Alex Wang 谈 Neuralink、AI 与人类未来

Alex Wang，欢迎来到节目，伙计。

Alex Wang, welcome to the show, man.

谢谢邀请，我很兴奋。

Yeah, thanks for having me. I'm excited.

我也是。就像早餐时跟你说的，我对科技了解不多，但自从 Joe 来了之后，我一直在努力理解这一切，这真是个迷人的话题。我现在很喜欢聊这个。所以，谢谢你来。

So am I. I like I was telling you at breakfast, I don't know a whole lot about tech, but ever since Joe came on, I've been trying to wrap my head around it all and it's just fascinating subject. I love talking about this subject now. So, thank you for coming.

嗯，这已经变得对国家安全和你热衷的那些事情至关重要。所以，我认为从根本上说，我们必须把技术搞对，否则事情会变得非常危险。

Well, it's becoming so critical to national security and all the stuff that you're very passionate about. So, I mean, I think fundamentally tech is like we got to get it right, otherwise stuff gets really dangerous.

是啊，把我吓坏了。事实上，我们刚才在楼下聊到你生孩子的事，你说你在等，然后提到了 Neuralink，我不得不打断对话。老兄，我挺担心 Neuralink 的，但你听起来很热衷。

Yeah. Scares the out of me. In fact, we were just having a conversation downstairs about you having kids and you were waiting and Neuralink came up and I had to pause the conversation. Dude, I'm like I'm worried about Neuralink, but it sounds like you're pretty gung-ho about it.

所以，有几点。我提到的基本上是我想要等到我们弄清楚 Neuralink 或其他脑机接口如何工作之后再生孩子。原因有几个。首先，在你生命的前七年，大脑的可塑性比任何其他时期都要高出一个数量级。有例子表明，如果一个孩子天生有白内障，看不见东西，并且在前七年都带着白内障生活，那么即使你在八九岁时摘除白内障，他们也无法学会看东西，因为在那前七年里，大脑学习解读眼睛信号至关重要。由于早期神经可塑性如此之高，我认为当我们有了 Neuralink 和其他技术时，天生就带着这些设备的孩子会以疯狂的方式学会使用它们——它们实际上会成为大脑的一部分，而这一点对于成年后植入 Neuralink 的人来说永远不会实现。所以这就是为什么我想等。

So yeah, a few things. So, I mean what I mention is basically I want to wait to have kids until we figure out how Neuralink or other brain-computer interfaces start working. Because there are a few reasons for this. First, in your first seven years of life, your brain is more neuroplastic than at any other point, by an order of magnitude. So there have been examples where if a kid is born with cataracts and they can't see through them, and they live their first seven years with those cataracts, then even if you remove them at age eight or nine, they won't learn how to see because it's so important in those first seven years that your brain learns how to read the signals from your eyes. So because neuroplasticity is so high in that early stage, I think when we get Neuralink and these other technologies, kids born with them will learn to use them in crazy ways—it'll actually become a part of their brain in a way that will never be true for an adult who gets a Neuralink hooked into their brain. So that's why I want to wait.

现在，Neuralink 这个概念——把大脑连接到电脑——我持务实态度。我的日常工作就是研究 AI。我相信 AI 会继续变得更聪明、更有能力、更强大。随着时间的推移，我们会有机器人和其他形式的 AI。人类进化的速度是固定的——以百万年为单位，因为自然选择很慢。所以如果往前推演，AI 会快速进步，而生物学进步缓慢。在某个时刻，我们需要自己接入 AI 的能力。我们需要让生物生命与硅基或人工智能并存，为了人类，我们需要接入它。所以最终，我认为我们需要某种直接连接大脑与 AI 和互联网的接口。这确实有潜在危险，也很可怕，但我们不得不这样做。AI 会这样发展，人类进步慢得多，我们需要接入那种能力。

Now, Neuralink as a concept—hooking your brain up to a computer—I take a pragmatic view. My day job is working on AI. I believe AI will continue becoming smarter, more capable, more powerful. We'll have robots and other forms for AI to take over time. Humans are only evolving at a certain rate—on the timescale of millions of years because natural selection is slow. So if you play this forward, AIs will keep improving quickly, while biology improves slowly. At some point, we need the ability to tap into AI ourselves. We need to bring biological life alongside silicon-based or artificial intelligence, and we'll want to tap into that for humanity's sake. So eventually, I think we'll need some interlink between our brains directly to AI and the internet. It is potentially dangerous and terrifying, but we just have to do it. AI will go like this, humans will improve much slower, and we'll need to hook into that capability.

我是说，你知道我已经表达过对此的恐惧，所以我正在克服自己的恐惧，我只是好奇，在你看来，可能会出什么问题？最明显的是某家公司黑进你的大脑。好吧，如果一家公司黑进你的大脑，那已经很糟了，但会是什么样？他们会直接把广告发到你的大脑里，或者让你想买他们的产品。但更糟的是，外国势力、恐怖分子、对手、国家行为者黑进你的大脑，窃取你的记忆或操纵你。那显然非常糟糕。

I mean, what you know that I've already expressed fear in this and so I'm curing my own fears, I'm just curious like what in your mind what could go wrong? I mean the obvious thing is that some corporation hacks your brain. Well, if a corporation hacks your brain, even that's pretty bad, but that'll be like what? They'll send ads directly to your brain or make you want to buy their products. But then even worse, a foreign actor, terrorist, adversary, state actor hacks into your brain and takes your memories or manipulates you. I mean that is obviously pretty bad.

是的，我认为这是一个巨大的风险。当然，如果你能直接连接某人的大脑，并能读取记忆、控制思想、读取思想，那非常糟糕。我和这个领域的很多科学家聊过，包括 Neuralink 的人，随着时间的推移，读心和控心将是技术发展的方向。所以我们必须小心对待，就像任何先进技术一样。但如果我们希望人类在 AI 不断进步时保持相关性，这将非常关键。

Yeah, I think that's a huge risk. For sure, if you have a direct link into someone's brain and the ability to read their memories, control their thoughts, read their thoughts, that's pretty bad. I've talked to a lot of scientists in this space, including folks at Neuralink, and mind reading and mind control are where the technology will go over time. So it's something we have to be careful about, like any advanced technology. But it's going to be pretty critical if we want humans to remain relevant as AI keeps getting better.

我采访过 Andrew Huberman 和 Ben Carson 博士。Huberman 告诉我，如果 Neuralink 能帮助盲人看见，那么他们能否在你的脑海中投射一个完全虚假的现实？也就是说，你看到天上不知道什么东西？他说是的，他们会有这种能力，而且不仅如此，他们还能操纵你的每一种感官——触觉、嗅觉、味觉，把恐惧等情绪植入你的大脑。

I interviewed Andrew Huberman and Dr. Ben Carson about this. Huberman told me that if Neuralink can help the blind see, then could they project a total false reality into your head? Meaning you're seeing who knows what in the skies everywhere? He said yes, they will have that ability, and not only that, they can manipulate every one of your senses—touch, smell, taste, insert emotions like fear into your brain.

我当时就想，‘天哪’。他们可以把你的整个现实操控成一个虚假的现实。你觉得呢？然后我问了本·卡森医生，他是世界知名的神经外科医生，他说，‘是的，绝对可以。’他又说，‘或者，他们也可以用它来做好事。’但他把问题抛回给我，他说，‘那你觉得会发生什么？’它会用于善事还是恶事？你怎么看？你觉得这真的可能吗？

And I was like, 'Holy shit.' Like, they could manipulate your entire reality into a false reality. I mean, you think that's... And then I asked Dr. Ben Carson about it and he said, you know, who's a world-renowned neurosurgeon, he said, 'Yes, absolutely. They...' He goes, 'Or, you know, they could use it for good.' But he goes, 'What?' He was kind of put it on me. He's like, 'Well, what do you think would happen?' And like, would it be used for good event? Or would it be used for evil? And I mean, what are your thoughts on that? You think that's a real possibility?

是的。首先，我们现在对大脑了解得还不多，但最终我们会了解的。科学会解决这个问题，对吧？你刚才提到的所有事情最终都会成为现实。操控情绪、操控感官。感官操控已经在发生了，比如在猴子身上，他们展示了可以——虽然不知道猴子主观感受如何——但能在猴子的视觉网格上投射，让它们非常可靠地点击正确的按钮。他们以某种方式接入了大脑中负责视觉处理的神经回路，把东西投射到猴子的视野里，让猴子总是点击你想要的按钮，然后给它奖励。天哪。所以，是的，操控视觉、操控感官、操控情绪。长期来看，还会利用记忆、操控记忆，这些都会成为现实。我觉得更令人兴奋的是，能够接入 AI，然后突然拥有关于一切的百科全书式知识，就像 ChatGPT 或其他 AI 系统那样。我可以以超人的速度思考，突然拥有更多信息可以处理，能瞬间理解世界上发生的一切。我认为这确实会让我们从认知上变得超人。但另一方面，正如你所说，风险也很大，那是一个巨大的攻击面。

Yeah. So, first of all, like we don't understand the brain too much today, but eventually we will. Like, science is going to solve this problem, right? And everything you just mentioned is ultimately going to be on the table, you know? Manipulating your emotions, manipulating your senses. The senses thing is already happening where I think in monkeys they've shown that like they can, you know, they don't know what it's like from the monkey's perspective, but they're able to project like on a grid of a monkey and get them to click on the right button really reliably. So they somehow hook into basically the neural circuits that are doing the visual processing, like image processing in the brain, and they're able to project things into their vision such that the monkey will always click the button that you want it to click. And then you give it a treat or something. Damn. And so yeah, manipulating vision, manipulating your senses, manipulating your emotions. This will be longer term but like leveraging your memories, manipulating your memories, manipulating like those are on the table. The other stuff that I think is more exciting is being able to hook into AI and all of a sudden I have encyclopedic knowledge about everything and just like ChatGPT or other AI systems do. I can think at superhuman speeds. I can all of a sudden I have way more information I can process, I can understand everything that's going on in the world and process that instantaneously. I think there's an element here where it'll legitimately turn us superhuman from a cognitive standpoint. But then to your point, the flip side of that is the risk the other way, which is that you're going to have a huge attack vector.

是的。我说过，我不是很懂技术，但你的公司 Scale AI，基本上——如果我说错了请纠正——Scale AI 就是 AI 用来得出答案、回应提示的数据库，对吧？

Yeah. I mean, like I said, I'm not super tech, but your company Scale AI, you basically, correct me if I'm wrong, Scale AI is basically the database that the AI uses to come up with its answers and answer your prompts and all of that. Correct.

是的。我们做几件事。我们帮助大公司和政府部署安全可靠的先进 AI 系统。我们参与流程的几乎每一步。但我们最早出名并且做得非常好的一点，正是你所说的：创建大规模数据集，我们称之为数据铸造厂，但就是为所有主要 AI 模型提供燃料的大规模数据生产。你知道，如果你在 ChatGPT 里提问，它能很好地回答很多问题，部分原因就是我们提供的数据。随着 AI 越来越先进，我们不断向这些模型注入更先进的科学信息和数据。我们还与最大的企业和政府合作，比如美国国防部和其他机构，利用他们自己的数据部署和构建完整的 AI 系统。我们公司的策略是专注于少数客户，这样我们能产生真正大的影响。所以我们与排名第一的银行、排名第一的制药公司、排名第一的医疗系统、排名第一的电信公司、排名第一的国家——美国合作。我们与他们一起，如何真正地改变他们今天的运营方式，利用 AI 从根本上改造他们今天的工作流程或运营。比如，如果你是全世界最大的医疗系统，必须为数百万患者提供护理，如何以最有效的方式做到？如何在物流上做得更好？如何改善诊断？如何改善所有患者的整体健康结果？这就是我们帮他们解决的问题。或者对于国防部，我们可以做很多事情来提高效率，最终实现更自动化的运作。我想你比任何人都更了解这一点。那么，如何开始用 AI 实施这些系统呢？

Yeah. So, we do a few things. We help large companies and governments deploy safe and secure advanced AI systems. We help with basically every step of the process. But the first thing that we were known for and we've done very well is exactly what you're saying, which is creating large-scale datasets and creating data foundry, as we call it, but creating the large-scale data production that goes into fueling every single one of the major AI models. And you know if you ask questions in ChatGPT, it's able to answer a lot of those questions well because of data that we're able to provide it. And as AI gets more and more advanced, we're continually fueling more advanced scientific information and data into those models. And then we also work with the largest enterprises and governments like the DoD and other agencies in the US to deploy and build full AI systems leveraging their own data. And our strategy as a company has been how do we focus on a small number of customers where we can have a really big impact. So we work with the number one bank, the number one pharma company, the number one healthcare system, the number one telco, the number one country, America. And we work with all of them to like how can you no kidding take how you are operating today and take the workflows that you're doing today or the operations that you have today and use AI to fundamentally transform them. So if you're the largest healthcare system in the world, and you have to provide care to millions of patients, how do you do so in the most effective manner? How do you do it logistically better? How do you improve your diagnosis? How do you improve the overall health outcomes of all your patients? That's a problem that we help solve with them. Or for the DoD, there's so much that we can do to operate more efficiently and ultimately in a more automated way. I mean, you'll know this, I think, better than anyone. And so, how do you start implementing those systems with AI?

我们稍后会更深入地讨论细节。我原本想说的是，如果最初是给 AI 提供数据，让它得出答案、回应提示，那么如果你脑子里有 Neuralink，它访问你的数据中心，那么把数据注入数据中心，再让所有有神经链接的人都接收到，这有多容易？所以可以是任何东西。举个例子，我是基督徒。很多人认为 AI 会篡改《圣经》，改变很多东西。那么，把篡改后的内容注入 AI 数据中心，然后它就成为新的真理，因为所有人访问的都是那个特定数据，这有多容易？

We'll dive way more in the weeds of that later in the interview. Kind of where I was going with this was if originally it was feeding the AI, you're giving the data to the AI to come up with the answers and answer the prompts, and so where I was going is if you have Neuralink in your head and it's accessing your data centers, how easy would it be to just feed into the data center that then feeds all everybody that has a neural link in their head? So it could be anything. I mean, here's an example. I'm a Christian. A lot of people think that AI is going to manipulate the Bible and change a lot of things. And so how easy would it be to just feed that into the AI data center and then that becomes the new truth because that's what everybody's accessing is that specific data.

是的。我认为这绝对是一个巨大的风险，这也是为什么我认为美国或其他民主国家在 AI 领域领先非常重要，而不是中国共产党或其他威权国家，因为即使是今天的 AI，也可以被用来进行大规模的宣传。但一旦有了神经链接或其他脑机接口，可以直接向人们的大脑植入思想，那将是前所未有的极端力量。那么，谁来掌控这种力量？谁来管理这项技术？谁来确保它被用于正确的目的？这些都是我们必须面对的最重要的社会问题。

Yeah. I mean, I think a yes for sure that's a huge risk and this is one of the reasons why I think it's really important that US or other democratic countries lead on AI versus the CCP like the Chinese Communist Party or other or Russia or other autocratic countries because the potential to utilize even AI today by the way you can use it to propagandize to a dramatic degree but yeah once you get towards you know you have neural link or other brain computer interfaces that can directly insert thoughts into people's brains. I mean, it's extreme power that has never existed before. And so, who governs that power? Who governs that technology? Who makes sure that it's used for the right purposes? Those are some of the most important societal questions that we'll have to deal with.

天哪，这该从何说起？你信任谁来控制你的思想？是啊，我觉得，嗯，这很有意思。

Man, I mean, where do you even start with that? Who do you trust to control your mind? Yeah. I mean, I think well, it's interesting.

我觉得有一件事，很多人现在都明白了，就是我们早上也聊到一点，就是如今连普通媒体都在多大程度上控制你的思想，控制你的观点和信念。我们谈到，媒体是不是在吹捧某些军事力量，让它们看起来比实际更可怕？现在有一些低级的宣传、操纵之类的东西，大概在 1 到 10 的尺度上处于 1 或 2 的水平。而一旦有了 Neuralink 或其他设备，就会达到 9 或 10。我觉得这真的很难。我认为没有任何国家准备好去治理像未来几十年我们将要开发的那么强大的技术。比如 AI，我不知道我们是否准备好了。脑机接口，我不知道我们是否准备好了。大规模机器人，我不知道我们是否准备好了。这些技术比以往任何东西都要强大得多。有时人们说 AI 就是新的移动技术，会和手机一样重要，但它会重要一千倍，影响更大。而且我们连手机都没监管好。所以，我们一定要把它做对。

I think the one thing that I think has been... a lot of people kind of understand it now... is like even the degree to which even just general media today kind of controls your mind or controls the opinions you have or the beliefs you have. And we were talking about like, does the media prop up certain military forces to make them seem far more fearsome than they actually are? There's some low-grade forms of propaganda, manipulation, all that stuff happening at a one or two level today. And once you have Neuralink or other devices, it's going to be like a nine or a ten. I think it's really hard. I don't think any country is prepared to govern technology as powerful as what we're going to be developing over the next few decades. Like AI, I don't know if we're prepared. Brain-computer interfaces, I don't know if we're prepared. Large-scale robotics, I don't know if we're prepared. These are technologies that are just so much more powerful than anything that has come before. Sometimes people say AI is the new mobile, it'll be as big as mobile phones, but it's going to be a thousand times bigger and more important and more impactful. And it's not clear that we did the best job regulating mobile phones even. So it's going to be really important that we get it right.

如果你觉得现代世界让你疲惫不堪，比如电磁场、人造光、种子油、微塑料、慢性压力，似乎我们的身体并不是为这些现代攻击而设计的，它们可能正在损害我们所有人。Armor colostrum 是一种生物活性全食物，可以帮助恢复细胞信号，从内部增强我们的健康。初乳是自然界最好的全食物，含有超过 400 种生物活性营养素，在细胞层面起作用，帮助增强免疫健康，强化肠道健康，并启动新陈代谢。自从他们寄给我试用后，我一直在用 Armor，现在精力好多了。它已经成为我日常的一部分。我早上和锻炼后服用。你准备好恢复健康了吗？我们为我的观众争取了一个特别优惠：首单 15%折扣。访问 tryarma.com/srs 或输入 SRS，即可获得首单 15%折扣。网址是 tr.com/srs。这些声明和产品未经 FDA 评估。这些产品不用于诊断、治疗、治愈或预防任何疾病。

If you ever feel like the modern world is wearing you down, things like EMFs, artificial light, seed oils, microplastics, chronic stress, it seems like our biology wasn't designed for these modern assaults, and it's probably taking a toll on all of us. Armor colostrum is a bioactive whole food that can help revive cellular signaling and help bolster our health from within. Colostrum is nature's finest whole food packed with over 400 bioactive nutrients that work at a cellular level to help strengthen your immune health, help fortify gut health, and help kickstart your metabolism. I've been using Arma ever since they sent me some to try and now I have a lot more energy. It's part of my daily routine now. I take it in the morning and after workouts. Are you ready to reclaim your health? We've worked out a special offer for my audience. Receive 15% off your first order. Go to tryarma.com/srs or enter SRS to get 15% off your first order. That's tr.com/srs. These statements and products have not been evaluated by the FDA. These products are not intended to diagnose, treat, cure, or prevent any disease or condition.

当你的新陈代谢正常工作时，你会在生活的方方面面感受到好处。Lumen 来帮忙了。Lumen 是世界上首批手持式代谢教练之一，可以通过你的呼吸测量新陈代谢。只需对着 Lumen 呼气，就能获得大量洞察。因为新陈代谢是你身体一切活动的核心，最佳代谢健康会带来更多精力、更好的健身效果等好处。我每天早上检查，应用程序会检测我是燃烧脂肪还是碳水化合物，并给出个性化指导，帮助改善营养、锻炼、睡眠甚至压力管理。这太棒了，给了我掌控健康所需的洞察。温暖的月份即将到来。让你的健康和健身恢复活力。访问 lumen.mers，获得 Lumen 10%折扣。网址是 lumme.me/srs，购买可享 10%折扣。感谢 Lumen 赞助本期节目。这些声明和产品未经食品药品监督管理局评估。这些产品不用于诊断、治疗、治愈或预防任何疾病。

When your metabolism is working properly, you can feel the benefits in literally every aspect of your life. Lumen is here to help. Lumen is one of the world's first handheld metabolic coaches that can help measure your metabolism through your breath. Just breathe into your Lumen and get a world of insight. Because your metabolism is at the center of everything your body does, optimal metabolic wellness translates to benefits like improved energy levels, better fitness results, and more. I check mine every morning, and on the app, it lets me know if it detects if I'm burning fat or carbs, and gives me tailored guidance to help improve my nutrition, workouts, sleep, and even stress management. It's amazing and gives me the insights I need to take control of my health. The warmer months are coming. Spring back into your health and fitness. Go to lumen.mers to get 10% off your Lumen. That's lumme.me/srs for 10% off your purchase. Thank you Lumen for sponsoring this episode. These statements and products have not been evaluated by the Food and Drug Administration. These products are not intended to diagnose, treat, cure, or prevent any disease or condition.

大家都明白我的意思。你基本上可以瞬间让整个军队、整个国家与你的思想、你的思维方式相连，然后操纵整个人群去做天知道什么事。希望是好事，但你知道事情通常的走向。但你对此很热衷。你会植入它吗？

Everybody that gets what I mean. You could basically instantaneously have an entire army, an entire nation that's linked into your thoughts, your way of thinking, and manipulate that entire population to do who the hell knows what. Hopefully something for good, but you know how things generally wind up going. But you're gung-ho about this stuff. Would you put it in?

我会植入，但在我愿意植入之前，有几件事需要发生。首先，我需要真正对网络攻防态势感到放心。我需要非常有信心能够防御任何攻击，任何针对我大脑接口的网络攻击。这是一个很大的门槛。其次，我需要有信心它不会以任何重大方式深刻改变我的意识。我认为你可以从其他使用者的数据中看到，并从其他人的采用中感受到。这两件事我需要非常非常确信。这是件大事。

I would put it in, but there are a few things that need to happen before I'd be willing to put it in. First, I would need to really feel good about the cyber offense defense posture. I need to have really good confidence that I would be able to defend from any attacks, any sort of cyber attacks into my brain interface. That's one big bar. And then I would need to feel confident that it wouldn't deeply alter my consciousness in any major way. I think you would see from data of other people who use it and get a sense from other people adopting it. Those would be the two things I would need to feel really really confident about. It's a big thing.

嗯，最后一件事，然后我们应该聊点别的。关于这个，最后一件事是……现在有很多关于人类如何永生，或者人类能否永生的讨论？如何不死？很多都集中在保持人体健康、治愈疾病，让人类活到几百岁。但我认为真正的终局是我们找到如何将意识从我们的肉脑上传到计算机。我有点把 Neuralink 或其他大脑与计算机之间的桥梁看作是第一步。

Well, the last thing, and then we should talk about other stuff, but the last thing about this is... there's a lot of talk right now about how humans will live forever, or can humans live forever? How do you not die? A lot of that is focused on keeping our human bodies healthy, curing diseases so that humans can live to hundreds and hundreds of years. But I think the actual endgame is that we figure out how to upload our consciousnesses from our meat brains into a computer. I kind of think about Neuralink or other bridges between your brain and computers as the first step there.

等等，什么？这又是另一个兔子洞了。你是说我们应该能够上传我们的意识，或者你想要能够将我们的意识上传到什么东西里？

Well, hold on. What? There's a whole another rabbit hole we can go down. So, you're saying that we should be able to upload our consciousness, or you want to be able to upload our consciousness into whatever?

是的，我觉得……现在我们是在科幻的深水区，但没错，我认为随着时间的推移会有……首先，我认为这项技术终将存在。我们现在还差得远，对吧？Neuralink 才刚刚有点用，对吧？所以还差得远，但将意识上传到计算机的技术将会存在。然后，好吧，假设我们坐在这里，比如 50 年后，这项技术存在了。

Yeah, I think... now we're on the deep end of sci-fi, but yeah, I think there will over time be... So, one, I think the technology will exist at some point. We're not close today, right? We barely have Neuralink kind of working, right? So we're not close, but the technology will exist to upload your consciousness onto a computer. And then, okay, let's say we're sitting here, it's like 50 years from now, this technology exists.

嗯，你在问这个问题，呃，你知道，人们会把自己的意识上传吗？首先，有很多人自然愿意这么做，比如身患绝症的人，呃，濒临死亡的人，嗯，你知道，还有那些非常前卫、正在尝试这项新技术的人，会有一类人最初就这么做。嗯，然后随着这种情况开始发生，他们上传了自己的意识，如果你有数字化的东西，你就有这些数字智能，嗯，它们就是真正的永生。这是你能得到的最接近真正永生的东西。嗯，所以我认为一旦技术存在，你知道，当它存在时，它会变得相当……它可能会成为大多数人类非常自然的选择。那么你认为，如果你上传了意识会发生什么？它会上传到什么地方，比如云端还是什么？

Um, and you're asking the question, uh, you know, are people going to upload their consciousness? Well, first off, there's a lot of people who who naturally would like people with terminal illnesses, um people near death, um you know, uh people who are like very fringe and you know, like experimenting with this new technology, there will be a class of people who will just initially do it. Mhm. And then um and then as that starts to happen and they upload their consciousness like the if you have a digital you have these sort of like digital intelligences um they're uh you know that's true immortality. That's the closest thing you'll get to to true immortality. Um and so the uh I think it's going to become like once the technology exists, you know, when it exists, it's going to become quite uh uh it's probably going to become a very natural path for most humans to go down. So what do you what do you think what do you think happens if you get your consciousness uploaded and what would it even be uploaded into like a cloud or something?

是的，会上传到云端。你怎么看？你认为通过将意识上传到云端，你能体验生活吗？

Yeah, it'd be uploaded to a cloud. What do you think? Do you think that you can experience life by uploading your consciousness to a cloud?

是的。所以，呃，这有几件事。首先，嗯，我非常相信机器人技术。我认为我们基本上正处于机器人革命的起点。嗯，我们还处于非常早期的阶段，但人们开始制造人形机器人。它们会变得非常非常出色。人们开始将它们应用于制造业、工业化和其他场景。嗯，我认为成本会大幅下降，所以最终，是的，如果你相信，如果你上传了，然后你可以下载或下行链接到一个人形机器人，那么你就能像在其他世界一样体验现实世界。嗯，或者你可以继续在某种模拟宇宙中，你几乎可以在云端玩电子游戏之类的东西，这可能是另一种选择。哇。你认为当你死的时候会发生什么？

Yeah. So, so uh yeah, this is uh few things. So, first um I'm a big believer in robotics. I think we're basically at the start of a robotics revolution. Um and we're in the very early innings of it, but people are starting to make humanoid robots. They're going to get really, really good. people are starting to apply them to manufacturing and industrialization and other contexts. Um I think the costs are going to come down dramatically and so eventually yeah if you you would believe that if you uploaded and then you could download or down link down to a down to a humanoid robot then you would kind of experience the real world like any other world. Um, or you would you could continue in some kind of like simulated universe in uh you could almost like play a video game in the cloud kind of thing and that could be like the other alternative. Wow. What do what do you think happens when you die?

呃，你知道，随着 AI 变得如此……嗯，埃隆总是谈论我们生活在一个模拟中，对吧？嗯，我记得我第一次听他这么说时，我想，啊不，我不相信，我不相信我们生活在模拟中。但是，嗯，随着 AI 在模拟世界方面越来越好，比如我不知道你有没有看过这些 AI 视频生成模型，比如 Sora 或 VO 或其他一些模型，但你知道它们可以制作完全逼真的视频，嗯，大多数人无法区分 AI 生成的视频和真实视频。随着这种情况的发生，我越来越觉得我们可能生活在一个模拟中。

Uh, you know, as AI has gotten so um so Elon always talks about how we're in a we live in a simulation right um uh and I remember when I first heard him talk about this I was like ah no this is like I don't believe that I don't believe we're in a simulation but um but as AI has gotten better and better at simulating the world like I don't know if you've seen these AI video um generation models like Sora or VO or some of these models but you know they can produce videos that are totally realistic um you would most people could not tell the difference between AI well we're seeing this AI generated video and uh and and real video and as that's happening it's making me think more and more that we probably live in a simulation no it.

是的。你怎么……这已经很有趣了。我们甚至还没开始采访。你怎么认为我们生活在模拟中？我的意思是，我知道他们说他们无法反驳这一点。是的。你不能……这有点像那种事情。没有办法证明或反驳你生活在模拟中。所以，但这就像任何来世想法或宗教思想一样，所有这些事情基本上都是无法证明的。但我认为这是事实的原因是，我认为在我们有生之年，我们将能够创建超逼真的现实模拟。我认为我们将有能力以超逼真的精度模拟我们世界的不同版本。嗯，这将在未来几十年内发生。如果我们能做到，就像《瑞克和莫蒂》的那一集，如果我们作为一个智能种族有能力产生数百万个模拟世界，嗯，那么很可能我们也是某个其他更智能或更有能力的物种的模拟。你认为当你死的时候，意识现在去了哪里？

Yeah. How do you just This is already fascinating. We haven't even got to the interview yet. How How do you think we're living in a simulation? I mean, I know they they say they they cannot disprove it. Yeah. You can't like It's kind of one of these things. There's there's no way to prove or disprove that that you live in a simulation. And so, but it's like it's like it's like any, you know, afterlife thought or religious thought like all these things that are like fundamentally unprovable. But the reason I think it's the case is I think in our lifetime we are going to be able to create simulations of reality that will be hyper realistic. Like I think we are gonna create the ability to um simulate different versions of our world with hyperrealistic accuracy. Um and uh and that will happen over the next few decades. And if if we can like it's kind of like that Rick and Morty episode where if we have the ability as an intelligent race to produce, you know, millions of simulated worlds, um then the likelihood is that we're, you know, we're probably also the simulation of some other uh more intelligent or more capable species. Where do you think consciousness goes right now when you die?

呃，如果我们就是超级先进的机器人呢？是的，我认为嗯，你的意识只是被下载到另一个身体世代。是的，没错。那会是……那是一种思考方式，就像，是的，这是一个运行中的大模拟，一旦你被下载或移除或从一个实体中退役，你就会上传到另一个实体。嗯，这似乎合理。我认为在另一个世界里，意识可能没那么重要。就像，随着模型越来越好，随着 AI 模型越来越好，嗯，你看着它们，嗯，你肯定会想，在某个时候，你会有真正有意识的模型，而且可能事实就是，你知道，这是可以工程化的东西，如果它是可以工程化的，那么嗯，那么一切皆有可能。

Uh, what if we are what if we are the super advanced robotics? Yeah, I think um and your consciousness just gets downloaded into another body generation. Yeah, that's true. That would be that's that's something like one way to think about it which is like yeah it's all this big simulation that's running and as soon as like you know you get you get kind of like downloaded or like taken off or like decommissioned from you know one entity you get like you know uploaded to another entity kind of thing. Um it's kind of that that's plausible. I think there's another world where like consciousness is like is consciousness may not like be that big a deal so to speak. like it could be the case that you know I definitely as as the models have gotten better and better as the AI models have gotten better and better um you look at them and uh you know you definitely wonder if at some point you're just going to have models that are properly conscious and it may just be the fact that like you know it's something that can be engineered and if it's something that can be engineered then um then all bets are off I think

该死，想想真疯狂。是的。是的。但是呃，我们进入采访吧。你准备好了吗？

Damn it's pretty wild to think about. Yeah. Yeah. But uh let's move into the interview. You ready?

是的。

Yeah.

好的。每个人都从这里开始介绍。那么，开始吧。Alex Wang，Scale AI 的创始人兼 CEO，这家公司是 AI 革命的支柱，提供驱动 AI 革命的数据和基础设施。神童，在新墨西哥州洛斯阿拉莫斯长大，周围都是科学家，父母是从事军事项目的物理学家。编程奇才，15 岁时就已经在 Kora 解决了让博士们困惑的 AI 问题。有远见的企业家，19 岁从麻省理工学院辍学，将 Y Combinator 初创公司转变为国家安全巨头，帮助美国在全球 AI 竞赛中保持领先。24 岁时成为世界上最年轻的白手起家亿万富翁，建立了一家估值近 250 亿美元的公司，同时专注于解决 AI 最大的瓶颈——高质量数据。敢于将中美 AI 竞争称为 AI 战争，警告像 Deepseek 这样的中国初创公司正在以比大多数人意识到的更快的速度缩小差距。你的使命是构建一个 AI 驱动进步、安全和机遇的未来。那么现在有一个大问题，每个人都在思考。AI 是下一个石油吗？

All right. Everybody starts off with an introduction here. So, here we go. Alex Wang, founder and CEO of Scale AI, a company that's backbone of the AI revolution, providing the data and infrastructure that powers the AI revolution. Child prodigy who grew up in Los Alamos, New Mexico, surrounded by scientists with parents who were physicists working on military projects. Coding wizard who by age 15 was already solving AI problems at Kora that stumped P PhDs. Visionary entrepreneur who dropped out of MIT at 19, turning a Y combinator startup into a national security powerhouse that's helping the US stay ahead in the global AI race. youngest self-made billionaire in the world by age 24, built a company valued at nearly 25 billion while staying laser focused on solving the biggest bottleneck in AI highquality data. Unafraid to call the US China AI competition and AI war, warning that the Chinese startups like Deepseek are closing the gap faster than most realize. guided by your mission to build future where AI drives progress, security, and opportunity. And so there's a big question right now that everybody's that everybody's thinking about. Is AI the next oil?

是的，我有一些想法。所以呃，AI 绝对是下一个……在某些方面它是下一个石油。AI 将从根本上成为任何未来经济、任何未来军事、任何未来政府的命脉。如果你推演下去，你会发现一个国家或经济体利用 AI 提高经济效率、自动化经济部分、进行自动化研发、推动科学进步的程度，将决定其成败。

Yeah, I think uh few thoughts there. So uh AI is definitely the next um some ways in which it's it is the next oil. AI will fundamentally be the lifeblood of any future economy, any future military, any future government. Like if you play it out, you're like the degree to which a country or economy is able to utilize AI to make its economy more efficient, to automate parts of its economy, um to do automated research and development, automate R&D, like you know, push forward in science um using AI.

这一切意味着，有效采用 AI 的国家将实现近乎无限的 GDP 增长，而没有采用的国家将被甩在后面。所以，它是推动每个国家未来的燃料。而且，我认为硬实力也是如此。如果你看看未来的军队或战争的样子，AI 是核心。我们肯定会谈到这一点。然后，它不像石油的地方在于，石油是有限的资源。发现大量石油储备的国家拥有那个储备，但总有一天会耗尽。比如挪威，它最终会枯竭。所以它在某个时期赋予国家权力和经济财富，然后你用完了，再寻找更多石油。而 AI 是一种会不断自我叠加的技术：AI 越智能，你获得的经济权力就越大，这意味着你能建造更智能的 AI，进而获得更多经济权力，如此循环。所以 AI 会形成一个持续加速的飞轮，它不是一种有时限的资源，而是一种会永远加速前进的东西。数据是其中的一部分，而且是核心部分。

All that stuff is going to mean that countries that adopt AI effectively will have like, you know, nearly infinite GDP growth and countries that don't adopt it are going to get left behind. Um, so it is sort of the fuel that will power the future of every country. And by the way, I think the same is true of hard power. Like if you look at what the militaries of the future are going to be like or what war looks like in the future, AI is at the core of what that is going to look like. I'm sure we'll get into that. Um and then the ways that it's not like oil is, you know, oil is this finite resource. Countries that stumble upon large oil reserves have that large oil reserve. At some point, it's going to run out. Like in Norway, it runs out at some point. And so it lends the country power and economic riches for a time period. Um and then you exhaust it and then you're looking for more oil. Whereas AI is going to be a technology that will just keep compounding upon itself and will keep, you know, the smarter AI, the more economic power you're going to get, which means you're going to build smarter AIs, which means you have more economic power and so on and so forth. And so there's going to be a flywheel that keeps going on AI, which means that it's not going to be a time-limited resource, let's say. It's going to be something that will just continue racing and accelerating for the entire perpetuity and data is part of that. Data is a big part of that. Data is the core part of it. Yeah.

其实很多时候，我喜欢把数据比作石油，而不是 AI。我刚才说错了，我本意是说数据。

So, a lot of times actually I like to compare data to oil versus AI. That's actually what I meant. I messed that up. I meant to say data. Yeah.

对，我觉得完全正确。数据。如果你思考 AI，它归结为三个部分：算法，也就是 AI 系统背后的实际代码，需要非常聪明的人来编写。我以前也写过一些算法。然后是算力，即计算能力，这归结为大规模数据中心。你有足够的电力吗？有足够的芯片吗？这就像一个大工业项目。最后是数据。你有所有的生命线，即输入算法供它们学习的数据吗？它确实是很多智能的原材料。这就是为什么我认为数据最像石油，因为它被输入算法和芯片，使 AI 如此强大。我们对 AI 的所有了解都表明，你在算法、算力和数据这三方面做得越好，你的 AI 就越强。关键就是在这三方面竞相领先。

Yeah. Well, I think that's totally true. Like data. If you think about AI, it boils down to three pieces. There's the algorithms, like the actual code that goes into the AI systems that really smart people have to write. I used to write some of these algorithms back in the day. Then there's the compute, the computational power which boils down to large scale data centers. Do you have the power to fuel them? Do you have the chips to go inside them? That's like a large scale industrial project. And then data. Do you have all the lifeblood, or all the data that feeds into these algorithms that they learn off of? It's really kind of like the raw material for a lot of this intelligence. And that's why I think data is the closest thing to oil because it is what gets fed into these algorithms and chips to make AI so powerful. And everything we know about AI is that the better you are at all three of these things—algorithms, computational power, data—the better your AI gets. And it's just all about racing ahead on all three of these.

那么当我们看到像 ChatGPT、Grok 这类东西时，它们是共享一个数据中心，还是各自拥有完全独立的数据中心？

So when we see like ChatGPT, Grok, these types of things, are they sharing a data center or are they all completely separate data centers?

它们都有独立的数据中心。这实际上是公司之间竞争的主要领域之一——谁有能力获得更多电力并建造更大的数据中心。因为最终，随着 AI 越来越强大，问题就变成了你能运行多少个 AI？假设我们有一个非常强大的 AI，能够进行自动网络黑客攻击。它可以登录任何服务器，或者尝试入侵某个网站或系统。那么问题来了：如果我拥有它，我能运行多少副本？一千个？一万个？一亿个？这完全取决于你拥有并运行多少个数据中心。而这又取决于你有多少电力来支撑这些数据中心，有多少芯片可以在这些数据中心运行，如何让它们尽可能长时间在线，以及什么数据在不断为这些模型提供燃料，使它们变得越来越好。这就是为什么 AI 公司之间竞争的主要方式之一——xAI（埃隆的公司）、OpenAI、谷歌、亚马逊、Meta 等——就是看谁现在在为五年、六年后的数据中心获取更多电力和土地。所以，五六年后的战役实际上在今天就已经打响了。

They all have separate data centers. This is actually one of the major lanes of competition between the companies—who has the ability to secure more power and build bigger data centers. Because ultimately, as AI gets more and more powerful, the question then becomes how many AIs can you run? So let's say for a second that we get to a really powerful AI that can do automated cyber hacking. So it can log into any kind of server or try to hack some website or some system. Then the question is: if I have that, how many copies of that can I run? Can I run a thousand copies? Can I run ten thousand? Can I run a hundred million? And that all boils down to how many data centers you have up and running. And that then boils down to how much power you have to fuel those data centers, how many chips you have to run in those data centers, how do you keep those online for as long as possible, and what data is constantly fueling those models to keep getting them better and better. And so this is one of the reasons why one of the major ways that AI companies compete—between xAI, Elon's company, and OpenAI, Google, Amazon, Meta, and all these companies—is just who right now is securing more power and more real estate for data centers five years from now and six years from now. And so the battles five, six years down the line are being fought literally today.

哇，太迷人了。在进入你的人生故事之前，还有几件事。给你带了份礼物。哦，每个人都有。太棒了。Vigilance Elite 小熊软糖。给你。美国所有 50 个州都合法。没有恶作剧，只是糖果。美国制造。还有一件事，我有一个 Patreon 账号，是一个订阅社区。他们从一开始就支持我，那时我还在阁楼里做这个。然后我们搬到了这里，现在又要搬到一个新工作室，团队规模是原来的 10 倍，最初只有我和我妻子。但这一切都归功于他们。所以今天我能坐在这里和你聊天，也是因为他们。我做的其中一件事是让社区成员有机会向每位嘉宾提问。这是 Kevin Omali 的问题：既然 AI 现在能够复制我们现实的许多方面，你是否预见未来所有在审判中呈现的视频或照片证据都会变得可疑，因为任何内容都可能通过 AI 工具复制？

Wow, man. That's fascinating stuff. Well, couple more things before we get into your life story here. Got you a gift. Oh, man. Everybody gets one. Love it. Vigilance Elite Gummy Bears. There you go. Legal in all 50 states. No funny business, just candy. Made here in the USA. Yeah. And um and then one other thing, got a Patreon account. It's a subscription account. It's turned into quite the community. And um they've been here with me since the beginning when I was running this thing out of my attic. And then we moved here and now we're moving to a new studio and the team's 10 times bigger than what it was, which was just me and my wife. But um it's all because of them. And so they're the reason I get to sit here with you today. And uh so one of the things I do is I offer them the opportunity to ask every guest a question. This is from Kevin Omali. With AI now able to essentially replicate so many facets of our reality, do you see a future where all video or photographic evidence presented in trials become suspect based on the ability for any of it to have been replicated through artificial intelligence tools?

嗯，是的，这又回到了我们刚才讨论的话题。我确实认为 AI 将实现疯狂的模拟水平，而我认为我们的法庭还没有准备好。就像 Kevin 说的，AI 将能够生成非常逼真的视频和图像，而我们甚至还没有达到那个阶段。现在，你还能分辨出这些视频或图像是 AI 生成的。但这种情况会不断改善，最终将无法与真实视频区分。那么，我们到底该如何辨别什么是真实的，什么是 AI 生成的呢？我认为有两件事。首先，人们需要非常好的检测器，极其出色的那种。

Uh yeah, so this goes back to what we were just talking about. I do think AI is going to enable you to do crazy levels of simulation and I don't think our courts are ready for it. I think that like Kevin was saying, AI will be able to generate very convincing video, very convincing images in a way that we're not even really at that point yet. Like right now, you can still tell when these videos or images are AI generated. That's going to keep getting better and it's going to be indistinguishable from real video. So, how the hell are we going to discern what's real and what's AI generated? I think that there's two things. I think first people are going to need really good detectors, like insanely good.

而且我觉得现在的孩子其实已经有更好的辨别能力了，因为他们从小就在互联网上长大，那里什么都有，所以他们自然学会了越来越好的辨别能力。但这是其一。第二，我认为这会是一个需要大力推动各种政策和监管的领域，但核心问题是：如果在审判中使用了伪造的视频或图像，并且被发现是伪造的，后果会是什么？我认为关键在于调整机制，让伪造证据成为最严重的罪行之一。这样，如果你正确设置了激励措施，就能有效阻止人们滥用这些工具。

And I think kids today, by the way, already have much better detectors because they grow up on the internet where there's just so much of everything that they already kind of learn to have better and better detectors. But that's one. And then the second is, I think there's going to be an area where I know there's a lot of push for various forms of policy and regulation, but this is going to be a major question: if there's fabricated video or imagery used in a trial and it's discovered that it was fabricated, what are the consequences? I think it's about tuning that such that if you fabricate evidence or fabricate things, then that's maybe the worst offense of all. Then you deter a lot of usage of those tools if you set up the incentives in the right way.

是啊。我首先想到的就是美国政府。就像我刚才在工作室里跟你说的，政府对待那些黑水公司员工的方式——他们删除了证据。但与其删除证据，他们完全可以制造新证据，比如在巴格达广场伪造一场枪战，证明那些人确实有罪。而背后就是政府在操控。你知道，我们在布拉德·吉里、埃迪·加拉格尔、黑水公司那些人身上都见过。就在我这个小圈子里，这种事已经见过无数次了。而且你看看整个欧洲的选举：他们取消了乔治斯库的资格，说他受俄罗斯影响；法国的玛丽娜·勒庞也被搞掉了。大概六个月前，他们还讨论过要取消德国某个人的资格。这太疯狂了，真的把我吓坏了。把我吓坏了，因为他们可以随便陷害任何人。

Yeah. I mean, the first thing that goes to my mind is the US government. I mean, just showing you around the studio and stuff, talking about what the government did to those Blackwater guys I was telling you about. They deleted the evidence. Well, instead of deleting the evidence, they could make new evidence that is a fake gunfight in the source square, Baghdad, that proves they're guilty. And then it's the government behind it. You know, we've seen it with Brad Giri. We've seen it with Eddie Gallagher. We've seen it with the Blackwater guys. We've seen it a ton just in my small network circle. And I mean, you see what's going on with the elections all over Europe. They pulled Georgescu, calling him under Russian influence. Marie Le Pen in France, done. I mean, they were talking about pulling somebody in Germany not too long ago, maybe about 6 months ago. And it's just crazy, you know, and it scares the hell out of me. Scares the hell out of me because then they can just frame anybody they want.

是的。我认为人工智能的一个后果是，今天拥有权力的机构将获得更多权力。它本质上不是民主化的，而是一种集权技术。所以我们需要建立机制，让我们能够信任这些机构，否则结局不会好。

Yeah. I think one of the outcomes of AI is that institutions that have power today will gain way more power. It's not naturally democratizing. It's a centralizing kind of technology. So we need to build mechanisms so that we can trust those institutions, otherwise it doesn't end well.

好了，我们来聊聊你的故事。我也有礼物。我喜欢礼物吗？好的，太好了。那么，有几样东西。我在新墨西哥州的洛斯阿拉莫斯长大。我的父母都是物理学家，在那里的国家实验室工作。这里是原子弹的诞生地。不知道你看没看过《奥本海默》，那部电影有一半场景就设在洛斯阿拉莫斯，我的家乡。所以我们有一顶洛斯阿拉莫斯的帽子。洛斯阿拉莫斯国家实验室的帽子。老兄，这太酷了。我们还有一些洛斯阿拉莫斯的硬币。一枚是关于原子弹的。一枚是关于实验室主任诺里斯·布拉德伯里的。还有一枚是关于原子弹之父的。给你。我们还有一份给科学家的手册副本，是从曼哈顿计划中解密的。哇。这个只是个好玩的东西——一个火箭模型套件，给你和你的孩子。哦，天哪，他们一定会喜欢的。谢谢，老兄。谢谢。这个放在工作室里会非常酷。太棒了。

Well, let's get to your story. I have gifts too. Do I love gifts? Okay, great. So, a few things. I grew up in Los Alamos, New Mexico. My parents were both physicists who worked at the national lab there. This is the birthplace of the atomic bomb. I don't know if you saw Oppenheimer, but half of that movie is set in Los Alamos, where I'm from. So, we got a Los Alamos hat. Los Alamos National Laboratory hat. Dude, it's very cool. We have some Los Alamos coins. One about the atom bomb. One about Norris Bradberry, who's a lab director. And then also a coin about the father of the atomic bomb. Here we go. We have a copy of the manual that they gave to the scientists that got declassified from the actual Manhattan Project. Wow. And this one's just a fun one. It's a rocket kit for you and your kids. Oh, man. They're going to love that. Yeah. Thank you, dude. Thank you. This is going to look awesome in the studio. That's very cool.

是啊，这有点超现实。每个人都说人工智能是下一个曼哈顿计划。而那就是我长大的地方，所以感觉挺奇怪的。

Yeah, it's been kind of surreal. I mean, everybody calls AI the next Manhattan Project. So it's been funny because that's where I grew up. It feels weird.

我敢肯定。那你小时候喜欢什么？

I'll bet it does. So, what were you into as a kid?

再说一遍，我的父母都是物理学家，我父亲的父亲也是物理学家。所以我是在一个纯粹的物理学家庭长大的。科学、技术、物理、数学——这些都是我小时候非常热衷的东西。我记得在餐桌上，我们会讨论黑洞、虫洞、外星生命、超新星、遥远的星系等等。这些东西对我来说非常迷人。可以说，我一直在思考理解宇宙这件事。然后我非常喜欢数学。四年级时，我参加了人生中第一次数学竞赛。那是新墨西哥州全州的比赛。我得了全州四年级组第一名。这激发了我的竞争基因，然后我就沉迷于数学竞赛、科学竞赛、物理竞赛。

So, again, both my parents are physicists and my dad's dad was a physicist as well. So I grew up in this pure physics family. Science, technology, physics, math—these were the things I was really excited about as a kid. I remember around the dinner table we would talk about black holes and wormholes and alien life and supernova and far away galaxies and all that stuff. That stuff was all very captivating to me. I was thinking about understanding the universe, for lack of a better term. And then I really liked math. In fourth grade, I entered my very first math competition. It was for the whole state of New Mexico. I scored the best out of any fourth grader in New Mexico. That activated this competitive gene in me, and then I just got consumed by math competitions, science competitions, physics competitions.

你四年级的时候在学什么数学？

What kind of math are you doing in fourth grade?

嗯。我父母在二年级就教了我代数，大概吧。可能是二年级到三年级之间。

Yeah. My parents taught me algebra in second grade, I want to say. Maybe between second and third grade.

真的吗？你二年级就掌握了代数？

Are you serious? You mastered algebra in second grade?

我不知道算不算掌握，但我确实在玩代数。他们教了我基础，然后我在二年级就整天琢磨这个。那时候大概七八岁吧？对。所以到四年级时，我已经能做一些基础代数和几何了。到了中学，我开始学微积分，还有大学水平的数学。高中时，我迷上了电脑，整天编程。我发现科学和数学很酷，但有了电脑和编程，你真的能做出东西来。这最终成了我的主要爱好。

I don't know if I mastered it, but I was playing around with algebra. They taught me the basics, and I would just spend all time thinking about it in second grade. It's like seven or eight years old, right? Yeah. So by the time I was in fourth grade, I could do some basic algebra and some basic geometry. Then in middle school, I was doing calculus and college-level math as well. In high school, I just became obsessed with computers and spent all day programming. I realized science and math are cool, but with computers and programming, you could actually make stuff. That ended up becoming the major obsession.

回到餐桌上的对话。是啊。洛斯阿拉莫斯那个地方有很多阴谋论和各种传闻。遥视之类的东西似乎都源自洛斯阿拉莫斯。但我父母都是洛斯阿拉莫斯的物理学家。

Back to the dinner table conversations. Yeah. I mean, Los Alamos, there's like a lot of conspiracies and all kinds of stuff going on about that place. Remote viewing, all this stuff seems to stem from Los Alamos. But I have two parents that are physicists in Los Alamos.

你们在聊黑洞和外星人。你怎么看？有外星人吗？

You guys are talking about black holes and aliens. What do you think? Are there aliens?

有个著名的悖论，费米悖论：我们生活在如此浩瀚的宇宙中，有数十亿、数千亿、万亿颗恒星和行星，但其中没有一个拥有智慧生命的概率有多大？我认为宇宙中其他地方肯定有智慧生命。但问题是我们相距太远，比如数百万或数亿光年，我们永远无法沟通。所以这说得通。还有暗黑森林假说，这是我最相信的。费米悖论说没有智慧生命的概率几乎为零。问题是我们为什么没看到外星人？暗黑森林假说最初来自一本科幻小说，意思是如果你推演博弈论，智慧生命不想广播自己的存在，因为那会让你成为超级侵略性生命体的目标。所以一旦你变得智慧并成为多行星物种，你就会意识到最好低调行事，保持孤立。外星人确实存在，但每个人的动机都是保持孤立。

So there's this famous paradox, the Fermi paradox, which is: what are the odds that we live in this vast universe with billions, hundreds of billions, trillions of other stars and planets, and none of them have intelligent life? I think definitely somewhere else in our universe there has to be intelligent life. But the issue is if we're really far apart, like millions or hundreds of millions of light years apart, there's no way we're ever going to communicate. So that's plausible. Then there's the dark forest hypothesis, which I actually believe the most. The Fermi paradox says the odds of no intelligent life are probably zero. The question is why aren't we seeing any aliens? The dark forest hypothesis, originally from a sci-fi novel, is that if you play out the game theory, intelligent life doesn't want to broadcast its existence because that makes you a target for hyperaggressive forms of life. So once you become intelligent and multiplanetary, you realize it's best to mind your own business and stay isolated. There are aliens out there, but everyone's incentive is to stay isolated.

有意思。我不知道。我以前相信，但后来采访了一群人。老实说，我觉得这都是分心的事。还有另一面：UFO 是阴谋，这样军方就能进行空中测试，然后被否定。我和人聊过，没有确凿证据。然后就是‘那是机密’。我不知道。有时我想，也许在任何特定时间点，只有一颗行星承载生命，当那颗行星过时，一切都会灭绝。也许它转移了——也许 50 亿年前是火星，然后生命在地球上发展。我一直在反复思考。

Interesting. I don't know. I used to believe in it. Then I interviewed a bunch of guys. I think all this stuff is a big distraction, to be honest. There's definitely the other portion: UFOs are a conspiracy so the military can do airborne testing and get discredited. I've talked to people, there's just no hard evidence. And then it's 'well, that's classified.' I don't know. Sometimes I think maybe at any given point in time, there is only one planet that holds life, and when that planet becomes obsolete, everything goes extinct. Maybe it moves—maybe it was Mars 5 billion years ago and then life developed on Earth. I go back and forth on this all the time.

对，完全同意。因为我们的恒星有生命周期，随着它演变，太阳系不同地方会有不同条件。所以这是个合理的理论。我认为这个和之前我们聊的——意识和来世——都是大问题，因为我们可能永远不知道答案。

Yeah, totally. Because our star has a life cycle, and as it goes through that cycle, different points of our solar system have different conditions. So that's a plausible theory. I think both that and what we were talking about before—consciousness and the afterlife—are some of the great questions because we'll probably never know the answers.

你父母在洛斯阿拉莫斯做什么工作？他们还在那里工作吗？

What were your parents working on at Los Alamos? Are they still working there?

是的，我妈妈还在那里工作。我爸爸不在了。他们是洛斯阿拉莫斯国家实验室里从事机密工作的部门成员。他们有安全许可。我妈妈有能源部的许可。我记得小时候，我以为他们在做很酷的物理研究。我以为洛斯阿拉莫斯是造原子弹的地方，但几十年后它只是一个先进的科学研究区。直到大学我才意识到它可能主要还是武器研究。我离开后，他们重启了核弹芯生产——制造核武器核心——大概在 2018-2019 年。然后我才明白它主要是研究新核弹头的设施。所以我猜我父母就是做那个的。

Yeah, my mom still works there. My dad doesn't. They were part of divisions at Los Alamos National Lab that worked on classified work. They had clearance. My mom has clearance with the DOE. I remember growing up, I just assumed they were working on cool physics research. I thought Los Alamos was where the atomic bomb was built, but decades later it's just an advanced scientific research area. It wasn't until college that I realized it's probably still mostly weapons research. Since I left, they restarted nuclear pit production—manufacturing the cores of nuclear weapons—around 2018-2019. Then I realized it's mostly a research facility for new nuclear warheads. So I guess my parents worked on that.

是啊，太疯狂了。哇。你小时候除了数学还喜欢什么？

Yeah. Damn, that's crazy. Wow. What else were you into as a kid other than mathematics?

我热爱数学、编程、科学，所有这些。我特别喜欢小提琴，每天练习一小时。很大程度上是因为完美有一种真正的美。我认为这在很多艺术、音乐，甚至几乎所有事情中都成立。我现在的生活和日常工作中也能看到这一点。但就是，如果你练习足够多，能完美演奏一首曲子，那就会很美。而在那之前，简直一塌糊涂。对我来说，这个概念很美：一旦你把某件事做到完全完美，它就变得美丽。我小时候对此非常着迷。

I loved math. I loved coding. I loved science. I loved all that stuff. I was really into violin. I would practice an hour of violin a day. A lot of that was because there is a real beauty to perfection. I think this is true in a lot of arts, a lot of music, a lot of frankly everything. I see it even in my current life, in my day-to-day job. But there was just, if you practice enough to play a piece perfectly, then it would be beautiful. And along the way, it's total dog until you get to perfection. There's a lot of beauty to that concept to me: once you get something totally perfect, it becomes beautiful. That was captivating when I was a kid.

所以你从小就是个完美主义者，现在也还是。

So you were a perfectionist from a young age and you're still a perfectionist today.

是的。我在完美中看到了很多美。但现在我会说，我认为我们没有奢侈的条件去做完美主义者。我现在务实多了。世界极其混乱。现实超级混乱。坏事不断发生，好事也很多，但完美不是一个可行的目标。我们永远达不到完美。所以我现在务实多了，但我确实在完美中看到了很多美。我也是个完美主义者，每天都在与之斗争。我有强迫症。我读过相关文章，看过演讲。我得出的结论是——我讨厌这么说，因为我骨子里是个完美主义者——完美主义会阻碍成功。

Yeah. I see a lot of beauty in perfection. Now I would say I don't think we have the luxury to be perfectionists. I'm much more pragmatic now. The world is extremely messy. Reality is super chaotic. There's a lot of bad going on constantly, a lot of good too, but perfection is not a plausible objective. We're never going to get perfection. So I'm a lot more pragmatic now, but I do see a lot of beauty in perfection. I'm also a perfectionist. I battle it every day. I'm OCD. I've read about it, watched talks. I came to the conclusion, which I hate saying because I am a perfectionist at heart, that perfectionism can get in the way of success.

你发现了吗？我是说，问这个问题甚至听起来很奇怪，因为你是世界上最年轻的亿万富翁，24 岁就做到了，现在你 28 岁，所以问完美主义是否阻碍了你听起来很奇怪，但确实如此吗？

Did you find that? I mean, it sounds weird even asking you the question because you're the youngest billionaire in the world at age 24 and you're 28 years old now, so it sounds weird saying did perfectionism hold you back, but did it?

我想在某个时刻，某个开关被拨动了，我意识到你必须多次运用 80/20 法则。你只需付出 20%的努力，达到 80%的效果，并且你必须接受这一点。而且你必须一遍又一遍地这样做。所以在某个时候，我内化了这一点，这与完美主义背道而驰，完全相反。所以现在我这样想：有些事情完美主义确实是正确答案，而有些事情你只需要接受不完美，速度是目标，而不是完美。老实说，现在我认为大多数事情速度是目标，而不是完美。所以我在这方面经历了一段完整的旅程。

I think at some point, some bit flipped and I realized you got to just do the 80/20 lots of times. You got to do 20% of the effort that's 80% as good and you just have to be okay with that. And you just have to do that over and over again. So at some point I internalized that and it's anathema to perfectionism, the exact opposite. So now I think about it as: there are some things where perfectionism really is the right answer, and some things where you just got to be okay with imperfection and speed is the objective versus perfection is the objective. Honestly, now I think most things are speed is the objective, not perfection. So I've had a whole journey with it.

是什么让你转变的？

What was it that flipped you?

埃隆·马斯克对他公司的人在危机情况下说过这样的话。他说，想象一下，你的身上绑着一枚炸弹，如果你不解决这个问题，它就会爆炸。那你该怎么办？大多数时候，当人们真正思考这个场景时，他们会集中注意力，振作起来，想出办法。很多时候初创公司就是这样。有太多生死攸关、压力巨大的时刻，你总是处于这些情况中，必须采取行动，否则就完蛋了，你只能找出最佳行动方案然后去做。所以，必须快速行动的现实久而久之重塑了我的大脑。

There's this thing that Elon says to people at his company when they're in a crisis situation. He says, imagine there was a bomb strapped to your body that will go off if you don't come up with a solution to this problem. Then what are you going to do? Most times when people actually think through that scenario, they focus and get their act together and figure out something to do. A lot of times startups are like that. There are so many moments that are so life and death and high pressure that you're in these situations all the time where you have to act and do something otherwise you're toast, and you just have to figure out the best plan of action and do it. So the realities of having to operate quickly over time remolded my brain.

有意思。你有兄弟吗？有兄弟姐妹吗？

Interesting. Do you have any brothers? Do you have any siblings?

有，我有两个哥哥。我大学辍学，而他们两个都有博士学位。大哥是经济学家，另一个哥哥是神经科学博士。他们都是聪明人。

Yeah, I have two brothers. Two older brothers. I dropped out of college and both my brothers have PhDs. My oldest brother is an economist. My other brother has a PhD in neuroscience. They're smart guys.

全家都是天才啊？

Whole lineage of geniuses, huh?

是啊，我想我父母可能还有点不高兴，因为我们谁都没成为物理学家。

Yeah, I think my parents are probably still a little miffed that none of us became physicists.

哦，天哪。不过我相信他们对现在的结果一定很满意。我是说，哇。

Oh, man. Well, I'm sure they got to be happy with how everything turned out. I mean, wow.

是啊，不，我觉得我父母非常为我骄傲。

Yeah, no, I think my parents are super proud of me.

你上的什么学校？是在家上学吗？

Where did you go to school? Were you homeschooled?

我上的是洛斯阿拉莫斯的公立中学和高中。那个镇大约有一万人。现在更多了，因为他们制造核芯。但我小时候，大约有一万到一万五千人。很小的镇子。有一所公立中学、一所公立高中、几所小学。我上的是公立学校。我很幸运；那些是很好的公立学校。但和其他公立学校一样。然后我每天回家，基本上每天都做数学和科学。

I went to Los Alamos public high school and Los Alamos public middle school. The town is about 10,000 people. Now it's more because they do manufacturing of nuclear cores. But when I was growing up, it was 10 to 15,000 people. Pretty small town. There's one public middle school, one public high school, a few elementary schools. I went to public school. I was lucky; those are amazing public schools. But it's public school like any other. Then I would get home every day and effectively do math and science every day.

普通二年级学生学什么？我是说，你说你二年级就学了代数。普通二年级学生学什么？我很久没上二年级了，但我很确定是基础加法。

What is the average second grader? I mean, you said you had learned algebra in second grade. What is an average second grader? It's been a long time since I've been in second grade, but I'm pretty sure it's basic addition.

是啊，我想是加法。也许学到乘法表。不确定。对，也许一些乘法表。

Yeah, I think it's addition. Maybe you get to your times tables. Not sure. Yeah, maybe some multiplication tables.

那么，你是怎么从前一天晚上学代数，到第二天学 2 加 2 等于 4 的？那是什么感觉？

So how do you go from studying algebra the night before to 2 plus 2 is four? What is that like?

我确实记得在学校里，就像很多孩子一样，基本上就是置身事外。你明白吗？就是走神、做白日梦，忽略课堂上发生的事情。

I definitely remember in school, like a lot of kids in general, just sort of buying out of the whole thing. Does that make sense? Just tuning out and daydreaming and ignoring what was happening in classes.

这确实开始发生了。我实际上会做的是回家后继续做数学题。我的意思是，你比老师还厉害。我记得有一次，我上的那所学校的好处是老师们真的很关心我的教育。我的很多老师都希望看到我茁壮成长并继续学习，这太棒了。我可以想象一个完全不同的学校，老师们不在乎，因为他们的生活一团糟，教室也一团糟，诸如此类。但我很幸运有真正关心我的老师。

That definitely started happening. What I would actually do or focus on is go back and then do math at home. I mean, you're more advanced than the teacher. I remember one time, the good thing about the school I went to is the teachers were really invested in my education. Many of my teachers wanted to see me thrive and continue learning, and that was awesome. I can imagine a totally separate school where the teachers don't care because their lives are chaotic, the classroom is chaotic, all that kind of stuff. But I was lucky to have teachers who really cared.

是啊，看起来结果不错。尽管你在 28 年里取得了这么多成功，但你是一个非常脚踏实地的人。我从来不知道和你们在一起会是什么样。吃早餐时，我印象非常深刻。我想，‘哇，这家伙真的很接地气。’而且看起来是个非常好的人。所以，太客气了。老兄，向你致敬。但是，嘿，我们快速休息一下。回来后，我们再聊 MIT。

Yeah. Seems like it worked out well. For all the success you have amassed in 28 years, you're a very grounded person. I never really know what I'm going to get with you guys. At breakfast, I was super impressed. I'm like, 'Wow, this guy's really grounded.' And seems like a really good person. So, too nice. Kudos to you, man. But hey, let's take a quick break. When we come back, we'll get into MIT.

你们听我提到 Patriot Mobile 已经有一段时间了。他们为那些相信信仰、家庭和自由值得为之奋斗的美国人挺身而出。他们是货真价实的。他们拥有尖端技术，切换起来很容易。保留你的号码，保留你的手机，或者升级。他们 100%的美国本土团队可以在几分钟内通过电话激活你。他们是少数能接入美国三大主要网络的运营商之一。这意味着出色的全国覆盖。他们甚至可以在手机上放一个不同网络的第二个号码。就像一部手机带两部手机。他们有无限数据套餐、移动热点、国际漫游、移动互联网设备和家庭互联网备份。今天就切换，体验不同。访问 patriotmobile.com/srs 或拨打 972 patatriot。现在，使用促销代码 SRS，注册即可获得一个月的免费服务。切换到 Patriot Mobile，用你的每一次通话和短信捍卫自由。网址是 patriotmobile.com/srs 或拨打 972 Patriot。

You've heard me talk about Patriot Mobile for a while now. They've stood in the gap for Americans who believe that faith, family, and freedom are worth fighting for. And they're the real deal. They've got cutting edge technology, and switching is easy. Keep your number, keep your phone, or upgrade. Their 100% US-based team can activate you in minutes right over the phone. They're one of the few carriers with access to all three major US networks. That means exceptional nationwide coverage. They can even put a second number on a different network on the phone. It's like carrying two phones in one. They have unlimited data plans, mobile hotspots, international roaming, internet on-the-go devices, and home internet backup. Make the switch today and experience the difference. Go to patriotmobile.com/srs or call 972 patatriot. And right now, use the promo code SRS for a free month of service when you sign up. Switch to Patriot Mobile and defend freedom with every call and text you make. That's patriotmobile.com/srs or call 972 Patriot.

夏天来了，如果你和我一样，冬天没有只是坐着不动。你保持敏锐，不断前进。现在是时候让你的装备跟上了。这就是为什么我想向你介绍 Roka。我一直在寻找能应对任何情况、兼具性能和风格的眼镜。让我告诉你，这些可不是普通的太阳镜。我在现实世界中测试过它们，从射击到钓鱼到越野，它们都经受住了考验。它们很轻，不会在脸上滑动，而且能承受撞击而不散架。最棒的是，它们看起来很好。干净现代，没有花哨的东西。只有不妥协的高性能优质眼镜。这是我尊重的一点，也是为什么每次出门，我都会拿起我的 Roka 太阳镜。Roka 总部在德克萨斯州奥斯汀。美国设计，不偷工减料。光学镜片清晰无比，能消除眩光，佩戴一整天都舒适。需要处方镜片？他们有太阳镜和光学镜两种选择。Roka 不仅有很棒的太阳镜，还有防蓝光眼镜。我每天晚上放松时都会戴，同时还要看手机、笔记本电脑或 iPad。它有助于你放松，准备入睡。他们是一站式眼镜店，能应对生活抛给你的一切。Roka 是货真价实的。准备好升级你的眼镜了吗？去 roka.com 看看，结账时使用代码 SRS 可享受全站 20%的折扣。网址是 roka.com。

Summer's here and if you're anything like me, you didn't spend the winter just sitting around. You stayed sharp and kept moving. And now it's time your gear caught up. And that's why I want to introduce you to Roka. I've been looking for eyewear that can handle any situation with performance and style. And let me tell you, these aren't your average shades. I've tested them in the real world from shooting to fishing to off-roading, and they hold up. They're lightweight, don't slide around on my face, and can take a hit without falling apart. And the best part, they look good. They're clean and modern. No frills here. Just premium eyewear that performs without compromise. That's something that I respect and that's also why every time I head out the door, I reach for my Roka shades. Roka is based in Austin, Texas. American designed, no cut corners. The optics are crystal clear, cut through glare, and the fit stays comfortable all day long. Need a prescription? They've got you covered with both sunglasses and eyeglasses. Not only does Roka have awesome shades, they also have these that protect you against blue light. I wear these every night when I'm winding down for the day and I still got to look at my phone or my laptop or my iPad. It just helps you wind down and get ready for bed. They are a one-stop shop for eyewear that's built to handle whatever life throws at you. Roka is the real deal. Ready to upgrade your eyewear? Check them out for yourself at roka.com and use code SRS for 20% off sitewide at checkout. That's roka.com.

好了，Alex，我们休息回来了。我们准备进入你上大学的部分。所以，你是在 MIT 开始的，对吗？

All right, Alex, we're back from the break. We're getting ready to move into you going to college. So, you started at MIT, correct?

是的。

Yep.

那怎么样？

How did that go?

嗯，让我想想。我先说说那之前的几年。实际上，我从高中辍学了。

Yeah, so let's see. I'll say the first few years before that. So, I dropped out of high school, actually.

哦，你从高中辍学了？

Oh, you dropped out of high school?

是的，我从高中辍学了。

Yeah, I dropped out of high school.

为什么？对你来说不够有挑战性吗？

Why? Wasn't challenging enough for you?

我提前一年辍学去 Quora 工作，这是一家科技公司。我想很多人都知道 Quora 是一个问答网站。但我去那家科技公司工作了一年。一年后，我决定是时候上大学了。所以我去了 MIT。

I dropped out a year early to go work at Quora, this tech company. I think a lot of people know Quora as the question-answer website. But I went to work at a tech company for a year. And after a year of that, I decided it was time to go to college. So I went to MIT.

15 岁，你就难倒了博士们？

15, you were stumping PhDs?

可能没那么早，但到 16、17 岁时，我已经更有能力了。

It was maybe not quite that early, but by 16 or 17, I was more competent by that point.

你在哪些方面难倒了他们？

What were you stumping these guys on?

那时是早期 AI。甚至还不叫 AI，更流行的术语是机器学习。是关于训练不同的算法来重新排序内容。都是社交媒体类东西的算法。比如，什么算法能产生最多的参与度，或者什么算法能让人们最沉迷于这些信息流。那就是我当时在做的事情。

At that point, it was early AI. It wasn't even called AI yet. It was called machine learning, the more popular term. It was about training different algorithms that would rerank content. It was all the algorithms for social media style things. Like, what algorithm creates the most engagement or what algorithm gets people most hooked on these feeds. That's what I was working on back then.

明白了。所以我工作了一段时间，然后去了 MIT。

Gotcha. So, I worked for a bit and then I went to MIT.

你 16、17 岁就难倒博士们，感觉如何？我是说，这对你来说就像正常生活吗？你有没有意识到‘天哪，我真的很聪明’？

What is it like for you to be 16, 17 years old stumping PhDs? I mean, is that just like normal life for you? Does it set in like 'holy, I'm really smart'?

我认为我很早就内化的一件事是专注非常关键。我不一定认为我本质上比很多其他人聪明得多，但我小时候极度专注于数学，然后极度专注于物理，高中时极度专注于编程。如果你极度专注，真正投入时间和精力，你可以取得非常快的进步。我长期以来相信的一件事是，如果你过度投入，比如你真的投入大量时间和精力，加倍努力，再加十倍努力，并且不断过度投入，那么你进步的速度会比其他人快很多倍。很多其他人可能只是没有加倍努力，或者没有这么专注，或者有点漫无目的。

I think something I internalized pretty early on was that focus was really critical. I didn't think necessarily that I'm way smarter fundamentally than a lot of these other people, but I was hyperfocused on math as a kid, then hyperfocused on physics, and then in high school I was hyperfocused on programming. If you're hyperfocused and you really invest the time and effort, you can make really fast progress. One of the things I've believed in for a long time is that if you overdo things, like you really invest lots of time and effort, go the extra mile, go the extra 10 miles, and constantly overdo things, then you will improve faster than anybody else by many times. A lot of other people maybe they're just not going the extra mile or maybe they're not as focused, or they're meandering a bit more.

所以对我来说，我觉得能取得这么多成就，很大程度上归功于专注和过度投入，付出额外的努力。我认为归根结底就是这样。

And so that's really like I definitely for me I think a lot of what I attribute being able to accomplish so much to is really about focus and overdoing it, going the extra mile. That's what I think boils down to.

你辍学时你父母怎么想？

What did your parents think when you dropped out of school?

你知道，我父母可能还是希望我读博、做科研。所以我觉得他们——我尊重这种信念——他们认为追求科学、追求知识高于一切。所以我总是跟他们说，这只是个小插曲，最终我会回来完成学位、拿到博士，走上正轨。我一直这么跟他们说，但后来这说法变得不可信了，我就不再说了。

You know, my parents I think still probably really want me to get a PhD and do scientific research. So I think they view, and I respect this belief, you know, I think they view the pursuit of science, the pursuit of knowledge as above all else. And so I would always tell them, hey, this is just a little detour, but ultimately I'm going to come back and finish my degree and get a PhD, and I'll be on the straight and narrow. So that's what I always told them, but then at some point it just wasn't believable, so I stopped telling them that.

你为什么决定去上学？

Why did you decide to go to school?

我上学有两个原因。一是真心想快速学习大量 AI 知识，我知道工作也能学，但最好的办法是去学校，把所有时间投入进去，快速学习。二是几乎所有人——不是所有人，但很多人——如果你问他们人生中最美好的时光，很多人会说是大学时代。所以我不想牺牲大学时光。于是我决定深入钻研 AI。我在 MIT 修了所有能修的 AI 课程。我只待了一年，但第一学期就选了最难的机器学习课。我的新生导师恰好是那门课的教授，她批准我所有课程。我选了课，她说：“你是新生，这门课对你来说太难了。”我说：“给我个机会，我对这个主题充满热情。”她说：“好吧，我们先让你上几周看看。”于是我进去了。我觉得压力很大，因为我想证明我能行。第一次考试来了，纯属运气，考的大部分内容我都理解得很好，我在几百人的班里得了最高分之一。之后教授就让我为所欲为了。于是我在 MIT 深入学习了所有 AI 课程。那一年 DeepMind 推出了 AlphaGo，第一个击败世界顶级围棋选手的 AI，这被认为是 AI 最难攻克的策略游戏。那是个大事。然后我开始自己捣鼓 AI。我想在冰箱里装个摄像头，告诉我室友是不是在偷我的食物。我开始捣鼓，很快意识到一切都会卡在数据上。无论你想让 AI 做什么，都依赖数据。我环顾四周，没人解决这个问题。很多人研究算法、芯片、算力，但没人研究数据。我那时 19 岁，很急躁，心想：“好吧，既然没人做，那我就自己做。”于是辍学，创办公司，开始狂奔。

I went to school for two reasons. One was genuinely I wanted to learn a lot about AI very quickly, and I knew I could kind of do that while working, but the best thing to do would be to go to school, invest all my time into it, and try to learn very quickly. And the second thing was that almost anyone, not anyone but many people, if you ask them what were the best years of your life, a lot of people will say their college years. So I wasn't going to sacrifice the college years. So I decided to go really deep into AI. I took all the AI courses I could at MIT. I was only there for a year, but I started out taking the hardest machine learning course my first semester. My freshman adviser, who approved all my courses, happened to be the professor of that course. I signed up for her course and she said, 'You're a freshman, this is going to be too much for you.' I said, 'Just give me a chance, I'm really passionate about the topic.' She said, 'Okay, we'll let you go for the first few weeks and see how you do.' So I got in. I felt the stakes were really high because I wanted to prove I could do it. The first test came around, and by sheer luck it happened to be mostly about things I understood well, and I got one of the top marks in a class of hundreds. After that, the professor let me do whatever I wanted. So I went really deep into AI coursework at MIT. That was the year DeepMind came out with AlphaGo, the first AI to beat the best Go players in the world, which was seen as the hardest strategy game for AIs. That was a big deal. Then I started tinkering with AI on my own. I wanted to build a camera inside my fridge that would tell me when my roommates were stealing my food. I started tinkering, and quickly realized that everything was going to be blocked on data. No matter what you wanted AI to do, it relied on data. I looked around and nobody was working on this problem. Plenty of people working on algorithms, chips, computational capacity, but nobody on data. I was impatient, 19 years old, and thought, 'Well, if nobody's going to do it, I might as well do it.' Dropped out, started the company, and was off to the races.

那你完善了那个冰箱 AI，能告诉你室友是否在偷你的食物吗？

So, did you perfect the refrigerator AI to tell you if your roommates are stealing your food?

我试着做，但发现数据远远不够。所以它总是误报，有假阳性和假阴性。然后我灵光一闪，意识到：“哦，如果真想做成这个，我需要的数据量是现在的百万倍。”而且任何 AI 应用都是如此。这就是这个想法的起源。

I was trying to build it, and then I realized I didn't have anywhere near enough data. So it always fired incorrectly, with false positives and false negatives. And then that was the light bulb moment. I realized, 'Oh, if I really want to make this, I need like a million times more data than I have now.' And that's going to be true for every AI thing anyone ever wants to build. So that was the genesis of the idea.

所以你离开了 MIT。

So you left MIT.

离开了 MIT。我直接从波士顿飞到旧金山创办公司。19 岁的我立刻开始在旧金山写代码。我参加了 Y Combinator 这个加速器项目，有点像创业界的饥饿游戏。夏天开始时有一百家初创公司，大家都在拼命工作，展示里程碑和进展。最后是演示日，每个人展示公司并争取投资。这简直就是饥饿游戏：你经历整个过程，最后如果拿到投资就赢了，否则就输了。这就是公司的开端。我们最终拿到了不错的投资。

Left MIT. I flew straight from Boston to San Francisco to start the company. I immediately went from being 19 years old to coding in San Francisco. I was part of this accelerator program called Y Combinator. It's kind of like the Hunger Games for startups. It starts with 100 startups at the beginning of the summer, and you're all grinding away, trying to show milestones and progress. It culminates at the end with a demo day where everyone presents their companies and tries to get investment. It literally is the Hunger Games: you go through this whole thing, and at the end if you get investment, you've won; if you didn't, you've lost. That was the beginning of the company. We ended up getting good investment.

你们做什么？

What did you do?

那时我们围绕 AI 的数据做文章。所以就是如何为人们想用 AI 构建的东西提供数据。但那时太早了，用例很蠢。比如我们帮一家 T 恤公司做检测之类的事。

At that time it was around data for AI. So it was all about how to fuel data for what people want to build with AI. But at that time it was so early that the use cases were pretty stupid. Like we were helping one company try to detect something for a t-shirt company.

他们做定制 T 恤设计，我们帮他们检测那些不适合印刷的设计，比如有血腥或非法内容。基本上就是识别非法 T 恤设计，现在听起来有点傻。然后我们帮一个家具市场用 AI 改进搜索算法。大概三个月后，我们开始和自动驾驶公司合作。这成了我们头三四年工作的核心。所以我们和通用汽车、丰田、Waymo 以及所有主要汽车制造商合作，帮他们造自动驾驶汽车。

They made custom t-shirt designs and we were helping them detect when people used a t-shirt design that was unfit to print, like it had gore or illegal stuff. Basically identifying illegal t-shirt designs. It sounds kind of stupid now. Then we helped a furniture marketplace improve their search algorithm with AI. A few months in, maybe 3 months, we started working with autonomous vehicle companies and self-driving companies. That ended up being the real meat behind our effort for the first 3-4 years. So we worked with General Motors, Toyota, Waymo, and all the major automakers in helping them build self-driving cars.

你们当时有多少竞争对手？

How many people were you competing against?

创业做什么都有几十个竞争对手。当时确实有几十个。这些领域竞争激烈，但我不介意竞争，毕竟从数学竞赛起就习惯了。所以我们专注于问题：如何为自动驾驶汽车获取最佳数据集？这很大程度上涉及传感器融合。有各种不同的传感器，如何组合它们得到一个输出？如果多个传感器感知到一个人，如何整合成一个人、一辆车、一辆自行车？那是我们的专长。然后我们就起飞了。公司发展到大约 100 人。

In anything you do in a startup, you have tens of competitors. There were definitely tens of competitors at that time. These are competitive spaces, but I don't mind competition from math competition days. So we were just really focused on the problem: how do you get the best possible datasets for these self-driving cars? A lot of that had to do with sensor fusion. There are so many different kinds of sensors, and how do you combine them to get one output? If multiple sensors sense a person, how do you collect that to say that's one person, one car, one bicycle? That was our specialty. Then we were off to the races. We grew the company to about 100 people.

我们往回说一点。你 19 岁从 MIT 辍学，独自去了旧金山。你当时还不成熟。你是怎么培养领导力的？怎么有知识和人脉来创办公司？

Let's go back a bit. You go to San Francisco by yourself as a 19-year-old who just dropped out of MIT. You're immature. How do you develop leadership skills? How do you have the knowhow and make connections to build a company?

早期关键是找谁投资。就我一个人参赛，没有团队。我每天写代码。然后我们拿到了 Y Combinator 的投资，还有 Excel——Facebook 的早期投资者。好的投资者帮我组建团队、招人。我主要招的是学校认识的人。因为他们能信任我。当时如果我去找旧金山 25 岁的工程师说‘我们一起干吧’，我毫无可信度。喝杯咖啡，他们就说‘好，我回去上班了’。所以早期我只对大学同学有可信度，我们是朋友。我招募了一批人。有些也辍学了，有些毕业了加入。那是早期的核心。然后我们开始和大型汽车公司及自动驾驶公司合作，势头起来了，团队也逐渐壮大。

Early on, it's about who you get investment from. It was just me with the competition, no team. I was coding every day. Then we got Y Combinator to invest, and then Excel, an early investor in Facebook. Good investors helped me build the team and find people to hire. I mostly hired people I knew from school. Because they could trust me. At the time, if I went to a 25-year-old engineer in San Francisco and said, 'Hey, we should work together,' I had no credibility. I'd get coffee and they'd say, 'Cool, I'm going back to my job.' So early on, I only had credibility with people I went to college with, who were friends. I managed to recruit a bunch of them. Some dropped out too, some finished school and joined. That was the early nucleus. Then we started picking up momentum working with large automotive companies and autonomous driving companies. As momentum picked up, we grew the team over time.

你的商业头脑从哪来？还是你雇了人管这些，你当主脑？

Where did you get your business sense? Or did you hire someone to run that while you were the mastermind?

大约一年后，我雇了一个人，头衔是业务主管。但在此之前，我只是自己学着做。

About a year in, I hired someone with the title head of business. But until then, I was just trying to learn it all.

你怎么把产品推出去的？

How did you get the product out there?

我全写好了代码，放到一个可以发布初创公司的网站上。它在 Twitter 上找新创业点子的人群中微传播了。那是让一切成长的早期种子。但过程很艰难。我整天写代码，偶尔发点东西到网上，然后求所有朋友点赞、给点热度。那就是早期。

I coded it all up and put it on one of those websites where you can launch startups. It went micro-viral among people on Twitter looking for new startup ideas. That was the early seed that enabled everything to grow. But it was tough going. I spent all my time coding, then occasionally posted something online and begged all my friends to upvote it, like it, give me some traction. That was the early days.

一开始就叫 Scale AI 吗？

Was it called Scale AI at the beginning?

一开始叫 Scale API，因为那个域名可用。大约一年半后改名为 Scale AI。早期创业真的很艰难。看看所有大公司早期是什么样，都很粗糙混乱。但最酷的是我们开始和汽车公司合作自动驾驶。这很快变得超级有趣，因为那是当时最伟大的科学和工程挑战之一。我们最终成功了：我们的客户之一 Waymo 现在已经在旧金山、洛杉矶、凤凰城等地推出大规模机器人出租车服务，还在扩展到更多城市。太棒了。

It was called Scale API at first because that website was available. It became Scale AI about a year and a half later. Early startups are so gnarly. If you look at all these big companies and think about what they were like in the early days, they were all pretty rough and tumble. But the coolest thing was we started working with automotive companies on self-driving. It quickly became hyper-interesting because it was one of the great scientific and engineering challenges of the time. We ultimately succeeded: Waymo, one of our customers, is now launched and driving large-scale robo-taxi services in San Francisco, LA, Phoenix, and more cities. It's pretty amazing.

公司发展有多快？五年后你成了世界上最年轻的亿万富翁。太疯狂了。

How fast did the company grow? Five years from starting, you became the youngest billionaire in the world. That's crazy.

是啊，当时并不觉得明显。

Yeah, that did not feel obvious.

第一年，头 12 个月，只有一到三个人。几乎没人。就我和另外一两个人做了一年。就这些。第二年，我们从一到三人开始招人，到了大概 15 人左右。第三年，从 15 人到了大概 100 人，然后就起飞了。100、200、500，一直增长。现在大概有 1000 人了。但一开始真的很慢。

The first year, for the first 12 months, it was like one to three people. It was almost nobody. It was me and one or two other people working on it for the first year. That's it. Then after the second year, we went from that one to three people and started hiring more. We got to maybe 15 or so people. Then that third year, we went from 15 to maybe 100, and then we were off. It was 100, then 200, then 500, and we kept growing. Now we're up to like 1000 people. But it was really slow going at first.

你们最初专注于什么？

What were you guys focused on initially?

先是自动驾驶，大约三年后，我们开始专注于国防，与国防部合作。

First it was autonomous driving, and then starting about three years in, we started focusing on defense and working with the DoD.

你们在国防领域做什么？

What are you guys doing in defense?

我们做几件事。最早之一是帮助国防部解决数据问题，以训练 AI 系统。他们想在卫星图像、SAR 图像和其他空中图像上做图像识别，但数据问题很大。就像我那个冰箱的例子，他们需要数据来检测图像中的目标。所以我们最初是为国防部提供数据集和数据能力。头几年都是这样。最近，我们开始与他们合作大规模部署 AI 能力。

We do a few things. One of the first things was helping the DoD with its data problem to train AI systems. They wanted to do image recognition on satellite imagery, SAR imagery, and other overhead imagery, but they had a huge data problem. Just like me with the fridge, they needed data to detect things in all that imagery. So the first thing we did was fuel the datasets and data capabilities for the DoD. That was true for the first few years. More recently, we've been working with them on large-scale fielding of AI capabilities.

国防部在图像中找什么？基本上，不需要人来检测核反应堆或导弹发射井之类的东西。AI 检测这些，减少人为错误和人力。更准确。我理解得对吗？

What kind of stuff is the DoD looking for in imagery? So basically, you don't need a human to detect something like a nuclear reactor or a missile silo. AI detects all that, reducing human error and manpower. It's more accurate. Am I on the right track?

对。而且可扩展。太空中的卫星数量激增，现在的感知数据远超人力处理能力。

Yeah. And it's scalable. The number of satellites in space has exploded, so we have much more sensing today, way more imagery than it's feasible for humans to work through.

你们怎么提供数据？

How do you fuel it?

分两部分。首先，要建一个数据工厂，一个生成大量数据来驱动算法的机制。很多是合成数据，用算法自己生成，但仍需人工验证。我们为此在密苏里州圣路易斯，国家地理空间情报局旁边建了一个设施，成立了 AI 数据处理中心，雇佣图像分析师验证 AI 系统的输出，确保准确、高完整性的数据反馈给 AI 系统。

There are two parts. First, you have to build a data foundry, a mechanism to generate lots of data to fuel the algorithms. A lot of it is synthetic, using the algorithms themselves to generate data, but you still need humans to validate and verify. One thing we did for this project was create a facility in St. Louis, Missouri, next to NGA, the National Geospatial Intelligence Agency. We produced a center for AI data processing where we hired imagery analysts to validate the outputs from the AI systems, ensuring accurate, high-integrity data to feed back into the AI systems.

然后我们开始与国防部合作更宏大、更大规模的 AI 项目。一个是名为 Thunderforge 的项目，用 AI 进行军事规划和作战规划。基本思路是用 AI 自动化军事规划的主要部分，从而在几小时内完成规划，而不是几天。

Then we started working with the DoD on more ambitious, larger-scale AI projects. One is a program called Thunderforge, which uses AI for military planning and operational planning. The basic idea is to use AI to automate major parts of the military planning process, so you can plan in hours instead of days.

这听起来像 Palantir。

This sounds like Palantir.

对，他们针对问题的不同部分，我们也是。最终我们合作得很好。这是更广泛概念的一部分，我们称之为智能体式战争：在战争中运用 AI 和 AI 智能体。思路是从人机回环到人机监督。不再是人工一步步传递工作流，而是 AI 智能体完成大部分工作，人类沿途检查验证。这是个大变化。当前模式中，每个步骤由拥有数十年单一领域经验的人类完成。而 AI 智能体拥有跨领域数千年的知识，任务速度快千倍。这加速了各个层面：感知与情报、作战规划、战术决策。核心是用 AI 智能体更快、更自适应，人类检查他们的工作。

Yeah, they target different parts of the problem, and we target different parts. Ultimately, we work together pretty well. This is part of a broader concept we call agentic warfare: the use of AI and AI agents in warfare. The idea is to go from humans in the loop to humans on the loop. Instead of workflows where a person does work, passes to the next person, and so on, AI agents do a lot of that work, and humans just check and verify along the way. It's a big change. In the current setup, you have individual humans with decades of single-domain experience doing each step. With AI agents, you have agents with thousands of years of knowledge across all domains, a thousand times faster at tasks. It's about accelerating every level: sensing and intel, operational planning, tactical decision-making. At its core, it's using AI agents to be faster and more adaptive, with humans checking their work.

说到任务规划，尤其是在战术环境中——我来自那个背景——能举个例子说明它如何加速吗？

When you talk about mission planning, especially in a tactical environment—I come from that background—can you give an example of how it speeds up the process?

当然。我们目前正与印太司令部和欧洲司令部合作，之后会更大范围部署。假设出现一个警报，有意外情况需要应对。比如，一艘意料之外的船出现了。这个警报流入一系列 AI 系统。第一步是感知。

Sure. We're working on this with INDOPACOM and EUCOM right now, and we'll deploy more broadly. Let's say there's an alert that pops up, something unexpected that we need to respond to. For example, a ship appears that we didn't expect. That alert flows into a bunch of AI systems. The first step is sensing.

所以，比如，我们审视所有传感能力，重新分析所有数据，弄清楚我们对那艘船了解多少。现在，一个人，比如分析师，会经历所有这些处理和分析工作。但理想情况下，你有 AI 智能体，它们可以查看所有历史传感器数据，发现雷达上出现了某个东西，卫星图像上也出现了某个东西，然后我们就能拼凑出这艘船的轨迹。好，你经历这个过程，试图理解发生了什么，然后找出可能的行动方案。一旦有了态势感知，针对这个特定场景有哪些行动方案？你可以让 AI 智能体直接提出行动方案。比如，在这个场景下，既然这艘船正在靠近，我们可以开火，也可以等待观察，或者重新部署以便更好地应对威胁。我们还可以重新定位一些卫星以获得更强的感知能力。有很多不同的行动方案可供选择。

So what like let's look through all of our sensing capabilities and let's like go reanalyze all of the data that we have and figure out how much do we know about that ship right so now a person would like an analyst would go through and like do all this you know all the ped and all the stuff to to be able to undergo this work but ideally you have AI agents that are just going they can look through all the historical sensor data they can figure out um oh actually there's like kind of a thing that showed up on this radar and there's kind thing that showed up on this satellite imagery and we can kind of like sketch together this like you know the trajectory of this of this ship. Okay. So you go through that process you try to understand what's going on and then you and then you go through and and figure out okay what are the what are the possible um courses of actions. So once you have situational awareness, then what are the courses of actions against this particular scenario and you can have an AI agent honestly just propose courses of actions. Um like hey in this scenario given this ship is is coming here you know we could fire at it. We could just wait to see what happens. We could reposition so that we're you know we're able to to um you know handle the threat better. you know, all sorts of we could we could reposition some satellites so we have greater sensing. You know, there's all sorts of different courses of actions um that we could take.

然后，一旦 AI 生成这些行动方案，它会通过模拟器运行每个方案。它会实时进行兵棋推演。没错，实时兵棋推演。它会通过模拟器说：“如果我们开火会发生什么？”比如，这是我们对红方力量的了解，这是我们对蓝方力量的了解。如果我们开火，这就是兵棋推演的结果。如果我们只是增强感知，红方可能对我们采取这些行动，这就是我们承担的风险。好处是，因为这一切都是自动化的，你可以运行这些兵棋推演和模拟一百万次。所以不仅仅是军事规划者像人类时间那样试图推演和规划。你可以运行一百万次模拟，因为你没有完美信息，没有完美知识。你需要根据情况的不确定性，找出所有可能的结果。哇。然后，你对每个行动方案运行一百万次不同的模拟。然后你可以直接给指挥官一份完整的简报和演示，基本上就是：这是我们考虑过的行动方案，这些是这些方案的可能结果。我们可以展示每个场景的模拟结果，比如代表性模拟，然后指挥官做出决定。哇。所以，这就是现状，这是正在做的事情，这些是可能的行动方案，这些是每个行动的后果，这是百分比。是的，没错。它能在几秒钟内输出结果。现在可能还需要几个小时，因为这些模型比未来慢得多，但相比人类，今天可能需要几天时间，这并非缺乏意愿、努力或能力，而是因为情况非常复杂，比如一艘船突然出现，有很多因素需要考虑。所以，真正的阶跃变化在于大幅加速态势感知，大幅加速理解不同行动方案、可能发生什么、后果是什么，并将这些呈现给指挥官。它会给出建议吗？

And then uh once the AI produces those course of actions, it'll run each of those different course of actions through a simulator. So it'll then run uh it war games at real time. Exactly. It'll war game at real time. And so then it'll run through a simulator and say, "Okay, what's going to happen if we fire at it?" like, you know, this is what we know about red forces. This is what we know about blue forces right now. Um, if we fire at it, this is like, you know, this is the war game of how that plays out. If we um just increase our sensing, like these are the things that that the red forces could do to us up. And like that's the risk that we take on. And um and then the benefit is because all this is automatic, you can run it these war games and these simulations a million times. So it's not just like one, you know, military planners just like trying to like war game and plan it out like you know in human time. It's like you could run a million simulations cuz you don't have perfect information. You don't have perfect knowledge. So you need to kind of figure out based on the uncertainties of the situation, what are all the potential outcomes that that pop out of that. Wow. And then so you run like a million different simulations of each of these different courses of action. And then you can give a commander direct like you just give them this whole like brief and presentation which is basically these are the courses of actions we considered. This is the this these are the likely outcomes in those courses of action. We can show you the simulated like outcome in each one of these scenarios. So we can like show you what it would look like in every one of those scenarios if it happened like representative um simulations and then the commander makes a call. Wow. So it's this is what it is. This is what it's doing. These are the possible courses of action. These are the consequences of each action. This is the percentage. Yeah. Exactly. And and it spits that out in what? A matter of seconds. Yeah. Now it takes a you know probably takes even now it probably takes a few hours cuz you know these models are a lot slower than they will be in the future but yeah I mean compare that to I mean depending on the situation like that could take you know that could take days for humans to do today like it's and and it's not from lack of will or effort or or capability. It's just it's a really complicated situation if a ship pops up out of nowhere like there's a lot of stuff you have to consider. Um, and so, uh, that's really the the the step change here is just like a, uh, like dramatically accelerating situational awareness, dramatically accelerating like an understanding of what the different course actions are, what could happen, what are the consequences, um, and surfacing that to commander. Does it make a recommendation?

嗯，这有点意思。我们在是否给出建议上反复权衡，因为最终我们不想让指挥官像梦游一样做决定。我们希望我们的军事指挥官——他们是世界上最好的人类——考虑所有这些行动方案的潜在后果，并基于这些后果做出决策。所以我认为我们要确保指挥官仍然运用自己的判断，而不是让他们更容易说“就按 AI 说的做”。有趣。哇。但接下来，想想会发生什么。这就变得非常诡异了。假设只有蓝方，只有美国拥有这种能力，那很好，我们会遥遥领先。但如果红方，比如中国、俄罗斯或其他对手也拥有这种能力呢？那么情况就变成了：我已经对整个局势进行了兵棋推演，他们也瞬间推演了全局。然后，蓝方和红方都知道对方拥有完美的兵棋推演场景。你选择哪条路？这就变成了一个非常复杂的、几乎是心理层面的情况，最终取决于我们的情报有多好。我们对那个指挥官的情报有多好？我们对他们的收集能力了解多少？我们对他们可能知道我们多少情报了解多少？反之亦然。这变得相当……所以，假设中国、俄罗斯等敌人拥有这种能力，我们也拥有这种能力，那么这就变成了我们现在处理的过程：谁的情报更好？只是速度更快，你更快地采取行动，敌人也更快地做同样的事。所以本质上和现在一样，只是更快。如果我们先开发出来，我们就能实现全球主导。我说得对吗？

Um, this is kind of an interesting thing. We we go back and forth if we want to make a recommendation because ultimately like we don't want um to just be like you know we don't want to let commanders kind of like sleepwalk if that makes sense. We want them to like, you know, our military commanders are the best humans in the world like considering all of the potential consequences of these different course of action and also considering you know um uh and and ultimately making a call based on those potential consequences. So I think we want to ensure that commanders are still exercising their judgment in these decisions versus just, you know, making it easier for them to just say, "Oh, go with what the AI says." Interesting. Wow. But this but then, okay, think about what happens next. So um and this is where stuff gets really freaky. So, let's say that um obviously in a world where just the blue force, just the United States has this capability, that's great. You know, we're going to we're going to be running circles around everyone else. Um but then what happens if the red force, you know, China, Russia, whomever also has that capability? Then you're in this situation where I've war gamed out the whole situation. you know, they've instantaneously wargamed out the whole situation and then it's like then then it I think I honestly think so then it's like we know and you know like blue forces, red forces, we both know that we both have like you know this perfectly war game scenarios. Which avenue do you pick? And then it becomes this really complicated almost like psychological you know kind of kind of situation which is like then it like all comes down to how good our intel is. So how good is our intel about that commander? How good is our intel about what their collection capabilities are? How good is our intel about you know what they likely know about us and vice versa. Um and it gets pretty so this is actually let's just so let's say China Russia our enemies have this capability we have this capability then it then it kind of becomes it's like the same process that we deal with now who has the better intel right it's just developing and in and you're going to a course of action quicker and the enem is doing the exact same thing quicker. So it's essentially it's the exact same thing that we're doing now but faster and so if we develop it first then we achieve basically global domination. Am I correct here?

我认为时机非常关键。如果我们比对手早一年获得这种能力，我们就能更快地做出反应。我常用的类比是：想象我们在下棋，你每走一步，我能走十步。我肯定会赢。这就是这种能力带来的不对称优势。一旦双方能力持平，就会变成你所说的那种基于情报和能力的对抗性冲突。我们如何阻止对手拥有这种 AI 系统？中国通过 DeepSeek 以及之后发布的模型已经证明，他们在 AI 领域会非常有竞争力。2024 年，中国的大语言模型 AI 公司与中国人民解放军之间签订了大约 80 份合同。这个数字在美国不是 80，要少得多。所以他们显然在非常迅速地将 AI 整合到国家安全和军事体系中。我认为目前我们无法现实地阻止他们拥有我描述的这种能力。

I think timing really matters here. If we get this capability a year ahead of adversaries, we'll be able to respond much faster. The analogy I often use is: imagine we were playing chess, but for every one move you take, I can take ten moves. I'm just going to win. That's the asymmetric advantage that comes out of this capability. Once it equalizes, it becomes this adversarial intel-based, capability-based conflict. How do we combat our adversaries from having this type of AI system? China has demonstrated with DeepSeek and models since then that they'll be very competitive on AI. In 2024, there were something like 80 contracts between large language model AI companies in China and the People's Liberation Army. That number is not 80 in the United States; it's way less. So they're clearly accelerating the integration of AI into their national security and military apparatus very quickly. I don't think at this point we can realistically stop them from having this capability I described.

那么接下来看下一层：情报。AI 如何影响情报？对抗性 AI 动态是怎样的？我们能用我们的 AI 破坏他们的 AI 吗？他们能用他们的 AI 破坏我们的吗？这实际上是 AI 对 AI 的战争。第一层分析是，这可能归结为我运行了多少个 AI 系统副本，而你运行了多少个。这变成了一个数量游戏。如果我有 10,000 个 AI 副本在运行，而你只有 100 个，我会轻松碾压你。假设你有 100 个 AI，我有 10,000 个。我会用一半的 AI——5,000 个——专门攻击你的 AI，寻找你信息架构和数据中心的漏洞。另外 5,000 个副本做我自己的军事规划。现在想想对手：你有 100 个 AI。如果你全部用于军事规划，你会被黑，因为你没有做任何网络防御。即使你全部用于网络防御，数量对比也很糟糕——100 个 AI 对 5,000 个。你很可能还是会被黑。所以数量非常重要。即使对手只有 2 倍的优势——我有 10,000 个副本，他们有 5,000 个——我也可以做同样的事：5,000 个副本专注于攻击他们的 AI，使其瘫痪或中毒，另一半做军事规划。对手就惨了，因为要妥善应对网络攻击，他们可能需要全部 5,000 个副本用于网络防御，这样就没有能力做军事规划了。

So then you go to the next layer down: Intel. How does AI impact intelligence? What is the adversarial AI dynamic? Can we use our AI to sabotage their AI? Can they use theirs to sabotage ours? It's AI-on-AI warfare effectively. The first-level analysis is that it probably boils down to how many copies of these AI systems I have running versus how many you have running. It turns into a numbers game. If I have 10,000 AI copies running and you only have 100, I'm going to run circles around you. Let's say you have 100 AIs, I have 10,000. I'll take half of my AIs—5,000—and focus them on hacking your AIs, looking for vulnerabilities in your information architecture and data centers. My other 5,000 copies will do military planning for myself. Now think about the adversary: you have 100 AIs. If you focus all on military planning, you'll get hacked because you're not doing any cyber defense. Even if you focus all on cyber defense, the numbers are bad—100 AIs versus 5,000 from me. You probably still get hacked. So numbers matter a lot. Even if the adversary has a 2x advantage—I have 10,000 copies, they have 5,000—I can do the same thing: 5,000 copies focus on hacking their AI to incapacitate or poison it, and the other half focus on military planning. The adversary is screwed because to properly deal with a cyber attack, they probably need all 5,000 copies on cyber defense, leaving no capacity for military planning.

哇。所以这真的变成了像指挥你的所有领域的力量来牵制或战胜敌人。你会对你的 AI 军队做同样的规划，你的资产分配。

Wow. So it really turns into this like commanding your forces across all domains to outmaneuver the enemy. You'll do the same kind of planning for your AI army, your allocation of assets.

没错。很多问题在于：我投入多少 AI 用于攻击和破坏对手？多少用于我自己的军事规划和兵棋推演？另一个关键部分是无人机，以及你分配多少 AI 用于战术层面的自主行动以完成任务目标。但我认为最终归结为谁拥有更多资源——大型数据中心、运行所有这些 AI 智能体的电力。以及由谁来决定将多少 AI 投入战术环境，多少用于网络安全？是人类还是另一层 AI 来输出你刚才说的那些？这是我们的情况，这是行动方案，这是后果。所以是不是一层又一层的 AI 在做所有这些模拟？

Exactly. A lot of it will be: how many am I dedicating towards hacking and sabotaging the opponent? How many towards my own military planning and wargaming? The other key component is drones and how many you allocate towards tactical mission-level autonomy to accomplish mission-level objectives. But I think it really boils down to who has more resources—large-scale data centers, power to run all these AI agents. And who makes the determination of how many AIs to put in tactical environment, how many to go after cybersecurity? Is that a human or another layer of AI that spits out exactly what you just said? This is our situation, here are courses of action, here are consequences. So is it just AI after AI after AI doing all these simulations?

是的，没错。你说得对。然后你有另一个 AI 来规划如何分配你的 AI 资源，根据你所了解的对手情况来应对。

Yeah, then yeah. No, you're exactly right. Then you have another AI that's planning out how to allocate your AI resources to deal with the adversary given what you know.

那么，哪些关键维度能让你获得优势？如果你的 AI 在某种程度上不同——对手很难确切知道你会如何行动，比如以不同的思维过程或推理方式实现战略突袭。另一个是实际资源数量的模糊性。如果我能让对手认为我的资源比实际少得多或比实际多得多，那将是战略突袭的关键要素。

So what are the key dimensions that would give you an edge? Well, if your AI is different somehow—it's hard for your adversary to know exactly how you would act, like strategic surprise in the form of a different thinking process or reasoning of the AI systems. The other one is ambiguity of how many resources you actually have. If I can make the adversary think I have way fewer or way more resources than I actually do, that'll be a critical element of strategic surprise.

AI 能否在知道自己被黑时发出警报？

Would an AI be able to alert if it knows it's been hacked?

是的，这是个好问题。目前来说，可能可以。

Yeah, this is a great question. Right now, probably yes.

但未来确实有可能有效地入侵系统或以某种方式投毒 AI 系统，并且这种活动相对难以追踪，因为你基本上会入侵那个 AI 系统。所以有两种方式。一种是投毒输入 AI 的数据。我不是入侵 AI 本身，只是投毒所有输入 AI 的数据，这样在未来的任何时刻，我都可以激活那个 AI，基本上无需主动入侵就能控制它。我之所以能做到，是因为我投毒了 AI 的数据，比如改变了决策过程。而最终的决策者——人类——不会意识到这一点。没错。所以数据投毒很可怕，这也是 DeepSeek 令人恐惧的原因之一。中国选择开源这个模型，美国很多大型企业都用了 DeepSeek，觉得模型好、免费，为什么不用？但 DeepSeek 本身可能已经被以某种方式破坏或投毒，中共或解放军知道一些我们不知道的激活方式或行为特征。所以 DeepSeek 很可怕。第一个领域就是数据投毒：能否投毒我们用来训练 AI 的数据，从而在你不察觉的情况下改变 AI 的行为，进而影响整个军事行动？这是其一。第二种是，如果你能足够快地完成整个操作，你入侵后销毁所有痕迹，用一个智能体清除入侵证据，在任何人警觉或通知之前。这可能更极端，但数据投毒在短期内更令人担忧。那么，如何防御？如果 AI 被入侵且你知道它被入侵了，那 AI 就完全没用了。但问题是我们仍然要依赖它做很多事情。所以最终还得靠人类思维。比如一艘船，你必须了解所有历史行动，这样 AI 才不会检测到你要用的战术，你必须做一些前所未有的事情来迷惑对手的 AI。你必须做出你都不知道是否有效的剧烈改变，这样 AI 才不会检测到“我们以前见过这个”。所以战略突袭很快成为关键。如何策划行动以最大化对敌方 AI 的战略突袭？这是其一。第二，很多问题实际上归结于你有多少副本在运行、数据中心有多大、工业产能有多强，能否在中央和边缘的所有战区、所有环境中运行这些 AI。

Uh but the you it's definitely possible in the future that you will be able to effectively hack into a system or somehow poison an AI system and uh have that activity be relatively untraceable because you would basically um you would you would hack into that AI system. So there's two ways you would do it. One is you poison the data that goes into that AI. So I'm not hacking into the AI itself. I'm just poisoning all the data that's feeding into that AI such that at any moment in the future I like I can activate that AI and basically hack it without any sort of active intrusion. But I can just do it because I've poisoned I've like poisoned the AI that go the data that goes into the AI such that if I like you know say it alters the decision-m process. Yeah. Exactly. But the but the the end decision maker which would be a human would not realize that. Yeah. Exactly. Okay. So so data poisoning is going to is but this is what's so terrifying about deepseek. One of the reasons why deepseek is really scary is um uh you know China chose to open source the model right so there's a lot of corporates large scale corporates in the United States that have chosen to use deepseeek because they're like oh it's a good model and it's a good AI and it's free why not use it um but deepseeek itself as a model could already be compromised could already be poisoned in some way such that, you know, there are characteristics or behavior or ways to activate deepseek that this the CCP and the PLA know about that um that we don't. Uh so so that's why deepseek is scary and why so so the first area is just data poisoning. So basically, can you poison the data that we're using to train the AIS such that to your point, I've altered the behavior of your AIS in a way that you don't know about and that's going to affect that's going to have cascading effects across your whole military operation. That's one. And then the second one is um uh is basically uh you know if if you're able to do the whole operation quickly enough you basically hack in and you uh kind of as we were talking about before you would like destroy the traces. You destroyed any sort of trace that like you had hacked in and you have an agent that like hacked in like removed that trace and the evidence of you hacking in. um uh before anybody before it was alerted or notified. That's maybe a bit more extreme, but definitely the data poisoning stuff is is more concerning in the near term. Damn. So, how would you how would you defeat it? I mean, it it's so if if it were to be hacked and you knew it was hacked, then AI becomes completely irrelevant. Correct. Well, the issue is we're still going to rely on it for lots of things. So, um, it would it would have to come down to the human mind again and you would have to you would have to, let's say it's a ship. You would have to know everything that you've done in the history so that it it doesn't detect what tactic you're going to use and do something just something that's never been seen before in order to confuse the adversar's AI. Correct. Yeah. So, you have to make a drastic change that you don't know know if it's actually going to work so that the AI doesn't detect, oh we've seen this before. this is what it's about to do. Yeah. Yeah. So, so to your point, yeah, strategic surprise becomes the name of the game very quickly. Um, and and how do you create an operation such that you maximize the amount of strategic surprise against an adversarial AI? That's one. And then honestly the second thing that's that's really critical is a lot of this will just plain up boil down to like straight up boil down to how many copies you have running and how large your data centers are and um how much industrial capacity you have to run these AIS both centrally and at the edge in all the war in all the theaters in all the the um in every in every environment.

它学习新技术的速度有多快？比如 Seronic 正在制造自主水面作战舰艇，Palmer Lucky 在做自主潜艇。假设我们与中国开战，中国拥有从二战以来我们所有能力的历史数据。当像 Seronic 的自主车辆或 Palmer 的火箭或潜艇这样的新事物引入战场时，AI 如何获取数据集来做出决策或提出行动方案、后果、概率？它学习新事物的速度有多快？

um how fast will it learn new technology? So, let's just take for example Seronic. They're making autonomous surface warfare vehicles or Palmer Lucky, you know, he's doing the autonomous submarines and and so when when am I trying to say here? So, let's say we're at war with China. China has all the data, all the history back from whatever World War II on different capabilities that we have. And what happens when a new when something new is introduced onto the battle space like Seronics autonomous vehicles or Eperus or uh or Palmer's rockets or his submarines? How how would the how would the AI get the data set to make a decision or or not make decisions but come up with what you're talking about courses of actions consequences what it's about to do p you know probability of what's going to happen how how fast will it be able to learn when something new is introduced onto the battle space?

这是个好问题。一般来说，第一次看到全新的东西，比如 USV 或 UUV，它无法预测会发生什么，因为它不知道速度、弹药、射程等关键信息，除非他们有很好的情报，已经通过入侵知道了。假设他们不知道。那么最初几次冲突，它无法搞清楚情况。这正是战略突袭的关键：总是拥有敌方兵棋推演技术无法模拟的新平台。这肯定是其中一部分。但到了某个点，它会知道硬件的性能，并能运行模拟来理解如何改变战局。最终，你会运行大规模模拟，它会发现这个新的无人水面舰艇有这些航程、速度、机动方式、弹药、连接性，容易受到这些电子战攻击，可以被这样干扰。这些都会成为模拟的参数。但最初它不会有任何建议，你会拥有战略突袭。所以武器能力上的 OBSAC（观察、判断、决策、行动）仍然至关重要。最终是否总是回到人类思维？是的，我相信如此。

Yeah, this is this is a great question. In general, so the so like the first time it sees a a totally new, let's say a USV or UUV or whatever it might be that that it's never seen before, um it won't be a, you know, it won't be able to predict what's going to happen like cuz, you know, it won't know how fast it's going to go. It won't know what what, you know, what um uh what munitions it has. It won't know what its range is. It It won't know all the key uh the key facts unless, by the way, they have really good intel and they already know all those things because they've hacked us. But um let's assume they don't know. So the first few conflicts, it's not really going to be able to to figure out what's happening. And that that's a that's a key component of strategic surprise is always having new platforms that won't be sort of simulatable, let's say, by enemy wargaming tech. Um, so that's that's definitely part of it. Um, but at a certain point it's going to know what the hardware are capable of and it's going to be able to run the simulations to to understand how that changes the calculus. Um because ultimately right what's going to happen is and some of this stuff like you know this is this is like you know some of the stuff is dissonant because obviously if you look at what happens today in the military it looks nothing like this but let's play the play the tape forward and like see what happens in the future. Ultimately, you're going to run large scale simulations and it's going to figure out, hey, this new, you know, uh, uh, unmanned surface vehicle has this much range. It can go this quickly. It can maneuver in this way. It has this kind of munitions. Um, it has this kind of connectivity. Uh, it is vulnerable to these kinds of, you know, EW attacks, whatever they may be. Um, it can be jammed in these ways. And those will all just be parameters for the simulation um to run. So I think but initially it would have no recommendations. Initially you'd have strategic surprise. So OBSAC when it comes to weapons capabilities is still just paramount and it will I mean will it always come back to the human mind? Uh yeah I believe so.

我相信我们经常讨论的一个概念是“人类主权”。AI 系统会变得更好，但我们如何确保人类保持主权？如何确保人类对重要事务保持真正的控制？比如控制我们的政治体系、军队、经济体系、主要产业等等。我认为这在军事领域至关重要。你不会想采取简单的做法，比如我们绝不会让 AI 拥有单方面发射核武器的能力。我们永远不会那样做。所以，真正关键的是信息的聚合、模拟、兵棋推演和规划，最终由人类做出正确的决策。而且，这很大程度上会渗透到需要做出的外交决策中，也会渗透到经济战中。我甚至认为这会延伸到国家间的关系建设。比如，如果我们与俄罗斯结盟，结果会怎样？有哪些行动方案？后果是什么？这渗透到一切：政治、盟友、对手、战争、经济，所有的一切。

I believe that you know we have this concept that we talk about a lot which is human sovereignty. So AI systems are going to get way better but how do we ensure that humans remain sovereign? How do we ensure that humans maintain real control over what matters? So maintain control over our political systems, maintain control over our militaries, maintain control over our economic systems, you know our major industries, all that kind of stuff. And so I believe it's pretty paramount in the military. You are not going to want to take certainly just as a simplistic thing. We're not going to give AI the capabilities to unilaterally fire nuclear weapons. Like we're never going to do that. And so ultimately so much of what is going to become really critical is the aggregation of information, simulations, wargaming, planning to humans to ultimately make the proper decisions. And by the way, so much of this will start bleeding into diplomatic decisions that need to be made. It'll bleed into economic warfare. I mean this goes all the way into relationship building between nations. Should we, you know, what are the outcomes if we become allies with Russia? What are the courses of action? What are the consequences? It bleeds into everything: politics, allies, adversaries, warfare, economics, all of it.

是的，完全同意。因为如果你最终归结起来，能力是什么？能力是感知和态势感知。我将能够处理海量数据，包括开源情报和其他各种情报源，了解当前状态、正在发生的事情、当前局势。它将能够聚合所有这些数据，提供对这些行为的全面视图，然后让你能够预测，并让你能够有效地推演你可能采取的每一个潜在行动，以及这些场景下会发生什么，并给出一些概率性的看法。然后，你会将其用于每一个重大决策。军队和政府应该将其用于我们做出的每一个重大决策。我们应该用于贸易政策、外交关系，以及对外方面，但老实说，我们也应该用于内部政策，比如医疗政策等等。所以，这种全领域感知加规划的能力将至关重要。

Yeah, totally. Because if you ultimately boil it down, what is the capability? The capability is sensing and situational awareness. So I'm going to be able to go through troves and troves of data, OSINT, other forms of open source intel, different kinds of various Intel feeds that I have and know what is the current status, what's going on, what is the current situation. It'll be able to aggregate all that data to provide a comprehensive view as to what those behaviors are and then it'll give you the ability to predict and it'll give you the ability to effectively play forward every potential action you could take, what would happen in those scenarios with some probabilistic view, some probabilities. And then yeah, you're going to use that for every major decision. The military and the government should use this for every major decision we make. We should do it for trade policies, diplomatic relations, and looking outwards but honestly we should also do it for internal policies like healthcare policies, all that kind of stuff too. So this capability of effectively all domain sensing plus planning is going to be paramount.

你是否看到一个世界，AI 变得如此强大，以至于它变得过时，我们又回到了 10 年或 20 年前，一切都由人类决策？它会超越自身吗？

Do you see a world where AI becomes so powerful throughout the world that it becomes obsolete and we're right back to where we were 10 years ago, 20 years ago where it's all human decision-making? Will it outdo itself?

我有几点想法。我认为将要发生的第一阶段就是我所说的：从人在回路中到人在回路上。现在，人类在经济和战争等各个领域做了大量 brute force 的人力工作。那是即将发生的第一层主要自动化。然后，就涉及到你的战略决策能力以及做出高判断决策的能力，这些决策要考虑长期、短期、中期等等。在某个时刻，随着 AI 不断改进，它将以前进的速度运行，人类很难跟上。这将首先发生在研发领域：AI 将能够进行大量科学研究、新武器系统的研发、新军事平台的研发等，速度远快于人类。然后人类只需检查他们的工作并做出决定。所以它会越来越快。然后会发生的是，它将给人类做出的少数决策带来巨大的权重。所以任何决策，比如极端情况下，总统或任何人决定是否让我的 AI 与另一个国家的 AI 合作？那将是一个具有巨大后果的决策，比今天类似的决策后果高得多。所以我认为，正如你所说，随着它加速，我们最终会到达一个点：一切都归结为人类决策，但这些决策将承载千倍以上的后果。

A few thoughts here. I think the first stage of what's going to happen is kind of what I'm saying: human in the loop to human on the loop. We're going to right now humans do a lot of brute force manpower work in all sorts of different places in the economy and in warfare, etc. That's the first level of major automation that's going to take place. So then it's about your strategic decision-making and your ability to make high judgment decisions that consider long-term, short-term, medium-term, all that kind of stuff. At a certain point, as the AI continues to improve and improve, it will operate at a pace that is very difficult for humans to keep up with. This will start happening in R&D first: AI will be able to start doing lots of scientific research, lots of R&D into new weapon systems, lots of R&D into new military platforms, etc., much faster than humans would be able to do. And then humans will just check over their work and decide. So it's going to sort of race faster and faster. And then what happens is it'll create dramatically more weight on the few decisions that humans make. So any decision, like all the way to the extreme, is the president or whomever making decisions about do I let my AI collaborate with another country's AI? That'll be a decision of dramatic consequence, much higher consequence than similar decisions today. So I think it almost to your point, as it accelerates, we'll end up at a place where you're right: it all boils down to human decision-making, but those decisions will carry a thousand times more consequence.

你如何决定与谁合作？我的意思是这是一家国际公司。你们和谁合作？

How do you decide who you're going to work with? I mean it's an international company. Who all are you working with?

嗯，首先，我们对合作对象非常挑剔，因为我们资源有限，构建这些系统和数据集非常复杂，正如我们讨论过的。所以我们的目标通常是如何与每个行业中最优秀的合作？如何与排名第一的银行、排名第一的制药公司、排名第一的电信公司、排名第一的军队等合作。我唯一想补充的是，当我们展望未来并讨论这一切时，我们认为重要的是：尽可能让世界运行在美国的 AI 堆栈上，而不是 CCP 的 AI 堆栈上。这变得非常重要。这不仅关乎意识形态、宣传和控制之类的东西，而且从纯粹的操作层面来看，我们也希望拥有尽可能扩展的 AI 能力。

Well, first thing is we're pretty picky about who we work with ultimately just because we only have so many resources and building these systems and building these datasets is pretty involved as we've discussed. So our aim generally is how do you work with the best in every industry? How do you work with the number one bank, the number one pharma, number one telco, number one military, etc. The only addition to this that I would say we viewed as important is how are we, as we play the tape forward and everything we're just discussing. It's really important that as much of the world runs on an American AI stack versus a CCP AI stack. That becomes really important. And it matters not only for ideology and propaganda and control and all that kind of stuff, but it also really matters just for a pure operational level: we're going to want to be able to have as extended AI capabilities as possible.

所以，我的理解是，你们与 X 国合作。你们给 X 国 AI 模型，让他们用于任何他们正在做的事情，比如战争。我们拥有模型，但他们必须接入美国的数据中心。我理解得对吗？所以，只要我们控制着为那个 AI 模型提供数据的数据中心，我们就基本上拥有它，而 X 国只能相信 Scale AI 会考虑他们的最佳利益。

So, the way I understand this is you're working with X country. You give country X the AI model to utilize for whatever they're doing, let's just say warfare. We own, but they have to tap into a US-based data center. Am I correct here? And so, as long as we control the data center that's feeding that AI model, we essentially own it and country X just has to trust that Scale AI has their best interest.

是的。这是下一个层次。

Yeah. It's next level.

如果他们改变立场，比如某个国家现在与中国结盟，决定不再与美国站在一起，那我们就把 AI 或者给 AI 提供数据的东西抽走，或者操纵那些数据，让它基本上被黑掉。我说得对吗？这就是我们保护自己的方式。

And if they change, let's say country X now forms an alliance with China, they decide they don't want to be a part of America, then we just yank the AI or the data that feeds that AI or manipulate that data to where it's essentially been hacked. Am I correct? And that's how we keep ourselves safe.

是的。另外，我认为至少我们今天是这样想的，很多人也这么认为：数据中心可以建在其他国家，只要它是美国拥有和运营的，因为这样我们在任何情况下都还能控制。我唯一想补充的是，我们最初更关注低风险的 AI 应用。比如，你能用 AI 帮助某个国家的教育行业、医疗行业，或者协助审批流程吗？我认为低风险用例最初更重要。但我确实认为我们有“地缘政治摇摆国”这个概念。现在世界上有一些国家，它们最终是站在美国还是中国一边，将对潜在冲突的形态产生巨大影响，甚至对长期的冷战格局也有影响。所以我把 AI 视为外交和国际博弈中一个关键元素，具有长期战略影响。

Yes. And then with the addition, I think the way that at least we think about it today and I think a lot of people think about it today is like it's okay for the data center to be located elsewhere, located in the country, as long as it's US owned and operated, because then we still have control in any sort of scenario that happens. And the only other thing I would say is we're much more focused initially on just low stakes uses of AI. So can you use AI to help the education industry in one of these countries or can you use it to help the healthcare industry or can you use it to aid in permitting processes? I think low stakes use cases matter a lot more initially. But I really do think we have this concept of geopolitical swing states. There are a number of countries right now in the world where whether they side with the US or China over time is going to have immense consequences for what a potential conflict scenario looks like, but also even what the long-term cold war scenario looks like. So I view AI as one of these key elements of diplomacy and long-term strategic impact in the international war game.

AI 如何被应用到我们的政府中？我不太记得你具体怎么说的了，用来运行我们的政治领域。那会是什么样子？因为那很大程度上关乎人们的价值观、信仰和立场。如今，这个国家可能比以往任何时候都更加两极分化。那么，当政府如此分裂、存在这么多不同意识形态时，你怎么让 AI 模型来运行政府？AI 模型怎么能做到？

How would AI be implemented into our government? I mean, I can't remember exactly what you said, implemented to run our political sphere. What does that look like? Because so much of that is people's values and what people believe in and stand for. Today, the country is probably more polarized than it's ever been. So how do you get an AI model to run government when it is this polarized and there's so many different ideologies? How would an AI model run that?

我们有“智能体式战争”、“智能体式政府”这个概念。所以，你能不能把政府中那些非常低效的流程，用 AI 相关功能替换掉，从而提升效率和改善结果？

We have this concept of agentic warfare, agentic government. So can you take these very inefficient processes in government and start replacing those with AI related functions so that you're just improving efficiency and improving outcomes?

给我一个具体的例子。

Give me a specific example.

一个非常简单的例子。目前，退伍军人在退伍军人事务部看医生的平均时间大约是 22 天。这太长了，部分原因是大量过时的流程和工作流。那个系统根本不行。那么，能不能用 AI 智能体来自动化部分流程，自动获取所有需要的审批和信息，让 22 天变成一两天？这毫无疑问是政府效率的纯收益。另一个大问题是审批流程。如果我想在某处建个新数据中心或者翻新房子，审批流程根据地点不同，可能真的需要好几年。部分原因是有太多不同的审批和工作流。如果我们把系统规则编码，让 AI 智能体自动走完审批流程，这样一天内就能拿到许可或被拒绝？类似的事情可以放大一百万倍。DOGE（政府效率部）发现的一件事是，退休档案存放在矿山里，铁山矿，所有联邦雇员的退休档案都是纸质版。我们能不能把这些落后两代的技术，用 AI 从落后两代直接跳到领先两代？能不能尽可能多地自动化这些流程？在让政府服务和流程更高效方面，有太多唾手可得的机会。我还没见过谁不这么认为。这些都是第一层次的事情，改善政府运作。

One super simple one. Right now, the average time it takes for a veteran to see a doctor in the VA is something like 22 days. It's way too long, and part of that is because of a host of antiquated processes and workflows. That system's not working. So can you use AI agents to automate some parts of that process, automatically get whatever approvals need to be gotten, get whatever information needs to be gotten, such that that 22 days becomes a day or two? That is a no-brainer, pure win for government efficiency. Another big one is permitting processes. If I want to build a new data center somewhere or remodel my home, the permitting processes depending where you are could take literally years. Part of that is there are so many different approvals and workflows. What if we codified the rules of the system and had an AI agent automatically go through that permitting process so you could get the permit or get it denied within a day? And just that times a million. One of the things from DOGE they found is that the retirements are stored in the mine, Iron Mountain mine, literally an iron mine, paper copies of retirements for all federal employees. Can we just take that, which is two generations behind in terms of tech, and use AI to go from two generations behind to two generations forward? Can we automate as much of those processes as possible? There's so much low hanging fruit in terms of making current government services and processes way more efficient. I haven't met anybody who doesn't think this is the case. That's all the level one stuff, improving how our government operates.

它最终会取代政客吗？

Would it eventually replace politicians?

这是个好问题。我认为最终，首先退一步说，政策、政策制定和立法的速度，以及政府对新技术反应的速度，肯定需要加快。我在华盛顿花了很多时间，试图确保我们国家能出台正确的 AI 立法和监管，让一切顺利。这已经努力了好几年。我们国家还没有真正搞清楚。什么是正确的 AI 监管框架？仍然悬而未决。

That's a good question. I think ultimately, first off, taking a step back, it's definitely the case that policy, the speed of policymaking and legislation, and the speed at which the government reacts to new technologies, that's going to have to speed up. I've spent a lot of time in DC trying to make sure that as a country we get the right kind of AI legislation and regulation to ensure this all goes well. It's been years of trying to get that done. We still haven't really figured that out as a country. What is the right AI regulatory framework? It's still undecided.

你怎么跟华盛顿那些老古董解释这些东西？我们有人在镜头前中风，有人真的死在办公室里。那里有些人可能连怎么打开邮件都不会。然后你来了，28 岁，创立了 Scale AI。我是说，回想扎克伯格在国会作证的时候。我不同意他做的所有事，但看着那个场景，我就想：你们这些在华盛顿的人，可能连自己的邮件都不会打开，却在对一个技术天才说话，他还要努力把东西讲得简单让你们理解。我跟你待一天才能勉强搞懂，而他们还有五千万件其他事要处理。他们对技术根本不了解。你怎么开始呢？

How do you even describe this stuff to the dinosaurs that are still sitting in DC? We've got people stroking out on camera, people literally dying in office. We got people up there that probably can't even figure out how to open an email. And then you come in, 28 years old, built Scale AI. I mean, going all the way back to when Zuckerberg was sitting there talking to Congress. I don't agree with everything he did, but I look at that and I'm like, you guys have been sitting in DC, probably don't even know how to open your own email, and you're talking to a tech genius who's trying to dumb this down and make you understand. I get one day with you to try to wrap my head around this, and they have 50 million other things they're dealing with. They're not up to speed on tech. How do you even begin to...

是啊。

Yeah.

接入？我是说，我认为很多人，圈内人都明白，很多具体决策其实是由幕僚做出的，对吧？而且我认为，一般来说，作为幕僚你必须非常能干，不管怎样，因为这是一份非常混乱的工作。事情很多，他们必须快速决策。另一件事是，我觉得类比很有用。我认为今天活着的每个人都看到了技术进步的节奏越来越快。我觉得你很难找到不相信 AI 会是改变世界的技术的人。至于它具体会如何改变世界，我觉得那更模糊，但它会是改变世界的技术。但问题是，政治体系反应不够快，对吧？那会很危险。我的意思是，我们需要对这些新技术做出非常快速的响应。所以我认为这会越来越明显。随着 AI 和其他技术加速，世界会变化得非常快，坦白说，我认为选民会要求更快的行动。所以我认为我们的政府需要加速，但这就是必须发生的事情。

Tap in? I mean, I think a lot of it... I think the first thing, and I think this is like a lot of people in the know understand this, like a lot of the minute decisions really end up being made by staffers, right? And I think like generally speaking, as a staffer you have to be extremely competent, no matter what, because it's a very chaotic job. There's a lot going on and they have to make very fast decisions. The other thing is I think analogies are pretty helpful. Like I think everybody alive today has seen the pace of technology progress just increase and increase and increase. Like I think you'd be hard-pressed to find anyone who doesn't believe that AI will be this world-changing technology. Now exactly how it'll change the world, I think that's where it gets fuzzier, but it will be world-changing technology. But the issue is the political system just doesn't respond very quickly, right? And that's going to be very harmful. I mean, we need to be able to respond very quickly to these new technologies. And so I think they'll become more and more obvious. Like I think as AI and other technologies accelerate, it'll be very obvious that the world will just change so quickly, and frankly I think voters are going to demand faster action. And so I think our government is set up to accelerate, but that's what needs to happen.

我们怎么给这一切供电？我是说，这是个很大的讨论，你知道，每个人似乎都对核能很犹豫。电网极其过时。我是说，我们大约 30 分钟前刚看到灯光闪烁。停电一直在发生。刚刚就有一场大停电。整个西班牙、葡萄牙、意大利。我是说，美国也一直在停电。我们怎么给这些东西供电？你希望看到什么？

How do we power all this? I mean, that's a big discussion, you know, and everybody seems so apprehensive to go nuclear. The grid is extremely outdated. I mean, we just saw the light flicker here about 30 minutes ago. Power outages happening all the time. There was just a big one. All of Spain, Portugal, Italy. I mean, it's happening all the time in the US. Power outages. How are we going to be able to power all this stuff? What would you like to see happen?

是的。我是说，首先，如果你看中国过去 20 年总发电能力的图表，对比美国过去 20 年的总发电能力，中国的图表是直线上升的。他们正在疯狂地增加电力。我认为他们在过去十年翻了一番。过去十年发电能力翻了一番。而美国基本持平。只增长了一点点。这就是现在的情况。中国大约每十年翻一番。美国基本持平。而我们需要，仅仅为了给 AI 公司知道他们想建的数据中心供电，我们就需要将能源容量翻倍，而且这需要非常非常快地发生，几乎立即。所以你必须相信我们的图表会从完全平坦变成垂直，比中国的能源增长更垂直。而与此同时，中国正在非常快速地增长。他们会加速。他们会给电网增加更多电力。我认为很难想象在没有激烈行动的情况下，美国能够比中国更快地增长其能源容量。

Yeah. I mean, first of all, if you look at a graph of China's total power capacity over the past 20 years versus US total power capacity over the past 20 years, the China graph is like straight up and to the right. They're just adding crazy amounts of power. They've doubled it in the last decade, I think. Doubled their power capacity in the last decade. And the United States is basically flat. It's grown a little bit. And so that's what's happening right now. China's doubling every decade or so. US is basically flat. And we're looking at, to just power the data centers that AI companies know they want to build, we're going to need something like a doubling of our energy capacity, and that needs to happen very, very quickly, almost immediately. And so you have to believe that our graph is going to go from totally flat to vertical, faster vertical than China's energy growth. And China in the meantime is just growing perfectly quickly. They'll accelerate. They'll add more power to their grid. I think it's very hard to imagine realistic scenarios where, without drastic action, the United States is able to grow its energy capacity faster than China.

所以如果中国直线上升而我们持平，你是说中国已经超过了我们的电力能力，还是我们仍然在他们之上，尽管他们在上升？

So if China's going straight up and we're flatlined, I mean, does that mean are you saying that China has surpassed our power capabilities or are we still above them even though they're on the rise?

他们肯定在我们之上，因为他们人口更多，工业更多。所以他们的总电力比我们多。发电能力更强。顺便说一句，原因并不复杂。如果你分解中国的电力来源，是因为煤炭占了大约 80%。是的，全是煤。就是大量的煤。而在美国，可再生能源增长了很多，但总体数字持平的原因是我们用可再生能源替代了煤炭、天然气等化石燃料。所以净结果，美国是持平的。而中国是直线上升。所以第一件事：我们需要激烈行动。政府有国家能源主导委员会。我们和他们谈过几次。我们必须采取激烈行动，至少开始匹配他们增加电网电力的速度，理想情况下超过它。这是第一件事。第二件事你提到的是我们的电网极其过时，这是一个重大的战略风险。我不知道西班牙停电的原因或来源，但有些人认为是外国行为或某种网络攻击。我向你保证，美国能源电网极易受到大规模网络攻击。这些网络攻击的复杂性有时非常愚蠢。就像如果你找到正确的发电厂登录终端，有时人们不改变默认的用户名和密码，就是 username 和 password。所以你可以找到怀俄明州某个仍然使用默认用户名和密码的电站。你登录进去，就可以关闭整个地区的电力。所以我们的电网，仅仅因为其过时和分散，就极易受到网络攻击，极易受到外国行动的影响。这现在很重要。如果你切断一个大城市的能源电网，人们会死。所以现在就很糟糕，但让我们回到我们刚才讨论的 AI。假设我们和中国进行大规模的 AI 对 AI 战争。他们直接切断电网，切断我们的数据中心和供电，那我们就成了坐以待毙的靶子。不仅如此，据我所知，中国实际上生产制造了我们电网中的许多主要部件，比如变压器。我们甚至不检查那些是否有恶意软件、特洛伊木马。事实上，能源部确实检查过一个，但从未公布结果，这可能意味着他们发现了什么。我只是不知道我们如何应对。

They're definitely above us because they have a bigger population and they have way more industrials. So they have more power total than us. More power generation capabilities. And by the way, it's actually not rocket science why that is. If you break that down to sources of that power in China, it's because coal is like 80% of that. Yeah, they're all coal. It's just tons of coal. And then in the US, renewables have grown a lot, but a lot of the reason the overall number is flat is because we're using renewables to replace coal, natural gas, like fossil fuels. So when you net it out in the US, we're flat. And then in China, it's straight up. So that's the first thing: we need drastic action. The administration has the National Energy Dominance Council. We've sat down with them a few times. We have to take drastic action to enable us to at least start matching their speed of adding energy to the grid and ideally surpass it. That's the first thing. The second thing you're talking about is our grid is extremely antiquated and that's a major strategic risk. I don't know what the cause or the source of the outage across Spain was, but some people think it was a foreign actor or some kind of cyber attack. I guarantee you the US energy grid is extremely susceptible to large-scale cyber attacks. The sophistication of these cyber attacks sometimes is so stupid. It's like if you find the right power plant login terminal, sometimes people don't change the username and password from the default, which is username and password. So you can just find some power station in Wyoming that still has the username and password as username and password. You log in and you can shut down the entire power in the entire region. So our grid, just because of how antiquated and decentralized it is, is hyper susceptible to cyber attacks, hyper susceptible to foreign action. And that matters now. If you take the energy grid in a major city, people will die. So it's bad now, but then let's go back to what we were just talking about with AI. Let's say we have large-scale AI on AI warfare with China. They just take out the power grid, take out our data centers and the power fueling those data centers, and then we're sitting ducks. Not only that, but it's my understanding that China actually produces and manufactures a lot of the major components that go into our grid, like the transformers. We don't even check those for malware, Trojan horses. In fact, DOE actually did an inspection on one and never released the results of what they found, which probably means they found something. I just don't know how we combat that.

我的意思是，就像……这还发生在哪里？看看“盐台风”，这是一个最近解密的黑客事件，是中国恶意软件和网络活动，基本上完全渗透了我们主要的电信运营商。我认为 AT&T 完全被这个来自中共的“盐台风”黑客攻击所攻陷。他们这样做是为了能够读取所有消息，所有短信、所有音频，作为情报收集行动的一部分。但如果他们能黑进我们的电信系统，他们肯定也能黑进我们的电网，显然也能黑进任何其他关键基础设施。这又回到我们刚才说的：电网方面，如果我们不能生产足够的电力，我们就完了；如果对手能随意切断我们的电力，我们也完了。所以，作为一个国家，我们在电网的网络安全态势上存在一个重大漏洞。我认为这是我们整个国家最明显、最直接的漏洞之一。这会造成社会动荡。想象一下，如果你切断休斯顿的电网，人们会死亡，你会引发各种混乱。但如果你再切断数据中心、军事基地、雷达系统，几乎任何本土基础设施，这些都会为对手创造巨大的战略突破口。

I mean just like the like what what is where did that happen elsewhere? Like look at um Salt Typhoon like uh this was a recent hack that was declassified which is that Chinese malware and cyber act activity like basically had fully infiltrated our um major telecom providers. I think AT&T was like entirely uh like um entirely compromised by this hack called Salt Typhoon um from the CCP and uh and that's they did that so that they could read all the messages like all the SMS all the audio um they were able to to capture as part of that as part of an intel gathering operation. Um, but if they're able to hack into our telco, they've sure as hell, you know, they're clearly capable of hacking into our energy grid, clearly h capable of hacking to any other any of our other critical infrastructure. And um, and it just goes back to what we're talking about, like the energy grid, a, if we can't produce enough power, we're hosed, and b, if the adversaries can take out our power at will, we're hosed. Mhm. Um, and so we have this major major vulnerability as a country on just like the cyber posture of our energy grid. I think it's like I think it's one of the biggest like very obvious like flatout um uh like clear vulnerabilities of our overall of our entire country. Um a just like you create civil unrest. you can like take, you know, imagine you took Houston's power grid out. People would die and uh you cause like all sorts of chaos. But then if you but then you take out these data centers, you take out um military bases, you take out radar systems, um you take out, you know, you name it, you can take out almost any piece of homeland infrastructure and that those create huge strategic openings for adversaries.

我的意思是，你在这个圈子里活动。你在建造大型数据中心，对吧？所以当你去华盛顿游说，说我们需要更多电力时，你见的是哪个协会？国家能源主导委员会？他们怎么说？他们完全同意。他们知道我们必须建设更多电力。然后就是下一层细节：我们如何加速核能？如何加速审批流程？有哪些已关闭的发电设施可以重新启用？我们列出了所有自然要做的事情。我认为我们知道该做什么。问题是我们能否不给自己设限，以及我们的电网是否如此陈旧，以至于这种漏洞意味着我们随时可能被摧毁。

I mean, what M you have to run in these circles. I mean, you're building massive data centers, correct? And so when you go to DC and you're advocating, hey, we need more power and you just I I didn't What's the association you met with? Uh the National Energy Dominance Council, I mean, what do they say? They totally agree. I mean, they know we have to build more power. And then it's about So then you get to the next layer of detail. It's like, okay, how can we how do we accelerate nuclear? How do we accelerate the permitting process? Um what are existing power generation capabilities that we turned off that we can turn back on? Um like you go through all the natural things to do. Like it's I mean I think I think we know what to do. The question is if we can get out of our own way and if and then if our grid is so antiquated that even that vulnerability like kind of means that we can be taken out any time.

我可能做了个假设。你们自己在建数据中心吗？我们本身不建数据中心，而是与那些正在建造世界上最大数据中心的公司合作。好的。我还听说这些大型数据中心开始自己发电了。这有道理吗？

I mean I may have made an assumption. Are you are you building data centers? We we ourselves are not building data centers. We partner with companies that yeah that are building you know the largest data centers in the world. Okay. And so I've I've also heard rumors that these major data centers are starting to just create their own power source. Is that is there any validity to that?

是的。所以现在很多设计都涉及为每个数据中心建造一个小型模块化反应堆（SMR），基本上就是让核反应堆与数据中心共址，为其供电。我认为这是个好主意。但问题是，中国在这方面会远远领先我们。世界上最大的核电站就在中国。所以，显然我们需要大力发展核能，也需要利用所有发电来源，采取全方位发电方式。但即便如此，我们也无法自信地超越中国，只能勉强追赶他们的水平。所以，这是一个巨大的问题。

Yeah. So a lot of designs these days involve can you just create an SMR a small um uh like a like a nuclear reactor per data center. can you basically like have a nuclear reactor colllocated with the data center um to uh to power that that data center's capacity um which I think is a good idea. The issue is like I mean China is going to be way ahead of us on that. The largest nuclear power plant in the world is in China. So um you know we're yeah you obviously we need to lean into nuclear that needs to happen obviously we need to to lean into all power generation sources. when you kind of an all the above approach to power generation. Um uh but even that doesn't get us to a posture where you're confidently exceeding China. You're just kind of catching up to where they are. And so um I mean this is a huge a huge issue.

好的。我们快速休息一下。回来后我想更深入地探讨中国的能力和我们自己的能力。

Yeah. Let's take a quick break. When we come back I want to I want to dive more into China's capabilities and and our capabilities.

明智的理财就是让每一美元发挥更大价值。使用 Rocket Money，你可以轻松找到被遗忘的订阅，并让他们为你协商账单，把钱放回你的口袋。有了这些节省，Rocket Money 几乎能自给自足。Rocket Money 是一款个人理财应用，帮助查找和取消不需要的订阅，监控支出，并帮助你降低账单，从而增加储蓄。使用 Rocket Money，当账单涨价、出现异常支出或接近超支时，你会收到提醒。新的目标功能会自动为你存钱，无需你操心。Rocket Money 拥有超过 500 万用户，通过取消订阅累计节省了 5 亿美元，使用所有高级功能的会员每年最多可节省 740 美元。立即取消不需要的订阅，更快实现财务目标。访问 rocketmoney.com/srs。网址是 rocketmoney.com/srs。

Smart money moves are all about getting more out of every dollar. With Rocket Money, you can easily find forgotten subscriptions and have them negotiate bills for you, putting money back in your pocket. And with all those savings, Rocket Money practically pays for itself. Rocket Money is a personal finance app that helps find and cancel your unwanted subscriptions, monitors your spending, and helps lower your bills so you can grow your savings. With Rocket Money, you get alerts if bills increase in price, there's unusual spending activity, or if you're close to going over budget. The new goals feature automatically saves money for you, so you don't have to think about it. Rocket Money has over 5 million users and has saved a total of 500 million in cancelled subscriptions, saving members up to $740 a year when they use all of the apps premium features. Cancel your unwanted subscriptions and reach your financial goals faster with RocketMoney. Go to rocketmoney.com/srs today. That's rocketmoney.comsrs. rocketmoney.comsrs.

当我开始这个播客时，似乎一切都得靠自己摸索。这让人不知所措。当你开始新事物时，待办事项清单似乎不断增长，可能占据你的全部生活。找到合适的工具对数百万企业来说可能是改变游戏规则的关键。这个工具就是 Shopify。Shopify 是全球数百万企业背后的电商平台，占美国所有电子商务的 10%，从家喻户晓的品牌到刚刚起步的品牌。我用 Shopify 来运营我自己的业务，这样我就能继续为你带来 Vigilance Elite 小熊软糖。Shopify 拥有数百个现成模板，帮助你建立在线商店，并配备有用的 AI 工具来加速内容创作。像营销团队一样，Shopify 可以在你的客户所在之处创建电子邮件和社交媒体活动。Shopify 是你的电商专家，在从库存管理到国际运输等各方面拥有世界级专业知识。如果你准备好销售，你就准备好使用 Shopify 了。在 Shopify 的支持下，把你的大生意想法变成现实。立即在 shopify.com/srs 注册，享受每月 1 美元的一个月试用期，开始销售。访问 shopify.com/srs。

When I started this podcast, it seemed like I had to figure it out all on my own. It was overwhelming. When you're starting something new, it seems like your to-do list just keeps growing and it can overrun your entire life. Finding the right tool can be such a gamecher for millions of businesses. That tool is Shopify. Shopify is the commerce platform behind millions of businesses around the world. and 10% of all e-commerce in the US from household names to brands just getting started. I use Shopify to power my own business so I can keep bringing you Vigilance Elite gummy bears. With hundreds of readytouse templates, Shopify helps you build online and is packed with helpful AI tools to accelerate your content creation. And like a marketing team, Shopify can create email and social media campaigns wherever your customers are. Shopify is your commerce expert with world-class expertise in everything from managing inventory to international shipping and beyond. If you're ready to sell, you're ready for Shopify. Turn your big business idea into with Shopify on your side. Sign up for your one month, $1 per month trial period and start selling today at shopify.com/srs. Go to shopify.comsrs. shopify.com/srs.

好了，Alex，我们休息回来了。我们准备讨论一下我们的能力与中国的能力对比。我们刚刚谈完了电力。在 AI 竞赛中，中国在其他领域是否领先美国？习近平本人甚至说过，AI 竞赛的赢家将实现全球主导。是的。

All right, Alex, we're back from the break. We're getting ready to discuss some of our capabilities versus China's capabilities. And you know, we we we just got done kind of talking about power. Is China leading the US in any other realms when it comes to the AI race? I mean Xi Jinping has even has said himself you know the the the winner of the AI race will achieve global domination. Yeah.

我认为，首先正如你提到的，要理解的是中国自 2018 年以来一直在按照 AI 总体规划运作。中共发布了一个广泛的、全政府的、军民融合的计划，以在 AI 领域取胜。

I think well the first thing almost as you're mentioning to understand is China has been operating against an AI master plan since 2018. They um the CCP put out a a broad whole of government you know civil military fusion plan to win on AI.

你提到习近平本人曾说过，人工智能将决定这场全球竞争的赢家。从军事角度看，他们明确表示相信 AI 是一种跨越式技术，意思是即使他们今天的军事实力不如美国，但如果他们过度投资 AI，拥有比美国更 AI 化的军队，他们就能实现跨越。所以他们投入巨大。现在，我认为描述当前局势的最佳方式是：他们在电力和发电方面遥遥领先。他们在芯片上落后，但正在追赶。他们在数据上领先我们。中国自 2018 年以来就有一项大规模行动来主导数据。2023 年，中国有超过 200 万人在数据工厂内工作，担任数据标注员或注释员，基本上是在为 AI 系统创造数据。相比之下，美国这个数字大约是 10 万。所以他们在数据上的投入是我们的 12 倍。他们有超过七个完整的城市作为专用数据中心，支撑着这种广泛的数据主导策略。在算法方面，我认为他们与我们持平，这要归功于大规模间谍活动。这是科技行业公开的秘密之一：中国情报机构基本上窃取了美国所有的知识产权和技术秘密。

You mentioned Xi Jinping himself has spoken about how AI is going to define the future winners of this global competition. From a military standpoint, they say explicitly that they believe AI is a leapfrog technology, meaning even though their military is worse than America's today, if they overinvest in AI and have a more AI-enabled military, they could leapfrog them. So they've been super invested. Right now, I think the best way to paint the current situation is they are way ahead on power and power generation. They're behind on chips but catching up. They are ahead of us on data. China has had, since 2018, a large-scale operation to dominate on data. In 2023, there were over 2 million people in China working inside data factories as data labelers or annotators, basically creating data to fuel AI systems. That number in the US is something like 100,000. So they're outspending us 12 to 1 on data. They have over seven full cities in China that are dedicated data hubs powering this broad approach to data dominance. On algorithms, I think they are on par with us because of large-scale espionage. This is one of those open secrets in the tech industry: Chinese intelligence basically steals all the IP and technological secrets from the United States.

有很多令人担忧的报告。一个是谷歌的一名工程师带走了谷歌设计 AI 芯片的所有设计和知识产权，然后搬到中国，利用这些设计创办了一家公司。他从谷歌企业云中窃取数据的方式很愚蠢：他只是把所有代码复制粘贴到 Apple Notes 中，导出为 PDF，打印出来，然后直接走出去。后来被发现了，但几个月里我们完全不知道他们窃取了所有这些关键知识产权。斯坦福大学，上周刚爆出，完全被中共特工渗透。根据中国法律，任何中国公民都必须配合中共情报收集。所以如果你是美国华裔，中国情报机构联系你，你就得提供你所看到和发现的东西。在所有顶尖大学、科技公司和 AI 实验室里都有大量中国公民。大约六分之一的中国留美学生持有中共资助的奖学金，他们必须向接头人汇报，否则奖学金会被取消。有一场针对美国科技行业的大规模情报行动，从我们最伟大的研究机构、大学、AI 实验室和科技公司收集所有信息和秘密。我认为这是中国如此迅速追赶的一个被低估的因素。DeepSeek 突然出现，所有人都惊讶于他们的模型能力以及他们如何学会所有这些技巧。其中有多少是他们自己发明的，又有多少是因为他们有一个精妙的间谍行动，窃取了我们所有的商业机密并在中国重新实现？

There are a bunch of very concerning reports. One is a Google engineer who took the designs and all the IP of how Google designed their AI chips and moved to China, then started a company using those designs. The way he stole the data out of Google's corporate cloud was stupid: he just copy-pasted all the code into Apple Notes, exported to a PDF, printed it, and walked out. That was later discovered, but for months we had no idea they had stolen all this critical IP. Stanford University, this just came out last week, is entirely infiltrated by CCP operatives. By law in China, any Chinese citizen must comply with CCP intelligence gathering. So if you're a Chinese citizen living in the US and Chinese intelligence reaches out, you have to give them what you see and find. There are tons of Chinese nationals across all major elite universities, tech companies, and AI labs. About a sixth of Chinese students in America are on CCP-sponsored scholarships, and they have to report back to a handler or their scholarships get revoked. There's an incredibly large-scale intelligence operation against the US tech industry, collecting all the information and secrets from our greatest research institutions, universities, AI labs, and tech companies. I think this is a very underrated element of how China caught up so quickly. DeepSeek came out of nowhere; everyone was surprised at how capable their model was and how they learned all these tricks. How much of that is because they came up with it on their own, or they had an exquisite espionage operation to steal all our trade secrets and reimplement them in China?

我们的间谍活动怎么样？远没有那么好。中共为 DeepSeek 做的一件事是，在它爆红、CEO 会见中国总理后，他们把研究人员都集中起来——我不该说关起来，但他们把所有人聚在一起，收走了所有护照。所以 DeepSeek 的 AI 研究人员没有一个能离开中国，也不能接触任何外国人。他们基本上封锁了整个研究工作，使得任何针对该行动的间谍活动都非常困难。还有一份报告称，十年或十五年前，许多美国中情局在华特工被杀，因为他们的通信渠道被中国情报机构攻破。所以我们在中国的间谍活动极其深入且风险巨大；我们被中国情报机构深度渗透。相比之下，我们的能力要弱得多，而且我认为他们设计得很难渗透进他们的 AI 工作。所以他们在数据上领先，通过间谍活动在算法上追赶，在电力上领先。我们在什么方面领先？目前，我们在芯片上领先。这是我们的救命稻草：英伟达芯片和整个堆栈是世界的骄傲，我们是最先进的。中国芯片正在追赶；最近的报告称华为芯片大约落后英伟达一代。所以他们很接近。所有这些都相当令人担忧。还有一份 CSIS 报告提到一个中国项目，叫做下一代大脑理解项目，他们基本上试图用 AI 来完全理解人类个性和心理行为。

What does our espionage look like? Nowhere close to as good. One thing the CCP did for DeepSeek is after it blew up and the CEO met with the Chinese premier, they locked up all the researchers—I shouldn't say locked up, but they huddled them together and took all their passports. So none of the AI researchers at DeepSeek can leave the country or come into contact with any foreigners. They basically locked down the entire research effort, making it very hard to conduct any espionage into that operation. There's also a report that a decade or 15 years ago, many CIA operatives in China were killed because their communication channel was compromised by Chinese intelligence. So our espionage in China is extremely deep and risky; we are deeply penetrated by Chinese intel. Comparatively, we have much less capability, and I think they've designed it so it's very hard to infiltrate their AI efforts. So they're ahead on data, they catch up on algorithms through espionage, and they're ahead on power. What are we ahead on? Right now, we're ahead in chips. That's our saving grace: Nvidia chips and the entire stack are the pride of the world, and we're the most advanced. Chinese chips are catching up; recent reports say Huawei chips are about one generation behind Nvidia. So they're close. All of this is pretty concerning. There was another CSIS report about a Chinese effort called the next-generation brain understanding project, where they're basically trying to use AI to fully understand human personality and psychological behaviors.

我想这最终实际上就是信息战。就像我们早餐时聊到的，中国有大规模的信息行动、大规模的信息战，而且已经搞了几十年，真的几十年，最早可以追溯到香港的线下行动。他们非常老练，而人工智能还会让他们行动更快。我们怎么应对？

I imagine that's ultimately for effectively like information warfare. As we were talking about at breakfast, I mean China has large-scale information operations, large-scale information warfare, and has been doing that for decades, literally decades, going back all the way to in-person operations in Hong Kong. They are so sophisticated, all that, and AI is going to enable them to just move much faster as well. How do we combat that?

嗯，我认为我们需要自己的信息行动。我觉得这非常关键。这是针对那一条线。然后我认为我们需要认识到，归根结底，我们是一个更具创新力的国家，但如果我们想在人工智能领域长期获胜，就必须大幅振作起来。我们需要将芯片制造回流本土。我们需要制造大量的芯片。我们不能依赖台湾来制造我们的高端芯片。

Well, I think we need our own information operations efforts. I think that's pretty critical. That's specifically on that thread. And then I think we need to acknowledge that at the end of the day, we are a more innovative country, but we have to dramatically get our act together if we want to win long-term in AI. We need to onshore chip manufacturing. We need to be manufacturing huge numbers of chips. We can't be dependent on Taiwan to manufacture our high-end chips.

我们开始做了吗？有任何规模吗？

Are we doing that yet? At any capacity?

规模极小。亚利桑那州有几家晶圆厂可以生产一些芯片，但绝大多数产量仍然来自台湾。我们需要大幅加强人工智能公司的安全。我们需要有适当的反间谍措施来了解这些公司内部的间谍风险。我们需要解决我们谈到的电力问题。我们需要投资于网络威胁，比如投资于大规模网络防御。我们需要投资于数据。我们需要自己的数据主导计划，以确保中国不会用更高质量和更大规模的人工智能数据集甩开我们。所以你可以逐一审视每个要素，为美国制定一个合适的获胜计划。但我们开始做了吗？我认为有些事情正在进行中，但不够，远远不够，无法确保美国会赢。绝对不够。

Extremely small capacity. There are a few fabs in Arizona that can produce some chips, but the vast majority of the volume still comes out of Taiwan. We need to tighten up security in our AI companies dramatically. We need to have proper counterintelligence on what is the espionage risk within these companies. We need to solve the power problem that we talked about. We need to be investing into cyber threats, like investing into large-scale cyber defense. We need to invest into data. We need our own programs around data dominance to ensure that China doesn't just run away with higher quality and greater AI datasets than us. So you can go through each of the elements and build the proper plan for the United States to win. But have we started any of that? I think some things are underway, but not enough, nowhere close to enough to be sure that the US will win. Definitely not.

而且他们还有一个根本优势。现在很多人常说：‘哦，我们在美国需要一个人工智能曼哈顿计划，把所有聪明人聚集起来，集中资源，搞一个大型项目。’但实际上，在美国很难做到这一点，而中国却可以轻松做到。中国可以说：‘嘿，所有最优秀的人工智能人才，你们现在都在一家公司工作。我们要集中你们所有的资源。我们要把你们安置在世界上最大的核电站旁边。我们要在这里建造世界上最大的数据中心。中国所有的芯片都将用于这个大规模人工智能项目。’他们有能力把所有资源集中起来，投入到赢得人工智能竞赛中。而在美国，我们有这么多公司。美国政府目前不会强迫所有这些公司合并。那会被视为政府权力的过度扩张。但正因为如此，我们会有五个分散的人工智能项目。也许总体上，我们会拥有更多的芯片、更多的电力和更多优秀的研究人员，但我们无法集中这些努力，而中国却能轻松集中所有力量。

And they also have a fundamental advantage. One of the things that people say a lot now is, 'Oh, what we need in the United States is an AI Manhattan Project where we collect all the brilliant minds together, collect our resources, and have one large effort in the United States.' Well, it turns out it's actually really hard to pull that off in the United States, but China can pull it off super easily. China can just say, 'Hey, all the best AI people, you now work in one company. We're going to pull together all of your resources. We're going to put you right next to the largest nuclear power plant in the world. We're going to build the largest data center in the world here. All the chips that China has are going to go towards building this large-scale AI project.' And they just have the ability to collect all of their resources together and throw it at winning the AI race. Whereas in the United States, we have all these companies. The United States government, as of yet, is not going to force all these companies to combine and merge. That would be viewed as such an overreach of government power. But because of that, we're going to have five fragmented AI efforts. And maybe in aggregate, we'll have way more chips, more power, more great researchers, but we're not going to be able to focus those efforts, whereas China is easily going to be able to focus all their efforts.

哇。你之前在楼下提到过核武器的事。

Wow. You had mentioned something downstairs about nuclear weapons.

是的。这就是国家安全变得非常诡异的地方。你可以清楚地想象到，先进的网络人工智能可能会使核威慑失效。我是什么意思呢？现在，没有人发射核弹，因为我们有“相互确保摧毁”（MAD）。如果我对另一个国家发动第一次打击，他们能在核弹还在空中时进行第二次打击，我们双方都会遭受毁灭。那会非常糟糕。所以由于这种二次打击能力，我们有了有效的威慑。但假如换一种情况，假设我是美国，拥有世界上最先进的人工智能网络黑客能力。我可以构建能入侵任何其他国家的人工智能智能体，关闭他们的电网，禁用他们的武器系统，禁用一切。那么我会怎么做呢？我发动第一次打击，或者先派出我的网络人工智能智能体部队。我派出我的网络人工智能力量，有效地禁用敌国的所有武器系统。因为我拥有如此强大的人工智能能力，我可以摧毁你所有的武器系统。然后我发动第一次打击，而你没有二次打击能力。如果发生这种情况，基本上人工智能和核武器的结合意味着你无法仅用核武器来威慑人工智能加核武器。这将迫使人工智能能力扩散，甚至小国也需要大量投资人工智能能力，因为他们的核武器不再足以构成威慑。

Yeah. So this is where stuff gets really weird for national security. You could clearly imagine scenarios where advanced cyber AI invalidates nuclear deterrence. What do I mean by this? Right now, nobody fires nukes because we have MAD, mutually assured destruction. If I do a first strike against another country, they're going to be able to, while that nuke is in the air, do a second strike, and we'll both have destruction on both sides. It'll be really bad. So because of this second strike capability, we have proper deterrence. Well, what if instead, let's say I'm the United States and I have the most advanced AI cyber hacking capabilities in the world. So I can build AI agents that hack into any other country, can turn off their energy grid, disable their weapon systems, disable everything. So what do I do instead? I launch the first strike, or first I send in my cyber AI agent capabilities. I send my cyber AI force effectively to disable all the weapon systems of the enemy country. And because I have so much AI capacity, I can take out all of your weapon systems. Then I send my first strike, and you don't have a second strike capability. So if that happens, basically the combination of AI and nuclear means you cannot deter AI plus nuclear with just nuclear. So this will force the proliferation of AI capabilities, and even small countries are going to need to invest in lots of AI capabilities because their nuclear weapons are no longer a sufficient deterrent.

天哪。那生物武器呢？

Jeez. What about bioweapons?

是的，这是目前被严重低估的一个因素。新冠病毒从武汉的一个病毒学实验室泄漏，基本上让世界停摆了两年。那还只是第一级的生物风险。那是一种相对无害的病原体，但仍然在全球导致至少 1000 万人死亡，让整个世界停摆了两年。而最近的新模型，新的人工智能模型，能够胜过 95%的麻省理工学院病毒学家。根据安全中心最近的一项研究，OpenAI 和 Google 的最新模型实际上比麻省理工学院 95%的病毒学家更聪明。所以现在，无论是现在还是几年后，利用人工智能能力来帮助设计强大的病原体将是可行的。而且更重要的是，你将能够设计这些病原体的某些特性。你可以调整它们的传播性、调整它们的致死率。此外，由于合成生物学的最新进展，你现在可以创造出专门针对特定 DNA 片段的病毒。

Yeah, this is the element that is really underrated right now. So COVID leaked out of a virology lab in Wuhan and basically shut the world down for two years. And that's like the level one biorisk kind of stuff. This was a relatively innocuous pathogen, but it still killed probably at least 10 million people globally and shut the whole world down for two years. Well, recent models, the new AI models, are able to outperform 95% of MIT virologists. So the newest models from OpenAI and Google are smarter than literally 95% of virologists at MIT, based on a recent study by the Center for Safety. So now, whether it's right now or in a few years, it will be feasible to use AI-based capabilities to help you design powerful pathogens. And what's more, you're going to be able to design certain characteristics of these pathogens. You'll be able to tune the virality, tune the lethality of them. Also, due to recent advancements in synthetic biology, you can now create viruses that specifically target certain segments of DNA.

所以我可以制造一种生物武器，只针对具有特定 DNA 片段的个体，这意味着我可以针对世界上任何人群、任何群体或任何亚群体。这非常非常糟糕。所以这种能力——首先，即使没有 AI，生物学和合成生物学也在取得巨大进展，生物武器或病原体、病毒泄漏的风险本身就存在。而有了 AI，虽然不是今天这些模型，但再经过几代，你将能够利用这些 AI 系统设计或构建下一代病原体。出于充分的理由，国际上有禁止生物战的条约，但想象一下这样的场景：有些国家没有核威慑，也没有资源建设大规模 AI 数据中心，我担心它们会转向生物武器作为威慑手段，这对世界来说非常不稳定。

So I could create a bioweapon that targeted any individual with a certain segment of DNA, which means I can target basically any population or any group or any subsegment of the population in the world. That is really really bad. So the ability—first, even without AI, biology and synthetic biology are making so much progress that there are all sorts of inherent risks of bioweaponry or leaks of pathogens and viruses. And then with AI, not literally today's models but a few generations down, you're going to be able to use these AI systems to design or build next-generation pathogens. For good reason, there are international treaties against biological warfare, but if you imagine scenarios where countries lack nuclear deterrence or the resources for large-scale AI data centers, I'm worried that countries will turn to bioweapons as their deterrence mechanism, which is highly destabilizing for the world.

哇，这太可怕了。但另一方面，也有新技术可以预防这些。西雅图 David Baker 实验室的研究——他刚因数字鼻子获得诺贝尔奖。基本上，这些设备可以自动检测空气中的蛋白质、化学物质或病原体。我认为生物和生物武器的真正攻防将最终表现为大规模部署数字鼻子，在每个空间、每个集装箱、每架飞机上，持续感知所有已知和新型病原体，检测并最终遏制它们。

Wow, that's scary. The flip side is there is new technology that can prevent this stuff. There's research coming out of David Baker's lab in Seattle—he just won a Nobel Prize on digital noses. Basically, these devices can detect proteins or chemicals or pathogens in the air automatically. I think the real offense-defense of bio and bioweaponry will end up looking like large-scale deployment of digital noses in every space, on every shipping container, on every plane, constantly sensing for all known and new pathogens, detecting and ultimately containing them.

有意思。实时嗅探所有这些。

Interesting. It's sniffing real time for all of that.

是的，没错。另一方面，如果 AI 在开发新型生物武器，那么 AI 也应该能够找出疫苗或解药，对吧？

Yeah, exactly. On the flip side, if AI is developing a new bioweapon, then AI should also be able to figure out the vaccine or antidote, correct?

完全正确。所以会有攻防元素，就像 AI 应用于指挥控制、网络和生物领域一样。希望我们最终能达成一个世界共识，即由于相互威慑，没有人会走这些路——没有人值得为此破坏稳定、拿人类冒险。这就是我们需要达到的状态。

Yeah, totally. So there will be an offense-defense element, just as with AI applied to command and control, cyber, and bio. The hope is that we end up in a world where everyone agrees not to go down any of these paths because of mutual deterrence—it's not worth it for anybody to destabilize and risk humanity. That's where we need to land.

哇。你对中台局势有多担心？我简直不敢相信他们还没动手。我原以为肯定会在上届政府末期发生，但以他们的芯片生产能力，你对中国拿下台湾有多担心？

Wow. How concerned are you about China and Taiwan? I can't believe they have not made a move yet. I thought for sure it would happen towards the end of the last administration, but with their chip production capabilities, how concerned are you about China taking Taiwan?

我认为如果要发生，就会发生在这十年，很可能在本届政府任内。中国有巨大的人口问题——独生子女政策导致的老龄化。这将在未来十年内显现，使他们更像日本，庞大的老龄化人口会瘫痪他们采取激进举措的能力，尤其是在军事工业产能方面。所以他们会想尽早行动。过去几十年他们进行了疯狂的军事建设。目前，中国的工业和制造能力远超美国，所以这对他们来说是一个窗口期。

I think if it's going to happen, it's going to happen this decade and probably this administration. China has huge demographic issues—an aging population from the one-child policy. That will play out over the next decade, making them look more like Japan, with a large aging population paralyzing their ability to make aggressive moves, especially in military industrial capacity. So they'll want to move sooner rather than later. They've had an insane military buildup over the past few decades. Currently, China has far more industrial and manufacturing capacity than the United States, so that's a window for them.

所以你认为他们因为人口老龄化而被迫行动？

So you think they're pressed to do it because of the aging population?

我认为有很多因素。习近平也在变老，这将成为他遗产的重要组成部分。老龄化人口会逐渐缩小他们的政治空间。而且他们正处于一个疯狂窗口期，拥有无与伦比的工业制造能力。2023 年，中国部署的工业机器人数量超过世界其他地区的总和。他们在自动化工厂方面比任何国家都跑得快。所以如果要动手，他们会很快行动。

I think a lot of factors. Xi is aging, and this will be an important part of his legacy. The aging population will minimize their political latitude over time. And they are in an insane window with incredible industrial manufacturing capabilities compared to anywhere else. In 2023, China deployed more industrial robots than the rest of the world combined. They're racing faster than any other country in automated factories. So if they're going to do it, they'll do it soon.

我们使用的芯片有多少来自台湾？95%的高端芯片在台湾制造。如果中国拿下台湾会怎样？

What percentage of chips come from Taiwan? 95% of high-end chips are manufactured in Taiwan. What happens if China takes Taiwan?

我们来推演一下。如果中国封锁或入侵台湾，那些晶圆厂价值连城。如果你相信 AI 进步的节奏，一切都归结于你有多少电力和多少芯片。如果他们拥有全球 95%的芯片制造能力，他们将一骑绝尘。

Let's war game it out. If China blockades or invades Taiwan, those fabs are incredibly valuable. If you believe in the pace of AI progress, everything boils down to how much power and how many chips you have. If they own 95% of the world's chip manufacturing capability, they will run away with it.

那么你看，台湾人会轰炸台积电的数据中心吗？或者美国会轰炸台积电的芯片工厂吗？或者其他国家会轰炸它们？我个人认为，台湾人不会这么做，因为即使他们被封锁或入侵，这些工厂仍然是台湾生存能力和作为实体重要性的巨大组成部分。所以我认为他们不会这么做。中国肯定不会这么做，因为他们入侵部分就是为了获得这些能力。那么，美国会轰炸它们吗？如果美国轰炸它们，那可能就是第三次世界大战了。很难想象这不会导致大规模升级。所以没有好的选择。每个人都非常关注这一点，但这确实是一个火药桶般的地区。

So then you look at that and you say, will the Taiwanese people bomb the TSMC data centers? Or will the US bomb the TSMC chip fabs? Or will some other country bomb them? I think my personal belief, I don't think the Taiwanese do it because even if they get blockaded or invaded, those fabs are still a huge component of Taiwan's survivability and Taiwan's relevance as an entity. So I don't think they do it. China definitely doesn't do it because they're obviously invading partially to gain those capabilities. And then, does the US bomb them? If the US bombs them, that's probably World War III. It's hard to imagine that not resulting in massive escalation. So there are no good options. Everyone's very focused on it, but it is a real powder keg of a region.

你认为这一切会如何结束？我们早餐时讨论过一点。

How do you think this all ends? We had a little discussion about this at breakfast.

是的。我认为如果在未来几年，比如 3 到 4 年内，台湾发生入侵或封锁，考虑到 AI 的重要性，美国很难在这种情况下不采取任何行动，而几乎所有行动都会升级为重大冲突。所以最好的情况是我们完全威慑住入侵或封锁。不陷入大规模世界大战符合每个人的利益，那会极具破坏性并导致大量人员死亡。所以从根本上我们应该能够威慑住这场冲突。但这就是为什么这一切如此重要。我们需要确保我们国家在 AI 能力上是世界最强的。我们需要确保我们的军事 AI 能力是世界最强的。我们需要确保对这种情景有明确的经济威慑。我们需要在各个方面投资来威慑这场冲突。真正出问题的地方在于，如果中国共产党的算计与我们的不同——如果他们的算计变成‘这会成功，我们可以拿下，我们足够强大’——而我们的算计相反。那就是世界大战发生的情景。所以我认为威慑是可能的，我们必须做很多事情来确保我们威慑住这场冲突。这应该是，而且我认为已经是，整个国防部 80%的焦点。

Yeah. I think if in the next handful of years, next 3 to 4 years, there's an invasion or blockade of Taiwan, given how important AI is, it's hard for the US to not take any action in that scenario, and almost all actions would escalate into a major conflict. So best case scenario is we deter the invasion or blockade altogether. It's certainly in everyone's interest to not get into a large-scale world war that's hugely destructive and kills lots of people. So fundamentally we should be able to deter that conflict. But that's why all this matters so much. We need to make sure our AI capabilities as a country are the best in the world. We need to make sure our military AI capabilities are the best in the world. We need to ensure clear economic deterrence of this kind of scenario. We need to invest in every way to deter this conflict. Where this will really break down is if the Chinese Communist Party's calculus diverges from our own—if their calculus becomes 'this is going to work, we can take this and we're strong enough'—and our calculus is the opposite. That's where the world war scenario happens. So I think it's possible to deter, and we have to do a lot of things to make sure we deter that conflict. That should be, and I think already is, 80% of the focus of the entire Department of Defense.

所以，我的意思是，我们可以威慑，但当你谈到人口老龄化时，他们正在变得绝望，而且听起来为了他们真正获胜，他们必须获得那些芯片工厂，对吗？他们已经拥有 250 倍的造船能力。他们人多得多。他们比我们更有力量。美国的征兵率处于历史最低点。所以我想说的是，你只能威慑一个绝望的实体这么久，直到他们孤注一掷，对吧？你同意吗？

So, I mean, it's just we can deter, but when you're talking about an aging population, they're getting desperate, and it sounds like in order for them to legitimately win, they have to acquire those chip fabs, correct? And so, they already have 250 times the shipbuilding capacity. They have way more people. They have more power than we do. Military recruitment in the US was at an all-time low. So I guess what I'm saying is you can only deter a desperate entity for so long before they throw a Hail Mary play, right? Would you agree with that?

是的。然后这取决于——在我看来，你必须投入整个军队来包围台湾才能有效做到这一点。

Yeah. And then it just depends on the—you would have to dedicate an entire military to surround Taiwan to effectively do that, in my opinion.

是的。我的意思是，我认为如果中国共产党和中国人民解放军评估认为台湾就是他们所需要的，他们会把全部军事能力集中在夺取台湾上，那么这就成了一个非常棘手的算计。我的意思是，他们为什么不呢？如果习近平认为 AI 竞赛的赢家将实现全球统治，而他正在变老。你刚刚谈到他的遗产对他有多重要，我相信你是对的。我不知道你如何威慑这一点。然后他们赢得了 AI 竞赛。

Yeah. I mean, I think that if the CCP and the PLA assess that Taiwan is all they need, they will focus their entire military capacity on seizing Taiwan, then that becomes a really tricky calculus. I mean, why wouldn't they? If Xi believes that the winner of the AI race achieves global domination, he's getting older. You just talked about how important his legacy is to him, which I'm sure you're right. I don't know how you deter that. And then they win the AI race.

我们唯一能做的，我认为这希望不大，但我觉得很重要，就是如果我们最终真的在 AI 上合作。我知道这听起来有点疯狂，但如果我们作为一个国家能够证明我们遥遥领先，并且存在 AI 自我改进或智能递归的想法。基本上，一旦 AI 变得足够好，你就可以开始利用 AI 来帮助你构建下一个 AI。你利用当前一代 AI 来更快地构建下一代 AI。所以在某个时刻，你的 AI 能力会实现某种指数级起飞。它们会变得非常非常快。如果有人比你落后哪怕 3 到 6 个月，他们就永远追不上了，因为你运行自我改进循环的速度比任何人都快。这是一个关键想法。目前这还有点理论化，但 AI 领域的很多人相信它。我可能也相信，我们将能够使用 AI 来帮助我们继续训练下一代 AI 并更快地改进事物。如果你相信这一点，那么如果我们比中国领先 3 到 6 个月，并保持这个优势并更快起飞，那么他们就会远远落后。然后最终我们处于一个很好的位置，可以说，‘嘿，我们遥遥领先，你们应该放弃努力。我们会给你们 AI 用于你们整个社会的经济和人道主义用途。我们同意不在军事 AI 上开战。’

The only thing that we can do, I think this is a long shot, but I think it's important, is if ultimately we actually end up collaborating on AI. And I know that sounds kind of crazy, but if we're able as a country to demonstrate we're so far ahead and there's this idea of AI self-improvement or intelligence recursion. Basically, once AIs get sufficiently good, you can start utilizing the AIs to help you build the next AI. You utilize your current generation AI to build the next generation AI faster and faster. So at some point, your AI capabilities enable some form of exponential takeoff. They get good really, really quickly. And if somebody's even 3 to 6 months behind you, then they're never going to catch up because you're running the self-improvement loop faster than anybody else. This is a key idea. It's a little bit theoretical right now, but a lot of people in AI believe it. I probably believe it too, that we will be able to use AIs to help us continue training the next AIs and improve things more quickly. If you believe that, then if we're 3 to 6 months ahead of China and we maintain that advantage and take off faster, then they're going to be way behind. And then ultimately we're in a great position to say, 'Hey, we're way ahead and you guys should quit your efforts. We'll give you AI for all of your economic and humanitarian uses throughout your society. And we agree we're not going to battle on military AI.'

需要什么才能把台湾拥有的芯片制造能力带到美国来保护它？

What would it take to take the chip building capabilities that Taiwan has and implement that here in the US to protect it?

首先，已经有数千亿美元投资于这些工厂和代工厂的建设，实际上是这些大规模芯片工厂的建立，以及其中的所有高端设备和工具。数千亿美元的投资。所以首先，美国需要数千亿美元的投资。这还不是最难的部分。

So the first thing is there's been hundreds of billions of dollars invested just into the buildout of those fabs and foundries, the buildup of these large-scale chip factories effectively, and all the high-end equipment and tooling inside of them. Hundreds of billions of dollars of investment. So first off, there needs to be hundreds of billions of dollars investment in the US. That's not the hard part.

第二部分才是真正的难点：这基本上是一个由经验丰富的高技能工人运营的大型工厂，整个流程像钟表一样精密运转。除非你能把这些人带到美国，否则你就得重建所有那些技术诀窍和能力，这需要很长时间。这就是问题之一——你觉得我们为什么没这么做？为什么我们没有激励这些聪明人来这里为我们做这件事？台积电（TSMC）在亚利桑那州建了几座晶圆厂，但他们遇到了问题：首先是许可和电力供应，还处理了一些环保局的问题。然后就是亚利桑那州的技术人员不如台湾的技术人员熟练或勤奋。所以他们尝试了，但我们的繁文缛节和电力基础设施达不到要求。繁文缛节、电力、劳动力。还有另一个关键点：从台积电的角度看，他们并没有太大动力在美国建立所有这些能力。一旦他们在美国建立这些能力，美国就没有动力保卫台湾了。而这是一家台湾公司——这是他们生存战略的关键部分。所以真正的症结在于：他们真的有动力在美国大规模建设芯片制造能力吗？我认为答案是否定的。

The second part that's really the hard part is it's basically a large-scale factory operated by highly skilled workers who are very experienced in those processes, and the whole thing operates like clockwork. Unless you can get those people to the US, you're going to have to rebuild all that knowhow and technical capability, and that takes a really long time. That's one of the things—why do you think we haven't done that? Why haven't we incentivized these brilliant minds to come here and do it for us? TSMC, the Taiwan Semiconductor Manufacturing Company, has built a few fabs in Arizona, but they cited issues: first, permitting and getting enough power, and they dealt with some EPA issues. Then they just have issues where the technicians working in Arizona aren't as skilled or don't work as hard as those in Taiwan. So they've tried to do it, but our red tape and power infrastructure aren't what they need to be. Red tape, power, workforce. And there's another key thing: from TSMC's perspective, they're not all that incentivized to stand up all these capabilities in the US. If they start standing up all these capabilities in the US, the US is not incentivized to defend Taiwan. And it's a Taiwanese company—it's a critical part of their survival strategy. So that's really where the rubber hits the road: are they actually incentivized to do a large-scale buildout of chip manufacturing capacity in the US? I think the answer is no.

有道理。我是说，必须达成某种协议，让他们处于我们的保护之下。

Makes sense. I mean, there would have to be some type of deal struck where they fall under our wing.

是的，你可以想象中美之间某种交易。那必须是最高层的外交协议，类似于：‘嘿，你们可以拥有台湾，但我们需要在美国大规模建厂’之类的。也许在某些情况下这种协议能达成。我不知道。但这也意味着美国必须说：‘我们现在只关心芯片制造，不关心中国人民和国家。’

Yeah, you could imagine some kind of deal between the US and China. It'd have to be a diplomatic deal at the highest levels, something along the lines of: 'Hey, you guys can have Taiwan, but we need large-scale fabs in the US' or something like that. Maybe there are worlds where that kind of deal could be drawn up. I don't know. But that would also mean the US would have to say, 'All we care about at this point is chip manufacturing, and we don't care about the Chinese people and the country.'

台积电和中国有合作吗？

Are they working with China at all, TSMC?

是的，理论上他们不应该，但很多华为——中国领先的公司之一——已经能够从台湾获得大量芯片。他们通常通过一个看似无关的壳公司在新加坡或马来西亚操作，那家公司从台积电购买大量芯片，然后邮寄回去。显然有很多台积电的高端产品流向了中国公司。

Yeah, they're technically not supposed to, but a lot of Huawei, one of the leading companies in China, has been able to get tons of chips from Taiwan. They usually do it through a cutout company that doesn't seem associated with them in Singapore or Malaysia, and then that company buys a bunch of chips from TSMC and mails them back. There's clearly been a lot of TSMC high-end outputs that have gone to Chinese companies.

哇，可怕。我是说，现在看看局势和所有动态，就像火药桶。非常不稳定，在很多方面问题重重。你只能相信必须努力寻求外交解决方案。因为战争对双方都非常糟糕。

Wow, scary. It gets—I mean, right now, if you look at the situation and all the dynamics at play, it's like a powder keg. Very volatile, highly problematic in many ways. And you just have to believe there's got to be some effort towards diplomatic solutions. Because war would be really bad for both sides.

是的。

Yeah.

我们如何与中国在 AI 方面协调？

How do we coordinate with China on AI?

现在，美国和中国绝对处于全面竞赛的动态中。我们将竞相构建最好的 AI 系统，他们也会竞相构建最好的 AI 系统。我们都全力以赴，竞相构建最先进的 AI 能力、最大的数据中心、最大的容量等等。如果你回想一下核能——核战争以及核能用于发电——那是全面启动，每个人都竞相建设能力和容量。然后切尔诺贝利和三哩岛事件发生了，引发了对该技术及其风险的大规模担忧。出现了国际条约和大型国际响应来协调核技术。但这使我们的国家在发电方面倒退了好几代。促成国际合作的是小规模灾难。你可以想象一个 AI 场景，由于我们讨论的所有这些，某个恐怖组织、非国家行为体或朝鲜决定以特别敌对或不人道的方式使用它，造成大规模灾难。例如，切断世界最大城市的电力，导致大量死亡，或者释放病原体杀死数千万人。类似的事情会让国际社会意识到我们必须协调，合作让 AI 改善我们的社会、经济和人民的生活。但我们需要协调其用于可怕目的，比如生物或网络战争，等等。

Right now, the US and China are definitely in an all-out race dynamic. We're going to race to build the best AI systems, and they're going to race to build the best AI systems. We're both all in on this approach, racing towards building the most advanced AI capabilities, the largest data centers, largest capacity, etc. If you recall how nuclear was—in nuclear war as well as application of nuclear towards power production—it was all systems go, everyone racing to build capacity and capability. Then Chernobyl and Three Mile Island happened, creating large-scale concern around the technology and its risks. There were international treaties and a large international response towards coordinating on nuclear technology. But that set our country back many generations in terms of power generation. What it took was small-scale disasters that were the forcing function for international cooperation. You can imagine a scenario with AI where, because of all the things we've been talking about, some terrorist group or non-state actor or North Korea decides to use it in a particularly adversarial or inhumane way, creating a disaster with large-scale fallout. For example, taking out power in one of the largest cities in the world, causing tons of deaths, or releasing a pathogen that kills tens of millions. Something like that would cause the international community to realize we have to coordinate on this, collaborate for AI to improve our societies and economies and lives. But we need to coordinate on its use towards scary things like bio or cyber warfare, and the list goes on.

长话短说，我认为真正的路径是某种 AI 石油泄漏或类似事件，让国际社会意识到必须开始协调。你说中国全力以赴投入竞赛，美国也是，但我们却在自缚手脚。你提到了繁文缛节、环保局、许可和电力问题。我们没有增加电力产出，而是停滞不前。据我所知，我们并没有消除这些障碍来启动这一切。这就像是在自断后路，对吧？现在，我们确实有很多工作要做。我们必须制定获胜的战略，实现能源主导、数据主导。在算法方面，我觉得没问题——虽然有间谍活动，但算法上我们能应对。我们需要确保长期保持芯片主导地位。还要确保这一切最终转化为军事优势。我完全同意。我们今天就需要确保有适当的战略，在这些领域保持领先。对美国来说，最坏的情况是中共在国内开展大规模曼哈顿式项目，利用我们讨论过的所有因素，开始超越美国 AI，从而获得极端军事优势，并用它来统治世界。这是最坏的情况。如果美中 AI 能力大致相当，我认为就形成了威慑。两国都不会冒险。如果美国遥遥领先，就能维持领导地位，世界也相对安全。所以最坏的情况是他们超过我们。

So long story short, I think the path really is some kind of AI oil spill or incident that causes the international community to realize we have to start coordinating. I mean, you say China is all out on the race, and the US is all out, but we're kneecapping ourselves. You mentioned the red tape, the EPA, permitting, and power. We're not producing more power; we're flatlined. We've established that, as far as I know, we're not getting rid of the red tape to launch this. It seems like we're cutting ourselves off at the knees here, right? Right now, we have a lot of work to do. We have to build strategies to win, to have energy dominance, to have data dominance. On algorithms, I think we'll be okay—there's espionage, but I think we'll be okay. We need to ensure we have chip dominance long term. We need to make sure all this lends itself to military dominance. I totally agree. We need to ensure today that we have proper strategies in place to stay ahead on all these areas. The worst case for the US is that the CCP does a large-scale Manhattan-style project inside their country, realizes they can start overtaking the US on AI due to all the factors we've discussed, which leads to extreme hyper military advantage, and they use that to take over the world. That's the worst case. If US and China AI capabilities are roughly on par, I think you have deterrence. Neither country will take the risk. If the US is way ahead, you maintain US leadership, and that's a safe world. So the worst case is they get ahead of us.

除了美国和中国，还有其他参与者吗？我们还需要留意谁？

Are there any other players other than the US and China involved in this? Who else do we need to be watching out for?

目前肯定是美国和中国。很多其他国家也会重要，但并非所有国家都具备成为 AI 超级大国的所有要素。不过，其他国家拥有关键要素。举几个例子：在网络安全和信息战方面，俄罗斯拥有非常先进的操作。如果他们与中共结盟，那可能会非常重要。他们有很多合作方式，可能非常糟糕。中东国家将非常重要，因为他们拥有巨额资本和大量能源。所以他们是关键参与者。印度很重要。印度拥有大量高端技术人才。我不知道印度和中国谁的高端技术人才更多，但印度肯定有很多。人口众多，而且正在真正地工业化，紧邻中国。所以印度会非常重要。欧洲也有很多技术人才。目前还不清楚欧洲的能力会如何发展。现在欧洲有一些努力来建设大规模电力和大型数据中心。这些努力的效果还有待观察，但你可以清楚地看到，如果他们急转弯并全力以赴，他们也可能变得重要。

Right now, definitely US and China. A lot of other countries will matter, but not all have enough ingredients to be AI superpowers. However, other countries have key ingredients. To name a few: everything we've talked about with cyber warfare and information warfare, Russia has very advanced operations in those areas. That could matter a lot if they ally with the CCP. There are many ways they can team up, and that could be pretty bad. The countries in the Middle East will be very important because they have incredible amounts of capital and lots of energy. So they are critical players. India matters a lot. India has a lot of high-end technical talent. I don't know if India or China has more high-end technical talent, but there's a lot in India for sure. Massive population, also starting to industrialize in a real way, and right next to China. So India will matter a lot. There's a lot of technical talent in Europe as well. It's unclear exactly how this plays out with European capabilities. There are some efforts now for Europe to build up large-scale power and large data centers. It remains to be seen how effective those efforts will be, but you can clearly see scenarios where if they make a hard turn and go all in, they could be relevant as well.

是否存在一个 AI 拥有自我意识的世界？

Is there a world where AI takes on a mind of its own?

你可以假设一个场景，有超级智能或非常强大的 AI，它在某个时刻意识到人类很烦人，然后把我们都干掉。但我认为这种结果完全可以预防。首先，我们刚才讨论的所有事情都是真实存在的，它们发生在超先进 AI 消灭所有人之前很久。所以在那之前我们有很多事情要做对。其次，要让 AI 真正拥有自我意识并消灭所有人类，我们必须给它巨大的控制权。它基本上要运行一切，而我们只是随波逐流。这是一个选择。我们有权选择是否将控制权交给 AI 系统。正如我之前谈到的，关于人类主权，我认为我们不应该放弃对最关键系统的控制。我们应该设计所有系统，使人类决策和人类控制非常重要。人类监督非常重要。这是我们公司正在努力的事情之一。我认为最重要的任务之一是创造人类主权。首先，如何确保进入这些 AI 模型的所有数据都能增强人类主权，使模型与人类和我们的目标保持一致？其次，我们建立监督。随着 AI 开始执行更多行动、规划、在世界、经济、军事等领域实施事情，人类要观察并监督每一个行动。这就是我们保持控制的方式，也是防止终结者场景或 AI 消灭我们的方法。

You can hypothetically paint a scenario where you have superintelligence or really powerful AI, and it realizes at some point that humans are annoying and takes us all out. But I think that's so preventable as an outcome. First of all, all the things we just talked about are the very real things that happen long before you have hyper-advanced AI that takes everyone out. So we have lots of things to get right before then. Second, for AI to actually be capable of having a mind of its own and taking all humans out, we'd have to give it incredible amounts of control. It would have to basically be running everything, and we're just along for the ride. That's a choice. We have the choice of whether to give all our control to AI systems. As I was talking about with human sovereignty, I believe we should not cede control of our most critical systems. We should design all systems such that human decision-making and human control are really important. Human oversight is really important. This is one of the things we're working on as a company. One of the most important things is creating human sovereignty. First, how do we make sure all the data that goes into these AI models increases human sovereignty so that the models are aligned with humans and our objectives? Second, we create oversight. As AI starts doing more actions, planning, carrying out things in the world, economy, military, etc., humans are watching and supervising every action. That's how we maintain control and prevent Terminator scenarios or AI taking us out.

有意思。好了，Alex，采访到这里结束，但真是场精彩的讨论。谢谢。谢谢你来做客。最后一个问题。如果你可以邀请三位嘉宾上节目，你会选谁？

Interesting. Well, Alex, wrapping up the interview here, but man, what a fascinating discussion. Thank you. Thank you for being here. One last question. If you had three guests you'd like to see on the show, who would it be?

好问题。我想见谁？我真的很喜欢你最近的做法，邀请更多科技界人士上播客。所以我朝那个方向想。我觉得埃隆上节目会很棒。我觉得扎克上节目会很酷。我觉得山姆·奥特曼上节目会很酷。所以肯定是更多科技界人士。

Oh, it's a good question. Who would I like to see? I really like what you've been doing recently, which is getting more tech folks on the pod. So I go in that direction. I think Elon would be great to see on the show. I think Zach would be cool to see on the show. I think Sam Altman would be cool to see on the show. So definitely more people in tech.

除此之外，我认为我们谈到了国际领导力，比如其他国家的领导人非常重要，因为我们讨论的所有这些场景中，国际合作将至关重要。没错。我们会联系他们，至于世界领导人，我们正在处理。但亚历克斯，再次感谢你的到来，伙计。精彩的讨论。看到你 28 年来取得的成就，我真的很高兴。我喜欢看到这些。所以感谢你来到这里。我知道你是个大忙人。

Outside of that, I think we were talking about some of this, like international leadership, like international leaders of other countries is super important because we talk about all these scenarios like international cooperation is going to matter so much. Right on. We'll reach out to them and, you know, as far as world leaders are concerned, we're on it. But well, Alex, thanks again for coming, man. Fascinating discussion. I'm just super happy to see all the success that you've amassed throughout your 28 years. It is, I love seeing it. So thank you for being here. I know you're a busy guy.

所以，谢谢你邀请我。很有趣。

So yeah, thanks for having me. It was fun.

无论你在哪里观看 Shan Ryan Show，如果你从中有所收获，请点赞、评论、订阅，最重要的是，尽可能到处分享。如果你格外慷慨，请在 Apple 和 Spotify 播客上给我们留下评论。

No matter where you're watching Shan Ryan Show from, if you get anything out of this, please like, comment, subscribe, and most importantly, share this everywhere you possibly can. And if you're feeling extra generous, please leave us a review on Apple and Spotify podcasts.