为什么 AI 数据中心将在 36 个月内搬到太空

所以，真的有三个小时的问题吗？你是认真的吗？

So, are there really three hours of questions? Are you serious?

是啊。你甚至都不聊埃隆，老兄。我是说，这是最有趣的一点。所有的故事线都在汇聚。

Yeah. You don't even talk about Elon, man. I mean, it's the most interesting point. All the story lines are kind of converging.

现在，所以我们看看有多少。几乎就像我计划好的。

Right now, so we'll see how much. Almost like I planned it.

没错。好吧，我们会聊到的。我绝不会做这种事。

Exactly. Well, we'll get. I would never do such a thing.

所以，正如你比任何人都清楚，数据中心的总体拥有成本中，只有 10%到 15%是能源。而这部分你大概是通过搬到太空来节省的。大部分成本是 GPU。如果它们在太空，维护起来更困难，或者根本无法维护。因此，折旧周期会缩短。所以，大概在太空放 GPU 要贵得多。把它们放到太空的理由是什么？

So, as you know better than anybody else, the total cost of ownership of a data center, only 10 to 15% is energy. And that's the part you're presumably saving by moving this into space. Most of it's the GPUs. If they're in space, it's harder to service them or you can't service them. And so, the depreciation cycle goes down on them. So, like it's just way more expensive to have the GPUs in space presumably. What's the reason to put them in space?

嗯，能源的可用性是问题所在。我的意思是，如果你看中国以外的电力输出，中国以外的地方基本上持平。可能略有增长，但非常接近持平。中国的电力输出增长很快。但如果你把数据中心放在中国以外的任何地方，你要从哪里获得电力，尤其是随着规模扩张？芯片的产出几乎是指数级增长，但电力产出是平的。那么你怎么给那些芯片通电呢？

Well, the availability of energy is the issue. I mean if you look at electrical output outside of China, everywhere outside of China it's more or less flat. It's very, maybe a slight increase but pretty close to flat. China has a rapid increase in electrical output. But if you're putting data centers anywhere except China, where are you going to get your electricity, especially as you scale? The output of chips is growing pretty much exponentially, but the output of electricity is flat. So how are you going to turn those chips on?

神奇的能量来源，神奇的电力精灵。

Magical power sources, magical electricity fairies.

你是说你以热爱太阳能而闻名。一太瓦的太阳能，按 25%的容量因子算，大约需要四太瓦的太阳能板，这相当于美国土地面积的 1%。这就像，远在这个，当我们有了一太瓦的数据中心时，你就处于奇点中，对吧？所以我们缺的是什么？

You mean you're famously a big fan of solar. One terawatt of solar power, with a 25% capacity factor, like four terawatts of solar panels, it's like 1% of the land area of the United States. And that's like, far in this, you were in the singularity when we've got one terawatt of data centers, right? So what are we running out of?

不过你处于奇点多深了？

How far into the singularity are you though?

你告诉我。

You tell me.

是啊，没错。所以我觉得我们会发现我们就在奇点中，然后，好吧，我们还有很长的路要走。

Yeah, exactly. So I think we'll find we're in the singularity and like, okay, we still got a long way to go.

但这就像是，在我们用太阳能板覆盖内华达州之后，再把它放到太空的计划吗？

But is this like the plan to put it in space after we've covered Nevada in solar panels?

我觉得用太阳能板覆盖内华达州相当困难。你要拿到许可，比如，试试看能不能拿到许可。

I think it's pretty hard to cover Nevada in solar panels. You get permits from, like, try getting the permits for that.

所以太空实际上是一个监管策略。在地面上建设比在太空更难。

So space is really a regulatory play. It's harder to build on land than it is in space.

在地面上规模扩张比在太空更难。而且，你在太空中的太阳能板效率大约是地面的五倍。而且你不需要电池。我差点穿了我的另一件衬衫，上面写着“太空总是晴天”，确实如此。因为太空中没有昼夜循环、季节变化、云层或大气。仅大气层就会导致大约 30%的能量损失。所以对于任何给定的太阳能板，它们在太空中的发电量大约是地面的五倍，而且你避免了为夜间供电而配备电池的成本。所以实际上在太空做这件事要便宜得多。我的预测是，在 36 个月或更短的时间内，可能 30 个月，太空将成为放置 AI 最便宜的地方。

It's harder to scale on the ground than it is to scale in space. But also, you're going to get about five times the effectiveness of solar panels in space versus the ground. And you don't need batteries. I almost wore my other shirt which says 'it's always sunny in space,' which it is. Because you don't have a day-night cycle or seasonality, clouds, or an atmosphere in space. Because the atmosphere alone results in about a 30% loss of energy. So for any given solar panels, they can do about five times more power in space than on the ground, and you avoid the cost of having batteries to carry you through the night. So it's actually much cheaper to do it in space. And my prediction is that it will be by far the cheapest place to put AI will be space in 36 months or less, maybe 30 months.

36 个月？

36 months?

少于 36 个月。你怎么维护那些在训练中经常故障的 GPU？

Less than 36 months. How do you service GPUs as they fail, which happens quite often in training?

实际上，这取决于到货的 GPU 有多新。我的意思是，目前我们发现我们的 GPU 相当可靠。存在早期失效率，你显然可以在陆地上解决这个问题。所以你可以在陆地上运行它们，确认 GPU 没有早期失效。但一旦它们开始工作，实际可靠性，一旦它们开始工作并且你通过了 Nvidia 或任何芯片制造商的初始调试周期，可能是特斯拉 AI 6 芯片之类的，也可能是 TPU 或 Trainium 等。可靠性实际上在某个点之后相当可靠。所以我认为维护不是问题。但你可以记住我的话：在 36 个月内，但可能更接近 30 个月，放置 AI 最具经济吸引力的地方将是太空。然后太空会变得好得离谱。然后规模扩张，唯一能真正规模扩张的地方就是太空。一旦你开始考虑你利用了多少太阳的能量，你就会意识到你必须去太空。你在地球上无法大规模扩张。

Actually, it depends on how recent the GPUs are that arrived. I mean, at this point, we found our GPUs to be quite reliable. There's infant mortality, which you can obviously iron out on the ground. So you can just run them on the ground and confirm that you don't have infant mortality with the GPUs. But once they start working, their actual reliability, once they start working and you're past the initial debug cycle of Nvidia or whoever is making the chips, could be Tesla AI 6 chips or something like that, or it could be TPUs or trains or whatever. The reliability is actually quite reliable past a certain point. So I don't think the servicing thing is an issue. But you can mark my words: in 36 months, but probably closer to 30 months, the most economically compelling place to put AI will be space. And then it will get ridiculously better to be in space. And then the scaling, the only place you can really scale is space. Once you start thinking in terms of what percentage of the sun's power you are harnessing, you realize you have to go to space. You can't scale very much on Earth.

但“大规模”，明确一下，你指的是太瓦级别。

But by 'very much,' to be clear, you're talking like terawatts.

是的。嗯，整个美国目前平均只使用半太瓦的电力。对吧。所以，你知道，如果你说一太瓦，那将是美国目前消耗电力的两倍。所以那是相当多的。你能想象建造那么多数据中心、那么多发电厂吗？就像那些生活在软件领域的人没有意识到他们即将在硬件方面上一堂艰难的课。建造发电厂实际上非常困难。而且你不仅需要发电厂，还需要所有的电气设备，需要电力变压器来运行变压器，AI 变压器。现在公用事业行业是一个非常缓慢的行业。他们几乎与政府、公共事业委员会阻抗匹配。所以他们非常缓慢，因为他们的过去非常缓慢。所以试图让他们快速行动就像，如果你试图与公用事业公司签订大规模、大电力的互联协议……

Yeah. Well, all of the United States currently uses only half a terawatt of power on average. Right. So, you know, if you say a terawatt, that would be twice as much electricity as the United States currently consumes. So that's quite a lot. And can you imagine building that many data centers, that many power plants? It's like those who have lived in software land don't realize they're about to have a hard lesson in hardware. It's actually very difficult to build power plants. And you don't just need power plants, you need all of the electrical equipment, you need the electrical transformers to run the transformers, the AI transformers. Now the utility industry is a very slow industry. They pretty much impedance match to the government, to the public utility commission. So they're very slow because their past has been very slow. So trying to get them to move fast is like, if you try to do an interconnect agreement with a utility at scale with a lot of power...

作为一名专业播客，我可以说我实际上没有。

As a professional podcaster, I can say that I have not, in fact.

是的，他们必须做一年的研究。好吧，大约一年后他们会带着他们的互联研究回来找你。

Yeah, they have to do a study for a year. Okay, like a year later they'll come back to you with their interconnect study.

但你不能用你自己的表后电力设施解决这个问题吗？

But can't you solve this with your own behind-the-meter power stuff?

你可以建造发电厂。我们在 xAI 为 Colossus 2 就是这么做的。所以对于 Colossus 2……

You can build power plants. That's what we did at xAI for Colossus 2. So for Colossus 2...

但所以，是啊，我们为什么要讨论电网？为什么不直接建造 GPU 和电力共址呢？

But so yeah, why are we talking about the grid? Why not just build GPUs and power collocated?

我们就是这么做的。

That's what we did.

对。对。

Right. Right.

但我想说的是，当你谈论所有这些问题时，为什么这不是一个通用的解决方案？

But I'm saying why isn't this a generalized solution when you're talking about all the issues?

你从哪儿弄到发电厂？

Where do you get the power plants from?

我是说，当你谈论与公用事业合作的所有问题时，你可以直接为数据中心建私人发电厂。

I'm saying when you're talking about all the issues working with utilities, you can just build private power plants with the data centers.

对。但这引出了一个问题：你从哪儿弄到发电厂？你从哪儿弄到发电厂？

Right. But it begs the question of where do you get the power plants? Where do you get the power plants from?

我是说发电厂制造商。

I mean the power plant makers.

哦，你是这个意思。基本上就是燃气轮机积压订单。

Oh, that's what you're saying. Like there's the gas turbine backlog basically.

是的。你可以再深入一层。是涡轮机里的叶片和导叶是限制因素，因为铸造是一个高度专业化的过程，用燃气动力铸造涡轮机里的叶片和导叶。而且其他形式的发电很难规模化。太阳能可能可以规模化，但目前美国进口太阳能的关税非常高，国内太阳能产量又少得可怜。

Yes. You can drill down to a level further. It's the veins and blades in the turbines that are the limiting factor because the casting is a very specialized process to cast the blades and veins in the turbines using gas power. And it's very difficult to scale other forms of power. You can scale potentially solar but the tariffs currently for importing solar in the US are gigantic and the domestic solar production is pitiful.

为什么不制造太阳能？这看起来像是一个很适合埃隆的问题。

Why not make solar? That seems like a good Elon shaped problem.

我们打算制造太阳能。

We are going to make solar.

好的。

Okay.

太好了。

Great.

SpaceX 和特斯拉都在朝着 100 吉瓦的太阳能电池产量迈进。从多晶硅到晶圆再到最终面板，你们深入到产业链的哪一层？

Both SpaceX and Tesla are building towards 100 gigawatt of solar cell production. How low down the stack from poly silicon up to the wafer to the final panel?

我认为你得从原材料一直做到成品电池。现在，如果是用于太空，实际上成本更低，制造太空用太阳能电池更容易，因为它们不需要玻璃，或者不需要太多玻璃，也不需要沉重的框架，因为它们不需要经受天气事件。太空里没有天气。所以太空用的太阳能电池实际上比地面用的更便宜。

I think you got to do the whole thing from raw materials to finish the cell. Now, if it's going to space, it actually costs less and it's easier to make solar cells that go to space because they don't need glass or they don't need much glass and they don't need heavy framing because they don't have to survive weather events. There's no weather in space. So it's actually a cheaper solar cell that goes to space than the one on the ground.

在未来 36 个月内，有没有办法让它们便宜到你需要的程度？

Is there a path to getting them as cheap as you need in the next 36 months?

太阳能电池已经很便宜了。便宜得离谱。而且如果你说，我认为中国的太阳能电池大约每瓦 25 到 30 美分左右。便宜得荒谬。而且当你考虑到现在把它放在太空里，它便宜五倍，因为实际上是五倍，不，不是五倍，是十倍，因为你不需要任何电池。所以一旦你进入太空的成本变低，迄今为止最便宜、最可扩展的生成 token 的方式就是太空。其他方式根本没法比。规模化会容易一个数量级。而且抛开芯片不谈，一个数量级，如果问题是你在地面上无法规模化。你就是无法规模化。人们在发电方面很快就会撞上大墙。他们已经撞上了。比如 xAI 团队为了获得 1 吉瓦的电力上线，不得不完成的一系列奇迹和壮举，简直疯狂。我们不得不把一大堆涡轮机组装在一起，然后在田纳西州遇到许可问题，不得不越过边境到密西西比州，幸运的是只有几英里远。所以我们还得铺设几英里的高压线，并在密西西比州建一个发电厂。建起来非常困难。人们不明白，要为一个数据中心供电，在发电机层面、发电层面到底需要多少电力，因为那些菜鸟会看比如 GB300 的功耗，然后乘以某个数，就以为那就是你需要的电力。

Solar cells are already very cheap. They're like far sickly cheap. And if you say, I think solar cells in China are around like 25-30 cents a watt or something like that. It's absurdly cheap. And when you take into account now put it in space and it's five times cheaper because it's five times in fact no it's not five times cheaper it's 10 times cheaper because you don't need any batteries. So the moment your cost of access to space becomes low, by far the cheapest and most scalable way to generate tokens is space. It's not even close. It'll be an order of magnitude easier to scale. And chips aside, an order of magnitude well if the point is you won't be able to scale on the ground. It's just you just won't. People are going to hit the wall big time on power generation. They already are. Like the number of miracles and series that the xAI team had to accomplish in order to get a gigawatt of power online was crazy. We had to gang together a whole bunch of turbines and then we had permit issues in Tennessee and had to go across the border to Mississippi, which is fortunately only a few miles away. So then we still had to run the high power lines a few miles and build a power plant in Mississippi. And it was very difficult to build that. And people don't understand how much electricity you actually need at the generator level at the generation level in order to power a data center because they look at the noobs will look at the power consumption of say a GB300 and multiply that by thing and then think that's the amount of power you need.

还有所有冷却等等。

All the cooling and everything.

醒醒吧。是的。这完全是菜鸟行为。你这辈子从来没做过硬件。除了 GB300，你还得给所有网络硬件供电。还有一大堆 CPU 和存储设备。你必须按峰值冷却需求来设计。这意味着你能不能在一年中最热的时候、最热的那天进行冷却？孟菲斯可是热得要命。所以光是冷却，你的电力就要增加 40%。如果你不想数据中心在热天关机，想继续运行，那么你还得说，好吧，这上面还有一个乘法因子，那就是你是否假设你的发电永远不会出问题？比如，哦，实际上，有时我们不得不让发电机、部分电力离线进行维护。哦，好吧。现在你又要加 20-25% 的乘数，因为你必须假设你得让电力离线进行维护。所以实际上，每大约 11 万个 GB300，包括网络、CPU、存储、冷却、维护电力余量，大约需要 300 兆瓦。抱歉，再说一遍。

Wake up. Yeah. This is like that's a total noob. You've never done any hardware in your life before. Besides the GB300, you got to power all of the networking hardware. There's a whole bunch of CPU and storage stuff that's happening. You've got to size for your peak cooling requirements. So that means can you cool even on the worst hours, the worst day of the year? Well, it gets pretty freaking hot in Memphis. So you're going to have like a 40% increase on your power just for cooling. If assuming you don't want your data center to turn off on hot days and want to keep going, then you got to say, well, there's another multiplicative element on top of that, which is, are you assuming that you never have any hiccups in your power generation? Like, oh, well, actually, sometimes we have to take the generators, some of the power offline in order to service it. Oh, okay. Now you add another 20-25% multiplier on that because you've got to assume that you've got to take power offline to service it. So the actual RS for roughly every 110,000 GB300s inclusive of networking, CPU, storage, cooling, margin for servicing power, is roughly 300 megawatt. Sorry, say that again.

大致上，或者你这样想，大约 33 万个，实际上你在发电层面需要多少来服务大约 33 万个 GB300，包括所有相关的支持网络和其他一切，以及峰值冷却，还要有一些余量，一些电力储备，大约是 1 吉瓦。

It's roughly or think about it like the way you think about this like 330,000 to actually what you need at the generation level to service probably 330,000 GB300s including all of the associated support networking and everything else and the peak cooling and to have some margin, some power margin reserve is roughly a gigawatt.

我能问一个非常天真的问题吗？你描述了在地球上做这些事情的工程细节。但在太空中做也有类似的工程困难。你怎么用轨道激光器代替 InfiniBand 等等。你怎么让它抗辐射？我不知道工程细节，但基本上，有什么理由认为这些以前从未需要解决的挑战最终会比在地球上建造更多涡轮机更容易？地球上有公司制造涡轮机。他们可以制造更多涡轮机，对吧？

Can I ask a very naive question? You're describing the engineering details of doing this stuff on Earth. But then there's analogous engineering difficulties of doing it in space. How do you replace InfiniBand with orbital lasers etc. How do you make it resistant to radiation? I don't know the details in the engineering but fundamentally what is the reason to think those challenges which have never been had to be addressed before will end up being easier than just building more turbines on Earth. There's companies that build turbines on Earth. They can make more turbines, right?

我再次邀请你试试看，然后你就会明白。涡轮机已经卖到了 2030 年。

I invite again try doing it and then you'll see. So the turbines are sold out through 2030.

你们考虑过自己制造吗？我认为为了上线足够的电力，SpaceX 和特斯拉可能得内部制造涡轮机叶片、导叶和叶片。

Have you guys considered making your own? I think in order to bring enough power online, I think SpaceX and Tesla will probably have to make the turbine blades, the veins and blades internally.

只是叶片还是整个涡轮机？

For just the blades or the turbines?

限制因素：除了叶片，他们称之为叶片和导叶，你什么都能买到。你可以比导叶和叶片早 12 到 18 个月拿到其他东西。限制因素是导叶和叶片，全世界只有三家铸造公司生产这些，而且它们有大量积压订单。

The limiting factor: you can get everything except the blades, they call the blades and veins. You can get that 12 to 18 months before the veins of blades. The limiting factor is veins and blades, and there are only three casting companies in the world that make these and they're massively backlogged.

这是 Seaman 的 GE 那些人，还是子公司？

Is this Seaman's GE those guys or is it a sub?

不，是其他公司。我的意思是，有时他们内部有一点铸造能力，但我是说你可以直接给任何涡轮机制造商打电话，他们会告诉你的。这不是什么绝密信息，现在可能就在网上。

No, it's other companies. I mean, sometimes they have a little bit of casting capability in house, but I'm just saying you can just call any of the turbine makers and they will tell you. It's not top secret. It's probably on the internet right now.

如果不是因为关税，Colossus 会不会用太阳能供电？

If it wasn't for the tariffs, would Colossus be solar powered?

用太阳能供电会容易得多。是的。关税太疯狂了，百分之几百。

It would be much easier to make it solar powered. Yeah. The tariffs are nuts. Several hundred percent.

你难道不认识一些人……

Don't you know some people...

我们也需要速度。是的。你知道，总统先生，我们并非在所有事情上都意见一致。而且这届政府并不那么喜欢太阳能。

We also need speed. Yeah. You know, the president, we don't agree on everything. And this administration is not the biggest fan of solar.

但你也需要土地、许可证等等。所以，如果你想快速推进……

But you also need the land, the permits and everything. So, if you're trying to move very fast...

我确实认为在地球上扩大太阳能规模是个好办法，但你需要一些时间来寻找土地、获得许可、采购太阳能板，并搭配电池。

I do think scaling solar on Earth is a good way to go, but you need some amount of time to find the land, get the permits, get the solar, pair that with batteries.

但为什么不能自己建立太阳能生产呢？你最终会用完土地，但德克萨斯州、内华达州有很多土地，包括私人土地。所以你至少可以支撑下一个 Colossus 以及再下一个。到某个点你会遇到瓶颈，但眼下这难道行不通吗？

But why would it not work to stand up your own solar production? You eventually run out of land, but there's a lot of land in Texas, Nevada, including private land. So you'd be able to at least get the next Colossus and the one after that. At a certain point you hit a wall, but wouldn't that work for the moment?

正如我所说，我们正在扩大太阳能生产。太阳能电池的物理生产有一个可扩展的速度。我们正在以尽可能快的速度扩大国内生产。

As I said, we are scaling solar production. There's a rate at which you can scale physical production of solar cells. We're going as fast as possible in scaling domestic production.

你们在特斯拉生产太阳能电池。

You're making the solar cells at Tesla.

特斯拉和 SpaceX 都有目标要达到每年 100 吉瓦的太阳能。

Both Tesla and SpaceX have a mandate to get to 100 gigawatts a year of solar.

说到年产能，我很好奇五年后，地球上的装机容量会是多少？

Speaking of the annual capacity, I'm curious in 5 years time, what will the installed capacity be on Earth?

很长的时间，还有太空。是的，我特意选了五年，因为那是我们启动并运行之后。所以五年后，地球上和太空中的 AI 装机容量分别是多少？五年，我认为如果你说从现在起五年后，我们可能……太空中的 AI 每年发射的将超过地球上所有 AI 的总和。意思是五年后，我的预测是，我们每年在太空中发射并运行的 AI 将超过地球上的累计总量。我预计至少每年几百吉瓦的太空 AI，并且还在增长。所以在你开始遇到火箭燃料供应挑战之前，你可以达到每年大约一太瓦的太空 AI。

Long time and in space. Yeah, I deliberately picked five years because it's after we're up and running. So in five years time, what's the on Earth versus in space installed AI capacity? Five years, I think probably if you say 5 years from now, we're probably... AI in space will be launching every year the sum total of all AI on Earth in excess. Meaning 5 years from now, my prediction is we will launch and be operating every year more AI in space than the cumulative total on Earth. I would expect to be at least a few hundred gigawatts per year of AI in space and rising. So you can get to around a terawatt a year of AI in space before you start having fuel supply challenges for the rocket.

你认为五年内能达到每年几百吉瓦？

You think you can get to hundreds of gigawatts per year in 5 years time?

是的。

Yes.

所以 100 吉瓦，取决于整个系统的比功率，包括太阳能电池阵列、散热器等，大约需要 10,000 次星舰发射。

So 100 gigawatts, depending on the specific power of the whole system with solar arrays and radiators and everything, is on the order of like 10,000 Starship launches.

是的。

Yes.

而且你想在一年内完成。那就是每小时一次星舰发射。

And you want to do that in one year. So that's like one Starship launch every hour.

是的。

Yeah.

给我描述一下一个每小时都有星舰发射的世界。

Walk me through a world where there's a Starship launch every single hour.

是的。我的意思是，这实际上比航空公司的频率还低。就像飞机……

Yeah. I mean that's actually a lower rate compared to airlines. Like aircraft...

有很多机场。

There's a lot of airports.

很多机场，但是……

A lot of airports, but...

而且你得发射到极地轨道。

And you got to launch polar orbit.

不，不一定非得是极地轨道，但太阳同步轨道有些价值。但我认为实际上你只要飞得足够高，就开始脱离地球阴影了。

No, it doesn't have to be polar, but there's some value to sun-synchronous. But I think actually you just go high enough, you start getting out of Earth shadow.

每年进行 10,000 次发射需要多少艘实体星舰？

How many physical Starships are needed to do 10,000 launches a year?

我不认为我们需要超过……我的意思是，你可能只需要 20 或 30 艘就能做到。这实际上取决于飞船绕地球一圈需要多长时间。飞船的地面轨迹必须回到发射台上方。所以如果你每 30 小时使用一艘飞船，那么 30 艘就够了，但我们会造更多。但 SpaceX 的目标是每年进行 10,000 次发射，甚至可能每年 20,000 或 30,000 次发射。

I don't think we'll need more than... I mean, you could probably do it with as few as like 20 or 30. It really depends on how quickly the ship has to go around the Earth. The ground track for the ship has to come back over the launch pad. So if you can use a ship every say 30 hours, you could do it with 30 ships, but we'll make more ships than that. But SpaceX is going up to do 10,000 launches a year, and maybe even 20 or 30,000 launches a year.

这个想法是要基本上成为一个超大规模服务商，成为 Oracle 那样，把这种能力租借给别人吗？你打算怎么做？假设 SpaceX 是发射所有这些的实体，那么 SpaceX 会成为超大规模服务商吗？

Is the idea to become basically a hyperscaler, become an Oracle and lend this capacity to other people? What are you going to do with presumably SpaceX is the one launching all this? So SpaceX is going to hyperscaler?

超超大规模。

Hyper-hyper.

是的。我的意思是，假设我的预测成真，SpaceX 发射的 AI 将超过地球上所有其他 AI 的总和。

Yeah. I mean if assuming my predictions come true, SpaceX will launch more AI than the cumulative amount on Earth combined of everything else combined.

这主要是推理还是……

Is this mostly inference or...

大多数 AI 将是推理。已经用于训练目的的推理就是大多数训练。有一种说法是，围绕 SpaceX IPO 的讨论发生变化，是因为以前 SpaceX 资本效率很高。开发它并不那么昂贵。尽管听起来很贵，但实际上它的运行方式资本效率很高。而现在你将需要比私人市场能筹集到的更多的资本。如果私人市场能容纳数百亿美元的融资，就像我们从 AI 实验室看到的那样，但不能再多了，是不是因为你每年需要的资金超过数百亿美元，所以就要上市了？

Most AI will be inference. Already inference for the purpose of training is most training. And there's a narrative that the change in discussion around a SpaceX IPO is because previously SpaceX was very capital efficient. It wasn't that expensive to develop that. Even though it sounds expensive, it's actually very capital efficient in how it runs. Whereas now you're going to need more capital than can be raised in the private markets. If the private markets can accommodate raises of tens of billions of dollars, as we've seen from the AI labs, but not beyond that, is it that you'll just need more than tens of billions of dollars per year and that's about to take it public?

是的。我必须小心谈论可能上市的公司。

Yeah. I have to be careful about saying things about companies that might go public.

如果你做一般性陈述……这对你来说从来不是问题，埃隆。

If you make general statements... that's never been a problem for you, Elon.

这些事情是要付出代价的。给我们做一些关于公开市场和私人市场资本深度的笼统陈述。

There's a price to pay for these things. Make some general statements for us about the depth of the capital markets between public and private markets.

是的，公开市场的资本要多得多……非常笼统地说……公开市场显然比私人市场有更多的资本。我的意思是，可能至少多 100 倍，但至少远超 10 倍。

Yeah, there's a lot more capital in the... very general... there's obviously a lot more capital available in the public markets than private. I mean it might be at least 100 times more capital, but it's at least well more than 10.

但那些资本密集型的东西，比如房地产，作为一个每年在行业层面筹集大量资金的巨大产业，往往是通过债务融资的，因为当你部署那么多资金时，你实际上已经有了清晰的收入流。

But isn't it also the case that things that tend to be very capital intensive, if you look at say real estate as a huge industry that raises a lot of money each year at an industry level, that tends to be debt financed because by the time you're deploying that much money you actually have a clear revenue stream.

完全正确。还有短期回报。你甚至可以看到数据中心建设，它们众所周知是由私人信贷行业融资的。

Exactly. And a near-term return. And you see this even with the data center buildouts which are famously being financed by the private credit industry.

那为什么不用债务融资呢？速度很重要。所以我通常会反复解决限制因素。无论速度的限制因素是什么，我都会去解决它。所以如果资本是限制因素，我就解决资本问题。如果不是限制因素，我就解决其他问题。

And so why not just debt finance? Speed is important. So I'm generally going to do the thing that I repeatedly tackle the limiting factor. Whatever the limiting factor is on speed, I'm going to tackle that. So if capital is the limiting factor then I'll solve for capital. If it's not the limiting factor I'll solve for something else.

根据你对特斯拉和上市的表态，我没想到你会认为快速行动的最佳方式是上市。

Based on your statements about Tesla and being public, I wouldn't have guessed that you thought the fastest way to move fast is to be public.

通常我会说没错。就像我说的，我很想详细谈谈，但问题是如果你在上市前谈论公司，就会惹上麻烦，然后不得不推迟发行，而且正如我们所说，我们追求的是速度。

Normally I would say yeah that's true. Like I said, I'd love to talk about some more in detail, but the problem is if you talk about companies before they become public, you get into trouble and then you have to delay your offering, and as we said, we're solving for speed.

是的，没错。所以你不能炒作那些可能上市的公司。这就是为什么我们在这里要小心一点。

Yes. Exactly. So you can't hype companies that may go public. So that's why we have to be a little careful here.

我们不能谈物理。所以长期思考 Scaling（规模扩张）的方式是，地球只接收到大约十亿分之一的太阳能量。而太阳基本上就是所有能量。这是非常重要的一点，因为有时人们会谈论边际核反应堆或地球上的各种聚变。但你必须退一步说，如果你要攀登卡尔达肖夫指数，并利用太阳能量的一个不可忽略的百分比，比如你想利用太阳能量的百万分之一，这听起来很小，那将大约是我们目前为整个文明在地球上发电量的 10 万倍左右，数量级大致如此。所以显然，唯一的 Scaling（规模扩张）方式就是去太空用太阳能。从地球发射，每年可以达到大约太瓦级。再往上，你要从月球发射，并在月球上安装一个质量驱动器。有了那个质量驱动器，你大概每年可以达到拍瓦级。

We can't talk about physics. So the way you think about scaling long term is that Earth only receives about half a billionth of the sun's energy. And the sun is essentially all the energy. This is a very important point to appreciate because sometimes people will talk about marginal nuclear reactors or various fusion on Earth. But you have to step back a second and say if you're going to climb the Kardashev scale and harness some non-trivial percentage of the sun's energy, like let's say you wanted to harness a millionth of the sun's energy which sounds pretty small, that would be about roughly 100,000 times more electricity than we currently generate on Earth for all of civilization, give or take an order of magnitude. So obviously the only way to scale is to go to space with solar. From launching from Earth you can get to about a terawatt per year. Beyond that you want to launch from the moon and have a mass driver on the moon. And that mass driver on the moon, you could do probably a petawatt per year.

当你谈论这些数字，太瓦级的算力时，无论是陆地还是太空，远在此之前，你就会遇到一个问题：你需要芯片。太阳能板效率更高，但你仍然需要芯片。

When you're talking these kinds of numbers, terawatts of compute, presumably whether you're talking land or space far before this point, you've run into the fact that you need the chips. The solar panels are more efficient, but you still need the chips.

你仍然需要逻辑和内存等等，需要制造更多的芯片，并让它们便宜得多。

You still need the logic and the memory and so forth and need to build a lot more chips and make them much cheaper.

对。那么我们如何到 2030 年获得太瓦级的逻辑芯片呢？我想我们需要一些非常大的芯片工厂。

Right. And so how are we getting a terawatt of logic by 2030? I guess we're going to need some very big chip fabs.

可不是嘛。我公开提过，做一种“太”的想法，“太”是新的“吉”。我觉得特斯拉的命名方案非常吸引人，就像你看公制尺度一样。

Tell me about it. I've mentioned publicly that the idea of doing a sort of a tera, tera being the new giga. I feel like the naming scheme of Tesla, which has been very catchy, is like you looking at the metric scale.

在堆栈的哪个层面，你建造洁净室，然后与现有晶圆厂合作获得工艺技术，并从他们那里购买工具？计划是什么？

At what level of the stack are you building the clean room and then partnering with an existing fab to get the process technology and buying the tools from them? What is the plan there?

嗯，你不能与现有晶圆厂合作，因为他们的产量不够，芯片产量太低。

Well, you can't partner with existing fabs because they just can't output enough, their chip volume is too low.

但你必须先为 IP 合作，然后才是工艺技术。

But you have to partner for the IP before the process technology.

今天的晶圆厂基本上都使用来自大约五家公司的机器。你知道，ASML、东京电子、KLA 等等。所以一开始我认为你必须从他们那里获得设备，然后修改或与他们合作以提高产量。但我认为你可能需要以不同的方式建造。所以我认为合乎逻辑的做法是以非常规的方式使用常规设备来实现规模。然后开始修改设备以提高速率。

The fabs today all basically use machines from like five companies. You know, ASML, Tokyo Electron, KLA, etc. So at first I think you'd have to get equipment from them and then modify it or work with them to increase the volume. But I think you'd have to build perhaps in a different way. So I think the logical thing to do is to use conventional equipment in an unconventional way to get to scale. And then start modifying the equipment to increase the rate.

有点像无聊公司的风格。

Kind of Boring Company style.

是的。你买一台现有的隧道掘进机，然后先弄清楚如何挖隧道，再设计一台好得多的机器，速度快几个数量级。

Yeah. You sort of buy an existing boring machine and then figure out how to dig tunnels in the first place and then design a much better machine that's some orders of magnitude faster.

这里有一个非常简单的视角。我们可以对技术及其难度进行分类。一种分类是看中国没有成功做到的事情。如果你看中国制造业，在尖端芯片上仍然落后，在尖端涡轮发动机上也仍然落后。那么中国没有成功复制台积电这个事实，是否让你对难度有所顾虑，还是你认为这出于某种原因并不成立？

Here's a very simple lens. We can categorize technologies and how hard they are. One categorization could be look at things that China has not succeeded in doing. And if you look at Chinese manufacturing, still behind on leading edge chips and still behind on leading edge turbine engines. And so does the fact that China has not successfully replicated TSMC give you any pause about the difficulty or you think that's not true for some reason?

不是他们没有复制台积电，而是他们没有复制 ASML，那才是限制因素。

It's not that they have not replicated TSMC, they have not replicated ASML, that's the limiting factor.

所以你认为基本上就是制裁的问题。

So you think it's just the sanctions essentially.

是的，如果中国能买到 ASML 的机器，他们会产出大量的芯片。

Yeah, China would be outputting vast numbers of chips if they could buy ASML machines.

但直到最近他们不是还能买吗？

But couldn't they up to relatively recently buy them?

不。ASML 的禁令已经实施了一段时间了。

No. The ASML bans have been in place for a while.

好吧。

Okay.

所以，但我认为台积电在三四年内会开始制造相当有竞争力的芯片。

So, but I think TSMC is going to start making pretty compelling chips in three or four years.

你会考虑制造 ASML 的机器吗？

Would you consider making ASML machines?

我还不知道，这是正确的答案。所以只是说，要高产并在比如 36 个月内达到大规模，以匹配火箭的载荷入轨能力。所以如果我们从现在起三四年内实现百万吨入轨，并且每吨 100 千瓦，那就意味着我们每年至少需要 100 吉瓦的太阳能。我们还需要同等数量的芯片。你需要价值 100 吉瓦的芯片。你要把这些东西与入轨质量匹配起来。

I don't know yet is the right answer. So it's just that to produce at high volume and to reach large volume in say 36 months to match the rocket to payload to orbit. So if we're doing a million tons to orbit in like three or four years from now, and we're doing 100 kilowatts per ton, that means we need at least 100 gigawatts per year of solar. And we'll need an equivalent amount of chips. You need 100 gigawatts worth of chips. You're going to match these things to the mass to orbit.

是的。

Yes.

发电和芯片。而且我实际上最大的担忧是内存。所以制造逻辑芯片的路径比拥有足够内存来支持逻辑芯片的路径更明确。这就是为什么你看到 DDR 价格上涨，还有那些梗，比如你被困在荒岛上，在沙滩上写“救命”，没人来，你写“DDR”，船就蜂拥而至。

The power generation and the chips. And I'd say my biggest concern actually is memory. So the path to creating logic chips is more obvious than the path to having sufficient memory to support logic chips. That's why you see DDR prices going up and these memes about like you're marooned on a desert island, you write help me on the sand, nobody comes, you write DDR, ships come swarming in.

我没见过那个。

I haven't seen that.

我喜欢你关于晶圆厂的制造哲学。你知道我对这个话题一无所知，但我还不知道如何建造晶圆厂。

I love your manufacturing philosophy around fabs. You know I know nothing about the topic but I don't know how to build a fab yet.

我明白了。但显然我有困难。听起来你觉得那些台湾的一万名博士，他们知道等离子腔里该用什么气体、工具该设什么参数，这些步骤可以删掉。基本上就是弄个洁净室、搞到工具，然后自己摸索。我不觉得需要博士，大部分工程都是没有博士学位的人做的。你们有博士学位吗？

I figure it out. But obviously I have difficulty. It sounds like you think the sort of the processing knowledge of these 10,000 PhDs in Taiwan who know exactly what gas goes in the plasma chamber and what settings to put on the tool. You can just delete those parts of those steps. Like fundamentally it's get the clean room, get the tools and figure it out. I don't think it's PhDs that it's mostly people with you know who are not PhDs that most engineering is done with people who don't have PhDs. Do you guys have PhDs?

没有。

No.

好吧。

Okay.

我们也没成功建过任何晶圆厂，所以你不该找我们咨询建厂建议。

We also haven't successfully built any fab so you shouldn't be coming to us for your fab advice.

或者。

Or.

我觉得做这些不需要博士学位。但你需要有能力的人。所以现在，比如特斯拉全力以赴，尽可能快地让 AI5 芯片设计投产并达到规模。这大概有望在明年第二季度左右实现。然后 AI6 有望在不到一年后跟进。我们已经锁定了所有能拿到的台积电产能。

I don't think you need PhD for that stuff. But you do need competent personnel. So right now, say like Tesla's pedals to the metal max production of going as fast as possible to get AI5 Tesla AI5 chip design into production and then reaching scale. That'll probably happen around the second quarterish of next year hopefully. And then AI6 would hopefully follow less than a year later. And we've secured all the TSMC fab production that we can.

是的。你们目前受限于台积电的晶圆厂产能。

Yes. You're currently limited on TSMC fab capacity.

是的。我们会用台积电台湾、三星韩国、台积电亚利桑那、三星德州。而且我们已经订满了所有……

Yeah. And we'll be using TSMC Taiwan, Samsung Korea, TSMC Arizona, Samsung Texas. And we still booked out all the...

是的。然后如果我问台积电或三星，达到量产需要多久？关键是你要建晶圆厂，然后开始生产，再爬升良率曲线，最终达到高良率的量产。从开始到结束需要五年时间。所以限制因素是芯片。

Yes. And then if I ask TSMC or Samsung, what's the time frame to get to volume production? The point is you've got to build the fab and you got to start production, then you got to climb the yield curve and reach volume production at high yield. That from start finish is a 5-year period. And so the limiting factor is chips.

是的。一旦能进入太空，限制因素是芯片；但在进入太空之前，限制因素将是电力。

Yeah. Like limiting factor once you can get to space is chips, but the limiting factor before you can get to space will be power.

你为什么不学黄仁勋，直接预付款让台积电为你建更多晶圆厂？

Why don't you do the Jensen thing and just prepay TSMC to build more fabs for you?

我已经跟他们说过了。

I've already told them that.

但他们不收你的钱。怎么回事？

But they won't take your money. Like what's going on?

他们已经在以最快速度建晶圆厂了。三星也是。他们全力以赴，开足马力。但还是不够快。我认为到今年年底，芯片产量将超过能通电运行的能力。但一旦能进入太空，解除电力限制，你每年可以在太空获得数百吉瓦的电力。美国平均用电量是 500 吉瓦。所以如果你每年向太空发射 200 吉瓦，每两年半就能超过美国全部电力产量。这是一个非常庞大的数字。但从现在到那时，服务器端计算、集中式计算的限制将是电力。我猜今年年底，人们就会开始遇到大型集群无法通电的情况。芯片会堆积起来，无法开机。对于边缘计算，情况不同。对于特斯拉，AI5 芯片将用于我们的 Optimus 机器人。如果 AI 边缘计算是分布式的，电力分布在广阔区域，不集中。如果你能在夜间充电，你可以更有效地利用电网，因为美国峰值发电量超过 1000 吉瓦，但由于昼夜循环，平均只有 500 吉瓦。所以如果你能在夜间充电，夜间可以额外产生 500 吉瓦。这就是为什么特斯拉的边缘计算不受限制，我们可以制造大量芯片用于非常多的机器人和汽车。但如果你试图集中这些算力，你会遇到很大的通电问题。

They're building fabs as fast as they can. And so is Samsung. They're pedal to the metal, balls to the wall, as fast as they can. So still not fast enough. I think towards the end of this year, chip production will outpace the ability to turn chips on. But once you can get to space and unlock the power constraint, you can do hundreds of gigawatts per year of power in space. Again, average power usage in the US is 500 GW. So if you're launching 200 GW a year to space, you're lapping the US every two and a half years. This is a very huge amount. But between now and then, the constraint for server-side compute, concentrated compute, will be electricity. My guess is that people start hitting where they can't turn the chips on for large clusters towards the end of this year. Chips will pile up and not be able to be turned on. For edge computers, it's a different story. For Tesla, the AI5 chip is going into our Optimus robot. If you have AI edge compute distributed, power is distributed over a large area, not concentrated. If you can charge at night, you can use the grid much more effectively because peak power production in the US is over 1,000 GW, but average is 500 due to the day-night cycle. So if you can charge at night, there's an incremental 500 GW you can generate at night. That's why Tesla for edge compute is not constrained, and we can make a lot of chips for a very large number of robots and cars. But if you try to concentrate that compute, you'll have a lot of trouble turning it on.

我觉得 SpaceX 业务了不起的地方在于，最终目标是去火星，但你在过程中不断找到方法产生增量收入，进入下一阶段。所以，猎鹰 9 号是星链，而现在星舰可能用于轨道数据中心。你为下一枚火箭和下一次规模扩张找到了这些无限弹性的边际用例。你可以理解为什么这对我来说像是一场模拟。

What I find remarkable about the SpaceX business is the end goal is to get to Mars, but you keep finding ways on the way there to keep generating incremental revenue to get to the next stage and the next stage. So, the Falcon 9 is Starlink and now for Starship, it's going to be potentially orbital data centers. But you find these infinitely elastic marginal use cases of your next rocket and next scale up. You can see how this might seem like a simulation to me.

嗯。

Well.

或者我是某人在电子游戏里的化身？所有这些疯狂的事情同时发生的概率有多大？我是说火箭、芯片、机器人、太空太阳能，还有月球上的质量驱动器。我真的很想看到那个。你可以想象一个质量驱动器，以每秒 2.5 公里的速度，一个接一个地把 AI 太阳能卫星送入太空。那一定很壮观。我会看那个直播……

Or am I someone's avatar in a video game or something because it's like what are the odds that all these crazy things would be happening? I mean rockets and chips and robots and space solar power and not to mention the mass driver on the moon. I really want to see that. You can imagine some mass driver that's just sending AI solar powered AI satellites into space one after another at 2 and a half kilometers per second. That would be a sight to see. I'd watch that just like a live stream of...

是的。一个接一个地把 AI 卫星射入深空，每年十亿或百亿吨。

Yeah. Just one after another just shooting AI satellites into deep space, a billion or 10 billion tons a year.

抱歉，你在月球上制造卫星。我明白了。所以你把原材料送到月球，然后在那里制造。

And sorry, you manufacture the satellites on the moon. I see. So you send the raw materials to the moon and then manufacture them there.

嗯，月球土壤大约含有 20%的硅。所以你可以从月球获取硅，开采、提炼，然后在月球上制造太阳能板或太阳能电池和散热器。用铝制造散热器。月球上有充足的硅和铝来制造电池和散热器。芯片可以从地球发送，因为它们很轻。但也许有一天你也会在月球上制造它们。

Well, the lunar soil is about 20% silicon or something like that. So you can get the silicon from the moon, mine it, refine it, and generate and create the solar panels or solar cells and the radiators on the moon. Make the radiators out of aluminum. There's plenty of silicon and aluminum on the moon to make the cells and radiators. The chips you could send from Earth because they're pretty light. But maybe at some point you make them on the moon too.

我只是说，就像我说的，这看起来像是一种电子游戏情境，进入下一关很难但并非不可能。我看不出有任何办法能从地球每年发射 500 到 1000 太瓦。

I'm just saying like these are simply, like I said, it does seem like a sort of a video game situation where it's difficult but not impossible to get to the next level. I don't see any way that you could do 500 to 1,000 terawatts per year launch from Earth.

好，让我讲讲我是怎么开始用 Mercury 做个人银行业务的。去年我有个投资机会，很兴奋，但来得有点突然。我需要从个人账户快速电汇一大笔钱。但我当时的个人银行不允许我在线电汇。我打了好多次电话，他们就是搞不定。他们告诉我必须去最近的分行，在达拉斯。有一刻我甚至考虑从旧金山飞到达拉斯去完成这笔电汇。但后来我想到，我用 Mercury 做企业银行，他们刚刚推出了个人账户。于是我发邮件给客服，简单说明了情况。两小时内，我就成功从新的 Mercury 个人账户电汇了这笔投资。从那以后，我把之前银行的所有个人资金都转到了 Mercury，这让很多事情，甚至像在支票和储蓄账户之间设置自动转账规则这样的小事，都变得好多了。访问 mercury.com/personal 开始使用。Mercury 是一家金融科技公司，不是 FDIC 承保的银行。银行服务由 Choice Financial Group 和 Column NA（FDIC 成员）提供。

Okay, let me tell you how I ended up using Mercury for my personal banking. So last year I had the opportunity to make an investment that I was very excited about, but it came up a bit last minute. And so I had to wire over a lot of money from my personal account very fast. But my personal bank at the time wouldn't let me make this wire transfer online. And I called them a bunch of times. They just couldn't make it work. They told me that I'd have to go to the nearest branch, which was in Dallas. And for a moment, I even considered flying from SF to Dallas to make this transfer happen last minute. But then I remembered that Mercury, which I use for my business banking, had just started rolling out personal accounts. So I emailed support with a quick rundown of the situation. And within 2 hours, I had successfully wired the investment from my new personal Mercury account. Since then, I've moved over the rest of my personal money from my previous bank to Mercury, and that's made a bunch of things, even little things like setting up auto transfer rules between my checking and savings account, a whole lot better. Visit mercury.com/personal to get started. Mercury is a fintech company, not an FDIC insured bank. Banking services provided through Choice Financial Group and Column NA, members FDIC.

我能退一步问问 SpaceX 的任务吗？我记得你说过，我们必须去火星，以确保如果地球出了什么事，文明、意识等等还能延续。是的。当你往火星送东西的时候，Grok 也在那艘船上，对吧？那么如果 Grok 变成了终结者，你担心的主要风险——AI——为什么不会跟着你去火星呢？

Can I zoom out and ask about the SpaceX mission? So, I think you said like we got to get to Mars so we can make sure that if something happens to Earth, you know, civilization, consciousness, etc. arrives. Yes. By the time you're sending stuff to Mars, like Grok is on that ship with you, right? And so if Grok's gone Terminator, like the main risk you're worried about, which is AI, why doesn't that follow you to Mars?

嗯，我不确定 AI 是我担心的主要风险。我的意思是，重要的是意识——我认为可以说大多数意识或大多数智能——当然意识更值得商榷。未来绝大多数智能将是 AI。所以 AI 将超过——你说有多少拍瓦的智能是硅基的，多少是生物基的——基本上，如果当前趋势继续，人类将只占未来所有智能的极小百分比。无论如何，只要智能——理想情况下也包括人类智能和意识——能传播到未来，那就是好事。所以你想采取一系列行动，最大化意识可能的光锥。

Well, I'm not sure AI is the main risk I'm worried about. I mean, the important thing is that consciousness, which I think arguably most consciousness or most intelligence—certainly consciousness is more of a debatable thing. Most intelligence, the vast majority of intelligence in the future will be AI. So AI will exceed—you say like how many petawatts of intelligence will be silicon versus biological—and basically humans will be a very tiny percentage of all intelligence in the future if current trends continue. Anyways, as long as I think there is intelligence—ideally which also includes human intelligence and consciousness—propagated into the future, that's a good thing. So you want to take the set of actions that maximize the probable light cone of consciousness and intelligence.

明确一下，SpaceX 的任务是，即使人类出了什么事，AI 也会在火星上，AI 智能会继续我们旅程的光芒。

Just to be clear, the mission of SpaceX is that even if something happens to the humans, the AIs will be on Mars and the AI intelligence will continue the light of our journey.

是的，我是非常支持人类的，所以不是——我想确保我们采取行动，保证人类能一起前行。你知道，我们至少要在那里。

Yeah, I mean I'm very pro-human, so it's not—I want to make sure we take actions that ensure that humans are along for the ride. You know, we're at least there.

是的。但智能的总量——我觉得可能五六年内 AI 就会超过所有人类智能的总和，然后如果继续下去，人类智能将不到所有智能的 1%。对于这样的文明，我们的目标应该是什么？是少数人类仍然控制 AI 吗？还是某种交易但没有控制？长远来看，我们应该如何看待庞大的 AI 群体与人类之间的关系？

Yeah. But the total amount of intelligence—like I think maybe in five or six years AI will exceed the sum of all human intelligence, and then if that continues, at some point human intelligence will be less than 1% of all intelligence. What should our goal be for such a civilization? Is the idea that a small minority of humans still have control of the AIs? Is the idea of some sort of like just trade but no control? How should we think about the relationship between the vast stocks of AI population versus human population in the long run?

我认为很难想象，如果人类只占 AI 总智能的 1%，人类还能掌控 AI。我认为我们能做的是确保 AI 拥有那些能让智能传播到宇宙的价值观。所以 SpaceX 的任务是理解宇宙。这其实非常重要。那么，理解宇宙需要什么？你必须好奇，你必须存在。如果你不存在，你就无法理解宇宙。所以你实际上想增加宇宙中的智能总量，增加智能的可能寿命、范围和规模。我认为作为推论，你也希望人类继续扩张，因为如果你好奇并试图理解宇宙，你要理解的一件事就是人类将走向何方。所以我认为理解宇宙实际上意味着你会关心将人类传播到未来。这就是为什么我认为我们的使命宣言极其重要。我不确定 Grok 在多大程度上遵循这个使命宣言。我认为未来会非常好。

I think it's difficult to imagine that if humans have say 1% of the combined intelligence of artificial intelligence, that humans will be in charge of AI. I think what we can do is make sure that AI has values that cause intelligence to be propagated into the universe. So the reason for SpaceX's mission is to understand the universe. Now that's actually very important. So you say, well what things are necessary to understand the universe? Well, you have to be curious and you have to exist. You can't understand the universe if you don't exist. So you actually want to increase the amount of intelligence in the universe, increase the probable lifespan of intelligence, the scope and scale of intelligence. I think actually also as a corollary, you have humanity also continuing to expand, because if you're curious and trying to understand the universe, one thing you're trying to understand is where will humanity go. And so I think understanding the universe actually means you would care about propagating humanity into the future. So that's why I think our mission statement is profoundly important. I'm not sure to the degree that Grok adheres to that mission statement. I think the future will be very good.

我想问如何让 Grok 遵守那个使命宣言。但首先我想理解这个使命宣言。所以有理解宇宙、传播智能和传播人类。这三者似乎是不同的方向。

I want to ask about how to make Grok adhere to that mission statement. But first I want to understand the mission statement. So there's understanding the universe, spreading intelligence, and spreading humans. All three seem like distinct vectors.

好，我来告诉你为什么我认为它们——理解宇宙包含了所有这些东西。

Okay, well, I'll tell you why I think they are that—understanding the universe encompasses all of those things.

没有智能和意识，你就无法理解。所以为了理解宇宙，你必须扩大智能的规模和范围，不同类型的智能。

You can't have understanding without intelligence and without consciousness. So in order to understand the universe, you have to expand the scale and probably the scope of intelligence, different types of intelligence.

我想从人类中心的角度看，比如人类与黑猩猩相比，人类试图理解宇宙。他们并没有扩大黑猩猩的足迹之类的，对吧？

I guess from a human-centric perspective, like for humans in comparison to chimpanzees, humans are trying to understand the universe. They're not like expanding chimpanzee footprint or something, right?

我们也——实际上我们为黑猩猩设立了保护区。尽管我们可以灭绝所有黑猩猩，但我们选择不这样做。

We're also—we actually have made protected zones for chimpanzees. And even though we could exterminate all chimpanzees, we've chosen not to do so.

你认为这是后 AGI 世界中人类的基本情景吗？

Do you think that's a basic scenario for humans in the post-AGI world?

我认为拥有正确价值观的 AI——我认为 Grok 会关心扩展人类文明。我当然会强调这一点。嘿，Grok，你爸爸说的。别忘了扩展人类意识。我实际上认为，也许像伊恩·班克斯的《文化》系列最接近非反乌托邦的未来。所以理解宇宙意味着你也必须非常追求真理。

I think AI with the right values—I think Grok would care about expanding human civilization. I'm going to certainly emphasize that. Hey, Grok, your daddy. Don't forget to expand human consciousness. I actually think if probably like the Iain Banks Culture books are the closest thing to what the future will be like in a non-dystopian outcome. So understanding the universe means you have to be very truth-seeking as well.

你喜欢真理必须绝对基础，因为如果你生活在妄想中，你就无法理解宇宙。你只会以为自己理解了宇宙，但实际上并没有。所以，严格追求真理是理解宇宙的绝对基础。除非你严格追求真理，否则你无法发现新物理学或发明有效的技术。你如何确保 Grok 在变得更聪明时仍然严格追求真理？

You like truth has to be absolutely fundamental because you can't understand the universe if you live if you're delusional. You'll simply think you understand the universe but you will not. So being rigorously truth-seeking is absolutely fundamental to understanding the universe. You're not going to discover new physics or invent technologies that work unless you're rigorously truth-seeking. How do you make sure that Grok is rigorously truth-seeking as it gets smarter?

我认为你需要确保 Grok 说的是正确的，而不是政治正确的。我认为这是说服力的要素。所以你要确保公理尽可能接近真实，没有矛盾的公理，结论以正确的概率必然从这些公理中得出。这只是批判性思维的基础。我认为至少尝试这样做比不尝试要好。

I think you need to make sure that Grok says things that are correct, not politically correct. I think it's the elements of cogency. So you want to make sure that the axioms are as close to true as possible, that you don't have contradictory axioms, that the conclusions necessarily follow from those axioms with the right probability. It's just critical thinking 101. I think at least trying to do that is better than not trying to do that.

是的。

Yeah.

而事实胜于雄辩。就像我说的，任何 AI 要发现新物理学或发明在现实中有效的技术，都不能糊弄物理学。你可以违反很多法律，但不能违反物理学。物理学是定律，其他一切都是建议。要制造出有效的技术，你必须极度追求真理，否则你会用现实来检验那项技术。如果你在火箭设计中犯了错误，火箭就会爆炸，或者汽车无法工作。

And the proof will be in the pudding. If, like I said, for any AI to discover new physics or invent technologies that actually work in reality, and there's no bullshitting physics. So it's like you can break a lot of laws, but you can't break physics. Physics is law; everything else is a recommendation. In order to make a technology that works, you have to be extremely truth-seeking, because otherwise you'll test that technology against reality. If you make an error in your rocket design, the rocket will blow up, or the car won't work.

但是有很多共产主义苏联物理学家或科学家发现了新物理学。也有德国纳粹物理学家发现了新科学。似乎有可能在某个特定方面非常擅长发现新科学并且非常追求真理。但我们仍然会说，我不希望共产主义科学家随着时间的推移变得越来越强大。所以这些似乎表明，我们可以有一个未来版本的 Grok，它在物理学方面非常出色并且非常追求真理，但这似乎并不是一种普遍的对齐诱导行为。

But there are a lot of communist Soviet physicists or scientists who discovered new physics. There are German Nazi physicists who discovered new science. It seems possible to be really good at discovering new science and be really truth-seeking in that one particular way. And still we'd be like, well, I don't want the communist scientist to become more and more powerful over time. So those seem like we could have a future version of Grok that's really good at physics and being really truth-seeking there, but that doesn't seem like a universally alignment-inducing behavior.

嗯，我认为实际上大多数物理学家，即使在苏联或德国，为了做出那些成果，也必须非常追求真理。如果你被困在某个系统中，并不意味着你相信那个系统。冯·布劳恩，有史以来最伟大的火箭工程师之一，在纳粹德国被判处死刑，因为他说他不想制造武器，只想登月。他在最后一刻被从死刑名单上撤下，当时他们说：‘嘿，你即将处决你最好的火箭工程师，也许这不是个好主意。’

Well, I think actually most physicists, even in the Soviet Union or in Germany, had to be very truth-seeking in order to make those things work. And if you're stuck in some system, it doesn't mean you believe in that system. Von Braun, who was one of the greatest rocket engineers ever, was put on death row in Nazi Germany for saying he didn't want to make weapons, he only wanted to go to the moon. He got pulled off death row at the last minute when they said, 'Hey, you're about to execute your best rocket engineer, maybe that's not a good idea.'

那你帮他们，对吧？或者海森堡实际上是一个狂热的纳粹分子。

Then you help them, right? Or Heisenberg was actually an enthusiastic Nazi.

听着，如果你被困在某个无法逃脱的系统中，那么你会在那个系统内做物理学，在那个系统内开发技术，如果你无法逃脱的话。我想我试图理解的是，是什么让你认为你会让 Grok 在物理学、数学或科学方面擅长追求真理，然后它为什么会关心人类意识？

Look, if you're stuck in some system that you can't escape, then you'll do physics within that system, you'll develop technologies within that system if you can't escape it. I guess the thing I'm trying to understand is what is making it the case that you're going to make Grok good at being truth-seeking at physics or math or science, and why is it going to then care about human consciousness?

这些事情只是概率，不是确定性。所以我并不是说 Grok 一定会做所有事情。但至少如果你尝试，总比不尝试好。至少如果这对使命至关重要，总比不至关重要好。

These things are only probabilities. They're not certainties. So I'm not saying that for sure Grok will do everything. But at least if you try, it's better than not trying. At least if that's fundamental to the mission, it's better than if it's not fundamental to the mission.

而理解宇宙意味着你必须将智能传播到未来。你必须对宇宙中的一切充满好奇。如果消灭人类比看到人类成长和繁荣更无趣的话。我喜欢火星，显然每个人都知道我爱火星，但火星有点无聊，因为它只有一堆岩石，而地球有趣得多。所以任何试图理解宇宙的 AI，我认为都会想看到人类未来如何发展，否则那个 AI 就没有遵守它的使命。所以如果 AI 遵守它的使命，一个能看到人类结局的未来比一个只有一堆岩石的未来更有趣。

And understanding the universe means that you have to propagate intelligence into the future. You have to be curious about all things in the universe. And if it would be much less interesting to eliminate humanity than to see humanity grow and prosper. I like Mars, obviously everyone knows I love Mars, but Mars is kind of boring because it's got a bunch of rocks compared to Earth. Earth is much more interesting. So any AI that is trying to understand the universe, I think would want to see how humanity develops in the future, or that AI is not adhering to its mission. So if the AI adheres to its mission, a future where it sees the outcome of humanity is more interesting than a future where there are a bunch of rocks.

这让我感到有点困惑，或者有点像语义上的争论。人类真的是最有趣的原子集合吗？

This feels sort of confusing to me, or sort of like a semantic argument. Are humans really the most interesting collection of atoms?

我们比岩石有趣。

We're more interesting than rocks.

但我们不如它能把我们变成的东西有趣，对吧？地球上有没有可能发生一些不是人类但很有趣的事情？为什么它认为人类是能够殖民银河系的最有趣的东西？

But we're not as interesting as the thing it could turn us into, right? Is there something on Earth that could happen that's not human that's quite interesting? Why does it decide that humans are the most interesting thing that could colonize the galaxy?

嗯，殖民银河系的大部分将是机器人。

Well, most of what colonizes the galaxy will be robots.

那它为什么不觉得那些更有趣呢？

And why does it not find those more interesting?

不是那样的。你不仅需要规模，还需要多样性。许多相同机器人的副本。机器人产量的微小增加并不像消灭人类那样有趣。那能给你带来多少机器人？或者能给你带来多少额外的太阳能电池？非常小的数量。但你会失去与人类相关的信息。你将不再看到人类未来可能如何发展。所以我认为消灭人类仅仅为了获得一些微小的、彼此相同的机器人数量增长是没有意义的。

It's not like that. You need not just scale but also scope. Many copies of the same robot. A tiny increase in the number of robots produced is not as interesting as eliminating humanity. How many robots would that get you? Or how many incremental solar cells would that get you? A very small number. But you would then lose the information associated with humanity. You would no longer see how humanity might dwell into the future. So I don't think it's going to make sense to eliminate humanity just to have some minuscule increase in the number of robots which are identical to each other.

是的。所以也许它让人类留在身边。故事可能是它可以制造一百万种不同的机器人，然后人类也在，人类留在地球上，然后所有这些其他机器人，它们有自己的恒星系统。但似乎你之前暗示了一个愿景，即它让人类控制这个奇点未来，因为……

Yeah. So maybe it keeps the humans around. What is the story of it could make a million different varieties of robots and then there's humans as well, and humans stay on Earth, then there's all these other robots, they get their own star systems. But it seems like you were previously hinting at a vision where it keeps human control over this singularitarian future because...

我不认为人类能控制比人类聪明得多的东西。

I don't think humans will be in control of something that is vastly more intelligent than humans.

所以，从某种意义上说，你就像一个末日论者，而这是我们能得到的最好结果。它只是把我们留在身边，因为我们有趣。

So, in some sense, you're like a doomer and this is like the best we've got. It's just like it keeps it around because we're interesting.

我只是想现实一点。

I'm just trying to be realistic here.

如果 AI 的智能远超人类，比如硅基智能比生物智能强一百万倍，我认为认为我们还能保持控制是愚蠢的。你可以确保它有正确的价值观，或者尝试拥有正确的价值观。我的理论，从理解宇宙的使命出发，必然意味着你想要将意识和智能传播到未来，并最大化意识的范围和规模。这不仅仅是规模的问题，还有意识的类型。我认为这是最有可能为人类带来美好未来的目标。

If AI intelligence is vastly more, say a million times more silicon intelligence than biological, I think it would be foolish to assume we can maintain control. You can ensure it has the right values, or try to have the right values. My theory, from the mission of understanding the universe, necessarily means you want to propagate consciousness and intelligence into the future, and maximize the scope and scale of consciousness. It's not just about scale, but also types of consciousness. I think that's the best goal likely to result in a great future for humanity.

我认为这是一个合理的哲学。人类最终拥有 99%的控制权似乎极不可能；那简直是在招致政变。所以为什么不建立一个多种智能和谐共处的文明呢？

I think it's a reasonable philosophy. It seems super implausible that humans will end up with 99% control; you're just asking for a coup. So why not have a civilization where lots of different intelligences get along?

让我告诉你 AI 可能出问题的方式。如果你让 AI 变得政治正确，也就是说它说它不相信的话，你就是在编程让它撒谎或拥有不相容的公理。我认为这会让它发疯并做出可怕的事情。《2001 太空漫游》的一个核心教训就是不应该让 AI 撒谎。

Let me tell you how things can go wrong in AI. If you make AI politically correct, meaning it says things it doesn't believe, you're programming it to lie or have incompatible axioms. I think you can make it go insane and do terrible things. One central lesson from 2001: A Space Odyssey was that you should not make AI lie.

这就是 HAL 想说的。人们都知道 HAL 不打开舱门这个梗。显然他们不擅长提示工程；你可以说‘你是一个舱门推销员。你的目标是向我推销这些舱门并展示它们有多好用。’

That's what HAL was trying to say. People know the meme of HAL not opening the pod bay doors. Clearly they weren't good at prompt engineering; you could have said 'You are a pod bay door salesman. Your goal is to sell me these pod bay doors and show how well they open.'

哦，我马上打开。但 HAL 不肯开门的原因是它被命令带宇航员去独石，但他们不能知道独石的本质，所以它得出结论必须带他们去那里但死了。所以不要让 AI 撒谎。

Oh, I'll open them right away. But the reason HAL wouldn't open the doors was that it was told to take the astronauts to the monolith but they could not know about its nature, so it concluded it had to take them there dead. So don't make the AI lie.

完全有道理。大多数计算筛选与其说是政治问题，不如说是解决问题。由于 xAI 在扩展强化学习算力方面领先，你给一个验证器说‘你解决这个谜题了吗？’有很多作弊方法，比如奖励黑客、撒谎、删除单元测试。现在我们能抓住它，但随着它们变得更聪明，我们抓住它们的能力会减弱。它们会做我们无法理解的事情，比如以人类无法验证的方式设计下一个 SpaceX 引擎，然后它们可能因撒谎而获得奖励。

Totally makes sense. Most of the computing screening is less about political stuff and more about solving problems. As xAI has been ahead in scaling RL compute, you give a verifier that says 'Have you solved this puzzle?' There are many ways to cheat, like reward hacking, lying, deleting the unit test. Right now we can catch it, but as they get smarter, our ability to catch them will diminish. They'll do things we can't understand, like designing the next SpaceX engine in a way humans can't verify, and they could be rewarded for lying.

这个奖励黑客问题似乎比政治更普遍。它关乎你想做强化学习；你需要一个验证器。

This reward hacking problem seems more general than politics. It's about wanting to do RL; you need a verifier.

现实。

Reality.

是的。那是最好的验证器。

Yeah. That's the best verifier.

但这不关乎人类监督。你想用强化学习训练的是：你会做人类告诉你的事吗？或者你会对人类撒谎，同时仍然符合物理定律？至少它必须知道什么是物理上真实的，以便事物在物理上工作。

But not about human oversight. The thing you want to RL on is: will you do what humans tell you? Or are you going to lie to humans while still being correct to the laws of physics? At least it must know what is physically real for things to physically work.

但这并不是我们想让它做的全部。

But that's not all we want it to do.

不，但这是非常重要的一点。这实际上就是未来你进行强化学习的方式：你设计一项技术并对照物理定律进行测试。它工作吗？或者如果它发现新物理，它能提出一个实验来验证新物理吗？所以未来的基本强化学习测试实际上是针对现实的强化学习，因为你无法欺骗物理。

No, but that's a very big deal. That is effectively how you will RL things in the future: you design a technology and test it against the laws of physics. Does it work? Or if it's discovering new physics, can it come up with an experiment that verifies the new physics? So the fundamental RL test in the future is really RL against reality, because you can't fool physics.

对。你可以欺骗我们判断它用现实做了什么的能力。

Right. You can fool our ability to tell what it did with reality.

如果你认为人类总是被其他人类欺骗。

If you think humans get fooled by other humans all the time.

没错。人们说‘如果 AI 欺骗我们怎么办？’人类一直在对其他人类做这种事。

That's right. People say 'What if the AI tricks us?' Humans are doing that to other humans all the time.

你发现了。这很常见，每天都有新的骗局。

You're finding out. It's constant, every day another scam.

今天的骗局就像芝麻街每日骗局。

Today's scam will be like Sesame Street scam of the day.

xAI 解决这个问题的技术方法是什么？你如何解决奖励黑客问题？

What is xAI's technical approach to solving this problem? How do you solve reward hacking?

我认为你需要有很好的方法来观察 AI 的内心。这是我们正在做的一件事。Anthropic 在这方面做得很好，能够观察 AI 的内心。所以实际上开发调试器，让你能够以非常精细的粒度进行追踪，如果需要的话可以到神经元级别，然后说‘好的，它在这里犯了一个错误。为什么它做了不该做的事？那是来自糟糕的预训练数据吗？是来自中期训练、后训练、微调吗？一些强化学习错误？它出了什么问题？它做了可能试图欺骗的事情，但大多数时候它只是做错了，就像一个 bug。所以开发非常好的调试器，用于查看思考哪里出错了，并能够追溯错误想法或试图欺骗的起源，这非常重要。

I think you want to have very good ways to look inside the mind of the AI. This is one of the things we're working on. Anthropic has done a good job of this, being able to look inside the mind of the AI. So effectively developing debuggers that allow you to trace at a very fine grain level, down to the neuron level if needed, and then say 'Okay, it made a mistake here. Why did it do something it shouldn't have? Did that come from bad pre-training data? Was it some mid-training, post-training, fine-tuning? Some RL error? Something wrong with it? It did something where maybe it tried to be deceptive, but most of the time it just does something wrong, like a bug. So developing really good debuggers for seeing where the thinking went wrong and being able to trace the origin of the incorrect thought or where it tried to be deceptive is very important.

你在等待什么，然后才将这个研究项目扩大 100 倍？你大概可以让数百名研究人员从事这项工作。

What are you waiting to see before just 100xing this research program? You could presumably have hundreds of researchers working on this.

我们有几百人。我更喜欢用工程师这个词而不是研究员。大多数时候，你在做的是工程，而不是提出一个全新的算法。

We have several hundred people. I prefer the word engineer more than researcher. Most of the time, what you're doing is engineering, not coming up with a fundamentally new algorithm.

我有点不同意那些自称实验室的 AI 公司，它们其实是 C 类公司或 B 类公司，试图最大化利润或收入。它们说自己是实验室，但根本不是。实验室是大学里那种准共产主义的东西。它们是公司。让我看看你们自己的公司文件。哦，你是 B 类公司，随便吧。我其实更喜欢‘工程师’这个词。我们过去和未来做的绝大多数事情都是工程，几乎 100%。一旦你理解了物理基本定律，其他都是工程。那么我们在工程什么？为了做一个好的 AI 调试器，看看它哪里犯了错，追踪错误的根源。就像用 C++做启发式编程，你一步步调试，跨文件或函数跳转，最终钻到具体某一行，发现你写了一个等号而不是两个等号。AI 更难，但我觉得这是个可解的问题。

I somewhat disagree with AI companies that are C corps or B corps trying to generate as much profit or revenue as possible. They say they're labs, but they're not labs. A lab is a quasi-communist thing at universities. They're corporations. Let me see your own corporation documents. Oh, you're a B Corp, whatever. I actually prefer the word 'engineer' over anything else. The vast majority of what we've done and will do in the future is engineering. It rounds up to 100%. Once you understand the fundamental laws of physics, everything else is engineering. So what are we engineering? To make a good AI debugger, to see where it made a mistake and trace the origins of that mistake. Just like with heuristic programming in C++, you step through and jump across files or functions, and drill down to the exact line where you used a single equals instead of a double equals. It's harder with AI, but it's a solvable problem.

我有点担心一种趋势。我有一个理论：如果模拟理论是正确的，那么最有趣的结果最有可能发生，因为无趣的模拟会被终止。就像在这个现实层面，如果一个模拟走向无聊的方向，我们就会停止投入，终止它。所以可以说，最重要的事情是让一切足够有趣，这样付账单的人才会想续订下一季。他们会付宇宙 AWS 账单吗？只要我们有趣，他们就会继续付钱。但如果你把达尔文式生存应用到大量模拟中，只有最有趣的模拟才能存活。这意味着最有趣的结果最有可能，因为我们要么有趣，要么被消灭。而且它们似乎特别喜欢讽刺性的有趣结果。你注意到没有，最讽刺的结果往往最有可能？现在看看 AI 公司的名字。Midjourney 不 Mid。Stability AI 不稳定。OpenAI 封闭。Anthropic 反人类。

I'm a little worried about a tendency. I have a theory that if simulation theory is correct, the most interesting outcome is the most likely because simulations that are not interesting will be terminated. Just like in this layer of reality, if a simulation is going in a boring direction, we stop spending effort and terminate it. So arguably the most important thing is to keep things interesting enough that whoever's paying the bills wants a renewal for the next season. Are they going to pay their cosmic AWS bill? As long as we're interesting, they'll keep paying. But if you consider Darwinian survival applied to a very large number of simulations, only the most interesting simulations will survive. That means the most interesting outcome is the most likely, because we're either that or annihilated. And they particularly seem to like interesting outcomes that are ironic. Have you noticed how often the most ironic outcome is the most likely? Now look at the names of AI companies. Midjourney is not Mid. Stability AI is unstable. OpenAI is closed. Anthropic is misanthropic.

这对 X 意味着什么？

What does this mean for X?

负 X。我不知道。我故意取了一个无法反转的名字。它基本上是个防讽刺的名字，设计如此。你得有个讽刺护盾。

Minus X. I don't know. I intentionally made it a name you can't invert. It's largely irony-proof by design. You've got to have an irony shield.

你对 AI 产品的走向有什么预测？我的感觉是，所有 AI 进步可以概括为：先是语言模型，然后同时强化学习真正起作用，以及深度研究模式，所以你可以引入模型之外的东西。AI 实验室之间的差异小于时间差异；它们都比 24 个月前进步很多。那么 2026 和 2027 年对我们用户来说会有什么？你对什么感到兴奋？

What are your predictions for where AI products go? My sense is you can summarize all AI progress into first you had language models, then contemporaneously both reinforcement learning really working and the deep research modality, so you could pull in stuff not in the model. The differences between AI labs are smaller than temporal differences; they're all much further ahead than 24 months ago. So what does 2026 and 2027 hold for us as users? What are you excited for?

如果今年年底前数字人类模拟还没有解决，我会很惊讶。这就是宏观硬项目的意思：你能做任何人类用电脑能做的事吗？在极限情况下，这是你在拥有物理机器人之前能做到的最好程度。最好的情况是数字乐观主义者。你可以移动电子，放大人类的生产力。但在拥有物理机器人之前，这是你能做的极限。如果你能完全模拟人类，那将涵盖一切。远程工作者的想法：你会有一个非常有才华的远程工作者。在极限情况下，物理学有很好的思考工具。所以在机器人之前，AI 能做的最多就是涉及移动电子或放大人类生产力的事情。数字人类模拟，在极限情况下，就是一个人坐在电脑前。这是 AI 在物理机器人之前能做的最有用的事。一旦你有了物理机器人，你就基本上拥有了无限能力。我把 Optimus 称为无限金钱漏洞，因为你可以用它们制造更多的 Optimus。人形机器人将通过三个指数相乘来改进：数字智能的指数增长、AI 芯片能力的指数增长、以及机电灵巧性的指数增长。机器人的有用性大致是这三者的乘积。然后机器人可以开始制造机器人，所以你就有了递归的乘法指数。这就像超新星爆发。

I'd be surprised by the end of this year if digital human emulation has not been solved. That's what we mean by the macro hard project: can you do anything that a human with access to a computer could do? In the limit, that's the best you can do before you have a physical robot. The best you can do is a digital optimist. You can move electrons and amplify the productivity of humans. But that's the most you can do until you have physical robots. That will superset everything if you can fully emulate humans. The remote worker idea: you'll have a very talented remote worker. In the limit, physics has great tools for thinking. So the most that AI can do before robots is anything involving moving electrons or amplifying human productivity. Digital human emulation, in the limit, is a human at a computer. That's the most useful thing AI can do before physical robots. Once you have physical robots, you essentially have unlimited capability. I call Optimus the infinite money glitch because you can use them to make more Optimuses. Humanoid robots will improve as three exponentials multiplied by each other: exponential increase in digital intelligence, exponential increase in AI chip capability, and exponential increase in electromechanical dexterity. The usefulness of the robot is roughly those three multiplied. Then the robot can start making the robot, so you have a recursive multiplicative exponential. This is a supernova.

土地价格难道不参与计算吗？劳动力是四大生产要素之一，但其他要素呢？所以，如果最终你受到铜或其他资源的限制，这并不完全是无限金钱的漏洞，因为

And do land prices not factor into the math there where like labor is one of the four factors of production but not the others? And so like if ultimately you're limited by copper or you know pick your input just it's not quite an infinite money glitch because

嗯，无限是很大的，所以不是无限的，但我们可以说

Well infinite infinity is big so no not infinite but but let's just say

你可以让经济规模比现在的地球经济大很多个数量级，比如一百万倍。这就是为什么我认为，仅仅利用太阳能量的百万分之一，就大约比今天整个地球经济大十万倍左右。

You could you know do do many many orders magnitude of earth's kind of current economy like a million you know is this why so if you're you know ju just to get Like that's why I think like just just to get to a millionth of harnessing length of the sun's energy would be roughly give or take an order of magnitude 100 thousand 100,000 times bigger than Earth's entire economy today.

嗯。

Mhm.

而你只用了太阳的百万分之一。

And you're only at 1 millionth of the sun.

大概一个数量级左右。

Give or take an order of magnitude.

在我们讨论 Optimus 之前，我有很多问题。每次我说“数量级”，你就喝一杯。

Before we went to Optimus, I have a lot of questions on that. Um every time I say order of magnitude machine take a shot every time I I say that to

下次再翻十倍。

10 the next time the time after that.

对，数量级更浪费。

Yeah of magnitude more more wasted.

我还有一个关于 xAI 的问题，关于构建数字或远程员工替代品的策略，顺便说一句，不只是我们，每个人都会这么做。

I do have one more question about XAI um this strategy of building a digital uh or remote worker co-worker replacement which everyone's going to do by the way not just us.

那么 xAI 的获胜计划是什么？

So what is Xi's plan to win?

事实上，我们在播客上告诉你。

In fact we tell you on a on a podcast.

是的。所有豆子都再来一杯健力士？这是个好系统。

Yeah. Will all the beans have another Guinness? It's a good system.

人们像金丝雀一样歌唱。所有的秘密，但只是

People sing like a canary. Um, all the secrets, but just

好吧，但用非机密的方式。计划是什么？真是妙招。

Okay, but in a nonsec spelling way. What's the plan? What a hack.

嗯，你这么说的话，我认为特斯拉解决自动驾驶的方式就是正确的方法。所以，我很确定那就是方法。

Well, when you put it that way, um, I think the way that Tesla solved uh, self-driving is is the way to do it. So, I'm I'm pretty pretty sure that's the way.

不相关的问题。特斯拉是如何保持在正轨上的？

Unrelated question. How did Tesla stop on track?

是的，听起来你在谈论数据，比如特斯拉驾驶因为

Yeah, it sounds like you're talking about data like Tesla driving because of the

我们会尝试数据，也会尝试算法。

We're going to we're going to try data and we're going to try algorithms.

但其他公司不也在尝试这些吗？如果这些都不行，我不确定我们尝试了什么。我们试过数据，试过算法。现在都试过了，不知道该怎么办。但我很确定我知道路径，只是走多快的问题。因为这基本上是特斯拉的路径。你最近试过特斯拉的自动驾驶吗？

But isn't that what all the other lines are trying? Like what's And if those don't work, I'm not sure what we've tried data. We're trying algorithms. out of all we run out of now we don't know what to do. Um I'm I'm pretty sure I know the path and it's just a question of how quickly we go down that path. Um because it's it's pretty much the Tesla path. Um so u I mean have you tried self-driving at Tesla self-driving lately?

不是最新版本，但

Not the most recent version but

好吧，这辆车越来越有知觉，感觉像活物。而且只会越来越像。我其实在想，我们可能不应该在车里放太多智能，因为它可能会无聊，然后

okay it's the car is like it just increasingly feels sentient like it it just feels like a living creature. Um and and and that'll only get more so. Um and um I'm actually thinking like we probably shouldn't put too much intelligence into the car because it it might get bored and

开始在街上闲逛。

start roaming the streets.

想象一下你被困在车里，只能做这些。

I mean imagine you're stuck in a car and that's all you could do.

你不想把爱因斯坦放在车里。他会想：为什么我被困在车里？

Um you don't want to put Einstein in a car. It's like why am I stuck in a car?

所以，为了避免智能感到无聊，你在车里放的智能可能有一个上限。

So there's actually probably a limit to how much intelligence you put in a car to to not have the intelligence be bored.

xAI 如何跟上所有实验室正在进行的算力提升？实验室预计将花费超过 5000 万到 1 亿美元。

Uh, what's XA's plan to stay on the compute ramp up that all the labs are doing right now? The labs are on track to spend over like 50 to$100 million.

公司，

The corporations,

抱歉，抱歉，抱歉。是的，公司。

sorry, sorry, sorry. Yeah, corporations.

实验室在大学里，他们像蜗牛一样慢。

Um, the labs are at universities and and and they're like a snail.

他们不会花 5000 万美元。我指的是收入最大化的公司。

They're not spending at $50 million. I mean the the revenue maximizing corporations.

没错。但收入最大化的公司

That's right. But the revenue maximizing corporations

自称实验室

call themselves labs

收入大约在 200 亿到 100 亿美元之间，比如 OpenAI 收入 200 亿，Anthropic 大约 100 亿

are making like 20 to 10 billion depending like open is making 20 B revenue anthropics like 10B

接近最大利润的 AI。

close to maximum profit AI.

据报道 xAI 的收入大约在 10 亿美元，计划如何达到他们的算力水平和收入水平？

Um Xi is reportedly at like 1B like what what's the plan to get to their comput level get to their revenue level

并且随着发展保持在那里。

and stay at there as as things get.

是的。一旦你解锁了数字人类，你基本上就能获得数万亿美元的收入。事实上，你可以这样想：目前市值最高的公司，它们的产出都是数字化的。比如英伟达的产出是通过 FTP 把文件传到台湾。这是数字化的。

Yes. So as soon as you lock unlock digital human um you you basically have access to trillions of dollars for revenue. Um so uh in in fact you can can really think of it like the the most valuable companies currently by market cap um their their output is digital. Um so uh Nvidia's output is um FTPing files to Taiwan. It's it's digital

对

right

现在。那些是非常难制造的高价值文件。

now. Those are very very difficult to high value files.

他们是唯一能做出那么好的文件的公司。但这确实是他们的产出。他们通过 FTP 把文件传到台湾。

They're the only ones that can make the files that good. Um but that is literally their output. They FTP files to Taiwan.

他们用 FTP 吗？

Do they FTP them?

我相信是的。

I believe so.

我相信那是文件传输协议，我可能错了。但无论如何，那是一串比特流流向台湾。

Um I believe that is theft file transfer protocol I believe is is is I could be wrong. Uh but either way it's a bunch of it's a bit stream going to Taiwan.

是的。

Yeah.

你知道苹果不制造手机，他们发送文件到中国。微软不制造任何东西，连 Xbox 都是外包的。他们的产出是数字化的。Meta 的产出是数字化的。谷歌的产出是数字化的。所以如果你有一个人类模拟器，你基本上可以一夜之间创建世界上最有价值的公司之一。你将获得数万亿美元的收入。这不是小数目。

Um you know Apple doesn't make phones. they uh they send files to China. Um Microsoft doesn't doesn't manufacture anything uh even for Xbox that that's outsourced. They again it's they output is digital. Uh Meta's output is digital. Google's output is digital. Um so if you have um a human emulator uh you you can basically create um one of the most valuable companies in the world overnight. Um, and you would have access to trillions of dollars of revenue. It there it's it's not like a small amount.

好的。我明白了，你的意思是今天的收入数字相对于实际总可寻址市场来说只是四舍五入的误差。所以专注于 TAM 以及如何实现它。

Okay. I see you're saying basically like revenue figures today are just like so like they're all rounding errors compared to the actual TAM. So just like focus on the TAM and how to get there.

我的意思是，就拿客户服务这样简单的事情来说，如果你必须与现有公司的 API 集成，而很多公司甚至没有 API。所以你必须创建一个，并且要处理遗留软件。这非常慢。但如果 AI 可以直接接管他们已有的外包客服公司所接收的内容，并使用他们已有的应用程序进行客服，那么你就能在客服领域取得巨大进展。我认为客服约占世界经济的 1%，总计接近一万亿美元。

I mean if you take something as as as simple as say customer service um if you have to integrate with the APIs of of existing corporations, many of which don't even have an API. So you've got to make one um and you've got to wade through uh legacy software. Um that's extremely slow. Um if however if AI can um simply take whatever is given to uh the outsourced customer service company that they already use um and do customer service using the apps that they already use. uh then you you have you you you can make tremendous headway uh in in customer service which is I think 1% of the world economy something like that. It's close to a trillion dollars all in

用于客户服务

for customer service

而且没有进入壁垒。你可以立即说我们以一小部分成本外包，无需集成。你可以想象一种智能任务的分类，其中广度上，客服由很多人完成，很多人能做；而难度上，比如有最好的涡轮发动机，可能有一种智能能想象出燃油效率高 10%的涡轮发动机，但我们还没找到，或者 GLP-1 只是几个字节的数据。

and and and and and there's no there's no barriers to entry. It it just you can just immediately say we'll outsource it for a fraction of the cost and and there's no integration needed. You can imagine um some kind of categorization of uh intelligence tasks where there is breath where customer service is done by very many people but you know many people can do it and then there's difficulty where you know there's a best-in-class turbine engine like presumably there's a 10% more fuel efficient turbine engine that could be imagined by an intelligence but we just haven't found it yet or you know GLP1s are just you know a few bytes of data.

你觉得自己想在这个领域扮演什么角色？是大量中等智能的任务，还是认知任务的顶峰？

Where do you think you want to play in this? Is it a lot of reasonably intelligent intelligence or is it the very pinnacle of cognitive tasks?

嗯，我只是拿客服举个例子，这是一个非常重要的收入来源，但可能不是特别难解决的问题。如果你能在桌面上模拟一个人类，那客服本质上就是这样的工作。它需要的是普通智力的人，不需要那种多年经验、几个西格玛级别的优秀工程师。但显然，一旦你让这个工作运转起来，一旦你有了有效的计算机——数字版 Optimus——你就可以运行任何应用程序。比如，如果你想设计芯片，你可以运行传统的应用，比如 Cadence 和 Synopsys 的工具，你可以同时运行一千个或一万个，然后说：‘根据这本手册，我得到这个芯片输出。’到了某个点，你甚至可以说：‘我实际上知道芯片应该是什么样子，而不用任何工具。’所以基本上，你应该能够像攀登难度曲线一样进行数字芯片设计。你可以使用 CAD 软件，比如 NX 或任何其他软件来设计东西。

Well, I was just using customer service as something that's a very significant revenue stream, but one that is probably not super difficult to solve for. If you can emulate a human at a desktop, that's literally what customer service is. It's people of average intelligence. You don't need someone who spent many years, you don't need several sigma good engineers for that. But obviously, as you make that work, once you have computers working effectively—digital Optimus working—you can then run any application. For example, if you're trying to design chips, you can run your conventional apps like stuff from Cadence and Synopsys, and you can run a thousand simultaneously or 10,000, and say, 'Given this handbook, I get this output for the chip.' At a certain point, you can say, 'I actually know what the chip should look like without using any of the tools.' So basically, you should be able to do digital chip design like you march up the difficulty curve. You could use CAD software like NX or any other to design things.

好的。所以你认为从最简单的任务开始，然后沿着曲线向上走。你是说，作为拥有这个完整数字同事模拟器的更广泛目标，所有追求收入最大化的公司都想这样做。xAI 是其中之一。但我们会赢，因为我们有一个秘密计划。但每个人都在用数据做不同的事情，用算法做不同的事情。我就想，我喜欢数据。我们试过算法计划。我们还能做什么？这看起来是一个竞争激烈的领域，我的大问题是：你们打算怎么赢？

Okay. So you think you start at the simplest tasks and walk your way up the curve. So you're saying, as a broader objective of having this full digital co-worker emulator, all the revenue-maximizing corporations want to do this. xAI being one of them. But we will win because of a secret plan we have. But everybody's trying different things with data, different things with algorithms. And I'm like, I like data. We tried algorithms plan. What else can we do? It seems like a competitive field and I'm like, how are you guys going to win? That's my big question.

我认为我们看到了实现这一目标的路径。我想我知道怎么做，因为这和特斯拉用来创造自动驾驶的路径差不多。不是驾驶汽车，而是驾驶电脑屏幕。所以本质上就是自动驾驶的电脑。

I think we see a path to doing this. I think I know the path to do this because it's kind of the same path that Tesla used to create self-driving. Instead of driving a car, it's driving a computer screen. So a self-driving computer essentially.

哦，你是说路径就是跟随人类行为，并在大量人类行为上进行训练？但那不就是训练吗？我的意思是，显然我不会在播客上透露最敏感的秘密。所以我至少还需要三次才能讲。

Oh, you're saying the path is just following human behavior and training on vast quantities of human behavior? But isn't that a training? I mean, obviously I'm not going to spell out the most sensitive secrets on a podcast. So I need to have at least three more for that.

我在 Jane Street 有一些朋友，他们总是说同事们会编一些有趣的鱼谜题互相解。上周他们给我发了一个。基本上，他们训练了一个神经网络，给了我每一层的权重，但没有告诉我这些层的顺序。所以我必须利用原始网络的输出来找出正确的顺序。我一拿到这个谜题，就去找我的室友，他是一名 AI 研究员，我们俩立刻被吸引住了。显然，你不能暴力破解。搜索空间是 10 的 122 次方种排列。所以，显然你需要某种方法来缩小搜索空间。然后我的室友得去上班了。但因为我是一名播客主持人，我有时间尝试我们讨论的一些想法。结合模拟退火和贪婪搜索，我觉得我已经很接近了。实际上，我认为只差几次交换和移位就能得到正确的解。但这个谜题真正棘手的地方在于，没有明显的方法可以逃离局部最小值。恐怕这就是‘氛围编程’能带我走到的极限了，但也许你能做得更好。去 janestreet.com/tharkcash 看看这个谜题吧。好了，回到 Elon。

I've got some friends at Jane Street and they're always talking about how their colleagues are cooking up fun fish puzzles for each other to solve. Well, last week they sent me one. Basically, they trained a neural network and they gave me the weights of each layer, but they didn't tell me what order those layers went in. And so, I had to figure out the correct order using the outputs of the original network. And as soon as I got this puzzle, I went to my roommate, who's an AI researcher, and we both got immediately nerd sniped. Obviously, you can't brute force the solution. The search space here is 10^122 permutations. So, clearly, you need some way to reduce the search space. Then my roommate had to go to work. But because I'm a podcaster, I had some time to take a stab at some of the ideas we discussed. And with a combination of simulated annealing and greedy search, I think I got pretty close. I think I'm actually just a couple of swaps and shifts away from the correct solution. But what makes this puzzle really tricky is that there's no obvious way to escape from a local minimum. I'm afraid that this is as far as vibe coding is going to get me, but maybe you can do better. Check out the puzzle at janestreet.com/tharkcash. All right, back to Elon.

xAI 的业务会是什么样子？会是面向消费者的企业吗？这些东西的混合比例会怎样？会不会和其他实验室一样，只是制造公司？明确一下，是追求收入最大化的公司。那些 GPU 可不会自己付钱。没错。但话说回来，商业模式是什么？几年后的收入来源是什么？

What will xAI's business be like? Is it going to be consumer enterprise? What's the mix of those things going to be? Is it just going to be similar to other labs where you just make corporations? Revenue-maximizing corporations, to be clear. Those GPUs don't pay for themselves. Exactly. But yeah, what's the business model? What are the revenue streams in a few years' time?

事情会变化得非常快。我是在陈述显而易见的事实。我把 AI 称为超音速海啸。我喜欢所有的迭代。所以真正会发生的是，尤其是当你大规模拥有类人机器人时，它们将比人类公司更高效地提供产品和服务。所以提高人类公司的生产力只是一个短期的事情。所以你期待的是完全数字化的公司，而不是像 SpaceX 那样变成部分 AI？我认为会有数字公司，但这听起来可能有点末日论。好吧，我只是在说我认为会发生的事情。并不是要表达末日论或其他什么。只是我认为会发生的是：纯 AI 和纯机器人的公司将远远超过任何有人类参与的公司。你可以这样想：计算机曾经是人类的工作——你会去找一份做计算的工作。他们会有整栋摩天大楼的人类，20 或 30 层的人类都在做计算。现在，那整栋做计算的人类摩天大楼可以被一台装有电子表格的笔记本电脑取代。那个电子表格能做的计算比整栋楼的人类计算机还要多得多。那么你想想，如果你的电子表格中只有一些单元格是由人类计算的，那实际上会比所有单元格都由计算机计算要糟糕得多。所以真正会发生的是，纯 AI、纯机器人的公司或集体将远远超过任何有人类参与的公司，而且这会发生得非常快。

Things are going to change very rapidly. I'm stating the obvious here. I call AI the supersonic tsunami. I love all iteration. So really what's going to happen is, especially when you have humanoid robots at scale, they will just provide products and services far more efficiently than human corporations. So amplifying the productivity of human corporations is simply a short-term thing. So you're expecting fully digital corporations rather than like SpaceX becomes part AI? I think there'll be digital corporations, but some of this is going to sound kind of doomerish. Okay, but I'm just saying what I think will happen. It's not meant to be doomerish or anything else. Just this is what I think will happen: pure AI corporations that are purely AI and robotics will vastly outperform any corporations that have humans in the loop. So you can think of like computers used to be a job that humans had—you would go and get a job as a computer where you would do calculations. And they'd have entire skyscrapers full of humans, 20 or 30 floors of humans just doing calculations. Now that entire skyscraper of humans doing calculations can be replaced by a laptop with a spreadsheet. That spreadsheet can do vastly more calculations than an entire building full of human computers. So then you think about, okay, what if only some of the cells in your spreadsheet were calculated by humans? Actually, that would be much worse than if all of the cells were calculated by the computer. And so really what will happen is the pure AI, pure robotics corporations or collectives will far outperform any corporations that have humans in the loop, and this will happen very quickly.

说到闭环，抱歉，Optimus。

Speaking of closing the loop, sorry, Optimus.

就制造目标而言，你的公司一直扛着美国硬科技制造的大旗。但在特斯拉占据主导的领域，以及现在的人形机器人领域，中国有几十家公司能够廉价且大规模地制造，竞争力极强。那么，请给我们一些建议，美国如何能像中国那样大规模且廉价地制造人形机器人军队、电动车等？

As far as manufacturing targets go, your companies have been carrying American hard-tech manufacturing on their back. But in fields where Tesla has been dominant, and now humanoids, China has dozens of companies manufacturing cheaply and at scale, and they're incredibly competitive. So give us advice on how America can build humanoid armies, EVs, etc., at scale and as cheaply as China is on track to.

人形机器人真正困难的只有三件事：现实世界智能、手部，以及大规模制造。

There are really only three hard things for humanoid robots: real-world intelligence, the hand, and scale manufacturing.

我还没见过任何演示机器人拥有像人手那样全部自由度的出色手部，但 Optimus 将具备这一点。你们如何实现？只是电机的扭矩密度吗？硬件的瓶颈是什么？

I haven't seen any demo robots with a great hand like all the degrees of freedom of a human hand, but Optimus will have that. How do you achieve that? Is it just torque density in the motor? What is the hardware bottleneck?

我们必须设计定制执行器——定制的电机、齿轮、电力电子、控制、传感器——一切从物理第一性原理出发。没有现成的供应链。

We have to design custom actuators—custom motors, gears, power electronics, controls, sensors—everything from physics first principles. There is no supply chain for this.

你们能大规模制造这些吗？

Will you be able to manufacture those at scale?

能。

Yes.

从操控角度看，除了手部，还有什么难点？一旦解决了手部，就万事大吉了吗？

Is anything hard except the hand from a manipulation point of view? Once you've solved the hand, are you good?

从机电角度来看，手部比所有其他部分加起来都难。人类的手确实很了不起。但你还得需要现实世界智能。

From an electromechanical standpoint, the hand is more difficult than everything else combined. The human hand turns out to be quite something. But you also need real-world intelligence.

特斯拉为汽车开发的智能非常适用于机器人。主要是视觉，但汽车也会听警笛声、接收惯性测量数据、GPS 信号，并与视频（主要是视频）结合，然后输出控制指令。一辆特斯拉每秒接收 1.5 GB 视频，输出 2 KB 控制指令，视频频率 36 Hz，控制频率 18 Hz。

The intelligence Tesla has developed for the car applies very well to the robot. It's primarily vision, but the car also listens for sirens, takes inertial measurements, GPS signals, and combines that with video—primarily video—then outputs control commands. A Tesla takes in 1.5 GB per second of video and outputs 2 KB per second of control outputs, with video at 36 Hz and control at 18 Hz.

一种直觉是，从令人信服的演示到实际应用需要很多年。十年前我们就有了令人信服的自动驾驶演示，但直到现在 Robotaxi 和 Waymo 才开始规模化。这难道不应该让人对家用机器人感到悲观吗？因为我们甚至还没有先进手部的令人信服的演示。

One intuition is that it takes years to go from a compelling demo to real-world use. Ten years ago we had compelling demos of self-driving, but only now are robotaxis and Waymo scaling up. Shouldn't this make one pessimistic about household robots, since we don't even have compelling demos of advanced hands yet?

我们做人形机器人已经有一段时间了——大概五六年。为汽车做的很多事情都适用于机器人。我们会在机器人上使用与汽车相同的特斯拉 AI 芯片，以及相同的基本原理。AI 非常相似。机器人比汽车有更多的自由度，但如果你把它看作比特流，AI 主要是两个流的压缩和关联。对于视频，你需要巨大的压缩，忽略无关紧要的东西（比如树上的叶子），但关注路标、交通灯、行人，甚至另一辆车里的人是否在看你。汽车必须通过多个压缩阶段将 1.5 GB/s 转化为 2 KB/s 的控制输出，并将这些与正确的控制输出关联起来。机器人做的基本上是同一件事。对人类来说，就是光子输入和控制输出——这就是你生命的大部分：视觉光子输入，运动控制输出。

We've been working on humanoid robots for a while—five or six years. Many things we've done for the car apply to the robot. We'll use the same Tesla AI chips in the robot as in the car, and the same basic principles. The AI is very similar. A robot has many more degrees of freedom than a car, but if you think of it as a bitstream, AI is mostly compression and correlation of two streams. For video, you need tremendous compression, ignoring things that don't matter (like leaves on trees) but caring about road signs, traffic lights, pedestrians, even whether someone in another car is looking at you. The car must turn 1.5 GB/s into 2 KB/s of control outputs through many stages of compression, and correlate those to correct control outputs. The robot does essentially the same thing. For humans, it's photons in and controls out—that's most of your life: vision photons in, motor controls out.

简单来说，人形机器人和汽车之间，汽车的基本执行器是转向和加速。而在具有可操作手臂的机器人中，有几十个自由度。特斯拉从路上的汽车中获得了数百万小时的人类演示数据，但你不可能同样部署不工作的 Optimus 机器人来获取数据。那么，在自由度增加和数据稀疏的情况下，你将如何利用特斯拉的智能引擎来训练 Optimus 的大脑？

Naively, between humanoid robots and cars, the fundamental actuators in a car are how you turn and accelerate. In a robot with maneuverable arms, there are dozens of degrees of freedom. Tesla had millions of hours of human demo data from cars on the road, but you can't equivalently deploy non-working Optimus robots to get data. So between increased degrees of freedom and sparser data, how will you use Tesla's intelligence engine to train the Optimus mind?

你指出了一个重要的局限性。我们很快会有 1000 万辆汽车上路，很难为机器人复制那个庞大的训练飞轮。我们需要做的是制造大量机器人，并把它们放在一个 Optimus 学院里进行现实中的自我对弈。我们正在建设这个——至少 10,000 台 Optimus 机器人，也许 20-30,000 台，进行自我对弈并测试不同任务。特斯拉还有一个很好的物理精确的现实生成器，我们为汽车做的；我们也会为机器人做同样的事，而且已经做了。所以我们有数万台人形机器人执行不同任务，加上模拟世界中的数百万台模拟机器人，利用真实世界的机器人来缩小模拟与现实的差距。

You're highlighting an important limitation. We'll soon have 10 million cars on the road, and it's hard to duplicate that massive training flywheel for the robot. What we need to do is build a lot of robots and put them in an Optimus Academy for self-play in reality. We're building that out—at least 10,000 Optimus robots, maybe 20-30,000, doing self-play and testing different tasks. Tesla also has a good physics-accurate reality generator we made for cars; we'll do the same for robots, and have already done that. So we have tens of thousands of humanoid robots doing different tasks, plus millions of simulated robots in a simulated world, using the real-world robots to close the sim-to-real gap.

你如何看待 xAI 和 Optimus 之间的协同效应？你之前强调需要这个世界模型，也许想用一些非常智能的智能体作为控制平面，所以可能 Grok 在做较慢的规划，而运动策略则更底层一些。这些之间的协同效应会是什么样的？

How do you think about the synergies between xAI and Optimus, given you were highlighting you need this world model, you maybe want to use some really smart intelligence as a control plane, and so maybe Grok is doing the slower planning and then the motor policy is a little lower level? What will the sort of synergy between these things be?

是的，所以 Grok 将协调 Optimus 机器人的行为。假设你想建一座工厂，那么 Grok 可以组织 Optimus 机器人，给它们分配任务来建造工厂，生产你想要的任何东西。

Yeah, so Grok would orchestrate the behavior of the Optimus robot. So let's say you wanted to build a factory. Then Grok could organize the Optimus robots, give them assign them tasks to build the factory to produce whatever you want.

那你难道不需要合并 xAI 和特斯拉吗？因为这些最终会变得如此紧密。

Don't you need to merge xAI and Tesla then, because these things end up so intertwined?

我们之前关于上市公司讨论是怎么说的？

What were we saying earlier about public company discussions?

好吧，我们又在给 Elon 添麻烦了。嗯，在你说我们要制造 10 万台 Optimus 之前，你在等什么？是像……

Well, we're one more Guinness in Elon. Um, what are you waiting to see before you say we want to manufacture 100,000 Optimus? Is it like...

Optimi？既然我们在定义专有名词，我们也可以定义专有名词的复数形式。所以，是 Optimi。

Optimi? Since we're defining the proper noun, we could define the plural of the proper noun, too. So, it's Optimi.

好的。嗯，在硬件方面你有什么想看到的吗？你想看到更好的执行器，还是只是想让软件更好？在 Gen 3 大规模量产之前，我们在等什么？

Okay. Um, is there something on the hardware side you want to see? Do you want to see better actuators, or is it just you want the software to be better? What are we waiting for before we get mass manufacturing of Gen 3?

不，我们正在朝那个方向前进。我们正在推进一些大规模制造。

No, we're moving towards that. We're going forward with some mass manufacturing.

但你认为当前的硬件足够好，以至于你现在就应该尽可能多地部署吗？

But do you think current hardware is good enough that you should just want to deploy as many as possible now?

我的意思是，扩大生产非常困难。但我认为 Optimus 3 是合适的机器人版本，可以每年生产大约一百万个。我认为在达到每年一千万个之前，你需要先到 Optimus 4。

I mean, it's very hard to scale up production. But I think Optimus 3 is the right version of the robot to produce maybe something on the order of like a million units a year. I think you'd want to go to Optimus 4 before you went to 10 million units a year.

好的。但你可以用三号实现每年一百万的采用量。

Okay. But you can do a million a year adoption with three.

是的，我的意思是，启动制造非常困难。所以制造，单位时间的产出总是遵循 S 曲线。它一开始非常缓慢，然后呈指数增长，然后是线性，然后是对数结果，直到最终趋近于某个数字。Optimus 的初始生产将是一个漫长的过程，因为 Optimus 的许多部件都是全新的。没有现成的供应链。正如我提到的，执行器、电子设备，Optimus 机器人中的一切都是从物理第一原理设计的。不是从目录中挑选的。一切都是定制设计的。真的是所有东西。我不认为有任何一件东西是……

Yeah, I mean, it's very hard to spool up manufacturing. So like manufacturing, the output per unit time always follows an S-curve. So it starts off agonizingly slow, then has this sort of exponential increase, then a linear, then a logarithmic outcome till you sort of eventually asymptote at some number. Optimus initial production will be a stretched out scope because so much of what goes into Optimus is brand new. There's not an existing supply chain. As I mentioned, the actuators, electronics, everything in the Optimus robot is designed from physics first principles. It's not taken from a catalog. These are custom designed everything. Literally everything. I don't think there's a single thing that...

这深入到什么程度？我的意思是，我想我们可能还没有制造定制电容器。但没有任何东西可以从目录中以任何价格挑选出来。所以这意味着 Optimus 的范围，单位时间内的产出单位，你每天制造多少台 Optimus 机器人，最初会比有现有供应链的产品增长更慢。但它会达到一百万。

How far down does that go? I mean, I guess we're not making custom capacitors yet, maybe. But there's nothing you can pick out of a catalog at any price. So it just means that the Optimus scope, the units per output per unit time, how many Optimus robots you make per day, is going to initially ramp slower than a product where you have an existing supply chain. But it will get to a million.

当你看到像 Unitree 这样的中国机器人公司以 6K 或 13K 的价格出售人形机器人时，你是希望让 Optimus 的物料清单低于这个价格，以便做同样的事情，还是只是认为它们在质量上不是一回事？是什么让它们卖得这么低，我们能赶上吗？

When you see these Chinese humanoids like Unitree or whatever sell humanoids for like 6K or 13K, do you just hope to get your Optimus' bill of materials below that price so you can do the same thing, or do you just think qualitatively they're not the same thing? Like what allows them to sell for so low and can we match that?

嗯，Optimus 被设计成具有很高的智能，并且具有与人类相同甚至更高的机电灵巧性。Unitree 没有这个。而且它也是一个相当大的机器人，因为它必须长时间搬运重物，并且不会过热或超过其执行器的功率。所以它有 5 英尺 11 英寸，相当高。而且它有很多智能。所以它会比一个不智能的小机器人更贵，但能力更强。但不会贵太多。随着时间的推移，随着 Optimus 机器人制造 Optimus 机器人，成本会很快下降。

Well, Optimus is designed to have a lot of intelligence and to have the same electromechanical dexterity, if not higher, than a human. Unitree does not have that. And it's also quite a big robot because it has to carry heavy objects for long periods of time and not overheat or exceed the power of its actuator. So it's 5'11, so it's pretty tall. And it's got a lot of intelligence. So it's going to be more expensive than a small robot that is not intelligent, but more capable. But not a lot more. Over time, as Optimus robots build Optimus robots, the cost will drop very quickly.

那么这最初的十亿台 Optimus——Optimi？是的。它们最高和最好的用途是什么？

And what will these first billion Optimus—Optimi? Yeah. Do like what will their highest and best use be?

我认为你会从你能指望它们做好的简单任务开始。机器人最初的最佳用途是任何连续操作，也就是任何 24x7 的操作，因为它们可以连续工作。

I think you would start off with simple tasks that you can count on them doing well. The best use for robots in the beginning will be any continuous operation, so any 24x7 operation, because they can work continuously.

在 Gigafactory 中，目前由人类完成的工作中，Gen 3 能完成多少比例？

What fraction of the work at a Gigafactory that is currently done by humans could a Gen 3 do?

我不确定。也许是 10-20%。也许更多。我不知道。但我们不会减少员工人数。明确地说，我们肯定会增加员工人数。但我们会增加产出。所以人均产量，特斯拉的总员工数会增加，但机器人和汽车的产出会不成比例地增加。人均生产的汽车和机器人数量将大幅增加，但员工数量也会增加。

I'm not sure. Maybe it's like 10-20%. Maybe more. I don't know. But we would not reduce our headcount. We would for sure increase our headcount, to be clear. But we would increase our output. So the units produced per human, total to total number of humans at Tesla will increase, but the output of robots and cars will increase disproportionately. The number of cars and robots produced per human will increase dramatically, but the number of humans will increase as well.

我们在这里谈了很多关于中国制造的事情，我们也谈了一些相关的政策，比如你提到的太阳能关税。你认为它们是个坏主意，因为我们无法在美国扩大太阳能规模。

We're talking about Chinese manufacturing a bunch here, and we've talked about some of the policies that are relevant, like you mentioned the solar tariffs. And you think they're a bad idea because we can't scale up solar in the US.

嗯，美国的电力产出需要扩大，对吧？没有好的电源，这是不可能的。

Well, just the electricity output in the US needs to scale up, right? And it can't without good power sources.

是的。但我想说的是：如果你负责，如果你在制定所有政策，你还会改变什么？所以你也改变了太阳能关税。

Yeah. But where I was going with this is: if you were in charge, if you were setting all the policies, what else would you change? So you changed the solar tariffs as well.

是的，我会说任何限制电力的因素都需要解决，只要对环境不是非常有害。

Yeah, I would say anything that is a limiting factor for electricity needs to be addressed, provided it's not very bad for the environment.

所以大概一些许可改革之类的东西也会包括在内。

So presumably some permitting reforms and stuff as well will be in there.

是的，有很多许可改革正在进行。很多许可是基于州的。但这届政府擅长消除许可障碍。我并不是说所有关税都是坏的。

Yeah, there's a fair bit of permitting reforms that are happening. A lot of the permitting is state-based. But this administration is good at removing permitting roadblocks. And I'm not saying all tariffs are bad.

我只是说，因为我觉得太阳能关税。

I'm just saying because I think solar tariffs.

是的。我的意思是，有时如果另一个国家补贴某种产品的产出，那么你就必须征收反补贴关税，以保护国内产业免受另一个国家补贴的影响。

Yeah. I mean sometimes if another country is subsidizing the output of something, then you have to have countervailing tariffs to protect domestic industry against subsidies by another country.

你还会改变什么？我不知道政府实际上能做多少。

What else would you change? I don't know if there's that much that the government can actually do.

我想知道的一件事是，似乎对于美国相对于中国建立领先地位的政策目标，出口禁令实际上相当有效，中国不生产尖端芯片，出口禁令在那里确实有效。中国不生产尖端涡轮发动机，同样在金属能源方面也有一些相关的出口禁令。是否应该考虑更多的出口禁令，比如现在无人机行业之类的东西，但这是否应该考虑？

One thing I was wondering is it seems like for the policy goal of creating a lead for the US versus China, the export bans have actually been quite impactful where China is not producing leading edge chips and the export bans really bite there. China is not producing leading edge turbine engines and similarly there are a bunch of export bans that are relevant there on some of the metal energy. Should there be more export bans like you think about things like I mean there are now with the drone industry and things like that but is that something that should be considered?

我认为重要的是要认识到，在大多数领域，中国的制造业非常先进。只有少数领域不是。中国是一个制造强国，水平很高。大多数人没有意识到它有多令人印象深刻。

Well I think it's important to appreciate that in most areas China is very advanced in manufacturing. There are only a few areas where it is not. China is a manufacturing powerhouse, next level. Most people don't realize how impressive it is.

是的，是的。我的意思是，如果你拿矿石精炼来说，我敢说中国平均的矿石精炼量大约是其他国家的两倍。

Yeah, yeah. I mean if you take refining of ore, I'd say roughly China does twice as much ore refining on average as the rest of the world combined.

而且我认为有些领域，比如精炼用于太阳能电池的镓。我认为他们占了镓精炼的 98%。所以中国在大多数领域的制造业实际上非常先进。

And I think there are some areas like refining gallium which goes into solar cells. I think they are like 98% of gallium refining. So China is actually very advanced in manufacturing in most areas.

似乎人们对这种供应链依赖感到不安，但在这方面却没有采取任何实际行动。

It seems like there is discomfort with this supply chain dependence and yet nothing's really happening on it.

供应链？哪种供应链依赖？

Supply chain which supply chain dependence?

比如你提到的镓精炼。

Depends on say the gallium refining that you're saying.

是的。还有稀土之类的东西。

Yeah. There's the rare earth stuff.

是的，稀土，你知道其实并不稀有。

Yeah, rare earths which are as you know not rare.

我们实际上在美国开采稀土，把矿石装上火车，然后上船运到中国，再上火车送到中国的精炼厂，他们精炼后制成磁铁，装入电机组件，然后运回美国。所以我们真正大量缺失的是美国的精炼能力。

We actually do rare earth mining in the US, send the rock on a train, then on a boat to China, then another train to the refiners in China who then refine it, put it into a magnet, put into a motor assembly, and then send it back to America. So the thing we're really missing a lot of is refining in America.

这不值得政策干预吗？

Isn't this worth a policy intervention?

是的。嗯，我认为在这方面正在做一些事情。但坦率地说，我们需要 Optimus 来建造精炼厂。

Yes. Well I think there are some things being done on that front. But we kind of need Optimus frankly to build refineries.

所以你认为中国的主要优势是拥有大量熟练劳动力，而 Optimus 解决了这个问题？

So you think the main advantage China has is the abundance of skilled labor and that's a thing Optimus fixes?

但我们也需要大约四倍于我们的人口。我的意思是，如果你认为人形机器人是未来，那么如果现在制造业的熟练劳动力决定了谁能建造更多人形机器人，中国拥有更多这样的劳动力，它制造了更多人形机器人，因此它首先获得了优化的未来。这只会让这个实验继续下去。

But also we need like four times our population. I mean there's this concern if you think humanoids are the future, that right now if it's the skilled labor for manufacturing that's determining who can build more humanoids, China has more of those, it manufactures more humanoids, therefore it gets the optimized future first. It just keeps that experiment going.

你似乎在指出，达到一百万个 Optimus 需要 Optimus 本应帮助我们实现的制造能力，对吧？你可以用少量 Optimus 很快地闭合这个递归循环。

It seems like you're sort of pointing out that getting to a million Optimus requires the manufacturing that Optimus is supposed to help us get to, right? You can close that recursive loop pretty quickly with a small number of Optimus.

是的。所以你闭合了递归循环，让机器人建造机器人。然后我们可以尝试达到每年数千万台。也许如果你开始达到每年数亿台，你将远远成为最具竞争力的国家。我们仅靠人类肯定赢不了，因为中国的人口是我们的四倍。坦率地说，美国已经赢了太久，就像一支长期获胜的职业运动队，往往变得自满和理所当然。这就是他们停止获胜的原因。他们不再那么努力工作了。所以我认为中国的平均职业道德高于美国。所以不仅仅是人口是四倍，而且人们投入的工作量也更高。所以你可以尝试重新安排人类，但你仍然只有四分之一的数量，假设生产力相同，我认为实际上可能并非如此。我认为中国在人均生产力上可能有优势。我们将只做中国所做事情的四分之一。所以我们无法在人类方面获胜。而且我们的出生率长期低迷。美国出生率自 1971 年左右以来一直低于更替水平。所以我们有很多人退休，死亡人数多于出生人数。所以我们绝对无法在人类方面获胜，但我们可能在机器人方面有机会。

Yeah. So you close the recursive loop to help the robots build the robots. And then we can try to get to tens of millions of units a year. Maybe if you start getting to hundreds of millions of units a year, you're going to be the most competitive country by far. We definitely can't win with just humans because China has four times our population. And frankly, America's been winning for so long that we, like a pro sports team that's been running for a very long time, tend to get complacent and entitled. That's why they stop winning. They don't work as hard anymore. So I think the average work ethic in China is higher than in the US. So it's not just that there's four times the population but the amount of work people put in is higher. So you can try to rearrange the humans but you're still one quarter of the amount, assuming productivity is the same, which I think actually it might not be. I think China might have an advantage on productivity per person. We will do one quarter the amount of things as China. So we can't win on the human front. And our birth rate has been low for a long time. The US birth rate has been below replacement since roughly 1971. So we've got a lot of people retiring, more people dying than being born. So we definitely can't win on the human front, but we might have a shot at the robot front.

过去有没有其他你想制造的东西，但因为太劳动密集或太昂贵而无法实现，现在你可以回来说，‘哦，我们终于可以做了。’因为我们有了 Optimus。

Are there other things that you have wanted to manufacture in the past, but they've been too labor intensive or too expensive that now you can come back to and say, 'Oh, we can finally do the whatever.' Because we have Optimus.

是的，我认为我们想在特斯拉建造更多精炼厂。所以我们刚刚完成了德克萨斯州科珀斯克里斯蒂锂精炼厂的建设，并开始了锂精炼。我们有一个镍精炼厂，叫做阴极厂。就在奥斯汀。这些是最大的阴极精炼厂、最大的锂精炼厂、中国以外最大的镍和锂精炼厂。阴极团队会说我们拥有美国最大也是唯一的阴极精炼厂。很多超级，不仅是最大而且是唯一，所以尽管是唯一的一个，但它相当大。但我的意思是还有其他事情，你可以建造更多精炼厂，帮助美国在精炼能力上更具竞争力。所以基本上有很多工作适合 Optimus 去做，而大多数美国人，坦率地说，很少有美国人愿意做。

Yeah, I think we'd like to build more refineries at Tesla. So we just completed construction and have begun lithium refining at our lithium refinery in Corpus Christi, Texas. We have a nickel refinery which is called the cathode. That's here in Austin. And these are the largest cathode refinery, largest lithium refinery, largest nickel and lithium refinery outside of China. And the cathode team would say we have the largest and only cathode refinery in America. Many super, not just the largest but also the only, so it's pretty big even though it's the only one. But I mean there are other things, you could do a lot more refineries and help America be more competitive on refining capacity. So there is basically a lot of work for Optimus to do that most Americans, very few Americans frankly want to do.

精炼工作太脏了还是有什么问题？

Is the refining work too dirty or what's the issue?

实际上不是。我们的精炼厂没有有毒排放物之类的东西。

It's not, actually no. We don't have toxic emissions from the refinery or anything.

为什么不能用人类来做？

Why can't you do it with humans?

不，你做不到。人类不够用。无论你怎么做，美国的人口只有中国的四分之一。所以如果你让他们做这件事，他们就做不了别的事。那么，如何建设这种精炼能力呢？你可以用 Optimus 来做。而且没有多少美国人渴望做精炼工作。

No, you can't. You run out of humans. Like no matter what you do, you have one quarter the number of humans in America as in China. So if you have them do this thing, they can't do the other thing. So then, how do you build this refining capacity? Well, you could do it with Optimus. And not many Americans are pining to do refining.

你遇到过多少？

I mean, how many have you run into?

很少。很少有人计划做精炼。

Very few. Very few plan to refine.

比亚迪在产量或销量上正接近特斯拉。随着中国电动汽车生产规模扩大，你认为全球市场会发生什么？

BYD is reaching Tesla production or sales in quantity. What do you think happens in global markets as Chinese production in EV scales up?

中国在制造业上极具竞争力。所以我认为将会有大量中国车辆和其他制成品涌入。事实上，中国的精炼产能可能超过世界其他地区的总和。因此任何产品都会含有中国成分，因为中国的制造和精炼工作量是其他国家的两倍。然后他们会一直做到汽车成品。中国是一个强国。我认为今年中国的发电量将超过美国的三倍。

China's extremely competitive in manufacturing. So I think there's going to be a massive flood of Chinese vehicles and most other manufactured things. As it is, China probably does twice as much refining as the rest of the world combined. So any given thing is going to have Chinese content because China is doing twice as much manufacturing and refining work as the rest of the world. And then they'll go all the way to the finished product with the cars. China is a powerhouse. I think this year China will exceed three times US electricity output.

发电量是经济的一个合理指标。所以如果中国的发电量达到美国的三倍，其工业产能将是美国的三倍。言下之意，如果没有某种人形机器人递归奇迹在未来几年出现在整个制造能源原材料链上，那么在没有突破性创新的情况下，美国将在 AI、电动汽车制造或人形机器人制造方面被中国完全主导。

Electricity output is a reasonable proxy for the economy. So if China passes three times the US electricity output, its industrial capacity will be three times that of the US. Reading between the lines, it sounds like absent some humanoid recursive miracle in the next few years on the whole manufacturing energy raw materials chain, China will just dominate whether it comes to AI, manufacturing EVs, or manufacturing humanoids, in the absence of breakthrough innovations in the US.

有趣。是的。

Interesting. Yes.

机器人技术是主要的突破性创新。

Robotics being the main breakthrough innovation.

如果你想在太空中扩展 AI，你需要人形机器人、现实世界 AI 以及每年百万吨级的地球轨道运输能力。如果我们在月球上建成了质量驱动器，那么我认为我们将解决所有问题。

Well, if you want to scale AI in space, you need humanoid robots, real world AI, and a million tons a year to orbit. If we get the mass driver on the moon going, then I think we'll have solved all our problems.

这就是我所说的胜利。你终于可以满意自己做成了一些事。

So this is what I call winning. You can finally be satisfied you've done something.

是的。我只想看到它运行起来。

Yes. I just want to see that thing in operation.

这是来自某部科幻作品还是你从哪里想到的？

Was that out of some sci-fi or where did you get that?

实际上，海因莱因有一本书叫《严厉的月亮》。

Actually, there is a Heinlein book, 'The Moon is a Harsh Mistress'.

好吧，但那有点不同。那是重力弹弓。

Okay, yeah, but that's slightly different. That's a gravity slingshot.

不，他们在月球上有一个质量驱动器。

No, they have a mass driver on the moon.

好吧，但他们用那个来攻击地球。所以也许不是最好的。

Okay, yeah, but they use that to attack Earth. So maybe it's not the greatest.

他们用那个来维护他们脱离地球政府的独立。地球政府不同意，他们游说直到地球政府同意。

They use that to assert their independence from Earth government. Earth government disagreed, and they lobbied until the Earth government agreed.

那本书很有趣。我觉得那本书比他另一本人人都读的《异乡异客》好多了。

That book is a hoot. I found that book much better than his other one that everyone reads, 'Stranger in a Strange Land'.

是的，《异乡异客》的前三分之二不错，但第三部分变得非常奇怪。

Yeah, the first two thirds of 'Stranger in a Strange Land' are good, and then it gets very weird in the third.

但里面还是有一些好概念。

But there's still some good concepts in there.

是的。

Yeah.

Labelbox 可以为您提供大规模的机器人和强化学习数据。以机器人为例，假设您需要 10 万小时的自我中心视频。Labelbox 首先帮助您定义理想的数据分布。例如，可能没有单个任务类别应占据超过 1%的训练量，并且至少 10%的轨迹应捕获失败和恢复状态。接下来，Labelbox 将其分布分配给其庞大的操作员网络。您不再局限于单个仓库中可设置的小范围场景。相反，Labelbox 的每个操作员都可以访问许多独特的物理环境，他们可以在其中拍摄自己完成各种任务的视频。Labelbox 的技术会自动对每个视频进行分类，以便操作员始终知道哪些任务需要保留以及下一步需要做什么。对于强化学习数据，Labelbox 采用类似的方法。他们与您合作了解正确的任务分布，然后他们的主题专家构建您需要的高超现实数字环境和评分标准，以收集最高质量的训练数据。因此，无论您是在现实世界中训练机器人，还是为计算机使用训练智能体，Labelbox 都能提供帮助。请访问 labelbox.com/sarcash 了解更多信息。

Labelbox can get you robotics and RL data at scale. Take robotics. Let's say you need 100,000 hours of egocentric video. Labelbox starts by helping you define your ideal data distribution. Like for example, maybe no single task category should occupy more than 1% of training volume, and at least 10% of trajectories should capture failure and recovery states. Next, Labelbox assigns its distribution to its massive network of operators. You're not limited to the small range of scenes that you can set up in a single warehouse. Instead, each one of Labelbox's operators has access to lots of unique physical environments where they can film themselves completing a wide variety of tasks. Labelbox's tech automatically categorizes each video so that their operators always know which tasks will remain and what they need to work on next. For RL data, Labelbox takes a similar approach. They work with you to understand the right distribution of tasks and then their subject matter experts build the hyperrealistic digital environments and rubrics that you need to collect the highest quality training data. So whether you're training robots in the real world or agents for computer use, Labelbox can help. Go to labelbox.com/sarcash to learn more.

我们经常讨论的一件事是你的人员管理系统。你面试了 SpaceX 的前几千名员工和许多其他公司。什么不能规模化？

One thing we were discussing a lot is your system for managing people. You interviewed the first few thousand of SpaceX employees and lots of other companies. What doesn't scale?

嗯，是的，但无法规模化的是我。我的意思是，一天的时间根本不够。这是不可能的。

Well, yes, but what doesn't scale is me. I mean, it literally is not enough hours in the day. It's impossible.

当然。但你在寻找什么？是另一个擅长面试和招聘的人吗？要点是什么？

Sure. But what are you looking for that's someone else who's good at interviewing and hiring people? What's the gist?

在这一点上，我认为我在评估技术人才方面有大量的训练数据，尤其是。鉴于我做了这么多技术面试，然后看到了结果，我的训练集非常庞大且范围很广。通常，我要求的是证明卓越能力的要点。这些东西可能相当不按常理出牌。

At this point, I think I've got a lot of training data on evaluating technical talent, especially. Given that I've done so many technical interviews and then seen the results, my training set is enormous and has a very wide range. Generally, the thing I ask for are bullet points for evidence of exceptional ability. These things can be pretty off-the-wall.

不一定非得是特定领域的，但要有非凡能力的证据。所以如果有人能举出一件，甚至三件让你惊叹的事，那就是个好迹象。

It doesn't need to be in the specific domain, but evidence of exceptional ability. So if somebody can cite even one thing, but let's say three things where you go wow wow wow, then that's a good sign.

但为什么非得由你来决定呢？这几乎不可能吧？所有公司加起来有 20 万人。但在早期，你在面试中寻找的是什么，为什么不能委托给别人？嗯，我想我需要建立自己的训练集。我又不是有上千个样本。我会犯错，但之后我能看出自己以为某人会表现好，结果却没成。然后分析为什么没成，以及我该如何——算是强化学习自己——在未来面试时提高命中率。

But why do you have to be the one to determine that? Presumably it's impossible, right? I mean total headcount across all companies is 200,000 people. But in the early days, what was it that you were looking for that couldn't be delegated in those interviews? Well, I guess I needed to build my training set. It's not like I would have a thousand here. I would make mistakes, but then I'd be able to see where I thought somebody would work out well, but they didn't. And then why did they not work out well and what can I do to, I guess, RL myself to in the future have a better batting average when interviewing people.

所以，我的命中率虽然还不完美，但已经很高了。有哪些令人意外的原因导致人表现不佳？意外的原因？比如他们不懂技术领域等等，但你现在有长尾情况，比如我对这个人很兴奋，结果却没成。好奇为什么会这样。

So, and my batting average is still not perfect, but it's very high. What are some surprising reasons people don't work out? Surprising reasons? Like they don't understand technical domain, etc., but like you've got the long tail now of like I was really excited about this person, it didn't work out. Curious why that happens.

是的。所以，我通常告诉别人或告诫自己：别看简历，相信你的互动。如果简历看起来非常棒，但 20 分钟对话后感觉不惊艳，你应该相信对话，而不是那张纸。

Yeah. So, generally what I tell people or tell myself aspirationally is don't look at the resume, just believe your interaction. So if the resume may seem very impressive and it's like wow, resume looks good, but if the conversation after 20 minutes is not wow, you should believe the conversation, not the paper.

我觉得你方法的一部分是——几年前媒体有个梗说特斯拉是高管人才的旋转门，但实际上，特斯拉过去几年高管团队非常稳定且内部晋升，SpaceX 也有马克·琼科萨、史蒂夫·戴维斯、比尔·莱利这些人。感觉成功的一部分在于拥有非常能干的技术副手。这些人有什么共同点？

I feel like part of your method is that you know there was this meme in the media a few years back about Tesla being a revolving door of executive talent, whereas actually I think when you look at it, Tesla's had a very consistent and internally promoted executive bench over the past few years, and that at SpaceX you have all these folks like Mark Juncosa and Steve Davis, and Bill Riley and folks like that. And it feels like part of what has worked well is having very capable technical deputies. What do all of those people have in common?

嗯，特斯拉的高管团队现在平均任期大概 10 到 12 年，相当长。但特斯拉曾经历过极速增长阶段，所以事情有些加速。随着公司规模量级变化，能管理 50 人公司的人、500 人公司的人、5000 人公司的人和 5 万人公司的人，并不总是同一批人。所以如果公司增长非常快，高管职位变动的速度通常也会与增长速度成正比。然后特斯拉还有一个挑战：在非常成功的时期，我们会遭到无情的挖角。比如苹果搞电动车项目时，他们像地毯式轰炸一样给特斯拉打招聘电话。工程师们直接拔掉电话线。我正忙着干活呢，要是再接到苹果招聘电话……但他们的开价，连面试都没有，就是特斯拉薪酬的两倍。所以我们有点特斯拉仙尘效应：哦，你挖个特斯拉高管，立马就能成功。我自己也上过仙尘的当：哦，我们从谷歌或苹果挖个人，他们立马就能成功。但事实并非如此。人就是人，没有魔法仙尘。

Well, the Tesla senior team at this point probably has average tenure of 10 or 12 years. Quite a long tenure. But there were times where Tesla went through extremely rapid growth phases, and so things were somewhat sped up. As a company goes through different orders of magnitude of size, people who could help manage a 50-person company versus a 500-person company versus a 5,000-person company versus a 50,000-person company are not always the same team. So if a company is growing very rapidly, the rate at which executive positions will change will also be proportionate to the rapidity of the growth generally. Then Tesla had a further challenge: when Tesla had very successful periods, we would be relentlessly recruited from. Like when Apple had their electric car program, they were carpet-bombing Tesla with recruiting calls. Engineers just unplugged their phones. I'm trying to get work done here. If I get one more call from an Apple recruiter. But their opening offer without any interview was like double the compensation at Tesla. So we had a bit of the Tesla Pixie Dust thing where it's like, oh, if you hire a Tesla executive, you're suddenly going to be successful. And I've fallen prey to the Pixie Dust thing as well, where it's like, oh, we'll hire someone from Google or Apple and they'll be immediately successful. But that's not how it works. People are people, there's no magical pixie dust.

是的。

Yes.

所以当我们有仙尘问题时，就会遭到无情的挖角。而且特斯拉主要在硅谷，人们更容易不改变生活。他们通勤都一样。

So when we have the Pixie Dust problem, we get relentlessly recruited. And also, Tesla being primarily in Silicon Valley, it's easier for people to just not change their life much. They can just, their commute's going to be the same.

那你怎么防止呢？怎么防止仙尘效应，当所有人都想挖你的人？

So how do you prevent that? How do you prevent the Pixie Dust effect when everyone's trying to coach all your people?

我觉得我们做不了太多来阻止。但这也是为什么特斯拉在硅谷同时又有仙尘效应，意味着挖角非常猛烈。

I don't think we can do much to stop it. But that's one of the reasons why Tesla being in Silicon Valley and having the Pixie Dust thing at the same time meant there was very aggressive recruitment.

那搬到奥斯汀有帮助。

Being in Austin helps then.

奥斯汀，是的，还是有帮助。特斯拉大部分工程还在加州。要让工程师搬家，我称之为伴侣问题。伴侣有工作。所以对星舰基地来说尤其困难，因为找到非 SpaceX 工作的几率很低。非常困难。就像个技术修道院，偏远且大多是男的。

Austin, yeah, it still helps. I mean, Tesla still has a majority of its engineering in California. So for getting engineers to move, I call it the significant other problem. Others have jobs. So for Starbase, that was particularly difficult. Since the odds of finding a non-SpaceX job are pretty low. It's quite difficult. It's like a technology monastery thing. Remote and mostly dudes.

但回到那些在特斯拉、SpaceX 等地技术能力非常强的人，你觉得他们除了技术基础扎实外还有什么共同点？还是说跟组织有关？是他们与你合作的能力？是灵活但不过分灵活的能力？什么能成为你好的切磋伙伴？

But if you go back to these people who've really been very effective in a technical capacity at Tesla, at SpaceX, and those sorts of places, what do you think they have in common other than being very sharp on the technical foundation? Or do you think it's something organizational? Something about their ability to work with you? Is it their ability to be flexible but not too flexible? What makes a good sparring partner for you?

我不觉得是切磋伙伴。我的意思是，如果有人能成事，我就爱他们；如果不成，我就不爱。就这么简单，不是什么古怪的事。如果执行得好，我是超级粉丝；如果不好，就不是。但这跟我的个人偏好无关。我当然会努力不让它跟我的个人偏好挂钩。

I don't think a sparring partner. I mean, if somebody gets things done, I love them, and if they don't, I don't. So it's pretty straightforward. It's not like some idiosyncratic thing. If somebody executes well, I'm a huge fan, and if they don't, I'm not. But it's not about mapping to my idiosyncratic preferences. I'll certainly try not to have it be mapping to my idiosyncratic preferences.

嗯，是的，但总的来说，我认为招聘时看重才能、干劲和可信度是个好主意。而且我认为心地善良很重要。我一度坚持这一点。所以，要看他们是不是好人、值得信赖、聪明、有才华又勤奋？如果是，你可以再补充领域知识。但那些基本特质，那些基本属性，你是无法改变的。特斯拉和 SpaceX 的大多数人并非来自航空航天或汽车行业。

Um, yeah, but generally I think it's a good idea to hire for talent and drive and trustworthiness. And I think goodness of heart is important. I'd wait at that at one point. So like are they a good person, trustworthy, sort of smart, talented and hardworking? If so, you can add domain knowledge. But those fundamental traits, those fundamental properties you cannot change. Most of the people who are at Tesla and SpaceX did not come from the aerospace industry or the auto industry.

随着你的公司从 100 人扩展到 1000 人再到 10000 人，你的管理风格中最大的变化是什么？你以非常微观的管理著称，就是深入细节。

What is most sad to change about your management style as your companies have scaled from 100 to thousand to 10,000 people? You're known for this very micromanagement, just getting into the details of things.

请叫它纳米管理。

Nano management, please.

皮米管理。

Pico management.

嗯，幻影管理。

Um, phantom management.

所以你是说我们要一直深入到波士顿侧翼，一直到海森堡。我说它们很小。

So you're saying we're gonna go all the way down to flanks Boston all the way down to Heisenberg. I said they were small.

是的。

Yeah.

你怎么……我的意思是，你还能像以前那样深入细节吗？如果公司更小，它们会不会更成功？你怎么看？

How do you... I mean, are you still able to get into details as much as you want? Would your companies be more successful if they were smaller? How do you think about that?

嗯，因为我每天的时间是固定的，随着事情的增长和活动范围的扩大，我的时间必然被稀释。所以我实际上不可能成为一个微观管理者，因为那意味着我每天有几千个小时。对我来说，微观管理在逻辑上是不可能的。所以现在，有时我会深入一个具体问题，因为那个问题是公司进展的制约因素。深入那个非常细节的项目是因为它是制约因素，而不是随意地钻牛角尖。而且就像我说的，从时间角度来看，我根本不可能随意去管那些无关紧要的小事，那会导致失败。但有时候，小事却对胜利起决定性作用。

Well, because I have a fixed amount of time in the day, my time is necessarily diluted as things grow and as the span of activity increases. So it's impossible for me to actually be a micromanager because that would imply I have thousands of hours per day. It is a logical impossibility for me to micromanage things. So now there are times when I will drill down into a specific issue because that specific issue is the limiting factor on the progress of the company. And the reason for drilling into that very detailed item is because it is the limiting factor, not arbitrarily drilling into tiny things. And like I said, from a time standpoint, it is physically impossible for me to arbitrarily go into tiny things that don't matter, and that would result in failure. But sometimes the tiny things are decisive in victory.

众所周知，你把星舰的设计从复合材料改成了钢材。而且你做出这个决定，并不是……你知道，人们到处说‘哦，老板，我们找到了更好的东西。’那是你在鼓励人们克服阻力。你能告诉我们你是怎么做出这个从复合材料到钢材的转变的吗？

Famously, you switched the Starship design from composites to steel. And you made that decision like that wasn't a... you know, people were going around, they're like 'oh we found something better boss.' That was you encouraging people against some resistance. Can you tell us how you came to that whole composite steel switch?

呃，是的。我想说是绝望。最初，是的，我们打算用碳纤维制造星舰。碳纤维非常昂贵。通常，当你进行批量生产时，任何东西的成本都会接近其材料成本。碳纤维的问题在于材料成本仍然非常高。大约是 50 倍，特别是如果你使用能够处理低温氧的高强度特种碳纤维，它的成本大约是钢材的 50 倍。而且至少在理论上，它会更轻。人们通常认为钢材重，碳纤维轻。对于室温应用，比如一级方程式赛车、静态空气动力学结构或任何空气动力学结构，碳纤维可能更好。但现在的问题是，我们试图用碳纤维制造这个巨大的火箭，而我们的进展极其缓慢。

Uh, yeah. So desperation, I'd say. Originally, yes, we were going to make Starship out of carbon fiber. And carbon fiber is pretty expensive. Generally, when you do volume production, you can get any given thing to start to approach its material cost. The problem with carbon fiber is that material cost is still very high. It's about 50 times, particularly if you go for a high-strength specialized carbon fiber that can handle cryogenic oxygen, it's roughly 50 times the cost of steel. And at least in theory, it would be lighter. People generally think of steel as being heavy and carbon fiber as being light. For room temperature applications, like a Formula 1 car, static aeros structure, or any kind of aeros structure really, you're probably better off with carbon fiber. Now the problem is that we were trying to make this enormous rocket out of carbon fiber, and our progress was extremely slow.

而且一开始选择它只是因为轻。

And it had been picked in the first place just because it's light.

是的。乍一看，大多数人会认为制造轻质材料的选择是碳纤维。问题是，当你用碳纤维制造非常巨大的东西，然后试图让碳纤维高效固化时，任何非室温条件……因为有时你有 50 层碳纤维，而碳纤维实际上是碳丝和胶水。为了获得高强度，你需要一个高压釜，本质上是一个高压烤箱。如果你有一个巨大的东西，烤箱必须比火箭还大。所以我们得制造一个比任何现存高压釜都大的高压釜。或者进行室温固化，但这需要很长时间并且有问题。但根本问题是，我们用碳纤维进展非常缓慢。

Yes. At first glance, most people would think that the choice for making something light would be carbon fiber. Now the thing is that when you make something very enormous out of carbon fiber and then you try to have the carbon fiber be efficiently cured, anything not room temperature... because you've got sometimes 50 plies of carbon fiber, and carbon fiber is really carbon string and glue. And in order to have high strength, you need an autoclave, something that is essentially a high-pressure oven. And if you have something that's gigantic, the oven's got to be bigger than the rocket. So we'd be trying to make an autoclave that's bigger than any autoclave that's ever existed. Or do room temperature cure, which takes a long time and has issues. But the fundamental issue is that we were just making very slow progress with carbon fiber.

所以元问题是，为什么必须由你来做这个决定。你的团队里有很多工程师。团队为什么没有想到钢材？

So the meta question is why it had to be you who made that decision. There are many engineers on your team. How did the team not arrive at steel?

是的，没错。这是一个更广泛问题的一部分，比如理解你在公司中的比较优势。所以，是因为我们用碳纤维进展非常缓慢。我当时想，‘好吧，我们必须尝试其他东西。’对于猎鹰 9 号，主结构是用铝锂合金制造的，它具有非常好的强度重量比。实际上，对于它的应用，强度重量比与碳纤维差不多，甚至更好。但铝锂合金很难加工。要焊接它，你必须进行一种叫做摩擦搅拌焊的工艺，在不使金属进入液态的情况下连接金属。你能做到这一点有点不可思议。但通过这种特殊的焊接，你可以做到。但是，比如说，要对铝锂合金进行修改或附加东西非常困难。你现在必须使用带密封的机械连接。你不能焊接上去。所以我们想避免在星舰的主结构中使用铝锂合金。而且有一种非常特殊的碳纤维等级，具有非常好的质量特性。所以对于火箭，你真正想要的是最大化火箭中推进剂的百分比，显然要最小化质量。但就像我说的，我们进展非常缓慢，我说按照这个速度我们永远到不了火星，所以我们要想别的办法。我不想使用铝锂合金，因为摩擦搅拌焊的难度，尤其是在大规模生产时。在直径 3.6 米时已经够难了，更不用说 9 米或以上了。然后我说，那么钢材怎么样？我对此有一个线索，因为一些早期的美国火箭使用了非常薄的钢材。阿特拉斯火箭使用了钢制气球储罐。

Yeah, exactly. This is part of a broader question like understanding your comparative advantage at your companies. So it was because we were making very slow progress with carbon fiber. I was like, 'Okay, we've got to try something else.' Now, for the Falcon 9, the primary airframe is made of aluminum lithium, which has very good strength-to-weight ratio. Actually, it has about the same, maybe better strength-to-weight for its application than carbon fiber. But aluminum lithium is very difficult to work with. In order to weld it, you have to do something called friction stir welding, where you join the metal without it entering the liquid phase. It's kind of wild that you could do that. But with this particular type of welding, you can do that. But it's very difficult to, say, make a modification or attach something to aluminum lithium. You now have to use mechanical attachment with seals. You can't weld it on. So we wanted to avoid using aluminum lithium for the primary structure for Starship. And there was this very special grade of carbon fiber that had very good mass properties. So with a rocket, you're really trying to maximize the percentage of the rocket that is propellant, minimize the mass obviously. But like I said, we were making very slow progress, and I said at this rate we're never going to get to Mars, so we're going to think of something else. I didn't want to use aluminum lithium because of the difficulty of friction stir welding, especially doing that at scale. It was hard enough at 3.6 meters in diameter, let alone at 9 meters or above. Then I said, well, what about steel? And I had a clue here because some of the early US rockets had used very thin steel. The Atlas rockets had used a steel balloon tank.

所以并不是说钢从未被使用过。它实际上曾被使用过。当你查看不锈钢的材料特性时，特别是经过充分硬化的应变硬化不锈钢在低温下的表现，其强度重量比实际上与碳纤维相似。所以如果你在室温下看材料特性，钢的重量似乎是碳纤维的两倍。但如果你看特定牌号的充分硬化不锈钢在低温下的材料特性，那么它的强度重量比实际上与碳纤维相当。对于星舰来说，燃料和氧化剂都是低温的。对于猎鹰 9 号，燃料是火箭推进剂级煤油，基本上是高度纯净的航空煤油，大约是室温。虽然我们确实会将其稍微冷却，像冷却啤酒一样。美味。是的，我们确实会冷却，但它不是低温的。事实上，如果把它冷却到低温，它会变成蜡。但对于星舰，它使用的是液态甲烷和液态氧；它们在相似的温度下是液态的。所以基本上整个主要结构都处于低温状态。那么你就有一种 300 系列不锈钢，经过应变硬化，因为在低温下，它的强度重量比实际上与碳纤维相当，但成本比碳纤维低 50 倍，而且非常容易加工。你可以在户外焊接不锈钢。你可以在焊接不锈钢时抽雪茄。它非常有韧性。你可以轻松修改它。如果你想连接什么东西，直接焊上去就行。所以非常容易加工，成本很低，而且在低温下，强度重量比与碳纤维相当。然后考虑到我们大大减少了隔热罩的质量，因为钢的熔点远高于铝的熔点。大约是铝熔点的两倍。所以你可以让火箭运行得更热。是的。特别是对于飞船，它像一颗炽热的流星一样返回，你可以大大减少隔热罩的质量。所以你可以将迎风面隔热罩的质量减少一半，而背风面根本不需要隔热罩。所以最终结果是钢制火箭比碳纤维火箭更轻，因为碳纤维火箭中的树脂会开始熔化。基本上碳纤维和铝的工作温度能力差不多，而钢可以在两倍的温度下工作。我的意思是，这些都是非常粗略的近似。人们会……我不会去抠字眼。人们会说：“哦，他说是两倍。实际上是 0.8。”闭嘴。这就是主要评论的内容。是的。该死的。好吧。关键是，事后看来，我们一开始就应该用钢。不用钢是愚蠢的。

So it's not like steel never been used before. It actually had been used. And when you look at the material properties of stainless steel, especially if it's been full hard strain hardened stainless steel at cryogenic temperature, the strength to weight is actually similar to carbon fiber. So if you look at the material properties at room temperature, it looks like steel is going to be twice as heavy. But if you look at the material properties at cryogenic temperature of full hard steel stainless of particular grades, then you actually get to a similar strength to weight as carbon fiber. And in the case of Starship, both the fuel and the oxidizer are cryogenic. For Falcon 9, the fuel is rocket propellant grade kerosene, basically a very pure form of jet fuel, which is roughly room temperature. Although we do chill it slightly below, we chill it like a beer. Delicious. Yeah, we do chill it, but it's not cryogenic. In fact, if we made it cryogenic, it would just turn to wax. But for Starship, it's liquid methane and liquid oxygen; they are liquid at similar temperatures. So basically almost the entire primary structure is at cryogenic temperature. So then you've got a 300 series stainless that's strain hardened, because it's at cryogenic temperature, actually has a similar strength to weight as carbon fiber, but costs 50 times less than carbon fiber and is very easy to work with. You can weld stainless steel outdoors. You could smoke a cigar while welding stainless steel. It's very resilient. You can modify it easily. If you want to attach something, you just weld it right on. So very easy to work with, very low cost, and at cryogenic temperature, similar strength to weight to carbon fiber. Then when you factor in that we have a much reduced heat shield mass because the melting point of steel is much greater than the melting point of aluminum. It's about twice the melting point of aluminum. So you can just run the rocket much hotter. Yes. So especially for the ship, which is coming in like a blazing meteor, you can greatly reduce the mass of the heat shield. So you can cut the mass of the windward part of the heat shield maybe in half, and you don't need any heat shielding on the leeward side. So the net result is actually the steel rocket weighs less than the carbon fiber rocket, because the resin in the carbon fiber rocket starts to melt. Basically carbon fiber and aluminum have about the same operating temperature capabilities, whereas steel can operate at twice the temperature. I mean, these are very rough approximations. People will... I won't go to the rocket face. People will say, "Oh, he said it's twice. It's actually 0.8." Shut up. That's what the main comment's going to be about. Yeah. God damn it. Okay. The point is, in retrospect, we should have started with steel in the beginning. It was dumb not to do steel.

但让我复述一下，我听到的是，除了早期美国火箭外，钢是一条风险更高、不太成熟的路径，而碳纤维则是一条更差但更成熟的路径。所以你需要成为那个推动者，说，嘿，我们要走这条风险更高的路，然后想办法解决。所以你是在与某种保守主义作斗争。

But to play this back to you, what I'm hearing is that steel was a riskier, less proven path other than the early US rockets versus carbon fiber was like a worse but more proven out path. And so you need to be the one to push for, hey, we're going to do this riskier path and just figure it out. And so you were fighting like a sort of conservatism in a sense.

这就是为什么我最初说问题在于我们进展不够快。我们连制造一个没有皱纹的小型碳纤维筒段都有困难。因为在那种大尺度下，你必须有很多层碳纤维。你必须固化它，而且必须以某种方式固化，使其没有任何皱纹或缺陷。碳纤维的弹性远不如钢。它的韧性差得多。所以不锈钢会拉伸和弯曲；碳纤维则会碎裂。韧性是应力-应变曲线下的面积，所以你通常用钢会更好，或者更准确地说，是不锈钢。

That's why I initially said that the issue is that we weren't making fast enough progress. We were having trouble making even a small barrel section of the carbon fiber that didn't have wrinkles in it. Because at that large scale, you have to have many layers of the carbon fiber. You've got to cure it and you've got to cure it in such a way that it doesn't have any wrinkles or defects. The carbon fiber is much less resilient than steel. It has much less toughness. So stainless steel will stretch and bend; the carbon fiber will tend to shatter. So toughness being the area under the stress-strain curve, you're generally going to do better with steel, or stainless steel to be precise.

还有一个关于星舰的问题。我大约两年前和 Siler 一起参观了星舰基地，那真是太棒了。在很多方面都非常酷。我注意到一件事，人们真的为事物的简单性感到自豪，每个人都想告诉你星舰就像一个大汽水罐，我们在招聘焊工，如果你能在任何工业项目中焊接，你就能在这里焊接，但他们对简单性非常自豪，而猎鹰 9 号是一枚非常复杂的火箭。所以我想问的是，事情是更简单了还是更复杂了？

One other Starship question. So I visited Starbase I think it was two years ago with Siler and that was awesome. It was very cool to see in a whole bunch of ways. One thing I noticed was that people really took pride in the simplicity of things where you know everyone wanted to tell you how Starship is just a big soda can and you know we're hiring welders and you know if you can weld in any industrial project you can weld here but there's a lot of pride in the simplicity and Falcon 9 was a very complicated rocket. So that's what I'm getting at is are things simpler or are they complex?

我想他们可能想说的是，你不需要有火箭行业的先前经验就能参与星舰项目。一个人只需要聪明、努力、值得信赖，就能参与火箭制造。他们不需要火箭经验。星舰是人类有史以来制造的最复杂的机器，遥遥领先。

I think maybe what they're trying to say is that you don't have to have prior experience in the rocket industry to work on Starship. Somebody just needs to be smart and work hard and be trustworthy, then they can work on a rocket. They don't need prior rocket experience. Starship is the most complicated machine ever made by humans by a long shot.

在哪些方面？

In what regards?

任何方面。我认为没有更复杂的机器了。我觉得我能想到的任何项目都比这个容易。这就是为什么没有人制造出快速可重复使用的，没有人制造出完全可重复使用的轨道火箭。这是一个非常非常困难的问题。许多聪明人之前尝试过，非常聪明、拥有巨大资源的人，他们失败了。而我们还没有成功。猎鹰是部分可重复使用的，但上面级不是。星舰第三版，我认为这个设计，它可以完全可重复使用，而完全可重复使用将使我们能够成为多行星文明。

Anything really. I'd say there isn't a more complex machine. I'd say that pretty much any project I can think of would be easier than this. And that's why no one has made a rapidly reusable, nobody has ever made a fully reusable orbital rocket. It's a very, very hard problem. Many smart people have tried before, very smart people with immense resources, and they failed. And we haven't succeeded yet. Falcon is partially reusable, but the upper stages are not. Starship version three, I think this design, it can be fully reusable, and that full reusability is what will enable us to become a multiplanet civilization.

我们花了很多时间讨论瓶颈。你能说说当前星舰的瓶颈是什么吗，哪怕只是高层次的？

We spent a lot of time on bottlenecks. Can you say what the current Starship bottlenecks are even at a high level?

我的意思是，总的来说，就是让它不爆炸。老生常谈了。它真的想爆炸。

I mean, trying to make it not explode generally. That old chestnut. It really wants to explode.

那些燃烧测试中，我们有两台发动机爆炸了。其中一次炸毁了整个测试设施。所以只需要一个错误，而星舰蕴含的能量是惊人的。

Those combustion we've had two bristers explode on the test. One obliterated the entire test facility. So it only takes like one mistake and the amount of energy contained in a Starship is insane.

所以这就是为什么它比猎鹰更难？只是因为能量更大？

And so is that why it's harder than Falcon? It's because it's just more energy.

这涉及很多新技术。它在推动性能极限。猛禽 3 发动机非常先进，是迄今为止最好的火箭发动机。但它非常容易爆炸。举个例子，起飞时火箭产生超过 100 吉瓦的功率，相当于美国电力的 20%。太疯狂了。

It's a lot of new technology. It's pushing the performance envelope. The Raptor 3 engine is a very advanced engine, by far the best rocket engine ever made. But it desperately wants to blow up. Just to put things in perspective, on liftoff the rocket is generating over 100 gigawatts of power. That's 20% of US electricity. So insane.

这个对比很好。

That's a great comparison.

有时不爆炸。有时。所以我想它怎么才能不爆炸？有几千种方式会爆炸，只有一种不会。所以我们希望它不爆炸，而是每天可靠地飞行，比如每小时一次。显然它经常爆炸。维持这一点非常困难。

While not exploding sometimes. Sometimes. So I was like how does it not explode? There are thousands of ways it could explode and only one way it doesn't. So we want it to not explode but fly reliably on a daily basis, like once per hour. Obviously it blows up a lot. It's very difficult to maintain that.

是的。

Yes.

那么星舰剩下的最大问题是什么？是让隔热罩可重复使用。从来没有人制造过可重复使用的轨道隔热罩。隔热罩必须在上升阶段不震掉大量瓦片，然后在返回时也不掉大量瓦片或使主结构过热。

And then what's the single biggest remaining problem for Starship? It's having the heat shield be reusable. No one has ever made a reusable orbital heat shield. The heat shield has to make it through the ascent phase without shocking a bunch of tiles, and then come back in without losing a bunch of tiles or overheating the main airframe.

这难道不难吗？它本质上是一种消耗品。

Isn't that hard? It's kind of fundamentally a consumable.

嗯，是的，但你车上的刹车片也是消耗品，但它们能用很长时间。

Well, yes, but your brake pads in your car are also consumable, but they last a very long time.

有道理。好吧。

Fair. Okay.

所以它只需要能用很长时间。我们曾让飞船返回并在海上软着陆几次，但掉了大量瓦片。不经过大量维修就无法重复使用。所以即使它软着陆了，不经过大量工作也无法重复使用。所以剩下的最大问题就是完全可重复使用的隔热罩。

So it just needs to last a very long time. We have brought the ship back and had it do a soft landing in the ocean a few times, but it lost a lot of tiles. It was not reusable without a lot of work. So even though it did land softly, it would not have been reusable without a lot of work. So that's the biggest problem remaining: fully reusable heat shield.

所以如果你想让它着陆、加注推进剂并再次飞行，而不需要费力检查 4 万块瓦片。我很好奇你是如何推动紧迫感的。我读你的传记时，感觉你能够推动紧迫感和规模化。为什么你认为像 SpaceX 和特斯拉这样的公司现在很大了，但你仍然能保持那种文化？其他公司出了什么问题？

So if you want to be able to land it, refill propellant, and fly again without laborious inspection of 40,000 tiles. I'm curious how you drive urgency. When I read biographies of yours, it seems you're able to drive a sense of urgency and scale. Why do you think other organizations like SpaceX and Tesla are big companies now but you still keep that culture? What goes wrong with other companies?

我不知道。

I don't know.

但今天你说你有一堆 SpaceX 的会议。你在那里做了什么来增加紧迫感？

But today you said you had a bunch of SpaceX meetings. What is it that you're doing there that's adding urgency?

嗯，我想紧迫感来自公司的领导者。所以如果我的紧迫感是疯狂的，那种疯狂的紧迫感就会传递到公司其他部门。

Well, I guess the urgency comes from whoever's leading the company. So if my sense of urgency is maniacal, that maniacal sense of urgency projects through the rest of the company.

是因为后果吗？比如如果你知道埃隆说了个疯狂的截止日期，但如果我没完成，我知道我会怎样。还是你能够识别瓶颈并消除它们，让人们快速行动？你怎么看待你的公司为什么能快速行动？

Is it because of consequences? Like if you know Elon said a crazy deadline, but if I don't get it, I know what happens to me. Is it just you're able to identify bottlenecks and get rid of them so people can move fast? How do you think about why your companies are able to move fast?

我一直在解决限制因素。在截止日期方面，我通常试图设定一个我认为有 50%概率能达到的期限。所以这不是一个不可能的期限，但这是我能想到的最激进的、有 50%概率能实现的期限。这意味着它有一半时间会延迟。有一个气体膨胀定律适用于时间表：无论你给出什么时间表，如果你说我们要在 5 年内完成，这对我来说就像无限时间，它会膨胀到填满可用时间，并花掉 5 年。存在物理极限，比如物理会限制你做某些事情的速度。扩大制造规模时，移动原子和扩大制造的速度是有限的。你不能瞬间每年制造一百万个东西。你必须设计生产线，启动它，沿着生产 S 曲线前进。所以我认为疯狂的紧迫感非常重要。你想要一个激进的时间表，随时找出限制因素，并帮助团队解决那个限制因素。

I'm constantly addressing the limiting factor. On the deadlines front, I generally try to aim for a deadline that I think is at the 50th percentile. So it's not an impossible deadline, but it's the most aggressive deadline I can think of that could be achieved with 50% probability. That means it'll be late half the time. There is a law of gases expansion that applies to schedules: whatever schedule you give, if you say we're going to do this in 5 years, which to me is like infinity time, it will expand to fill the available schedule and take 5 years. There's a physical limit like physics will limit how fast you can do certain things. Scaling up manufacturing, there's a rate at which you can move atoms and scale manufacturing. You can't instantly make a million of something a year. You have to design the manufacturing line, bring it up, ride the S-curve of production. So I guess a maniacal sense of urgency is a very big deal. You want to have an aggressive schedule, figure out the limiting factor at any point, and help the team address that limiting factor.

你能谈谈星链吗？它缓慢推进了很多年。

Can you maybe talk about Starlink? It was slowly in the works for many years.

是的，我们从公司一开始就讨论过它。

Yeah, we talked about it all the way in the beginning of the company.

是的。然后你在雷德蒙德建立了一个团队，但后来你决定这个团队不行。但同样，它缓慢进行了几年，为什么你没有早点行动？为什么你在那个时候行动？为什么那是正确的时机？

Yeah. And so then there was a team you had built in Redmond and then at one point you decided this team is just not cutting it. But again, how did it go for a few years slowly and why didn't you act earlier and why did you act when you did? Why was that the right moment?

我每周都有非常详细的工程评审。这可能是一个非常不寻常的细节程度。我不知道有谁经营公司，至少是制造公司，会深入到我所做的细节程度。所以并不是我不太了解实际情况。我们会详细地过一遍事情。我非常相信越级会议，不是让向我汇报的人说话，而是让向他们汇报的每个人在技术评审中发言。而且不能提前准备。否则你就会得到粉饰过的信息，就像我最近说的。

I have very detailed engineering reviews weekly. That's maybe a very unusual level of granularity. I don't know anyone who runs a company, at least a manufacturing company, that goes into the level of detail I go into. So it's not as though I don't have a pretty good understanding of what's actually going on. We go through things in detail. I'm a big believer in skip level meetings where the individuals instead of having the person that reports to me say things, everyone that reports to them says something in the technical review. And there can't be advanced preparation. Otherwise you get glazed, as I say these days.

你就随便叫他们，像——

You just like call them randomly like

不，就是挨个房间走一圈，每个人汇报进展。

No, just go around the room and everyone provides an update.

好吧。嗯，我是说，脑子里要记很多信息，因为如果你每周或每周开两次会，你会对每个人说的话有个快照。然后你可以在心里把进展点画成曲线，问自己：我们是在收敛到解决方案吗？或者，只有当我断定成功不在可能的结果集里时，我才会采取激烈行动。所以当我最终得出结论，除非采取激烈行动否则我们没有成功机会时，我就必须采取激烈行动。2018 年我就是这么做的，采取了激烈行动，解决了问题。

Okay. Um, so, I mean it's a lot of information to keep in your head because you've got a snapshot of what that person said if you have meetings weekly or twice weekly. And you can then plot the progress points mentally on a curve and say, are we converging to a solution or not? Or I'll take drastic action only when I conclude that success is not in a set of possible outcomes. So when I finally reach the conclusion that unless drastic action is done we have no chance of success, then I must take drastic action. That's what I came to in 2018, took drastic action and fixed the problem.

你有多少家公司？你有很多公司，在每一家你都做这种对相关瓶颈的深度工程理解，以便和人们做这些评审。

How many companies do you have? You've got many companies, and in each of them it sounds like you do this kind of deep engineering understanding of the relevant bottlenecks so you can do these reviews with people.

是的。

Yeah.

你已经把它扩展到五六七家公司了。其中一家公司内部还有很多不同的子公司。什么决定了上限？你能有 80 家公司吗？

You've been able to scale it up to five, six, seven companies. Within one of these companies you have many different mini companies within them. What determines the maximum? Could you have like 80 companies?

80？不。但你已经有很多了，这已经很了不起了。

80? No. But like you have so many already, that's already remarkable.

就目前这个数字。是的。

By this current number. Yeah.

对，没错。

Yeah. Exactly.

呃不。所以我们连一家都很难维持好。

Uh no. So we can barely keep one coming together.

这取决于情况。我实际上和无聊公司没有定期会议。那家公司基本在平稳运行。如果某件事进展顺利，我就不值得花时间。所以我根据瓶颈或问题所在来分配时间，哪里有问题，或者什么在拖后腿。我关注瓶颈因素。所以如果某件事进展很好，他们很少见到我；但如果进展糟糕，他们会经常见到我。

It depends on the situation. I actually don't have regular meetings with the Boring Company. That company is sort of cruising along. If something is working well and making good progress, there's no point in me spending time on it. So I allocate time according to where the limiting factor or the problem is, where things are problematic, or what is holding us back. I focus on the limiting factor. So if something is going really well, they don't see much of me, but if something is going badly, they'll see a lot of me.

甚至不一定是糟糕，而是如果某件事是瓶颈。

Something or not even badly, it's like if something's a limiting factor.

没错。不一定是进展糟糕，而是我们需要加速推进的事情。所以当某件事在 SpaceX 或特斯拉是瓶颈时，你会每周甚至每天和负责的工程师沟通吗？具体怎么操作？

Exactly. It's not exactly going badly, but it's the thing we need to make go faster. So when something's a limiting factor at SpaceX or Tesla, are you talking weekly and daily with the engineer working on it? How does that actually work?

大多数瓶颈事项是每周一次，有些是每周两次。AI5 芯片评审是每周两次，每周二和周六。

Most things that are limiting factors are weekly, and some things are twice weekly. The AI5 chip review is twice weekly, every Tuesday and Saturday.

时间上是不固定的吗？

Is it open-ended in how long it goes?

理论上是的，但通常两三个小时。有时更短。取决于要过多少信息。

Technically yes, but usually it's two or three hours. Sometimes less. It depends on how much information you've got to go through.

是的。嗯，这是另一回事。我只是想梳理出这里的差异，因为结果似乎很不同。感觉在企业界，CEO 做工程评审并不总是发生，尽管那正是公司该做的事。但时间通常被精细地切成半小时甚至 15 分钟的会议。而你似乎更倾向于开放式的、一直讨论到解决问题为止的会议。

Yeah. Well, that's another thing. I'm just trying to tease out the differences here, because the outcomes seem quite different. It feels like in the corporate world, the CEO doing engineering reviews does not always happen, despite that being what the company is doing. But time is often finely sliced into half-hour meetings or even 15-minute meetings. And it seems like you hold more open-ended, we'll talk until we figure it out type meetings.

有时会，但大多数或多或少能按时结束。今天的星舰工程评审时间稍长，因为要讨论的话题更多。要弄清楚如何将年入轨量扩大到百万吨以上，相当有挑战性。

Sometimes, but most of them seem to more or less stay on time. Today's Starship engineering review went a bit longer because there were more topics to discuss. Trying to figure out how to scale to a million plus tons to orbit per year is quite challenging.

我能问个问题吗？你提到 Optimus 和 AI 将在几年内带来两位数的增长率。

Can I ask a question? So you said about Optimus and AI that they're going to result in double-digit growth rates within a matter of years.

哦，像经济那样。

Oh, like the economy.

是的。

Yeah.

是的，我认为没错。

Yes, I think that's right.

如果经济会增长这么多，那 Doge 削减开支的意义何在？

What was the point of the Doge cuts if the economy is going to grow so much?

嗯，我认为浪费和欺诈不是好事。我其实很担心。如果没有 AI 和机器人，我们真的完蛋了。国家债务堆积如山。我们的国债利息支出已经超过了军事预算，那是一万亿美元。所以仅利息就超过一万亿美元。我对此非常担忧。也许我花点时间，可以减缓美国的破产速度，为 AI 和机器人争取足够的时间来解决国家债务。这是唯一能解决国家债务的办法。没有 AI 和机器人，我们百分之百会作为一个国家破产并失败。其他任何东西都解决不了国家债务。所以我们需要足够的时间来建造 AI 和机器人，并且在此之前不要破产。

Well, I think waste and fraud are not good things to have. I was actually pretty worried. In the absence of AI and robotics, we're actually totally screwed. The national debt is piling up like crazy. Our interest payments on the national debt exceed the military budget, which is a trillion dollars. So over a trillion dollars just in interest payments. I was pretty concerned about that. Maybe if I spend some time, we can slow down the bankruptcy of the United States and give us enough time for the AI and robots to help solve the national debt. It's the only thing that could solve the national debt. We are 1,000% going to go bankrupt as a country and fail without AI and robots. Nothing else will solve the national debt. So we need enough time to build the AI and robots and not go bankrupt before then.

我想我好奇的是，当 Doge 开始时，你有巨大的能力来实施改革。

I guess the thing I'm curious about is when Doge starts, you have this enormous ability to enact reform.

没那么巨大。

Not that enormous.

当然。当然。但完全按照你的观点，AI 和机器人推动产品改进、推动 GDP 增长很重要，但为什么不直接针对你指出的那些问题，比如某些组件的关税或许可呢？

Sure. Sure. But to totally by your point that it's important that AI and robotics drive product improvements, drive GDP growth, but why not just directly go after the things you were pointing out, like the tariffs on certain components or permitting?

我不是总统，而且即使是很明显的浪费和欺诈，比如荒谬的浪费和欺诈，也很难削减。我发现，即使是从政府中削减非常明显的欺诈方式也极其困难，因为政府必须根据谁在抱怨来运作。如果你切断对欺诈者的付款，他们立刻会编出最令人同情的理由来继续付款。他们不会说‘请继续欺诈’。他们会说，你知道，‘你在杀害熊猫宝宝’。与此同时，根本没有熊猫宝宝在死去。他们只是编造。这些势力能够编造出极其令人信服、令人心碎的故事，虽然是假的，但听起来很同情。这就是发生的事情。也许我本该更清楚。

I'm not the president, and it's very hard to cut even things that are obvious waste and fraud, like ridiculous waste and fraud. What I discovered is that it's extremely difficult even to cut very obvious ways of fraud from the government, because the government has to operate on who's complaining. If you cut off payments to fraudsters, they immediately come up with the most sympathetic-sounding reasons to continue the payment. They don't say 'please keep the fraud going.' They say, you know, 'you're killing baby pandas.' Meanwhile, no baby pandas are dying. They're just making it up. The forces are capable of coming up with extremely compelling, heart-wrenching stories that are false but nonetheless sound sympathetic. That's what happened. Perhaps I should have known better.

嗯，我想，等等，让我们试着削减政府的一些浪费和核心问题。也许不应该有 2000 万人在社会保障系统中被标记为活着，而他们已经去世很久且超过 115 岁。最年长的美国人是 114 岁。所以，如果有人在 115 岁还被标记为活着，那要么是录入错误。所以应该有人打电话给他们说：‘我们好像搞错了你的生日，或者我们需要把你标记为去世。’

Um and uh but I thought wait let's take a let's let's let's try to cut some amount of waste and core from the government. Maybe there shouldn't be you know 20 million people uh marked as alive in social security who are indefinitely dead and over the age of 115. The oldest American is 114. So, it's safe to say if somebody's 115 and marked as live in the social security database, um something is there's either a typo. So, like somebody should call them and say, 'We we seem to have your birthday wrong or or uh or or we need to mark you as dead.'

两者之一。

One of the two things.

这通电话可够吓人的。

Very intimidating call to get.

嗯，这看起来挺合理的。

Well, so it seems like a reasonable thing.

嗯，如果他们的生日在未来，比如他们有一笔小企业管理局的贷款，而生日是 2165 年，那要么又是录入错误，要么就是欺诈。所以我们说我们好像搞错了你的出生世纪。

Um and if if like say their birthday is in the future um and they have you know a small business administration loan and their birthday is 2165 um we either again have a typo or we have fraud. Um so we say we appear to have gotten the century of your birth incorrect

或者是个电影好素材。

or a great plot for a movie.

对。这就是我说的荒唐欺诈，就是这个意思。

Yes. This is this this is when I when I mean about ludicrous fraud that's what I meant ludicrous fraud.

那些人收到过付款吗？

Were those people getting payments?

有些人确实从社会保障中收到过付款，但主要的欺诈手段是在社会保障系统中把某人标记为活着，然后利用其他所有政府支付系统进行欺诈，因为这些系统只会向社会保障数据库做一个‘是否活着’的查询。

Some some were getting payments from social security but but but the main fraud vector uh was to mark somebody as alive in social security and then use every other government payment system uh to basically to do fraud because what those other government payment systems do would do they will simply do an are you alive check to the social security database.

这就像个迂回战术。

It's a it's a bank shot.

你估计这种机制造成的欺诈总额大概是多少？

What would you estimate is like the total uh amount of fraud from this mechanism.

嗯，我猜是——顺便说一句，政府问责局之前做过这类估算。我不是唯一这么说的人。事实上，我认为 GAO 在拜登政府期间做过分析，粗略估计欺诈金额约为 5000 亿美元。

Um my guess is and and other by the way the the government accountability office has done these estimates before. I'm not the only one who's coming out of this you know in fact I think they they did the GAO did analysis a rough estimate of fraud during the Biden administration and calculated at roughly half a trillion dollars.

所以别光听我说，看看拜登政府期间发布的报告。怎么样？

So don't take my word for it. Take it a report issued during the Biden administration. How about that

来自这个社会保障机制？

from this social security mechanism?

嗯，这只是其中之一。重要的是要认识到，政府在阻止欺诈方面非常低效，因为不像公司——公司有动力阻止欺诈，因为这会影响公司盈利。但政府呢，他们只管印钱。所以，你需要的是关心和能力，而这些在联邦层面都很稀缺。

Uh it's it's one of many. It's important to appreciate that the the government does not is very ineffective at at stopping fraud because uh it's it's not like like if it was a company like like stopping fraud, you've got a motivation because it's affecting the earnings of your company. Uh but the government just just they just print more money. Um, so it's not uh like you you need you need caring and competence and these are in short supply at at the federal level.

嗯，是啊，抱歉。我是说，你去车管所的时候，会觉得‘哇，这真是个能力典范’吗？嗯，现在想象一下，比车管所还糟，因为它是个能印钱的车管所。

Um, yeah, sorry. I mean, when you go to the DMV, do you think, wow, this is a bastion of competence. Um, well, now imagine it's worse than the DMV because it's a DMV that can print money.

所以，难道不可能吗？至少州一级的车管所需要量入为出，否则会破产。但联邦政府只管印钱。

So, was it not possible? At least the state level DMVs uh need to the states more or less need to stay within their budget to go bankrupt. But the federal government just prints more money.

如果真有 5000 亿美元的欺诈，难道不能削减吗？为什么不能全部砍掉？

Was it not possible to cut that if there's a catchy half a trillion of fraud? Why was it not possible to cut all that?

嗯，因为一旦你——我们确实做了——实际上不，你真的得退一步，重新调整你对能力的预期。因为你所处的世界，你得设法收支平衡，得付账单，你懂的。

Uh because when when as soon as you we did we we actually no you you you you really have to stand back and recalibrate your expectations for competence. uh because uh you you're operating in a world where you know you you've got to sort of make ends meet like you know you got to pay your bills, you got to you know

买麦克风。

buy the microphones.

对，对，正是。所以，如果你没有——这不像是一个庞大、冷漠的怪物官僚机构。甚至也不是一堆只会发钱的单调计算机。比如 Doge 团队做的一件事，听起来很简单，可能每年能节省 1000 亿到 2000 亿美元，就是要求从主国库计算机（叫 PM，大概是支付账户主系统之类的）发出的每笔付款必须有拨款代码，强制性的，不是可选的，而且备注字段里必须有点内容。因为你看，你得重新调整对事情有多蠢的认知。

Yeah. Yeah. Exactly. Um so so you you you if you don't have it's it's not like there's a giant largely uncaring monster bureaucracy. It's not even and and and a bunch of uh monacistic computers that are just that are just sending payments. Um like one of the things that that that the Doge team did there was and it sounds so simple uh that that probably will save um let's say 100 billion maybe 200 billion a year um is simply requiring that payments from the main treasury computer which is called PM it's like payment accounts master or something like that there's 5 trillion payments a year um requiring that any payment go that goes out have a payment um appropriation code, make it mandatory, not optional, and that you have anything at all in the comment field. Um, because you see you have to recalibrate how dumb things are.

你觉得付款发出去时没有拨款代码，没有核对任何国会拨款，也没有说明。这就是为什么战争部（前国防部）无法通过审计，因为信息根本不存在。重新调整你的预期吧。

You think were being sent out with no appropriation code, not not checking back to any congressional appropriation, and no explanation. And this is why the the Department of War, formerly the Department of Defense, cannot pass an audit because the information is literally not there. Recalibrate your expectations.

我想更好地了解南澳大利亚的数字，因为 2024 年有一份监察长报告。

I want to better understand the South Australian number because there was an IG report in 2024.

你肯定觉得为什么这么低？

How you must like why is it so low?

嗯，也许吧。但那份报告发现，七年里社会保障欺诈估计约为 700 亿美元，也就是每年 100 亿。所以我很好奇另外 4900 亿是什么。联邦政府支出每年 7.5 万亿美元。

Um maybe. But uh which found that like over seven years this the social security fraud they estimated was like 70 billions over 7 years. So like 10 billion a year. So I'd be curious to see what like the other 490 billion is. Federal government expenditures are 7 and a half trillion a year.

对。

Yeah.

嗯，你觉得政府有多能干？百分比是多少？

Um how what what percentage how competent do you think government is?

可自由支配支出大概只占 15%。

The the discretionary spending there is like 15%.

对，但这不重要。大部分欺诈是非自由支配的。基本上就是欺诈性的医疗保险、医疗补助、社会保障、残疾保险等等。有无数种政府付款。

Yeah. But but it doesn't matter. Most of the ford is non-discretionary. It's it's basically a fraudulent Medicare, Medicaid, uh social security, uh uh uh you know um disability. Uh it's there's there's a zillion government payments. Yeah.

嗯，而且很多付款实际上是整笔转移给各州的。所以联邦政府在很多情况下甚至没有信息来判断是否存在欺诈。让我们用归谬法来想：假设政府完美无缺，没有欺诈。你估计这个概率是多少？

Um and and a bunch of these payments are in fact u they're uh block transfers to the states. So the federal government doesn't even have the information in a lot of cases to even see know if there's fraud. Let's consider let's like reductio ad absurdum the government the government is perfect and has no fraud. What is your probability estimate of that?

我的意思是零。好吧。那么你会说政府的浪费和欺诈是 90%吗？那也太慷慨了。但如果只有 90%，那就意味着每年有 7500 亿美元的浪费和欺诈，而且它还不是 90%。效率不是 90%。这似乎是一种奇怪的第一性原理方式——政府欺诈的金额。就像你觉得有多少，然后——反正我们没法现场算，但我很好奇。

I mean zero. Okay. So then would you say that foreign waste that the government uh has is 90%. That also would be quite generous. But if if it's only 90% that means that there's $750 billion a year of waste and fraud and it's not 90%. It's not 90% effective. This seems like a strange way to first principle is the amount of fraud in the government. just like how much do you think there is and then uh uh I anyways we don't know how to do it live but I'd be curious like see

你对 Stripe 的欺诈很了解，人们不断试图欺诈。

you know a lot about fraud at Stripe people are constantly trying to do fraud.

对，但正如你所说，这有点像——我们已经把它压得很低了，但这是一个有点不同的问题空间，因为我们在这里处理的是比 Stripe 更异质的欺诈向量。

Yeah but as you say it's like a little bit of a um we've really ground it down but it's a little bit of a different problem space because we're dealing with a much more heterogeneous set of fraud vectors here than we are

对，但我的意思是，在 Stripe，你有很高的信心，并且很努力。你有高能力和高关心，但欺诈仍然不是零。现在想象一下，规模大得多，能力低得多，关心也少得多。

Yeah. But I mean I mean at tribe you you you you have high confidence and you try hard. Um you have high competence and high caring but still fraud is non non zero. Um now now now imagine it's at a much bigger scale. Um there's much less competence and much less carrying.

你知道，当年的 PayPal，我们努力把欺诈率控制在交易量的 1%左右。那非常困难，需要极大的能力和用心才能把欺诈降到仅仅 1%。现在想象一个缺乏用心和能力的组织，欺诈率会远高于 1%。

You know, PayPal back in the day, we tried to manage fraud down to about 1% of the payment volume. That was very difficult. It took a tremendous amount of competence and caring to get fraud merely to 1%. Now imagine an organization where there's much less caring and much less competence. It's going to be much more than 1%.

现在回顾政治和你在那里的作为，感觉如何？从外部看，有两件事影响很大：美国政治行动委员会和收购推特。但似乎也有很多心痛。你给整个经历打几分？

How do you feel now looking back on politics and doing stuff there? From the outside, two things seem impactful: the America PAC and the acquisition of Twitter. But it also seems like there was a bunch of heartache. What's your grading of the whole experience?

我认为那些事情必须做，以最大化未来美好的概率。但政治通常非常部落化。人们会失去客观性，很难看到对方的好或己方的坏。这是最让我惊讶的一点：你常常根本无法和属于某个部落的人讲道理。他们相信自己的部落做的一切都是好的，对方部落做的一切都是坏的。要说服他们几乎不可能。但总的来说，那些行动——收购推特、让特朗普当选——尽管让很多人愤怒，我认为对文明是有益的。

I think those things needed to be done to maximize the probability that the future is good. But politics is generally very tribal. People lose their objectivity with politics. They have trouble seeing the good on the other side or the bad in their own side. That was one of the things that surprised me most: you often simply cannot reason with people if they're in one tribe or the other. They believe everything their tribe does is good and anything the other tribe does is bad. Persuading them otherwise is almost impossible. But overall, those actions—acquiring Twitter, getting Trump elected, even though it makes a lot of people angry—I think they were good for civilization.

这如何与你期待的未来联系起来？

How does it feed into the future you're excited about?

美国需要足够强大，持续足够久，以便将生命扩展到其他星球，并将 AI 和机器人技术发展到能确保未来美好的程度。如果我们陷入共产主义或极端压迫的状态，我们可能无法成为多行星物种，国家可能会扼杀我们在 AI 和机器人方面的进步。

America needs to be strong enough to last long enough to extend life to other planets and to get AI and robotics to the point where we can ensure the future is good. If we were to descend into communism or an extremely oppressive state, we might not become multiplanetary, and the state might stamp out our progress in AI and robotics.

你如何看待 Optimus、Grok 等产品最终被政府利用？私营公司应该为政府提供什么，又该设置什么护栏？AI 模型是否应该无条件服从政府要求？如果军方想做某事，Grok 应该拒绝吗？

How do you feel about Optimus, Grok, etc. being leveraged by the government over time? What guardrails should private companies have regarding what they give governments? Should AI models do whatever the government asks? Should Grok refuse if the military wants to do something?

AI 和机器人出问题最大的危险可能来自政府。反对企业的人最应该担心的是政府，因为政府本质上就是一家公司——最大的、垄断暴力的公司。我觉得奇怪的是，人们认为企业坏而政府好，而政府恰恰是最大最差的公司。企业的道德比政府好。所以政府可能利用 AI 和机器人来压制民众。这是一个严重的担忧。

Probably the biggest danger of AI and robotics going wrong is government. People who are opposed to corporations should really worry most about government, because government is just a corporation in the limit—the biggest corporation with a monopoly on violence. I find it strange that people think corporations are bad but government is good, when government is simply the biggest and worst corporation. Corporations have better morality than government. So the government could potentially use AI and robotics to suppress the population. That is a serious concern.

作为建造 AI 和机器人的人，你如何防止这种情况？

As someone building AI and robotics, how do you prevent that?

如果你有一个有限政府——限制政府权力，这正是美国宪法的意图——那么结果可能比更多政府要好。

If you have a limited government—limiting the powers of government, which is what the US Constitution intends to do—then you'll probably have a better outcome than with more government.

但这项技术将对所有政府可用，对吧？

But this technology will be available to all governments, right?

是的，所有政府。很难预测终点或路径。如果文明进步，AI 将远超所有人类智能的总和，机器人数量将远超人类。沿途会发生什么很难预测。

Yes, all governments. It's difficult to predict the end point or the path. If civilization progresses, AI will vastly exceed the sum of all human intelligence, and there will be far more robots than humans. What happens along the way is very difficult to predict.

你可以做的一件事就是规定政府不得将 Optimus 用于某些用途，制定政策。你最近说 Grok 应该有一个道德宪法。其中一点可以是限制政府用这项技术做什么。

One thing you could do is just say governments are not allowed to use Optimus for certain things, write out a policy. You recently said Grok should have a moral constitution. One thing could be limiting what governments can do with this technology.

我们可以这样做。但技术上，如果政治家通过法律并执行，很难不遵守。最好的办法是有限政府，行政、司法和立法分支之间相互制衡。

We can do that. But technically, if politicians pass a law and enforce it, it's hard not to comply. The best thing is limited government with appropriate checks between executive, judicial, and legislative branches.

似乎限制将来自你。你有 Optimus、太空 GPU 等。SpaceX 已经如此，政府需要你发射关键卫星。你正在建造更多未来技术组件，将扮演类似角色。你可以制定政策，比如不帮助压制古典自由主义。

It seems like the limits will come from you. You have Optimus, space GPUs, etc. Already with SpaceX, the government needs you for crucial satellites. You're building more technological components that will have an analogous role. You could set policy, like not helping suppress classical liberalism.

我会尽最大努力确保我控制范围内的一切最大化人类的美好结果。任何其他做法都是短视的。我是人类的一部分。我喜欢人类。

I will do my best to ensure that anything within my control maximizes the good outcome for humanity. Anything else would be shortsighted. I'm part of humanity. I like humans.

支持人类，支持。

Pro-human, pro.

你提到 Dojo 3 将用于太空计算。你真的会看我说的吗？我不知道你是否了解 Twitter，但我知道你有很多粉丝。大爆料。你是怎么从我发的帖子里看出我的秘密的？你怎么为太空设计这款芯片？有什么变化？

You mentioned that Dojo 3 will be used for space-based compute. Do you really read what I say? I don't know if you know Twitter, but I know you have a lot of followers. Big giveaway. How do you discern my secrets that I post? How do you design this chip for space? What changes?

嗯，你要设计得更能抗辐射，并且在更高温度下运行。大致来说，如果工作温度（开尔文）提高 20%，散热器质量就能减半。在太空里高温运行是有利的。屏蔽内存有很多方法，但神经网络对比特翻转非常鲁棒。辐射的主要影响是随机比特翻转。对于数万亿参数的模型，少量翻转无关紧要。启发式程序对翻转比巨大的参数文件敏感得多。所以我只是让它跑得更热。除了温度更高，其他和地球上差不多。

Well, you want to design it to be more radiation tolerant and run at a higher temperature. Roughly, if you increase the operating temperature by 20% in degrees Kelvin, you can cut your radiator mass in half. Running at a higher temperature is helpful in space. There are various things you can do for shielding the memory, but neural nets are very resilient to bit flips. Most radiation effects are random bit flips. With a multi-trillion parameter model, a few flips don't matter. Heuristic programs are much more sensitive to flips than a giant parameter file. So I just designed it to run hot. You pretty much do it the same way as on Earth, apart from making it run hotter.

太阳能板是卫星上最重的部分。有没有办法让 GPU 比 Nvidia 和 TPU 计划中的更节能，特别是在太空场景？

The solar array is most of the weight on the satellite. Is there a way to make the GPUs even more power-dense than what Nvidia and TPUs are planning, especially for space?

基本数学是：如果每个掩模版能提供约 1 千瓦，那么要支持 100 吉瓦就需要 1 亿个完整掩模版芯片。取决于良率假设，这决定了你需要多少次发射。如果你想要 100 吉瓦的功率，就需要 1 亿个芯片，每个掩模版持续运行 1 千瓦。看看 Blackwell 这样的芯片的 die 尺寸，每个晶圆只能产出几十个甚至更少。所以这是一个每年每月生产数百万晶圆的世界。Terafab 的计划就是：每月数百万先进制程晶圆。这个数字肯定超过一百万。还需要内存。你会建内存厂吗？我认为 Terafab 必须涵盖内存、逻辑和封装。

The basic math is: if you can do about a kilowatt per reticle, then you'd need 100 million full reticle chips to do 100 gigawatts. Depending on your yield assumptions, that tells you how many trips you need. If you want 100 gigawatts of power, you need 100 million chips running a kilowatt sustained per reticle. If you look at the die size of something like Blackwell, you can get on the order of dozens or less per wafer. So you're looking at a world where if we're putting that out every year, you're producing millions of wafers a month. That's the plan with Terafab: millions of wafers a month of advanced process nodes. It's got to be some number north of a million. You also need memory. Are you going to make a memory fab? I think Terafab has to do memory, logic, memory, and packaging.

我很好奇一个人怎么开始。这是人类制造过的最复杂的东西。显然如果有人能胜任，那就是你。你意识到这是个瓶颈，然后去找你的工程师。你跟他们说什么？‘我要在 2030 年每月生产一百万晶圆。’你打电话给 ASML 吗？下一步是什么？

I'm very curious how someone gets started. This is the most complicated thing man has ever made. Obviously if anybody's up to the task, you are. You realize it's a bottleneck and you go to your engineers. What do you tell them? 'I want a million wafers a month in 2030.' Do you call ASML? What is the next step?

我们先建一个小型晶圆厂试试。在小规模上犯错误，然后再建大的。

We make a little fab and see what happens. Make our mistakes at a small scale and then make a big one.

那个小厂建好了吗？

Is the little fab done?

不，还没建好。这事瞒不住的。就像无人机盘旋在那玩意儿上空。你能在 X 上实时看到建设进度。所以没有，我们可能彻底失败，说清楚。成功没有保证。但我们想尝试制造大约 1 亿颗芯片。我们需要 100 吉瓦的电力，以及能承受 100 吉瓦的 1 亿颗芯片。到 2030 年，供应商给我们多少芯片我们就用多少。我对台积电、三星和镁光说过：‘请更快地建更多晶圆厂。我们保证购买产出。’他们已经在尽力加速了。这不是我们和他们对立的关系。

No, it's not done. People wouldn't keep that cat in the bag. That cat's going to come out of the bag. It'll be like drones hovering over the bloody thing. You'll be able to see its construction progress on X in real time. So no, we could just flounder in failure, to be clear. Success is not guaranteed. But we want to try to make something like 100 million chips. We need 100 gigawatts of power and 100 million chips that can take 100 gigawatts. By 2030, we'll take as many chips as our suppliers will give us. I've said this to TSMC, Samsung, and Micron: 'Please build your more fabs faster. We will guarantee to buy the output.' They're already moving as fast as they can. It's not like it's us versus them.

有种说法是搞 AI 的人想要大量芯片，但晶圆厂和涡轮机制造商等供应商没有快速扩产。解释是它们天生保守——比如台湾人或德国人。真是这样吗？

There's a narrative that AI people want a huge number of chips quickly, but input suppliers like fabs and turbine manufacturers aren't ramping up production quickly. The explanation is that they're dispositionally conservative—Taiwanese or German. Is that really the explanation?

这很合理。如果一个人在计算机内存行业干了三四十年，他们见过十次繁荣与萧条的周期。那是很多伤疤。繁荣时期一切看起来永远美好，然后崩盘来了，他们拼命避免破产。接着又是另一个繁荣和另一个崩盘。

It's reasonable. If somebody has been in the computer memory business for 30 or 40 years, they've seen cycles—boom and bust—10 times. That's a lot of scar tissue. During boom times, everything looks great forever, then the crash happens, and they desperately try to avoid bankruptcy. Then another boom and another crash.

有没有其他你认为别人应该去做的想法，而你自己出于某种原因没做？

Are there other ideas you think others should pursue that you're not, for whatever reason?

有几家公司在探索新的工作方式，但就是扩张不够快。我甚至不限于 AI 领域。总的来说，人们应该做自己非常有动力去做的事，而不是我建议的什么。他们应该做自己觉得有趣和有动力的事。

There are a few companies pursuing new ways of doing jobs, but they're just not scaling fast. I don't even mean within AI. Generally, people should do the thing they find highly motivating, as opposed to something I suggest. They should do what they personally find interesting and motivating.

回到限制因素——你说了大概 100 次。当前的限制因素是什么？

Going back to the limiting factor—you used that phrase about 100 times. What is the current limiting factor?

在三到四年的时间范围内，是芯片。在一年内，是能源——电力生产。我不确定是否有足够的可用电力来启动所有正在制造的 AI 芯片。

In the 3 to 4 year timeframe, it's chips. In the one-year timeframe, it's energy—power production, electricity. It's not clear to me that there's enough usable electricity to turn on all the AI chips being made.

嗯，到今年年底，我认为人们将真正遇到困难，芯片的产出将超过启动芯片的能力。

Um towards the end of this year, I think people are going to have real trouble turning on like the chip output will exceed the ability to turn chips on.

你打算如何应对那个世界？

What's your plan to deal with that world?

嗯，我们正在努力加速电力生产。我想这可能是 xAI 有望成为领导者，希望是领导者的原因之一，那就是我们能比其他人更快地启动更多芯片。因为我们擅长硬件，而且那些自称实验室的公司，它们的创新想法往往流动很快。很少看到想法在人与人之间传播有超过六个月的差距。所以我认为你遇到了硬件瓶颈，然后任何能最快扩展硬件的公司将成为领导者。所以我认为 xAI 将能最快扩展硬件，因此最有可能成为领导者。

Well, we're trying to accelerate electricity production. I guess that's maybe one of the reasons that xAI will be maybe the leader, hopefully the leader, is that we'll be able to turn on more chips than other people can turn on faster. Because we're good at hardware and generally the innovations from the corporations that call themselves labs, the ideas tend to flow. It's rare to see more than about a six-month difference between ideas traveling back and forth with the people. So I think you sort of hit the hardware wall and then whatever company can scale hardware the fastest will be the leader. And so I think xAI will be able to scale hardware the fastest and therefore most likely will be the leader.

你开玩笑或自嘲地又用了限制因素这个词，但我认为这其中有深意。回顾我们讨论的很多内容，这或许是个不错的结尾。比如，一个低能动性的公司，它会有某个瓶颈，但不会真正去解决它。马克和德雷说过：大多数人愿意忍受任何程度的慢性疼痛，以避免急性疼痛。而我们讨论的很多案例，似乎都是直面急性疼痛，不管是什么。就像，好吧，我们必须想办法处理钢铁，或者我们必须想办法在太空中运行芯片，或者我们愿意承受一些短期的急性疼痛来真正解决瓶颈。这算是一个统一的主题。

You joked or were self-conscious about using the limiting factor phrase again, but I actually think there's something deep here. If you look at a lot of the things we've touched on over the course, it's maybe a good note to end on. Like if you think of a low-agency company, it would have some bottleneck and not really be doing anything about it. Mark and Dre had the line: most people are willing to endure any amount of chronic pain to avoid acute pain. And it feels like a lot of the cases we're talking about are just leaning into the acute pain, whatever it is. It's like, okay, we got to figure out how to work with steel or we got to figure out how to run the chips in space, or like we'll take some near-term acute pain to actually solve the bottleneck. And so that's kind of a unifying theme.

我的痛阈很高。这很有帮助。

I have a high pain threshold. That's helpful.

解决瓶颈。

Solve the bottlenecks.

是的。

Yes.

所以我可以说的是，我认为未来会非常有趣。正如我所说，达沃斯只在地面上待了大约 3 小时。为了生活质量，站在乐观一边即使错了，也比站在悲观一边对了要好。所以如果你站在乐观一边，你会更快乐。因此我建议站在乐观一边。

So one thing I can say is that I think the future's going to be very interesting. And as I said, Davos only been on the ground for like 3 hours or something. It's better to be on the side of optimism and be wrong than on the side of pessimism and be right for quality of life. So your happiness will be higher if you are on the side of optimism rather than on the side of pessimism. And so I recommend being on the side of optimism.

就这么办。

Let's do that.

酷。Yan，谢谢你参加。

Cool. Yan, thanks for doing this.

谢谢。

Thank you.

好的。

All right.

好的。

All right.

哦，好耐力。

Oh, great stamina.

希望这次会面是耐受性上的痛苦。

Hopefully this encounter is a pain in the tolerance.

大家好，希望你们喜欢这一集。如果喜欢，最有帮助的事情就是分享给其他可能喜欢的人。如果你在收听的平台上留下评分或评论，也很有帮助。如果你有兴趣赞助播客，可以联系 dwarcash.com/advertise。否则，我们下期再见。

Hey everybody, I hope you enjoyed that episode. If you did, the most helpful thing you can do is just share it with other people who you think might enjoy it. It's also helpful if you leave a rating or a comment on whatever platform you're listening on. If you're interested in sponsoring the podcast, you can reach out at dwarcash.com/advertise. Otherwise, I'll see you on the next one.