本周AI动态：OpenAI寻求纾困/谷歌与苹果10亿美元合作/TPU崛起

所有人都在问，我们是否身处泡沫之中？

Everyone's asking, are we in a bubble?

这个泡沫何时会破裂？

When is the bubble going to pop?

在一切崩塌之前，AI能发展到多大规模？

How big can AI get before it all comes crumbling down?

而昨天，OpenAI的首席财务官可能已经透露了这种情况发生的方式和原因。

And yesterday, the OpenAI CFO may have delivered some hints as to how and why this happens.

她用了这个词——说出来我都有点难为情，感觉像是个不好的词——她用了‘兜底’这个词。

She used the word, and I'm I'm like ashamed to say it out loud, this feels like a bad word, she used the word backstop.

如果你熟悉‘兜底’的含义，就会想到2008年许多银行的遭遇，政府在那里为这种扩张可能带来的任何负面后果提供资金支持。

And if you're familiar with the word backstop, it's what a lot of banks were experiencing in 2008, where the government kind of sits there to fund any sort of negative repercussions that happen as a result of this build out.

看完这个采访后，不禁要问：我们看到的英伟达、微软、AMD之间的这种循环经济现象，究竟是资本主义效率的体现，还是泡沫即将破裂的早期征兆？

And then after seeing this interview, it kind of begs the question, are these circle like, circular economics we've been seeing among NVIDIA, Microsoft, AMD, is it really a capitalist efficient thing, or is this is are we really starting to see early signs of a bubble starting to burst?

接下来我们将深入探讨所有这些话题。

So we're gonna get into all of that.

我们还要讨论苹果刚确认的一笔交易——他们向谷歌支付巨额资金，将AI计算任务外包以构建新模型。

We are also talking about a deal that was just confirmed with Apple, where they are paying Google a tremendous amount of money to actually offload their AI compute to build a new model.

所以现在我们有了一笔官方交易，AI领域出现了一款新硬件设备，还有一堆很酷的太空相关消息。

So we have an official deal there, we have a new hardware device in the world of AI, and we have just a bunch of cool space related stuff.

那么EJS，让我们进入今天的重磅新闻：OpenAI正在寻求‘兜底’保障。

So, EJS, let's get into the big news of the day, which is OpenAI asking for a backstop.

这听起来很可怕。

That sounds scary.

请详细说明发生了什么，因为我感觉有些信息在传递过程中丢失了。

Walk us through exactly what happened because I feel like some things get lost in translation here.

他们具体说了什么？这对我们意味着什么？

What exactly did they say, and and what does it mean for us?

是的。

Yeah.

让我先铺垫一下背景。

So so let me set some context.

Sarah Fryer是OpenAI现任首席财务官，她接受了《华尔街日报》采访，自然谈到了OpenAI的收入模式以及如何支付数万亿美元的计算费用。

Sarah Fryer is the current CFO at OpenAI, and she did this interview with the Wall Street Journal where, obviously, they get into the economics of OpenAI's revenue model and how they're gonna pay for all these trillions of dollars worth of compute.

她说了几句很关键的话。

And she had a few choice words.

最关键的正是你提到的那个保障机制，Josh。

The most choicest was around this backstop that you're talking about, Josh.

嗯。

Mhmm.

让我给你详细解释一下。

Let me let me kinda paint it out for you.

目前OpenAI已签署了价值1.4万亿美元的计算协议。

So OpenAI has currently signed $1,400,000,000,000 worth of compute deals.

这是OpenAI同意向计算供应商或芯片供应商（如英伟达、AMD等）支付数千亿美元购买GPU，用于训练下一代最先进的人工智能模型——通用人工智能。

This is OpenAI agreeing to pay compute providers or chip providers or GPU providers like NVIDIA, AMD, hundreds of billions of dollars in order to buy their GPU so that they can train the next best AI model, AGI.

但存在一个重大问题。

There's one massive problem.

他们没有足够的现金。

They don't have enough cash.

他们没有足够的钱。

They don't have enough money.

事实上，他们目前正处于亏损状态。

In fact, they are currently running at a loss.

他们既没有计划，也没有近期能实现的目标来支付这些东西的费用。

There's no plan or near term kind of goal that they can achieve to pay for this stuff.

那么问题就来了，如果他们付不起这些东西的钱会怎样？

So then the question becomes, what happens if they can't pay for this stuff?

如果Chattypati未来不像他们预测的那样盈利了会怎样？

What happens if Chattypati stops becoming profitable in the future as they've projected?

到那时会发生什么？

What happens then?

对此莎拉有一个简单的答案，那就是政府可以救助我们。

And Sarah had one simple answer to that, which was the government can bail us out.

具体来说，她描述的是一个生态系统：如果OpenAI真的违约或无力支付，政府会出面支付OpenAI购买这些GPU所需的资金。

And so specifically what she described was an ecosystem where the government basically pays OpenAI and gives them the money that they need to buy these GPUs if they happen to default, if they happen to not have enough money to pay for these things.

这反过来会给那些最初同意贷款给OpenAI购买GPU的私募基金和银行吃下定心丸——他们会没事的。

This in turn will give the private equity firms and the banks that are agreeing to loan OpenAI money in the first place to buy these GPUs, the sanctity and peace of mind that I'm gonna be okay.

如果OpenAI无法兑现承诺，我至少能拿回我的钱。

I'm gonna get my money back if OpenAI doesn't deliver through.

最疯狂的是，乔什，我感觉自己像在房间里被弹来弹去整整两周——因为山姆·奥尔特曼从运营非营利组织，变成了运营一个有点像私营公司的非营利组织，还把部分业务变成了营利性质。

The craziest part about this, Josh, is I feel like I've just been ricocheted across the room for two weeks straight because Sam Altman went from operating a nonprofit to operating a nonprofit that is kind of a private company where they kind of turned into a part of their company into a for profit.

所以他们某种程度上是在撒谎。

So they're kind of like lying about it.

然后又散布关于进行一万亿美元IPO的谣言。

Then kind of breaking rumors around doing a $1,000,000,000,000 IPO.

你对此的第一反应是什么？

What what do you what is your immediate gut reaction to this?

这是谎言吗？

Is it lies?

这是真的吗？

Is it real?

到底发生了什么？

What's happening?

我脑子里一直在重复'大而不倒'这个词。

I just like I keep repeating the words too big to fail in my head.

大而不倒。

Too big to fail.

大而不倒。

Too big to fail.

看起来这就是他们想要达到的目标。

It seems like that's that's where they want to get to.

我对此有两种想法。

I'm like, I'm of two minds of this.

一方面，人工智能确实关乎国家安全。

One is that, well, AI is a matter of national security.

正确且尽可能快速地推进此事至关重要。

It is very important to get this right and to move as fast as possible.

如果这需要政府提供一定帮助，在某种程度上可能是合理的。

If that requires some government help, that probably makes sense to an extent.

其次，如果你正在寻求政府援助，在这种情况下可能并不是件好事。

And then the second thing is, well, if you are asking for government help, that's probably not a good thing in this instance.

而政府援助实际上意味着纳税人的资助。

And government help really means taxpayer help.

就像，是我们为这些东西提供资金。

Like, we we fund this stuff.

甚至要为Sam Altman提供兜底保障，而我们却因公司高度私有化无法获得公开股票敞口。

And to to to pay a backstop for Sam Altman when we don't even get public stock exposure because it's so privately held.

另一点是她在此承认，他们实际上近期几乎没有上市的兴趣。

And another thing is she admitted here that they actually have little to no interest in IPOing anytime soon.

现在公众根本没有途径分享上涨收益，只能在他们自成立以来一直标榜非营利的情况下承担下跌风险。

So now there's no real trajectory for the public being able to own any upside and only participate in the downside after they've been signaling that they are not for profit for since inception.

这些信号混乱得可怕，完全看不出明确意图。

It's just like these really horrifically mixed signals with no clear intention.

问题很严重。

It's a lot.

是啊。

Yeah.

你描述得像在房间里来回反弹——确实有这种感觉，有点令人不安。但我怀疑这是否正是我们看到的循环经济现象的核心：OpenAI与谷歌、微软轮流达成交易。

I guess the the way that you described it being like ricocheted across a room, like, it kinda feels that way, and it's a little disturbing, but I I wonder if this is kind of at the core of what we've been seeing with this circular economic thing happening where OpenAI sends a deal with Google, with Microsoft.

这一切感觉非常近亲繁殖，但也许是因为这确实关系到国家安全，政府某种程度上允许他们做许多传统上不被接受的事情。

It's all very it feels very incestual, but maybe it's because it really is a matter of national security, and the government's just kind of allowing them to do a lot of things that wouldn't traditionally have been acceptable.

是啊。

Yeah.

但看起来我们有一些澄清。

But it looks like we have some clarifications.

对吧？

Right?

像是...对。

Like Yeah.

我是说也许我们

I mean Maybe we

应该对此持保留态度。

should take this with a grain of salt.

没错。

Yeah.

需要谨慎看待的是，提出这些说法的首席财务官莎拉·弗莱在采访中收回了她的说法，明确指出OpenAI并不寻求政府为其基础设施承诺提供担保。

I so a grain of salt being that Sarah Fry, the same CFO that kind of made these claims, walked back her claims in the interview specifying that OpenAI is not seeking a government backstop for our infrastructure commitments.

她接着给出了官方声明，说OpenAI是盈利的，完全有能力自行承担所有费用。

And she goes on to give this official statement about, you know, OpenAI, you know, being profitable and be gonna be able to pay their way through this all.

乔什，我对这件事的质疑在于，我认为这不是诚实的声明。

The issue that I have with this, Josh, is I don't think it's an honest statement.

为什么我认为这不是诚实的声明？

Why don't I think it's an honest statement?

因为她的老板、CEO萨姆·奥特曼曾在多次采访和自己撰写的博客中声称，他毫不介意向政府寻求救助、设置安全网，以帮助他填补赤字和债务，如果他无力承担的话。

Because her boss, CEO Sam Altman, has claimed so many times in interviews and his own blog posts that he's written that he has no issue asking the government for a bailout, for a backstop, to help him kind of like cover his deficits and his debts if he's not able to pay for it.

我这里摘录了一段采访内容，他在其中说道：‘某种程度上，当事情变得足够重大时，无论是否在纸面上，联邦政府就像是最后的保险商。’

I have a an excerpt from an interview pulled up here where he goes, at some level, when something gets sufficiently huge, whether or not they are on paper, the federal government is kind of an insurer of last resort.

正如我们在各种金融危机和保险公司搞砸事情时所见，他们基本上会为我们兜底。

As we've seen in various financial crises and insurance companies screwing things up, they basically cover us.

所以，这再次——也是本周我们第二次讨论萨姆·奥特曼言行不一的性格矛盾，他说的是一套，想的却是另一套。

And so, this is again and and this is the second time we've spoken about this this week about Sam Altman's character inconsistencies where he says one thing and means another.

本周早些时候我们做了一期节目，讨论了一份52页的证词，他的联合创始人兼前首席科学官伊利亚基本上表示，萨姆就像是在不断撒谎，而这正是导致他在2023年11月被罢免的原因。

We did an episode earlier this week where we covered a 52 page deposition where his cofounder and former chief science officer, Ilya basically says that Sam was, like, incessantly lying, and that that's what led to his ousting, his firing in November 2023.

快进到现在，看起来萨姆还在继续这种行为，乔希。

Fast forward to to present day, it seems like Sam's still at it, Josh.

嗯。

Mhmm.

存在很多性格矛盾，而让情况更糟的是，从他创业第一天起直到

There's a lot of character inconsistency, and what's certainly not helping the case is all of his cofounders from day one to

直到几年前

to up until a few years

的所有联合创始人，现在都在作证指控他。

ago, they're all testifying against him.

埃隆 是啊。

Elon Yeah.

强烈反对萨姆的立场。

Strongly disagrees with Sam.

伊利亚和米拉，他们对萨姆极度不信任。

Ilya and Mira, they strongly distrust Sam.

所以这不仅仅是公众舆论的负面信号，连认识他本人的人也这样认为，这确实不是什么好现象。

So this is not this is poor signal coming from the public perception, but also from people who know him personally, which is not really making a good case.

我们拭目以待吧。

So we'll see.

我们会继续关注事态发展。

We'll continue to monitor the situation.

关于OpenAI的新闻，目前情况大致如此。

That's kind of where we're at with the OpenAI news.

抱歉打断一下，乔什，我想快速了解另一方的观点，特别好奇你会怎么看待这个问题。

I just wanna, sorry, I think we need to just look at the other side very quickly, Josh, I'm curious if you will take care specifically.

好的，让我们进入泡沫这个话题。

Yeah, let's get into the bubble.

从中得出的结论就是：天啊，我们正在做循环投资。

So the take from this is, oh my god, we're doing circular investing.

这是个巨大的泡沫。

This is a massive bubble.

很明显它即将破裂。

It's so obviously gonna pop.

他们甚至在崩盘前就要求政府直接救助。

They're asking for a literal government bailout before the crash actually happens.

我们在新冠疫情中就见识过这种情况。

We're seeing this with COVID.

我们在2008年就看到了这种情况。

We're seeing this in 2008.

我们在这里做什么？

What are we doing here?

对此的反驳论点是，如果你看看其他所有超大规模企业，或许暂时先把OpenAI放一边。

And the counterargument to that is if you look at every other hyperscaler, maybe not put put OpenAI aside for a second.

看看Meta、Google和微软这些公司，尽管他们正在投入数千亿美元，并承诺在未来几年持续投入，但他们的损益表上还没有出现足以让人担忧的疯狂亏损。

If you look at, Meta, if you look at Google, if you look at Microsoft, although they're spending hundreds of billions of dollars and committing to do that over the next couple of years, they still haven't made a a crazy enough dent where you should be getting worried, on their profit and loss sheet.

别忘了，这些公司都拥有其他利润极其丰厚的业务，他们正在赚取海量的利润。

Remember, these companies all have other businesses that are massively profitable, and they're making tons and tons of money.

从技术上讲，他们目前在计算领域的投资规模，如果与那些利润相比，其实并不算太疯狂。

The money that they're investing in compute right now, technically, if you weigh it up against that, isn't too crazy.

这太疯狂了。

It it is crazy.

过去十年间他们在其他方面的支出确实是个大窟窿，但这仍不足以弥补他们当前的收入缺口。

It is a big dent that we've seen them spend on anything else over the last decade, but it's still not overcompensating for what they're earning right now.

而这是唯一的反对理由，就像在说：嘿，我们看到企业客户有需求。

And that's the only argument against it, which is like, hey, we're seeing demand with our enterprise customers.

我们看到零售客户也有需求。

We're seeing demand with our retail customers.

因此我们投资这些计算资源是合理的。

And so it makes sense for us to invest in this compute.

不知道你对此怎么看。

Don't know what you think about this.

是的。

Yeah.

比如，如果AI扩展定律明天突然停止，我们突然发现，哎呀，糟了。

Like, in in the case that AI scaling laws stop tomorrow, where suddenly we figure out, oh oh, no.

就像，这这实际上行不通。

Like, this this isn't actually gonna work.

花更多的钱。

Spending more money.

市场会受到严重损害，但还不至于灾难性。

The the market gets harmed very badly, but it's not catastrophic.

这还没达到经济衰退级别的泡沫。

It is not a recession level bubble.

如果这种情况持续延长，像OpenAI这样没有像微软和谷歌那样资本结构的公司——这些科技巨头存在于AI领域之外，现在正用资产负债表来支付这些费用——我想那时候可能就会开始出问题了。

In the case that this continues to prolong, companies like OpenAI that don't have a cap table like companies like Microsoft and Google who exist outside of AI and are now using their balance sheet to pay this off, I think that's probably when you start to see problems.

所以在某种程度上我同意你的观点，我们目前还是安全的。

So I kind of agree with you in the sense that we're still good.

就像，情况仍然良好。

Like, things are still good.

我我并不是特别担心短期内会出现泡沫。

I'm I'm not particularly concerned of a short term bubble happening here.

我想强调，OpenAI并不是干坐着不开发其他创收方式。

I wanna emphasize that, like, OpenAI isn't just kind of sitting on their hands and not coming up with other ways to turn on revenue.

她在采访中透露的另一件事，Josh，我想是屏幕上第六条这里。

One other thing that she revealed in this interview, Josh, was I think it's number six on the screen here.

此外，他们将进行创新的商业交易。

Additionally, they will do creative commercial deals.

他们的意思是，如果有公司，比如一家制药公司使用ChatGPT并通过它发现癌症治疗方法，他们会与这家制药公司签订协议，从而从利用AI技术研发的药物中分得一定比例的利润——这种分成关系可能会持续到天知道多久以后。

What they mean by that is if there's a company, say a pharmaceutical company that uses ChatGPT and finds a cure for cancer using ChatGPT, they're signing a deal with that pharmaceutical company such that they get a percentage of profits from that drug that they create using their AI intelligence, that will like occur for like, God knows how long after that.

因此他们会从那些将ChatGPT作为产品来创造其他产品的用户那里抽取利润或收入分成。或者如果你是通过ChatGPT销售产品的公司，比如目前已在ChatGPT上线的Etsy就是个典型案例。

And so taking a percentage of profit or revenue share from people who are using it as a product to create other products, Or if you're a company that sells products via ChatGPT, Etsy is a common example that is live on ChatGPT right now.

他们能获得一定比例的利润。

They get a percentage of profits.

然后是最显而易见的方式——OpenAI即将在ChatGPT上投放广告。

And then there's the obvious one, which is ChatGPT is gonna OpenAI is just gonna turn on ads.

一旦开启广告业务，天知道能带来多少收入。

And when they turn on ads, who knows how much money that's gonna bring in.

所以他们正在朝这个方向努力。

So they are making efforts towards it.

我不想把这说得像是——嘿——在诋毁OpenAI，但考虑到他们承诺投入的巨额资金，这种盈利模式恐怕不太现实。

I don't want this to be a like a, hey, like bad OpenAI thing, but it's just unlikely given the amount that they've committed to spend.

你提到的第一点关于通过ChatGPT取得医疗突破并抽成的模式，听起来就像一场即将发生的灾难。

That first point you mentioned around, like, health breakthroughs through ChatGPT getting a percentage, that sounds like a train wreck waiting to happen.

这种盈利结构非常混乱，还有很多需要评估的地方。

That's a very messy monetization structure, so a lot still to be evaluated.

不过Ejaz，我现在想谈谈苹果公司。

But, Ejaz, I wanna talk about Apple now.

我们昨天做了一期关于苹果和谷歌的节目，讨论它们之间的关联性，特别是涉及苹果某笔见不得光交易的部分。

We had an episode yesterday with Apple and Google and how they kind of relate to each other, particularly around a deal in which Apple kind of sucks to the eye.

他们确实不太擅长这个，嗯。

They're really just not good at it, and Mhmm.

他们需要帮助。

They need help.

这时Gemini前来救援了。

And here is Gemini coming to the rescue.

我们正式达成了一项非正式官方的协议，看起来苹果每年将向谷歌支付10亿美元，以获得一个为苹果定制的1.2万亿参数的谷歌Gemini模型。

We officially have a deal that is unofficially official, it looks like they're gonna be paying $1,000,000,000 a year, Apple, to Google in order to get a 1,200,000,000,000.0 parameter Google Gemini model custom for Apple.

这可是件大事。

This is a really big deal.

苹果现在很吃力。

Apple is struggling.

苹果在AI领域一直无所作为，突然间却拥有了如此强大的模型。

Apple has not done anything in the world of AI, and suddenly, they have this really powerful model.

所以至少在苹果的利好情况下，这似乎将变得非常重要。

So this seems like it's going to be very important for the case of the bull case for Apple, at least.

乔什，我有几个理由为此感到高兴。

So I'm happy about this for a few reasons, Josh.

因为通常我会觉得好笑。

Because normally I'd be laughing.

对吧？

Right?

我会觉得，这显然，苹果失败了。

I'd be like, This obviously, Apple's failed.

他们在这场竞争中迟到了。

They're late to the game.

但我对此感到高兴，因为在苹果看来，这某种程度上是个明智之举。

But I'm happy about this because, in Apple's sense, it's kind of a smart move.

想想看。

Think about it.

对吧？

Right?

他们无需花费数千亿美元投资GPU并训练复杂的人工智能模型。

They haven't spent hundreds of billions of dollars trying to invest in GPUs and train a complex AI model.

他们完全没有承担这些风险。

They haven't taken on any of that risk.

他们只是拍拍做了所有苦工的Google的肩膀说：嘿，如果我每年付你十亿美元，你愿意为我们定制一个苹果AI模型吗？这个模型可以接入Siri并在我们的私有云实例上运行。

They just tap Google on the shoulder who's done all the hard work and say, yo, are you down if I pay you a billion dollars, per year and you make our own custom version of an Apple AI model that I can plug into Siri and will run on our private cloud instance.

所以你看，Google不一定能获得全部权限，但他们每年都能拿到这笔钱。

So, you know, Google won't necessarily get access to all of it, but they just get the payment every year.

听起来这交易相当划算。

That sounds like a pretty sweet deal.

我喜欢的另一点是，这可不是随便什么普通模型。

The other thing I like about this is this isn't just any kind of like model.

这是一个拥有1.2万亿参数的模型。

It's a 1,200,000,000,000 parameter model.

这规模已经堪比市面上最大的那些模型了。

That is like up there with like one of the biggest models that would be out there.

再加上我认为苹果将会整合到Siri中的那种个性化功能，以及在手机上使用文本和其他应用时的用户体验，这对我来说非常有吸引力。

And to combine that with the kind of personalization that I'm presuming Apple is gonna integrate into Siri and in the consumer experience with using text and other apps on the phone, that's pretty attractive to me.

另一方面，我觉得谷歌这边非常酷的是，乔什，他们能以经济可行的成本运行一个1.2万亿参数的模型——也就是说他们还能从中盈利——这充分展示了谷歌实现的惊人工程成就。

The other thing that I thought was super cool on the Google side here, Josh, is to be able to run a 1,200,000,000,000.0 parameter model at economically viable cost, AKA they're making money from that, just goes to show that there's some pretty crazy engineering that Google has achieved.

这间接告诉我，他们的芯片设计和TPU架构已经做到了极致，才能实现这样的壮举。

This indirectly tells me that they have absolutely nailed their chip design and their TPU architecture to be able to pull this off.

简直太酷了。

Just super cool.

如果想了解更多关于苹果的信息，可以收听我们昨天发布的节目。

There's a lot more info if you wanna find out about Apple on our episode that we released yesterday.

另外，我今天在通讯稿中发表了一篇文章，当你看到这个视频时，文章会详细分析苹果采取这些举措的经济逻辑和合理性。

And also, I am publishing an essay in the newsletter today when you're watching this all about the economics and why this makes sense for Apple to do.

如果你对这些深度分析感兴趣，想看到更多这样的思考，可以去看看通讯稿。

If you're interested in hearing more of these takes, like more thoughtful takes, check it out on the newsletter.

我...我对这篇文章真的非常自豪。

I I'm, like, very proud of this article.

我觉得反响会非常好。

I think it'll do really well.

看着苹果的战略阴差阳错地让他们陷入这种绝佳境地，虽然绝非刻意为之，但他们就是莫名其妙地占据了优势，这现象实在引人入胜。

It's just really fascinating to see the Apple strategy kind of accidentally step into this amazing situation for them, where it was certainly not by design, but they somehow managed to put themselves in a really good place.

不过关于谷歌的话题，我还想谈谈他们刚发布的新硬件——Ironwood TPU芯片。

But on the topic of Google, I also wanna talk about the new hardware that they just announced, which is their new Ironwood TPUs.

再次提醒，昨天的节目里我们详细讨论过什么是TPU以及它和GPU的关系，那期内容很值得一看。

Now again, in yesterday's episode, this is a good one, We talked about what a TPU was and how it relates to a GPU.

今天有个新消息，我们迎来了新的TPU，EJS。

And today, in some new news, we got new TPUs, EJS.

你能详细介绍一下这些Ironwood是什么、有什么功能、为何如此出色吗？

So can you walk us through what these Ironwoods are, what they do, why they're impressive?

Ironwood是谷歌最新的TPU。

So Ironwood is Google's latest TPU.

TPU全称是张量处理单元。

TPU stands for tensor processing unit.

你只需要知道，谷歌的TPU相当于英伟达的GPU，但具备更多优势。

All you need to know is that the TPU of Google is the equivalent of the GPU from NVIDIA, but with some additional perks.

它更专精于适配谷歌的软件和应用套件，对吧？

It is more specialized and custom fit towards Google software and app suite, right?

正是它驱动并训练了他们所有的模型。

It's the thing that has powered and trained all their models.

他们实际上从未依赖过英伟达来训练和推理他们的模型。

They've never actually relied on NVIDIA at all to kind of train and inference a bunch of their models.

他们就像是个独行侠。

They've been kind of like this lone entity.

为什么这很酷呢——我们在昨天的节目里解释过（推荐回看）——谷歌一直保持高度独立，实现了多项突破，使他们能以更低成本训练出与OpenAI和微软同类型的模型，且能大规模扩展更多自建模型。

And why this is so cool, and we explained this on yesterday's episode, which you should check out, is Google's been super independent and they've been able to make several breakthroughs, which have allowed them to train the same type of models that OpenAI and Microsoft produce, but much cheaper or cost efficient and can scale massively with a greater number of models that they build.

现在他们发布了名为Ironwood的新型号，本质上是TPU的下一代产品。

And now they've released this new model called Ironwood, which is basically the next generation of their TPU.

它比前代快四倍，还能以更高效的方式堆叠组合，这意味着大规模训练更大型号将变得容易得多。

It is four times faster than the prior version, and it can basically clamp together as one singular stack in a much more feasible way, which means that training larger models at scale is gonna be much easier.

但对我来说最重要的消息，Josh，是他们将开始销售这些TPU，让其他人更容易购买他们的TPU，并训练或影响自己的人工智能模型。

But the biggest news about this for me, Josh, is they're gonna start selling these TPUs and making it more accessible for anyone else to buy their TPUs and train or influence their own AI models.

为什么这是重大新闻？因为这标志着他们正式进入竞技场与英伟达展开竞争。

Why this is such big news is this means they're formally stepping into the ring to compete with NVIDIA.

别误会我的意思。

Now don't get me wrong.

他们目前的规模还无法与英伟达相提并论，但迄今为止，还没有出现能真正挑战英伟达的对手。

They're not doing this at the scale that NVIDIA is currently doing it, but currently, there has been no feasible challenger to NVIDIA.

而现在谷歌入场了，他们拥有强大的分销网络和技术专长。

And now you have Google entering the ring, has a lot of distribution and technical expertise.

这很值得注意，或许也暗示着谷歌的市值应该远远高于当前水平。

It's notable and probably a hint that the Google market cap should be much, much higher.

是的。

Yeah.

我不会说没人能与英伟达竞争。

I wouldn't say no one is competing with NVIDIA.

还有AMD的芯片。

There there are AMD chips.

中国正在创造自己的替代品。

China's creating their own alternatives.

所以我认为目前正处于试图竞争的水平，但要成为真正的英伟达竞争者，还有一座非常陡峭的山需要攀登。

So I'd say it's at that level where it's trying to compete, but there is going to be a very steep mountain in order to get there to become a real NVIDIA competitor.

另外，我们这周拿到了一款新的人工智能硬件设备，它名叫Sandbar，采用环形设计。

In other news, we got a new AI hardware device this week, and it goes by the name of Sandbar, and it comes in the form factor of a ring.

现在我手上戴着一个Oura戒指。

Now I have an Oura Ring on it.

我超爱我的Oura戒指。

I love my Oura Ring.

它是很棒的无感睡眠追踪器。

It's great non intrusive sleep tracker.

这完全重新定义了戒指的概念，因为它能接入我们称为边缘节点的设备。当你与AI系统交互时，这些传感器可以向真正的AI系统发送请求，这是一种全新的传感器。

This is a totally new take on a ring because it gets into these things that we call edge nodes, and when you deal with AI systems, there are sensors that can then send requests back to the actual AI system, and this is a new sensor.

现在屏幕上播放的是这个戴戒指人士的宣传视频。

So what we're seeing on screen is a promo video of this person who has a ring.

按下戒指上的按钮，你就可以对着麦克风说话。

You press a button on the ring, and you can speak into a microphone.

这个麦克风充当了你与AI系统之间的交互界面，你可以让它做各种事情，比如设置提醒、回答问题、记录对话，这确实是AI硬件设备中非常独特且新颖的形态。

Now this microphone acts as an interface between yourself and the AI system, and you could ask it to do things like remind you, to query questions against it, to record conversations, and it's this really unique and, I guess, somewhat novel form factor in the world of AI hardware devices.

至少对我来说这很有趣，Ejaz，因为我特别好奇OpenAI明年会和Johnny Ive合作推出什么AI硬件，这个戒指形态的实验也很有意思。

This was interesting to me, Ejaz, at least, because I'm so fascinated about what OpenAI is going to make next year with Johnny Ive in terms of AI hardware, and this is an interesting experiment to kind of see how the ring form factor would work.

它的工作原理是通过麦克风将音频被动传输到你的耳机或手机，这种使用AI的方式既有趣又直观，完全不会造成干扰。

So the way this kind of exists is it's a microphone that has passive audio back to your earbuds or, I assume, back to your phone, and it's this really fun and somewhat intuitive way of using AI without AI getting in the way.

我认为在智能手机逐渐淡出的过程中，我们会看到大量这种环境设备，让你随时随地与AI互动。而戒指是绝佳的载体，因为它几乎不会妨碍日常活动。

So I think a lot of the things that we're gonna start to see in this removal of the smartphone is this this suite of ambient devices where you can just kind of engage with AI wherever you are at any time, And a ring is a really neat form factor for this because it's it doesn't really get in the way.

它就那样一直存在着。

It's just kind of always there.

想使用时召唤它，不想用时甚至完全感觉不到它的存在。

If you want to engage with it, you summon it, and if you don't, you don't even think about it.

我觉得这是个有趣的实验，我有点想尝试一下。

And I I think this is this is an interesting experiment, and it's something I'd kind of want to try this.

倒不是说它是个成功的产品，但我觉得它展示了未来AI设备可能的样子。

Not that I think it's a successful product, but I think it's an it's an interesting take on what the future AI devices could look like.

是啊。

Yeah.

我看到这个时，第一反应是她举起手对着戒指说话的样子。

Where my mind immediately leaps when I I look at this is her holding up her hand to speak into the ring.

看起来要么像是要咳嗽，要么就像电影里保镖需要通话时会摸耳朵的特勤人员那样。

It either looks like she's about to cough or, like, you know, in the movies where the the bodyguards are, like, touching their ear when they need to speak in turn secret service becomes in the secret service.

感觉有点像那种场景。

It kind of seems like something like that.

所以从习惯角度，我很好奇这种东西如何融入社会，不过我同意你的观点。

So habitually, I'm kind of curious as to how this kind of integrates into society, but I agree with you.

我觉得这款戒指非常低调不显眼，能在不盯着屏幕的情况下便捷使用技术，同时保持与现实生活的互动。

I think, like, the ring is super subtle and kind of nonobtrusive, and it makes it super convenient to kind of engage with this technology without needing to stare at another screen and whilst maintaining the ability to kind of interact with real life.

没错。

Yeah.

这让我深究了一番，因为我们聊到苹果公司，我真心希望苹果能开始做这类产品。

This led me down a rabbit hole because we were on the topic of Apple, and it really I'd love for Apple to start doing things like this.

比如苹果现在有1000亿美元现金储备。

Like, for example, Apple has a $100,000,000,000 in cash.

他们可以去收购Aura。

They go acquire Aura.

现在他们从谷歌获得了定制的Gemini模型。

Now they get the custom Gemini model from Google.

突然间，AI出现了。

Suddenly, have AI.

他们推出了这款兼容新型Gemini模型的AI戒指，还发布了一款配备视觉传感器的AirPods，让你能收集外界数据，于是你开始拥有这套设备——虽非iPhone，但其功能日益强大，逐渐接近iPhone所能实现的一切。

They ship this this AI ring that is compatible with these new Gemini models, they ship a new set of AirPods that have visual sensors on them so you could collect data from the outside world, and you start to get this suite of devices that isn't an iPhone, but is increasingly capable and more powerful and approaching what we can do with an iPhone.

所以我希望这能成为一种趋势，下一代iPhone级别的设备不再是一个单一设备。

So I hope this is a trend that we see where the next iPhone level device isn't a device.

而是一套设备组合。

It's a suite of devices.

也许戒指是其中之一，也许不是，但探索这种可能性会是个有趣的实验。

Maybe the Ring is one of them, maybe it's not, but it's an interesting experiment to see what it could look like if that became the case.

我想聊聊AI竞赛中的黑马——Anthropic。

I wanna talk about the dark horse of the AI race, which is Anthropic.

乔什，现在我要昂首挺胸地说。

Now I'm gonna hold my heads up here, Josh.

我曾多次批评Anthropic，甚至称其为'守规矩的AI模型'。

I have given Anthropic a lot of flack, and I've kind of called them the NOC AI model.

比如，他们遵守规则。

Like, they follow the rules.

他们去找政府，他们运作

They go to the government, they run

到政府那里说，嘿，就像

to the government and say, hey, like,

你能给我们这个交易还是什么吗？

can you give us this deal or whatever?

而且我一直有点看不起Claude，因为他们似乎在编码代理层面保持了某种对等。

And I've kind of like looked down on Claude since they kind of like maintained their parity at the the coding agent level.

所以我就觉得，随便吧，我为什么要用Claude呢？

So was kind of like, whatever, like, why would I use Claude?

结果证明我大错特错了。

Turns out I was very wrong.

泄露的信息显示了一份关于Anthropic预期收入的报告，基本上把他们放在了与OpenAI同等的水平，在某些方面甚至更好，Josh。

So the information leaked a report on projected revenue for Anthropic, and it basically puts them at the same level as OpenAI, in some cases better, Josh.

所以主要结论是，他们预计到2028年将实现700亿美元的收入和4000亿美元的估值。

So the major takeaway from this is they're projecting $70,000,000,000 worth of revenue by 2028 and a $400,000,000,000 valuation.

请记住，OpenAI目前的估值大约是5000亿美元。要从他们现在的2000亿美元跃升到这个水平，简直是疯狂的跳跃。

Bear in mind that OpenAI is currently valued at, I think, $500,000,000,000 So to make that leap from what they are where they are currently, which I think is $200,000,000,000 is a a crazy jump.

但第二点，这意味着他们将比OpenAI更早实现盈利。

But number two, this would assume they then become profitable way earlier than OpenAI.

与此同时，OpenAI在2028年技术上会比Anthropic赚更多钱，但不会盈利。

OpenAI at the same time, 2028, will be making technically more money than Anthropic, but won't be profitable.

所以我立刻问自己的问题是，他们打算如何实现这个目标？

And so the immediate question that I asked myself was, well, like, how are they planning to do this?

到目前为止，他们在零售销售上是亏损的。

Like so far they're losing on the retail sales.

所以他们有什么绝招吗？

So like, do they have a hail Mary?

我把线索串联起来的答案就在这里，Josh，在企业AI市场份额中，Anthropic已经悄然超越了OpenAI。

And the answer I connected to the dots, Josh, comes right here in the enterprise AI market share, where Anthropic has sneakily surpassed OpenAI.

我认为这张图表有点过时了，但他们目前占据了企业API市场约25%的份额。

They currently command I think this chart is a little outdated, but they currently command around 25% of the enterprise API share.

这之所以非常重要，是因为虽然企业客户数量可能较少——比如OpenAI有8亿周活跃用户，而Anthropic可能只有几十万企业用户——但每个企业用户的付费能力远超普通零售用户。

And why this is super important is although there may be fewer enterprise customers in terms of numbers, so, you know, OpenAI has 800,000,000 weekly active users, Anthropic might have, I don't know, a couple 100,000 enterprise users, Each enterprise user pays way, way, way more than the average retail user.

所以这是我没预料到的情况，Josh。

And so it's just something that I didn't see, Josh.

他们正在幕后与众多企业展开合作。

They're like engaging with a lot of enterprises behind the scenes.

他们签署着数十亿美元的合同，并真正将这些合同转化为企业实际使用的实用产品。

They're signing these multi billion dollar contracts and they they're actually like translating these contracts into useful products that these businesses are using behind the scenes.

至于如何验证这一点，我也不确定。

How we prove that, I don't know.

或许可以通过这些合作企业随时间推移的经济GDP来证明，但我觉得这个发现很有意思。

Maybe it's like the economic GDP over time from a bunch of these different companies that they've signed deals, but I thought this was cool to point out.

确实。

Yeah.

这在我看来完全合乎逻辑。

This to me intuitively makes sense.

就像提到编程就会想到Claude，我认为这在企业界已是普遍共识。

It's like when you you think of coding, you think of Claude, and I think that's kind of the universal truth amongst corporations.

当你需要编写代码的机器人时，你用的就是Claude。

And when you want a bot to write code for you, you are using Claude.

如果你正在调用一个为你编写代码的API，很可能你用的是Claude。

And if you're pinging an API that is writing code for you, odds are you're using Claude.

尽管普通大众对Claude和Anthropic的观感并不那么乐观，因为它不如ChatTubyTe或Gemini实用，但现实是，如果你在编写代码，或者是一家想要拥有出色代码生成模型的公司，你用的正是Claude和Anthropic。

So while the retail general public facing sentiment isn't that optimistic around Claude and Anthropic because it's just not as useful as ChatTubyTe or Gemini, the reality is that if you're writing code and if you are a company that wants a model that writes fantastic code, you are using Claude and Anthropic.

嗯。

Mhmm.

Anthropic正在悄然收获大量红利，却避开了许多公开的舆论风波，因为这些交易都是私下进行的。

And Anthropic is just collecting a lot of the upside without a lot of the public facing noise because those are just private entities.

他们只需刷卡购买代币，就能获得所需的一切资源，心满意足地继续前行。

They're just swiping their credit cards, they're getting all their tokens, and they're just happy, and they're they're on their way.

所以，是的，我为Anthropic感到高兴。

So, yeah, I mean, I'm I'm happy for Anthropic.

希望这种模式能持续下去。

I hope this is durable.

我很欣赏这些公司各自在市场中找准定位的趋势。

I like the fact that we're starting to see each of these companies kinda slot themselves into a portion of the market.

我喜欢Anthropic专注代码开发的策略。

So I like that Anthropic is just working on code.

我认为这很棒。

I think that's great.

不必事事争第一，只需打造最优秀的代码生成模型，看看能从中创造多少财富。

Don't try to be the best at everything, try to make the best coding model, and look how much money you could print from it.

这是个正和博弈。

This is a positive sum game.

这块蛋糕正在飞速膨胀，所以如果能像Anthropic那样占据一小块角落，那真是再好不过了。

The pie is continuing to grow so quickly, so if you could just own a small corner of it like Anthropic is doing, all the power to them.

继续保持。

Keep it going.

有些人可没Anthropic这么顺风顺水。

Someone who's not faring as well as Anthropic.

那匹白马。

The white horse.

我们的老朋友Meta和扎克伯格先生，本周财报后因多重原因遭遇了重创。

Our good friends over at Meta and mister Zuckerberg, they got absolutely crushed this week after earnings for a series of reasons.

这其实并不令人意外。

This doesn't really come as a shock.

总体来说，我们对Meta一直持相当悲观的态度。

I think we've been pretty bearish on Meta as a whole.

经历了那场智能眼镜的惨败后，我对Meta这家公司及其在AI时代的执行力越来越看衰。

After that whole Glasses debacle, I became increasingly bearish on Meta as a company and their ability to execute in this world of AI.

不过Aegis，你对Meta财报后的情况有什么看法吗？

But, Aegis, do you have any takes on what happened after the Meta earnings report?

我...我有的。

I I I do.

我们大概两周前做过一期关于Meta看涨理由的节目。

So we put out an episode, I think, two weeks ago on the Meta bull case.

所以很多人可能会问：你们现在还坚持部分看涨观点吗？

And so the question a lot of you might have is, like, you know, do you still maintain some of that?

简而言之答案是肯定的，但需要更长的时间周期。

The short answer is yes, but over a longer time period.

比如，事实是这样的。

Like, here's the facts.

Meta上周公布了财报，由于一个特定原因，结果不如他们预期的那样理想。

So earnings came out for Meta last, week, and it wasn't as great as they had hoped for one particular reason.

他们本季度在AI上的支出简直荒谬。

Their spend on AI was ludicrous for the quarter.

该公司在员工身上花费了数百亿，这属于过度开支。

They had The company spending tens of billions on employees is overspending.

没错。

Exactly.

我完全没想到会这样。

I would have never guessed.

正是如此。

Exactly.

他们花费了数十亿美元，有些方式简直疯狂——就为了招聘几个人，同时还要为Hyperion数据中心的资本支出买单，投资开发各种应用程序，然后又解雇大批员工。

So they had spent billions and billions of dollars, in some cases, in crazy ways, just to hire a couple of people, but also to pay for CapEx investments for their Hyperion data center to invest in different apps being built, to fire a bunch of people.

是啊。

Yeah.

支出的金额简直离谱。

Just ludicrous amounts being spent.

我认为Meta股东们还有点2022年元宇宙时期留下的创伤后应激障碍，当时他们把整个公司从Facebook更名为Meta，结果他们押注的元宇宙并没有成功。

And I think that meta shareholders had a bit of PTSD from, the metaverse days in 2022, where they renamed their entire company from Facebook to, to meta based on this metaverse they're in ended up not panning out.

对吧？

Right?

NFTs当时还不存在。

NFTs weren't a thing.

所以他们有点创伤后应激障碍，看到扎克伯格花这么多钱却没有任何实际投资回报。

And so they have a bit of PTSD where they're seeing Zuck spending all this money, but no real ROI.

一个明显的例子是他们推出了AI助手，但根本没人用。

Some one clear example is they launched an AI assistant and no one really uses it.

现在没什么人用Facebook了，而且它也没能很好地整合到现有产品中。

No one's really on Facebook, and it's not really integrated well into their existing products.

人们只用ChatGPT。

People just use ChatGPT.

然后他们推出了对标Sora的产品，叫Meta Vibes应用。

They then launched a Sora competitor, like called the Meta Vibes app.

那个也没人用。

No one uses that either.

你还记得吗？

Do you remember that?

对吧？

Right?

接着他们又推出了...他们开创了...

And then they launched, they, they pioneered.

他们说：听着，我们不再抄袭别人了。

They said, listen, gonna stop copying people.

我们要走自己的路。

We're gonna do our own thing.

他们推出了自己的硬件设备——AI眼镜。

And they launched their own hardware device, which is the AI glasses.

死忠粉丝的接受度和反馈是他们听过最糟糕的。

The reception and feedback from their diehard fans was the worst they've ever heard it.

他们痛恨整个体验过程。

They hated the entire experience.

他们认为这产品比市面上任何同类产品都要差。

They think it's a it it is a lesser product than anything else on the market.

基于以上种种原因，人们根本不相信扎克有能力投入并兑现承诺。

So for all of these reasons and much more, people just don't have faith in Zuck's ability to spend and deliver on this.

这直接反映在了股价上。

And it's reflecting in the share price.

乔希，他们在24小时内市值蒸发了2500亿美元。

They lost $250,000,000,000 in twenty four hours, Josh, on market over.

简直疯狂，暴跌15%。

Just completely insane, down 15%.

现在说到元宇宙业务，我认为他们能重整旗鼓。

So now going on to the ball case, I do think they pick themselves out.

其实我觉得现在可能是买入Meta股票的好时机。

I actually think Metja is probably like a really good buy at this point.

一个核心原因是：我不认为扎克愿意输掉这场竞赛，他会不惜一切代价达成目标。

And for one solid reason, which is I don't think Zuk wants to lose this race, and he's willing to figure any and all out to to get himself to a point.

他能否开发出一系列利用AI技术实现类似他现有分发渠道的酷炫应用？

Is he able to produce like a bunch of really cool apps that leverage kind of like the distribution that he has using AI?

就像谷歌和OpenAI那样？

Like Google has, like OpenAI has?

我不知道。

I don't know.

待定。

TBD.

不过，乔希，你有什么想法吗？

But yeah, Josh, do you have any thoughts?

现在有个有趣的现象：我对使用Meta的任何产品都完全提不起兴趣，这在其他公司身上是不存在的。

There's this interesting phenomenon happening where I feel absolutely zero inclination to use any of Meta's products, and that doesn't exist with any other company.

基本上所有产品我都尝试过。

Like, I've experimented with pretty much everything.

Meta的AI产品堆栈中没有任何一个部分能引起我丝毫兴趣。

There is not a single part of any of my Meta's AI product stack that I'm remotely interested in.