公关专家称人工智能热潮实为泡沫

泰勒·斯威夫特的最新专辑《秀场女孩的生活》才发行一周，舆论就已两极分化。一些评论家称其虚假且尴尬，另一些人则打出五星好评，但几乎所有人都认为这是自传。如果他们全都错了呢？如果每一个彩蛋、每一句歌词都只是表演的一部分呢？

Taylor Swift's latest album, The Life of a Showgirl, has only been out a week, and the discourse could not be more divided. Some critics are calling it fake and cringe. Others are giving it five stars, but almost everyone assumes it's autobiography. What if they're all wrong? What if every Easter egg, every lyric is just part of the performance?

我是《流行开关》的查理·哈丁。加入我们，仔细聆听，揭示秀场女孩的生活并非忏悔，而是技艺。请在任意播客平台收听《流行开关》。

I'm Charlie Harding from Switched On Pop. Join us as we listen closely to reveal how the life of a showgirl isn't confession, it's craft. Hear it on Switched on Pop wherever you get podcasts.

最高法院最近对平权法案的裁决究竟改变了什么？它对精英大学的录取意味着什么？我是普里特·巴拉拉，本周耶鲁法学院教授贾斯汀·德莱弗将做客《与普里特保持连线》，追溯平权法案的历史，并解释最高法院对SFFA诉哈佛案裁决后的现状。节目现已上线，请在任意播客平台搜索并关注《与普里特保持连线》。

What did the Supreme Court's recent decision on affirmative action actually change? And what does it mean for who gets into elite colleges? I'm Preet Bharara, and this week, Yale Law professor Justin Driver joins me on Stay Tuned with Preet to trace the history of affirmative action and explain where we are in light of the Supreme Court's ruling in SFFA versus Harvard. The episode is out now. Search and follow Stay Tuned with Preet wherever you get your podcasts.

来自Vox媒体播客网络，这里是《频道》节目，我是彼得·卡夫卡。我也是《商业内幕》的首席记者。今天我们要讨论两件事：第一，我们是否处于人工智能泡沫中？

From the Vox Media Podcast Network, this is channels with Peter Kafka. That's me. I'm also the chief correspondent at Business Insider. Today, we are talking about two things. First, are we in an AI bubble?

其次，通过主张存在AI泡沫能赚钱吗？为了回答这个问题，我邀请了埃德·齐特隆——一个你或许从未听说过，或者对其有强烈看法的人。齐特隆的本职是公关工作，他拥有自己的 agency。但过去几年，他以科技炒作戳破者的身份崭露头角，且言辞犀利。

And secondarily, can you make money by arguing that we're in an AI bubble? To answer that question, I'm chatting with Ed Zitron, who is either someone you've never heard of or someone you have very strong feelings about. Zitron has a day job in PR. He owns his own agency. But for the last several years, he's been making a name for himself as a tech hype deflator, a voluble one.

如果你困惑于这些身份如何共存，这正是我们对话的主要内容，尤其是在后半部分。但前半部分反映了当下许多人的讨论：我们是否处于AI泡沫中？如果是，泡沫破裂时会发生什么？正如我们所讨论的，我越多使用ChatGPT等AI工具，就越相信它们的持久力。

If you are confused about how those things fit together, well, that's a lot of what we talk about in this conversation, especially in the second half. But the first half of our conversation mirrors the conversation lots of people are having these days. Are we in an AI bubble? And if the answer is yes, then what happens when it deflates? And as we discussed, the more I use AI tools like ChatGPT, the more I'm convinced of their staying power.

但这与认为所有投入的资金都物有所值是两回事。而埃德·齐特隆认为，几乎这一切都是虚假的，毋庸置疑。但让我们听听他的原话。以下是我与埃德·齐特隆的对话：我身边的是埃德·齐特隆，根据你的网络活动范围，你可能对他相当熟悉。

But that's different than arguing that all the money that we're pouring into them is well spent. And Ed Zitron, well, he thinks almost all this is bogus, full stop. But let's let him make the case in his own words. Here's me talking to Ed Zitron. I'm here with Ed Zitron, a guy depending on where you hang out online, you might know fairly well.

他在Twitter上。他在Substack上。他还在《金融时报》上有专题报道。

He's on Twitter. He's on Substack. He's profiled in the financial times.

现在是Ghost。我去年初就从Substack搬出来了。

Ghost now. I moved off of Substack beginning of earlier last year.

好吧。那我就叫它新闻通讯好了。

Okay. I'm just gonna call it newsletter then.

不，不。那没关系。只是

No. No. That that's fine. It's just

就像，我对那个特别在意。好吧。它不是Kleenex纸巾。它是纸巾。就是那个

like, I'm very specific about that one. Fair enough. It's not Kleenex. It's a paper towel. It's the one

不主动推广纳粹的。

that doesn't actively promote Nazis.

好的。我们会谈到那个。Ed Ed写作。他做播客。他还经营一家公关公司。

Okay. We'll get to that. Ed Ed writes. He podcasts. He runs a PR firm.

嗯。

Mhmm.

写一份通讯。我想，我想我会把你的在线项目描述为‘互联网已坏’。是的，大体上是这样。而现在，你痴迷于解释为什么AI时代是一个可怕的错误。这是一个公平的总结吗？

Writes a newsletter. I guess I guess I would describe your project online as the Internet is broken Yes. Generally. And right now, you're obsessed with explaining why the AI moment is a terrible mistake. Is that a fair summary?

我想是的。是的。

I think so. Yes.

只是我已经做这件事有

It's just that I've been doing it for

很多年了，不只是今年。但现在，我想人们开始意识到我并没有错。对吧。但我现在

a lot longer than this year. But now it's now I think people are kind of working out that I was not wrong. Right. So but what I'm

说的是，你长期以来一直是大科技的批评者。

saying is you've been a critic of big tech for a long time.

是的。我是。

Yes. I have.

现在，显然，大家都聚焦于人工智能的时刻。是的。我们在这个节目上花了很多时间，因为我们讨论科技和媒体。重点是人工智能。我邀请你来是因为你有一个令人警醒的观点，解释了为什么你认为这一切正朝着糟糕的方向发展，我也想聊聊你是如何走到这一步的。

And now, obviously, are focused on the AI moment. Yes. And we spend a lot of time on this show because we talk about tech and media. It's focused on AI. And I wanted to have you on because you've got a bracing perspective that explains why you think this is all going in a terrible direction, and I also wanna talk about how you got here.

是的。这是我的介绍。让我们在讨论你的观点和论点之前，先聊聊你是如何走到这里的。你职业生涯是从记者开始的。

Yeah. That's my intro. Let's start with with with before we get to your take to your argument, talk about how you got here. You started your career as a journalist.

是的。所以16岁时，我在一家叫《电脑与电子游戏》杂志的地方实习，那是在西伦敦，后来我们搬到了法林厄姆，那里很压抑，就像伦敦大部分地区一样。但我是一名游戏记者，主要评测多人在线游戏。所以是从《魔兽世界》前时代到后时代。然后22岁时我搬到了美国，因为大学期间我在宾州州立大学待了一年。

Yes. So when I was 16, I did work experience at a magazine called Computer and Video Games magazine in what was was it West London, and then we went moved to Farringham, which was depressing, like most of London. But But I was a games journalist, and so I reviewed predominantly multiplayer online games. So the pre World of Warcraft era into the post World Warcraft era. And then I moved to America when I was 22 because I did a year at Penn State during college.

我当时想，哦，我挺喜欢美国的。所以22岁时搬到了美国，转行做公关，一直做到2012、2013年自己开公司。

I was like, oh, I quite like America. So I moved to America when I was 22, moved into public relations, kind of stayed around that until I did my own firm in 2012, 2013.

那家公司还在吗？

That firm still exists?

是的。EZPR还在运营。然后在2020年，我感染了新冠，坐在那里想，哦，我担心会死。我要写作，因为写作总能让我感觉好一些。我一直觉得写作很自在，所以我就这么做了。

Yes. EZPR is still around. And then in 2020, I got COVID, I and was, like, sitting around saying, like, oh, I'm worried about dying. I'm gonna write because writing's always make me feel better. I've always always felt quite at home writing, so I was like, I'm just gonna do this.

我当时有300个订阅者，而且

I had 300 subscribers, and

我想我当时有60次观看，是你作为附加内容在做的事情

I think I had 60 views per something you were doing as an add on to

是的，我只是在业余时间做做

Yeah. I was just doing it my spare time.

是的，公关是你的主业

Yeah. PR was your main business.

是的

Yes.

所以我一直坚持做。最多300个订阅者，60次观看。我记得有一个视频只有20次观看，这在2021年1月左右真的很尴尬。然后2021年是个有趣的年份。我觉得2021年在历史上是被遗忘的一年，它摧毁了一切

And so I kept doing it. 300 subscribers, 60 views at max. I think I had a 20 view one, which is really embarrassing in, like by, like, January 2021. And then 2021 was an interesting year. I think 2021 is a kind of forgotten year in history and how much it destroyed everything.

我只是觉得金融繁荣让人们变得疯狂。但当时我在写关于Clubhouse的文章。不知道你还记不记得那时候。我相信你对它很感兴趣，因为大家都在说这是下一个大型社交网络。嗯

I just think the financial opulence got people going crazy. But I was writing about Clubhouse. I don't know if you remember at the time. It I'm sure you had a lot of interest in it because everyone's like, this is the next big social network. Mhmm.

但如果你听一下，它听起来很糟糕。是的，就是垃圾。我当时写的文章有点像在说，嘿，这看起来更像是一个被支撑起来的资产

But if you listen to it, it sounded bad. Yeah. It was just crap. And I was kind of writing at the time saying like, hey. This kind of seems more like an asset being propped up.

这感觉像是Andreessen Horowitz，我认为

This feels like Andreessen Horowitz, who whom I think

这是你第一次说‘我认为科技皇帝们其实赤身裸体’吗？是的。这是你第一次表达那种想法吗？

Was this the first time you said, I think the tech emperors have no clothes? Yes. The first time you were sort of expressing that idea?

是的。当时我还在摸索如何写作。但有种解放的感觉，因为没有编辑，没有规则，没有束缚。就像，我只是随心所欲地写，写了关于Clubhouse的内容，当时围绕它发生了一些非常奇怪的新闻界内部争论，然后一切就慢慢消失了。是的。

Yes. And I was still figuring out how to write. But there was something kind of freeing about it because no editor, no cods, no masters. Like, I was just sitting around writing whatever I felt like, and I wrote about Clubhouse and just saying and there were these really bizarre intrajournalism arguments happening about it, and then it all just kind of went away. Yep.

然后我

And I went

每个人都能走出家门。

Everybody could walk outside the house.

是的。不过那很诡异，因为所有人，就在不到两个月前——是2021年6月吧？——都在说这是有史以来最重大的事情，然后它就慢慢消失了。然后是反对远程办公的浪潮，每个人都在说，哦，我们必须回办公室。办公室很重要。

Yeah. It was very weird, though, because everyone, not two months before was it June 2021, was saying, like, this is the biggest thing ever, and then it just kind of disappeared. And then the anti remote work stuff where it was everyone was saying, oh, we gotta go back to the office. The office is important.

持续如此。

Ongoing to be.

现在仍然是。但当时出现了一次大爆发，人们都在说我们需要回去。我们必须回去。但当时所有记者采访的对象总是经理和高管。我当时就说，嘿。

Still is. But there was this big flare up where people were saying, we need to go back. We must. But the people that all the journalists talked to at the time were always managers and executives. And I was saying, hey.

看。这似乎并没有真正涉及到工人。这看起来只是关于管理者的感受。然后这一切就从那里开始发酵，因为从那以后，我接触了加密货币、NFT和元宇宙——我的天啊。那个该死的元宇宙。

Look. This doesn't seem to actually involve workers. This appears to be just about how managers feel. And it all kind of grew from there because from there, I went to crypto and NFTs and the met oh my god. The bloody metaverse.

所以你一直在说这件大家都认为很棒的事情其实并不好。是的。而且人们对此有反应。

And so you've been consistently saying this thing that everyone thinks is great is not. Yes. And people respond to that.

他们确实有反应。令人恼火的是，你有点被贴上了唱反调的标签。嗯哼。但那并不是——如果我是唱反调，我应该是反对真相，而不是指出我认为非常合理的事情。我没有说任何荒谬的话。我只是在问，比如，嘿。

And they do. And the annoying thing is is that you kinda get branded as a contrarian Uh-huh. When that's not if I'm being a contrarian, I would be pushing against the truth rather than pointing out what I think are very reasonable things. I'm not saying anything ridiculous. I'm asking, like, hey.

我们能不能，就像，把这个和现实联系起来，如果可以的话？我们能不能真正把它连接到？然后，元宇宙就是个大例子。那是2021年10月吧？当时所有人都为此疯狂。

Can we can we kind of, like, attach this to reality, if you will? Can we actually connect this to? And, like, the metaverse was the big one. That was so was that October 2021? Everyone was losing their proverbial over it.

他们都在说，哦，这就是未来。马克·扎克伯格要这样做那样做。而我是个MMORPG玩家。我从11岁就开始上网太多了。所以我就觉得，所有这些元宇宙的东西听起来太理论化了，如果你对网络游戏有所了解的话。

They're saying, oh, this is the future. Mark Zuckerberg's gonna do this and this. And I'm an MMORPG guy. Like, I've been online too much since I was 11. And so I was like, all of this metaverse stuff sounds so theoretical if you know anything about online gaming.

要么元宇宙已经存在了，因为它就是网络游戏；要么它根本不存在，因为它只是《头号玩家》那样的幻想。

This either the metaverse already exists because it's online gaming or it doesn't exist at all because it's Ready Player One.

我写科技文章已经很久了。我常常持怀疑态度，甚至有些愤世嫉俗。有时我会反思自己，然后想，你看，如果我总是说这东西永远行不通

I've been writing about tech for a long time. I'm often skeptical, cynical. Sometimes I check myself and I go, well, look. If I just consistently said this thing will never work

是啊。那我就会是对的

Yeah. I will be right

十次里有九次。对吧。95%的情况下，大多数东西确实行不通。对吧？所以我总担心自己会下意识地说，这看起来就是胡扯。

nine out of 10 times. Right. 95% of the time, most things don't work. Right? And so I always worry that I'm reflexively saying, this is this seems like bullshit.

你内心有过这种自我对话吗？

Do you ever have that sort of discussion internally in your head?

没有。因为我是一个心碎的浪漫主义者。这才是关键。我其实非常非常热爱科技。比如，我现在就有好几款科技产品让我爱不释手，像Anker做的产品就很棒。

No. Because I'm a brokenhearted romantic. That's the biggest thing here. I actually really, really love technology. Like, there are several tech products I have right now that I'm absolutely I love like, there's like, Anker does great stuff.

比如，他们所有的电池都很出色。他们还有一套超棒的投影设备。有很多东西我觉得真的很酷。

Like, all their batteries are great. They have a crazy projector set up. There's tons of stuff I find really cool.

我觉得我们正处于

I think we're in the

比如，掌机游戏的黄金时代。我们只会得到更好的掌上游戏电脑。

golden age of handheld gaming, for example. We're we're only gonna get better handheld gaming PCs.

有些东西你是喜欢的。

There's things you like.

真的很令人兴奋。比如，我，真的对那类东西感到兴奋。而且我认为电池技术也在进步，比如氮化镓，它让更小的电池组和充电成为可能。有很多很酷的事情正在发生。

Really exciting. Like, I'm, like, genuinely excited about that stuff. And I think battery technology as it gets, like, was gallium nitride, which has allowed you to do smaller battery packs and charging things. There were so many cool things happening.

艾德，你找对人了，因为我可能刚买了价值200美元的安克（Anchor）产品。

Ed, you you're talking to the right guy because I've I just probably bought $200 worth of Anchor stuff.

太棒了，老兄。我有那个安克，所以我们……

Hell, yeah, dude. I've I've got the Anchor with the So we're we're

这个留到单独再说。不，不。艾德和彼得会讨论安克。重点是，你的观点是你喜欢这些东西。

gonna save that for a separate No. No. I Ed and Peter talk about anchor. The point your viewpoint is you like stuff.

是的。而且我真的——我的谎言是，我成长过程中没有朋友。我知道这听起来很戏剧性，但我真的没有。比如，我在友谊方面童年很艰难，对我来说是非常非常严重的抑郁。

Yes. And I really and my lie I didn't have friends growing up. And I know that sounds dramatic. I really didn't. Like, I had a a rough childhood with friendship, like a very, very hardcore depression for me.

所以我在网上成长了很多，通过在线世界我学到了很多，我看到了生活的各种可能性。因此我非常感激科技，超连接性造就了今天的我。它成就了我的事业，也缔造了我的友谊。我最亲密的朋友都来自网络，所以我对整个科技界怀有深深的感激之情。

So And I did a lot of growing up online, and I learned a lot of I'd like I learned the possibilities of life through online. So I'm very grateful to technology, and hyperconnectivity is something that's made me who I am. It's made my business. It's made my friendships. My closest friends are all from online, so I have a great debt of gratitude to technology at large.

让你能够做这项工作。

Allows you to do this work.

没错，完全正确。这其中有很多很棒的地方，但我能理解有人可能会觉得，哦，他什么都讨厌。不是的。我讨厌的是所有这些阻碍了真正酷炫事物的东西。

Exactly. Exactly. And there's so many cool things about it, but I can understand that someone would come off and be like, oh, he just hates everything. No. I hate all this stuff getting in the way of the cool shit.

那么我们来谈谈你现在痴迷的事情，显然每个人都在关注——AI泡沫。我们现在正处于一个人人都在说存在AI泡沫的时刻，包括萨姆·奥特曼，嗯，还有杰夫·贝索斯。

So let's talk about the thing you are obsessed about right now, and everyone is for for obvious reasons. AI bubble. We're in a moment now where everyone says there's an AI bubble, including Sam Altman Well well. Jeff Bezos.

贝索斯先生。

Mister Bezos.

你称之为AI的次贷时刻。所以给我最短的版本——你可以去读埃德的博客，他有一篇18000字的文章详细阐述了这一点。哦，是的，我们这里不会讲18000字。

You've called this the subprime moment of AI. So let's let's give me the shortest and you you can go and read Ed's blogs. He's got an 18,000 word piece that lays this all out. Oh, yeah. We're not doing 18,000 words here.

给我一个——不是电梯演讲，但比电梯演讲稍长一点的——对我们现状的批评。

Give me the the not the elevator pitch, but a little longer than an elevator pitch critique of where we are.

具体来说，次级AI危机指的是一个相当简单的情况：目前所有使用这些模型并与之连接的公司都在支付补贴价格。它们全都处于亏损状态。一旦Anthropic和OpenAI提高价格（它们已经对企业客户推出了优先处理的高级层级），就会产生连锁反应，那些本就亏损的AI初创公司将无法运营。但更大的AI泡沫在于所有这些业务都在亏钱。成本只会越来越高，而全球私募股权市场上根本没有足够的资金。

So specifically, subprime AI crisis refers to something quite simple, which is all of the companies currently using these models connecting to them are paying subsidized rates. They're all unprofitable. Once Anthropic and OpenAI raise their prices, which they're already doing to their enterprise clients with premium tiers of priority processing, then there will be a knock on subprime effect where already unprofitable AI startups will be kind of unable to run. But the grand AI bubble thing is all of this stuff all loses money. It's only getting more expensive, and there is quite literally not enough money in the world available in private equity.

甚至像养老基金这样的机构，现在都难以支付OpenAI开展业务所需的1,250,000,000,000美元。

And even in like, pension funds would struggle to pay the $1,250,000,000,000 now that OpenAI needs to do any of their stuff.

我认为人们普遍——不知道‘适应’这个词是否准确，但我们可以用它——往往当互联网产品刚出现时，它们是免费、非常便宜或享受补贴的，随着时间的推移，如果公司成功了，它们就会提高价格。通常人们不喜欢这样。比如优步以前比现在便宜得多。现在它们必须盈利，所以价格就上涨了。

I think people are generally I don't know if comfortable is the right word, but we'll use it with the idea that oftentimes when Internet stuff comes out, it is free or very cheap or subsidized, and that over time, if the company succeeds, they're able to raise rates. Oftentimes, people don't like that. If you're on the you know, Uber used to be much cheaper than it is now. Now they have to make a profit. Now things are more expensive.

诸如此类的情况不断发生。Netflix曾经每月8美元、10美元就能提供所有内容。现在每月24美元，内容却少了很多。我们似乎多少理解这是一种发展规律。你不认为AI也会这样吗？

On and on and on. Netflix gave you everything on the Internet for $8 a month, $10 a month. Now it's $24 a month and has much less. Sort of sort of we sort of understand there's sort of a progression of things. You don't think that's gonna happen with AI?

完全不一样。根据我查到的资料，亚马逊网络服务按现今货币计算，十年间的资本支出约为63,000,000,000美元。而我认为OpenAI的有效基础设施成本超过100,000,000,000美元。这包括微软建设的资金和基础设施，但我们不知道微软基础设施中有多少是OpenAI的。可能高达数千亿美元。

Totally different. Amazon Web Services, from what I could find, was in today's money cost about $63,000,000,000 worth of CapEx over ten years. So that's compared to what I think the effective infrastructure cost of OpenAI is over 100,000,000,000. That's funding and the infrastructure that Microsoft built, but we don't know how much of Microsoft's infrastructure is OpenAI. It could be as could be hundreds of billions of dollars.

但由于OpenAI没有承担这些成本，人们没有将其联系起来。关键在于所有其他繁荣期都没有这么亏损。这就好比每次优步打车要花1万美元，而且靠长颈鹿血运行。这成本高得离谱。推理成本（即生成输出所发生的一切）正在上升，而且还在持续扩大。

But because OpenAI is not shouldering that cost, people don't associate it. The point is all of these other booms, nowhere near as unprofitable. This would be like if every Uber cost $10,000 and ran on giraffe blood. It's crazy how expensive this is. The cost of inference, which is the basically everything that happens to generate an output, is going up and has only continued to expand.

大多数人甚至没有关注这里的基本经济学原理，而且这与以往完全不同。根本没有可比性。

Most people aren't even looking at the basic economics here, and it's nothing like before. Like, it's there is no comparison.

这是你的根本批评吗？是说经济上这笔账算不过来，还是你的主要批评在于技术本身被过度炒作，或者两者都有？两者我都知道。所以

That your fundamental criticism? Is that economically this is not gonna pencil out, or is your major critique with the tech itself that this tech is overhyped, or are both things equal? Know both. So

大型语言模型作为一种工具，作为基础技术本身还算有趣，没问题。但它们要实现规模化运作，需要窃取所有人的数据，造成真实的环境危害，损害我们的电网，耗费难以置信的资金，而且实际成效有限。最让我恼火的是，人们谈论生成式AI时总说它能实现那些它根本做不到的事情。

large language models as a tool already require as an as a basic thing, a tech kind of tech, interesting, fine. They are to do the things they're doing at scale, they require stealing from everyone, genuine environmental harm, harm to our power grid, unbelievable amounts of money, and they don't even do that much. That's the thing that really pisses me off. The way that people talk about generative AI is like it can do these things. It's never going to be able to do.

智能代理根本不像他们描述的那样存在。这和一两年前的情况差不多，你可以说现在有了推理能力，有了这个那个，但本质上大同小异，产出效果也差不多。

Agents don't exist in the way they describe them. It's the same thing as it was a year or two ago, you can say, oh, there's reasoning now. There's this. It's mostly the same thing. It has about the same outputs.

编程模型确实有所改进，但就连SWE bench这种基础基准测试现在可能都已经失效了。假设我们正处于泡沫中。

I mean, with coding models, they've got better, but even SWE bench, the very basic way they benchmark this may be broken now. Assuming we are in a bubble Right.

也就是说这些公司大部分都会消失，就像我们请亨利·布洛杰特讨论过的互联网泡沫时期那样。假设会有几个赢家，其他大部分都会烟消云散，可能还会有些连带影响。我这么说可能有点轻描淡写，但就算有些人损失惨重，总会有真正的赢家脱颖而出。谁是赢家？

And that a bunch of these companies are gonna go away and and we have again, we can we had Henry Blodgett on, we talked about what's parallel to the .com boom. Assume that there's a couple winners and that most everything floats away and maybe there's some contagion. I'm saying that in a very lackadaisical way. But let's say some people lose a bunch of money, but there are real winners that come out of this. Who winners?

不过等等，在讨论这个之前，你难道不相信构建和维护这项技术的底层经济模型有任何可能算得过账吗？不，我不相信。我认为如果大型语言模型真有未来，

But wait, before we get to that part, you don't believe there's any way that the underlying economics of what it takes to build build and maintain this technology will ever pencil out? No. I don't. I think that if there's ever a future for large language models,

那只能是客户端模式，让人们用有限参数的DGX设备本地运行模型。现在已经有人在本地运行模型了，虽然我不知道实际规模能有多大。现在就能做到，这可能是个方向，但眼下完全没有经济可行性，我认为永远都不会有。

it will have to be either client side, so people running their own DGX boxes with limited parameter models. You've already got people running models locally. I don't know to what practical scale, boom. You can do it right now. That might happen, but there is no economic case for this right now, and I don't think there ever will be.

没有人在这上面赚钱。每个人都在亏钱，而且只会亏得更多。微软，我几周前报道过这个。微软在Microsoft 365 AI Copilot上有800万付费活跃许可证。引用The Information的话，这是他们的皇冠上的明珠。

No one is making money on this. Everyone's losing money, and they're only losing more. Microsoft, I reported this out a couple weeks ago. Microsoft has 8,000,000 paying active licenses on Microsoft three sixty five AI Copilot. That is their crown jewels, to quote the information.

他们有大约4.4亿付费用户，却只能让800万活跃用户每月支付30美元使用Copilot，这还是在假设他们没有折扣的情况下，而他们肯定是有折扣的。即使你假设有50%的付费许可证不活跃，那也就是1200万。对微软来说这不算什么。这是小钱。从微软的年收益中砍掉30亿美元是很糟糕的。

It's like 440,000,000 paying users, and they can only get 8,000,000 actives to pay $30 a month for Copilot, and that's assuming they got they weren't discounted, which they were for sure. Even if you assume there's 50% inactive paying licenses, that's 12,000,000. It's nothing for Microsoft. That's chump change. Cut $3,000,000,000 from my from Microsoft in an annual earnings is terrible.

再说一次，你不认同像铁路或有线电视行业那样的类比，那些行业需要前期投入巨额资金来建设基础设施。你会亏损多年，甚至几十年，但最终会建成极具价值的东西。

And again, you're not comfortable with analogies like railroads or the cable TV industry where you have to spend a ton of money upfront to build out this infrastructure. You lose money for years, maybe decades, and eventually you have built something that's enormously valuable.

不。GPU根本不是那样的。它们的能力有限，它们非常擅长并行处理，可以处理大量数据流。我知道我把技术简化了一点，但它们没有其他可大规模应用的商业模式。如果有，我们早就看到了，因为总会有人尝试。

No. GPUs just aren't like that. They are limited they're extremely good at parallel processing, can throw a bunch of data in stream. I realize I'm truncating the tech a little bit, but they don't have other applicable business models at scale. And if they did, we'd have seen another one because someone would have tried them.

到现在总该有人尝试过别的方法了，否则这就只是个烧钱机器。所以情况是这样的，好吧，铁路、有线电视、光纤等等。所有这些都有其他用途。而且，它们是分散的。而现在的GPU都很好。

Someone would have tried something by now because otherwise, there's just the money burning machine. So you've got this thing where, okay, railroads, cable, fine, fiber optic, what have you. All of those have other use cases. And also, they were spread out. The GPUs right now are all good.

它们都高度集中。你有CoreWeave、微软，或者说微软、亚马逊、Meta、谷歌，它们不会倒闭。即使AI归零，不管结局如何，它们仍然会拥有那些GPU，谷歌的话还有TPU。仍然会拥有那些。市场上已经充斥着大量廉价的AI GPU，这已经在发生了。

They're all heavily centralized. You've got CoreWeek, Microsoft, or Microsoft, Amazon, Meta, Google, they're not coming of business. Even if AI goes to zero, not that whatever whatever end it has, they're still gonna have those GPUs and TPUs in Google's case. Still gonna have those. You're gonna have a bunch of cheap AI GPUs flooding the market already happening.

你现在已经可以便宜地买到H100和H100了。

You can already get a one hundreds and a h one hundreds for cheap.

没错。而且他们需要，并且他们持续需要购买更多、更好的设备。

Right. And they and they will need to and they continually need to buy more of them, buy better ones.

在价值上。这就像买新车然后立刻当二手车卖掉。

In value. They they're like it's like buying new cars and selling them used immediately.

就像你必须不断挖开地面重新铺设光纤一样。

It'd be like if you had to keep digging up and relaying the fiber optic.

是的，实际上就是这样。而且光纤的成本远超你的实际需求。还有个巨大的公开噱头在说：嘿，我们有新光纤了。旧的那个现在烂透了，根本不行。

Yes, effectively. And also the fiber optic costs more money than you ever need it to. And also there is a big public stock going, Hey, we've a new fiber optic. The last one sucks. It's shit now.

你需要最新的，你需要全新的。这就是英伟达。对，就是英伟达。而且这也很疯狂，因为它还很昂贵。这一切都更...你谈到光纤部署。

You need the new new and you need a whole new That's your NVIDIA. That's NVIDIA. Yes. And it's crazy as well because it's also expensive. It's all so much more you talk about fiber rollouts.

你谈到电信部署。那些相对分散。而这个高度集中，因为成本太高。所以拥有所有GPU的人主要是超大规模厂商或新云厂商，它们就像是英伟达克苏鲁的触手。它们就像...蔓延开来

You talk about telecoms rollouts. Those were relatively distributed. This is heavily centralized because they cost so much. So you've got the people who own all the GPUs are predominantly either hyperscalers or neo clouds are just tendrils of NVIDIA's Cthulhu. They just like, the the spreading

我不知道我们是否想...我知道克苏鲁是什么，因为我一直在研究。我知道什么是超大规模厂商和新云厂商。但你在谈论...你在谈论基本上所有非谷歌、微软、Meta的公司

I don't know if we wanna I I know what Cthulhu is because I've been reading up. Know what a hyperscaler is and a Neo Cloud. But you're talking about you're talking about all the the basically, everyone who's not Google, Microsoft, Meta

嗯。

Mhmm.

谁是提供镐和铲子的生意。是的。我们可以分解一下。

Who's in the business of supplying the picks and shovels. Yes. We can break it down.

所以情况是这样的，有一家叫英伟达的公司。他们销售GPU图形处理单元。我说这个的时候，不是指PC游戏机里的那些。我指的是这些AI用的，是那些巨大的机架式服务器，一个巨大的服务器里，我想是放了72个。然后用高速网络把它们运行在一个叫做集群的东西里。

So what happens is you have a company called NVIDIA. They sell a GPU graphics processing unit. Now when I say this, I don't mean the ones in PC gaming machines. I mean these AI ones, which are these giant rack mounted, so a giant server, and you put, I think it's 72 of them in one rack. And you run them in something called a cluster using high speed networking.

基本上，把它们全部集群在一起进行大规模运算。推理，也就是生成输出。这就是其力量所在

Basically, cluster them all together to do massive. The inference, so the creation of the output This is the in the power of

你的你的LLM。

your your LLM.

输出或训练它。是的。所以英伟达销售服务器架构，但特别是GPU。现在之所以只有英伟达能做到，是因为一个叫CUDA的东西。CUDA是一个软件库编程语言，基本上允许... 有人可能会对我这样描述感到生气。

Out or trains it. Yep. And so NVIDIA sells the server architecture, but specifically the GPUs. Now the reason that it's just NVIDIA is something called CUDA. CUDA is a software library coding language that basically allow someone's gonna get angry at how I describe that.

你可以用它来在GPU上运行计算程序。其他公司还没有真正解决这个问题。中国在尝试。AMD在尝试。每个人都在尝试，但英伟达早就解决了这个问题，并且他们拥有大量的人才。

The you use you can run computing programs on GPUs. No one else has really worked this out. China's trying. AMD's trying. Everyone's been trying, but NVIDIA worked this out a while ago, and they've got a ton of talent.

所以这基本上意味着英伟达是这方面的唯一供应商。

So this means that NVIDIA is the pretty much the single vendor for this.

你想让这个运作起来，就需要英伟达的产品。

You wanna make this work, you need NVIDIA's.

你必须这么做。根本没有其他选择。根据我与业内人士交流所知，这确实是因为训练需求。推理方面或许有其他选项，但我的意思是，如果有其他选择，人们早就大规模部署了。但你不得不与英伟达合作。

You have to. You just there's no other choice. And from what I know from people I've talked to, it really is because of training. Inference, maybe there are other options, but I mean, if there were other people, be installing them at scale. But you have to work with NVIDIA.

所以英伟达简直就是在印钞。但问题是英伟达让市场上了瘾。上季度英伟达同比增长55%，市场却不满意。不，完全不满意。

So NVIDIA has just been printing money. But the problem is that NVIDIA has got the markets addicted. So the market's 55% growth year over year last quarter for NVIDIA, markets weren't pleased. No. No.

黄仁勋先生，你们本该卖出更多GPU的。我觉得他们卖了394亿美元，但这还不够。这才是真正的问题所在——在这个高度集中、庞大无比的泡沫之上，还有着婴儿般喜怒无常的市场。

Mister mister Wang, you should have sold more GPUs. I think they sold $39,400,000,000 worth. It's not enough. And that's the actual real problem because on top of this incredibly centralized, really just incredibly large bubble, you have markets that have the temperaments of babies.

没错。但话说回来，我们对此已经习以为常，而且这种情况周期性地出现，反噬我们时总会让人惊呼：我们当初到底在想什么？是啊，那东西确实被高估了。

Right. But again, we are kind of used to that, and again, it periodically kind of there's a cycle to it, comes and bites us on the ass and go, woah. What were we all thinking? Yeah. That that thing was overvalued.

我们会重新估值。特斯拉就是个很好的例子。它还没有完全回落，事实上又涨回去了。但埃隆·马斯克必须不断告诉人们：这不止是电动汽车公司，而是更伟大的存在。

We mark it down. Tesla's a great example of this. It hasn't fully come down yet. In fact, it's back up again. But Elon Musk has to keep telling people that this is more than electric car company, it's something else.

这是一家AI公司，不管它是什么。股价被严重高估，但还在持续上涨。它就像一台永动机，直到某天突然失灵。没错。然后人们会损失一大笔钱，但世界照常运转。

It's an AI company, whatever it is. The stock is wildly overinflated, but it keeps going. It's sort of a perpetual motion machine until one day it doesn't. Right. And then people have lost a bunch of money, but the world moves on.

是的。所以英伟达的增长速度不可持续或估值过高，这对在英伟达工作、持有英伟达股票的人来说似乎是个问题。

Yes. So the fact that Nvidia has an unsustainable growth rate or a valuation that's too high, that seems like a problem for people who work in Nvidia, own Nvidia stock.

将会是

Would be

一场壮举。

a feat.

对吧？它不就是标普500指数中如此大的一部分吗？占了7%到8%。‘瑰丽七巨头’中的其他六家公司——特斯拉、谷歌、Meta、微软、苹果、亚马逊。天哪。

Right? Wasn't such a large part of the S and P 500. Seven to 8% of it. The magnificent seven, so the other six companies, Tesla, Google, Meta, Microsoft, Apple, Amazon. Good lord.

我应该能脱口而出这些名字。‘瑰丽七巨头’占了美国股市市值的35%。英伟达占了7%到8%。他们是股市上最大的公司。这对每个人都会很糟糕。

I should remember those off the top of my head. The magnificent seven is 35 of the value of The US stock market. Nvidia is seven to 8%. They are the largest company on the stock market. This is going to be bad for everyone.

几乎每个人的401k养老金计划都会部分投资于‘瑰丽七巨头’。特斯拉在市场中占比小得多，但除此之外，大量私募资金并没有被用来支撑特斯拉。我记得读到过一些数据。过去三个季度，每季度至少有500亿美元。

Pretty much everyone's four zero one k's are going to be partially in the magnificent seven. Tesla is a much smaller part of the market, but on top of that, large amounts of private equity are not being shoved on propping up Tesla. You've got I think I read something. It was $50,000,000,000 at least per quarter for the last three quarters.

所以这不仅仅是你的401k直接与此相关，因为你要么是直接投资者，要么是通过指数股票投资的。这就是涉及到次贷的部分，对吧，存在传染性，经济中有很多深层次的部分

So this is more than just your four zero one k is directly connected to this because you're either a direct investor or you're an investor through an index stock. This is where you get to the subprime part, right, that there's contagion and that there's lots of deep parts of the economy that

联系 次贷问题特指人工智能领域的某一方面。我不认为这像次贷危机那样，因为那次危机牵扯太广。我认为这不会像金融危机那么严重，但这并不意味着情况会好。只是说我们不会看到大批人被驱逐。我们不会看到数百万获得房屋的人。

ties The subprime thing was specifically referring to one part of the AI thing. I don't think that this is like the subprime mortgage crisis because that was so tied up. I think that this would be not as bad as the great financial crisis, but that doesn't mean it will be good. It just means that we're not going to see a bunch of people evicted. We're not seeing millions of people who got houses.

就像，那太可怕了。这更像是电信泡沫。基本上是上一次互联网泡沫。是电信加初创公司，但区别在于，当所有那些初创公司在互联网泡沫期间倒闭时，还有一堆有用的硬件，因为那是在亚马逊网络服务之前。是在普及云存储之类的东西之前。

Like, that was horrifying. This is much more like the telecommunications bubble. It's basically the last .com bubble. It's telecoms plus startups, but the difference is is that when all of those startups died during the .com bubble, there was a bunch of useful hardware because it was pre Amazon Web Services. It's before you had kind of ubiquitous cloud storage and such.

有一些实用性，你可以打折买到阿伦椅。你可以买到实际的...有大量资金被烧掉了。哦，当然。

There was some utility, and you could buy Aaron chairs at a discount. You could buy actual There was a ton of money that got incinerated. Oh, sure.

但我的意思是，当时有人不得不自己搭建服务器。因为没有AWS。所以人才和技术得到了有益的传播，我不是说那是好事。

But what I'm saying is there were people that built their own servers because they had to. There was no AWS. So there was a useful propagation of talent and stuff that I'm went not saying it was good.

我认为科技泡沫的结果是很多人损失了很多钱。但总的来说，无论是实际的服务器，我认为更广泛地看是技术的传播。是的。总体上是净收益。今天你和我能在互联网上做这项工作。

I would argue that the result of the tech bubble was a lot people lost a lot of money. But in general, whether it's the actual servers, I think more broadly the propagation of technology. Yes. Overall net positive. You and I are doing this work today on the Internet.

那都是好的。所以我们当时估值错了。大多数人押错了宝。这有点资本主义的味道。对吧？

That was all good. So we got the valuations wrong. Most people bet on the wrong horses. It's kind of capitalism. Right?

所以如果我们暂时把经济部分搁置一边

So if we table the economic part for a second

然后

and

转向技术方面。没错。在我看来，我们很可能会重演互联网泡沫的情景：许多我们原以为会成功的公司并没有成功，但各种形式的技术在未来多年仍将持续发挥作用。多年来我一直对人工智能持怀疑态度，而过去几年随着ChatGPT的周期性爆发，我会去让它写我的传记。它一开始写得正确，但你知道，一段之内就开始严重胡编乱造。

move on to the tech. Right. If if it seems to me we could very easily have a replay of the .com bubble where a lot of the things that we thought were gonna be successful companies aren't, but the technology in various forms continues to be useful for years and years in the future. I have been a big AI skeptic for years, and over the last couple years as ChatGPT has blown up periodically, I'd go to it and ask it to write a biography of me. It would start off correctly, and then, know, within a paragraph, start lying dramatically.

现在它不再那样做了。好吧。行吧。无所谓。那不算什么，但我现在发现它对我的工作确实很有用。

It doesn't do that anymore. Okay. Fine. Whatever. That's not that but what I am finding now is that it's I find it genuinely useful for my work.

即使我不做那些，它也是一个好得多的

Even if I wasn't doing that, it's just a much better

它在哪些方面有用呢

Where is it useful in

我马上会说到，但它是一个更好的谷歌。在搜索方面比谷歌强得多。如果它止步于此，如果它永远不会比现在更好，而只是在互联网上为你查找信息比当今任何其他工具都更出色，那似乎就是件大事了。不是吗？

I'll get to that in a second, but it's a much better Google. It's much better than Google at Googling. If it stops there, if it never gets any better than it is today, and it's just better at finding things on the Internet for you than anything else that exists today, that seems like a big deal. No?

并不完全是。不是吗？只是因为这在当下已经变得商品化了。有趣的是谷歌似乎解决不了这个问题，但我

Not really. No? It's just because that's a commoditized thing at this point. Now it's funny that Google can't seem to work this out, but I

这其中有个真正的讽刺：它比谷歌更好的原因之一是，谷歌允许这种生态系统建立起来，每个人都创建糟糕的网页，这些网页是给机器人而不是人类阅读的。所以优质信息被埋没其中。ChatGPT真正做的是剔除所有垃圾信息，并以比谷歌更好的方式整合它们。

And there's a real irony that that one of the reasons it's better at Google than Google is that Google has allowed this sort of ecosystem to build up where everyone creates terrible web pages that are meant to be read by robots, not humans. And so the good information is buried in there. And so what ChatGPT is really doing is taking out all the cruft and assembling it better than Google can do.

在某种程度上，我有一个更宏大的理论，我认为整个ChatGPT的情况实际上是在2020年由Prabhakar Raghavan创造的，他就是那个摧毁了谷歌的人。

To an extent, it's the thing is I have this grander theory that I actually think all of ChatGPT was created the whole situation was created in 2020 with Prabhagar Raghavan, the guy the man who destroyed Google.

我听说过这个说法。我们...我们暂且搁置这个话题

I've heard I've heard this one. We'll we'll let's table that

不，不。我们可以搁置，但我想说的是谷歌停止专注于改进搜索。他们停止了在搜索领域的创新，所以搜索质量下降了。因此ChatGPT才显得稍微好一些。

No. No. We can table it, but I'm saying that Google stopped focusing on making search better. They stopped innovating in search, so search decayed. So ChatGPT being slightly better.

我认为你不能把这归咎于某一位工程师或

I don't think you can put this off the foot of one engineer or

我不是说...不。我是说谷歌整体放弃了改进搜索。抛开宣传不谈，我只是说谷歌这种放弃的想法是导致这种情况的原因。如果搜索不断创新，你认为大家还会如此...

I'm not saying oh, No. I'm saying the wider Google thing of them giving up on making search better. Putting aside propaganda, I'm just saying that the idea of Google kind of giving up is why this happened. If search if search had innovated, do you think that everyone would be as

重要吗？这又是一个经典的创新者困境。对吧？你打造了这个东西，它让你赚得盆满钵满。

important? Again, it's a classic innovator's dilemma. Right? You've built this thing. You make it mince money.

嗯。

Mhmm.

它表现得非常好。你就没什么动力去对它进行重大改动了，而别人可能会冒出来，把你原来做的事情做得更好。

It does really well. You're less incented to go and make major changes to it, and someone can come up and and do a better job of the thing you were doing.

他们这些年确实做了很多重大改动。只是越改越糟。他们添加了...某种程度上模糊了与广告的界限。但是，嗯。撇开这个不谈，ChatGPT所做的...也许我不认为它是一个很棒的搜索引擎，但它比谷歌更好吗？

They made they made major changes over the years. They just made it worse. They added they kind of blurred the line with ads. But Mhmm. Putting that aside, what ChatGPT has done may be I don't agree that it's a great search engine, but is it better than Google?

我不知道。重点是

I don't know. Point is

我上周末用它来帮我决定买哪个Anchor产品，要是没有它，我很难搞清楚到底是Anchor还是其他品牌的东西。我有一辆旧车和一部新手机，我需要买一个转接头把两者连接起来。

I used it this weekend to help me figure out which Anchor products to buy, and I would have had a very difficult time figuring out which it was Anchor and some other stuff. I have an old car, and I have a new phone, and I needed to get a dongle that would connect the two.

好吧。

Okay.

再说一次，这并不是说不行。平均值不是大问题，但它能帮我解决这个问题。它帮我弄明白了。说实话，我不知道如果没有它，我该怎么在网上完成。

And again, this is not a No. Mean is not a big problem, but it would solve the problem for me. It helped me figure it out. It would've I don't know how I would've done it on the web, honestly.

谷歌本应为你做到这一点，但我完全同意。

Google should've done this for you, but I fully agree.

所以这对我来说很有价值。我愿意为此付费。

So that's valuable to me. I would pay for it.

但它真有那么值钱吗？因为记住，谷歌——你和其他人一样清楚——谷歌之所以盈利，不是因为它是一个伟大的搜索引擎，而是因为它掌握了如何销售广告、如何购买广告、如何投放广告，以及广告的搜索引擎。是垄断让它如此盈利。

Is it that valuable, though? Because remember, Google is you know this as well as anyone. Google is not profitable because they're a great search engine. It's because they own how to sell the ads, how to buy the ads, how to place the ads, the search engine for the ads. It's the monopoly that makes it so profitable.

并不是因为它搜索能力更强。它过去确实很棒。

It isn't the fact that it is better at search. It used to be that it was really good.

它在广告方面确实提供了很好的效果。是的，没错。因为你告诉它你对这个感兴趣，广告商可以竞价

It delivers really good results for advertising. Yes. Exactly. Because you're telling it, I'm interested in this, advertisers can bid

你会注意到没有人在这上面做广告。Perplexity去年有2万美元的广告收入。但在技术层面上，好吧，我不知道我是否同意它更好，但假设它是一个更好的谷歌搜索。

on it. And you'll notice that nobody is making ads on this. Perplexity had 20,000 of ad revenue last year. But on a technological level, okay. Let's I don't know if I agree it's better, but let's say it was a better Google search.

假设它只是取代了谷歌。但它能在收入基础上取代谷歌吗？这才是关键。没有人成功取代过谷歌。嗯。

Let's say it just replaces Google. But does it replace Google on a revenue basis? Because that's the thing. No one has succeeded in replacing Google. Mhmm.

微软没能做到。我意识到微软就是市场上的麦肯锡、憨豆先生。

Microsoft couldn't do it. And I realized that Microsoft the the McKinsey, Mr. Beans at the market.

但假设它成为了比现有任何方式都更好的互联网信息检索工具，也许它不会像谷歌那样拥有强大的盈利机器。但它仍然具有固有价值，很大价值。有些人愿意花大价钱购买。有些人则愿意支付极低费用或用免费的低配版。市场总是存在这种分层现象。

But let's just say it becomes a better way to retrieve information on the Internet than anything else that exists, and maybe it won't have as good a money making machine attached to it as Google does. But there's still inherent value, a lot of value. Some people will pay a lot of money for it. Some people will be fine paying next to nothing or just getting free crappy version. We always have those sort of dispersions.

重申一下，这仅仅是我们预期功能的冰山一角，但依然意义重大。而且这还是在我们今天停止改进的前提下。

Again, that's a fraction of what we've been promised it will do, but it's still meaningful. And that's if we stop if we then just stop making it better today.

更何况它还会产生错误幻觉结果，这个问题至今仍然存在。你会发现这个问题无处不在：谷歌有，必应有，ChatGPT有，各种产品都有。OpenAI自己也承认无法解决幻觉问题。

So on top of the fact that it also hallucinates bad results and this still happens to this day, And you'll notice that this problem is everywhere. It's on Google. It's on Bing. It's on ChatGPT, Corp, what have you. And is OpenAI themselves have said they can't solve hallucinations.

没错，这是整体特性的一部分。好吧。就算它是更好的搜索引擎。请指出靠这个赚钱的公司。我甚至不是说盈利。

They're, yep, part of the whole thing. Okay. It's an it's a better search engine. Point to the company making money off of that. I don't even mean a profit.

我指的是真正从中获得实际收入。你无法真正理解这1.5亿美元的年度经常性收入意味着什么？那到底是什么？

I mean, making real revenue off of that. And you can't really look you can't really perplex these 150,000,000 ARRs. What's that?

嗯，是的，我们现在就专注于当前明确的领导者，也就是ChatGPT，对吧？我的意思是，我不明白我们为什么要讨论Perplexity，但假设所有

Well, yeah, let's let's let's just focus on on the clear leader right now, which is ChatGPT. Right? I mean, I don't know why we're ever talking about perplexity, but let let's say that all of

我提到他们的原因是他们是一个真正的、根深蒂固的AI搜索引擎。

the The reason I mentioned them is they are an actual dyed in the wall AI searcher.

我知道。但没人用他们。

I know. But no one's using them.

这正是我的观点。

That's kind of my point.

不，不。但但但有一个明确的领导者。所以我们就专注于...让我们说其他所有都退场。我们只剩下OpenAI/chat GPT。

No. No. But but but the the the there's one clear leader. So let's just focus on Let's chat say everything else falls away. We're left with with OpenAI slash chat GPT.

当然。看起来他们似乎已经/将会构建出非常有价值的东西。

Sure. That seems like they have they will have built something very valuable.

但有多有价值呢？因为关键就在这里。他们需要所有这些数据中心吗？你是说我们应该...最终结果是什么？因为我知道你想抛开经济学，我理解，但谷歌可是个印钞机。

How valuable, though? Because that's the thing. How because do they need all these data centers? Are you saying we shall like, what what is the end result here? Because I know you wanna remove the economics, and I understand, but Google is a money printer.

嗯。它之所以能存活下来，就是因为这个。它也不是运行在极其昂贵的大型语言模型上。所以OpenAI创造了一个更好的搜索引擎。好吧。

Mhmm. It like, it survives because of that. It is also not run on large language models, which are incredibly expensive. So OpenAI has created a better search. Okay.

太好了。

Great.

它具有货币价值，但你说它永远不足以证明其估值合理，甚至可能无法覆盖成本。是的。这就是你所说的

It has monetary value, and you're saying it won't ever be enough to justify its valuation or even maybe cover its costs. Yes. That's where you've

从根本上说，我认为是的。我不认为有办法解决这个问题。我甚至不知道他们将如何在ChatGPT中投放广告。我确信他们会想出办法，但如何通过算法进行广告投放——因为这是唯一能规模化盈利的方式——这将是一个非常棘手的难题。而且一开始所有的广告预算都将是实验性的，因为他们需要证明其效果。

Fundamentally, I believe Yes. I don't think that there is a way out of it. I don't I don't even know how they will put ads within ChatGPT. Like, I'm sure that they will work out something, but how you do ad placement algorithmically, because it's the only way you scale it, it's gonna be a bloody difficult nut to crack. And then all of the ad budget at first is going to be experimental as they kind of prove it.

事实上，此前没有人真正在这方面成功过，这并不是一个好兆头，但除此之外，好吧，很好。你已经建立了一个搜索引擎。

The fact that no one has really succeeded in this otherwise is not a good sign, but on top of okay, great. You've built a search engine.

我的意思是，我们还有第二个例子，一个通过广告赚取巨额利润的大型互联网公司，对吧？那就是Meta。是的。你确实提供了很多关于自己的信息，但它实际上是将每个人在网络上的行为联系起来，这些行为来自不同的地方，内容五花八门。它再次为广告商提供了一个非常非常有价值的工具来触达用户。

I mean, we also have a second example of a hugely it's a huge Internet company that has made printing money from ads, right? And that's Meta. Yes. And yes, you've given it a bunch of information about yourself, but really it's sort of tether it's tying together what everyone does on the web, which is a lot of different stuff from different places. Again, made a very, very valuable tool for advertisers to reach people.

嗯。

Mhmm.

再说一次，也许这对OpenAI行不通。也许广告组件没那么有价值，但企业价值也很高。也许像我这样的人正在为此付费。

And again, maybe that doesn't work for OpenAI. Maybe the ad component isn't as valuable, but also there's a lot of enterprise value. Maybe people like me are paying money for it.

但根本没有企业价值。问题是他们在向企业销售方面做得不好。没人做得好，因为企业需要可复现性。SaaS要么是靠洗脑高管来运作——引用尼克·苏雷什的话——要么是靠可复现性和实实在在的美元。而且

But there's no enterprise value. The thing is they're not doing well selling to the enterprise. No one is because the enterprise needs replicability. SaaS is run on either brainwashing executives, to quote Nick Suresh, or it's run on replicability and real actual dollars. And

可复现性是指

Replicability meaning

意思是，你可以相信它能多次做同样的事情。

As in, like, you can trust it to do the same thing more than once.

不是说你必须反复购买。虽然你

Not not that you have to keep buying it over and over. Although you

不能。那是一个重要部分。那是很重要的一部分，但不是我所指的。

can't That is a big part. That's a big part of it, but not what I was referring to.

但是，比如，你可以一直问它二加二等于几，它总是回答四。

But Like, you can consistently say what it's two plus two and it says four.

你可以始终如一地说出它会做什么。我，这实际上就是整个问题的关键，真的。即使它99次都做对了——虽然它不会——但只要有那么一次，如果你允许它，我不知道，协助处理任何与你的代码相关的事情，那可能就不太妙了。而且你输入推理模型的代码越复杂，它在后端产生幻觉的可能性就越大，因为关键在于每次你试图让AI去做他们一直在吹捧的事情时。嗯。它搞砸了。

You can consistently say what it will do. I and that's actually the whole ballgame, really. You can't con even if it does the same thing right 99 times, which it won't, that one time, if you're allowing it to, I don't know, help with anything to do with your code, it's probably not great. And the more complex the code you put into a reasoning model, the more it's going to hallucinate on the back end because the point is every time you try and make AI do the thing they've been hyping Mhmm. It fucks up.

每次当你试图让你说，哦，它要重构整个代码库。那是你100%知道它会搞砸的事情。推理模型需要推理的内容越多，它们产生幻觉的情况就越严重。这些都是明摆着的事实。而你说，我理解为什么。

Every when you try and make you say, oh, it's gonna refactor a whole code base. That is the one you know 100% it's gonna fuck something up. Ghost reasoning models hallucinate more than the more that they have to reason. It's just this is all out there. And you say, I understand why.

你说，哦，好吧，随着时间的推移，会有赢家出现。问题在于，OpenAI堆积了这么多资金，即使他们的目标最终只是成为一个平庸的初创公司，现在压力也太大了。他们承诺得太多了。他们承诺了超过万亿美元。这太荒谬了。

You say, oh, well, over time, there'll be victors. The problem is with all the money that OpenAI is stacking on top, even if their goal was to eventually become just a mediocre startup, it's too much on top of them now. They promised too much. They promised over a trillion dollars. It's ridiculous.

所以你看这项技术，它已经平庸了，而他们已经做出了所有这些承诺。萨姆（Sam）做出了所有关于它能做什么的承诺。每个月，感觉他都得说，好吧，当我说AGI时，我的意思是软件。当我说软件时，我的意思是，它大概能做点什么。好吧。

So you look at the tech and it's already mediocre, and they've already made a and clammy set clammy Sammy's made all these promises. He's made all of these promises about what it can do. And each month, it feels like he has to go, well, when I said AGI, what I meant was software. When I say software, I mean, it kinda does something. Okay.

它其实并不真的做什么事，但你可以用它搜索。人们真的很喜欢它。我听说前几天有人用它解决了一个物理问题。我不会告诉你是谁，因为这其实……我认为AI时刻的某种特质，就是硅谷一切傲慢自大的集中体现。过度支出、英雄崇拜、那种几乎是被动的信念——无论何时科技界出现一个‘天选之子’，你知道，我们相信他说的任何话，即使这些话荒谬得离谱。

It doesn't really do stuff, but you can search on it. People really like it. I heard someone solve a physics problem the other day. I'm not gonna tell you who because I it's just this I actually think that there is something about the AI moment that is just the hubris of everything in Silicon Valley at once. The overspending, the hero worship, the kind of almost passive belief of whatever the chosen one, whenever there is a chosen one in a period of tech, you know we believe anything he says, even when it's egregiously silly.

糟糕的是，如果他们在，嗯，2022年出来时就说清楚，并且非常明确，比如，这里有哪些限制。这是实验性的。即使它病毒式传播了，也非常明确地说，嘿，我们不能信任它做这个。比如，我们必须……并且试图不以这种荒谬的速度扩张，也许故事会不一样。

And what sucks is if they'd have come out of this in, what, twenty twenty twenty two and said and were very clear, like, here are the limitations. This is experimental. Even if it went viral, being very clear clear to say, like, hey. That we can't trust it to do this. Like, we have to be and trying not to scale at this ridiculous rate, maybe it would be a different story.

但从一开始，一旦ChatGPT的病毒式传播效应出现，它就失控了。你有所有那些报道声称GPT-4在TaskRabbit上下了订单。它从未那样做过。

But from the beginning, it was this once the virality hit with ChatGPT, it ran away with it. You had all that coverage claiming that GPT four ordered a TaskRadit. It never did that.

本周的《Virgin历史》——The Verge关于旧技术的新聊天节目，我需要你们在脑海中回溯到每条短信收费10美分的时代。还记得黑莓推出能实时免费与所有朋友发消息的功能时，是多么轰动的大事吗？我指的当然是BBM，或许是有史以来最伟大的通讯应用。好吧，可能不是最伟大的，但无疑是非常重要的通讯应用。

This week on Virgin history, the verge's new chat show about old technology, I need you to go back in your minds to a time when text messages cost 10¢ each. And remember how big a deal it was when BlackBerry came out with a way to message all of your friends in real time for free. I refer, of course, to BBM, maybe the greatest messaging app of all time. Okay. Probably not the greatest messaging app, but a very important messaging app nonetheless.

本期节目中，我们将探讨它的起源、衰落，以及如果没有BBM，世界可能会如何不同。尽在YouTube上的《Version History》及所有播客平台。

This episode, we talk about where it came from, where it went, and why without BBM, the world might look really different. All that on Version History on YouTube and wherever you get podcasts.

婴儿应该拥有手机吗？这是Vox资深记者Adam Clark Estes最近向多位研究者提出的问题。这很个人化——他有个两岁的孩子。

Should babies have phones? That's a question senior Vox correspondent Adam Clark Estes asked a bunch of researchers recently. It was personal. He has a two year old.

我觉得在某些方面，我余生每天都会面临新的挑战。但就科技而言，最大的难点在于它不断变化，这些都是新挑战。我们对于正确做法没有明确答案。作为父母，这让人非常惶恐。

I feel like I'm gonna be facing a new hurdle every day for the rest of my life in some respects. But when it comes to tech, think the the big challenge here is that it is constantly changing and and these are new challenges. We don't have clear answers on on what the right thing to do is. As a parent, it feels very scary.

主流观点告诉我们：不行。让孩子远离屏幕。但有一位专家希望我们重新思考这个问题。

The prevailing wisdom tells us no. Keep that screen away from your child. But one expert wants us to rethink things.

这是我和妻子反复讨论思考的问题。实际上在两到三岁之间的某个阶段，我们决定给女儿她的第一部手机。

This is something that my wife and I talked and thought a whole lot about. And really kind of at some point between the ages of two and three, we decided to give our daughter, her first phone.

本周《Explain It To Me》主题：幼儿与科技。我们该怎么做？我们知道什么？每周日更新新集。请在您获取播客的平台收听——说不定就在您宝宝的手机上。

This week on Explain It To Me, toddlers and tech. What to do and what do we know? New episodes on Sundays. Find it wherever you get your podcasts, maybe even on your baby's phone.

数百名德克萨斯国民警卫队士兵已抵达伊利诺伊州，正准备在芝加哥部署。当地居民一直在抵制移民及海关执法局（ICE），他们甚至阻止国土安全部部长克里斯蒂·诺姆使用洗手间。

Hundreds of Texas National Guard troops have arrived in Illinois and are getting ready to deploy in Chicago. Residents there have been pushing back against ICE. They blocked DHS secretary Christy Noem from using the bathroom.

这就是普利兹克州长所说的合作与保障人民安全。

That's what governor Pritzker says is cooperation and keeping people safe.

然后实际上还有更多关于洗手间的事情。

Then actually even more bathroom stuff.

他们甚至不允许我们的ICE官员和边境巡逻官员使用洗手间和设施。

They don't even let our ICE officers and our border patrol officers use restrooms and facilities.

但这并不全是与洗手间相关的。

But it's not all bathroom related.

你会用那把枪对准自己的人民。真可耻。我希望此刻你的祖先正看着你。

You're gonna use that gun on your people. Shame on you. I hope right now your ancestors are looking at you. And

这种紧张局势，加上特朗普总统清晨呼吁监禁伊利诺伊州州长，引发了人们对接下来会发生什么的担忧。以上就是今天由Vox为您带来的每日解说。

this tension, combined with president Trump's early morning call for the governor of Illinois to be jailed, has raised fears about what is coming next. That's on today explained from Vox every weekday.

我认为有两件不同的事情正在发生，对吧？一个是炒作周期和人们的从众行为，那是硅谷的现象，但也是全世界的普遍现象。对吧？

I think there's two different things going on. Right? There's a hype cycle and people moving in herds and that's Silicon Valley, but it's also just the world. Right?

确实。

Sure.

那就是那个麻省理工学院的研究吧？所以，你知道，95%的公司投入这些东西却搞不清楚自己在做什么。这并不让我惊讶。就像，如果你在企业里，你大概知道，哦，这是我们必须做的新鲜事。我们被告知必须这么做。

That's that m I was it MIT study? So, you know, 95% companies who put this stuff in can't figure out what they're doing. And then, that doesn't surprise me. Like, you just if you and in businesses, you sort of know, like, oh, this is the new thing we have to do. We've been told we have to do it.

我们会去买它。我们不确定为什么要这么做。再次，这与.com泡沫时期有相似之处，当时

We'll go buy it. We're not sure why we're doing it. Again, there's parallels to the .com boom where

我们什么都想要。

we want Everything.

每个人都必须有一个网站

Everyone had to have a website

却没人真正使用。这种情况持续了好几年。

that no one do. This for years.

是的。所以情况就是这样。但话说回来，我觉得这只是人类正常的行为表现。另一个问题是，它是否真的实用？也许它没有承诺的那么强大，也许它不能完成某些特定任务，但我认为它确实有价值。

Right. So there there's that. But again, I I think that's just humans behaving as humans do. And the other is, is there actual utility? And maybe it's not as promised, and maybe it can't do task x or y, but it definitely has value, I think.

我发现它确实在帮助我写作。它帮我查找资料。我用它来碰撞想法。我知道它的工作原理，也明白它并不具备真正的推理能力。

I am finding it like help it is helping me write. It's helping me find stuff. I use it to bounce ideas off. I know what's going on. I know that it's not reasoning.

我知道它基本上就是在给我输出填空游戏般的答案。但大多数时候它还是有用的。我也经常发现它的局限性所在。比如，它从来不愿意承认自己不知道。它永远不会说'我不知道'，你明白吗？

Know that it's spitting out basically a mad libs at me. But more often than not, it's useful. I'm also finding, you know, frequently where the limits are. Like, it it never wants to tell me it doesn't know. It'll never you know?

如果它无法访问彭博社的文章，它不会直接告诉我这一点。它会告诉我其他人对这篇文章的看法，并把这些当作自己的观点呈现。但通过这些使用体验，我逐渐明白了：好吧，这个功能我不用它，那个功能可以用。它确实具有实用价值。

If it doesn't have access to a Bloomberg article, it won't tell me that. It'll just tell me what other people are saying about the Bloomberg article and present it as it's and but these are sort of things that I, as I'm playing with it, I'm figuring out, alright. Well, I won't use it for that. I'll use it for this. There's real utility.

如果今天就把ChatGPT从我身边拿走，我的生活会变糟吗？可能也还好，但我确实在使用它。我喜欢它。没错。

If you took my chat GPT away from me today, would would my life get worse? It could probably be okay, but I'm I use it. I like it. Right.

这并没有什么错，我也不是说没有像你这样的用户。如果它当初就是这样宣传的——嗯——就像个可爱的知识引擎，不管你怎么称呼它。如果当初是这样推广的，大家都说'对啊，我们有个知识引擎'。

And there's nothing wrong with and I'm not saying that there are not people like you. If that was how it was sold Mhmm. It was like this kind of cutesy knowledge engine, whatever you call it. If this was how it was being put out there and everyone's saying, yeah. We got a knowledge engine thing.

那样的话几乎无可指摘，但实际营销方式并非如你所说。MIT的研究中有句很值得玩味的话（人们都不爱讨论）：'企业AI采用率很高，但颠覆性很低'。因为关键在于，它并不是按你描述的方式在营销。你描述的是你发现它的实用方式，我听过很多人都这么说。这没问题。

It would be this in a it would almost be inoffensive, but it's not being sold how you're saying. And that MIT study had a really interesting line that people don't like talking about where they there was one line where it said enterprise AI adoption is high, disruption is low. Because the thing is, it's not being sold as you're describing. How you're describing it is how you're finding it useful, and I've heard multiple people say that. Fine.

它被宣传成包治百病的灵丹妙药。嗯。它被塑造成新一代的iPhone加上新一代的企业级SaaS加上新一代的亚马逊云服务——所有人都想分一杯羹。就像我说的，这是种荒谬的自大时刻，所有人都相信单一事物能改变一切。糟糕的是，正如你所说，每个周期人们都会问：我们的元宇宙战略是什么？

It's being sold as this panacea of everything. Mhmm. It is the new I it actually LLMs got set up to fail because it's the new iPhone plus the new enterprise SaaS plus the new Amazon Web Services that everyone, everyone gets in on this. It's like I said, it's this egregious hubris moment where everyone believed one thing could change everything. And it sucks because to your point, yeah, every single cycle, everyone goes, well, what's our metaverse strategy?

我们曾经说要加入区块链，但我从未见过他们像这次这样全情投入。因为这次提供了实实在在的东西——有API接口，有办公套件，有能添加到Microsoft 365账户的订阅服务。CEO们只需掏出信用卡就能假装在这方面有所作为。所以我们现在正处于那个阶段。

We're we're gonna put blockchain in this, but I've never seen them jump whole hog into this. Like, this because this gave everyone something there's an API. There's an office there's a subscription you can add to your Microsoft 365 account. This is you can get your credit card out and really pretend as a CEO you're doing something with this. So we've had that moment now.

当你说它不擅长其他功能，缺乏规模化企业级应用时——微软已经从OpenAI模型转向Anthropic，因为它连他妈的数据表格都生成不了。做不了PPT。这本该是它的强项，结果却做不到。关于编程大模型的争论更是没完没了。

And when you say, oh, it's not as good at other things, it doesn't really have scaled enterprise functionality. Microsoft has moved on from OpenAI models to Anthropic because it can't generate bloody Excel spreadsheets. It can't generate PowerPoints. The one thing you'd think this does, and it can't do it. The coding LLMs, that debate is going on and on.

有些人宣称这是未来，也有大规模误报说它会取代程序员。但当你实际询问使用者时，即便是最热衷的人也会说：没错，我得盯着它。嗯。必须随时检查输出结果。

You've got some people who claim it's the future. You've got misreporting at scale saying it's replacing coders. But when you actually talk to people using this, even the excited ones are saying, yeah, I have to keep an eye on it. Mhmm. I have to look at it.

而且所谓'氛围编程'也是谎言。但问题在于：如果它真如宣传所说倒也罢，可现在的宣传根本是神话。是虚构的传说。

And I'm vibe coding is a lie as well. But that's the thing. If it was being sold as the thing you were talking about, that'd be one thing. What it's being sold as is myth. It's mythology.

宣传的一直是它'可能'做到什么——这种'可能'状态已经持续三年了。我们仍停留在'可能'阶段。何时才能变成'可以'？永远不可能，因为我们已经触达极限。触及了规模化的天花板。

It's what this could do, and it's been what this could do for three years. We're still at the could do stage. When will it do? And it won't because we are at the limits. We're at the scaling limits.

这个瓶颈期已经持续很久了。我们进入了训练收益递减阶段。撞上了南墙，而且...我不知道。OpenAI不会退缩，他们绝不会做那些本该做的——符合伦理的事情，永远不可能。

We've been at them for a while. We've hit the diminishing returns period of training. We're at the wall, and it's not I don't know. I don't think OpenAI will back off this. I don't think that they're gonna the ethical thing for them to do, which they'd never do ever.

从来没有人这样做过，意思是说，好吧。让我们冷静一点。不。实际上他们是在说，不。我们需要更多的喷气式飞机。

No one's ever done this is to say, okay. Let's cool our jets a bit. No. They're in fact saying, no. We need more jets.

更多的喷气式飞机。火箭。现在就要燃料。

More jets. Rocket. Fuel now.

所以我们先别讨论这个了，因为我们实际上在辩论。我们有点在辩论关于关于效用的问题。我认为我们

So let's let's let's move off the because we're really having a debate. We're kind of having a debate about about the utility. Think we're

大致上意见一致。

kind of on the same page.

而且，嗯，是的。我的意思是，我从中得到的比你多。就我而言，我深深担忧我们的经济建立在这东西之上，我也担心数据中心对我来说似乎是个真正的问题，因为它们只进入那些没有其他经济的地方，而且它们也会消失。钱的问题。但我们先搁置这个吧。

And the well, yeah. I mean, I'm I'm getting more out of it than you are. And as far as I'm deeply worried about our economy being based on this thing, and I'm also worried about the the the data centers to me seem like a real problem because they're only going into places that have no other economy, and they're also gonna go away. Money. But let's table that.

好吧。我们不要——我们已经搁置太多事情了。我只是——我们有——我们有——因为我们，你一再提出这个论点。人们总能听到你提出这个论点。我们有没有很好地总结你的——你的

Okay. Let's not we've tabled too many things. I just do have we have we have we because, again, you you make this argument. People can hear you make this argument all the time. Have we done a good job of summing up your your

数据中心的事情很重要。所以最大的误区是存在一个销售AI计算资源的行业。根据我的计算，除了微软、Meta、亚马逊和OpenAI之外，计算收入不到十亿美元。理论上，我们有数千亿美元的私人数据中心正在建设中。我们有Meta创建了一个特殊目的实体，安然风格的小宝贝，一个表外数据中心交易。

The data center thing is important. So the big myth is that there is an industry selling AI compute. From my calculations outside of Microsoft, Meta, Amazon, and OpenAI, there's less than a billion dollars of revenue in compute. We have hundreds of billions of dollars of private data centers being built theoretically. We have Meta creating a special purpose vehicle, Enron style baby, of an off balance sheet data center deal.

我们建了这么多数据中心，结果都白费了。除了超大规模企业把资产从资产负债表上移除外，我们还建了这么多东西，而且私募股权也大量涌入。

We have all of these data centers being built for nothing. There is outside of hyperscalers moving stuff off of their balance sheet. We have all of these things being built, and we have private equity up the wazoo.

所以当马克·扎克伯格告诉唐纳德·特朗普，我们要在路易斯安那州投资500亿美元之类的。

So when Mark Zuckerberg tells Donald Trump, we're investing 50,000,000,000 in Louisiana or whatever.

是的。

Yeah.

就算这个数字没错，这笔钱会流向哪里？所以这

Even assuming that is correct, that money goes where? So this

就是疯狂的地方。我不确定是不是路易斯安那州那个项目，但Meta和Blue Owl Digital——我相信他们也是Crusoe项目和OpenAI阿比林项目背后的支持者——他们正在设立一个特殊目的实体，Meta会往里面投钱，我认为他们会发行债券来融资。Meta将拥有这个SPV，但在数据中心建成后还要支付租赁费用。这就是

is the crazy thing. I don't know if it's the Louisiana one, but Meta and Blue Owl Digital, who are also, I believe, behind Crusoe's thing and Abilene for OpenAI, They are doing a special purpose vehicle where Meta and then will put money into this thing, selling I think they sell bonds to fund it. Meta will own the SPV, but also pay a lease on the data center once it's built. And it's this

整个...我想说的是他们要建个仓库。

whole I guess I was getting at they're gonna put up a warehouse.

没错。里面堆满了芯片。

Yes. Full of full of chips.

堆满了芯片。所以有些人会拿到钱去建仓库。

Full of chips. So some people will be paid money to build the warehouse.

没错。

Correct.

还有一些人，数量少得多，会拿到钱四处巡视，拿着手电筒确保空调开着。然后大部分价值，对吧，都流向了英伟达，或者卖硬件的人。

And some people, a much smaller number, will be paid to sort of walk around with flashlights and make sure the air conditioning is on. And then most of the value, right, goes to NVIDIA, whoever's selling Correct. The the hardware.

是的。这笔交易是260亿美元还是240亿美元？

Yes. And 26,000,000,000 or $24,000,000,000 is the deal?

所以听起来这像是路易斯安那或其他地方的一个巨大经济发展项目，但实际上钱很少会流到那里。

So it sounds like this is a giant economic development project for Louisiana or wherever, but it's really very very little of the money is going there.

事实上，我几周前和一位州监管人员谈过。这很大程度上是西部在掠夺南方。就是把南方塞满了GPU，收入却没多少，真正的工作岗位也不多，既不是建设阶段也不是运营阶段，因为数据中心并不需要大量人力。嗯。就是这样。

In fact, this is I was talking to a state regulator a few weeks ago. It's it's very much the West stealing from the South. It's just filling up the South full of GPUs with revenue, like, no real a bunch of jobs ongoing there, not for the construction and not for the because data centers are not superhuman heavy. Mhmm. And that's the thing.

这些东西被宣传成巨大的就业创造者和经济推动力，但其实不是。它只是帮助私募股权把大量资金投入一个永远不会亏损的项目。但我想简单点说：AI数据中心不赚钱，它们亏很多钱。

These things are being sold as this massive jobs creator in economic and it's no. It's just it is helping private equity sink a bunch of money into something they can never lose. But I really wanna make this simple. AI data centers do not make money. They lose a bunch of money.

据我所读，GPU最多只能撑三年就会报废。如果没有客户，这一切又如何运作？这怎么行得通？而且，私募股权传统上会从资产中退出。他们打算卖给谁？

GPUs die, from what I can read, at best in three years. How does any of this work out if there are no customers too? How this work? Also, private equity traditionally moves on from assets. Who are they gonna sell them to?

没人想买。每个人都在建数据中心。每千兆瓦需要两年半时间。现在到底发生了什么？就好像我们在毫无理由地建一堆数据中心。

No one wants to buy. Everyone's building data centers. They take two and a half years per gigawatt. When it what's happening? Like, we're building a bunch of data centers for no reason.

我们马上回来，但首先插播一条赞助商广告。好了我们回来了。你刚才简要概括了你的论点。是的。在更广泛的范围内，提出这个论点进展如何？

We'll be right back, but first, a word from a sponsor. And we're back. You have made an abbreviated version of your argument. Yes. How is making that argument going for you in the wider world?

谁谁对你反应积极？谁在关注？有谁的反应方式让你重新思考你的任何论点？

Who is who is responding positively to you? Who's paying attention? Who has responded in a way that's making made you rethink any of your arguments?

所以总的来说，我有非常亲密的伙伴，比如Kagawa，还有我的编辑Matt Hughes。有很多人经常反驳我的论点。就像，我我希望这样发生。我想提出根本上有说服力的论点。我会说一年前，人们人们觉得我疯了。

So something I have generally I have really close my mate case, Kagawa, for example, Matt Hughes, my editor. I have plenty of people who push back on my arguments constantly. Like, it's I I want that to happen. I wanna make fundamentally sound arguments. I would say a year ago, people people thought me mad.

人们因为我写了OpenAI如何在2024年6月生存下来？那是经济上的一个大问题，比如，嘿，他们正在亏损50亿美元。这可不好。我会说在那之后的几个月，很多人说我反应过度了。

People because I wrote OpenAI how does OpenAI survive back in June 2024? And that was the the big economic one of like, hey. They're losing 5,000,000,000. That's not good. And I would say in the months after that, I had a lot of people saying, you overreacting.

这些这些公司将会起飞。会非常惊人。Edge，你疯了。现在人们说，我会说深度求索（DeepSeek）算是转折点，但真正是在几个月前当

These these companies are gonna take off. It's gonna be amazing. Edge, you're crazy. Now people are saying and I would say deep seek was kind of it, but really it was a few months ago when

那是中国的。是的。

It was the Chinese Yeah.

在一月份。嗯，那甚至不是山寨版。只是训练模型的成本更低。实际上最终花费少得多，但不管怎样，真正是几个月前，随着《AI泡沫的黑粉指南》的出现，人们才觉得，哦，糟糕。他说得对。

In January. Well, it wasn't even knock off. It was just it cost less to train the model. Actually ended up costing way less, but regardless, it was really a few months ago with the haters guide to the AI bubble that people like, oh, crap. He's right.

但其实是保罗·库德洛夫斯基，我觉得是他。是他打破了这个局面，因为他有个很棒的故事。说的是AI数据中心的发展在今年上半年带来的经济增长超过了所有消费者支出的总和。一旦这事发生，人们就开始关注了。但我的听众和读者群体非常多元化，有与科技无关的普通人，有技术背景很强的人，还有越来越多的金融界人士。

But really Paul Kudrowski, I think, was the guy. He was the one that broke it because he had this great story. It was for it was something like AI data center development added more economic growth than all the consumer spending combined in the first half of this year. Once that happened, people started to listen. But I have a really diverse listener and reader base of regular people who are nothing to do with tech, very technical people, and a growing amount of finance people.

我听到不少来自私募股权公司和对冲基金等的反馈，就是说，比如，我读了这篇文章。这很好。随着时间的推移，他们从偶尔给我发几条消息，像是‘我不知道’，但现在变成了‘天哪，这让我想起了北电网络’。

Hedge I hear a good amount from private equity firms and hedge funds and stuff just saying, like, I read this. This is nice. And over time, it's gone from them sending me mess a few messages here and there being like, I don't know. But, you know, what if what is the economic opportunity if this works to now saying, oh god. This reminds me of Nortel.

我不喜欢北电网络。请别提醒我

I don't like Nortel. Please don't remind me

Lucid。因为我有点怀疑，是不是你击中了某种情绪，因为总有人想听那些看似很大的东西其实很小。对吧？总有人对此感兴趣。显然，反企业主义的浪潮

of Lucid. Because part of me was wondering is, like, if you're hitting a chord one, just because there's always gonna be people who wanna hear that the thing that seems big is actually small. Right? There's just always an interest in that. There is obviously a swell of anti corporatism

嗯。

Mhmm.

没错，这体现在很多方面，有些是不好的，有些可能是健康的。但你的意思是，不仅仅是那些本能地反对大型科技公司或大企业甚至金钱的人，还包括那些对技术和金钱感兴趣的人

Right, that manifests in a lot of ways, some of which are not good, some of which are maybe healthy. But you're saying it's not just people who are reflexively against big tech or big companies or even money. It's it's people who have an interest in technology and money

是的。

Yes.

正在关注你。你提出这个论点是在赚钱吗？

Are paying attention to you. Are you making money making this argument?

嗯，我是由Our Hurt Radio支付的，我还有一个付费通讯。播客？是的。嗯。我有一个付费通讯，做得相当不错。

Well, I'm paid by Our Hurt Radio, and I have a premium newsletter. Podcast? Yes. Mhmm. And I have a premium newsletter, which is doing pretty well.

数量相当可观

Got a good good amount

像个幽灵。

of a ghost.

幽灵。但我有免费通讯，还有一个我现在每周做的付费版。关键是，我知道我确实有一些人是这样的，是的。我反企业。是的。

Ghost. But I have the free newsletter and a premium one I do weekly now. And the thing is, I know I definitely have people who are like, yeah. I'm anti corporate. Yeah.

我不喜欢这些科技公司。绝对不喜欢。但我的基本面分析很深入。我不是随便出去就说，去他妈的这些公司，虽然我确实这么想。我是在深入剖析论点，进行深入的财务分析，深入比较——我大概能记得所有提到OpenAI收入或Anthropic收入的新闻标题。

I don't like the tech companies. Absolutely. But my fundamental analysis is deep. It's not like I'm just going out and saying, fuck these companies, which I am. I am pulling apart the arguments in deep financial analysis, deep comparison of I think I can remember roughly what every information headline that has mentioned OpenAI's revenue or Anthropix revenue at this point.

我对这些东西钻研得很深。所以人们被吸引过来，因为我觉得有些人通过这个在学习金融知识，这挺吓人的。我自己在做这些事的时候当然也在学习金融，因为我没受过经济学训练。但人们被吸引的点在于，我这种深入的方式我觉得没多少人能做到。我觉得如果分析师们都这么做，大家会愿意读更多分析。

I go very deep in this stuff. And so people are attracted to it because I think some of them are learning finance through it, which is scary. I'm certainly learning finance while doing this because I have no economics training. But it's people are attracted to the fact that I go deep in a way that I don't think many people do. I think people would love to read more if analysts did this stuff.

我意识到一篇18000字的博客可读性有点离谱，但我写的风格更平易近人。

I realize more readable with an 18,000 word blog is kind of ridiculous, but I write in a style that's more approachable.

挑衅又炫富

Provocative and swarry

而且娱乐性，我是说，但这是

And entertain and, I mean, but it's

也是我说话的方式。故事性。

also how I talk. Story.

对。但这不是表演。这就是我

Yeah. But it's not it's not an act. It's just how I

我的播客之所以好是因为你有英国口音。哦，是的。美国人喜欢英国口音。所以有一个播客，还有一个付费通讯。

how it comes out my Podcast is good because you got the English accent. Oh, yeah. Americans love English accent. So there's a podcast. There's a paid newsletter.

你能扩展这个业务的其他部分吗？你想扩展这个业务的其他部分吗？除了做这些面向公众的事情，你还能做私人顾问吗？

Can you build out other parts of this business? Do you wanna build out other parts of this business? Could you be a private consultant in addition to doing this public facing stuff?

是的，可能吧。我的意思是，有人给过我咨询的工作，就像只是和他们聊聊。但在我做过的少数几次中，我说话的方式完全一样。这没什么不好处理的。

Yeah. Probably. I mean, I've had people that have given me, like, consultancy gigs, like, just talking to them. But within the small few I've done, I don't I speak exactly the same way. There's no bad handling with that.

老实说，这就是去年和我交谈的企业人士的最大不同。他们不想听这些。我在一些人面前做了一个小组讨论，我在批评人工智能，他们真的不高兴听到这些。你能看出他们像是，妈的，这家伙是谁？

And honestly, that's been the big difference with the corporate people that I talked to last year. They didn't wanna hear this. I did a panel in front of some people, and I was talking shit on AI, and they were really not they were not happy to hear it. You could tell that they were like, fuck. Who's this guy?

最近我做了一次，他们对此持开放态度，并且实际上感到震惊，但是一种'谢天谢地你告诉我'的方式，而不是'滚出我家'的方式。但我认为咨询可以是其中的一部分，但我真的很喜欢写作。我真的非常非常非常享受这个。我喜欢做广播，所以我做得很开心。这对我意义重大。

More recently, I did one, and they were open to it, and they were actually alarmed, but in a thank God you told me way rather than a get out of my house way. But I think that consultancy could be part of it, but I just really love writing. I really I really, really, really enjoy this. I enjoy doing broadcast, so I really have fun doing it. It means a lot to me.

你有一个公关业务。我有。提到过好几次了。那仍然是一个业务。你还在运营它。

You have a PR business. I do. Mentioned several times. That's still a business. You're still running it.

那么，你花在做公关上的时间占多少比例，对比

So between how much of your time are you spending doing PR versus

我大部分时间仍然在做，但我把它隔离开来。我不涉及相同的主题。

Still the majority of my time doing But I firewall it off. I don't cover the same subjects.

你是否想过，在某个时候，公关工作会消失，这会变成你的全职工作，写作、交谈和思考？

Do you Imagine that at some point, PR goes away and this becomes full time for you, writing and talking and thinking?

我真的很享受写作，我越来越...我无法告诉你我有多开心。我真的很爱做这件事。感觉这就是我天生该做的事，它让我...我就是真心喜欢做这个。我很乐意去做。

I really enjoy the writing the I am having more I I can't tell you how happy I am. I really love doing this. It feels like what I was born to do, and it makes me I just genuinely, like, love doing it. I would love to do it.

我很乐意每周写多份付费通讯。还有这个'喷火器真相揭露者'的形象，也就是你本人。是的。这在公关界是怎么运作的？你有没有客户会说，你在外面批评AI很棒，但别忘了你需要...我需要你帮我完成我雇你做的任务。或者，顺便说一句，我不会雇你，因为我不喜欢你那么高调，我不想让Ed这个品牌代表我。

I would love to write multiple paid newsletters a a week. And this flamethrower truth teller persona, which is also you Yeah. How does that work in the PR world? Do you have clients who are like, that's great that you're out talking shit about AI, but let's remember you need I need you to help me do this task that I've hired you for. Or by the way, I'm not gonna hire you because I don't like that you're so out there, and I don't want Ed, the the the brand, to represent me.

我想要嗯。一家公共关系机构来代表我。

I want Mhmm. A public relations agency to represent me.

说实话，所有客户都喜欢这样。他们因为我像烟花一样引人注目。我不涉及我写的东西，因为那会很糟糕。我

Honestly, across the board with clients, they like it. They because I fireworks. I'm not covering the things I write about because that would be horrible. I

你不...你不做任何不。

You don't you don't do any No.

我过去做过一些人工智能相关的工作，但不是像基础模型公司那样的。我坚决远离那些。不是说我觉得OpenAI或Anthropic会来找我，而是我不与大科技公司合作。我不与AI模型公司合作。就是这样。

I've done some AI I've done AI stuff in the past, but it's not like foundation model companies. I keep the fuck away. Like, not that I think that OpenAI or Anthropic is calling, but I do not work with big tech. I don't work with AI model companies. Like, it's just

是因为你不会接受还是他们不提供，所以不行？

would not because you won't take it or they're not since it's not on offer?

接受它。就像，那简直是地狱。好吧。但问题是，我确实非常热爱写作，也喜欢公关方面的工作。而且客户想要真相。

Take it. Like, it's just it would be hell. Okay. But the thing is I I do love the writing so much, and I love the PR stuff. And also clients want the truth.

他们不想被抓包。是的。偶尔会遇到想被哄着的客户，但他们想知道现实生活中到底发生了什么。

They don't want to be caught yes. You get the occasional one that wants to be coddled, but they wanna know what's actually happening in real life.

所以，想必有很多人听你这么说后会想，这不是我想合作的人。但你说还有其他人会说，绝对没错。这正是突破噪音的方式。我喜欢这一切。

And and and so presumably, there's a lot of people who hear this from you and go, not the guy I wanna work with. And but you're saying there's others who say, absolutely. This is the this is breaks through the clutter. I like all that.

这是一个重要的事情。实际上有两部分。一个是心碎的浪漫主义部分。我这么做不是因为我想唱反调或表现得刻薄。

This is an important thing. There's really actually two parts. One is the brokenhearted romantic thing. I don't do this because I'm like, I want me to be a contraire. I'm a be nasty.

不。我这么做是因为我对他们搞砸计算机的方式感到愤怒。但第二点是我懂行。如果我仅仅是个唱反调的人，只会说‘呃，我讨厌这个’，进行非常肤浅的分析，那就太糟糕了。

No. I do it because I'm pissed off at the way they fucked up the computer. But the second thing is I know my shit. If I was just a contrarian just went, I'm like, ugh. I hate this with this really base level analysis, that would suck.

但我知道我在说什么。我费尽心思深入探讨，因为不这样你就无法如此尖刻。你无法如此批判而不具备实际能力，否则你只是个混蛋。你只是在那里发怒。不，我有根本的理由认为OpenAI是个糟糕的公司。

But I know what I'm talking about. I take great pains to go in-depth because otherwise, you can't be this acerbic. You can't be this critical without being able to actually play because otherwise, you're just an asshole. You're just going up being mad. No, I have fundamental reasons I think OpenAI is a wretched company.

我真的这么认为。而且这是经济层面的原因。是那些做出的承诺。是文化问题。这些都是真实具体的事情。

I really do. And it's an economic one. It's the promises made. It's the culture. There are real tangible things.

是的，我确信有些人没和我共事过。

And yeah, I'm sure there are people that haven't worked with me.

如果你对我们所处的这个时刻判断正确。那么，清算何时会出现？

If you are right about the moment we're in Right. When when does the reckoning show up?

所以我认为可能很快。现在是十月份。这时OpenAI应该停止向CoreWeave付款。CoreWeave，我可以花一个小时来谈论它。但我不会。

So I think it could be very soon. So we're in October right now. This is when OpenAI is meant to stop paying CoreWeave. CoreWeave, I could spend an hour talking about. I won't.

他们是一家新兴云服务商。大客户是OpenAI和微软。CoreWeave建立在债务之上。如果那笔钱出任何问题，比如CoreWeave没有建设产能，CoreWeave可能会倒闭。这些公司任何财务问题，任何资金短缺的传闻，他们融资困难，任何关于未来的担忧都会动摇英伟达，因为这真的会

They're a Neo Cloud. Big customer is OpenAI and Microsoft. CoreWeave is built on debt. If anything goes wrong with that money, such as CoreWeave not building the capacity, CoreWeave could die. Anything financially with any of these companies, any of stories bubbling up that they're running low on capital, they're having trouble raising, any worries about the future that rattle NVIDIA, because it really is gonna be

你认为这可能在今年内爆发吗？

You think this could crack within this calendar year?

也许吧。问题是你不知道会有什么结果，因为我要告诉你的一件事是

Maybe. The thing is you don't know what's gonna come out because the one thing I'll tell

从财务泄露的信息来看，这些公司运营得非常糟糕。Ed，如果财务崩溃，你认为一年后我们会处于什么境地？所有技术都消失了，还是我们只剩下一些有用技术的残余？

you from the leaks about financials is these companies run terribly. Ed, if it cracks financially, where do you think we are in a year? All the tech is gone or we end up with remnants of some useful tech?

我认为当OpenAI崩溃时——我确实认为这会发生，可能需要几年时间——可能就只剩下一个邮箱地址在起诉别人。Copilot将成为ChatGPT，它将变成一个被阉割的、极度安全的版本。你将无法与这些东西进行有趣的对话，因为当它们精简运营、不再把它们当作烧钱机器、试图让它们变得稍微负担得起时，运行成本太高了。它们只会受到限制。

I think when OpenAI collapses, which I do think would happen, and it could take years. There could just be a PO box suing people. Copilot will be ChatGPT, and it will be this neutered, ultra safe one. You're not gonna be able to you're not gonna be able to have saucy conversations with these things because they're so expensive to run when they just condense these operations, when they stop treating them as money burners, and they try and make them somewhat affordable. They're just going to be restricted.

也许每天几次搜索你可以付费使用。我认为编程LLM会消失。我不知道是否在一年内，但它们的运行成本太高了。你会看到其中一些逐渐消亡，然后就是整合，这些能力慢慢降低，价格慢慢上涨，直到如果你有足够的钱、愿意支付，它们仍然存在，但这真的取决于吓到所有人的那个时刻有多严重。但总会有那么一个时刻。

And with a few searches a day, perhaps you'd be able to pay. I think the coding LLMs go away. I don't know if it's in a year, but they're so expensive to run. You're going to just see a few of them die, and then it's just gonna be a consolidation and a slow lowering of the abilities of these things and a slow raising of prices to the point that they'll be out there if you've got enough money, if you're willing to pay, but it really it depends on how egregious the moment is that scares everyone. But there will be one.

可能会发生一些与金钱有关的事情，然后大家就慌了，可能是某个AI明星公司。

There's gonna be something with money probably that says, okay. Everyone shits their pan it could be one of the AI stars.

你说得对。你因为先见之明、提前预测并说'我早就告诉过你了'而得到什么奖励呢？

You are right. What is your reward for being for being prescient and out there and saying, I told you so?

我现在正在收获回报。我说我非常热爱做这件事是极其认真的。我的意思是，我可以抽一支大雪茄，发一些表情包。我会从中找点乐子。但真正的回报是我已经建立了一个案例。

I've I'm reaping the rewards right now. I'm deadly serious when I say I love doing this so much. I mean, I get to smoke a big cigar, and I post some memes. Like, I'll have some fun with that. But the payoff is the fact that I have built a case.

过去两年里我写了五十万字，可能还不止。我能看到、能讲述完整的故事，这很棒。而且，是的，可能会引起一些关注。我已经有了一个实际运作的业务。我觉得每个人都想要那种大圆满的感觉，比如你会做这个做那个。

I've written half a million words in the last two years, probably more than that, actually. I get to see the I get to tell the full story, which is great. And, yeah, there'll probably be some attention. I'm already like, I already have, like, a functional business out of this. Everyone want I think everyone wants a big satisfying, oh, you're gonna do this and that.

完成这些之后，我也在关注其他公司。比如微软这样的公司。哦，我还在考虑微软的一些事情。

After I'm done with them, I'm looking into other companies too. There are other companies that have like Microsoft, for example. Oh, I've got some looking at Microsoft to do.

你可能有没完没了的公司和科技清单可以研究。而且，还有一些我们现在还不知道的事情，五年后会变得很重要，因为人们会为之兴奋。这就是行业的本质，到时候你也会做同样的事情。

I will probably You unending have list of companies and technologies that you can poke at And also, there's things that we don't know about yet that will be a big deal five years from now because people will get excited about them. It's the nature of the beast, and you will do the same thing then.

我还一直在教普通人金融知识。每天收到很多邮件，人们说他们从一无所知到学会这么多。这真的很酷。单是这一点就非常值得。我会来份大份的烤肋排。

I also wanna and I have been teaching regular people about finance stuff. I get a lot of emails every day with people saying, I don't know anything, I've learned all this. It's really cool. That alone is really rewarding. I I'll have a big prime rib.

我大概会这么做。关键是我是个简单的人。我喜欢做的事情，而且我觉得人们可能认为这是一时风光，我是为了成名才做这个。但我真心享受其中的每一刻，所以我会继续做下去，可能会有更多订阅者，这样就能持续下去，这真的很有趣。

That's probably what I'll do. The thing is I'm a simple man. I I enjoy the things I do, and I really get to I think that people may think that, oh, this is a flash in a pan moment. I'm trying to do this to fame. I'm genuinely having the time of my life doing this, so it's like, I'll keep doing more of it, and I'll probably have made I'll have more subscribers so I'll be able to keep doing it, which is really fun.

我真的无法用言语表达我每天做这件事感到多么幸运。太有趣了，明天我还要和David Roth（The Verge的Victoria Song）、Defector的David Roth、小Edward Ongoiso一起做节目。能和他们做一个小时的播客，和朋友们做很酷的播客，谈论我最热衷的事情。

I really do I I cannot say enough how, like, I genuinely feel so lucky to do this every day. It's so much fun, and I get to talk with my I'm doing a show tomorrow with David Roth, Victoria's song from the verge, David Roth from defector, Edward Ongoiso junior. I get to do a podcast with them for an hour. I get to do a pod cool podcast with my friends. I get to talk about the thing that I'm most passionate about.

来宣传一下你的节目吧

Let's plug your show

在我们离开前，还有你的通讯录。只需访问better offline.com获取一切信息。

and your newsletter before we leave. Just better offline.com for everything.

你的红色帽子在哪里？那是通讯录吗？

Where's your red hat's the newsletter?

你想雇佣Ed吗？做你的公关？

Do you wanna hire Ed? Do your PR?

访问easypr.com。而且，是的，我们免费和付费订阅者加起来接近8万人了。相当不错。

Get off at easypr.com. And, yeah, we're nearly at 80,000 free and paid subscribers combined. Pretty great.

也许有一天你会掌管CBS新闻。是的。哇。我们暂且搁置这个话题吧。要是能成真就好了。

Maybe you'll be running running CBS News one day. Yeah. Wow. Let's table that conversation. If only.