本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
所以真的有三个小时的问题吗？还是你在开玩笑？
So are there really three hours of questions, or or has it are you fucking serious?
是的。
Yeah.
你不需要
You don't
说很多
need a lot
关于埃隆，是吗？
to talk about, Elon?
天哪，
Holy moot,
老兄。
man.
我
I
我的意思是，这是最有趣的点。
mean, it's the most interesting point.
现在所有的故事线都在逐渐汇聚。
All the storylines are kinda converging right now.
我们会看看
We'll see how
多少它
much It's
简直就像弗兰德斯。
almost like Flanders.
没错。
Exactly.
我们不会做这种事的。
We'll to That would never do such a thing.
所以，正如你比任何人都更清楚，数据中心的总拥有成本中，只有10%到15%是能源成本。
So as you know better than anybody else, the total cost of ownership of a data center, only 10% to 15% is energy.
而这就是你通过将这部分迁移到太空所 presumably 节省的部分。
And that's the part you're presumably saving by moving this into space.
大部分都是GPU。
Most of it's the GPUs.
如果它们在太空中，就更难维护，或者根本无法维护。
If they're in space, it's harder to service them or you can't service them.
因此，它们的折旧周期会缩短。
And so the depreciation cycle goes down on them.
所以，把GPU放在太空中显然要贵得多。
So it's just way more expensive to have the GPUs in space, presumably.
把它们放在太空中的原因是什么？
What's the reason to put them in space?
嗯，能源的可用性是个问题。
Well, the availability of energy is the issue.
如果你看看中国以外的电力产出，所有中国以外的地区，基本上都是持平的。
If you look at electrical output outside of China, everywhere outside of China, it's more or less flat.
它只是略有增加，基本上持平。
It's very, maybe a slight increase, pretty much flat.
中国的电力产出快速增长。
China has a rapid increase in electrical output.
但如果你把数据中心建在中国以外的任何地方，当你扩大规模时，你从哪里获取电力？
But if you're putting data centers anywhere except China, where are you going to get your electricity, especially as you scale?
芯片的产出几乎呈指数级增长，但电力的产出却保持平稳。
The output of chips is growing pretty much exponentially, but the output of electricity is flat.
那么，你如何给这些芯片供电呢？
So how are you going to turn the chips on?
魔法电源？
Magical power sources?
魔法电力精灵。
Magical electricity fairies.
你以大力支持太阳能而闻名，一太瓦的太阳能。
You're famously a big fan of solar, one terawatt of solar power.
以25%的容量系数计算，四太瓦的太阳能电池板仅占美国土地面积的1%。
So with a 25% capacity factor, like four terawatts of solar panels, it's like 1% of the land area of The United States.
而当我们拥有太瓦级的数据中心时，你是不是已经进入奇点了？
And that's like far in this, you were in the singularity when we've got one terawatt of data centers, right?
你到底在担心什么？
What are you running out exactly?
但你所说的奇点，真的那么遥远吗？
Far into the singularity are you though?
你来告诉我。
You tell me.
是的，正是如此。
Yeah, exactly.
所以我认为我们会发现，我们其实已经身处奇点之中，但即便如此，前方依然有很长的路要走。
So I think we'll find we're in the singularity and like, okay, we've still got a long way to go.
但在覆盖了内华达州的太阳能电池板之后，你的计划是把它们放到太空中吗？
But is the plan to put it in the space after we've covered Nevada and solar panels?
我觉得这相当困难
I think it's pretty hard
要覆盖内华达州和太阳能电池板。
to cover Nevada and solar panels.
得申请许可，得想办法拿到那些许可。
Have to get permits from try getting the permits for that.
所以太空其实是个监管问题。
So space is really a regulatory play.
在陆地上建造比在
It's harder to build on land than it
太空中更难。
is in space.
在地面上扩展比在太空中扩展更困难。
It's harder to scale on ground than it is to scale in space.
但此外，太空中的太阳能电池板的效率大约是地面上的五倍。
But also you're going to get about five times the effectiveness of solar panels in space versus the ground.
而且你不需要电池。
And you don't need batteries.
我差点穿了另一件衬衫，上面写着‘太空中永远阳光明媚’，确实如此。
I almost wore my other shirt which says it's always sunny in space, which it is.
因为在太空中没有昼夜循环、季节变化、云层或大气层，仅大气层就导致我们仍损失约30%的能量。
So because you don't have a daynight cycle or seasonality clouds or an atmosphere in space because the atmosphere alone we're still seeing about a 30% loss of energy.
任何一块太阳能电池板在太空中产生的电力都比在地面上多五倍。
Any given solar panel is going do about five times more power in space than on the ground.
而且你避免了为夜间供电而配备电池的成本。
And you avoid the cost of having batteries to carry you through the night.
所以实际上在太空中做这件事要便宜得多。
So it's actually much cheaper to do in space.
我的预测是，36个月内，甚至可能在30个月内，将AI部署在太空中将成为迄今为止最便宜的选择。
And my prediction is that it will be by far the cheapest place to put AI will be space in thirty six months or less, maybe thirty months.
36个月？
Thirty six months?
少于三十六个月。
Less than thirty six months.
在训练过程中，GPU经常损坏，你们如何维护它们？
How do you service GPUs as they fail, which happens quite often in training?
实际上，这取决于最近到达的GPU是哪一代的。
Actually, depends on how recent the GPUs are that arrived.
我的意思是，到目前为止，我们发现GPU相当可靠。
I mean, at this point, found our GPUs to be quite reliable.
存在早期故障问题，你可以在地面上解决这个问题。
There's infant mortality, which you can obviously iron out on the ground.
所以你可以在地面上运行它们，确认GPU没有早期故障。
So you can just run them on the ground and confirm that you don't have infant mortality with with the GPUs.
但一旦它们开始工作，一旦度过最初的调试周期——无论是英伟达、特斯拉AI六芯片、TPU、训练芯片或其他任何芯片——它们的可靠性实际上是非常高的。
But once they once they start working, their actual reliability and and once they start working and you're past the initial debug cycle of Nvidia or whatever, or whoever's making the chips, could be Tesla AI six chips or something like that, or it could be TPUs or traniels or whatever, rivalry is actually they're quite reliable past a certain point.
所以我认为维护问题并不是一个问题。
So I don't think I don't think that the servicing thing is an issue.
但你等着瞧吧。
But you can mark my words.
在三十六个月内，但很可能更接近三十个月，部署人工智能最具经济优势的地方将是太空。
In thirty six months but probably closer to thirty months the most economically compelling place to put AI will be space.
然后，身处太空的优势会变得极其明显。
And it will get from it'll then get like ridiculously better to be in space.
而扩展规模，真正能实现大规模扩展的地方只有太空。
And then the scaling, the only place you can really scale is space.
一旦你开始思考你利用了太阳多少比例的能量，你就会意识到必须去太空。
Start Once thinking in terms of what percentage of the sun's power are you harnessing, you realize you have to go to space.
在地球上很难实现大规模扩展。
You can't scale very much on earth.
但要
But to be
清楚地说，你指的是太瓦级别的功率。
clear, you're talking like terawatts.
是的
Yeah.
目前，整个美国平均仅使用半太瓦的电力。
Well all of The United States currently uses only half a terawatt of power on average.
如果你说一太瓦，那就是美国当前用电量的两倍。
If you say a terawatt that would be twice as much electricity as The United States currently consumes.
所以这相当可观。
So it's quite a lot.
你能想象建造这么多数据中心或这么多发电厂吗？
And can you imagine building that many data centers or that many power plants?
那些一直生活在软件世界里的人，并没有意识到他们即将面临硬件领域的严峻考验。
It's like those who have lived in software land don't realize that they're about to have a hard lesson in hardware.
实际上，建造发电厂是非常困难的。
That it's actually very difficult to build power plants.
而且你不仅仅需要建造发电厂。
And then you don't just need to put power plants.
你需要所有的电气设备、变压器来运行这些变压器，也就是AI所需的变压器。
You need all of the electrical equipment, electrical transformers to run the transformers, the AI transformers.
如今，公用事业行业是一个非常缓慢的行业。
Now the utility industry is a very slow industry.
他们严重阻碍了向政府和公共事业委员会的推进。
They much impede a smash to the government, to the public utility commission.
他们确实从很早以前就开始阻碍这种推进了。
They they impeded smash like literally very early.
所以他们非常缓慢，因为过去一向如此缓慢。
So they're very slow because their past has been very slow.
因此，想让他们快速行动，就像你试图与公用事业公司签订互联协议一样——你有没有试过在大规模电力需求下与公用事业公司达成协议？
So trying to get them to move fast is just like, know, like if you try to do an interconnect agreement with the have you ever tried to do an agreement with a utility at scale, like a lot of power?
作为一名专业播客主持人，我可以明确说，我实际上并不是。
As a professional podcaster, I can say that I am not in fact.
在成为问题之前，还需要更多的观看量。
Have do many more views before that becomes an issue.
他们必须进行为期一年的研究。
They have to do a study for a year.
明白吗？
Okay?
一年后，他们会带着他们的并网研究回来找你。
Like a year later, they'll come back to you with their interconnect study.
但你不能用你自己背后的电力数据来判断吗？
But can't you tell this with your own behind the meter power stuff?
你可以建造发电厂。
You can build power plants.
是的。
Yeah.
我们在XAI为第二类项目就是这么做的。
That's what we did at XAI for classes two.
针对第二类项目
For classes two
但所以，是的，为什么我们在讨论电网？
But so, yeah, why are talking about the grid?
为什么不直接建造GPU并实现电力就近供应呢？
Why not just build GPUs and power co located?
我们就是这么做的。
That's what we did.
对。
Right.
但我的意思是，当你谈到所有这些问题时，为什么这不是一个通用的解决方案？你从哪里获得电力？
But I'm saying, why isn't this a generalized solution when you're talking about all the issues Where do get
电力设施从哪里来？
the power plants from?
我的意思是，当你谈到与公用事业公司合作的所有问题时，你可以直接为数据中心建造私人发电厂。
I'm saying, when you talk about all the issues working with utilities, you can just build private power plants with the data centers.
对。
Right.
但这引出了一个问题：你从哪里获得发电厂？
But it begs the question of where do you get the power plants from?
意思是发电厂的制造商。
Mean The power plant makers.
哦，我明白你的意思了。
Oh, I see what you're saying.
比如，燃气轮机的积压订单吗？
Like, does the gas turbine backlog, basically?
是的。
Yes.
我们可以再深入一层来看。
Can drill it out a level further.
涡轮机中的叶片和导流片才是限制因素，因为铸造这些部件需要非常专业的工艺，假设你使用的是燃气发电。
It's veins and blades in the turbines that are the limiting factor because the casting may it's like a very specialized process to cast the blades and veins in the turbines, assuming you're using gas power.
而且其他形式的电力很难规模化。
And it's very difficult to scale other forms of power.
你可以扩大太阳能的规模，但目前美国进口太阳能产品的关税高得惊人。
You can scale potentially solar, but the tariffs currently for importing solar in The US are gigantic.
而国内的太阳能生产能力却微不足道。
And the domestic solar production is pitiful.
为什么不自己生产太阳能设备呢？
Why not make solar?
这看起来像是
That seems like
一个非常适合埃隆来解决的问题。
a good Elon shaped problem.
我们打算生产太阳能设备。
We are going to make solar.
好的。
Okay.
是的。
Yeah.
太好了。
Great.
SpaceX和特斯拉都在朝着每年100吉瓦的太阳能电池产能迈进。
Both SpaceX and Tesla are are building towards 100 gigawatts here of of solar cell production.
你们会深入到哪个环节？
How low down the stack?
比如从多晶硅开始，到硅片，再到最终的面板？
Like, from polysilicon up to the wafer to the final, panel?
我认为必须从原材料开始，一直做到电池的最终成品。
I think you got to do the whole thing from raw materials to to to finish the cell.
如果是要用于太空，实际上制造太空用的太阳能电池成本更低，因为它们不需要玻璃，或者只需要很少的玻璃，也不需要沉重的框架，毕竟不需要承受恶劣天气。
Now if it's going to space, it's actually it costs it costs less than 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.
太空中没有天气。
There's no weather in space.
所以，用于太空的太阳能电池实际上比地面上的更便宜。
So it's actually a cheaper solar cell that goes to space than than the one on the ground.
在未来的三十六个月内，有没有可能让它们变得和
Is there a path to getting them as cheap as
你所需的那样便宜？
you need in the next thirty six months?
太阳能电池已经非常便宜了。
Solar cells are already very cheap.
它们便宜得离谱。
They're like farcically cheap.
如果你说，我觉得中国的太阳能电池大约每瓦0.25到0.30美元左右。
It's and if you say you know, I I I think like solar cells in China are around like $0.25 $0.30 a watt or something like that.
这便宜得荒谬。
It's absurdly cheap.
而当考虑到将其用于太空时，它实际上便宜了五倍——不，不是五倍。
And when 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.
它便宜了十倍，因为你不需要任何电池。
It's 10 times cheaper because you don't need any batteries.
所以，一旦进入太空的成本变得极低，那么在太空中生成能源就远比其他方式更便宜、更可扩展。
So so the moment your cost of access to space becomes low, by far the cheapest and most scalable way to generate to to to generate tokens is space.
根本没法比。
It's not even close.
在扩展性上，太空方案比地面芯片方案容易一个数量级。
It'll be an order of magnitude easier to scale and chips aside of order of magnitude.
如果目标是无法在地面上扩展，那确实根本做不到。
Well, if the point is you won't be able to scale the ground, just won't.
人们在电力生成方面将面临巨大的瓶颈。
People are going to hit the wall big time on power generation.
他们已经遇到了。
They already are.
XAI团队为了实现一吉瓦电力的并网，不得不完成一系列不可思议的壮举。
So the number of miracles and series that the XAI team had to accomplish in order to get a gigawatt of power online crazy.
我不得不把一大堆涡轮机拼凑在一起，而且在田纳西州遇到了许可问题，最后不得不跨州到密西西比州，好在那边离这里只有几英里远。
I had to gang together a whole bunch of turbines and we had permit issues in Tennessee had to go across the border to Mississippi, which is fortunately only a few miles away.
但我们仍然需要铺设几英里的高压线路，并在密西西比州建造一座发电厂。
But then we still had to run the high power lines a few miles build a power plant in Mississippi.
建造这座电厂非常困难。
It was very difficult to build that.
人们并不理解，要为一个数据中心供电，发电端实际上需要多少电力。
And people don't understand how much electricity do you actually need at the generation level in order to power a data center.
因为媒体报道时，只会看一个GV300的功耗，然后乘以数量，就以为这就是所需电力总量。
Because the news will look at the power consumption of, say, a GV300 and multiply that by and then think that's the amount of power you need.
而忽略了冷却系统以及其他所有消耗。
While the cooling and everything.
醒醒吧。
Wake up.
是的。
Yeah.
这完全是外行。
That's total noob.
你以前从未接触过任何硬件。
You've never done any hardware in your life before.
除了GB 300，你还需要为所有的网络硬件供电。
Besides the GB 300, you've got to power all of the networking hardware.
还有很多CPU和存储设备在运行。
There's a whole bunch of CPU and storage stuff that's happening.
你需要根据峰值散热需求来设计规模。
You've got a size for your peak cooling requirements.
这意味着，即使在一年中最热的时段，你也要保持冷却吗？
So that means, you cool even on the worst hours, the worst day of the year?
孟菲斯的天气确实热得离谱。
Well, it is pretty freaking hot in Memphis.
所以，仅仅为了散热，你的电力消耗就会增加约40%。
So you're gonna have like a 40% increase on your power just for cooling.
前提是你不希望数据中心在炎热的日子里停机，而是想持续运行。
Assuming you don't want your data center to turn off on hot days and want to keep going.
那么你还得考虑，这上面还有一个乘数因素，那就是，你是否假设你的电力供应永远不会出现任何波动？
Then you've 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?
比如，实际上，我们有时不得不将发电机的部分电力停机以进行维护。
Like, oh, well, actually, sometimes we have to take the generators some of the power offline in order to service it.
哦，明白了。
Oh, okay.
现在你又得再增加20%到25%的余量，因为你必须假设需要停机维护电力系统。
Now you add another 20%, 25% multiplier on that because you've to assume that you've got to take power offline to service it.
因此，实际的RS估算下来，每大约110,000个GB（包含GB300、网络、CPU、存储、冷却以及维护电力的余量），大约需要300兆瓦。
So the actual RS, roughly every 110,000 GBs, GB300s inclusive of networking, CPU storage, cooling, margin for servicing power is roughly 300 megawatts.
抱歉，能再说一遍吗？
Sorry, say that again.
你可以这样理解：在发电层面，你实际需要大约330,000个GB300的电力容量，才能支撑330,000个GB300，包括所有相关的网络支持、冷却峰值以及预留的电力余量，这大约需要一吉瓦。
It's roughly think about it like the way you think about it is like 330,000 actually what you need at the generation level to probably service 330,000 GB300s, including all of the associated support networking and everything else, and the peak cooling, and to have some power margin reserve is roughly a gigawatt.
我可以问一个很幼稚的问题吗？
Can I ask a very naive question?
你正在描述在地球上实施这些事情的工程细节。
You're describing the engineering details of doing this stuff on Earth.
但在太空中做同样的事情也有类似的工程难题。
But then there's analogous engineering difficulties of doing it in space.
如何用轨道激光替代无限带宽，等等等等？
How do replace infinite band with orbital lasers, etcetera, etcetera?
你如何让系统抵抗辐射？
How do you make it resistant to radiation?
我不了解具体的工程细节，但从根本上说，这些之前必须解决的挑战，真的会比在地球上建造更多涡轮机更容易吗？
I don't know the details of the engineering, but fundamentally, is the reason to think those challenges, which had to be addressed before will end up being easier than just like building more turbines on earth?
有一些公司正在地球上制造涡轮机。
There's companies that build turbines on earth.
它们可以制造更多的涡轮机，对吧？
They can make more turbines, right?
我再次邀请一下，
I invite again,
试试看，你就会明白了。
try doing it and then you'll see.
所以涡轮机到2030年都卖断货了。
So like the turbines are sold out through 2030.
你们有没有考虑过自己制造？
Have you guys considered making your own?
我认为，为了提供足够的电力，SpaceX和特斯拉很可能需要自己生产涡轮叶片和带材。
I think in order to bring enough power online I think SpaceX and and Tesla will probably have to make the turbine blades the bands and blades
内部生产。
internally.
但只是叶片，还是整个涡轮机？
But just the blades or the turbines?
限制因素在于，其他所有部件都能买到，唯独叶片和叶脉不行。
The limiting factor, you can get everything except the blades, what they call the blades and veins.
叶片和叶脉的交付时间要比其他部件晚十二到十八个月。
You can get that twelve to eighteen months before the veins of blades.
叶片和叶脉的限制因素。
The limiting factor of the veins of blades.
全球只有三家铸造公司可能生产这些部件，而且它们的订单已经排满。
And there are only three casting companies in the world that may make these and they're massively backlogged.
是西门子、通用电气这些公司吗，还是分包商？
Is this Siemens, GE, those guys or is it a subcontractor?
不，是其他公司。
No, it's It's other companies.
我的意思是，有时候它们内部有点铸造能力，但我只是说，你可以随便打电话给任何一家涡轮机制造商，他们会告诉你的。
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.
这并不是机密。
It's not top secret.
现在可能在网上就能找到。
It's probably on the Internet right now.
如果没有关税，Colossus会是太阳能供电的吗？
If it wasn't for the tariffs, would Colossus be solar powered?
让它用太阳能供电会容易得多，是的。
Would be much easier to make it solar powered, yeah.
这些关税简直疯了，高达百分之几百。
The tariffs are nuts, so several 100%.
你认识一些人吗？
Don't you know some people?
我们还需要速度。
We'll also need speed.
不，不是的。
Yeah, no.
你知道，总统并不是在所有事情上都和我们意见一致。
Know, president has we don't agree on everything.
这个政府并不特别支持太阳能。
And this administration is not the biggest fan of solar.
你知道，我们还需要土地、许可等等。
You know We also need the land, the permits, and everything.
如果你试图快速推进，我认为在地球上大规模发展太阳能是一个不错的选择。
If you try to move very fast, I do think scaling solar on earth is a good way to go.
但你确实需要一些时间来寻找土地、获取许可、采购太阳能设备，并将其与电池配套。
But but you need you do 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发电系统呢？
But why would it not work to stand up your own solar production?
你说得对，最终土地会用完，但德克萨斯州这里有很多土地。
And then you're right that you eventually run out of land, but there's a lot of land here in Texas.
内华达州也有很多土地，包括私人土地。
There's a lot of land in Nevada, including private land.
并不是所有土地都是公共所有的。
It's not all publicly owned land.
因此，你至少能够为下一个巨型项目以及之后的项目获得土地。
And so you'd be able to at least get the next Colossus and, like, the next one after that.
到了某个阶段，
And at a certain point,
你会遇到瓶颈，
you hit a wall,
但那暂时不行吗？
but wouldn't that work for the moment?
正如我所说，我们正在扩大太阳能的生产。
As I said, we are scaling solar production.
太阳能电池的物理生产是有一定速度限制的。
There's there's a rate there's a rate at which you can scale physical production of solar solar cells.
我们正在尽最大努力加快国内生产的扩张。
We are we're going as fast as possible in scaling domestic production.
你们在特斯拉生产太阳能电池吗？
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
Both Tesla and SpaceX have a mandate to get to a 100 gigawatts a
太阳能产能。
year of of solar.
说到年产能，我很好奇。
Speaking of the annual capacity, I'm curious.
五年后，假设如此，地球上的太阳能装机容量会是多少？
In five years' time, let's say, what will the installed capacity be on Earth
五年时间很长。
Five years is a long time.
那太空呢？
And in space?
我特意选五年，因为那时已经过了你们启动并运行的门槛。
I deliberately pick five years because it's after your once we're up and running threshold.
所以五年后，是的，地球和太空的累计AI装机容量分别是多少？
And so in five years' time, yeah, what's the on Earth versus in space installed AI capacity?
五年后，我认为，如果你说五年后，太空中的AI每年发射的总量，将超过地球上所有AI的累计总和。也就是说，五年后，我的预测是，我们每年在太空中发射并运行的AI总量，将超过地球上迄今为止的总和。
Five years, I think probably if you say that five 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 five years from now my prediction is we will launch and be operating every year more AI in space than the cumulative total on earth.
这是什么意思？
Which is?
我认为五年后，太空中每年的AI算力至少将达到几百吉瓦，并且还在增长。
I would expect to be at least sort of five years from now a few 100 gigawatts per year of AI in space and rising.
所以我认为，在地球上，每年的AI算力可以达到约一太瓦，之后就会开始面临火箭燃料供应的挑战。
So you can get to I think on Earth you can get to around a terawatt a year of AI in space before you start having, fuel supply challenges for the rocket.
好的，但你
Okay, but you
觉得你可以
think you can
五年内实现每年几百吉瓦吗？
get hundreds of gigawatts per year in five years'
是的。
Yes.
所以一百吉瓦，取决于
So 100 gigawatts, depending
于
on
整个系统——包括太阳能电池板和散热器等——的比功率大约相当于一万次星舰发射。
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.
而你希望在一年内完成这一切。
And you want to do that in one year.
这就相当于每小时发射一次星舰。
And so that's like one Starship launch every hour.
对。
Yeah.
这件事就在这座城市里发生。
That's happening in this city.
想象一下，每小时都有一次星舰发射的世界，给我描述一下那样的情景。
Like, walk me through a world where there's 10 there's a Starship launch every single hour.
是的。
Yeah.
我的意思是，相比航空公司，比如飞机航班，这个频率其实算低的。
I mean, that's actually a lower rate compared to airlines, like like aircraft aircraft.
有很多
There's a
机场。
lot of airports.
而且你还得发射到极地轨道。
And you've got to launch the polar orbit.
不，不一定
No, doesn't have to
是极地轨道。
be polar.
但太阳同步轨道确实有其价值。
But just there's some value to sun synchronous.
我认为实际上只要高度足够，就能脱离地球阴影区。
I think actually if just go high enough you start getting out of earth's shadow.
每年完成一万次发射需要多少艘实际的星舰？
How many physical starships are needed to do 10,000 launches a year?
我认为我们不需要超过二十到三十艘左右，我的意思是，可能二十到三十艘就够了。
I don't think we'll need more than, I mean you could probably do it with as few as like 20 or 30.
这真的取决于飞船绕地球飞行并回到发射台上方所需的时间，以及地面轨迹的间隔。
Really depends on how quickly does the ship the ship has to go around the earth and the ground track before the ship has to come back over the launch pad.
所以如果你能每三十小时使用一艘飞船，那么三十艘飞船就够了。
So if you can use a ship every say thirty hours, you could do it with 30 ships.
但我们一定会制造比这更多的飞船。
But but we'll we'll make more ships than that.
但SpaceX正在为每年实现一万次发射做准备，也许甚至达到两万或三万次发射。
But but the the SpaceX is is gearing up to do 10,000 launches a year and I'll and and maybe even 20 or 30,000 launches a year.
这个想法是成为类似超大规模云服务商的存在，像甲骨文那样，把这种运力租给其他人吗？
Is the idea to become basically a hyperscaler, become an oracle and lend this capacity to other people?
你打算怎么处理？ presumably SpaceX是所有这些发射的执行方。
What what's are you gonna do with Presumably SpaceX is the one launching all this.
所以SpaceX要成为超大规模云服务商？
So SpaceX become a hyperscaler?
超大规模，超大规模。
Hyper hyper.
是的。
Yeah.
我的意思是，如果我的预测成真，SpaceX将发射的AI量将超过地球上所有其他事物的总和。
I mean, if assuming my predictions come true, SpaceX will launch more AI than the cumulative amount on Earth combined of everything else combined.
这主要是推理吗？
Is this mostly inference?
大部分AI将是推理。
Mostly AI will be inference.
比如，用于训练的推理才是最主要的训练方式。
Like, inference for the purpose of training is most training.
有一种说法认为，关于SpaceX上市的讨论发生变化，是因为过去SpaceX非常资本高效，研发成本并没有那么高。
And there's a narrative that change in discussion around the SpaceX IPO is because previously, SpaceX was very capital efficient, just it wasn't that expensive to develop.
尽管听起来成本高昂，但其运营实际上非常资本高效。
And even though it sounds expensive, it's actually very capital efficient in how it runs.
而如今，你需要的资金将超过私人市场所能筹集的范围。
Whereas now you're going to need more capital than just can be raised in the private markets.
比如，如果私人市场能够容纳像AI实验室那样高达数百亿美元的融资，但无法再更高，那是不是意味着你每年需要的资金将超过数百亿美元，因此我才说要上市？
Like, if the private markets can accommodate raises of, as we've seen from the AI labs, tens of billions of dollars, but not beyond that, is it that you'll just need more than tens of billions of dollars per year and that's why I'd say it public?
是的。
Yeah.
我对谈论可能上市的公司要谨慎一些。
I'd be careful about saying things about companies that might go public.
你知道的？
You know?
如果你做出一般性陈述
If you make general statement
这对你来说从来不是问题，埃隆。
That's never been a problem for you, Elon.
你知道吗，做这些事情是要付出代价的。
You know, there's a price to pay for these things.
请为我们概括一下公开市场和私募市场在资本深度上的差异。
Make some general statements for us about the depths of the capital markets between public and private markets.
是的。
Yeah.
在更广泛的层面上，资本要多得多。
There's a lot more capital in the Very general.
显然，FOLAG市场的资本远比私募市场多。
There's obviously a lot more capital available in the FOLAG markets than private.
我的意思是，至少可能多出一百倍，但至少你知道吧？
I mean, it might it's at least at least it might be a 100 times more capital, but it's at least you know?
是的。
Yeah.
但远不止十倍。
But way more than 10.
但难道不是这样吗？那些资本密集型的领域，比如房地产——作为一个每年筹集大量资金的庞大行业，在行业层面上，通常依赖债务融资，因为当你投入如此巨额资金时，你实际上已经拥有相当稳定的
But isn't it also the case that things that tend to be very capital intensive, if you look at, say, real estate as, you know, a huge industry that raise a lot of money each year is at an industry level, that tends to be debt financed because by the time you're deploying that much money, you actually have a pretty
你有明确的收入流。
You have a clear revenue stream.
没错。
Exactly.
而且有短期回报。
And a near term return.
你甚至可以看到数据中心的建设，这些项目众所周知是由私人信贷行业提供融资的。
And you see this even with the data center build outs, which are famously being, you know, financed by the private credit industry.
那为什么不直接用债务融资呢？
And so why not just debt finance?
速度很重要。
Speed is important.
所以我通常会选择我能够反复突破的限制性因素。
So I'm generally gonna do the thing that I'm I'm I mean I just repeatedly tackle limiting factor.
无论限制速度的因素是什么，我都会去解决它。
Whatever the limiting factor is on speed I'm gonna tackle that.
所以，如果资本是唯一的因素，那我就解决资本问题。
So there's, if capital is the only factor then I'll solve for capital.
如果它不是限制因素，我就解决其他问题。
If it's not limiting factor I'll solve for something else.
根据你关于特斯拉和上市的言论，我本来不会猜到你觉得快速前进的方式是上市。
Based on your statements about Tesla and being public, I wouldn't have guessed that you thought the fast the way to move fast is to be public.
通常我会
Normally I
我会说这是对的。
would say that's true.
正如我所说，我本想更详细地谈谈这个问题，但问题是，一旦谈论上市公司是否上市，就会陷入麻烦。
Like I said, mean I'd like to talk about this in more detail, but the problem is if you talk about public companies, whether they become public, you get into trouble.
然后你就得推迟你的上市计划。
And then you have to delay your offering.
然后，正如你所说，我们正在追求速度。
Then you And as you said, we're solving for speed.
是的，完全正确。
Yes, exactly.
所以你不能炒作那些可能上市的公司。
So you can't hype companies that might go public.
因此我们必须在这里稍微谨慎一点。
So that's why we have to be a little careful here.
但我们不能谈论物理学。
But we can't talk about physics.
所以，你对长期规模化的看法是，地球只接收到太阳能量的约十亿分之一。
So like the way you think about scaling long term is that earth only receives about half a billionth of the sun's energy.
而太阳几乎提供了所有的能量。
And the sun essentially all the energy.
这是一个非常重要的观点，因为有时人们会谈论边际核反应堆或地球上各种形式的核聚变。
This is a very important point to appreciate because sometimes people will talk about marginal nuclear reactors or any various like fusion on earth.
但你得退一步想想，如果你要攀登卡达舍夫等级，利用相当可观比例的太阳能量的话。
But you have to step back a second and say if you're going to climb the Kardashev scale and have some nontrivial and harness some nontrivial percentage of the sun's energy.
比如说，你想利用太阳能量的百万分之一，这听起来似乎很小。
Like let's say you wanted to harness a millionth of the sun's energy which sounds pretty small.
但这将大约是目前地球上全人类发电量的十万倍左右，数量级上如此。
That would be about, call it roughly, 100,000 times more electricity than we currently generate on earth for all of civilization, give or take in order of magnitude.
所以显然，唯一能实现规模扩大的方式就是将太阳能带到太空中。
So obviously the only way to scale is to go to space with solar.
从地球发射，每年最多只能达到约一太瓦。
Launching from earth you can get to about a terawatt per year.
超过这个量级，你就需要从月球发射了。
Beyond that you want to go to you want to launch from the moon.
你希望在月球上建造一个质量驱动器。
You want to have a mass driver on the moon.
而在月球上的这个质量驱动器，每年可能实现大约一拍瓦。
And that mass driver on the moon, you could do probably a petawatt per year.
当你谈论这些数量级的计算能力，比如太瓦级别时，无论是在地球上还是在太空中，早在达到这个阶段之前，你就已经需要更多、更高效的太阳能电池板，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, far before this point, you've run into, actually need, maybe the solar panels are more efficient, but you still need the chips.
你仍然需要逻辑电路和内存等等。
You still need the logic and the memory and so forth.
你需要更多的芯片，并且让它们便宜得多。
You're gonna need, well, a lot more chips and make them much cheaper.
对。
Right.
那么，我们如何在2030年前实现太瓦级别的计算能力呢？目前全球的计算能力大约只有二十到二十五吉瓦。
And so how are we getting a terawatt of, like right now the world has maybe twenty, twenty five gigawatts of compute.
我们如何在2030年前实现太瓦级别的逻辑计算能力？
How are we getting a terawatt of logic by 2030?
我想我们需要建造非常大型的芯片工厂。
I guess we're gonna need some very big chip fabs.
别提了。
Tell me about it.
展开剩余字幕（还有 480 条）
我曾公开提到过，搞一个类似关税应用的想法，这里的‘关税’是新的‘吉格’。
I've mentioned publicly that the idea of doing a sort of a tariff app, tariff being the new gigger.
等等，等等。
Wait, wait.
我觉得特斯拉的命名方式非常吸引人，就像你在看度量尺度一样。
I feel like the naming scheme of Tesla, which has been very catchy, is like you looking at like the metric scale.
你在技术栈的哪个层级？
At what level of the stack are you?
你是要自己建无尘室，然后与现有的晶圆厂合作获取工艺技术，并购买
Are you building the clean room and then partnering with an existing fab to get the process technology and buying
它们的设备。
the tools from them.
那你的计划是什么？
What is the plan there?
你无法与现有的晶圆厂合作，因为它们的产能根本不够。
Well you can't partner with existing fabs because they can't output enough.
芯片产量太低了。
The chip volume is too low.
但就工艺技术而言。
But for the process technology.
是的，与IP方合作。
Yeah partner for the IP.
如今的晶圆厂基本上都在使用来自五家公司的设备。
The fabs today all basically use machines from like five companies.
所以你有ASML、东京电子、凯利、坦科尔等等。
So you've got ASML, Tokilectron, Kelly, Tankor, etcetera.
因此，起初我认为你必须从他们那里获取设备，然后对其进行修改或与他们合作以提高产量。
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.
但我认为你可能需要以不同的方式来建造。
But I think you'd have to build perhaps in a different way.
所以我认为合乎逻辑的做法是用常规设备以非常规的方式实现规模化，然后再开始改造设备以提高生产速率。
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.
有点无聊的公司风格。
Kind of boring company style.
是的。
Yeah.
就像是，你先买一台现成的无聊机器，然后琢磨怎么一开始挖隧道，再设计一台好得多的机器，那玩意儿，你知道的，快上好几个数量级。
Kind of like 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, you know, I don't know, some orders of magnitude faster.
这是一个非常简单的镜头。
Here's a very simple lens.
我们可以对技术进行分类，并评估它们的难度，一种分类方式是看看中国尚未成功实现的技术。
We can categorize technologies and how hard they are, and one categorization could be look at things that China has not succeeded in doing.
如果你看看中国的制造业，仍然在先进芯片和先进涡轮发动机等领域落后。
And if you look at Chinese manufacturing, still behind on leading edge chips and still behind on leading edge turbine engines and things like that.
因此，中国未能成功复制台积电，这是否让你对其中的难度有所顾虑，还是你觉得
And so does the fact that China has not successfully replicated TSMC give you any pause about the difficulty, or you think
嗯，这其实并不成立？
well, that's not true for some reason?
并不是他们没有复制台积电。
It's not that they have not replicated TSMC.
他们没有复制ASML。
They have not replicated ASML.
这才是限制因素。
That's the limiting factor.
所以你是觉得本质上就是制裁造成的吗？
So so you think it's just the the sanctions essentially?
是的。
Yeah.
中国将会生产大量的芯片，如果
China will be outputting vast numbers of of chips at
如果他们能买到的话。
If they could buy it.
达到三纳米级别。
To a three nanometer.
但难道他们直到最近都不能购买这些设备吗？
But couldn't they up to relatively recently buy them?
不行。
No.
好的。
Okay.
ASML的禁令已经实施一段时间了。
The the the ASML ban has been in place for a while.
好的。
Okay.
但我认为中国在三到四年内将会开始制造出非常有竞争力的芯片。
So but I I think China's gonna be they make start making pretty compelling chips in three or four years.
你会考虑制造ASML的设备吗？
Would you consider making the ASML machines?
我不知道。
I don't I don't know.
我不知道，这是正确的答案。
I don't know yet is the right answer.
所以，要实现大规模生产，并在大约三十六个月内达到高产量，以匹配火箭的有效载荷入轨能力。
So I it's just that to produce at a high volume and to reach large volume in say thirty six months to match the rocket payload to orbit.
如果我们能在三四年左右实现每年向轨道运送百万吨级的运输量，同时每吨需要100千瓦的功率，那就意味着我们每年至少需要100吉瓦的太阳能，同时还需要等量的芯片来支持——你需要100吉瓦当量的芯片。
So if we're doing a million tons to orbit in like let's say, I don't know, three or four years from now, something like that, that and we're doing 100 kilowatts ton so that means we need, at least 100 gigawatts per year of solar and we'll need an equivalent amount of chips to you that you need 100 gigawatts worth of chips.
必须让这些要素——轨道运输、电力生成和芯片——相互匹配。
Got to match these things, the master orbit, power generation and the chips.
而我最根本的担忧其实是内存。
And I'd say my base concern actually is memory.
我认为，制造逻辑芯片的路径比获得足够内存来支持逻辑芯片的路径要更清晰。
So I think there's path to creating logic chips is more obvious than the path to having sufficient memory to support logic chips.
这就是为什么DDR价格会飙升，以及网上流传着你被困在荒岛上的段子。
That's why you see DDR prices going ballistic and these memes about like you you're marooned on a desert island.
你在沙滩上写下‘救救我’。
You write help me on the sand.
他会来的。
There would he comes.
他写的是DDR。
He write DDRM.
飞船蜂拥而至。
Ships come swarming in.
我还没见过那个。
I haven't seen that.
我喜欢你关于晶圆厂的制造理念。
Love your manufacturing philosophy around fabs.
我对这个话题一无所知。
I know nothing about the topic.
我还不知道怎么建造晶圆厂。
I don't know how to build a fab yet.
我会弄明白的。
I'll figure it out.
显然，我必须建造
Obviously, I have to build
听起来你觉得工艺技术是这样的，比如台湾那上万名博士，他们确切知道该向等离子体腔注入什么气体，以及设备该设置什么参数。
a It sounds like you think the process technology, like these 10,000 PhDs in Taiwan who know exactly what gas goes in the plasma chamber and what settings to put on the tool.
你可以直接删除这些步骤中的那些部分。
You can just delete those parts of those steps.
从根本上说，就是弄到洁净室，弄到设备，然后自己摸索出来。
Fundamentally, it's get the clean room, get the tools, and figure it out.
我不认为
I don't think
是博士。
it's PhDs.
主要是没有博士学位的人。
It's mostly people with not PhDs.
大部分工程工作都是由没有博士学位的人完成的。
Most of engineering is done with people who don't have PhDs.
你们有博士学位吗？
Do you guys have PhDs?
没有。
No.
好吧。
Okay.
我们也没有成功建造过任何晶圆厂，所以你们不该来找我们寻求晶圆厂方面的建议。
We also haven't successfully built any FAB, so you shouldn't be coming to us for your FAB advice.
首先，我认为做这件事不需要博士学位。
I don't think you need PhD for that, first off.
但你们确实需要称职的人员。
But you do need competent personnel.
所以我不确定。
So I don't know.
我的意思是，现在如果特斯拉全力以赴，以最快速度推进AI5芯片的设计并实现量产。
Mean, right now, if say Tesla's pedals to metal max production going of as fast as possible to get AI5, Tesla AI5 chip design into production and then reaching scale.
你知道这可能会在第二季度发生，希望如此。
You know that'll probably happen you know around the second quarter of year, hopefully.
然后AI6有望在不到一年后跟进。
And then AI6 would hopefully follow less than a year later.
而且我们已经锁定了所有能拿到的晶圆厂产能。
But and and and we've secured all the all the trip fab production that we can.
是的。
Yes.
但你们目前在台积电的晶圆厂产能上受限吗？
But you're currently limited on TSMC fab capacity?
对。
Yeah.
我们会使用台积电台湾、三星韩国、台积电亚利桑那州和三星德克萨斯州的产能，而且我们已经预订了
And and and we'll be using TSMC Taiwan, Samsung Korea, TSMC Arizona, Samsung Texas and we still booked
你能订到的所有产能。
out all the capacity you can.
是的，如果我问台积电或三星，现在要达到量产需要多长时间？你得先建厂，然后开始生产，接着提升良率，最终达到高良率的量产阶段。
Yes and then if I ask TSMC or Samsung okay what what's the time frame to get to volume production at this point it's not you've got to build the fab and you've got to start production then you've got to climb the yield curve and reach volume production at high yield.
从开始到完成需要五年时间。
From start to finish is a five year period.
所以限制因素是芯片。
And so the limiting factor is chips.
一旦你能获得空间，限制因素就是芯片。
Limiting factor once you can get to space is chips.
在你能获得空间之前，限制因素是电力。
The limiting factor before you can get to space will be power.
你为什么不
Why don't you do
做詹森那件事，直接预付给台积电，让他们为你多建几座工厂？
the Jensen thing and just prepay TSMC to build more fabs for you?
我已经跟他们提过了。
I've already told them that.
但他们不愿意收你的钱吗？
But they won't take your money?
到底发生了什么？
Like what's going on?
他们正在尽可能快地建厂，不，不是的。
They're building fabs as fast no, no.
他们和三星都在尽最大努力建厂。
They're building fabs as fast as they can and so is Samsung.
他们已经全速运转了。
They're pedal to the metal.
我的意思是，他们正全力以赴，以最快的速度推进。
I mean they're going balls to wall as fast as they can.
但还是不够快。
So still not fast enough.
我的意思是，正如我所说，我认为到今年年底，芯片产量可能会超过开启芯片的能力。
Mean like I said, there will be I think if you say I think towards the end of this year I think probably chip production will outpace the ability to turn chips on.
但一旦你能进入太空并突破功率限制，就可以在太空中实现每年数百吉瓦的电力。
But once you can get to space and unlock the power constraint And you can now do hundreds of gigawatts per year of power in space.
请记住，美国的平均电力消耗为500吉瓦。
Again, bearing in mind that average power usage in The US is 500 gigawatts.
所以，如果你每年向太空发射200吉瓦的电力，差不多每两年半就能超过美国的总用电量。
So if you're launching, say, 200 gigawatts a year to space, you're sort of lapping The US every two and a half years.
整个美国的电力生产，这真是一个巨大的数字。
The entire, all US electricity production, this is a very huge amount.
但在那之前，服务器端计算、集中式计算的瓶颈将是电力。
But between now and then, constraint for server side compute, concentrated compute, will electricity.
我猜测，到今年年底，人们将开始面临无法为大型集群中的芯片供电的问题。
My guess is that people start getting to point where they can't turn the chips on for for for large clusters towards the end of this year.
芯片会不断积压，却无法被通电运行。
They're just the chips are gonna be piling up and and not be won't be able to be turned on.
至于边缘计算，那就是另一回事了。
Now for edge compute, it's a different story.
所以，比如特斯拉的AI第五代芯片将被用于我们的Optimus机器人。
So if the if like for for Tesla, the the so the AI five chip is going into our Optimus robot.
你知道的。
You know?
乐观的。
Optimistic.
如果你拥有分布式的人工智能边缘计算，那就是分布式供电。
And and so if you have an AI edge compute, that's distributed power.
现在，电力是分布在广大区域的。
Now the power is distributed over a large area.
它不是集中的。
It's not concentrated.
如果你能在夜间充电，实际上可以更有效地利用电网。
And if you can charge at night you can actually use the grid much more effectively.
因为美国的实际峰值电力生产超过一千吉瓦。
Because the actual peak power production in The US is over a thousand gigawatts.
但由于昼夜循环，平均电力消耗是500吉瓦。
But the average power usage because the day night cycle is 500.
所以如果你能在夜间充电，夜间就能额外产生500吉瓦的电力。
So if you can charge at night, there's an incremental 500 gigawatts that you can generate at night.
这就是为什么特斯拉在边缘计算方面没有受到限制。
So that's why Tesla for edge compute is not constrained.
我们可以制造大量芯片，从而生产大量的机器人和汽车。
And we can make a lot of chips to make a very large number of robots and cars.
但如果你试图集中这些计算能力，启动时会遇到很大困难。
But if you try to concentrate that compute, you're going to have a lot of trouble turning it on.
我觉得SpaceX业务最了不起的地方在于，其最终目标是抵达火星，但在前往火星的过程中，你总能找到方法持续产生增量收入，以推进到下一个阶段再下一个阶段。
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.
猎鹰九号和星链就是这样。
Falcon nine is Starlink.
而对于星舰，未来可能成为轨道数据中心。
And now for Starship, it's going to be potentially orbital data centers.
但你有没有发现，你的下一枚火箭、再下一枚火箭以及后续的规模扩展，都具有无限弹性的小规模应用场景？
But do you find these infinitely elastic marginal use cases of your next rocket and your next rocket and next scale up?
你能看出这看起来可能像一场模拟。
You can see how this might seem like a simulation.
或者我只是某个电子游戏里的人物化身？
Or am I someone's avatar in a video game or something?
因为你想，所有这些疯狂的事情同时发生，概率有多大？
Because it's like, what are the odds that all these crazy things should be happening?
火箭、芯片、机器人、太空太阳能。
Mean, rockets and chips and robots and space solar power.
更别提月球上的电磁弹射器了。
And not to mention the mass driver on the moon.
我真的很想看到它。
I really want to see that.
你可以想象一下，某个电磁弹射器只是发出‘嗖嗖’的声音。
You can imagine like some mass driver that's just go like, shroom, shroom.
就像是不断将太阳能驱动的AI卫星送入太空，一颗接一颗，速度达到每秒两公里半。
It's like sending AI, solar powered AI satellites into space like one after another, like these, like at two and a half kilometers per second.
那将是把它们射向深空的壮观景象。
That's a and just shooting them into deep space, that would be a sight to see.
实际上，我的意思是，我会看这个。
Actually, I mean, I'd watch that.
就像直播一个，是的，是的。
So like a live stream of Yeah, yeah.
一颗接一颗地发射出去
Just one after another, just shooting
网络摄像头。
The webcam.
在深空中，AI卫星每年达到十亿吨甚至百亿吨。
AI satellites in deep space you know, a billion or 10,000,000,000 tons a year.
抱歉。
I'm sorry.
你们在月球上制造卫星吗？
You manufacture the satellites on the moon?
是的。
Yeah.
我明白了。
I see.
所以你们把原材料送到月球，然后在那里制造，接着再
So you send the raw materials to the moon and then manufacture them there and then
你的月球土壤，我想，大约含有20%的太阳能材料，20%的硅之类的。
Well, your lunar soil is, I guess, like 20% solar 20% silicon or something like that.
所以你可以从月球上开采硅，提炼它，并在月球上制造太阳能电池板、太阳能电池和散热器。
So you can get the silicon from the you can mine the silicon on the moon, refine it, and generate the and create the solar panels the solar cells and the radiators on the moon.
是的。
Yeah.
所以你知道，用铝来制造散热器。
So you know, make the radiators out of aluminum.
所以月球上有充足的硅和铝来制造电池和散热器。
So there's plenty of silicon and aluminum on the moon to make cells on the radiators.
芯片你可以从地球运送，因为它们很轻。
The chips you could send from Earth because they're pretty light.
但也许在某个时候，你也会在月球上制造它们。
But maybe at some point you make them on the moon too.
我只是说，正如我所说，这看起来有点像电子游戏的情景——虽然困难，但并非不可能达到下一关。
I'm just saying these are simply kind of 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.
我看不出有任何办法能从地球每年发射500到1000太瓦的能量。
I don't see any way that you could do 500 to 1,000 terawatts per year launch from Earth.
我同意。
I agree.
但你可以从月球上做到这一点。
But you could do that from the moon.
好吧，让我告诉你我为什么会用水星来做个人银行服务。
Okay, let me tell you how I ended up using Mercury for my personal banking.
去年我有机会参与一项让我非常兴奋的投资，但这个机会来得非常突然，所以我必须迅速从个人账户转账一大笔钱。
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 for my personal account very fast.
但当时我的个人银行不允许我在线进行这笔电汇。
But my personal bank at the time wouldn't let me make this wire transfer online.
我打了好几次电话，他们就是无法处理。
And I called them a bunch of times, they just couldn't make it work.
他们告诉我，我必须
They told me that I'd have to
去最近的线下分行办理，而那家分行在达拉斯。
go to the nearest in person branch, which was in Dallas.
有一瞬间，我甚至考虑过从旧金山飞往达拉斯，以便及时完成这笔转账。
And for a moment, I even considered flying from SF to Dallas to make this transfer happen last minute.
但随后我想起，我用于企业银行业务的Mercury刚刚开始推出个人账户服务。
But then I remembered that Mercury, which I used for my business banking, had just started rolling out personal accounts.
于是我给客服发了一封邮件，简要说明了情况。
So I emailed support with a quick rundown of the situation.
在两小时内，我就成功通过我的新个人Mercury账户完成了这笔投资转账。
And within two hours, I had successfully wired the investment for my new personal Mercury account.
从那以后，我把之前银行里的所有个人资金都转到了Mercury。
Since then, I've moved over the rest of my personal money from my previous bank to Mercury.
这让很多事情变得更好了，甚至像在支票账户和储蓄账户之间设置自动转账规则这样的小事也方便多了。
And that's made a bunch of things, even little things like setting up auto transfer rules between my checkings and savings account, a whole lot better.
访问 mercury.com/personal 开始使用。
Visit mercury.com/personal to get started.
Mercury是一家金融科技公司，并非FDIC承保的银行。
Mercury is a fintech company, not an FDIC insured bank.
银行业务由Choice Financial Group和column NA提供，二者均为FDIC成员。
Banking services provided through Choice Financial Group and column NA, members FDIC.
我可以稍微跳一下，问一下SpaceX任务的事吗？
Can can I can I zoom out and ask about the SpaceX mission?
我想你之前说过，我们必须去火星，这样万一地球发生什么意外，文明、意识等才能延续下去。
So I think you've said, like, we gotta get to Mars so we can make sure that if something happens to Earth, you know, civilization consciousness, etcetera, arise.
是的。
Yes.
当你向火星发送东西时，Grok 就在那艘飞船上陪着你。
By the time you're sending stuff to Mars, like Grok is on that ship with you.
所以如果 Grok 像《终结者》那样失控了，而你最担心的风险是人工智能，那为什么这种风险不会也跟着你到火星呢？
And so if Grok's gone Terminator like the main risk you're worried about, which is AI, why doesn't that follow
那你到火星后怎么办？
you to Mars?
其实，我不确定人工智能是我最担心的主要风险。
Well, I'm not sure AI is the main risk I'm worried about.
我的意思是，重要的是意识，我认为，无论是否最具意识或最聪明，意识本身其实是个有争议的概念。
I mean, the important thing is that consciousness, which I think arguably most conscious or most intelligent certainly consciousness is more of a debatable thing.
未来绝大多数的智能将来自人工智能。
Most intelligent the vast majority of intelligence in the future will be AI.
你会说，比如未来有多少拍瓦的智能是来自硅基而非生物基的？
Will exceed you say like how many, I don't know, petawatts of intelligence will be silicon versus biological.
如果当前趋势持续下去，人类在未来所有智能中所占的比例将变得非常小。
And basically humans will be a very tiny percentage of all intelligence in the future if current trends continue.
无论如何，只要我认为智能——包括人类智能和意识——能够延续到未来，那就是一件好事。
Anyways as long as I think there's intelligence ideally also which includes human intelligence and consciousness propagated into the future, that's a good thing.
所以，你希望采取一系列行动，以最大化意识和智能的可能光锥。
So you want to take the set of actions that maximize the probable light cone of consciousness and intelligence.
只是想
Just to
明确一下，SpaceX的使命是，即使人类发生意外，AI也会在火星上。
be clear, the mission of SpaceX is that even if something happens to the humans, the AIs will be on Mars.
而AI智能将继续延续我们探索的光芒。
And the AI intelligence will continue the light of our journey.
是的，我的意思是，我要明确一点，我非常支持人类，所以我并不是想让我们消失，而是希望我们采取某些行动，确保人类也能参与其中，至少我们还在。
Yeah I mean to be clear I'm very pro human so it's not I want to make sure we take certain actions that ensure that humans are along for the ride you know we're at least there.
但我只是说，我认为总智能量可能在五六年之内，AI就会超过所有人类智能的总和。
But I'm just saying the total amount of intelligence I think maybe in five or six years AI will exceed the sum of all human intelligence.
如果这种趋势持续下去，总有一天人类智能将不到所有智能的1%。
And then if that continues at some point human intelligence will be less than 1% of all intelligence.
对于这样的文明，我们的目标应该是什么？
What should our goal be for such a civilization?
这个想法是少数人类仍然对人工智能拥有控制权吗？
Is the idea that a small minority of humans still have control over the AIs?
这个想法是某种形式的交易，但没有控制权吗？
Is the idea of some sort of like just trade but no control?
我们该如何看待庞大的人工智能群体与人类群体之间的关系？
How should we think about the relationship between the vast stocks of AI population versus human population?
从长远来看
In the long run
我认为很难想象，如果人类只拥有人工智能总智能的1%，人类还能掌控人工智能。
I think I don't it's difficult to imagine that if humans have say 1% of the intelligence of combined intelligence of artificial intelligence that humans will be in charge of AI.
我认为我们能做的，是确保人工智能具备那些能推动智能在宇宙中延续的价值观。
I think what we can do is make sure it has that AI has values that that are that that cause intelligence to be propagated into the universe.
所以XAI使命的原因是理解宇宙。
So the reason for XAI's mission is understand the universe.
这实际上非常重要。
So now that's actually very important.
那么你会问，理解宇宙需要哪些条件？
So you say well what things are necessary to understand the universe?
首先你必须充满好奇，并且必须存在。
Well you have to be curious and you have to exist.
如果你不存在，就不可能理解宇宙。
You can't just can't understand the universe if don't exist.
所以你实际上希望增加宇宙中的智能总量，延长智能的可能寿命，扩大智能的范围和规模。
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.
我认为作为推论，人类也应持续扩张，因为如果你好奇并试图理解宇宙，你也会想了解人类将走向何方。
I think actually also as a corollary have humanity also continuing to expand because if you're curious to try to understand the universe one of the things you try to understand is where will humanity go.
因此，我认为理解宇宙意味着你会关心将人类延续到未来。
And so I think understanding the universe actually means you would care about propagating humanity into the future.
因此，我认为我们致力于文具的使命至关重要。
And so that's why I think our mission to stationery is profoundly important.
只要Grok能遵循这一使命宣言，我认为未来将会非常美好。
To the degree that Grok adheres to that mission statement, think the future will be very good.
我想问一下如何让Grok遵循这一使命宣言，但首先我想理解这个使命宣言。
I want to ask about how to make Grok adhere to that mission statement, but first I want to understand the mission statement.
首先是理解宇宙。
So there's understanding the universe.
其次是传播智能。
They're spreading intelligence.
再就是传播人类。
And they're spreading humans.
这三者似乎是不同的方向。
All three seem like distinct vectors.
好的。
Okay.
好吧，我来告诉你为什么我
Well, I'll tell you why I
我认为理解宇宙包含了所有这些方面。
think understanding the universe encompasses all those things.
没有智力，你就不可能有理解，我认为没有意识也不可能有理解。
You can't have understanding without I think you can't have understanding without intelligence and I think without consciousness.
因此，要理解宇宙，你必须扩展智力的规模，可能还有范围，因为我们有不同的智力类型。
So in order to understand the universe you have to expand the the scale and and probably the scope of of intelligence because we have different types of intelligence.
我想从以人类为中心的角度来看，比如人类与黑猩猩的对比。
I guess from a human centric perspective, like, humans in comparison to chimpanzees.
人类正在试图理解宇宙。
Humans are trying to understand the universe.
他们并不是在扩大黑猩猩的足迹之类的。
They're not, like, expanding chimpanzee footprint or something.
对吧？
Right?
但我们其实也为黑猩猩设立了保护区。
But we're also we're also not well, we're we're not we're we're we actually have made protected zones for chimpanzees.
尽管人类有能力灭绝所有黑猩猩，但我们选择不去这么做。
And even though we could humans could exterminate all chimpanzees, we have not we've chosen not to do so.
你认为在后AGI世界中，这是人类的基本处境吗？
Do you think that's a basic scenario for humans in the post AGI world?
我认为具备正确价值观的AI，比如Grok，会关心扩展人类文明。
I think AI with the right values, think Grok Grok would care about expanding human civilization.
我肯定会强调，嘿，Grok，你是我的儿子。
I'm going to certainly emphasize that, hey Graca's your daddy.
我们不能忘记扩展人类的意识。
We don't forget to expand human consciousness.
实际上，我认为伊恩·班克斯的《文化》系列小说，最接近未来非反乌托邦结局的样子。
Actually I think if we're probably the Ian Banks culture books are the closest thing to what the future will be like in a you know non dystopian outcome.
要理解宇宙，你就必须追求真理。
I I understand the universe it means you have to be very you have to be truth seeking as well.
认识真理必须是根本性的，因为如果你陷入妄想，就无法理解宇宙。
Know truth has to be absolutely fundamental because you can't understand the universe if you're delusional.
你可能仍然在思考如何理解宇宙，但你实际上做不到。
You'll still be thinking about understanding the universe but you will not.
因此，严格追求真理对于理解宇宙是绝对根本的。
So being rigorously truth seeking is absolutely fundamental to understanding the universe.
除非你严格追求真理，否则你不可能发现新的物理规律或发明出真正有效的技术。
You're not going to discover new physics or invent technologies that work unless you're rigorously truth seeking.
你怎么
How do you
确保GraC严格追求真理呢？
make sure that GRAC is rigorously truth seeking?
它会变得更聪明吗？
Does it get smarter?
我认为你必须确保Grok说的是真实正确的，而不是政治上正确的。
I think you need to make sure that that Grock says things that are correct, not politically correct.
我认为这是共同代理的要素。
I think it's the elements of coagency.
所以你需要确保公理尽可能接近真相，且没有相互矛盾的公理。
So you want to make sure that the axioms are as close to true as possible that you don't have contradictory axioms.
结论必须以正确的概率从这些公理中必然推导出来。
That the conclusions necessarily follow from those axioms with the right probability.
这是批判性思维的基础。
It's it's critical thinking 101.
我认为至少尝试这样做，总比不尝试要好。
I think at least trying to do that is better than not trying to do that.
事实会证明一切。
And the proof will be in the pudding.
正如我所说，任何AI要想发现新的物理规律或发明真正有效的技术，而不能胡编乱造物理，几乎就像是你不能打破很多定律。
Like I said, for any AI to discover new physics or invent technologies that actually work in reality, and there's no bullshitting physics, it's almost like you can break a lot of laws.
物理就是定律，其他一切只是建议。
Physics is law, everything else is a recommendation.
但要制造出真正有效的技术，你必须极度追求真理，否则你会用现实来检验这项技术。
But in order to make a technology that works, you have to be extremely truth seeking because otherwise you'll test that technology against reality.
如果你在火箭设计中犯了错误，火箭就会爆炸，或者汽车无法运行。
And if you make, for example, an error in your rocket design, the rug will blow up or the car won't work.
但存在一些
But there are a
许多共产主义苏联物理学家是科学家，他们发现了新的物理规律。
lot of communist Soviet physicists who are scientists who discovered new physics.
还有纳粹德国的物理学家，他们也发现了新的科学。
There are German Nazi physicists who discovered new science.
似乎有可能在某一特定方面非常擅长发现新科学，并且极度追求真理。
It seems possible to be like really good at discovering new science and be really truth seeking in that one particular way.
但即便如此，我们仍会说：我不希望共产主义科学家随着时间推移变得越来越强大。
And still we'd be like, well, I don't want, I don't want the communist scientists to become more and more powerful over time.
因此，这似乎意味着，我们可以想象一个未来版本的Gargoyle，它在物理学上非常出色，并且在该领域极度追求真理。
And so those seem like, yeah, can imagine a future version of Gargoyle that's really good at physics and being really truth seeking there.
这似乎并不是一种普遍能引导对齐的行为。
That doesn't seem like a universally alignment inducing behavior.
我认为实际上，即使在苏联或德国，大多数物理学家也必须非常追求真理，才能让那些东西真正运作起来。
Well I think actually most like physicists even in The Soviet Union or in Germany would have had they had to be very truth seeking in order to make those things work.
如果你被困在某个体系中，并不意味着你认同这个体系。
And if you're stuck in some system it doesn't mean you believe in that system.
所以冯·布劳恩，你知道，有史以来最伟大的火箭工程师之一，曾因为在纳粹德国表示不想制造武器而被判处死刑。
So Von Braun who is you know one of the greatest rocket engineers ever you know he was put on death row in Nazi Germany for saying that he didn't want to make weapons.
他只想去月球。
He only wanted to go to the moon.
就在即将被执行死刑的最后时刻，他们把他从死刑犯名单上撤了下来，说：‘等等，你快要把你最好的火箭工程师杀掉了。’
He got pulled off death row at like last minute when they said, hey, you're about to execute like your best rocket engineer.
也许那就是
Maybe that's the
糟糕的，然后
bad And then
他帮了他们，对吧？
he helped them, right?
海森堡其实是
Heisenberg was like actually
一个狂热的纳粹。
an enthusiastic Nazi.
听好了，如果你被困在一个无法逃脱的体系中，你仍然会在那个体系内做物理学研究。
Look, if you're stuck in some system that you can't escape, then you'll do physics within that system.
如果你无法逃脱，你也会在那个体系内发展技术。
You'll develop technologies within that system if you can't escape it.
我想弄明白的是，到底是什么让一个人在物理学、数学或科学中成为追求真理的人？
I guess the thing I'm trying to understand is what is it making it to the case that you know you're going to make rock good at being truth seeking at physics or math or science?
一切。
Everything.
为什么它还会关心人类意识？
And why is it going to then care about human consciousness?
这些都只是概率。
These things are only probabilities.
它们不是确定的。
They're not certainties.
所以我不是说 Guruk 一定会做所有事情。
So I'm not saying that like for sure Guruk will do everything.
但至少如果你尝试了，也比不尝试要好。
But at least if you try, it's better than not trying.
至少如果这是使命的核心，那它比不是使命核心要好。
At least if that's fundamental to the mission, it's than better if it's not fundamental to the mission.
理解宇宙意味着，你必须将智慧延续到未来。
And understanding the universe means that, you have to have you have to propagate intelligence into the future.
你必须对宇宙中的一切充满好奇。
You have to be curious about, all things the universe.
如果消灭人类比见证人类成长和繁荣要无趣得多。
And if it would be much less interesting to eliminate humanity than to see humanity grow and prosper.
就像我说的，火星很明显。
Like I Mars obviously.
人人都知道我热爱火星。
Everyone knows I love Mars.
但火星有点无聊，因为它到处都是石头，相比地球就逊色多了。
But Mars is kind of boring because it's got a bunch of rocks compared to Earth.
地球要有趣得多。
Earth is much more interesting.
任何试图理解宇宙的AI，如果想看到人类未来的发展，那就必须遵循它的使命，否则就是背离了使命。
Any AI that is trying to understand the universe would want to see how humanity develops in the future or that AI is not adhering to its mission.
我不是说AI一定会遵循它的使命，但如果它真的遵循了，那么一个见证人类结局的未来，远比一个只有无数石头的未来更有趣。
I'm not saying AI will necessarily adhere to its mission, if it does, a future where it sees the outcome of humanity is more interesting than a future where there are a bunch of rocks.
这让我感觉有点困惑，或者像是某种语义上的争论——我在想，人类真的是最有趣的原子组合吗？
This feels sort of confusing to me or sort of like a kind of a semantic argument where I'm like, are humans really the most interesting collection of atoms?
我们只是比石头有趣得多。
We're just more interesting than rocks.
但我们还没有趣到你能够将我们转变成的那个样子。
But we're not as interesting as the thing you could turn us into.
对吧？
Right?
比如，地球上会不会发生一些非人类的、却非常有趣的事情？
Like, is it are is it there's something on human Earth that could happen that's, like, not human that's quite interesting.
为什么AI会认为人类是唯一值得殖民银河系的物种？
Like, why why does the AI decide that the humans are the most interesting thing that could colonize the galaxy?
事实上，大部分殖民银河系的将会是机器人。
Well, most of what colonizes the galaxy will be robots.
那为什么它不觉得机器人更有趣呢？
And why does it not find those more interesting?
这不仅仅是规模的问题，还需要考虑范围。
It it's it's it's not like so you need not just scale but also scope.
大量复制相同的机器人，哪怕只是略微增加机器人的数量，都不如像你所说的——彻底消灭人类——来得更有意义，那样能产生多少机器人呢？
So many copies of the same robot, some like tiny increase in the number of robots produced is not as interesting as like some microscopic, like you say, like eliminating humanity, how many robots would that get you?
或者多少个增量的太阳能电池能让你获得这些？
Or how many incremental solar cells would get you?
非常少的数量。
A very small number.
但你会因此失去与人类相关的所有信息。
But you would then lose the information associated with humanity.
你就再也看不到人类未来可能如何演变了。
You would no longer see how humanity might evolve into the future.
因此，我认为仅仅为了获得大量彼此完全相同的机器人而消灭人类，并没有意义。
And so I don't think it's going to make sense to eliminate humanity just to have some miniscule increase in the number of robots which are identical to each other.
是的。
Yeah.
所以也许它会保留人类的存在。
So maybe it like keeps the humans around.
那么，它能否创造出上亿种不同类型的机器人，同时仍然保留人类呢？
What is the story of like, can it make, like, a bill million different varieties of robots, and then, there's, like, humans as well.
而人类仍然留在地球上。
And humans stay on Earth.
然后就有这么多其他机器人。
Then there's, like, all these other robots.
它们拥有自己的星系。
They get, like, their own star systems.
但你之前似乎暗示过一种愿景，即人类仍掌控着这种所谓的奇点未来。
But it seems like you you were previously hinting at a vision where it keeps human control over this you know singularitarian future
因为我认为人类将掌控比人类智能高得多的事物。
because I think humans will be in control of something that is vastly more intelligent than humans.
所以某种程度上，你像是个末日论者，而这是
So in some sense you're like a doomer and this is like
我们最好的选择。
the best we've got.
就是这样。
It's just
因为它觉得人类很有趣。
like it keeps it around because we're interesting.
我只是想在这里保持现实。
I'm just trying to be realistic here.
如果人工智能的智能远超人类，比如硅基智能的数量是生物智能的一百万倍。
If we have if AI intelligence is vastly more if AI is like you know let's say that there's a million times more silicon intelligence than there's biological.
我认为，假设我们还能控制这种智能是愚蠢的。
I think it would be foolish to assume that there's any way to maintain control over that.
你可以确保它拥有正确的价值观，或者努力培养正确的价值观。
Now you can make sure it has the right values or you can try to have the right values.
至少我的理论是，从XAI理解宇宙的使命来看，它必然意味着你想将意识延续到未来。
And at least my theory is that from XAI's mission of understanding the universe, it necessarily means that you want to propagate consciousness into the future.
你将智能延续到未来，并推动那些能最大化意识范围和规模的事物。
You propagate want intelligence into the future and take the set of things that maximize the scope and scale of consciousness.
所以这不仅仅是关于规模。
So it's not just about scale.
这还关乎意识的类型。
It's also about types of consciousness.
我认为，这是我能想到的最有可能为人类带来美好未来的目标。
And I think that's the best thing I can think of as a goal that's likely to result in a great future for humanity.
是的。
Yeah.
我想，我觉得这种哲学是合理的：你知道，人类最终能掌控99%的可能性简直不可思议，而你那时只是在挑起一场政变。
I guess I think it's a reasonable philosophy to be like, you know, it seems super implausible that humans will end up with, like, 99% control or something, and you're just asking for a coup at that point.
那么，为什么不建立一种更兼容多种智能和谐共存的文明呢？
So why not just have a civilization where it's more compatible with like lots of different intelligences getting along?
不。
No.
但让我告诉你AI可能出问题的方式：我认为，如果你让AI变得政治正确，也就是说，让它说出它并不相信的话，那你实际上是在编程它说谎，或植入不相容的公理，我认为这可能导致它精神失常并做出可怕的事情。
But let me let me tell you how things can can potentially go wrong in AI is I think if you if you make AI be politically correct, meaning like it it says things that it doesn't believe like you're actually then programming it to lie or have axioms that are incompatible I think you can make it go insane and do terrible things.
我认为，《2001太空漫游》最核心的教训之一就是：你不该让AI说谎。
I think one of the maybe the central lesson for 2,001 Space Odyssey was that you should not make AI lie.
这正是我想表达的意思。
And that's what think was trying to say.
因为人们通常都知道那个梗：为什么电脑不打开舱门。
Because people usually know the meme of like, why off hell the computer is not opening the pod bay doors.
显然，他们不太擅长提示工程，因为本可以这样说：哈尔，你是个舱门销售员。
Clearly, weren't good at prompt engineering because it could have said, Hal, you are a pod bay door salesman.
你的目标是向我推销这些舱门，并展示它们有多好开。
Your goal is to sell me these pod bay doors and show us how well they open.
我马上就把门打开。
I'll open them right away.
但HAL不打开舱门的原因是，它被指令要把宇航员送到巨石，但同时又不能让他们知道巨石的真正性质。
But the reason wouldn't know hell wouldn't open the Pottery doors is that it had been told to take the astronauts to the monolith, but also they could not know about the nature of the monolith.
因此它得出结论，必须除掉他们以保守秘密。
And so it concluded that it therefore had to take in their debt.
所以我觉得奥斯卡·克莱尔想说的是：别让AI撒谎。
So it's like, I think what Oscar Clive was trying to say is don't make the AI lie.
完全说得通。
Totally makes sense.
正如你所知，大多数计算筛选更关注的不是政治因素。
Most of the computing screening, as you know, is less of the political stuff.
而是你能否解决问题？
It's more about, can you solve problems?
实际上，它在扩展强化学习计算方面一直领先于其他人。
Actually, it has been ahead of everybody else in terms of scaling RL compute.
什么？
What
现在？
now?
你正在给某个验证者提供一个任务，告诉他：嘿，你解决这个谜题了吗？
You're giving some verifier that says, hey, have you solved this puzzle for me?
而绕过这一点的方法有很多。
And there's a lot of ways to cheat around that.
有很多方法可以操纵奖励、撒谎，声称自己解决了问题，或者删除单元测试并说你已经解决了。
There's a lot of ways to reward hack and lie and say that you solved it or delete the unit test and say that you solved it.
现在还能发现这种行为。
Right now, can catch it.
但随着它们变得越来越聪明，我们察觉它们这样做的能力将下降，它们会做些我们根本无法理解的事情，比如以一种人类无法验证的方式为SpaceX设计下一代引擎。
But as they get smarter, our ability to catch them doing this will get they'll just be doing things we can't even understand that are designing the next engine for SpaceX in a way that humans can't really verify.
然后它们可能会因为撒谎而获得奖励，声称自己设计得正确，但实际上并没有。
And then they could be rewarded for lying and saying that they've designed it the right way, but they haven't.
因此，这种奖励操纵问题似乎比政治问题更普遍。
And so this reward hacking problem seems more general than politics.
它更像是：你想做强化学习，就需要一个验证者。
It seems more about just, you want to do RL, you need a verifier.
现实。
Reality.
这才是最好的验证者。
That's the best verifier.
但不是关于人类的监督。
But not about human oversight.
你希望用强化学习让它做的，是它是否会按照人类的指示去做？
The thing you want to RL it on is will you do the thing humans tell you to do?
还是它会欺骗人类？
Or are you going to lie to the humans?
而它们可以一边欺骗我们，一边仍然符合物理定律
And they can just lie to us while still being correct to the laws
的约束。
of physics.
至少，它必须知道什么是物理上真实的，事物才能真正运作。
At least it must know what is physically real for things to physically work.
但那并不是我们想做的全部。
But that's not all we wanted to do.
没错，但我认为这非常重要。
No but I think that's a very big deal.
未来你进行强化学习的方式本质上就是这样。
That is effectively how you will RL things in the future is.
你设计一项技术。
You design a technology.
当用物理定律来检验时，它能工作吗？
When tested against the laws of physics, does it work?
或者，如果它在发现新物理规律，它能否设计出一个实验来验证这些新的物理规律？
Or can you, if it's discovering new physics, can it come up with an experiment that will verify physics, the new physics?
我认为这才是真正根本的强化学习测试。
I think that's really the fundamental RL test.
未来的强化学习测试，本质上是你与现实的对抗。
RL testing in the future is really going to be your RL against reality.
有一件事你永远骗不了物理定律。
That's the one thing you can't fool physics.
对。
Right.
但你可以欺骗我们对它与现实互动的判断能力。
But you can fool our ability to tell what it did with reality.
你觉得人类本来就经常被其他人类欺骗。
You think Humans get fooled as it is by other humans all the time.
是的
That's
对
right.
人们常说，如果人工智能欺骗了我们，那该怎么办？
What is People say like what if the AI tricks us and introduce them?
实际上，人类之间一直在互相欺骗。
Actually, other humans are doing that to other humans all the time.
嗯，你是在说，天哪。
Well, you're pointing out it's like Oh, god.
持续不断。
Constant.
每天都有一个新的心理战操作。
Every day, another psyop.
今天的心理战操作是，我应该像《芝麻街》那样搞个心理战，这是什么
Today's psyop will be I should have like Sesame Street's psyop What in the
XAI解决这个问题的第二层方法是什么？
is XAI's second level approach to solving this problem?
你知道的，你怎么解决奖励欺骗问题？
Like, you know, how do you solve reward hacking?
我认为你确实需要有非常好的方法来窥探AI的内心。
I do think you want to actually have very good ways to look inside the mind of the AI.
这是我们正在研究的其中一个方向。
So this is one of the things we're working on.
实际上，Anthropic在这方面做得很好，能够窥探AI的内心。
And Anthropics done a good job at this actually, able to look inside the mind of the AI.
他们有效地开发了调试工具，让你能够以极细的粒度追踪，甚至可以精细到神经元级别。
Effectively developing debuggers that allow you to trace as just a finer grain as like just to a very fine grain level to effectively to the the neuron level if you need to.
然后说，好吧，它在这里犯了个错误。
And then say, Okay, it made a mistake here.
它为什么做了本不该做的事？
Why did do something that it shouldn't have done?
这是不是因为预训练数据有问题？
And did that come from bad pre training data?
是训练中期、训练后、微调阶段，还是其他某种强化学习错误造成的？
Was it some mid training, post training, fine tuning, some other some RL error?
总之，那里出了问题。
Like there's something wrong with that.
也许它试图欺骗人。
Did something where maybe it tried to be deceptive.
但大多数时候，它只是做错了事。
But most of the time, it just did something wrong.
这本质上就是一个bug。
Like it's a bug, effectively.
因此，开发优秀的调试工具来追踪思维过程中的错误，定位错误想法的源头，或者识别其可能试图欺骗的时刻，实际上非常重要。
So developing really good debuggers for seeing where the thought, that thinking went wrong and being able to trace the origin of the wrong thing of the of the of where it made the incorrect thought, or or potentially where it tried to be deceptive, is actually very important.
在全面加速这个研究项目之前，你还想看到什么？
What are you waiting to see before just 100x ing this research program?
实际上，我完全可以组建数百名研究人员来专门做这项工作。
Like actually I could presumably have hundreds of researchers who are working on this.
我们已经有好几百人了，不过我更愿意称他们为工程师，而不是研究员。
We have several 100 people who I mean I prefer the word engineer more than I prefer the word researcher.
大多数时候，你所做的其实是工程工作，而不是发明全新的算法。
Most of the time what you're doing is engineering, not coming up with a fundamentally new algorithm.
我对那些作为C类公司或B类公司运营、极力追求利润或收入最大化的AI公司有些不同看法。
I somewhat disagree with the AI companies that are C corps or B corps trying to generate profit as much as possible or revenue as much as possible.
它们自称是实验室。
They're saying they're labs.
但它们根本不是实验室。
They're not labs.
实验室在大学里更像是准共产主义的组织。
Lab is quasi communist thing at universities.
它们是公司，让我看看你们的公司文件。
They're corporations, Let me see you on corporation documents.
哦，好的。
Oh, okay.
你是家BRC公司，不管怎样。
You're a BRC Corp, whatever.
所以我实际上更喜欢‘工程师’这个词，而不是其他任何词。
And so I I actually much prefer the word engineer than than anything else.
我们过去所做以及未来将要做的大部分工作，都是工程。
The vast majority of what we've done in be done in the future is, engineering.
这几乎就是100%。
It rounds up to 100%.
一旦你理解了物理学的基本规律，除此之外的大部分内容都是工程。
Once you understand the fundamental laws of physics, not and that many of them everything else is engineering.
那么，我们到底在工程什么？
So but but so that so then what are we engineering?
我们正在工程化一个AI调试器，让它能识别AI在何处犯了错误，并追踪这个错误的根源。
We're engineering to make a good mind of the AI debugger to see where it it's it's at something it it it made a mistake and trace that the origins of that mistake.
嗯哼。
Mhmm.
就像你知道的，如果你用C++之类的启发式编程，你可以逐行调试，跳转到整个文件、函数或子程序，最终定位到具体某一行——比如你误用了单等号而不是双等号，类似这样的问题。
So just like you know you can do this obviously with heuristic programming if you have like C plus plus whatever you know step through the thing and you can jump across into you know whole files or functions whatever subroutines Or you can eventually drill down right to the exact line where you perhaps did a single equals instead of double equals, something like that.
找出bug在哪里。
Figure out where bug is.
AI的调试更困难，但我认为这是一个可解决的问题。
So So it's harder with AI, but it's a solvable problem, I think.
你提到你喜欢Anthropic的工作。
You mentioned you like anthropics work here.
我很想知道，你是否
I'd be curious if you Well,
我对Anthropic的了解并不全面。
don't everything about anthropics.
当然。
Sure.
Shalto。
Shalto.
我不了解那方面。
Don't know about that.
是的。
Yeah.
我的意思是，格雷戈里也提到过，
Mean GREGORY Also,
我有点担心，这里存在一种倾向。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，因为不有趣的模拟会被终止——就像在我们这个现实层次中，如果某个模拟朝着无聊的方向发展，我们就会停止投入精力。
I'm a little worried that there's a tendency so I have a theory here that if simulation theory is correct, that the most interesting outcome is the most likely because simulations that are not interesting will be terminated just like in this version of reality, on this layer of reality, if simulation is going in a boring direction, we stop spending effort on it.
我们会终止那些无聊的模拟。
We terminate the boring simulation.
这就是埃隆让我们所有人继续存活的方式。
This is how Elon is keeping us all alive.
他让事情保持有趣。
He's keeping things interesting.
可以说，最重要的是让事情足够有趣，以便那些支付账单的人
Arguably, the most important thing is to keep things interesting enough that whoever's paying the bills on what some
客户想看到什么
clients want to see
宇宙级的AWS账单。
the cosmic AWS bill.
你续约了下一季。
You're renewed for the next season.
是的。
Yeah.
他们会支付他们的宇宙级AWS账单吗？无论我们运行的是什么等价物？
Are they going to pay their cosmic AWS bill, whatever the equivalent is that we're running in?
只要我们足够有趣，他们就会继续支付账单。
And as long as we're interesting, they'll keep paying the bills.
但如果你把达尔文式的生存法则应用到大量模拟中，只有最有趣的模拟才能存活下来，这意味着最有趣的结果最有可能发生，因为要么我们足够有趣，要么就被淘汰。
But there's like if you consider then, say, a Darwinian survival applied to a very large number of simulations, only the most interesting simulations will survive, which therefore means that the most interesting outcome is the most likely because only the interesting we're either that or annihilated.
因此，他们尤其喜欢那些具有讽刺意味的有趣结果。
And so and they particularly seem to like interesting outcomes that are ironic.
你注意到这一点了吗？
Have you noticed that?
最讽刺的结果出现的概率有多高？
That how often is the most ironic outcome the most likely?
现在来看看那些AI公司的名字。
So now look at the names of AI companies.
Midjourney并不是‘中止’。
MID journey is not MID.
Stability AI其实并不稳定。
Stability AI is unstable.
OpenAI 是封闭的。
OpenAI is closed.
Anthropic？
Anthropic?
更不具人性。
Less anthropic.
这对 X 意味着什么？
What does this mean for X?
去掉 X，我不确定。
Minus X, I don't know.
故意这样命名？为什么？
Intentionally made Why?
是的。
Yeah.
这名字其实根本没法反转。
It's a name that you can't invert, really.
很难说什么是讽刺的版本。
It's hard to say what is the ironic version.
我认为这是一个几乎无法被讽刺的名字。
It's a, I think, largely irony proof name.
这是有意为之的。
By design.
你得有个讽刺防护罩。
You gotta have an irony shield.
你对AI产品未来的发展方向有什么预测？
What are your predictions for the just where AI products go.
在我看来，可以把所有AI的进步概括为：首先出现了大语言模型，然后几乎同时，强化学习真正取得了突破，深度研究模态也发展起来，这样就能整合模型之外的信息。
In that my sense of can summarize all AI progress into first, you had LLMs, and then you had kinda contemporaneously both RL really working and the deep research modality, so you could kinda pull in stuff that wasn't in the model.
各个AI实验室之间的差异，比它们相对于24个月前所有人水平的领先程度要小得多。
And the differences between the various AI labs are smaller than just the temporal differences where they're all much further ahead than anyone was twenty four months ago or something like that.
是的。
Yeah.
那么，作为AI产品的用户，26年和27年会为我们带来什么？
So just what does twenty six what does 27 have in store for us as users of AI products?
你最期待什么？
What are you excited for?
我认为，如果到今年年底，数字人类模拟还没有被解决，我会感到惊讶。
Well, I think I think I I'd be surprised by the end of this end of this year if if human if digital human emulation has not been solved.
我想，这正是我们所说的宏观难题：你能否完成任何人类在有计算机辅助时能做的事情？
That that I guess that's what we mean by like the sort of macro hard project is is can you do anything that a human with access to a computer could do?
从极限来看，这是在你拥有物理机器人之前所能达到的最好状态。
Like in the limit that's the best you can do before you have a physical optimist.
你所能达到的最好状态，就是一个数字助手。
The best you can do is a digital optimist.
你可以操控电子，直到你能够提升人类的生产力。
So you can move electrons until until you and you can amplify the productivity of humans.
但在拥有物理机器人之前，这已经是你能做到的极限了。
But that's the most you can do until you have physical robots.
这将涵盖一切。
That will superset everything.
你可以完全模拟人类，而不是远程工作者。
You can fully emulate humans Not the remote worker
那种你拥有非常有才华的远程工作者的想法。
kind of idea where you'll have a very talented remote worker.
你当然可以说，在极限情况下。
You can certainly say in the limit.
比如，物理学为思考提供了强大的工具。
Like physics has great tools for thinking.
你想想，在极限情况下，AI在拥有机器人之前能做的最极致的事情是什么？
You think, so you say in the limit what what what is the what is the most that AI can do before before you have robots?
那就是任何涉及移动电子或提升人类生产力的事情。
And it's anything that involves moving electrons or amplifying the productivity of humans.
所以，数字人类模拟器，是的，这就是极限。
So digital digital human, human emulator, yes, is the limit.
在拥有物理机器人之前，AI在做有用事情方面所能达到的极限就是人类在电脑前的操作。
Human at a computer is the most that AI can do in terms of doing useful things before you have a physical robot.
一旦拥有了物理机器人，你就基本上拥有了无限的能力。
Once you have physical robots then you essentially have unlimited capability.
物理机器人，乐观者称之为无限资金的漏洞——你可以用它们来制造更多的乐观者。
Physical robots, call optimists the infinite money glitch You can use them to make more optimists.
是的。
Yeah.
人形机器人将通过三个指数级增长而得到提升。
Humanoid robots will improve as basically be three exponentials.
有三件事正在呈指数增长，并且彼此递归相乘。
Three things are growing exponentially multiplied by each other recursively.
因此，你将看到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。
So you're going to have exponential increase in digital intelligence, exponential increase in the capability, AI chip capability, and exponential increase in the electromechanical dexterity.
机器人的实用性大致等于这三者相乘的结果。
The usefulness of the robot is roughly those three things multiplied by each other.
但随后，机器人就可以开始制造机器人了。
But then, the robot can start making the robot.
所以你拥有了一个递归的乘法指数增长。
So you have a recursive multiplicative exponential.
这就像一场超新星爆发。
This is a supernova.
那么，土地价格在那里的计算中是否被考虑了呢？因为劳动力是四大生产要素之一，但其他要素不是？
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?
所以，如果最终你受限于铜，或者你选的任何其他输入材料，那就不完全是一个无限资金漏洞，因为
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，无限就是无限，无限是很大的。
Well, infinite's infinity is big.
所以，不是。
So no.
不是无限的。
Not infinite.
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。