇很棒的文章，我喜欢写那篇文章的 Julian。非常好。但就像，我们是不是只是稍微美化了一下？那只是未来：人们用 AI 高效地做事。这就像是对它的整个美化。有点过火了，不是吗？

Can I— we see it, you know, I call it word wank, which is where, like, you know, VCs come up with a coined theory of the world. And then it's like, you know, the trillion-dollar services thing. You know, I think Sequoia did a brilliant piece on it and I love Julian who did it. It's very good. But like, are we not just slightly jazzing it up? That's just the future: people using AI to do things efficiently. This is like a whole glamorization of it. A bit overkill, isn't it?

我觉得你说得对，但我也认为——

I think you're right, but I also think—

你这么做不就是在投资一家律师事务所吗？

Aren't you just investing in a law firm when you do that?

实际上是的，但它只是一个更好版本的律师事务所，对吧？那整个——

Effectively, yeah, but it's just a better version of a law firm, right? That's a whole—

而你愿意这么做吗？因为我遇到很多公司做的正是你说的，而我会想，我不想做律师事务所。就像，我很乐意做 Legora 或 Harvey，它们卖给律师事务所，但它们实际上不是律师事务所。你明白我的意思吗？

And you'd be okay to do that? 'Cause I meet a lot of companies doing exactly what you say and I'm like, I don't want to do a law firm. Like, I'm happy doing a Legora or a Harvey which sell to law firms but they're not actually a law firm. You get me?

嗯，但我们确实看到——一些大型律师事务所现在直接投资 AI。所以，我认为在极限情况下，你会看到 Harvey 基本上卖给律师事务所，因为律师事务所有渠道。那个渠道会持久吗？我不知道。但我也能想象他们说，‘你知道吗？我们可以直接卖给公司。’就个人而言，我会买那个。

Well, but we are seeing that— like some of the large law firms are now investing in AI directly. So, I think in its limits you'll see basically Harvey sells to the law firms because the law firms have the channel. Is that channel going to be enduring? I don't know. But I could also see them saying, 'Well, you know what? We could just sell directly to companies.' Like, personally, I would buy that.

如果 Harvey 说，‘嘿，我有一个解决方案，能处理你所有的保密协议，审查你的合同，所有这些事情，而且我能确信它有效’，我会把我 80% 的法律工作负载交给他们。

If Harvey said, 'Hey, I have a solution that does all your NDAs, it reviews your contracts, all these things and I can gain confidence that it works,' I would send 80% of my legal workloads to them.

我很感兴趣，因为你谈到一些公司正在构建自己的 AI。我想你看到法律领域的 K&E 承诺投入约 5 亿美元。Fireworks 的 Lynn，我是那里的投资者，他非常公开地谈到拥有自己的智能，以及公司利用自己的数据构建自己的模型的重要性。你认为那是一个非常可行的未来，还是实际上从一个非常特定的立场出发的有点技术乐观主义的预测？

I am intrigued because you spoke about some firms that are building their own. I think you saw K&E in the legal space commit like $500 million to it. Lynn at Fireworks, where I'm an investor, has spoken very openly about owning your own intelligence and the importance of companies leveraging their own data to build their own models. Do you think that is a very viable future or actually a slightly techno-optimist kind of prediction from a very specific standpoint?

我认为这有点技术乐观主义，因为我认为你必须拥有一定的规模才能构建自己的智能并拥有它。继续以律师事务所为例，如果你是 Cooley 或这些大所之一，是的，你可以投资这些东西，你可以让它实现。如果你是内布拉斯加州中部的一家夫妻店，你做不到。所以，你会把它作为一种服务来使用，对吧？你可能会用那种服务增强你的律师，这就是你为客户提供服务的方式。所以，我不认为每家公司都会以利用过去数据来改进自己模型的方式使用 AI。他们不会那样做。这对他们来说太难了。他们想要一个开箱即用的解决方案。所以，我认为销售这些开箱即用解决方案是有市场的。但它会是这个长尾。大玩家会开始定制，因为他们有规模来证明其合理性。这说得通。这又回到了这个想法：我们世界的每个部分都是对抗性的。这取决于你在和谁竞争，对吧？如果你在和另一家大型律师事务所竞争，而那家大所使用 AI 来提升他们的利润率，那么你也得这么做。夫妻店可能不会这么做，他们竞争的人也不会这么做。所以，他们会在另一个领域停留一段时间。

I think it's a bit techno-optimist because I think you have to have a certain amount of scale in order to build your own intelligence and own it. Keep going with law firms, if you're Cooley or one of these big firms, yeah, you can invest in these things and you can make it happen. If you're a mom-and-pop shop in the middle of Nebraska, you can't do that. So, what you're going to use that as a service of some sort, right? You may augment your attorneys with that service and that's how you'll provide service to your clients. So, I don't see this as every company will use AI in a way of leveraging their past data to make their own models better. They're not going to do that. It's just too hard for them. They want an out-of-the-box solution. So, I think there is a market for selling those out-of-the-box solutions. But it's going to be this long tail. The big players will start to customize because they have the scale to justify it. It makes sense. And it goes back to this idea that every part of our world is adversarial. Like it depends on who you're fighting, right? If you're fighting another big law firm and that big law firm goes and uses AI to supercharge their margins, well, you kind of have to do the same. The mom-and-pop shop probably isn't doing that and the people they're fighting against are not doing that. So, they're going to stay in this other realm for some time.

我认为支撑大量法律编程的一件事是，我们将把工资的百分比转化为技术支出。

I think one thing that underpins a lot of law coding is the percentage of salary that we will then transition to technology spend.

是的，没错。

Yeah, that's right.

当我们看今天的律师时，你知道，我认为美国法律市场大约 3000 亿美元，你看着它想，那么这些提供商将吃掉多少？我认为你在编程领域也看到同样的问题。你看到 Benioff 说，他们将在 60 亿美元的开发者工资预算中花费 3 亿美元在核心编程工具上，显著地。我的问题是，对开发者来说，这个比例会保持在 5% 吗？会上升到 30%、50%，还是会降到 1%？因为那是关键问题。

When we look at lawyers today, you know, I think it's about a $300 billion legal market in the US and you're looking at that going, well, how much of that will be eaten by these providers? I think you see that same question there with coding. You saw Benioff say that they'll spend $300 million of a $6 billion developer salary budget on core coding tools significantly. My question to you is will that stay at 5% for developers? Will that go up to 30%, 50% or will it come down to 1%? 'Cause that is the mother lode question.

我实际上认为百分比可能会保持稳定，因为我们最终会遇到这个验证问题。所以，这很难，因为训练人类知道什么是好的实际上是困难的部分。所以，开发者，实际上不是关于你今天直接产出什么。不是像代码行数那样。

I actually think the percentages will remain maybe sort of stable because we end up with this verification problem. So, it's hard because training humans to know what's good is actually the hard part. So, developers, it's actually not about what your direct output is on this day. It's not like lines of code.

实际上重要的是代码质量。那些非常有经验、顶尖的开发者，那些 10 倍开发者，他们培养出了对什么是好的、什么看起来不错、什么在美学上正确的感觉，并且能快速让东西运转起来。在这个新世界里，那些编码者、那些知道什么是好的人，实际上可以验证模型的输出并变得更高效。所以，我认为发生的是我们只是生产更多。市场规模，比方说我们支付人类加 AI 的费用，并没有剧烈变化，只是他们的产出大大改变。这就是我的看法，因为让我们稍微展开一下。假设我在一家公司每年花 1000 万美元开发我的软件。现在，突然之间，我的软件客户开始有不同需求。10 年前，他们对 1 年的发布周期感到满意。现在，他们说：‘嘿，我想在一周内修复我的 bug。’所以，唯一的方法是实际上将 AI 与人类结合使用。所以我并没有做得更少。我实际上做得更多。我在一周内修复那些 bug，最终在一小时内修复它们，但我仍然支付 1000 万美元来继续开发整个东西。我需要那些能验证输出并实际朝着有用目标前进的人类，修复 bug、添加功能。所以，我认为交付速度上升是因为需求方开始要求它。我认为这里被忽略的关键点是：并不是每个人的工作都消失，因为同样的任务可以用 AI 加人类以更低成本完成，而是现在——是的，那是真的，但我现在会做更多那些任务，因为需求方增加了。你不想只得到一个而没有另一个。当供应变得更便宜时，它也会改变需求方。所以，我认为将会发生的是我们最终会做更多。我们有更多的发布周期、更多的变化。就像这里有一个人们可以理解的物理例子。假设今天我们有建造建筑的速度。如果我想建一座摩天大楼，比方说需要 2 年的建设时间。那是基于 X 数量的人类拥有 Y 套技能。那么，假设突然我有机器人可以完成其中 80%。会发生什么？好吧，可能会有短暂的体力劳动中断。但然后你会说，你知道吗？突然之间我建造同样摩天大楼的成本和时间——原来是 2 年，但现在 80% 的任务由机器人完成，我可以把整个时间缩短到 8 个月。所以，并不是你只是建造同样数量的摩天大楼，只是用 8 个月完成，你会建造两倍的摩天大楼。会有更多东西在运转，更多建筑被拆除，更多劳动力进入那个市场。所以，项目总数会上升。投入所有这些项目的美元金额可能相同，但将是人类和机器人的组合来建造它们。

It's actually like quality of code that matters. So, very experienced, very top-notch developers, those 10x developers, whatever, they develop the sense of what is good, what looks good, what's aesthetically right, and they can make things work fast. In this new world, those coders, those people who know what's good, can actually verify the output of models and become more efficient. So, I think what happens is we just produce more. And the market size, let's say that we're paying for humans plus AI, doesn't change drastically, just their output greatly changes. That's the way I look at it because let's play it out a little bit. So, I'm spending in a company, let's say, $10 million a year to develop my software. Now, all of a sudden, my customers of that software start having different demands. 10 years ago, they were happy with 1 year release cycles. Now, they're saying, 'Hey, I want my bug fixed in a week.' And so, the only way to do that is actually to use AI with the humans. So, I'm not doing less. I'm actually doing way more. I'm fixing those bugs in a week and eventually fixing them within an hour, but I'm still paying $10 million to keep developing this whole thing. And I need the humans that can verify those outputs and actually move to a goal that's useful, fixing the bug, adding features. So, I think the pace of delivery goes up because the demand side starts to demand it. And I think that is the key thing here that's being left out: it's not that everyone's job goes away because the same task can be done for less with AI plus humans, it's that now I—yes, that is true, but I will now do way more of those tasks because the demand side has increased. You don't want to get one without the other. When the supply gets cheaper, it changes the demand side as well. So, I think that's what's going to happen: we just end up doing more. We have more release cycles, more changes. Like, here's a physical example that people can get their heads around. Let's say today we have a rate of building a building. If I want to build a skyscraper, it takes 2 years of construction, say. And that's based upon X number of humans with Y sets of skills. So, let's say all of a sudden I have robots that can do 80% of that. What's going to happen? Well, okay, there might be a disruption in physical labor for a moment. But then you'll say, you know what? All of a sudden my cost and time to build that same skyscraper—it was 2 years, but now 80% of this task is done by robots, I can bring the whole thing down to 8 months. So, it's not that you're just going to build the same number of skyscrapers and just do it in 8 months, you're going to build twice as many skyscrapers. There's going to be more things churning, more buildings being torn down, more labor going into that market. So, the total number of projects will go up. The amount of dollars going into all of those projects may be the same, but it'll be a combination of humans and robots to build them.

Navin，你说你显然做天使投资——你天使投资非常成功。你有点轻描淡写了。我认为我今天投资的地方，一个逆势的想法是我不喜欢机器人投资。我认为我们离我们需要达到的地方还很远，而且我确实认为时机很重要。现在，我不是说驱动机器人的基础设施——我认为围绕它们的数据集非常有趣。我们看到很多人像 Mecha 或 Micro AGI，他们实际上在为此建设基础设施。但就像实际的应用层，我们看到像机器人清洁工、机器人建筑工，我们离我们需要达到的地方还很远。我不想在那里，我不想承担时机风险。我错了吗？我天真吗？你比我聪明得多。

Navin, you said that you obviously angel invest—you angel invest very successfully. You kind of downplayed it. I think where I invest today, a contrarian thought for me is I do not like robotics investments. I think we are so far from where we need to be and I do think that timing is important. Now, I'm not saying the infrastructure to power robotics—I think the data sets around them are very interesting. We see a lot of people like Mecha or Micro AGI, they're actually in the infrastructure build out for it. But like the actual application layer where we see like robot cleaners, robot builders, we are so far away from where we need to be. I don't want to be there and I don't want to take the timing risk. Am I wrong? Am I naive? You're much smarter than me.

嗯， again，我对此非常坚信，所以我会强烈不同意。

Well, again, I'm pretty pilled on this one, so I am going to disagree strongly on it.

好吧，当你说我们离得很远时，把它放在时间框架里。你认为时间线是什么？

Okay, when you say we're so far away, put that in some of the time. What do you think is the timeline?

呃，五年以上。

Uh five plus years.

好吧，我们在这点上分歧不大。我认为我们将在未来两到三年内有一些任务的工作原型。市场将在五到七年内出现。我认为这是对的。

Okay, we're not so far apart on this. I think we'll have working prototypes for some of these tasks in the next two to three years. And the market will emerge in five to seven years. I think that's right.

我认为人形机器人和非人形机器人之间也有明确的界限。我的意思是，非人形机器人绝对能看到更清晰的路径。人形机器人我真的只是震惊——每当我認為某件事真的很明显时，我总是想我一定大错特错。

I think there's also a clear line of delineation between our humanoids and non-humanoids. I mean like non-humanoids absolutely see that clearer pathway. The humanoids I'm really just struck—Whenever I think something's really obvious, I'm always like I must be very wrong.

因为人形机器人显然离得很远。是的。

Because the humanoids is so obviously so far off. Yeah.

你知道，我认为我们将处于这个奇怪的世界，我们生活在指数发展周期中。真的很难应用我们的旧模型。这就是问题所在。

You know, I think we're going to be in this weird world where we live in exponential development cycles. It's really hard to apply our old models. That's the problem.

是的。

Yeah.

确实如此。所以，今天看起来很远的东西，在指数曲线下最终会比你想象的快得多。就像 TAM——我是说市场规模——你知道，就像我认为推理市场。我现在知道什么，我希望 24 个月前就知道？推理市场就像无界的一样。

It is. So, something that looks like so far away today with an exponential curve ends up being much sooner than you thought. And like TAMs—I mean, market size—you know, like I think the inference market. What do I know now that I wish I'd known 24 months ago? That the inference market was just like unbounded.

它是无界的。它是数万亿美元。就像如果你 10 年前问我，什么是万亿美元市场，那是基本需求。是住房、食物、交通，对吧？你永远不会想到科技中的某个东西会成为万亿美元市场。而且我认为仅仅 AI 模型本身在未来几年就会成为一个万亿美元市场。而推理市场将在 10 年内成为一个 10 万亿美元市场。所以想想这很疯狂，对吧？而针对你的观点，我们已经——我们正在将部分劳动力市场和部分价值生成栈转移到技术。所以这就是我们获得这些巨大 TAM 的地方。所以你知道，我看待这个的方式，我认为这是一个神奇市场的非常规原因是，如果我们在交付一个 token 的效率上敲掉一个零，一个数量级，那在某些方面有效地将市场扩大 10 倍，对吧？所以，某个东西曾是万亿美元市场，现在突然你有更多的适用性。潜在用途的基础增加了超过 10 倍。这是杰文斯悖论。所以，我认为我们处于这个位置，我们可以建立这些指数并释放巨大的价值和市场。所以我实际上认为机器人将比智能体式编码大 100 倍。

It's unbounded. It's multi-trillions. Like if you asked me 10 years ago, you know, what were trillion-dollar markets, it's basic needs. It's housing, food, transportation, right? You would never have thought something in tech would be a trillion-dollar market. And I think just AI models themselves will be a trillion-dollar market in the next few years. And the inference market is going to be a 10 trillion-dollar market within 10 years. So this is insane to think about, right? And to your point, we have—we're moving some of that labor market and part of the value generation stack to technology. So that's where we're getting these giant TAMs. So you know, the way I look at like unconventional reason that I think is a fantastic market to be in is that if we knock off a zero, one order magnitude efficiency on delivery of a token, that effectively 10x the market in some ways, right? So, something was a trillion-dollar market, now all of a sudden you have way more applicability. The base of potential uses just increases more than 10x. This is Jevons paradox. So, I think we're in this place where we can build these exponentials and just unlock massive value and market. So, I actually think robotics is going to be 100 times larger than agentic coding.

嗯，那有点不同意我，Naveen。在我们快速提问之前，我们能展开一下吗，因为我不想完全愚蠢。你为什么这么说？

Well, that is slightly disagreeing with me, Naveen. Can we just like before we do a quick fire, can we just unpack that cuz I don't want to be completely moronic. Why do you say that?

我认为我们仍然在如何看到智能输出的表面上 scratching。

I think we're still scratching the surface of how to see the output of intelligence.

所以，物理世界是我们从未真正涉足智能的领域。我认为今天的机器人，比如工业机器人，并不是真正智能的。它们是写死的。我们把自己拥有的智能——比如如何完成一项任务——灌输给它们，然后它们只是照做。那不是真正新颖的问题解决。我们将在不久的将来进入物理问题解决的领域。想象一个世界，我有数十亿机器人在四处运行，它们真正动态地解决问题。比如，那是个问题。那东西刚坏了。我要去清理它。我要去替换它。在我们的世界里，清理一切、建造新建筑、解决物理问题的交付成本不会降到零。我讨厌那些降到零的说法。它会大大降低。所以，我认为那个市场将是巨大的，这就是为什么我认为它会比我们今天在虚拟世界中看到的要大得多。

So, the physical world is something we've just not been in at all for intelligence. I would argue robotics today, like industrial robotics, are not really intelligent. They're wrote. We're imbuing them with intelligence that we have, like how to do a task, and we're just doing it wrote. That's not really novel problem solving. We will be in the realm of physical problem solving not in the not-so-distant future. Imagine a world where I have billions of robots running around and they actually solve problems dynamically. Like, that's a problem. That thing just broke. I'm going to go clean it up. I'm going to replace it. The cost of delivery of cleaning everything up in our world, building new buildings, solving physical problems will not drop to zero. I hate those drop-to-zero kind of statements. It will greatly reduce. So, I think that market will be enormous, and that's why I think it's just going to be much bigger than what we see today in the virtual world.

我能直说吗？那么你具体投资机器人的哪个环节？比如，驱动它们的模型公司，驱动模型公司的数据公司，消费者应用层，哪里？

Can I be blunt? Where specifically then do you invest in robotics? Like, in the model companies that power them, the data companies that power the model companies, the consumer application layer, where?

我认为还不是应用层。我更多在技术栈。所以，我实际上认为那些做模型的人还没有真正弄清楚它们长大后会长什么样。这没关系。我不是说他们需要。我不认为价值会在模型里。我在投资那些构建模型的公司。我认为他们可能最终会构建端到端设备。他们会构建带有模型的机器人，该模型能非常非常出色地完成某些任务。他们就是这样销售的。我个人的看法是，它会结合在一起。不会有分离。

I don't think the application layer yet. I'm more in the technology stack. So, I actually think these guys who are doing the models haven't really figured out what they're going to look like when they grow up. And that's okay. I'm not suggesting they need to. I don't think the value will be in the model. I am investing in companies that are building models. I think they are probably going to end up building end-to-end devices. They're going to build the robot with a model that does something very, very well. That's how they're going to sell it. My personal take is that it's going to go together. You're not going to have a separation.

当 TAM 如你所说时，你在意价格吗？

Do you care about price when the TAM is what you said it was?

不太在意。不。我只是把它看作，我需要把我财富的一部分投入到那个领域，因为它很重要。那里会有大量增长。所以，有趣的是，我们处于这样一个地方，价值投资者或随便你怎么称呼，比如传统风投会说：“哦，你知道，那定价过高了。”伙计们，当我们谈论万亿级以上的 TAM 时，价格几乎无关紧要。

Not really. No. I just see it as like, I need to put some percentage of my wealth into that space because it's important. And that's where there's going to be a ton of growth. And so, it's interesting because we're in this place where value investors or whatever you want to call it, like, traditional VCs are like, "Oh, you know, that's way overpriced." Like, guys, when we're talking about trillion-plus dollar TAMs, the price is almost irrelevant.

我完全同意你。挑战在于，我们经常也有客户，或者老板，就是 LP。而 LP 是一个更滞后的指标，他们认为所有权和价格非常重要。当你出现在年度股东大会上说：“嘿，所有权下降了，价格上升了。”是的。他们会说：“去你的。”

I completely agree with you. The challenge is very often we have a customer, too, or a boss, too, and it's LPs. And the LPs are an even lagging indicator who think that ownership and price are very important. And when you rock up to your AGM and you say, "Hey, ownership's gone down and price has gone up." Yeah. They go, "Fuck you."

他们确实会。直到他们开始看到大公司也有同样的趋势，他们就会说：“啊，好吧，也许这就是趋势，对吧？”

They do. Until they start to see the big firms also have the same trend and they're like, "Ah, okay, maybe that's the way it's going, right?"

如果你可以把所有钱投入一家不是你自己的公司，你会选哪家？

If you could put all of your money into one company that is not your own, what would it be?

嗯，可能像 Physical Intelligence 这样的公司。他们把它当作模型来思考。我认为还有几家公司也在这样做。比如，我觉得 Rota 和其他几家也在这样思考。在我看来，他们正在解决关于如何定义任务然后执行的根本问题。所以，将任务转换为电机输出。对我来说，这是目前物理世界要解决的问题。所以，除了我们正在做的事情和构建新计算机之外，我认为那是下一个大问题。

Yeah, probably like a physical intelligence. They are thinking about it as a model. I think there's several companies doing that as well. Like, I think Rota and a few others are thinking about it like this. I think they are solving, in my view, something fundamental about how you define a task and then execute it. So, conversion of a task to motor output. To me, that's the problem to solve right now for the physical world. So, outside of what we're doing and building a new computer, I think that is the next big problem.

你在多家公司都取得了令人难以置信的成功。关于拥有金钱，没有人告诉过你但你希望有人告诉过你的是什么？

You've been incredibly successful across companies. What does no one tell you about having money that you wish you had been told?

我认为它不会改变你的价值观。你仍然有有限的资源，但同时你仍然以某些方式看重某些东西。比如我在某些事情上仍然有点小气。其实没有理由这样。

I think it doesn't change your values. You still have a finite amount of resources, but also at the same time you still value certain things in certain ways. Like I'm still kind of cheap when it comes to certain things. There's no reason for it to be.

你在哪些方面小气？

What are you cheap with regards to?

嗯，如果你问我妻子，可能是衣服之类的。我就会想：“什么？我才不会为一条牛仔裤付那个价？”我觉得旅行也是。比如，我到了会尝试优化时间的地步，但总体上我会想：“啊，那太贵了。我就不做了。”我花不花那个钱其实没什么区别。我只是看不到价值。所以，我觉得就是这样。我把钱投入我看到价值的东西，而且我可以在那些东西上投入更多钱。但在那些我看不到价值的地方，我实际上不会多花一分钱。

Well, if you ask my wife, probably clothes, that kind of a thing. I just like, "What? I'm not going to pay that for this for a set of jeans?" I think travel still. Like, I'm getting to a point where I will try to optimize my time, but overall I'm like, "Ah, that costs too much. I'm just not going to do it." It wouldn't have mattered if I spent the money or not. I just don't see value. So, I think that's the thing. I put money into things where I see value, and I can put a lot more money into those things. But, the places where I don't, I actually don't spend any more.

伙计，我们来做快速问答。我说一个简短陈述，你给我你的即时想法。过去 12 个月你改变了什么想法？

Dude, we're going to do a quick fire round. So, I say a short statement, you give me your immediate thoughts. What have you changed your mind on in the last 12 months?

是的，我认为在投资方面，我改变了对有用进展指标的看法。我们总是把它看作：“哦，我们需要把市场验证作为主要指标。”我认为那是一种旧的看法。事实上，我认为它很大程度上来自 SaaS。因为当没有深层技术根源时，市场验证是有意义的。比如，当你说：“嗯，我在构建一个服务于某种目的的产品”，那全是关于市场验证。但是，当我拥有某种几乎一眼就能看出是真实的东西时，比如我要治愈癌症。你不需要验证市场。你只需要验证技术问题的进展。现在，投资者分为两个阵营。一些人说，好吧，让我们大力投资这些公司，估值见鬼去吧。你知道，我们就是那些公司之一。另一边则说，天哪，那是不自律的。就像，你没有市场验证。我实际上认为那根本上是错误的。许多成长型投资者仍然属于那一类。所以，我成长过程中读了很多关于如何看待投资风险的书，我在这方面的想法发生了相当大的改变。不仅仅是因为我的公司在这样做，而是我认为现在我们处于解决真正深层技术问题的位置。机器人、计算、生物学。这不是关于市场验证。市场就在那里。市场非常明显在那里。你解决了那个问题，100% 的人会购买它。这是关于解决问题的进展。

Yeah, I think in terms of investing, I've changed my mind on useful metrics for progress. We always looked at it as like, "Oh, we need to look at market validation as the main metric." And I think that's an old way of looking at it. In fact, I think it comes a lot from SaaS. Because market validation makes sense when there's not a deep technology root. Like, when you're saying, "Well, I'm building a product that serves a purpose," it's all about market validation. But, when I have something that is on its face almost true by inspection, like I'm going to cure cancer. You don't need to validate the market. You just need to validate the progress on the technical problem. Right now, investors are in two camps. Some are saying, well, let's invest heavily in these companies, valuation be damned. You know, we're one of those companies. And the other side is like, oh my god, that's undisciplined. It's like, you don't have market validation. I actually think that's fundamentally wrong. And a lot of growth investors are still in that bucket. So, I kind of grew up reading all the books about how you look at risk in investing, and I've changed my mind quite significantly on this. Not just because my company is doing it, but I think now we're in this place where we're solving really deep technical problems. Robotics, computing, biology. And it's not about market validation. The market is there. The market is very clearly there. You solve that problem, 100% people are going to buy it. It's about progress of solving a problem.

是的，尤其是在欧洲，这是一种罕见的心态。如果你在职业上完全不受资源限制，你会做什么不同的事？

Yeah, that's a rare mindset to have in Europe especially. If you were completely resource unconstrained, professionally speaking, what would you do differently?

我会开始建立自己的制造。因为我想越来越接近原材料的源头。

I would start to build my own manufacturing. Because I want to go closer and closer to the source of raw materials.

我认为我们现在所处的世界，从硬件生态系统的角度来看，有太多的供应限制。如果我不受限制，可以建造任何东西，老实说，这就是你想做的事情。是的，我们应该建造自己的晶圆厂，因为我可以针对我的需求对那一层进行超优化。我不需要拿一个为其他 100 种用途而制造的商品来适应我的需求。我会建造我需要的精确的东西，并打造更好的产品。坦率地说，我认为计算机的所有价值都来自于提供智能。如果我能一直超优化到处理从工厂运来的硅锭以及如何切割它，我绝对想这样做。

I think we're in this world now where there are so many supply constraints from the hardware ecosystem point of view. If I were unconstrained and I could build anything, this is kind of what you want to do, to be honest. Yeah, we should build our own sorts of fabs because I can hyper-optimize that layer for what I need. I don't need to take a thing that is a commodity that's been built for 100 other purposes and make it fit my needs. I would build exactly the thing I need and build a better product. And frankly, I see it as all the value of a computer comes from delivering intelligence. If I can hyper-optimize that all the way down to how I process the silicon ingot when it comes from the factory and how I slice it, I absolutely want to do that.

很好的回答。你认为当今 AI 中最被高估的领域是什么，你不会投资哪里？

Great answer. What area of AI do you think is most overrated today, where you would not invest?

所以，对我来说，我认为客户支持 AI 是最过度泛滥、缺乏差异化的领域之一。它就像一个试图杀死 Salesforce 的切入点，这相当无趣。

So, like for me, I think customer support AI is like one of the most over-proliferated, non-differentiated spaces. It's just like an entry point to try and kill Salesforce, which is quite uninteresting.

是的，我认为那些快速见效的领域，比如你有一个标准的事情由人类完成，我只是尝试自动化它。我认为那是现在最糟糕的投资领域。我认为它非常被高估。我实际上确实认为智能体式编程在市场方面相当被高估，因为我认为它转移到模型提供商。我不认为你从其他人那里获得很多价值。我知道 Cursor 有一个很好的退出，那很棒。但模型提供商，模型本身才是大部分价值所在。我不认为它是工具链或 IDE 那么多。它不是超级粘性的。我在自己的公司里看到，人们在这些东西之间流动得相当流畅。所以，重要的是模型以及它提供的能力。所以，我认为智能体式编程作为一个市场，可能，你知道，除了模型提供商之外，是我今天不会投资的。

Yeah, I think those quick hits where you have a standard thing that's done by humans, I'm just going to try to automate it. That's the worst place to invest right now, I think. I think it's very overrated. I actually do think that agentic coding stuff is quite overrated in terms of a market, because I think it moves to the model provider. I don't think you get much value from the others. I know Cursor had a great exit and that's wonderful. But the model providers, the model itself is where most of the value is coming from. I don't think it's the harness or the IDE so much. It's not super sticky. I've seen it in my own company, people move between these things pretty fluidly. So, it's the model that matters and what capabilities it delivers. So, I think agentic coding as a market might, you know, modulo the model providers, is something I wouldn't invest in today.

3 年后，开放模型与封闭模型的 token 流量各占多少百分比？

What percent of token traffic will flow through open versus closed in 3 years' time?

我认为 80% 将是开放的。

I think it'll be 80% open.

那在美元流量方面意味着什么？

What does that translate to in terms of dollar traffic?

那实际上是个更好的问题。我认为它可能会收敛到非常像手机市场的样子。粗略地说，iPhone 占总市场的 20%，但占利润的 80%。类似这样。我们可能会看到，有相对少数非常重要的问题需要前沿模型，你会为那些东西支付更多。然而，你做的所有常规任务将由开放模型处理。所以，我实际上认为支出上大致是 50/50 平分，但就 token 而言，只有 20% 会流向前沿模型。

That's a better question, actually. I think it's going to probably converge on something that looks very much like the cell phone market. Just roughly speaking, iPhone is 20% of the total market, but 80% of the margin. Something similar to this. We'll probably see that there's a relatively small number of very important problems that frontier models are needed for, and you'll pay much more for those things. However, all of the routine tasks that you do will be handled by open models. So, I actually think there'll be a rough 50/50 split on spend, but in terms of tokens, it'll be only 20% to the frontier models.

你参加过的最棒的一次投资者会议是和谁，为什么？

Who is the single best investor meeting you've ever had and why?

我可以先讲一个小故事。

I can give a bit of a story for it first.

好的，请讲。

Yeah, go on.

我想，你知道，自从我开始 Nirvana 以来，我大概开过，我不知道，一千次投资者会议，对吧？所以，在第 800 次之后，第 801 次就不再留下印象了。所以，很早的时候，我和 Steve Jurvetson 开了一次会，他现在已经成为我的好朋友，我当时紧张得要命，因为我听过关于这个人的太多神奇的事情，超级聪明，他会把我拆穿，你知道，我被告知的就是这样。我和他进行了一次对话，我们只是开始对数值表示极客式地讨论。我当时想，这家伙是投资者？然后我发给他几篇论文，他像批注一样标记并提问。我把回复发给他，我想，这从来不是我期望从投资者那里得到的。你知道，这家伙在某种意义上像技术同事一样和我交谈。所以，对我来说，那是我有过的最好的投资者会议之一，因为我们做的事情如此新，还没有人推销过。2014 年，10 年来没有人推销过硬件初创公司。而这个人几乎迫不及待地想投资这样的东西，因为他自己一直在思考。所以，那就像是，砰，我们走吧。

I think, you know, since I started Nirvana, I've probably had, I don't know, a thousand investor meetings, right? So, after the 800th one, the 801st one doesn't really leave an impression anymore. So, very early on, I had a meeting with Steve Jurvetson, who has now since become a good friend of mine, and I was nervous as hell because I'd heard so many amazing things about this guy, super smart, he's going to take me apart, you know, that's kind of what I was told. I had a conversation with him which we just started geeking out on numeric representations. And I was like, this guy's an investor? And then I send him a couple papers, he marks them up and asks questions. I'm sending responses back to him and I'm like, this is not what I expected from an investor, ever. You know, like, this guy's talking to me like a technical colleague in a sense. So, to me, that was one of the best investor meetings I had because we were doing something so new that no one had pitched it yet. No one had pitched a hardware startup in 10 years in 2014. And this guy was almost chomping at the bit to invest in something like this because he'd been thinking about it himself. And so, that was just like, boom, let's go.

很好的回答。我喜欢 Steve。我记得请他上过节目。他太棒了。你说到愿意非共识，这就是我喜欢他的地方。

It's a great answer. I love Steve. I remember having him on the show. He was fantastic. You talk about being willing to be non-consensus, is what I love about him.

没错。你必须这样才能做出会得到回报的大赌注。根据定义，共识已经是错的。对吧？它已经过时了。所以，你必须是非共识的。

That's right. You have to be to make big bets that will pay off. By definition, the consensus is already wrong. Right? It's already old. So, you have to be non-consensus.

不过，我在欧洲发现的挑战是，你有想法的商品化。我的意思是，现在当创始人是一件相当游客化的事情。每个人都想为了体验而成为创始人。但想法的质量降低了。我认为我不认为美国人天生更好、更聪明、更有才华，但我确实认为你们在想法选择上更具创新性。那可能是因为 exposure，你看到什么，你如何思考，在结构中哪里玩。你知道，我们还在用 AI 做费用管理。

The challenge I find with being in Europe though is that you have like the commoditization of ideas. And what I mean by that is actually, now being a founder is quite a touristy thing to do. Everyone wants to be a founder for the experience. But the quality of ideas has reduced. And I think I don't think Americans are inherently better, smarter, more talented, but I do think you are more innovative around idea selection. And that's probably because of exposure, what you see, how you think, where to play in the structure. You know, we're still doing expense management with AI.

是的。不，完全同意。我认为在这里我们被允许想得更大。

Yep. No, totally. I think we are given the permission to think bigger here.

是的。在欧洲，你不是，因为整个投资者网络和人们的思维方式是保守的。你知道，如果你带着一个疯狂的想法进来，他们会说，“那家伙疯了。我不会投资他。”而在这里，就像是，“那家伙很棒。他在想大的。”对吧？

Yeah. In Europe, you're not because the whole investor network and the way people think is conservative. You know, if you come in with a crazy idea, they're like, "That guy's mad. I'm not going to invest in him." Whereas here, it's like, "That guy's awesome. He's thinking big." Right?

是的。不，我完全同意你。这绝对正确。困难的是，我遇到了这个问题，我有一个真正的……好吧，让我们这样做。现在有一个 A 轮，他们都在建造太空相关的东西。我完全不懂。团队很棒。他们很厉害。但是，老兄，从 4 亿美元的基金中开出 2000 万美元的支票。那是 5%。你不能那么容易地失去它。老兄，这是太空领域的超级高风险游戏，我完全不明白。

Yeah. No, I completely agree with you. It's absolutely right. The hard thing is and I've got this problem I've got this genuine Okay, let's do this. There's a series A now and they're all building space stuff. I've got no clue. The team is great. They're awesome. But dude, it's 20 million check from a 400 million fund. That's a It's 5%. You can't lose that that easily. Dude, this is super risky play in space and I do not understand it at all.

我有点把这些看作，你知道，在技术周期的早期，像我们所有技术专家都会给你一个故事，而我们需要做的就是构建一个叙事。那个故事会是错的。100% 会是错的。我不知道别的，但我确实知道它是错的。我不知道它具体错在哪里。我不知道它会错得多离谱，但它是错的。而你在那个阶段投资的，我认为是开车的人，而不是车本身。对吧？所以，想法是对的，方向是对的，团队才是你关心的。他们能开动它吗？他们能做出必要的转型和正确的技术决策，真正把车开到重要的地方吗？

I sort of look at these as, you know, in the early part of the technology cycle, like, all of us who are technologists are going to give you a story and that's what we need to do is build a narrative. That story is going to be wrong. 100% will be wrong. I don't know anything else, but I do know it's wrong. I don't know exactly where it's wrong. I don't know how wrong it'll be, but it is wrong. And what you're investing in, I think, at that stage is someone to drive the bus, not the bus itself. Right? So, the idea is right, the direction is right, and the team is what you care about. Can they drive that? Can they make the requisite pivots and right technology decisions to actually drive the bus to something that matters.

那么在这个市场里，挑战实际上变成了你没有时间和他们相处。你可能只有两到三次会议。

The challenge actually then becomes in this market, you don't have the time with them. You probably got two to three meetings.

是的，我很想说没错，但我不知道。你说得对。你得尽可能多做功课，和他们认识的人聊聊等等，但这很难。是的，我同意。

Yeah, I would love to say yeah, I don't know. You're right. And you got to do as much work as you can, talk to people that they know and whatnot, but it's hard. Yeah, I agree.

最后一个问题，老兄。你最兴奋的是什么？我喜欢乐观。我不喜欢我们对泡沫的冷嘲热讽，你的是什么？比如“我就是为此感到兴奋”？

Final one for you, dude. What are you most excited for? I like optimism. I don't like the cynicism that we have around the bubbles and what's your like I'm just pumped for this?

我把 AI 看作一种工具和能力，它使我们能够提出关于宇宙的更大问题。科学中有太多我们不了解的东西，这是因为它们难以接近。它们难以接近是因为能思考那个问题并真正取得进展的人太少了。我们可以着手并开始解决的问题范围开始扩大。对我来说，这极其令人兴奋。我不认为“天哪，AI 会杀死所有人类”。这是我听过的最愚蠢的世界观。对我来说这简直是白痴。我现在完全厌倦了这种说法。就像，不不，伙计们，AI 是使人类真正开始破解这些更大问题的东西。比如，我们如何探索更多的宇宙？当我们开始理解物理学更深层次的部分时，我们就能做到。我们有虫洞之类的想法。这些都是理论构造。我们现在可以开始使用 AI 并开始取得进展。我真的能建造出利用那种物理现象的东西吗？或者它是真的吗？所以，这些是我们开始可以问的问题。在过去，我们就是做不到。我们没有足够多的人有能力接近这些问题。所以，对我来说，推动科学前进和推动人类理解前进的能力是如此令人兴奋。我甚至无法告诉你。它让我感觉像个孩子。外面有一个我们甚至无法想象的神奇世界。我们被自己的能力的局限所吓倒，认为这就是我们能看到的全部。而现在突然间，整个世界都向我们敞开了。

I look at AI as a tool and capability that enables us to ask much bigger questions about the universe. There's so many things we don't understand in science, and it's because they're unapproachable. They're unapproachable from just the number of humans who could think about that problem and actually make progress. The surface area of problems we can go after and start to solve starts to increase. And to me, that's extremely exciting. I don't look at, oh my god, AI is going to kill all humans. That's the stupidest worldview I've ever heard. It's just idiotic to me. I actually completely tired of it now. And it's like, no no guys, AI is the thing that enables humans to actually start to crack these bigger problems. Like, how can we explore more of our universe? We can do that when we start to understand deeper parts of physics. We have these ideas of like wormholes and these things. These are all like theoretical constructs. We can now start to use AI and start making progress. Can I actually build something that exploits that? That physical phenomenon. Or is it real or not? So, like these are questions we can start to ask. In the past, we just couldn't. We didn't have enough humans that were capable of approaching these questions. So, to me, the ability to move science forward and to move human understanding forward is so exciting. I can't even tell you. It makes me feel like a kid. There's a world of wonder out there that we just couldn't even think about. We got browbeaten into our own capabilities of like this is all we can look at. And now all of a sudden that whole world is open to us.

我的意思是，像这样的节目就是为什么我做了 10 年仍然喜欢做这个节目。就像你的知识广度真是太棒了。老实说，能见证这一点对我来说太神奇了。所以感谢你如此愿意参与如此广泛的领域。你太棒了，老兄。

I mean, shows like this are why I still love doing this show after 10 years. Like the breadth of your knowledge is just like wonderful. It's amazing for me to witness, honestly. And so thank you for being so willing to engage across such a broad spectrum. And you've been amazing, dude.

非常有趣。我喜欢这些对话，我们可以即兴讨论这么多不同的话题。

It's a lot of fun. I love these conversations where we can riff on so many different topics.