AI 该搬到太空去吗？Should A.I. Move to Space?

Hey there, it’s Stephen Dubner. Let’s start with a puzzle. I’m going to name some industries, some varied industries, and you have to figure out what they all have in common. Okay? Here’s the list: agriculture, ride-share services, global banking, meteorology, military communication, insurance, shipping logistics, disaster management, broadcasting. So what do they all have in common? All of these industries, and many more, are heavily reliant on infrastructure that is based in space. Which means that space-based infrastructure is an industry of its own. Here is Will Marshall:

大家好，我是 Stephen Dubner。我们先来猜个谜。我会说出一些行业，一些五花八门的行业，你们来找出它们有什么共同点。准备好了吗？清单如下：农业、网约车服务、全球银行业、气象学、军事通信、保险、航运物流、灾害管理、广播。那么，它们有什么共同点呢？所有这些行业，以及更多其他行业，都严重依赖基于太空的基础设施。这意味着基于太空的基础设施本身就是一个产业。下面有请 Will Marshall：

Everything from the rockets, the satellites, communications is a big one. You add all of that up, it’s around 300 billion a year.

从火箭、卫星，到通信，通信是个大头，所有这些加起来，一年大约有 300 billion。

Marshall got his Ph.D. in physics at Oxford, where he conducted experiments in quantum superposition under Dirk Bouwmeester and future Nobel Prize winner Roger Penrose. But Marshall’s true love was space. So he went to work for NASA at its research center in Mountain View, California. Among other things, NASA has been developing and managing satellites for decades.

Marshall 在牛津大学获得了物理学博士学位，师从 Dirk Bouwmeester 和未来的诺贝尔奖得主 Roger Penrose，从事量子叠加实验。但 Marshall 真正热爱的是太空。于是他去了 NASA，在加州山景城的研究中心工作。几十年来，NASA 除了其他事情，一直在研发和管理卫星。

We actually started a project called PhoneSat because my colleague at NASA kept on holding up his phone, because it was pretty new at the time, first iPhone had come out, and they were like, these new smartphones have a lot of what you need in a satellite. They have batteries, and they have sensors, and they have cameras, and they have fast processors. They even have GPS to know where they are. They are kind of incredible, and it’s all stuffed in a little box, and it costs $500 bucks.

我们实际上启动了一个名为 PhoneSat 的项目，因为我在 NASA 的同事老是举着手机，当时手机还很新，第一代 iPhone 刚出来，他们说，这些新的智能手机具备卫星所需的很多功能。它们有电池、传感器、摄像头和快速处理器。它们甚至有 GPS 来定位。它们简直不可思议，所有东西都塞在一个小盒子里，而且只要 $500 美元。

This led Marshall to ask himself an obvious question:

这让 Marshall 问了自己一个很自然的问题：

Why are we spending $500 million on satellites? What are the extra six zeros doing for us?

我们为什么要花 $500 million 造卫星？多出来的那六个零给我们带来了什么？

Marshall eventually left NASA to help run a startup that made small, inexpensive satellites, which they called “doves.” They had thin rectangular wings and a big central camera for an eye. Like any good startup, they made their first prototype in a garage. Or, if you are Will Marshall, in a garage:

Marshall 最终离开了 NASA，去帮助运营一家制造小型廉价卫星的初创公司，他们把这些卫星称为"鸽子"。它们有薄薄的长方形翼板和一只巨大的中央摄像头作为眼睛。像所有优秀的初创公司一样，他们的第一个原型是在车库里造出来的。或者，对 Will Marshall 来说，是在车库里：

Literally in a garage, We leveraged smartphones. And we plonked the satellite down on the desk of the venture capitalist that we went to see as opposed to have any PowerPoint deck. That was certainly a shock to people. And enough people were crazy enough to give us a check.

真的就是在车库里。我们利用了智能手机。然后我们直接把卫星放在我们要去见的那位风险投资人的桌子上，而不是搞什么 PowerPoint 演示。那确实让人很震惊。而且有足够多的人疯狂到愿意给我们开支票。

Those checks helped fund what became Planet Labs, a publicly traded company that is today worth around $10 billion. Their vast array of tiny satellites constantly scan the Earth’s surface and provide data for all sorts of public agencies and private firms. So that has been a success for Will Marshall, but he wants to go way beyond that. He thinks the next frontier for space technology is to provide computing power for A.I. Today on Freakonomics Radio: my Freakonomics friend and co-author Steve Levitt explores how a breakthrough technology is entering its space age. That’s in four, three, two, one, liftoff.

那些支票帮助资助了后来的 Planet Labs，这是一家上市公司，如今价值约 $10 billion。它们庞大的微型卫星群不断扫描地球表面，为各种各样的公共机构和私营公司提供数据。这对 Will Marshall 来说算是成功了，但他想要的远不止于此。他认为太空技术的下一个前沿，是为人工智能提供计算能力。今天在 Freakonomics Radio：我的 Freakonomics 朋友兼合著者 Steve Levitt 将探讨一项突破性技术如何进入其太空时代。倒计时，四，三，二，一，点火升空。

Hi, I’m Steve Levitt. From 2020 to 2025, I had my own podcast on the Freakonomics Network called People I (Mostly) Admire. And I had a guest on my show in 2024 who changed the way I thought about artificial intelligence. Since then, A.I. has only become more relevant — and so has this guest’s work.

嗨，我是 Steve Levitt。从 2020 年到 2025 年，我在 Freakonomics 网络上有自己的播客，叫 People I (Mostly) Admire。2024 年，我节目里来了一位嘉宾，他改变了我对人工智能的看法。自那以后，AI 只会变得更具相关性——这位嘉宾的工作也是如此。

I’m Blaise Aguera y Arcas. I am a vice president and fellow at Google and the C.T.O. of technology and society. So just this morning I’ve been using A.I. to work on some artificial life simulations that I have been playing with, so it’s a combination of coding and paper writing.

我是 Blaise Agüera y Arcas。我是 Google 的副总裁兼院士，也是技术与社会的 C.T.O.。就在今天早上，我一直在用 AI 做一些我一直在玩的人工生命模拟，这既是编程也是写论文。

Blaise has had an unusual life. Let me give you one example. One of his first jobs was with a U.S. Navy research center, where he reprogrammed the software that stabilizes aircraft carriers to reduce the seasickness of the crew. At the time, he was only 14 years old. When I talked to him in 2024, we had a conversation about how you can know if AI is truly intelligent.

Blaise 的人生非同寻常。我给你们举一个例子。他最初的一份工作是在美国海军的一个研究中心，重新编写稳定航空母舰的软件，以减少船员的晕船情况。那时，他才 14 岁。2024 年我跟他交谈时，我们进行了一场关于如何判断 AI 是否真正智能的对话。

The moment you start to have something that behaves intelligent, the question, like, “Well, but is it really intelligent?” is an interesting one to ask. A lot of philosophers talk about this as the philosophical zombie problem. Could it be that something could behave like a person, but is dead on the inside, has no inner life? There’s no there there; there’s nobody home. Is that meaningful? But the trouble with the whole philosophical zombie question is that it’s almost like a frontal attack on the whole idea of doing science.

And this is one of the things that led Alan Turing to propose the Turing test. He was basically saying, look, if all of the tests check out, then that is your answer. And I’m kind of with Turing on that one. It can be a shock. Maybe intelligence is this much more general accessible faculty. It doesn’t have to depend on any particular biological substrate. You know, it doesn’t have to be a person in the sense that we’ve always understood persons. But I’m sure that we’re going to be debating this one for a long time to come because it really cuts to the quick of what we consider to be our special sauce as humans.

一旦你开始拥有行为上表现出智能的东西，“嗯，但它真的智能吗？”这个问题就有意思了。很多哲学家把这说成是哲学僵尸问题。有没有可能，某个东西行为像人，但内在是死的，没有内在生命？里边空空如也，没有人在家。这有意义吗？但整个哲学僵尸问题的麻烦在于，它几乎像是对整个科学研究理念的正面攻击。这也是促使 Alan Turing 提出图灵测试的部分原因。他基本上是在说，瞧，如果所有测试都通过了，那这个就是你的答案。在这一点上，我有点赞同 Turing。这可能让人很震惊。也许智能是一种更普遍、更容易获得的能力。它不必依赖于任何特定的生物基质。你知道，它不必是我们传统意义上理解的人。但我确信我们还会为此争论很长时间，因为它真的触及了我们自认为的人类"独门秘方"的要害。

So when Aguera y Arcas told me he had founded a new A.I. project at Google called the Paradigms of Intelligence team, I wanted to know more.

所以，当 Agüera y Arcas 告诉我他在 Google 创立了一个名为 Paradigms of Intelligence 团队的新 AI 项目时，我想了解更多。

At its heart, it’s an A.I. research team. But we have a few philosophers on the team. We have neuroscientists on the team. We have people from other disciplines, sociology. And part of the reason for that is that I’ve always thought that understanding intelligence generally is key to making progress with A.I., and that A.I. in turn can help us to understand ourselves. I see the study of intelligence as a connected field, both in its analysis of how we work and in its attempts to synthesize intelligence outside the human brain.

本质上，这是一个 AI 研究团队。但我们团队里有几位哲学家。我们有神经科学家。我们有来自其他学科的人，比如社会学。这么做的部分原因是，我一直认为，从总体上去理解智能是推动 AI 进步的关键，而 AI 反过来又能帮助我们理解自身。我把智能研究看作一个互联的领域，既包括分析我们自身如何运作，也包括试图在人脑之外合成智能。

So mostly when people say they work on A.I. at a place like Google, it means you’re part of a huge team that’s building enormous models and training them on endless amounts of data for the next release. But that’s not what your team is doing. You’re working on more philosophical issues, foundational issues. Could you give some examples of the kinds of specific problems that your team is tackling?

所以大多数时候，当人们说在 Google 这样的地方做 AI，意思是你是某个庞大团队的一部分，这个团队在构建巨大的模型，用无尽的数据训练它们，为下一个版本做准备。但你的团队做的不是这个。你们在研究更哲学性的问题、更基础的问题。你能不能举几个例子，说一下你们团队正在解决的具体问题？

Some of the problems are philosophical, but they’re also very practical. A lot of the breakthroughs in A.I. have come from people who study neuroscience and study the brain and are working at the intersection of those things. So we’re very much in that tradition, too. Examples of some of the more heady problems that we’re looking at now are not just how individual intelligence agents work, but how multiple agents in interactions with each other lead to larger collective intelligences.

And even how, when you look at what we think of as single models, they look internally like societies of sub-agents. We think of this as the social intelligence hypothesis, meaning that intelligence is inherently social, whether you think about it at the scale of societies or even at the scale of brains.

有些问题确实是哲学性的，但同时也非常实际。AI 领域的很多突破都来自研究神经科学和大脑的人，他们正是在这些学科的交叉点上工作。所以我们很大程度上也延续了这个传统。我们目前正在研究的一些比较深奥的问题，比如，不只是单个智能体如何运作，而是多个智能体在相互交互中如何形成更大的集体智能。甚至还包括，当你审视我们认为的单一模型时，它们内部看起来就像是由子智能体组成的社会。我们把这种想法称为社会智能假说，意思是智能本质上是社会性的，无论你是在社会的尺度上还是在单个大脑的尺度上看待它。

So an example of what you might do would be to create a set of five or six different A.I. agents who were taking on different roles or personalities. One is the skeptic, one is the optimist, or what specifically are you talking about?

那么，你具体会做的事情可能是，创建一组五六个不同的 AI 智能体，让它们扮演不同的角色或个性。一个是怀疑论者，一个是乐观主义者，你具体指的是什么？

Yes, so we can definitely, and we do do those sorts of experiments, where we take agents and we put them in societies or institutions or conversational structures with each other, and you can ask questions like if they all are set up to agree with each other versus set up to be misaligned with each other, to disagree with each other, which of those kinds of setups leads to better problem-solving of the collective, and it turns out there that it’s better if you have them disagreeing with each other than if they’re all preaching to the choir, as it were.

But then you can also ask the same question of an individual model, and this is a little bit mind-bending, but if you look at the chain of thought, which is to say the inner monologue of an individual reasoning model, then one of the things you find in there is that there are actually voices or characters that have been developed that do exactly the same thing inside those models. It’s like the model is talking to itself, arguing with itself. Kind of like Gollum versus Smeagol in Lord of the Rings or something, but often with more characters.

And you get that behavior just from training reasoning models to do a good job of reasoning. So I think that there is a really profound lesson in this. And it’s a practical one in terms of how to build A.I. models. But it’s also a philosophical one in the sense that it suggests that intelligence is fundamentally social. Both at the scale of brains and societies.

是的，我们完全可以做，也确实在做这类实验，我们把智能体放进社会、机构或彼此的对话结构中，你可以问这样的问题：如果把它们都设置为彼此认同，还是设置为彼此不对齐、彼此不同意，哪一种设置会带来更好的集体问题解决能力？结果发现，让它们互相不同意比让它们对着唱诗班说教更好。但你也可以对单个模型问同样的问题，这有点令人费解，但如果你看一下思维链，也就是单个推理模型的内心独白，你会发现其中确实已经形成了各种声音或角色，它们在这些模型内部做着完全相同的事情。就像是模型在自言自语，跟它自己争论。有点像《指环王》里的咕噜对战史麦戈之类的，但通常角色更多。而你只要训练推理模型做好推理工作，就能得到这种行为。所以我认为这其中有一个非常深刻的教训。它对于如何构建 AI 模型来说是一个实际教训，但它同时也是哲学性的，因为它表明智能从根本上说是社会性的，无论是大脑的尺度还是社会的尺度。

I don’t know anything about neuroscience, but I know my own experience is that my brain works exactly like you’ve just described. I’ve got four or five different parts of my brain that stake out different territory and some are risk-taking and some are cautious and some skeptical. And so I definitely feel like my own brain is in no sense unified. It’s some kind of a competition for, for time on the stage. What I’ve never been able to figure out at all, though, is who is the “it” that’s behind the scenes, who’s deciding who the winner is among these different voices.

我对神经科学一无所知，但我自己的经验是，我的大脑运作方式跟你刚刚描述的一模一样。我大脑里四五个不同的部分，各自圈定不同的领地，有的爱冒险，有的很谨慎，有的充满怀疑。所以我绝对能感觉到，我自己的大脑在某种意义上根本不统一。那是某种竞争，争夺登台发言的时间。但是，我一直完全搞不懂的是，幕后那个"它"到底是谁，是谁在决定这些不同声音中哪个胜出。

Well, that is the place where I think a lot of folk psychology and philosophy also goes wrong in assuming that there is a homunculus, you know, that there’s a mini-Steve, you know, the mini-Blaise you know inside our brains, when in fact, what you are is that collection of voices. I’m really excited to hear that that’s how you think of yourself, how you model yourself internally. It’s how I think of myself, too, but many people think of themselves as being very unified, as having just one self.

But we know from neuroscience and we know from psychology that your perspective is closer to what is really going on if you, for instance, do radical brain surgery on people and split their brains in half, which used to be done quite a lot as a last-ditch surgery for addressing epilepsy. You can see that the two halves, the two hemispheres of the brain are working independently of each other, and yet people’s sort of impression of themselves remains unified. So, it’s kind of like every part of the brain is working as a member of a team, they all know that they’re on Team You, and they’re modeling themselves as a whole, and yet also, in a kind of competition, as you’re describing, for who says the thing that is going to come out of the mouth.

And that’s how neural nets work, too. You know, artificial neural nets have these softmax layers, as they’re called, which are essentially an internal competition for which part of the network will get to emit the behavior.

嗯，这正是我认为很多大众心理学和哲学走偏的地方，它们假设存在一个小矮人，你知道，我们大脑里有一个迷你版的 Steve，迷你版的 Blaise，而事实上，你就是那个声音的集合体。听你说这就是你对自己、对内心自我模型的理解，我真的很兴奋。我也是这么看待自己的，但很多人认为自己是非常统一的，只有一个自我。不过我们从神经科学和心理学中知道，你的视角更接近真实情况。比如，如果你对人们做极端的脑部手术，把大脑切成两半——这过去经常作为治疗癫痫的最后一招——你就能看到，大脑的两个半脑在独立运作，但人们对自我的印象却依然保持统一。所以，这有点像大脑的每个部分都作为团队的成员在工作，它们都知道自己在"你"这个团队里，并且把自己作为一个整体来建模，但同时也存在你所描述的某种竞争，争夺谁能说出即将从嘴里出来的那句话。神经网络也是这么运作的。你知道，人工神经网络有所谓的 softmax 层，它本质上就是一种内部竞争，看网络的哪一部分可以发出行为。

Now, you published a book last year based in part on the work that you and your team are doing at the Paradigms of Intelligence team. It’s called, What is Intelligence?: Lessons From A.I. About Evolution, Computing, and Minds. It’s full of a lot of very bold ideas. Let me just hit on a few of those. For instance, you argue that the primary function of intelligence, and life itself, is prediction. What do you mean by that?

你去年出版了一本书，部分基于你和团队在 Paradigms of Intelligence 团队所做的工作。书名是《什么是智能？：从 AI 看进化、计算和心智》。书里有很多非常大胆的想法。我只挑其中几处来说。比如，你论证说智能，乃至生命本身的首要功能是预测。你这话是什么意思？

This is in many ways an old idea. One of the long-standing critiques of A.I. models is that they’re just auto-complete on steroids, that they are just predicting the next token. I always kind of thought in my heart, as it were, that that could not be the secret of intelligence. Surely it’s not just predicting the next token. So I was as shocked as anybody when we made really large scale next-token predictors, and they started to get the answers right to hard mathematical word problems and write poetry and all this kind of stuff.

I found that very surprising as far as my intuition went, but if you really start to look at what it means to predict, the surprise dissipates a bit. The reason we’ve got brains is because we live in a complex world, and our actions can have effects on that world and our own futures that can either hurt us or help us. And so, in order to be able to distinguish between actions that will lead to good places or to bad places, that’s exactly what you have to do, you have to predict.

Not just predict the world in your absence, but do what statisticians call conditional prediction, meaning, predict the world conditional on action A and action B, and then decide which prediction you like better. In some sense, it’s almost a tautology, like of course that’s what brains are there for. If you couldn’t act to change your own future, then there would really be no point in having a brain. And if you couldn’t predict the effects of your actions, then there would also be no point in having a brain.

这从很多方面来说都是一个老想法。长期以来对 AI 模型的一个批评是，它们不过是打了激素的自动补全，只是在预测下一个 token。我心里其实一直觉得，这不可能是智能的奥秘。智能肯定不只是预测下一个 token。所以，当我们做出大规模的下一个 token 预测器，它们开始能正确解答复杂的数学文字题、写诗、做所有这些事情时，我和所有人一样震惊。根据我的直觉，这非常令人惊讶，但如果你真正开始审视“预测”意味着什么，这种惊讶就会消散一些。我们之所以有大脑，是因为我们生活在一个复杂的世界里，我们的行动会对这个世界和我们自己的未来产生影响，这些影响可能会伤害我们，也可能帮助我们。因此，为了能够区分哪些行动会导向好的结果、哪些会导向坏的结果，这正是你必须要做的事——你必须预测。不仅仅是预测没有你参与的世界会怎样，还要做统计学家所说的条件预测，也就是说，在采取行动 A 和行动 B 的条件下分别预测世界会怎样，然后决定你更喜欢哪个预测。从某种意义上说，这几乎是个同义反复——当然，大脑就是为此而存在的。如果你无法通过行动改变自己的未来，那长个大脑就真的毫无意义。而如果你无法预测自己行动的后果，那长个大脑同样也毫无意义。

What people might find confusing about this, though, is that we’ve built these neural nets, these A.I. programs, and it is bizarre that the objective is to predict the next word. And I think what you’re arguing is that you can’t predict the next word well at all unless you have a pretty complete model of how the entire world works, because you don’t know. (Agüera y Arcas: Exactly. If you don’t have the context, you’d just be terrible at that. And at least my own experience, as I’ve interacted more and more with A.I., is to really reconsider who I am and to think that, yeah, that kind of, what A.I. is doing feels a lot like what I do, that if I’m really tired or maybe I had too much to drink, I start making the exact same kind of mistakes that A.I. makes when it doesn’t exactly understand what’s going on.

It’s been interesting. I wouldn’t have thought that I would come to see myself as being like A.I., but I really have.

但人们可能对此感到困惑的是，我们构建的这些神经网络、这些 AI 程序，它们的目标是预测下一个词，这很奇怪。而我想你的论点是，除非你对整个世界的运作方式有一个相当完整的模型，否则你根本无法很好地预测下一个词，因为你不知道的事情太多了。（Agüera y Arcas：完全正确。）如果你没有上下文，你的预测就会非常糟糕。至少根据我自己的经验，随着我与 AI 的互动越来越多，我真的开始重新思考自己是谁，并意识到，对，AI 在做的事情，很大程度上跟我做的事情很像。如果我特别累，或者可能喝得有点多，我就会犯和 AI 在没完全搞懂情况时一模一样的那种错误。这很有意思。我原本不会想到我会渐渐把自己看作像 AI 一样，但我确实已经这么觉得了。

Yeah, that’s exactly what I meant when I wrote the title of the book, what A.I. is teaching us about intelligence generally, including about ourselves. So I think that’s real. It doesn’t mean that you’re just babbling, that you are just saying stuff. In a way, when people say, oh, are people really just next-token predictors, the error is in the word “just,” because a lot is implied by that. And in particular, what I think trips a lot of people up is that they imagine that the mind that is doing that predicting or the process that is doing that predicting is outside the thing that is being predicted.

But the key thing to understand there is that the sequence of events that you’re predicting include what you are going to do. They include your own actions. They are part of that stream of events that you experience too. And what that means is that in order to do a good job of predicting the future, any future that involves your own actions, you actually have to be predicting yourself as well, modeling yourself. And I actually think that that’s where consciousness comes from.

Not consciousness in some kind of mystical sense, but in the sense that knowing what it’s like to be you and knowing what you’re likely to do in the future and what others think about you and what you think they think and so on. You know, I think that is the functional essence of what consciousness really is.

是的，这正是我给这本书起名的用意——AI 在教我们什么是智能，广义上的智能，包括我们自身的智能。所以我认为这是真实的。但这并不意味着你只是在胡言乱语，只是在随口说些东西。当人们说“哦，人真的只是预测下一个 token 的东西吗”时，某种意义上，错误出在“只是”这个词上，因为这个词蕴含了太多东西。具体来说，我觉得让很多人卡住的一点是，他们想象那个进行预测的心智或过程，是独立于被预测的事物之外的。但这里需要理解的关键是，你所预测的这一连串事件，包含了你将要做什么。它们包括了你自己的行动。这些行动也是你所经历的事件流的一部分。这意味着，为了做好预测未来的工作，任何一个涉及你自己行动的未来，你实际上也必须预测你自己，模拟你自己。而且我真的认为，这就是意识的来源。不是那种神秘意义上的意识，而是指：知道作为你是什么感觉、知道你将来可能会做什么、知道别人怎么看你、知道你觉得他们怎么看，等等。你知道，我认为这就是意识其功能性的本质。

Now someone might get the idea from our conversation so far that you’re a dreamer or a philosopher, but that’s so far from the truth. You’ve got a long track record of building things that really work. And one of the projects that your Paradigm of Intelligence team is working on now is called Project Suncatcher. And if you pull it off, it will be perhaps the most ambitious engineering undertaking of all time. Would you say that’s a fair assessment of the degree of difficulty or am I exaggerating?

到现在为止，有人可能从我们的对话中得出印象，觉得你是个梦想家或哲学家，但这与事实相去甚远。你有着长期的、真正做出了实际成果的工程履历。而你的“智能范式”团队目前正在进行的其中一个项目，叫做 Project Suncatcher。如果你们能做成，它可能会成为有史以来最雄心勃勃的工程壮举。你觉得这个难度评估公允吗，还是我说得夸张了？

Yeah. Yeah, I think it is a fair assessment, actually. As much as it makes me very afraid to say that, yes. I think that’s correct.

对。对，我觉得这个评估其实很公允。尽管这么说让我非常害怕，但没错。我认为是对的。

There’s a lot of controversy around A.I. data centers and their power consumption. But one thing that’s not disputed is that that consumption is growing. The International Energy Agency expects that about half the increased demand for electricity in the U.S. through 2030 will come from data centers.

围绕 AI 数据中心及其能耗，存在很多争议。但有一点没有争议，那就是能耗在增长。国际能源署预计，到 2030 年，美国大约一半的新增电力需求将来自数据中心。

I was really worried about this a few years ago, because even though the number is still small, how much electricity is being used today on A.I., that is still a very small number. But it is growing so quickly. And so if you think about that exponential growth over the course of the next 10, 20 years, then you really start to get worried. My trajectory emotionally about all of this is that I started off really concerned about it, and then I kind of reassured myself that we were working on a lot of things that I believe are going to be able to gain us at least a factor of 1,000 in efficiency.

And that reassured me that it’s gonna be fine. We’re gonna solve this problem through increased efficiency.

几年前，我对此非常担忧，因为尽管这个数字仍然很小，也就是如今用于 AI 的电力消耗量，仍然是一个非常小的数字，但它的增长速度太快了。所以，如果你想想未来 10 年、20 年这种指数级增长，你就会真正开始担忧。我对这件事的情绪轨迹是，一开始我真的很担心，然后我某种程度上说服了自己，说我们正在做的很多事情，我相信能够将效率至少提升一千倍。这让我安心，觉得会没事的。我们会通过提高效率来解决这个问题。

So we wouldn’t need that much electricity in the end because even though we’ll build bigger and bigger models and use A.I. more and more, we’ll get so good at saving the energy that it won’t really be a problem.

所以最终我们不需要那么多电力，因为尽管我们会构建越来越大的模型，并且越来越多地使用 AI，但我们在节能方面会做得非常好，以至于这真的不会是个问题。

That it’ll be a wash, that’s right.

这会抵消掉，没错。

Then something happened.

然后发生了一些事。

Right. So that was the first and second parts of my emotional journey. But then there was a part three, which is that a factor of a thousand in efficiency is, in an exponential landscape, that only buys you maybe a decade. And that sounds crazy to say, that a factor a thousand only buys a decade, but I mean, if you just look at Moore’s law, you can see how many factors of 1,000 we’ve gotten in amount of computing since the 1940s. And it really is only a decade. And I see no reason to believe that the demand for A.I. is somehow going to saturate in the next 10 years and will be done.

So that’s what really got me thinking about the supply side of energy, not just how much energy A.I. is using, but how we could think about obtaining more energy for A.I.

没错。那就是我情绪历程的第一和第二部分。但还有第三部分，就是在一个指数级的图景里，一千倍的效率提升，大概只能为你争取十年时间。说一千倍只能争取十年听起来很疯狂，但我的意思是，你只要看看摩尔定律，就能看到自 1940 年代以来，我们在计算能力上已经翻了多少个一千倍了。而每次跨越确实也就十年左右。而且我看不出有任何理由认为，人们对 AI 的需求会在未来十年内饱和、就此结束。所以，这让我真正开始思考能源的供给侧——不只是 AI 用了多少电，而是我们怎么为 AI 获取更多能源。

And you’re obviously not the only one thinking about this. All the big A.I. companies have been scrambling to find reliable power sources. (Agüera y Arcas: Yes. Amazon, Meta, Google, Microsoft, they’re all making big investments in nuclear power in various forms. Microsoft even has plans to restart Three Mile Island, which seems like one of the odder choices. But this all has to strike people as being pretty surprising given that nuclear power has been mostly out of favor for a long time, at least in the U.S.

Another option would be to look at solar power, which has been expanding like crazy. What are the pros and cons of solar power when it comes to these massive data centers?

显然，不是只有你在想这件事。所有大型 AI 公司都在争相寻找可靠的电力来源。（Agüera y Arcas: 是的。）Amazon、Meta、Google、Microsoft，它们都在以各种形式大举投资核能。Microsoft 甚至计划重启三里岛，这看起来是其中比较奇怪的选择之一。但这一切肯定让人很意外，因为核能长期以来大多不受欢迎，至少在美国是这样。另一个选择是看太阳能，太阳能的扩张一直非常猛。对于这些巨型数据中心来说，太阳能的优缺点是什么？

I don’t think any of these things should not be pursued. We should be pursuing all of them. But the trouble, of course, with solar power on Earth is that the sun is only up half the day. And also, they have to be oriented toward the sun. There is cloud cover. There’s the changing angle of the sun, which means that you get less when it’s not overhead. So intermittency is really the big issue, and the fact that battery technology is still not anywhere close to where it needs to be in order to smooth that power curve from solar.

So at least on Earth, it won’t be able to fulfill the entire demand until or unless we make really radical changes in how we store power. And even if we do that, we also need to think about how much of the Earth’s surface we’re willing to cover with solar panels. I think we can certainly afford to cover a lot more of it than we have today, but there is a limit, which is not just human habitation and so on, but the rest of nature. The whole planet is solar powered, and so there is ultimately going to be a zero-sum game there.

我不认为这些事情里有哪一样不该去推进。我们应该全部推进。但地球上的太阳能当然有一个问题，就是太阳一天只出来一半时间。此外，太阳能板必须朝向太阳。还有云层遮挡。太阳角度在不断变化，这意味着太阳不在头顶正上方时，发电量就会减少。所以间歇性是真正的大问题，而且电池技术远远还达不到能抚平太阳能发电曲线的水平。所以至少在地球上，太阳能无法满足全部需求，除非我们在储能方式上做出真正彻底的变革。而且即便做到了，我们也得想想，我们愿意用太阳能板覆盖多少地球表面。我觉得，我们肯定可以覆盖比现在多得多的面积，但存在一个限度，这个限度不只是人类居住等等问题，还有自然界的其他部分。整个地球就是靠太阳能驱动的，所以这最终会是一场零和博弈。

So, you’ve got a different solution in mind and it’s one that at first sounds completely ridiculous, I have to say. What is Project Suncatcher?

所以，你心里有另一个解决方案，而且我不得不说，这个方案乍听起来完全荒谬。什么是 Project Suncatcher？

Basically the concept is that you move A.I. into space and you use solar power, but rather than solar power on Earth, you have orbiting solar panels that are in sunlight almost all the time, in something called a sun-synchronous orbit. These are orbits that go around the Earth in such a way that the solar panel is facing the Sun. So those sun-synchronous orbits allow a solar panel to gather about eight times the energy of a solar panel on the ground because there’s no nighttime and there’s no atmosphere.

基本概念是，把 AI 搬到太空中去，并使用太阳能，但不是用地球上的太阳能，而是用几乎一直处于阳光照射下的轨道太阳能板，放在一种叫做太阳同步轨道的轨道上。这些轨道以某种方式绕地球运行，使太阳能板始终面向太阳。因此，太阳同步轨道能让一块太阳能板收集的能量大约是地面太阳能板的八倍，因为没有夜晚，也没有大气层。

So solar power works really great in space. But the problem we’ve always had is that beaming energy, in the same way that moving people or materials through long distances, is not very efficient. So what good does it do you to have this energy in space?

所以太阳能在太空中确实非常好用。但我们一直面临的问题是，能量传输，就像长距离运送人员或物资一样，效率不高。那么，太空中有了这些能量又有什么用呢？

The good that it does you is if you move the actual computation to space also. So if you want your farm-to-table distance, as it were, to be as short as possible, then the way to solve for that is by putting the data centers in space with the solar panels.

用处就在于，如果你把实际的计算工作也搬到太空中去。如果你想让你从农场到餐桌的距离——可以这么说——尽可能短，那么解决的办法就是把数据中跟太阳能板一起放在太空中。

I’m just trying to imagine this, because I’ve seen these data centers and they are not small buildings. I mean, these are some of the biggest buildings that we ever built. How are you gonna do that in space?

我正试着想象这个场景，因为我见过那些数据中心，它们可不是小建筑。我的意思是，这是我们建造过的最庞大的建筑之一。你怎么在太空里做这件事？

So there’s a short-term answer and there’s a longer-term answer. The short-term answer is that, you know, of course a data center is a giant building and we can’t be talking about literally moving those buildings up into space. This is not the Death Star. It’s very energetically costly to lift mass into space. So none of this makes sense unless you think, (a), that it’s going to continue to get cheaper to move mass into orbit, as it has been. There’s been an incredible decline in cost over the last 10 years, thanks in large part to SpaceX.

And we believe that will continue to be the case for some years to come. So you have to believe in declining costs to launch. But you also have to believe that you can do an orbiting data center in a much more physically lightweight way than a terrestrial data center.

这有一个短期答案和一个更长期的答案。短期答案是，你当然知道，一个数据中心是一栋巨型建筑，我们不能说要真的把那些建筑搬上太空。这不是死星。把质量送进太空的能源成本非常高。所以这一切都不成立，除非你认为，(a) 把质量送入轨道的成本会持续降低，就像过去那样。过去十年成本下降得非常惊人，很大程度上要归功于 SpaceX。而且我们相信未来一些年这种情况还会持续。所以你必须相信发射成本会下降。但你还必须相信，你可以用一种比地面数据中心在物理上轻得多的方式来建造轨道数据中心。

So what would your data centers look like? They’re not buildings, they’re something very, very different.

那么你的数据中心会是什么样子？它们不是建筑，是非常非常不一样的东西。

No, they’re not buildings. In the beginning, they look like dragonflies. So with giant, but very, very thin solar wings, and with a body in the center where the computing happens, and the body should be as lightweight as possible. Basically, you need as much area as you can for solar collection. You also need area in order to radiate heat. But then you want your mass to be as small as possible. And when you put all of those constraints together, it turns out that you can design satellites using today’s technologies that put you on the right side of the economics, with some pretty modest assumptions, over the next decade.

对，它们不是建筑。一开始，它们看起来像蜻蜓。有巨大但非常非常薄的太阳能翼，中心是进行计算的主体，主体要尽可能轻。基本上，你需要尽可能大的面积来收集太阳光。你也需要面积来散发热量。但你要让质量尽可能小。当你把所有这些约束条件放在一起，就会发现，用今天的技术，基于一些相当温和的假设，在未来十年内，你设计的卫星在经济效益上是可行的。

And so this data center that you’re imagining, it’s not a single satellite.

那么你设想的这个数据中心，它不会是一颗单独的卫星。

No, it’s a swarm.

不会，它是一个集群。

How do you imagine controlling a swarm of satellites that are rocketing through space at many tens of thousands of miles an hour? Can you just give us the broadest view of how you pull this off?

你怎么想象去控制一个以每小时数万英里速度在太空中飞驰的卫星集群？能不能先给我们一个最宏观的概念，这件事你怎么做到？

You have to obviously work through a lot of the details of the orbital mechanics of this. There are also many other things to think about, like safety of the low Earth orbit environment. You want to make sure that you’re not creating space junk, because with massive numbers of satellites, you want to be very careful that collisions are extremely, extremely rare because those are super costly. They create a lot of pollution in low Earth orbit. And you want to be sure that if the worst happens, if there is a collision, it doesn’t result in a catastrophe where we end up with all kinds of junk that makes that extremely valuable real estate around the Earth — that poisons the well for us and for everybody else.

So there are many important considerations. But basically, the swarm has to be self-organizing in a lot of ways. So you need to imagine those spacecraft being intelligent in their own right, not just being all controlled from the ground, you have to imagine them being really robust and communicating with each other via light. So, laser, so-called free space optics, is a really important part of the puzzle. And you can always use radio to communicate with the swarm, but ultimately we’re probably going to be using a combination of radio and light also for ground-to-space communication.

你显然得把这里面大量的轨道力学细节都啃透。还有很多其他事情要考虑，比如近地轨道环境的安全问题。你得确保不会制造太空垃圾，因为卫星数量一庞大，就必须极其小心，让碰撞的概率降到极低极低，因为碰撞的代价太大了。它们会在近地轨道产生大量污染。而且你得确保，如果发生最坏的情况，如果真的发生了碰撞，它不会酿成一场灾难，弄出各种碎片，把地球周围那片极宝贵的空间给污染了，害了我们自己，也害了所有人。所以有很多重要的考量。但基本上，这个卫星集群必须在很多方面做到自组织。所以你得想象那些航天器自身具备智能，不是所有行动都由地面控制，你得想象它们非常皮实，并且通过光来互相通信。所以，激光，也就是所谓的自由空间光通信，是拼图中非常重要的一块。你始终可以用无线电与集群通信，但最终，我们很可能也会用无线电和光的组合来进行天地通信。

How many satellites would be working together in this swarm in your vision?

在你的设想里，这个集群里会有多少颗卫星协同工作？

Well, it’s an exponential. I can’t give you a single number because if you really zoom out over the 50, 70-year time scale that I think you need to in order to understand where this will go, I know that it’s a little bit tricky for me to say that because it sounds like science fiction if I talk about things that are 50 or 70 years out, but that’s what today would have sounded like to somebody in 1945 or the dawn of the computing age. And I think we’re going to undergo a similar exponential explosion.

We’re talking about near-Earth orbits in the beginning, but by the time we get toward the end of this century, I’m sure that we’re gonna be talking about very, very thin orbiting structures, both around the Earth and around the sun, that require scientific notation to describe. So, really, really large numbers.

嗯，这是指数级的。我没法给你一个单一的数字，因为如果你真的把时间尺度拉长，用我认为理解这件事最终走向所必需的那种50年、70年的眼光来看，我知道这么说有点棘手，因为如果我谈论50年或70年后的事，听起来就像是科幻小说，但你想想，今天的一切对1945年或计算机时代黎明时分的人来说，听着也一样像科幻。我认为我们将经历一场类似的指数级爆发。我们一开始谈的是近地轨道，但到了本世纪末，我敢肯定我们谈论的将是极其极其纤薄的轨道结构，环绕地球和环绕太阳的都有，得用科学记数法才能描述。所以，真的是非常非常巨大的数字。

I think most people’s reaction, including mine, when you first hear about this is like wow, that’s that’s kind of crazy that — could that actually work, right? Like it seems like it shouldn’t work.

我想大多数人，包括我自己在内，第一次听到这个想法时的反应都是：哇，这有点疯狂啊——那真的能行吗？对吧？就感觉它好像不应该行得通。

I’m Travis Beals, Senior Director, Product Management, and what I do is I lead Project Suncatcher.

我是Travis Beals，产品管理高级总监，我负责领导Project Suncatcher。

Project Suncatcher is Google’s effort to put solar-powered data centers in Earth’s orbit. Beals runs the project from his home on an island off the coast of British Columbia, Canada, near where he grew up.

Project Suncatcher是Google将太阳能供电的数据中心送入地球轨道的项目。Beals在加拿大不列颠哥伦比亚省海岸外的一个小岛上，在他长大的地方附近，远程管理这个项目。

How did you come to be part of the Paradigms of Intelligence team? Did you volunteer or were you drafted?

你是怎么成为Paradigms of Intelligence团队一员的？是你自愿加入的，还是被征召的？

I volunteered for that. Back in 2016, I started working on Google’s consumer hardware effort. So like Pixel and later Nest and so on. And Blaise was also part of the leadership team. At some point Blaise, I think this would have been early 2024, Blaise had the idea for doing A.I. in space, and he talked to me about that and, I thought it was really interesting. We convened a small group of folks and this was almost all people’s part-time or 20 percent time, to try to figure out why it wouldn’t work, let’s find the reason this is not going to work.

And we couldn’t. We kept finding all these things we’re like, well, maybe this, this won’t work. And then actually, no, no, there’s a way to solve this. There’s no physics reason this is not going to work. And the economics of it seemed like maybe that could work. At some point we got this far enough along that it’s like, okay, we need to start doing this for real.

我自愿加入的。早在2016年，我就开始参与Google的消费硬件业务了。比如Pixel，后来是Nest等等。Blaise也是领导团队的一员。大概在2024年初吧，Blaise有了做太空AI的想法，他跟我聊了这件事，我觉得非常有意思。我们召集了一小群人，这几乎全是大家利用业余时间或20%时间来做的，就是想找出这事为什么行不通，我们要找到它不行的理由。结果我们找不到。我们总是找出各种觉得"嗯，可能这个不行"的地方，然后发现，诶，不对，不对，这个问题有办法解决。从物理上看，没有任何它行不通的理由。而且经济上看，好像也有可能行得通。到某个阶段，项目推进的程度让我们觉得，好了，我们需要开始真刀真枪地干了。

Now I’m wondering, somebody at the top of Google must have said yes to this at some point. How did that conversation go?

现在我好奇的是，Google最高层一定有人在某个节点对这个项目说了"行"。那场对话是怎么进行的？

I don’t know if I can get into everything that went down in that room, but I can tell you a little bit about that. So, obviously this is a big deal trying to present a seemingly crazy sounding idea like this to company leadership. Blaise and I are all set to go present this to Sundar Pichai, the CEO of Google and Alphabet. And so I walk into the room, and then Sergey, who’s one of the co-founders of Google, walks in as well. And I was not expecting that.

我不确定我能不能把那房间里发生的所有事都讲出来，但我可以跟你说一点点。很明显，要向公司领导层推销一个像这样听起来很疯狂的想法，是件大事。Blaise和我都准备好要向Sundar Pichai，也就是Google和Alphabet的CEO，做汇报了。然后我走进房间，结果Sergey，Google的联合创始人之一，也走了进来。这是我完全没有预料到的。

And so you’re nervous? This is high stakes now, suddenly.

所以你当时很紧张？这一下子赌注就变大了。

Yeah, but Sergey was actually also supportive of doing this, which was great. He’s also realistic about, and I think this is true for all the leadership, that this is hard too, right? This is not an easy thing to go do. We’ve gone from this being a thing that sounds crazy and impossible to: Actually this is possible. This could make sense. It’s just going to be really hard.

是啊，不过Sergey实际上也支持我们做这件事，这太棒了。他也很现实，而且我觉得所有领导层都这么认为，就是这事也非常难，对吧？这可不是件容易办成的事。我们已经把这个想法从"听起来疯狂且不可能"，推进到了"实际上，这是可能的。这可能行得通。只是会非常非常难"。

I can’t help whenever I think about Project Suncatcher, the Manhattan Project always pops into my mind. And obviously they have very different objectives, but they have this feeling of trying to apply physics in a way that seems unreal, but making it real. Do you feel like Robert Oppenheimer sometimes?

每当我想到Project Suncatcher时，我总忍不住想起曼哈顿计划。显然，它们的目标截然不同，但它们都有那种感觉：要试着用一种看似虚幻的方式来应用物理，并把它变成现实。你有没有时侯觉得自己有点像Robert Oppenheimer？

I thought you were going to go with the Apollo Project there, I mean that would have been the more straightforward kind of moonshot project here, to make an analogy to.

我以为你要拿阿波罗计划来类比呢，我是说，那样类比更直接嘛，就那种登月式项目。

Could you talk about moving data through space? Because the linchpin of this whole thing is the fact that everything is hard to send through space, except data. We’re really awesome at sending data through space.

你能谈谈在太空中传输数据这件事吗？因为整件事的关键就是，往太空里送什么都难，唯独数据不难。我们在太空里传输数据，那可真是超级擅长。

Yeah, and we keep getting better at that. So the way we send lots of data around on Earth, we usually do that through light in the infrared, through optical fiber. And if you want to send data around in space between satellites, kind of awkward to imagine stringing fiber between satellites. But it turns out you don’t need to. You can just beam the lights directly from one satellite to another. In a sense, that can actually work even better through free space than through glass.

There’s two reasons behind that. The first reason is that the speed of light through vacuum is faster than through glass. And the second reason is you can use more optical bandwidth, right? So I mentioned that when you’re sending light through an optical fiber, you do it in the infrared and that’s because outside of these narrow ranges, impurities in the glass will absorb the light. So you’re limited in how much optical bandwidth you have through an optical fiber. If you want to send light through space, you could, in principle, use the entire infrared and visible spectrum, and that’s about 30 times more spectrum.

So there’s hard engineering problems here, but if you solve those, in principle, sending data through space can be even better than sending it through glass on Earth.

没错，而且我们在这方面做得越来越好。在地球上，我们传输大量数据通常是靠光纤里的红外光来实现的。如果想在太空中、在卫星之间传输数据，给卫星之间拉光纤，想想还挺不切实际的。但事实证明，你根本不需要。你可以直接把光从一颗卫星射到另一颗卫星上。从某种意义上说，这种通过自由空间传输的方式，效果甚至可以比通过玻璃传输更好。这背后有两个原因。第一个原因是，光在真空中的速度比在玻璃中更快。第二个原因是，你可以使用更多的光学带宽，对吧？我提到过，当你通过光纤发送光时，用的是红外波段，这是因为在这些窄波段之外，玻璃中的杂质会吸收光。所以光纤能利用的光学带宽是有限的。如果你想在太空中发送光，原则上，你可以使用整个红外和可见光谱，这大约是光纤可用频谱的30倍。这里面当然有艰难的工程问题，但如果你解决了这些问题，原则上，在太空中传输数据甚至可以比在地球上通过玻璃传输更好。

The idea behind Project Suncatcher is that the AI computing happens in space, powered by the sun, and then that data is beamed back down to earth. But why bother to send all those computers up there? Why not bring the power back here?

Project Suncatcher 背后的想法是，AI 计算在太空中进行，由太阳能驱动，然后数据被传回地球。但为什么要费那么大劲把那些计算机送上去呢？为什么不把电力传回来？

The idea of doing space-based solar power has been around for a long time. The earliest mention I could find of this was an Isaac Asimov short story, I think it was from the forties. It is an interesting idea and maybe someday we see people doing that but you have this challenge: you’ve got to send that power, and it’s a lot of power we’re talking about for this to be worthwhile, down to Earth and somehow safely receive it there. And that power transmission and reception problem is a really hard problem.

So in some sense, the insight behind something like Suncatcher is like, hey, space is a great place to do solar power, but with AI, maybe we actually have a really important and useful way to use that power in space. The thing that’s different about AI versus some other applications you can imagine, like, I don’t know, running a steel mill or something like that, you’re not moving a lot of atoms around, right? Like you just need to send data up and down to these satellites, you don’t need to send, I don’t know, iron ore, if you were going to imagine a steel mill in space.

关于太空太阳能发电的想法其实由来已久。我最早能找到的相关记载，是 Isaac Asimov 的一个短篇小说，我想是四十年代的。这是个有趣的想法，也许有一天我们会看到人们这样做，但这里有一个挑战：你必须把电送下来，而且我们讨论的是非常大的电量，这事才值得做，并且还要安全地在地面上接收。而这个电力传输和接收问题是个极难的问题。所以从某种意义上说，Suncatcher 这类项目背后的洞见就像是：嘿，太空是个做太阳能发电的好地方，但有了 AI，我们可能真的有一种非常重要且有用的方式，在太空中使用这些电力。AI 与你能想象到的其他应用不同的地方在于，比如，我不知道，开个钢厂什么的，你不会去移动大量的原子，对吧？你只需要向卫星上传和下载数据，而不需要运送，我不知道，铁矿石，如果你想象的是在太空建钢厂的话。

Now, by my rough calculations, to generate five percent of U.S. electricity demand via solar energy on Earth would require covering an area of about 1,000 square miles, about the size of Rhode Island. You’ve said that solar panels are eight times as efficient in space, but that still implies that you’d need something like 100 square miles of solar panels in orbit to produce five percent of U.S. electricity.

根据我的粗略计算，要在地球上通过太阳能满足美国 5% 的电力需求，需要覆盖大约 1000 平方英里的面积，差不多是罗德岛州那么大。你说过，太阳能板在太空的效率是地球上的八倍，但即便如此，这仍然意味着，要生产美国 5% 的电力，还是需要在轨道上部署约 100 平方英里的太阳能板。

I’m not going try to real-time check your math on that one here, but yeah, it’s a lot of area. Here’s the thing. There’s a lot of space in space. Space is just mind-bogglingly big. The first time I really got a sense of that actually was, I was in high school and I was taking a computer animation class, and I decided I wanted to do a fly-through of the solar system, but to scale. And when you tried to do it to scale, it’s so hard to make anything visible because everything is so small.

Like the planets are so small compared to the vastness, even just of our solar system. And so you talk about these enormous areas of solar panels that you would need and yes, it’s true. And it would be smaller than what you would need if you’re trying to generate the same amount of power down on Earth. But then if you compare that area versus the amount of area that you could, in principle, build in low-Earth orbit, in this dawn-dusk, sun-synchronous, low-Earth orbit, it’s nothing.

It’s like a tiny, tiny drop in the bucket. We tried to build a visualization of what it would look like if you sort of zoomed back from Earth and you’re looking at Earth and you’re looking at, this seemingly enormous amount of satellites there, and it looks like nothing. You can’t see the satellites because they’re such a tiny percentage of the total area, the total volume there.

我现在不会去实时验算你的计算，不过没错，那确实是非常大的面积。但问题是，太空里的空间多得很。太空大得超乎想象。我第一次真正体会到这点，还是在高中上电脑动画课的时候，我决定做一个按真实比例穿越太阳系飞行的动画。当你试着按比例做的时候，很难让任何东西在画面里被看见，因为一切都太渺小了。那些行星，相比于我们太阳系的广袤空间，都是那么小。所以你说到需要部署的太阳能板面积如此巨大，是的，确实是这样。但这会比你要在地球上发同样多的电所需的面积要小。然而，如果你再把这个面积，和原则上你可以在近地轨道——也就是这种晨昏、太阳同步的近地轨道——上建造的总面积相比，那根本不值一提。简直就是九牛一毛。我们试着做了一组可视化的图像，展示当你从地球向后拉远镜头，看着地球，看着这些仿佛数量庞大的卫星群时，结果看起来什么也没有。你根本看不到那些卫星，因为它们的总面积、总体积占比实在是太小了。

The problem is getting it there, of course. You’re going to need a lot of rockets to get these things into space.

当然，问题在于怎么把它送上去。你需要很多很多火箭才能把这些东西送入太空。

Or potentially you need maybe still a decent number of rockets, but you need to reuse them a lot. And reuse of rockets is a great way to make rockets cheaper, right? That’s how you can get a lot of launches with a reasonable number of rockets.

或者说，你可能仍然需要相当数量的火箭，但你需要让它们被重复使用很多次。火箭的重复使用是降低成本的一个绝佳途径，对吧？这样你才能用合理数量的火箭，去完成大量的发射任务。

So in the short run, from what I’ve heard, the biggest barrier to doing this is not even really the physics. It sounds like you’re pretty confident of the physics, it’s just that for orbiting data centers to be economically viable, you need the cost of rocket launches to fall. How much does the cost of launching these rockets have to fall where this now becomes economically sensible?

所以从短期看，就我所听到的，做这件事的最大障碍甚至不在物理层面。听起来你对物理学原理挺有信心的，问题在于，要让轨道数据中心具备经济可行性，你必须大幅降低火箭发射的成本。发射这些火箭的成本需要降到多低，这件事在商业上才算合理？

In the draft paper that we put out, we talk about $200 a kilogram as being an important milestone, I mean, that’s not the only thing that you have to make progress on by far, but that’s a good sort of milestone to think about.

在我们发布的那篇论文草稿里，我们提到每千克 $200 是一个重要的里程碑，我的意思是，这远不是你唯一需要取得进展的方面，但这是一个很好的考量标杆。

What are the launch costs right now?

目前的发射成本是多少？

They’re quite a bit more than that, but they’ve been coming down at, you know, a really impressive pace, right? So this is, in some sense, it’s not about where things are today. It’s about skating to where the puck is going.

比那个高不少，但它们的下降速度非常惊人，对吧？所以，从某种意义上说，这无关乎今天的情况如何，关键在于你得滑向冰球要去的地方。

Spoken like a true Canadian.

这话说得像个地道的加拿大人。

Project Suncatcher plans to launch its first satellites in 2027. To build the prototypes they’re partnering with Planet, the company founded by former NASA scientist Will Marshall, who we heard from earlier. Here’s Travis Beals again.

Project Suncatcher 计划在 2027 年发射其首批卫星。为了制造原型机，他们正在和 Planet 公司合作，这家公司由我们之前听到的前 NASA 科学家 Will Marshall 创立。下面是 Travis Beals 再次发言。

We want to launch two satellites and the reason for that is one of the things we want to test is an optical link between the two satellites so that we can run a compute workload that involves multiple satellites because that’s one of important aspects of the system. So we want to test out the simplest possible version of that. And in order to do that, you need two satellites, because you need both endpoints. There’s a couple other things we want to test as well. One is just making sure all the thermal solutions behave the same way in space that we expected them to, that we tested in the lab.

And then another aspect of this is tolerance to the radiation environment in space. So that’s something else we also tested on Earth, but, you know, doing things in space is never exactly the same as trying to test things out in a lab on Earth.

我们打算发射两颗卫星，原因是我们要测试的其中一项就是两颗卫星之间的光学链路，这样我们就能跑一个涉及多颗卫星的计算工作负载，因为这是系统中很重要的一个方面。所以我们想测试这种能力最简单可行的版本。而要做这个测试，你就需要两颗卫星，因为两端都得有。另外还有几件事我们也要测。一个是确保所有的散热方案在太空中的表现，都和我们在实验室里测试时预期的一样。再一个方面，是对太空辐射环境的耐受能力。这个我们在地球上也测过，但是，你知道的，在太空里做事，跟在实验室里试着模拟测试，永远不可能完全一样。

Doing things in space is where Will Marshall comes in.

在太空里做事，这就该 Will Marshall 出场了。

We’ve got the record for the most satellites put into space. We’ve also got the record for the most blown up going to space.

我们保持着送入太空卫星数量最多的纪录，同时也保持着进入太空过程中——炸毁数量最多的纪录。

Were you one of those kids who loved outer space from an early age?

你是不是那种从小就热爱外太空的孩子？

Yeah, I saved up my pocket money and bought a pair of binoculars when I was eight or nine and eventually built my own telescope because I couldn’t afford buying one. I wouldn’t have been so surprised if you’d asked my 16-year-old self that was building a telescope that I’d be building telescopes when I’m in my mid-40s just in space looking down. I’d say, okay, well maybe I’m looking the wrong way but fine.

是啊，我八岁还是九岁的时候攒下零花钱买了一副双筒望远镜，后来因为买不起，干脆自己动手造了一台望远镜。如果你去问我那个 16 岁正在造望远镜的自己，说你四十多岁的时候还是在造望远镜，只不过是在太空里往下看，他大概不会感到特别惊讶。他可能会说，好吧，方向可能反了，但也行。

You may remember Will Marshall from the beginning of the episode. He’s the former NASA scientist who started a company by prototyping satellites in his garage. That company, Planet, is the one now prototyping Project Suncatcher’s satellites, but that isn’t their main business — not at all. I asked Marshall what Planet does.

你可能还记得这集节目一开头的 Will Marshall。他就是那位前 NASA 科学家，在自家车库里做卫星原型创办了一家公司。那家公司，Planet，现在就是在为 Project Suncatcher 的卫星做原型，但那不是他们的主业，完全不是。我问 Marshall，Planet 是做什么的。

Basically, we monitor the planet, the whole Earth, every day to help people make more informed decisions across security, civil government, commercial applications, nonprofits. And we do this having launched hundreds of satellites that image literally the whole earth landmass every single day, and with advances in AI, help people track changes, answer questions they need in more or less real time. We think of it a bit like Google indexed the internet to make it searchable. We’re sort of — Planet — indexing the Earth to make it searchable.

基本来说，我们每天监测这个星球，整个地球，帮助人们在安全领域、民用政府、商业应用、非营利组织等各方面做出更明智的决策。我们做到这一点，靠的是发射了数百颗卫星，每天都把整个地球的陆地全部真正地成像一遍，然后借助 AI 的进步，帮助人们追踪变化，近乎实时地回答他们需要了解的问题。我们某种程度上把它想成——就像 Google 给互联网建立了索引让它变得可搜索，我们，Planet，在给地球建立索引，让它变得可搜索。

Do you see yourself more as a data company or are you a satellite company?

你更多把你们看作一家数据公司，还是一家卫星公司？

No, definitely see ourselves first and foremost as a data company. Calling Planet a satellite company would be a bit like calling Google a server company, and they’ve got lots of servers, but that’s not what you see and that’s not the products you use, right? The products we give to people are data services, Earth imagery itself, as well as analytics on top of that. If you’re a farmer, we’re going to give you intelligence about fields. If you’re an intelligence official, we’ll give you information about what’s over the corner in some theater.

If you’re a civil government actor, we might be helping you to stop deforestation across large areas where you can’t monitor. We’re giving actual information to help people make smarter decisions day to day. So, in that sense, we think of ourselves more like a Bloomberg terminal, but for Earth data. People set up their data feeds. It helps them make smarter decisions day to day. And the satellites are a backend. Now, they’re pretty cool and sweet. And I can geek out on that because that’s my area.

But in the end, they’re a backend to those services.

不，我们绝对首先且最重要的，把自己看作一家数据公司。把 Planet 叫做卫星公司，就有点像把 Google 叫做服务器公司，他们确实有很多服务器，但这并不是你看到的东西，也不是你使用的产品，对吧？我们给人们的产品是数据服务，是地球影像本身，以及建立在此之上的分析。如果你是一位农民，我们会给你提供关于田块的情报。如果你是一名情报官员，我们会给你展示某个区域转角处有什么的信息。如果你是一位民用政府部门的工作人员，我们可能在帮你监测大面积的森林砍伐，这些地方靠人力是监测不过来的。我们在提供真实的信息，帮助人们日复一日做出更明智的决策。所以，从这个意义上说，我们更像是 Bloomberg 终端，只不过是地球数据的 Bloomberg 终端。人们设定好自己的数据推送，它就能帮助他们日常做更明智的决策。而卫星是后端。当然，它们非常酷炫，我也能沉浸在技术细节里，因为那是我的老本行。但归根结底，它们只是支撑那些服务的后端。

How many satellites do you need to take a picture of every piece of land on Earth every day?

要拍下地球上每一片陆地每天的照片，你需要多少颗卫星？

Well, yeah, we figured this out when we left NASA. We would need about 100 imaging satellites. We’ve actually got about twice that number in space. It’s the largest Earth-imaging constellation. And basically they act as a line scanner for the planet. So, imagine the orbit of the Earth, you have a line of satellites in a ring that are going in a polar orbit, so over both poles. And the Earth ends up rotating underneath that line. The line actually stays fixed with respect to the sun, and the Earth rotates underneath, and each one takes a strip of images, but the Earth is rotated slightly by the time the next satellite comes down, and therefore you end up systematically line-scanning the Earth once every 24 hours.

呃，是的，这个我们离开 NASA 的时候就算过。我们大概需要 100 颗成像卫星。实际上我们在太空里现在有这个数量的两倍左右，是最大的对地成像星座。它们基本上就像一个对星球的线扫描仪。你可以想象，环绕地球的轨道上，有一圈卫星沿着极地轨道运行，所以会飞过南北两极。然后地球就在那条线下自转。这条线本身相对于太阳是固定的，地球在底下自转，每颗卫星都会拍下一条带状影像，但等下一颗卫星过来的时候，地球已经又转了一点点，这样你最终就能在 24 小时内有条不紊地对整个地球做一次线扫描。

When I think about what your company could do and be, there’s this real dichotomy, right? You could imagine all of these wonderful things that would be good for the world and, and helpful, whether it’s fighting forest fires, or helping in droughts. But then there’s also sort of this specter of surveillance, which I know makes a lot of people nervous. Who are your customers? And how do you think about that problem?

当我想到你们公司能做什么、成为什么的时候，确实存在一种截然相反的两面性，对吧？你可以想象出所有这些美好的事情，会对世界有益，无论是在对抗森林火灾，还是在干旱中提供帮助。但同时也存在着一种监控的阴影，我知道这让很多人感到不安。你们的客户是谁？你怎么看待这个问题？

So let me start with the customers. We service customers in civil government. So, that’s agencies like NASA that buy it for science research. Or it’s the Brazil Federal Police that are trying to stop deforestation. We work with commercial applications. So, insurance, for example. After disasters, instead of sending a person out to see if that farm or that house has been affected by the flood or the storm, they can just look at our imagery instead and save themselves a lot of time and effort.

And then there’s defense and intelligence applications. For example, we work with the U.S. Navy monitoring the South China Sea, where they’re tracking illicit activities, North Korean evasion of oil sanctions. We work with Ukraine on trying to help them monitor Russian activities. And then we work also with a lot of different other kinds of actors that are less prominent like we work with think-tanks, NGOs like the Red Cross, or Amnesty International, Human Rights Watch, who are tracking events like refugee camps around the world.

We also work with news media to shed light on events as they go on around the world. To the surveillance question, I think of this as that we are helping people have better information to help them make better decisions. That’s especially true in sustainability, stopping deforestation, illegal fishing, and all that. It’s kind of obvious, but even in the security realm, our theory of change is kind of simple. Greater transparency leads to greater accountability leads to greater security, better decision-making.

If you look back in history, the history of conflict is highly correlated with people having a lack of information. The Cuban Missile Crisis was a situation when the U.S. didn’t know and got surprised by the Soviet Union putting missiles in Cuba. It was because they didn’t know that they were there, that the Cold War almost became hot. Or vice versa when the U.S. had put missiles in Turkey without the Soviet Union knowing. When we have better information about what each other are doing around the planet it tends to diffuse situations, it is less risky, less miscalculations, and so on.

那么让我先从客户说起。我们为民用政府客户提供服务。比如像 NASA 这样的机构，购买我们的数据用于科学研究。还有巴西联邦警察，他们在努力制止森林砍伐。我们也服务商业应用，以保险业为例，灾难发生后，与其派人去现场查看那个农场或那栋房子是否被洪水或风暴影响，他们可以直接看我们的影像，省下大量时间精力。然后是国防与情报方面的应用。比如，我们与美国海军合作，监测南海，追踪非法活动、朝鲜规避石油制裁的行为。我们与乌克兰合作，协助他们监测俄罗斯的活动。此外，我们还与许多其他不那么显眼的行动方合作，比如智库、非政府组织，如红十字会、大赦国际、人权观察，他们在追踪全球各地难民营等情况。我们也与新闻媒体合作，为世界各地正在发生的事件提供光照。关于监控的问题，我认为，我们是在帮助人们获取更优质的信息，从而做出更好的决策。在可持续发展、制止森林砍伐和非法捕捞等领域，这一点尤为明显。这几乎是显而易见的，但即便在安全领域，我们关于改变的逻辑也很简单：更高的透明度带来更高的问责制，进而带来更高的安全性、更好的决策。回顾历史，冲突的历史与信息缺失高度相关。古巴导弹危机的局面，就是因为美国不知情，被苏联在古巴部署导弹打了个措手不及。正是因为不知道它们在那里，冷战差点变成了热战。反过来，美国在土耳其部署导弹时，苏联同样不知情。当我们对彼此在全球各地在做什么有了更清晰的了解，往往就能缓和局势，风险更低，误判更少，诸如此类。

Your company made the headlines recently when you made the decision to restrict access to some satellite data in areas affected by the war with Iran.

你的公司最近登上了新闻头条，因为你们决定限制对伊朗战争影响区域的部分卫星数据的访问。

Correct.

没错。

Was that an easy decision or a hard decision?

那是一个容易的决定，还是一个艰难的决定？

It’s always a very hard decision. We always get in a situation where there’s active conflicts whether Ukraine, Gaza, Iran, where there is a tension between ongoing operations and trying to ensure the legitimate interests of people not to be exposing ongoing operations that could put military personnel or civilians in harm’s way, and at the same time, trying to provide transparency and accountability in the public interest. And what we’ve typically done is put a delay around some of our imagery so that it ensures that we can’t have civilians and military operations put in harm’s way, but at the same time, ultimately provides transparency and accountability.

这始终是一个非常艰难的决定。每当出现像乌克兰、加沙、伊朗这样的活跃冲突时，我们总会陷入一种两难境地：一方面是正在进行的军事行动，要力求确保不让相关行动暴露，以免将军事人员或平民置于险境；另一方面，又努力提供符合公众利益的透明度和问责制。这两者之间存在张力。我们通常的做法是，对部分影像设置一个延迟期，这样既能确保我们不会让平民和军事行动置于险境，同时又最终提供了透明度和问责制。

How do you get your satellites into orbit? You don’t actually do launches yourself, right?

你们是如何把卫星送入轨道的？你们自己不实际执行发射，对吧？

Well, I’ve been practicing with my right arm for quite a while now, and, uh, no. Obviously we use rockets. We’ve launched on now 41 rockets to space. Just to give you a sense that 38 of them got to orbit. So, three failures in our time where the rocket blew up somehow. So, it’s a risky business still, but, yes, it’s becoming more routine. We’ve launched over a dozen times on SpaceX rockets, we’ve launched on an Indian PSLV rocket, we’ve launched on various rockets from around the world.

So, basically we stick up our thumb at the launch site and hitch a ride because normally we’re not the main payload. There’s a big billion dollar satellite going to space or something, and we’re hitching on the side of those with lots of little smaller satellites. Interesting fact, we’ve got the record for the most satellites put into space. We’ve also got the record for the most blown up going to space. A common misunderstanding of what happened in the launch sector, a lot of people know that launch costs have come down.

But for the first more than a decade of launches for SpaceX, they really brought U.S. launch costs down to be comparable with international launch costs. Really, the U.S. launch costs have been kept arbitrarily high by a central monopoly because the U.S. government satellites had to be launched on U.S. rockets by law from Congress, so Boeing and Lockheed teamed together to ensure that they had a really expensive rocket because they had to be used because they were the only rocket provider.

But we as a commercial company weren’t limited to U.S. rockets, and we went around the world. And, like the Indian PSLV, very reliable at least at the time, and it was quite cheap. SpaceX ultimately got not just to that point but then brought it down even further about four or five x further than that, and that’s really, really, really a boon. But that wasn’t how we first got to space at least. And that has been a further help. But remember even a four or five x reduction in cost, which really helps the whole space industry and has helped spawn it, is less significant, purely from an economic standpoint, than the 100 to a 1,000x improvement in cost performance of satellites that are going in.

So, in each kilogram that you put up into space, if you can get a hundred times more data, because you’ve made the satellite way smaller, then you have won big time, even if the launch cost didn’t change one iota.

嗯，我练习用右臂把卫星扔上天已经有一阵子了，呃，当然不行。显然我们用的是火箭。到目前为止我们已经执行了 41 次太空发射。给你一个大致的概念，其中 38 次成功入轨。所以，在这段时间里，我们遇到过三次失败，火箭因为某种原因爆炸了。所以，这仍然是一个高风险的行业，不过，是的，它正在变得更加常态化。我们搭载 SpaceX 的火箭发射过十几次，也搭载过印度的 PSLV 火箭，还有世界各地其他各种火箭。所以，基本上，我们就在发射场竖起大拇指，搭个便车，因为我们通常不是主要载荷。通常是有一颗价值 10 亿美元的大型卫星要上天之类的，我们就在它们旁边，带着许多更小的小卫星一起搭车。说个有趣的事，我们保持着送入太空的卫星数量最多的记录。我们也保持着在升空过程中被炸毁数量最多的记录。关于发射领域的一个常见误解是，很多人知道发射成本已经下降了。但在 SpaceX 开展发射业务的头十多年里，他们实际上是把美国的发射成本降低到了与国际发射成本相当的水平。事实上，美国的发射成本之所以被虚高维持，是因为一个核心垄断者，因为，依照国会法律，美国政府卫星必须用美国火箭发射，所以 Boeing 和 Lockheed 联手确保他们拥有非常昂贵的火箭，因为必须用，而他们是唯一的火箭提供商。

但作为商业公司，我们不受限于美国火箭，我们走遍了全球。比如，印度的 PSLV，至少在当时非常可靠，而且相当便宜。SpaceX 最终不仅达到了那个水平，还进一步把成本又拉低了大约四到五倍，这真的是非常非常非常大的利好。但至少，我们最初进入太空靠的不是这个。这项进展给了我们进一步的帮助。但要记住，即使成本降低了四五倍，这对整个太空产业帮助巨大、促进了其蓬勃发展，但从纯经济角度来看，它与进入太空的卫星在性价比上实现 100 到 1,000 倍的提升比起来，就没那么显著了。因为，在你送入太空的每千克载荷中，如果你能通过把卫星做得小得多，来获得一百倍的数据量，那么，哪怕发射成本一丁点都没变，你也已经大获全胜了。

So, you’ve been working with Google’s Project Suncatcher team to build and test prototypes of orbital data centers. Do you think over the next 50 years, orbital data centers are a long shot, a sure thing, somewhere in between?

所以，你们一直在与 Google 的 Project Suncatcher 团队合作，建造并测试轨道数据中心的原型。你认为在未来 50 年里，轨道数据中心是希望渺茫、是板上钉钉，还是介于两者之间？

I actually think it’s quite clearly going to happen, and here’s why. We did a calculation with Google back eight or nine years ago now. And looking at all the cost of data centers on the ground, all the costs of data sensors in space, and we sort of did a modeling and figured out that by around costs of launch coming to $200 to $300 a kilogram, it would just be cheaper. Purely on cost grounds, it’d be cheaper to put them in orbit than to put them on the ground. And I remember Sergey and Larry saying to me, well, let’s come back in about 2030 when we had predicted that launch costs would come down to there and start this.

I said, no, let’s come back five years before that, because it’s going to take a little bit of time to actually build these technologies — the thermal technology and other things — and that’s exactly what happened. So, last year, Google came back to us and said, let’s start this project. And so, it’s an early days project, it’s a moonshot as they call it, but like Waymo, or quantum computing, which turned out to be quite successful. It’s an ambitious project, but ultimately, I think it’s inevitable that, within a 10-ish year timeframe, most computers being built on the Earth will be going to space.

I actually think not only is it going to be economically cheaper, but it’s going to be more sustainable. Earth’s life is so precious, it’s incredible, but yet we’re whittling away a lot of it. We’re doing deforestation for lithium, for cows, for whatever, we’re doing it. And we need to put our energy-intensive infrastructure into orbit, if we can, to not have a collision course with the biodiversity on the planet, as Jeff Bezos has said, we need to zone the Earth — rural and light urban — and intensive stuff in orbit where there’s lots of energy, and we don’t conflict with that precious biodiversity.

我其实认为这件事非常明确地会发生，原因如下。大概八九年前，我们和 Google 一起做过一个计算。我们把地面上所有数据中心的成本、太空中所有数据传感器的成本都算了进去，做了一种建模，得出一个结论：当发射成本降到每千克 $200 到 $300 左右时，在轨道上建数据中心就会更便宜。纯粹从成本角度看，把它们放在轨道上会比放在地面上更便宜。我记得 Sergey 和 Larry 当时对我说，嗯，等差不多 2030 年我们再回头来谈这件事吧，我们当时预测到那时发射成本会降到这个水平，然后再启动。我说，不行，我们得提前五年就开始，因为真正建造这些技术——热控技术还有其他东西——还需要一点时间，结果也确实如此。于是，去年 Google 回头来找我们说，我们启动这个项目吧。所以，这是一个早期项目，他们管它叫登月式项目，但就像 Waymo 或量子计算一样，这些项目后来都证明相当成功。这是一个雄心勃勃的项目，但归根结底，我认为在十年左右的时间框架内，地球上建造的大多数计算机都将进入太空，这是不可避免的。我其实认为，这不仅在经济上更便宜，而且会更可持续。

地球上的生命太珍贵了，简直不可思议，可我们却在大量消耗它。我们为了锂、为了养牛、为了各种原因砍伐森林，我们正在这样做。我们需要把我们能源密集型的基础设施放到轨道上去，如果做得到的话，以避免与这个星球上的生物多样性发生冲突。正如 Jeff Bezos 所说，我们需要对地球进行分区——乡村和轻城市——而把高强度的东西放在轨道上，那里有大量的能源，而且我们不会与那珍贵的生物多样性发生冲突。

So, you’re part of this big space economy right now. If these orbital data centers do turn out to be viable, where will they fit into the space economy? Will they be a small part? Will they dominate the space economy?

所以，你已经是这个庞大太空经济的一部分了。如果这些轨道数据中心最终真的可行，它们会在太空经济中处于什么位置？会是一个很小的部分，还是会主导整个太空经济？

Roughly the latter. So, if you think about the rough maths, the current space sector has all of 300 billion annually. But if you just look at compute spend, it adds up to more than that every year that it’s being spent, and it’s being projected to be spent. Again, we’re talking years out from now, I’m not talking about tomorrow, but, yeah, in 10 years’ time I would imagine it would probably be bigger than all of the rest of the space economy combined.

大概会是后者。如果你算一算大致的数字，目前整个太空行业每年的规模也就 300 billion。但如果你只看计算支出，每年花在上面的钱加起来就已经超过这个数了，而且未来的支出预测也是这样。再说一次，我说的是几年以后的事，不是明天的事，但是，没错，我觉得十年之后，它的规模大概会比太空经济中其他所有领域加起来还要大。

Until then, take care of yourself — and, if you can, someone else, too.

在那之前，照顾好你自己——如果做得到的话，也照顾一下别人。

Freakonomics Radio is produced by Renbud Radio. You can find our entire archive on any podcast app; it’s also at freakonomics.com, where we publish transcripts and show notes. This episode was produced by Augusta Chapman and edited by Gabriel Roth; it was mixed by Jake Lummus with help from Jeremy Johnston. The Freakonomics Radio Network staff also includes Dalvin Aboagye, Eleanor Osborne, Ellen Frankman, Elsa Hernandez, Ilaria Montenecourt, Pete Madden, and Theo Jacobs. Our theme song is “Mr. Fortune,” by the Hitchhikers; and our composer is Luis Guerra. As always, thanks for listening.

Freakonomics Radio 由 Renbud Radio 制作。你可以在任何播客应用上找到我们的全部节目存档；也可以访问 freakonomics.com，我们在那里发布文字记录和节目笔记。本期节目由 Augusta Chapman 制作，Gabriel Roth 编辑；由 Jake Lummus 混音，Jeremy Johnston 协助。Freakonomics Radio Network 的团队成员还包括 Dalvin Aboagye、Eleanor Osborne、Ellen Frankman、Elsa Hernandez、Ilaria Montenecourt、Pete Madden 和 Theo Jacobs。我们的主题曲是 the Hitchhikers 的《Mr. Fortune》；我们的作曲是 Luis Guerra。一如既往，感谢收听。