AI 先驱吴恩达谈开放模型与美国竞争力

如果美国的对手想拖慢我们的脚步，我认为他们几乎找不到比这些关于建设数据中心的愚蠢暂停令更好的办法了。我一直对少数企业进行的游说感到震惊，他们说开放模型是危险的。我认为这是错误的。我认为蒸馏是一个主要因素的说法被夸大了。有时世界上最强大的力量是看不见的。一个驱动可能性的关键智能层。这是美光内存和存储。发现新疗法、模拟气候变化、建设更智能、更安全的城市的智能。AI 的每一次突破都由数据驱动。而数据存在于美光内存和存储中。大家好，我是《华盛顿邮报》副观点编辑詹姆斯·H·霍姆曼。欢迎收听《邮报》的“建设美国”播客，探讨美国如何更好地建设、制造和创新未来。今天，我们邀请到安德鲁，讨论投资 AI 基础设施将如何塑造美国的长期竞争力。安德鲁是 AI 领域的真正先驱。他领导了谷歌大脑团队，该团队启动了这家科技巨头的大部分 AI 工作。他曾指导斯坦福 AI 实验室，并撰写了超过 200 篇关于 AI 及相关问题的论文。此后，他共同创立了 deeplearning.ai，提供技术培训。他还是一家风险投资基金的执行普通合伙人。他同时也是亚马逊的董事会成员，其执行董事长杰夫·贝索斯拥有《邮报》。安德鲁，欢迎你。

If an adversary of the United States wanted to slow us down, I think they couldn't wish for almost anything better than these silly moratoriums on building our data centers. I've been alarmed at the amount of lobbying that a handful of businesses have been doing, saying that open models are dangerous. I think that's false. I think the concept that distillation is a major factor has been overstated. Sometimes the most powerful forces in the world go unseen. A critical layer of intelligence driving possibilities. This is Micron Memory and Storage. Intelligence that uncovers new cures, models changing climates, and build smarter, safer cities. Every breakthrough in AI is driven by data. And data lives in Micron Memory and Storage. Hello everyone. I'm James H. Homeman, deputy opinion editor at the Washington Post. Welcome to the Post's Building America podcast about how America builds, manufactures, and innovates better for the future. Today, we're speaking with Andrew about how investing in AI infrastructure will shape America's long-term competitiveness. Andrew is a true pioneer in AI. He led the Google Brain Team, which launched much of the tech giant's AI work. He directed the Stanford AI lab and has authored over 200 papers on AI and related issues. He has since co-founded deeplearning.ai, which provides technical training. And he's managing general partner of a venture fund. He is also on the board of directors of Amazon whose executive chairman Jeff Bezos owns the Post. Andrew, welcome.

谢谢你，詹姆斯。能来《邮报》总是很高兴。

Thank you, James. It's always good to be with the Post.

对于那些不了解你精彩故事的人，带我们回到起点吧。你进行了神经网络研究，促成了谷歌大脑的成立。2010 年，你向谷歌创始人拉里·佩奇推销的是什么？

Well, for those who don't know your incredible story, take us back to the start. You conducted research on neural networks that led to the formation of Google Brain. What was it that you pitched to Google founder Larry Page back in 2010?

当时我有一个有争议的想法：如果我们构建非常大的神经网络，并给它们喂大量数据，它们会学到令人难以置信的东西。所以当我启动谷歌大脑团队时，我给团队的首要方向是：构建非常非常大的 AI 系统，并喂入大量数据。凭借这个过于简化的方向，我们最终取得了一些惊人的成就。我认为谷歌大脑团队的文化后来促成了 Transformer 神经网络的发明，这确实是谷歌大脑团队公开发布的一项关键技术，OpenAI 采用了它，许多其他机构也采用了，从而引发了现代 AI 革命。这确实是由我在斯坦福早期数据中看到的一个想法驱动的：你构建的 AI 系统越大，它们就越聪明。而这个想法至今仍在推动我们前进。

At that time I had this controversial idea that if we build really big neural networks and feed a lot of data to them, they will learn incredible things. So when I started the Google Brain team, the number one direction I gave the team was let's build really, really big AI systems and feed a lot of data. And with that oversimplified direction, we wound up having some incredible accomplishments. I think that culture of the Google Brain team led it to later invent the Transformer neural network, which is really the one key technology that the Google Brain team openly published and that OpenAI adopted, many others adopted, and sparked off the modern AI revolution. It was really driven by this one idea that I saw in my very early data at Stanford that the bigger you build AI systems, the smarter they get. And that idea is still carrying us forward today.

我的意思是，快进到 2026 年，你对达到这个地步的速度感到惊讶吗，还是你有点预料到我们会走到这一步？显然，很多都是理论性的，而现在它已成为世界上最重要的事情。

I mean, fast forwarding to 2026, are you surprised by the speed in which we've gotten to this place, or did you kind of expect that this would be where we'd end up? Obviously, a lot of this was theoretical and now it's the biggest thing in the world by far.

即使在那个时候，我对它的发展方向非常乐观，而且我仍然对未来能达到的高度非常乐观。但至于我们目前进展到哪一步，以及还能走多远，有些日子我很兴奋，有些日子我则想知道为什么我们还没有走得更远。我认为我们看到的是，尽管 AI 技术令人惊叹，而且确实令人惊叹，但企业还没有感觉到它对他们有帮助。挑战之一是，没有公司仅仅通过购买 ChatGPT 或 Microsoft Copilot 许可证就能获得竞争优势。问题在于找到正确的用例。这是一个人员变革管理问题，让合适的人理解技术，将其与实际用例结合起来，当然还需要你提到的基础设施的支持。

Even back then, I was very optimistic about where this would go and I remain very optimistic about where we can get in the future. But exactly how far we are and how far we could go, some days I'm excited, some days I wonder why we aren't even further along. And I think what we're seeing is that as amazing as AI technology is, and it is amazing, businesses do not feel like it's helping them yet. And one of the challenges is no company ever gains competitive advantage just by buying a ChatGPT or Microsoft Copilot license. The problem is finding the right use cases. And that's a people change management issue, having the right people understand the technology, marry it to the actual use cases, and of course supported also by the infrastructure that you talked about.

是的，我看到你刚刚推出了 OpenWorker，一个开源的桌面智能体，它产生完成的工作，而不仅仅是对话。这是你所说的答案的一部分吗？

Yeah, I saw that you've just unveiled OpenWorker, an open-source desktop agent that produces finished work rather than just a conversation. Is that part of the answer to what you're talking about?

是的。所以我和我的朋友拉哈布·普拉萨德一直在开发一个开放版本的 AI，你不仅可以用它聊天，它实际上还能为你工作。所以很多人仍然会使用 ChatGPT、Claude 或 Gemini 之类的工具来聊天并获得答案。但今天 AI 可以更进一步。它不只是和我聊天，而是可以生成一份完成润色过的文档，而不是只给我可以复制粘贴的东西；或者它可以实际为你发送电子邮件，或者在 Slack 或消息系统中为你发送一条短消息，或者为你构建一个仪表板。所以 OpenWorker 是一个完全开源、免费的软件，让人们可以在自己的电脑上做到这一点。我实际上对 Claude Code、ChatGPT Work 和 Gemini Antigravity 等工具感到兴奋，它们都是好工具。而 OpenWorker 是一个任何人都可以使用的免费开源版本。

Yeah. So my friend Rahab Prasad and I have been working on an open version of AI that you don't just chat with, but that actually does work for you. So a lot of people still will use ChatGPT or Claude or Gemini or tools like that to chat and get back an answer. But today AI could go further. Instead of just chatting with me, it can produce a finished polished document rather than just give me things to copy paste, or it can actually send an email for you, or send a short message in Slack or messaging system for you, or build a dashboard for you. And so OpenWorker is a fully open source, free piece of software to let people do that on their computer. And I'm actually excited about tools like Claude Code and ChatGPT Work and Gemini Antigravity, which are all good tools. And OpenWorker is a free open source version that anyone can use.

现在在华盛顿特区肯定有很多关于开放与封闭、开放权重模型与封闭模型的讨论。你在开放方面做了很多工作。你看到很多前沿模型在开发这些封闭模型，人们必须为 token 付费。这已经变得有点……你看到公司开始抵制这种情况，因为成本和账单开始上升。你怎么看待开放与封闭的辩论，如果可以这样称呼的话？

There's a lot of talk right now in DC certainly about open versus closed, open-weight models, closed. And you've done a lot on the open side. You see a lot of the frontier models developing these closed models where people have to pay for tokens. That's become sort of... You're seeing companies push back on that as the costs and the bills start to run up. How do you think about the open versus closed debate, if you can call it that?

是的。所以，为了维持美国的竞争优势，我们必须做的最重要的事情之一就是支持和维持开放模型。为了澄清术语，顺便说一句，我认为我是唯一一个山姆和达里奥都为之工作过的人。所以实际上，你知道，我支持 OpenAI、Anthropic，我真的希望他们做得好，并拥有出色的 IPO 等等。所以我真的希望这些伟大的美国企业能够成功。与此同时，这些企业的成功不能以切断其他人对开放模型的访问为代价。所以，OpenAI、Anthropic、谷歌在训练封闭专有模型方面做得很好，他们以相当高的价格出售访问权。开放模型是人们在互联网上发布的模型，任何人都可以免费使用。这些模型性能极高，非常智能，而且通常便宜得多。所以，我一直对少数企业进行的游说感到震惊，他们说开放模型是危险的。我认为这是错误的。或者说开放模型不如封闭模型，他们就是不明白。开放模型给企业更多选择。这意味着你不会被锁定在一个封闭的专有提供商上。事实上，我经常给商业领袖的一条建议是，我无法预测一年甚至六个月后什么会是顶级模型。所以在我们建立业务时，最重要的事情之一是保留可选性。这样我就不会被锁定。

Yeah. So, to sustain competitive advantage in America, one of the most important things we have to do is support and sustain open models. So just to clarify terminology, by the way, I think I'm the only person that both Sam and Dario have worked for. And so actually, you know, I support OpenAI, Anthropic, I really hope they do well and have fantastic IPOs and so on. So I really hope these great American businesses will succeed. At the same time, the success of these businesses cannot be at the cost of shutting down everyone else's access to open models. So, OpenAI, Anthropic, Google have done a great job training closed proprietary models that they sell access to for a decently high price. Open models are models that people have published on the internet, free for anyone to use. And those are extremely performant, very intelligent, and often much cheaper. And so, I've been alarmed at the amount of lobbying that a handful of businesses have been doing, saying that open models are dangerous. I think that's false. Or saying that open models are somehow not as good as the closed models, which they just don't get. Open models give businesses more choice. It means you're not locked in to one closed proprietary provider. In fact, one piece of advice I often give to business leaders is I can't forecast in a year or even six months what is going to be the top model. So as we build our businesses, one of the most important things to do is to preserve optionality. So I'm not locked in.

我用 OpenAI、Claude、Gemini，还有很多不同的模型，但我不让自己被任何一个锁定，因为一年后我们想用最好的那个，无论是更便宜的开源模型还是闭源专有模型。所以保持可选择性非常重要。而且全球供应链都在用开源模型，所以确保美国继续能接触到这些，对我们来说也很重要。

I use OpenAI, I use Claude, I use Gemini, and a lot of different models, but I don't let myself be locked into any one of them because a year from now we want to use the best one, whether it's an open model that's cheaper or a closed proprietary model. So preserving optionality is very important. And the world's supply chain is using open models, so making sure that America continues to have access to that will be important for us as well.

是的，我是说，你提到的那个游说活动，他们基本上想把开源模型和中国模型划等号，然后提出国家安全论调，但这不一定是这样，也不一定一直这样。我还好奇中国体制内开源模型的张力，以及中国政府是否会长期允许这种做法。

Yeah, I mean, part of the lobbying campaign that you're referring to, they're trying to basically portray open models as one and the same as Chinese models, and then they're making this national security argument, but it doesn't necessarily need to be or stay that way. And I also wonder about the tension in the Chinese system of having open models and whether the government there will continue to allow that over the long term.

几年前 ChatGPT 刚发布时，美国在生成式 AI 技术上明显领先中国。从那以后，中国打得一手好牌。中国做得特别好的一点是拥抱开源模型，因为事实证明，当你免费发布模型供任何人使用时，它帮助了全世界。是的，但它对你的帮助比对全世界的帮助更大。所以开放性、自由交流、知识的快速传播意味着中国已经迅速加速，正在接近——也许还没完全持平，但正在接近领先的中国模型水平。相比之下，因为美国很多工作都投入到了闭源专有模型上，所以如果你有问题，很难知道该给谁打电话去了解怎么做这个建模。

When ChatGPT was first released a few years ago, America was decisively ahead of China in generative AI technology. Since then, China has played its hand really well. One of the things that China did really well was embrace open models, because it turns out that when you release models freely for anyone to use, it helps the whole world. Yes, but it helps you even more than it helps the whole world. So the openness, freedom of communication, fast diffusion of knowledge has meant China has rapidly accelerated and is approaching, maybe not quite at par, but approaching par with the leading Chinese models. In contrast, because a lot of American work has gone into closed proprietary models, it's just much harder to know, if you have a question, who should I call up to understand how to do this modeling thing.

嗯，也许一家公司以很高的价格挖走另一家公司的工程师，然后知识有一点扩散。但我认为这种反对开源模型的游说阻碍了美国 AI 的发展。我们确实有一些很棒的美国开源模型。我认为英伟达的 Nemotron 模型很强。Thinking Machines 的 Inkling Lab 看起来非常出色。我希望看到更多这样的。我认为专有模型和开源模型都有成功的空间。

Um, and maybe one company recruits another engineer for a very high price tag, and then there's a little bit of diffusion of knowledge. But I think this lobbying against open models has hampered American AI development. We do have some great open American models. I think Nvidia's Nemotron models are strong. The Thinking Machines Inkling Lab feels like a very strong showing. I'd love to see more of this. I think there's room for proprietary models and open models to succeed.

重要的是保持公平的竞争环境，这样任何模型——你知道的——这样人们就有选择去用最好的模型。

And the important thing is to maintain a level playing field so that any model, you know, so that people have choice to use the best model.

你认为中国开源模型取得的进展，有多少是蒸馏的结果，又有多少是它们自身进步和开源模型民主化特性的结果？

How much do you think that the gains that the open Chinese models have made are the result of distillation versus their own progression and the kind of the democratized nature of an open model?

很明显，美国有伟大的技术创新，中国也有伟大的技术创新。事实上，许多美国前沿实验室都在积极阅读中国实验室发表的大量开源研究。我的意思是，不读才是傻。所以美国实验室肯定从大量中国研究实验室的创新中受益。我认为蒸馏是主要因素的说法被夸大了，被大大夸大了。有一个问题：世界应该认为什么是公平的。全世界的 AI 实验室都从开放互联网上获取数据，把互联网的知识蒸馏到他们的 AI 模型里。他们反过来又说“我把互联网蒸馏进了我的模型，如果任何人从今往后蒸馏我的模型，那就不公平”，这公平吗？我认为这是一个开放的问题，社会应该认为什么公平或不公平。有些说法，比如 DeepSeek 最近发布了，我想 V3 也快出来了。考虑到 Fable 模型只开放了很短一段时间，说 DeepSeek 主要是通过蒸馏 Fable 训练的，我觉得这很难让人相信。因为不可能有那么多 Fable 数据，而且 DeepSeek 怎么可能在这么短的时间内训练出来？

So it is clear that we have great technical innovations in America and China has great technical innovations. In fact, many of the US frontier labs are actively reading a lot of the open research that the Chinese labs published. I mean, you have to be dumb not to. So American labs are definitely benefiting from tons of Chinese research lab innovations. I think the concept that distillation is a major factor has been overstated, vastly overstated. There's a question of what the world should consider fair. AI labs all around the world took data off the open internet and used it to distill knowledge from the internet into their AI models. Is it fair for them to turn around and say, "I've distilled the internet into my model. If anyone distills my model from here on out, that's not fair." I think that's an open question of what society should consider fair or not. Some of the claims that, you know, DeepSeek was released recently, I guess with V3 coming soon. Given that the Fable model was available only for a short period of time, the idea that DeepSeek was trained primarily by distilling Fable, I just find that very hard to believe. Given there just couldn't have been that much Fable data, and how could DeepSeek have been trained in such a short time?

这真的是一个很好的视角，因为我觉得这绝对不是华盛顿讨论的一部分。但这是个很好的观点。

That's really a really good perspective, because I think that's definitely not part of the conversation in DC. But that's a great point.

是的，有些势力急于把能扔的东西都往墙上扔，来攻击开源模型，包括中国的，但同样，当你拿一个开源模型，不管是中国实验室还是美国实验室发布的，在美国基础设施上运行，你知道，基本上美国基础设施就能控制那个模型。比如，当我在美国发布代码，被其他国家使用，那么那个国家就从那时起控制了它，或者不是。所以我认为那些为了减缓美国采用而散布的恐惧、不确定性和怀疑（FUD）是不幸的。全世界都会用开源模型。很多国家都会用开源模型。所以美国在开源模型上的投资不足意味着很多国家，比如在非洲，DeepSeek 的采用率飙升。我们在美国不太看到这种情况，但在那些对价格更敏感的地方，中国模型确实获得了巨大的市场份额。这很遗憾，因为我希望美国的开源模型或美国模型在全球范围内更有竞争力。这是一个巨大的竞争劣势。如果一个开源模型能让你以，我不知道，五分之一、三分之一或十分之一的成本获得智能，那么对于每个想构建 AI 应用的人来说，如果你的智能供应成本贵三倍，那是一个非常根本的商业劣势。

Yeah, there are certain parties that are eager to throw whatever they can at the wall to attack open source models, including Chinese, but also when you take an open model, whether it's released from a Chinese lab or American lab, and run it on American infrastructure, you know, it basically becomes American infrastructure can control that model. For example, when I in the US publish code online that is used by some other country, well, that other country controls it from then on, or no. And so I think the amount of FUD to slow down American adoption is unfortunate. The whole world's going to use open models. A lot of the world is going to use open models. And so American underinvestment in open models means that there are many nations, for example, in Africa, DeepSeek adoption is through the roof. We don't see this that much in the US, but in places where they're a little bit more price sensitive, Chinese models have really gained tremendous market share. And this is regrettable, because I would like American open models or American models to be more competitive all around the world. And it is a massive competitive disadvantage. If an open model allows you to get intelligence at, I don't know, one-fifth or one-third or one-tenth the cost, then for everyone wanting to build AI applications, if your supply of intelligence costs three times more, that's a very fundamental business disadvantage.

我有点好奇，你怎么维持开源模型的商业模式？如果你按 token 收费，如果你是 Anthropic 或 OpenAI，那显然有帮助，有一些利润，但最终也是希望如此，而且你也在帮助进行这种大规模建设。当你谈到开源 AI 模型时，资本支出的钱从哪里来？

I guess I'm sort of curious, how do you sustain a business model for an open model? If you're charging for tokens, if you're Anthropic or OpenAI, that's helping, obviously there's some profit, but it's also eventually hopefully, but you're also helping do this big buildout. Where does the money come from for the capital expenditures when you're talking about open AI models?

你知道，Bill Gurley 写了一篇文章，我想是最近发表在《华盛顿邮报》上，谈到开源是一个成熟的商业模式。Red Hat 通过开源 Linux 操作系统建立了伟大的业务，许多企业都是通过开源建立起来的。如何用开源模型做到这一点，我认为细节还在摸索中。但有一些上市的中国公司，至少从股市来看，有聪明的投资者，他们的开源战略似乎做得很好。所以训练开源模型的资本支出确实比编写传统软件的资本支出高，但我认为他们的商业模式还有待完善，而且只要开源模型能给你带来根本性的成本优势，这就是值得关注的。

You know, Bill Gurley wrote an article, I think it was in the Washington Post recently, talking about how open source is a well-established business model. Red Hat built a great business by open sourcing the Linux operating system, and many businesses have been built with open source. The details of how to do this with open models, I think, are still being worked out. But there are publicly traded Chinese companies that seem to be doing just fine, at least in the stock market, with smart investors, with their open source strategy. So that's true that the capex of training open models is higher than the capex of writing traditional software, but I feel like their business models are to be worked out, and to the extent that open models give you a fundamental cost advantage, that's something to pay attention to.

完全同意。

Totally.

我们这一系列访谈的重点之一是，AI 不仅是数字故事，也是物理故事。而且你知道，正在进行大规模建设。显然，这确实推动了经济增长。你认为你所看到和参与的所有突破，在多大程度上仍然依赖于实体世界？我的意思是，我们谈论的是训练、运行、使用 AI 模型、数据中心、电网、半导体、冷却设备和其他基础设施。你如何看待物理空间和数字空间之间的互动？

One of the focuses of this series of interviews that we're doing is that AI is not just a digital story but also a physical one. And that you know there is this huge buildout. Obviously it's really kept the economy growing. How much do you think all the breakthroughs that you're seeing and working on depend still on the built world? I mean, we're talking about training, running, using AI models, data centers, grids, semiconductors, cooling equipment, other infrastructure. How do you see the physical and the digital space in conversation with one another?

世界正在投资建设更多更大的数据中心，我认为这是好事。我觉得故事中有一部分被低估了，那就是在数据中心建设、购买 GPU、基础设施层的资本支出上投入了这么多。我认为我们确实需要更多的 AI 推理能力，也就是在模型训练后提供智能的能力。我们似乎永远都不够用。但故事中更有价值且被大大低估的部分，是建立在这个基础设施之上的所有应用的价值。例如，当互联网出现时，思科做得很好。我的意思是，思科不错。我在思科有很多朋友，为他们高兴。但建立在互联网之上的应用才变得更有价值。而且必须如此，因为需要开发应用的人产生足够的收入来支付基础设施提供商的费用。所以，我在 AI Aspire 与我的商业伙伴 Kirsty Tan 花了很多时间，帮助大型企业思考哪些应用不仅仅是，你知道，让我们和聊天机器人聊天、复制粘贴、修正我的电子邮件语法，而是思考哪些自动化场景能让你改变商业模式，不仅节省成本，更令人兴奋的是获得增长。人们仍然低估了这一点，它将比我们同样关注的基础设施故事更有价值。

The fact that the world is investing in building more bigger data centers, I think that's a good thing. I feel like there's one part of the story that's underappreciated, which is there's so much investment in data center buildout, buying GPUs, the capex for the infrastructure layer. And I think we definitely need more AI inference capacity, which is delivering the intelligence after the model trains. So we definitely can't seem to get enough of that. But the part of the story that is going to be even more valuable and that is vastly underappreciated is the value of all the applications that'll be built on top of this infrastructure. So for example, when the internet came up, you know, Cisco did well. I mean, good for Cisco. Happy a lot of friends at Cisco. But it was the applications built on top of the internet that became even more valuable. And it had to be that way because you need the people building applications to generate enough revenue to pay the infrastructure providers. And so, a lot of the time that I spend at AI Aspire with my business partner Kirsty Tan is helping large businesses think through what are those applications that are not just, you know, let's just chat with a chatbot and copy paste, fix my email grammar, but to think through what are the automation scenarios that let you change the business model and not just get cost savings, but more excitingly get growth. And people still underappreciate that that's going to be even more valuable than this infrastructure story that we're also focused on.

我想你知道，当你和一些投资者交谈时，你显然会面临这样的问题：如何避免投资一家可能正在开发产品或应用的公司，而这些产品或应用基本上会被 Anthropic 或 OpenAI 的下一代产品一击即溃，因为它们能实现相同的功能。你在做投资决策时是怎么考虑这个问题的？

I think you know when you talk to some investors you're obviously like in this how do you avoid investing in a company that might be making a product or an application that could just be basically oneshotted by the next Anthropic rollout or the next OpenAI rollout you know where they can perform the same function. How do you think about that when you're making investment decisions?

是的。所以，我认为对于小型初创公司和大型现有企业来说，我们有时会思考是什么护城河让业务具有防御性并长期持续。对于大型企业，我发现他们的许多现有资产是经过几十年甚至更长时间积累起来的，实际上很难被取代。但需要做的大量工作是，要有正确的自上而下的领导力，来思考鉴于我们建立的所有有价值的东西——所有客户关系、所有技术、所有数据、所有供应链，所有这些——我们在哪里插入 AI 来提高效率，或者更令人兴奋的是推动业务增长。所以，也许举个例子，本周早些时候我访问了摩根士丹利，在那里我与 CEO Ted Pick、Dan Sinowitz 和 Andy Saperstein 进行了交谈。我发现那是一家伟大的金融机构，世界上最重要的机构之一，它从高层就定下了使用 AI 的基调，提供培训给全球所有员工，然后执行他们称之为“挑战实验室”的一系列练习，在那里他们召集业务领导和技术领导合作寻找用例。我发现业务和技术领导之间的这些合作往往能识别出真正有价值的用例，而不仅仅是渐进式的效率提升。

Yeah. So, I think for both small startups as well as large incumbent companies, we sometimes think about what is the moat that makes the business defensible and sustained for a long time. For large businesses, which I find that a lot of their existing assets have been built up over decades or sometimes even more than decades and are actually really difficult to displace. But a lot of work that needs to be done is to have the right top-down leadership to think through given all the valuable things we built, all the customer relationships, all the technology, all the data, all the supply chain, all of these things, where do we insert AI to get efficiencies or maybe even more excitingly drive business growth. So maybe one example earlier this week I was visiting Morgan Stanley where I was chatting with the CEO Ted Pick, and then Dan Sinowitz and Andy Saperstein. And I find that that's one example of a great financial institution, one of the most important institutions in the world, that has set a tone from the top of using AI, offering to train everyone in the company globally, and then executing a set of exercises that they call challenge labs where they assemble business leaders and technology leaders to collaborate to find use cases. And I find that those partnerships between business and tech leaders can often identify really valuable use cases that are not just incremental efficiency shift.

在那里你思考企业如何创造价值，然后你不只是让它稍微便宜一点，而是可能思考如何将这项服务做 100 倍并推动增长，或者将服务速度提高 10 倍，这样你就能更快地回复客户并提供更有价值的产品。

Where you think about how does the business create value and then you don't just make it a little bit cheaper to do but maybe think about how to do this service 100 times more and drive growth or do the service 10 times faster so you can get back to the customer faster and offer more valuable products.

我发现这些事情非常令人兴奋，而且将比人们谈论的业务效率成本节省更有价值。

I find those things really exciting and will be even more valuable than the business efficiency cost savings that people also talk about.

这很令人振奋。我认为这也有助于解释一些估值。当你想到可扩展性时，我的意思是，我们在基础设施方面继续看到这些真正惊人的资本支出。我们刚刚听到谷歌基本上说他们今年将在资本支出上花费几乎美国军事预算的四分之一。你知道，那是整个海军陆战队预算的四倍。我的意思是，这真是令人难以置信的数额。你对那些担心存在某种泡沫或我们投资过快过多的人有什么要说的？

That's heartening. And I think it helps explain some of these valuations too. When you think about the scalability, I mean we are continuing on the infrastructure side to just see this really amazing capital expenditures. We just heard Google basically say they're going to spend almost a quarter of the US military budget on capex this year. You know, that's four times the budget of the whole Marine Corps. I mean, it's just incredible sums. What do you say to people who worry that there's some kind of bubble or that we're going to have too much investment too fast?

所以，我不是在给任何人投资建议。不，我知道。而且我认为我知道，我觉得有一件事我很确定，那就是我们需要更多的推理能力。也就是说，AI 为我们生成词元、生成输出。

So, I'm not giving anyone investment advice. No, I know. And I think I know, I feel like one thing I'm confident about is we need a lot more inference capacity. Meaning AI to generate tokens, generate outputs for us.

以软件工程为例，这是 AI 加速最明显的领域。它被报道最多，因为 AI 正在编写大量软件。

And take software engineering, which is the sector that AI has accelerated the most. It's covered the most because AI is writing tons of software.

有两个有趣的观察。渗透率仍然很低。很多软件工程师还没有完全拥抱 AI 工具，但那些已经使用的人速度更快、生产力更高，而且坦白说，我认为他们也更有乐趣。而且我们已经无法获得足够的推理能力。所以，随着渗透在软件领域加深，我们只是需要更多的 AI。这并不意味着人们不会在资本支出上亏钱，但我认为无论我们建设什么都会被使用。另一个好消息是，关于 AI 最激烈的叙事之一是会出现某种工作末日或 AI 工作末日，有些人说我们 50% 的工作可能会被取代，人们可能会上街暴动。那不会发生。

Two interesting observations. Penetration is still low. A lot of software engineers have not yet fully embraced AI tools, but the ones that have are just so much faster, more productive, and frankly, I think have more fun as well. And already we just can't get enough inference capacity. So, as penetration goes deeper in software, we just need more AI. Doesn't mean people won't lose money building capex, but I think it will get used whatever we can build out. The other piece of good news for that is one of the fiercest narratives for AI is there'll be some sort of job apocalypse or AI job apocalypse where some people have said 50% of the people we other job may be rioting in the streets. That's not going to happen.

我认为相反的情况在软件工程中已经非常明显，AI 正在自动化大量工作。

I think the opposite is already very visible in software engineering, where AI is automating so much work.

坦白说，我们无法获得足够多的熟练 AI 工程师。我的意思是 AI 服务器工程师，这使软件工程师更有价值。因此，软件工程职位发布在增加，行业健康且增长。

Frankly we can't get enough skilled AI engineers. I mean AI server engineers, that's made software engineers even more valuable. And so software engineering job postings are up, and the industry is healthy and growing.

所以我认为，就 AI 先渗透软件领域、随后再渗透其他领域（比如营销、招聘等）而言，平均来看会有少数例外，但对大多数职业类别来说，因为 AI 无法包办一切，人们需要去引导和补充 AI。挑战在于人们需要掌握新技能。我们不能再用两年前的方式写软件，却还指望能保住工作。但通过转变技能组合、学习新技能，我们却找不到足够的软件工程师和 AI 工程师。所以随着 AI 开始进入更多领域，我认为软件将证明是我们在其他领域也会看到的趋势的先兆或先行者——你需要推动技能转变，但人们将能够做更多事，希望也能获得更高报酬。但这实际上才是真正的挑战——不是应对就业末日，对吧？挑战在于如何帮助每个人获得他们需要的新技能。

So I think to the extent that AI is infiltrating software first and then other sectors later, be it marketing, recruiting, and so on, I think on average there'll be small exceptions, but on average for most job categories, because AI can't do everything, people are needed to steer and complement AI. The challenge will be that people will need new skills. We can't write software the way we did two years ago and expect to have a job. But by shifting the skill mix and learning new skills, we can't find enough software engineers and AI engineers. So as AI starts to enter more fields, I think software will prove to be a harbinger or a forerunner of a trend we'll see in other sectors as well, where you do need to grow the skill shift in skills, but people will be able to do more, hopefully get paid more. But that'll be actually the challenge—not dealing with the job apocalypse, right? The challenge is how to help everyone gain the new skills they will need.

是的。我的意思是，当你谈到发展推理能力时，实际上这意味着什么？是建更多的数据中心吗？我们如何从物理上实现这一点来获得更多的推理能力？

Yeah. I mean, when you talk about developing inference capacity, in practice, what does that mean? Is that building more data centers? How do we physically do that to get more inference capacity?

一种方式是建更多的数据中心。我认为坦率地说，建更多的数据中心——它们会被用上的。具体的盈利或亏损数字，我不知道。

One way would be to build more data centers. I think frankly, build more data centers—they will get used. The exact amounts of profit or loss, I don't know.

对。

Right.

这更难确定，但我认为对推理数据中心的需求非常非常高。然后，是否会有更有创意的方式、新的技术突破来进一步降低成本？我实际上相当确信会有，仅从我看到研究人员正在做的事情来看。但我认为有了 AI，计算机和数据中心变得比以前更有价值，所以建很多数据中心是完全合理的。具体需要多少，我们需要校准，但显然应该建更多。然后，如果我再想想基础性部分，一个大的建设方向——我刚刚和我在 AI 中心的合作者聊过这个——就是数据架构。事实证明 AI 是由数据驱动的，我看到很多大型企业正在重新思考如何组织数据，使数据不仅供人类使用，也供智能体使用。所以实际上，我和我的团队花了很多时间思考未来数据将由人类使用，绝对如此，但也会被我们的 AI 智能体使用。释放企业大量价值的关键方式之一就是让你的数据架构为智能体做好准备。但智能体使用数据的方式与人类非常不同。你访问数据的频率更高。模式有时更混乱，但你需要自动化接口。例如，如果我的智能体想要访问数据，我不能让它每 60 秒就打断我，让我输入密码，对吧？所以就是诸如此类的小管道问题。但我认为这是另一个能释放企业大量价值的东西。

Harder to pin down, but I think there is very, very high demand for inference data centers. And then, will there be more creative ways, new tech breakthroughs, to bring down the cost even further? I'm actually pretty confident there will, just from the things I see researchers working on. But I think with AI, computers and data centers have become even more valuable than before, so building a lot of them just makes sense. Exactly how much we need to calibrate, but there should clearly be a lot more of them. And then, if I think about foundational pieces as well, one of the big buildouts—and I was just chatting with my collaborators at the Center for AI about this—is data architecture. So it turns out AI is fueled by data, and I see a lot of large businesses rethinking how we organize our data so that data is not ready only for humans to use but for agents to use. So actually, in my team and I, we actually spend a lot of time thinking through in the future data will be used by humans, absolutely, but it'll also be used by our AI agents. One of the key ways to unlock a lot of value in businesses is to make your data fabric agent-ready. But agents use data very differently than humans. You access it a lot more. The patterns are sometimes more chaotic, but you need automated interfaces. For example, if my agent wants to access the data, I can't have it stop me every 60 seconds to have me type in the password, right? So just little plumbing things like that. But I think that's another thing that unlocks a lot of value in businesses.

这太有趣了。关于数据中心，你怎么看待我们开始看到的抵制、反对和恐惧？我们如何克服这些？像纽约暂停一年这样的做法，对建设我们所需的推理能力、实现你所说的所有这些伟大成就，构成了多大的威胁？

That's so interesting. On data centers, what do you make of the backlash that we're starting to see, the pushback, the fear? How do we get past that? How much of a threat is something like the moratorium for a year in New York to the buildout, to getting the inference capacity that we need to be able to achieve all the great stuff that you're talking about?

你知道，如果美国的对手想拖慢我们的步伐，我认为他们几乎不可能指望比这些愚蠢的暂停建设数据中心更好的事情了。我不想忽视人们的担忧。我觉得是的，数据中心在某些地方确实有点碍眼。你知道，也许我们应该想办法让它们更美观。但我认为很多其他事情被过度炒作。事实证明，我们能为环境做的最好的事情之一就是把算力集中在数据中心。

You know, if an adversary of the United States wanted to slow us down, I think they couldn't wish for almost anything better than these silly moratoriums on building our data centers. I don't want to dismiss the concerns that people have. I feel like yes, data centers are kind of an eyesore in some places. You know, maybe we should figure out a way to make them prettier. I think a lot of other things have been overhyped though. It turns out that one of the best things we could do for the environment is to concentrate our compute in the data center.

所以你知道，我办公室里有一架服务器。坦率地说，我的那架服务器在能源使用、水消耗或任何方面都比那些被设计成超高效的数据中心效率低得多。所以，我们能为环境做的最好的事情之一就是把所有本地设备都搬进数据中心，因为集中——而且我认为因为我们把大量算力集中在一起，数据中心确实消耗能源和水，但这比分散部署的替代方案要好。我认为真正的问题是我们在使用如此多的算力，这开始对环境产生影响。但如果我们打算使用这么多算力，显然最好放在数据中心。我怀疑对数据中心的抵制更多是对 AI 的抵制或不适，其程度不亚于甚至超过对数据中心本身的抵制。

So you know, I have a rack of servers in my office. Frankly, my rack of servers is so much less efficient from energy use or water consumption or any point of view than data centers which have been engineered to be hyperefficient. So one of the best things we can do for the environment is move all the on-premises stuff into data centers where the concentration—and I think because we put a lot of compute together, a data center does use energy and water, but it's better than the alternative of having it distributed. I think the real question is we're using so much compute, so that starts to have an environmental impact. But if we're going to use this much compute, it's clearly better to put it in a data center. I suspect that backlash against data centers is more backlash or discomfort against AI as much as or even more than data centers per se.

而且我担心 AI 会让很多人的生活变得更好。我认为它将帮助人们学习、改善医疗保健。它将推动商业成果。我觉得我个人生活更有趣，因为 AI 为我做了一些不那么有趣的事情。所以我把 AI 看作一股非常积极的力量，但不知何故，那些奇怪的反对 AI 的游说活动已经过度了，如果我们让这种情况继续下去，这真的会损害美国。

And I am worried that AI is going to make so many people's lives better. I think it will help people learn, improve health care. It will drive business outcomes. I feel like my life personally is more fun that AI does some of the less fun things for me. So I see AI as a very positive force, but somehow the amount of weird anti-AI lobbying has been excessive, and this will really damage America if we let this continue.

说到人才和技能，你有 deeplearning.ai，这是一家教育科技公司，帮助人们在机器学习和人工智能领域建立职业生涯和技能。你认为这对培养年轻人填补你所说的这些需要填补的职位最有帮助的地方在哪里？

Speaking of talent and skills, you have deeplearning.ai, which is an education technology company that's helping people build careers and skills in machine learning and artificial intelligence. Where do you see that being most helpful for preparing young people to fill all these roles that you're talking about will need filling?

许多国家面临的挑战是传统教育体系适应新技术缓慢。现实是，我们不想为 2022 年的工作培养软件工程师，坦率地说，我们甚至不想为 2026 年的工作培养他们。我们应该为 2028 年及以后的工作培养他们。但如何让教育体系适应软件和 AI 工程的未来走向，对我来说越来越清晰。所以我认为我能够对人们需要哪些新技能有自己的看法。而且我认为我实际上也在努力为所有其他职位角色弄清楚这一点。你知道，我们不希望每个人都成为软件工程师。那会很有趣，但有点奇怪。但对于记者、营销人员、招聘人员、人力资源专业人士、运营专家来说，他们的工作将会改变。而且我认为他们的工作不会消失。我们仍然绝对需要很多人来做这些工作。但成为下一代营销人员、下一代招聘人员、下一代记者，他们需要哪些新技能？我对此有看法。但我认为把这些想法具体化，给人们一条清晰的学习这些技能的路径。

The challenge that many countries have is that the traditional educational system is slow to adapt to new technologies. And the reality is we don't want to train up software engineers for the jobs of 2022, or frankly we don't even want to train them up for the jobs of 2026. We should be training them for the jobs of 2028 and beyond. But how do we get the educational system to adapt where the future of software and AI engineering is going is becoming clearer to me. So I think I'm able to have a view on what are the new skills people need. And I think I'm actually working hard to figure out for all of the other job roles as well. You know, we don't want everyone to be a software engineer. It'd be fun but kind of weird. But for the journalists and the marketers, the recruiters, the HR professionals, the operational specialists, their jobs will change. And I don't think their jobs will go away. We still absolutely need lots of people to do this. But what are the new skills that they will need to be the next-gen marketer, the next-gen recruiter, the next-gen journalist? I have a view on that. But I think firming that up to give people a clear path to learn those skills.

是的。它们值得一看，我相信它们会越来越好。最后，我们在“建设美国”项目中会问每位受访者一个问题：你最钦佩哪位美国建设者或创新者，无论健在还是已故，为什么？

Yeah. And they're worth checking out and I'm sure they'll just keep getting better. To close out, something we're asking everyone we're talking to as part of this Building America project: who is an American builder, living or dead, or innovator, who you most admire and why?

你知道吗，你说话时我脑海里浮现出一个名字，但听起来可能有点矛盾，但说实话，杰夫·贝索斯是我最钦佩的领导者之一。我关注他多年，一直远远地关注，你知道，在我有幸与亚马逊打交道之前，我听过他很多演讲。从他身上学到了很多。然后最近还有安迪·贾西。但我认为杰夫·贝索斯是美国最杰出的风险投资人之一。

You know, I'll tell you the name that popped in my head as you were talking, but it'll sound so conflicted, but honestly, Jeff Bezos is one of the leaders I most admire. Followed him for years, and just from afar, you know, before I had the privilege to pass through Amazon, I was trying to listen to a lot of talks. Learned so much from him. And then also more recently, Andy Jassy. But I think that Jeff Bezos has been one of the most amazing VCs America has seen.

是的。

Yeah.

而且显然他在亚马逊所建立的事业惠及了很多人。

And then clearly what he built at Amazon is something that is benefiting a lot of people.

客观来说，这一点无可争辩，仅从影响力来看。

Objectively, you can't argue with that, just in terms of impact.

好的，安德鲁，非常感谢你今天来做客。

Well, Andrew, thank you so much for coming on today.

谢谢。能来这里真的很有趣，而且我认为这对美国来说是一个重要的话题。确保美国的竞争力，我希望我们伟大的美国企业能够获胜，但让美国整体，而不仅仅是 AI 领域，获胜的最佳方式之一就是保留开放模型，这样所有美国公司都能发展得好。

Thank you. This is really fun to be here, and I think this is an important topic for America. So ensuring American competitiveness, where I want our great American businesses to win, but one of the best ways for America broadly, not just the AI sector, to win is preserving open models so that all American companies can do well.