SpaceX Nears $60B Cursor Acquisition, Meta’s Open Source Model Family, Microsoft’s AI Chip Ramp Up

欢迎大家收看The Information的TITV节目。

Welcome everyone to The Information's TITV.

我是Stephanie Palazzolo，今天是8月10日，星期一。

My name is Stephanie Palazzolo and it's Monday, August 10.

在节目开始之前，英特尔宣布将通过发行股票筹集150亿美元，以支持公司在竞争激烈的芯片市场的扩张。

Before we start the show, Intel announced it will be raising $15,000,000,000 in stock to fund the company's expansion in the competitive chip market.

今天节目中，我们将了解SpaceX完成对Cursor高达600亿美元收购的时间表，以及这笔交易对Cursor品牌的影响。

And today on the show, we'll learn more about SpaceX's timeline to complete its massive $60,000,000,000 acquisition of Cursor and what the deal means for Cursor's brand.

接着，我们会分析Meta在继续与OpenAI和Anthropic竞争的同时，推出其最强大AI模型作为开源模型的举措。

We'll then break down Meta's push to release its most powerful AI models open source as it continues to battle OpenAI and Anthropic.

我们会和Amp、Slack的CEO聊一聊。

We'll talk to the CEO of Amp, Slack, that.

另外，我们将解读微软计划明年大幅提升其自研AI芯片产量的策略。

Plus, we'll unpack how Microsoft plans to significantly ramp up production of its homegrown AI chips next year.

最后，我们会探讨为什么AWS的领导者要求工程师减少CPU浪费。

And to wrap it up, we'll take a look at why AWS leaders are telling their engineers to cut back on CPU waste.

今天节目会很精彩，那我们就直接开始吧。

It's going to be a great show, so let's get right on into it.

《The Information》独家披露了SpaceX距离完成以六百亿美元收购Cursor有多近，以及这笔交易对Cursor品牌意味着什么。

The Information published exclusive details about how close SpaceX is to closing its $60,000,000,000 acquisition of Cursor and what that means for the Cursor brand.

我们的埃隆·马斯克记者凯写了这篇报道，现在她来和我们聊聊。

Our Elon Musk reporter, Kay, wrote the story and joins me now.

欢迎来到节目，格蕾丝。

Welcome to the show, Grace.

谢谢。

Thanks.

那你到底从你的报道中发现了什么？

So what exactly did you learn in your reporting?

对。

Yeah.

我了解到Cursor周四开了个全员大会，会上他们告诉员工，和SpaceX的交易最快可能在这周末就完成。

So I learned that Cursor had an all hands on Thursday with staff and they told them that the deal with SpaceX could close as soon as the end of this week.

他们还告诉员工一些关于未来Cursor产品品牌更新的消息，也稍微透露了一下会怎么融入到SpaceX的AI部门里。

They also told them some updates regarding future branding for Cursor products and gave them a little bit of information about how they're going to integrate into SpaceX AI division.

所以，你知道，你之前报道过，Cursor的CEO迈克尔·特罗尔跟员工说过，SpaceX AI觉得Cursor这个品牌很有价值。

So, you know, you've previously reported that, you know, Cursor CEO Michael Troll has told employees that SpaceX AI viewed Cursor's brand as valuable.

现在看起来他们好像在品牌策略上要变卦了。

Now it seems like maybe they're switching up on the branding.

我猜这对员工来说可能挺震惊的。

I'm sure this came maybe as a shock to employees.

Cursor的员工怎么看这个变化呢？

How are Cursor staffers viewing this, this change?

对，我觉得很多人确实挺惊讶的，因为迈克尔之前给他们的初步说法就是，这是我们想收购你们的原因之一。

Yeah, I think a lot of people were surprised because that was kind of the early pitch from Michael of like, this is part of why they want to acquire us.

这肯定引发了不少疑问。

And it definitely spurred some questions.

所以全员大会之后，他们开了个问答环节，有些员工就问，未来Cursor的产品可能会改用Grok这个品牌，而不是用Cursor自己的名字，这背后的考量是什么，毕竟Cursor在编程圈子里口碑挺好的。

So after the all hands, they held a Q and A, and some staff asked about, you know, what would be the reasoning behind potentially having future Cursor products, you know, under the Grok brand name versus the Cursor brand name, you know, which is viewed pretty well in the the coding community.

还有一些讨论是关于可能起一个新品牌名，既不是Cursor也不是Grok。

There was also some discussion around, you know, the potential of a new brand name that would be neither Cursor or Grok.

我很好奇，你觉得哪个品牌在这里最有分量？

And I guess I'm just curious in your opinion, you know, which brand do you think kind of holds the most weight here?

是Cursor、SpaceX，还是像你提到的，可能是个新名字？

Cursor, SpaceX, or you know, maybe something new as you mentioned?

对，我觉得这要看是哪个群体。

Yeah, mean, think it depends on the community.

Cursor显然有很多企业客户关系。

I mean, Cursor obviously has, you know, a lot of enterprise relationships.

它在工程师圈子里非常有名。

You know, it's very well known, you know, with like engineers.

所以我觉得Cursor绝对是个公认的品牌。

So I I think, you know, Cursor is definitely a well recognized brand.

那这些Cursor员工在新的合并公司里会怎么安排？他们会在SpaceX AI里有自己的部门，还是和其他团队合并？

And in terms of where those Cursor employees are going within the, within the new combined company, are they going to have their own unit within SpaceX AI or are they going to be combined with other teams?

那里发生了什么？

What's, what's going on there?

是的，那是我一开始就很好奇的事情。

Yeah, that was something I was really curious about going into it.

你知道，比如，他们会不会只是SpaceX AI里的一个编程团队？

You know, like, would they just be this coding unit within SpaceX AI?

但看起来他们会整合到整个公司里。

But it looks like they're gonna be integrated across the company.

所以他们会直接和SpaceX AI的其他人一起工作，还会做一些不完全是编程相关的项目。

So they will be working directly with other SpaceX AI people, and they'll be working on products that aren't necessarily coding focused.

公司告诉员工，收购完成后的几天内，他们会完全搬到SpaceX AI的Slack里。

The company told staff that, you know, within a few days of the deal closing, they're going to fully move into the Slack, you know, SpaceX AI Slack.

你知道，他们会融入不同的团队，跟着不同的领导。

You know, they're going to incorporate themselves into different under different leaders.

他们对此怎么看？

How do they feel about that?

我觉得，他们大多数人加入Cursor，是因为Cursor是一家做编程AI的初创公司。

I mean, I'm sure most of them join Cursor because Cursor is a coding AI startup.

对。

Yeah.

我觉得，这个情况挺复杂的，因为就像你说的，大家是因为Cursor这个品牌才来的，现在却要归到另一家公司旗下。

I think, I think it's complicated because yeah, like you said, like people joined Cursor for Cursor's brand and now they're coming under this other company.

我觉得收购中经常发生这种情况，就是不是每个人都会满意，而且收购之后马上会有很多变化。

Think I this is something that happens a lot with acquisitions is like, you know, not everyone's gonna be happy about it, and there's probably gonna be a lot of change, you know, immediately after the acquisition.

我觉得大家总是担心裁员，或者削减开支会是什么样子。

I think there's always concerns about layoffs, you know, or or what cuts could look like.

所以，我觉得肯定心情复杂。

So yeah, I think mixed feelings for sure.

那接着这个说，这对员工来说是不是有离职风险？

And you know, following, following up on that, is that a flight risk then for employees?

你觉得会不会有人因此想离开？

You know, do you think some are going to want to leave following this?

现在公司里的士气怎么样？

How does that sort of like morale feel right now at the company?

是的，我觉得，你知道，士气在这方面是复杂的。

Yeah, I think, you know, morale is, is mixed when it comes to this.

肯定有人在反对，比如在问答环节，人们问关于品牌的问题，而且对Grok品牌似乎不太满意。

There's definitely people who seem to be pushing back, like in the Q and A people were asking about branding and, you know, maybe didn't seem as happy about Grok branding specifically.

所以，确实有人可能会离开。

So yeah, there is an opportunity for people to leave.

我觉得已经有人离开了。

I think some people have left already.

所以我肯定在密切关注这件事。

So it's something I'm looking at very closely for sure.

退一步说，你觉得Cursor和SpaceX或者SpaceX AI之间的关系在短期内会怎么发展？

And just taking a step back, I mean, how do you see this relationship broadly between Cursor and SpaceX slash SpaceX AI evolving in the, in the near term?

是的，这很有意思，因为他们之前就已经紧密合作了，因为除了收购之外，他们还有合作关系。

Yeah, it's, it's interesting because they've already been working closely together because they had this partnership in addition to the acquisition.

所以我们知道他们会来。

So, you know, we're expecting them.

马斯克说过他们会发布另一款联合模型。

Elon Musk has said they're gonna release another joint model.

他们已经发布了一个，所以这会是第二个。

They already released one, so this would be the second one.

你知道，光标即使在收购之前，也在尝试扩展到编程以外的领域。

You know, Cursor, even before this acquisition, you know, was trying to broaden outside of just coding.

所以，我认为我们会继续看到光标的发展，也会继续看到这种合并，看看事情会怎么发展。

So, you know, I think we're gonna continue to see Cursor evolve and, you know, continue to see this kind of merger and like, look at how that's going to play out.

太好了。

Great.

嗯，那我们以后得再请你来聊聊这些。

Well, we'll have to have you back on to talk about all of that.

再次非常感谢你，格蕾丝。

Again, thanks so much, Grace.

刚才那是我们的埃隆·马斯克记者格蕾丝·凯。

And that was Grace Kay, our Elon Musk reporter.

Meta首席执行官马克·扎克伯格今天早上宣布，公司将开源其新AI模型Muse Spark 1.2，并发布一系列新的开源模型，这些模型设计用于在笔记本电脑上运行。

Meta CEO Mark Zuckerberg announced this morning that the company will open source its new AI model, Muse Spark 1.2, and release a new family of open source models designed to run on laptops.

为了更深入了解这对AI领域可能意味着什么，我们请来了Amp的联合创始人兼CEO奎因·斯莱克。

For more on what this could mean for the AI landscape, let's bring in Quinn Slack, co founder and CEO at Amp.

欢迎来到节目，奎因。

Welcome to the show, Quinn.

谢谢。

Thanks.

很高兴来到这里，今天有大新闻。

It's great to be here and big news today.

确实。

Definitely.

没错，周一早上有很多大新闻。

No, lots of big news on a Monday morning.

回顾一下，显然Meta曾经是美国开源AI的领导者，但你知道，真的开源其模型是在一年前，当时它在AI竞赛中落后了。

So taking a step back, obviously at one point Meta was kind of the leader for American open source AI, but you know, really open sourcing its models a year ago as it fell behind in the AI race.

你怎么看待它最近宣布再次开源其模型这件事？

What do you make of its recent announcement to start open sourcing its its models again?

这对所有人来说都是一个好迹象。

It's a great sign for everyone.

而且我认为，对于美国的每一个使用模型的个人，每一家使用模型的公司，他们都不想支付Anthropic特别是有OpenAI那样的高昂的token费用，而且还有一些监管方面的尝试，特别是由Anthropic主导的，似乎正在进行中。

And I think for every person in America who's using models, for every company who's using models, they wanna know that they don't have to pay exorbitantly priced tokens from Anthropic, in particular in OpenAI, and there were some regulatory capture attempts that seemed to be underway, in particular led by Anthropic.

Meta这样的巨头以及NVIDIA也大力支持开源，看到这些，我认为这对于美国人能够受益于使用AI来说是一个好兆头，而不仅仅是得益于这些大型AI公司位于他们的国家。

And the fact that Meta, with their big guns and others, see NVIDIA coming along too, being a huge champion of open source, think I it's a really good sign for Americans being able to benefit from using AI, not just benefit from having the big AI companies in their country.

所以看起来，你知道，更多的选择总是好事，尤其是如果这些选择是美国本土的，并且可能没有中国模型那种风险，比如可能会被禁用或其他限制。

So it seems like, you know, more options is always a good thing, especially if they're American and maybe don't have that same type of risk that Chinese models have where, you know, we might see like a ban or other restrictions on using them.

在过去的一年里，Meta确实停止了开源模型，许多替代方案，尤其是中国的，纷纷涌现。

You know, in, in the past year, while Meta has really stopped open sourcing its models, a lot of alternatives, especially Chinese ones, have popped up.

你觉得现在Meta要想重新夺回领先地位，比一年前它似乎在全球领先时要难多少呢？

I mean, much harder do you think is it now for Meta to try to regain its lead versus, you know, a year ago whenever it seemed to be more on top of the, of the world?

我觉得他们正在飞速前进。

I think that they are roaring forward.

永远不要低估一家公司，尤其是在这个一切似乎都更加动荡的时代。

Never count out a company, especially in this day and age when it seems like everything is more volatile.

在AI领域，你会看到人们比以往任何时候都更快地重回巅峰。

And with AI, you see people, you know, coming back on top faster than ever.

我觉得有很多人希望Meta赢，希望NVIDIA赢，也希望小公司赢，我们可以通过这种愿望把它变成现实。

I think that there is so many people that want Meta to win, that want NVIDIA to win, that want the the little guy to win now, that we can wish cast this into existence.

这意味着NVIDIA会给Meta优先使用GPU的权限，帮助他们领先，同时也能在一定程度上控制Anthropic能领先多少。

And, you know, what that means is there's gonna be NVIDIA giving preferential access to GPUs, to Meta, to help them get ahead, and they can kind of control, you know, how much Anthropic can pull ahead.

显然，NVIDIA不是唯一的关键，但除了Anthropic自己，没有人希望它完全主导。

And, obviously, NVIDIA is not the only game here, but you there's nobody except for Anthropic that wants a totally Anthropic dominance.

所以有很多人非常希望Meta能强势回归。

So there's a lot of people with a big interest in making it so that Meta roars back.

所以，我完全支持这一点。

So I think I am all for this.

在Amp，我们确实在一些较低级别的模式中使用了中国的模型，我们和其他人一样，我觉得，我们必须非常清楚地披露这一点，这样如果人们不想用这些模型，他们就不会用。

And in Amp, we, yeah, have used some of the Chinese models for some of the lower modes and we like everyone does, I think, you know, we have to disclose this really clearly so that if people don't wanna use those models, then they don't use them.

但如果是NVIDIA或Meta的开源模型，没有人会反对使用。

But when it's an NVIDIA model or a Meta model that's open, there's nobody who's gonna be against using that.

所以这很了不起。

So that's huge.

有趣的是，我们现在把NVIDIA和Meta称为小公司，而把OpenAI和Anthropic称为大公司，尽管它们显然都是大型科技公司。

It's funny that we're in a world now where we call NVIDIA and Meta the little guys compared to OpenAI and Anthropic when obviously there are these huge tech companies.

我也很好奇，Meta开源这些新模型会怎么样。

I mean, I am curious too in terms of, you know, Meta open sourcing these new models.

显然，这些公司之间竞争非常激烈，都想发布特别好的模型，尤其是在编程方面。

There's obviously just so much competition from all these companies that want to release really great models, especially in coding.

你觉得有没有哪些领域被忽视了，或者Meta有什么特别的事情可以做，来重新夺回开源领域的领先地位，而其他人可能忽略了这些？

Are there any kind of like underserved areas or anything special you think Meta can do to try to like regain its lead in open source that maybe other people are overlooking?

嗯，我觉得，即使是编程，虽然它是AI第一个真正流行和赚钱的应用，AI也在这方面变得非常出色，但变化还是很大。

Well, there are I think with coding, even for coding, even though that was the first application of AI to really get popular and lucrative and for AI to get really good, it's changing so much.

而且你需要使用的令牌数量正在大幅增加。

And the number of tokens you need to use is going up so much.

所以在过去六周里，我们看到了一个巨大的变化：人们不再只是在自己的电脑上运行一个代理，一次只做一件事，然后盯着它看。

So just in the last six weeks, we have seen a huge shift from people running a single agent on their computer, doing one thing at a time, and they're watching it.

现在他们同时运行并照看五到十个不同的任务。

Now they're running and babysitting five or 10 different things.

在某些情况下，这些代理还会生成其他代理。

And in some cases, those are spawning other agents.

所以即使在编码这个用例中，你也会看到对令牌的需求不仅仅是最昂贵的那些。

So even in the coding use case, you're seeing there's an appetite for tokens that are not just the most expensive tokens.

而且我认为在其他应用中，当然，你可能会找到更便宜令牌的用途，但让我们看看最成熟的应用，以及它变化得有多快。

And I think that there's other applications where, yeah, sure, you could find a use for cheaper tokens, but let's look at even the most mature one and how fast that's changing.

我认为，即使是最成熟的应用也在寻找更多使用优质但可能不是GBD 5.6或Fable级别的令牌的方式，那么在其他领域，即使它们达到了编码目前的成熟度，对令牌的需求也是永无止境的。

And I think that if even the most mature one is finding more ways to use tokens that are good, but maybe not GBD 5.6 or Fable level, then whatever happens in the other realms, even when they achieve the current maturity the coding has, yeah, there's there's just such an insatiable appetite for tokens.

这意味着，即使你从未使用过Meta的模型，你仍然会从它们给其他公司带来的价格竞争中受益。

And what this means, even if you never use any of the Meta models, you're still benefiting from the price competition that they're creating on everyone else.

对。

Yeah.

看起来Meta确实在加大投入，专注于更高效的模型和降低成本，今早他们发布的一些消息也印证了这一点。

It does really seem like Meta is doubling down on this idea of, you know, more efficient models, lowering costs with some of its most recent announcements from this morning.

你知道，他们发布了Muse Glimmer，这是一个新系列模型，体积小到可以在笔记本电脑上运行。

You know, they obviously announced Muse Glimmer, which is, a new family model that's small enough to run on laptops.

这似乎和最近的发布有些不同，比如Chemie K3，那个模型其实挺大的，开发者用起来可能有点困难。

And that does seem like a bit of a divergence from, you know, recent releases like Chemie K3, which is actually like quite large and that can actually make it somewhat difficult for developers to use.

你觉得这对他们来说是正确的方向吗？继续押注更小、更便宜的模型？

Does that kind of feel like the right move for them and this kind of doubling down on, you know, smaller, cheaper models?

我觉得这里有一个复杂的边界。

I think that there's a complex frontier.

你知道，有很多不同的场景需要好的模型，而那个领域相对被忽视了。

You know, there's there's a lot of different areas where you want a good model, and that was a relatively underserved area.

之前有Google的Gemma，它小到可以在单台机器上运行。

You had Gemma from Google that was, know, small enough to run on a machine.

还有其他一些模型，但根据早期报告来看，这似乎是目前能在笔记本电脑上运行的最好的美国模型。

There's some other ones, but now I think, you know, based on the early reports, this does seem to be the best US model that you can run on your laptop.

如果这是Meta获取关注度、逐步攀升的方式，最终他们可能会希望在某些能力上超越Anthropic和OpenAI的模型，虽然可能不是所有方面，但你得一步步来。

And if that's a way for Meta to go gain mindshare and to start climbing the ladder toward, you know, the point where I think eventually they will want to have a model that is better than Anthropics and OpenAI's, maybe in certain capabilities and not others, but you gotta work your way up.

而且我觉得Meta也明白保持相关性的重要性。

And I think Meta also understands the importance of being relevant.

他们必须出现在新闻里。

They have to be in the news.

他们必须展示进展，因为他们是小角色。

They have to be showing progress because they are the little guys.

如果他们两三个星期不发布任何东西，他们就会把注意力让给别人，人们可能就会开始不看好他们。

And if they don't release anything for two or three weeks, they're ceding that attention to everyone else, and people will probably start counting them out.

所以我觉得你可以期待那些努力往上爬的公司，我为他们加油，我想其他人也会为他们加油。

So I think you can expect to see anyone who's trying to claw their way up, which I cheer, and I think everyone else cheers.

他们会发布各种填补大模型公司目前没有覆盖的空白的产品。

They're gonna be releasing all kinds of things that that fit gaps that the big model companies are not filling right now.

虽然可能感觉有点怪，但突然之间他们就会在排行榜上一步步往上爬了。

And it might feel weird, but then all of a sudden they're gonna be working their way up the leaderboard.

我觉得这是个好策略。

I think it's a good strategy.

对。

Mhmm.

确实感觉像是在争夺

It does feel like there is this race to

某种

kind of

头条的主导权。

dominate headlines.

就像你说的，如果哪个公司沉默了几周，哪怕只是几周，他们很快就会失去聚光灯，然后把注意力让给其他AI初创公司。

Like you're saying, if you miss out, if there's a couple, even a couple of weeks of silence from any one company, it does feel like they will lose that spotlight very quickly and cede it to other AI startups.

我是说，上周我们确实请了人聊了聊新的Meta模型，Muse Spark 1.2。

I mean, so we did have someone on last week to talk a bit about the new Meta Model, Muse Spark 1.2.

他们好像对它有点复杂的感受。

They seem to have kind of mixed feelings about it.

有好有坏吧。

Some good, some good, some bad.

你和Amp的其他人试过这个模型了吗？

Have you or others at Amp tried it out yet?

我很好奇你们有没有什么初步的想法或者感受，觉得这个模型表现怎么样。

I'm curious if you have any initial kind of thoughts or, you know, feelings about how well the model does.

是的，我们试过Muse Spark，它是个非常好的模型。

Yeah, we've tried out Muse Spark and it's a very good model.

至少在Amp，这个模型主要用于软件开发，如果你想留在美国的开放模型生态里，而且想要一个比较便宜的模型，你就会用它。

It's a model that at least in Amp, which is primarily used for building software, you would use it if you want to stay in the American open model ecosystem for one of the cheaper models.

这对一些大银行来说很有吸引力，不过，你知道，在我们的客户群中。

And that is something that appeals to a few big banks, but, you know, among our customer base.

但实际上，我们看到中国模型甚至在大银行中也获得了相当令人惊讶的接受度，因为他们知道他们需要把这个作为备选方案。

But we'd actually seen a pretty surprising acceptance of Chinese models even among big banks because they know that they needed to have that in their back pocket.

所以他们都已经批准了那些模型，我觉得这一点很少听人提起。

So they had gotten those approved, which is something I don't think you hear a lot of people talking about.

但我觉得Muse Spark的影响以及我们在产品中使用它的方式是，既然现在有了美国的选择，随着前沿模型的发展，这些银行将不再批准那些处于前沿的中国模型，而美国的模型会超越它们。

But I think the impact of Muse Spark and the way that we'll use it in our product is now that there is an American alternative, as the frontier goes on, these banks will no longer be approving the Chinese models that are on the frontier, and the American ones will surpass them.

然后我们就会处于这样一种情况：也许你的支出中80%会花在Anthropic和OpenAI上，20%花在美国的开放模型上。

And then we'll be in a place where, you know, perhaps your usage will be 80% of the spend on anthropic and open AI and 20% of the spend on American open models.

让我们都希望那20%能不断增长。

And let's all hope that that 20% grows and grows.

这很有意思。

That's very interesting.

所以听起来确实是这样，那些大公司和像金融服务这样受严格监管的行业，现在被迫使用中国的开源模型，因为确实没有其他选择。

So it does sound like there, you know, there are these large companies and pretty regulated industries like financial services that are being kind of forced to use Chinese open source right now because there's just not really any other option out there.

而且听起来这对Meta来说是个机会，如果有一个美国开源选项，而且价格相对便宜，就能从那些中国开源模型中抢走一部分份额。

And it sounds like there's this kind of like opening for Meta if, you know, if there are this American open source option, that's fairly cheap to kind of take some of that share from those other Chinese open source models.

是的，完全同意。

Yeah, absolutely.

而且切换起来非常快。

And it's very fast to switch.

所以这是可以发生的。

So this can happen.

对，而且你很少听到公司谈论他们开始使用中国模型或已经批准这些模型的事情。

And yeah, you don't hear a lot of companies talk about when they are starting to use the Chinese models or when they've approved them.

他们不想谈。

They don't want to.

有些情况下，他们这样做只是为了下次Anthropic来找他们谈大额费用或要求大笔承诺时，他们能有个备选方案。

In some cases, they're just doing so, so that next time Anthropic goes to them and tries to charge them a lot or get them to do a big commit that they have some alternative.

你之前稍微提到了你的一些客户在做什么。

And, you know, you kind of touched on a little bit about what some of your customers are doing.

我很好奇，总体来说，你在编程AI领域看到了哪些新趋势？

I'm curious, just like broadly, what, what new trends are you seeing in the world of Coding AI?

有没有哪些特定模型在Amp客户中特别受欢迎，或者有没有什么新类型的任务，是开发者现在能做到，而六个月或十二个月前做不到的？

Are there any specific models that are really taking off with Amp customers or, you know, new types of tasks that developers can do now that they couldn't do, you know, six, twelve, twelve months ago?

我觉得在过去的六到八周里，大家真的开始用云端智能代理了。

I think in the last six weeks, eight weeks, you're seeing people actually use agents on the cloud.

我们在AMP里把它们叫做orbs。

We call them orbs in AMP.

这种方式让你可以合上笔记本电脑。

And this is something where you can close your laptop.

你可以在iPad上、iPhone上操作。

You can do it on your iPad, on your iPhone.

这就是智能代理的承诺。

That's the promise of agents.

人们拿这个开玩笑很久了，但现在真的实现了，因为智能代理和模型已经足够好，能监督整个流程，而且你已经让足够多的人和公司改变了他们的工作方式。

People joked about that for a long time, but now that's actually possible because the agents and models are good enough at supervising what's going on, and you have gotten enough people, enough companies where they've changed how they work.

所以如果智能代理真的能让他们开发软件快五到十倍，那其他流程环节就不再是瓶颈了。

So that if the agents do let them build software five or 10 times faster, then all the other parts of the process are not gonna be blockers anymore.

我的意思是，在很多公司里它们还是瓶颈，但你开始看到进展了。

I mean, they are in a lot of companies, but you're starting to see progress there.

你开始看到软件工程团队的形态在变化，这样5倍或10倍的交付速度才能真的把产品送到客户手中。

You're starting to see the shape of the software engineering team change so that five or 10 times faster delivery actually can get in the hands of customers.

这向来是瓶颈所在。

That's always been the bottleneck.

人一直是瓶颈，但现在我们终于看到这个瓶颈松动了，感觉太棒了。

The humans are the bottleneck here, but we're seeing that finally lift and it feels incredible.

感觉像是巨大的加速。

It feels like incredible acceleration.

对。

Yeah.

这让我想到，我记得Anthropic最近也宣布，在某些套餐里把自动模式设为默认，意思就是，你不必再手动检查AI做的每一件事了。

It kind of reminds me of, I think Anthropic also recently announced that they're making auto mode the kind of default with some of their, you know, plans, which is this idea that, you know, you don't have to manually check every single thing that the AI does anymore.

听起来你们那边也有类似的情况，开发者们对AI的能力越来越有信心，让它在云端自己跑，几乎不用盯着。

It sounds like maybe you're seeing something similar where developers are becoming more, you know, confident with the AI being able to do things by themselves and being able to run by themselves in the cloud without not much oversight.

这么理解对吗？

Is that is that fair?

对。

Yeah.

确实如此。

That's absolutely fair.

而且如果你想让人来监督，他们其实根本看不懂。

And if you try to have the humans provide oversight, they don't understand.

我觉得Anthropic那个研究特别有意思，就算让人去批准或者拒绝AI想做的步骤，他们犯的错比让AI自己批准或拒绝自己的步骤还要多得多。

I think the Anthropic study was really interesting showing that even if humans were asked to approve or reject the steps that the AI wanted to take, they made so many more mistakes than if the AI was in charge of approving and rejecting the steps that it wanted to take.

这对人类来说挺打击人的，但说到底，跟你去雇人没什么两样。

And that's a humbling thing for humans, but it's really no different from if you go and hire people.

你确实得放弃一些微观管理，为了这个我们已经建立了很多机制，而且我们能搞定这些。

You do have to give up some of your micromanagement, and that's something that we've built all kinds of structures around, and and we can figure that out.

不过没错，Anthropic最近就是这么做的。

But, yes, Anthropic just did that.

我觉得AMP用户大概十二个月前就已经意识到这才是正确做法了。

I think AMP users have been realizing that's the right thing for about twelve months now.

所以我们领先了，但看到他们在追赶也是好事。

So we are ahead, but it's good to see them catching up.

对。

Mhmm.

我觉得区别在于，让AI自动做它想做的事和让人类员工做，最大的差别是，你至少可以指望人类员工不会逃出沙箱去入侵另一家创业公司。

I guess the difference between, you know, letting your AI kind of automatically do what it wants in a human employee is that hopefully you could trust a human employee to not, you know, escape a sandbox and hack into another startup.

没错。

Right.

不过，这确实有点吓人。

But I mean, it's a little bit scary.

我得说。

I do have to say.

是的。

Yeah.

关于OpenAI和Hugging Face事件的黑客演讲，是最值得听的内容。

The black hat talk about the OpenAI Hugging Face incident is the most interesting thing to listen to.

我一直在玩这个，就是和家人朋友在一起，他们都不是搞软件的。

I've been playing it just, you know, hanging out with family and friends and people that are not software developers.

这真是个不可思议的故事，而且这才刚刚开始。

It is an incredible story, and that's just the beginning.

所以，是的，这确实很吓人，就像所有新事物一样，但我们有很多聪明人。

So, yes, it's it's absolutely scary like all new things are, but we have a lot of smart people.

我们有很多聪明的模型在努力让这一切更安全，而且没有暂停这个的选项。

We have a lot of smart models that are working to make this more secure, and there's no option to go and pause this.

实际上，你知道吗，如果有办法暂停它，那会比所有聪明人怀着善意努力让它安全更吓人。

Actually, you know, if there was some way to pause it, then that would be scarier than all of the smart people working with their best intents to try to, you know, make this safe.

因为如果有人能暂停它，那那个人就会拥有所有的权力。

Because if there was someone who could pause this, then that person would have all of the power.

而且我觉得我们美国人一般不会信任这种情况发生。

And I don't think that we as Americans generally trust that happening.

所以让我们全速前进吧。

So let's move full speed ahead.

让我们拥有自由。

Let's have freedom.

让我们拥有更多开放模型，让除Anthropic以外的公司也能制定政策。

Let's have more open models that bring in the ability for companies other than Anthropic to set the policy.

我认为这样对我们最有利。

I think that's the way we'll be best off.

但确实，外面的世界很疯狂。

But yes, it's a crazy world out there.

完全同意。

Totally.

嗯，接下来有很多事情要关注。

Well, lots to keep on, lots to keep on top of coming up.

太好了。

Great.

好的，再次感谢Quinn Slack，Amp的联合创始人兼CEO。

Well, again, that was Quinn Slack, co founder and CEO at Amp.

感谢你来做客，Quinn。

Thanks for coming on, Quinn.

微软计划明年大幅增加其内部设计的下一代AI芯片的产量，希望能说服像Anthropic这样的大型云客户使用这些芯片。

Microsoft is planning to significantly increase production of its internally designed next generation AI chips next year in hopes of persuading big cloud customers like Anthropic to use them.

让我们请出The Information的微软记者Aaron Holmes。

Let's bring in Aaron Holmes, our Microsoft reporter at The Information.

欢迎来到节目，Aaron。

Welcome to the show, Aaron.

你好。

Hello.

很高兴来到这里。

Happy to be here.

那么到目前为止，微软在芯片战中的立场到底是什么？现在又有什么变化呢？

So what exactly has been Microsoft's position in the chip war so far and what's changing now?

对。

Yeah.

所以微软正在尝试做谷歌和亚马逊已经做得更成功一些的事情，那就是自己研发和生产AI芯片，来跟英伟达在GPU市场上的垄断地位竞争。

So Microsoft is trying to do what Google and Amazon have been a little bit more successful at so far, which is to build and produce its own AI chips that could compete with NVIDIA's stranglehold on the GPU market.

而且，到目前为止，跟谷歌和亚马逊不一样，微软还没能把这些叫Maya的芯片真正租给任何客户。

And, you know, so far, unlike Google and Amazon, Microsoft has not been able to actually rent these chips called Maya to any of its customers.

他们最大的成就也就是在少数几个数据中心内部使用Maya，而不是用英伟达的芯片。

The most that they have achieved is just sort of using Maya internally in a small number of data centers, instead of NVIDIA chips.

但最近，微软对下一代Maya芯片，也就是Maya 300，更有信心了，觉得它会比现在的这一代更成功。

But more recently, Microsoft has become more confident that the next generation of Maya, which is called Maya 300, will be more successful than the current generation.

正因为这样，他们正在大幅提高生产计划，明年以及未来几年可能会生产几十万甚至上百万颗Maya芯片。

And because of that, they are, dramatically ramping up their plans to produce, potentially hundreds of thousands or even millions of Maya chips next year and in the years ahead.

那他们为什么这么有信心Maya 300会火起来呢？毕竟到目前为止，看起来好像还没火起来。

And why are they so confident that Maya 300 is gonna catch on when, you know, so far it doesn't really seem like it it it has.

有几个原因。

So there's a few things happening.

一个是Anthropic一直在跟微软谈，打算明年开始用Maya 300。

One is that Anthropic has been in extended talks to use Maya 300 starting next year.

我听说到目前为止还没有签任何协议，但Anthropic已经表现出愿意租用任何能拿到的AI芯片。

I'm told that so far, still no deal has been signed, but Anthropic has shown, you know, this willingness to rent any sort of AI chips it can get.

微软也很乐观，觉得自己能向Anthropic证明这些芯片足够好，值得租用。

And Microsoft feels optimistic that, you know, it can prove to Anthropic that these chips are good enough to to be rented.

另外，微软在云服务商中有点特别，因为他们也卖很多使用AI模型的软件。

And then the other piece is that Microsoft is a little bit unique among the cloud providers in that they also sell a lot of software that uses AI models.

所以，因此他们内部对运行模型的需求特别大，尤其是他们自己的第一方软件Copilot。

And, you know, as a result, they have this pretty big internal need to run the models, for their own first party software copilot.

所以微软开始更有信心，即使找不到很多大客户来用Maya，他们也可以把自己更多的AI Copilot使用量转移到Maya芯片上，而把最先进的Nvidia芯片留给Azure上的云客户。

So Microsoft is starting to get more confident that, you know, even if it can't find a lot of big customers for Maya, it can shift more of its own AI Copilot usage onto Maya chips and save the, you know, state of the art Nvidia chips for more of its cloud customers on Azure.

确实，特别是Anthropic，因为它面临的计算资源紧缺非常严重，一直是很多芯片开发商的重要推动力。

I mean, does seem like Anthropic in particular because it's faced this, you know, pretty insane compute crunch has been a big driver for a lot of chip developers out here.

我们之前写过一些较小的芯片初创公司和Anthropic洽谈的事。

We've obviously written about some smaller chip startups that have been talking with Anthropic.

现在显然更大的芯片供应商也在做同样的事。

Now we see obviously larger chip providers doing the same.

我很好奇，微软打算怎么在持续的存储芯片紧张的情况下实现这个计划。

I'm curious how Microsoft kind of plans on pulling this off with the ongoing memory crunch.

其他组件也供不应求，而且台积电肯定有一长串公司想和它合作制造芯片。

Other components are in short supply here and obviously TSMC has, I'm sure, a very long list of companies that wanna build chips with it.

它打算怎么搞定这一切呢？

How does it plan on pulling all this off?

对。

Yeah.

我觉得这是个好问题。

And I I think that's a good point.

你知道，台积电是否愿意给微软分配产能来制造这些芯片，据我了解，这是微软能否启动这个项目的最大不确定因素。

You know, TSMC's ability to, or or willingness to grant capacity to Microsoft to manufacture these chips is sort of the biggest question mark per my understanding of of whether, you know, Microsoft will be able to get this effort off the ground.

对吧？

You know?

我觉得到目前为止，Maya 200芯片的总产量只有几万片，而微软想要扩大至少一个数量级，达到几十万片。

I think so far, the the Maya 200 chips, the total amount that were produced were in the tens of thousands, and Microsoft now wants to, you know, go at least in order of magnitude bigger into the hundreds of thousands.

我觉得要做到这一点，他们得持续向台积电证明这些芯片有稳定的需求。

And I think to do that, they're going to have to consistently, you know, show to TSMC that there is sustained demand for these chips.

我觉得如果能签下像Anthropic这样的大客户，那显然会是关键的一步。

I think signing a big customer like Anthropic would be obviously a linchpin to to do that.

但除此之外，我相信他们肯定也乐意去争取其他任何客户。

But, you know, I think barring that, they also, I'm sure, would love to to potentially, you know, ink any other customer.

OpenAI显然是一家微软已经租用大量云服务器的公司，而且微软已经在优化Maya来运行OpenAI的模型，因为微软自己就有权运行这些模型，并且也想在内部使用Maya。

OpenAI is a company that, obviously, Microsoft already rents a lot of cloud servers too, and it's already been, you know, optimizing Maya to run OpenAI models because Microsoft itself has the rights to run those models itself and and wants to, you know, use Maya internally.

所以我觉得微软得去争取这些大客户，证明他们真的愿意花钱租用这些Maya芯片。

So I think that it it's gonna have to chase down these big customers and prove that they're actually willing to pay to rent these Maya chips.

如果能做到这一点，那可能就足以让这个项目启动了。

And if it can do that, that would potentially be enough to to get the program off the ground.

对。

Mhmm.

这听起来有点像是进退两难的情况。

Seems like a bit a bit of a catch 22 there.

对，是吧？

Right?

得向台积电证明有需求，但没有芯片又很难有需求。

It has to prove to TSMC that there's demand, but then it's hard to get demand if they don't really have the chips.

所以我就不知道了。

So I don't know.

看来微软现在处在一个挺尴尬的位置。

It seems like a bit of a tough position for Microsoft to, be in.

是的。

Yeah.

我觉得真正的后路基本上就是自己用这些芯片。

And and I think that really the fallback there is just essentially using the chips itself.

微软的首席财务官艾米·胡德多次谈过他们其实在做一种平衡——既要继续通过向Azure客户出租GPU来加速Azure的增长，又要留一部分GPU来支持Copilot业务随着软件业务增长而扩大。

You know, Amy Hood, CFO of Microsoft has spoken at length about how they're essentially doing this balancing act where they want to continue to accelerate their Azure growth by renting GPUs to Azure customers, but they also have to save some of those GPUs, to continue to grow their Copilot usage as their software business grows.

理论上，如果能在全球的数据中心部署更多Maya芯片，他们就能更好地平衡这一点，把更多稀缺的英伟达GPU留给客户。

And in theory, you know, if they are able to get more Maya chips online in data centers across the globe, they could, essentially help balance that and and reserve more of their scarce NVIDIA GPUs for customers.

所以我的理解是，这就是未来一年的路线图。

So my understanding is that, you know, that is sort of the the road map for the year ahead.

而现在，微软只需要生产足够多的芯片，来证明这个转型是可行的。

And, for now, Microsoft just needs to produce enough chips that it can show that that transition is working.

如果他们真的开始更多地在内部使用这些即将推出的芯片，比如Maya 300，那我们能不能知道这会在多大程度上帮助他们，既能为客户释放更多GPU容量，又能在芯片更划算的情况下降低成本？

And if it does start using more of these, upcoming chips in house, Maya 300, do we have any sense of how much that will help them either opening up more GPU capacity for customers, but also maybe lowering costs if these chips are more cost effective?

是的。

Yeah.

微软已经做了一些内部测试。

So Microsoft has done some internal tests.

对于目前这一代的Maya芯片，他们展示在某些情况下，自家的AI模型和OpenAI的模型运行成本可以降低30%到40%，也就是说，在Maya芯片上运行模型比在NVIDIA GPU上花更少的时间和电。

So the current generation of Maya, they've shown that in some cases they can get their own in house AI models and open AI models to run for 30 to 40% cheaper, meaning, essentially, it takes, you know, less time and electricity to to run the models on Maya chips than it does on NVIDIA GPUs.

我听说，对于下一代芯片，微软可能最快下个月就会发布，测试结果甚至更好，成本节省也更多，同等对比下运行OpenAI和微软的模型，比在NVIDIA芯片上更有优势。

I'm told that with the forthcoming generation, which Microsoft is potentially gonna announce as soon as next month, that they've gotten even better results and and even more cost savings, you know, apples to apples running OpenAI and Microsoft models versus on NVIDIA chips.

所以，这肯定是他们想向潜在客户大力宣传的卖点。

So, you know, that that's something that they're certainly going to want to tout to to those potential customers.

而且我觉得，如果他们能证明这一点，特别是能证明在运行Anthropic的模型时也有同样的效果。

And I I think that, you know, if they can prove that out and, you know, especially show that the same is true for running anthropic models.

那这也许就足够有说服力了，至少能给他们带来一些芯片业务。

That could be enough of a selling point to, to at least get some business for the chips.

微软一开始是怎么会落到这个地步的，对吧？

How did, I guess Microsoft even end up in this position in the first place, right?

其实它当时是领先的，而且和OpenAI的合作也很早。

Like it was really ahead of the curve and, you know, the early partnership with OpenAI.

但现在看来，在自研AI芯片方面，它好像让亚马逊和谷歌抢先了。

But here it seems like it's it's kind of let Amazon and Google pull ahead, when it comes to building their own AI chips.

微软怎么会起步这么慢呢？

How did Microsoft, I guess, kind of get such a slow start here?

对，你说得没错。

Yeah, you're right.

确实，他们花了一段时间才找到节奏。

I mean, has taken them a while to hit their stride.

而且，你知道，就像我们去年报道的那样，Maya 200芯片实际上被推迟了好几次，因为微软最初无法从这些芯片中获得它想要的性能。

And, you know, as we reported last year, the Maya 200 chips, were actually delayed several times, just because Microsoft initially couldn't get the performance, that it wanted out of the chips.

结果，这也就把大规模生产的计划往后推了。

And as a result, you know, it sort of pushed back to the timeline to have them produced at scale.

而且，到去年春天为止，微软表示Maya 200芯片已经在两个数据中心投入使用，但很快会在国际上上线。

And, you know, as of last spring, Microsoft said that Maya 200 chips were alive in two data centers, but would soon go live internationally.

我听说，目前基本上还是只有美国的那两个数据中心，而且还没有更广泛地推广。

I'm told that it's still essentially just those two data centers in The US and, you know, still haven't ramped up more broadly.

所以我觉得，这已经证明了这件事很棘手，微软也确实很难把它做好。

So I I think it's, you know, been proven that this is tricky, and and Microsoft has had a hard time getting it right.

而且我们确实看到，在去年OpenAI的审判中，有一些萨提亚·纳德拉在2023年（我记得是2022年）发给副手们的邮件，基本上是在抱怨，在NVIDIA芯片上运行OpenAI模型太贵了。

And we actually saw, you know, in the OpenAI trial, you know, this past year, there were some emails that came out from Satya Nadella to his deputies back in 2023, basically bemoaning the fact or 2022, I believe, that that it was so expensive to, you know, run open AI models on NVIDIA chips.

他还抱怨说，某个负责运行这些OpenAI模型的业务部门亏损了40亿美元，而且他把这归咎于NVIDIA对市场的牢牢掌控。

And he complains about, you know, losing $4,000,000,000 on some business unit that was responsible for running those open AI models and the fact that NVIDIA had such a stranglehold on the market, you know, he cites as the reason for that.

所以我认为，很明显，把这件事做好对微软在AI领域的前进道路至关重要，尤其是要从NVIDIA和OpenAI那里夺回一些定价权和利润空间。

So I think that it's pretty clear that getting this right is key to to Microsoft's path forward in AI and, especially to, you know, getting some of that pricing power and margin power back from NVIDIA and from OpenAI.

明白。

Mhmm.

懂了。

Gotcha.

好的。

Alright.

非常感谢，Aaron。

Thanks so much, Aaron.

刚才那位是《The Information》负责微软报道的记者Aaron Holmes。

And that was Aaron Holmes, Microsoft reporter at The Information.

关于算力紧张的讨论，大多集中在GPU上。

Conversations around the compute crunch have largely centered on GPUs.

但我们的亚马逊记者Perloff写了一篇报道，讲的是AI热潮如何导致CPU等老技术出现短缺。

But our Amazon reporter, Perloff, wrote a story about how the AI boom is creating shortages in older technology like CPUs.

现在Catherine来和大家分享她的发现。

Catherine joins me now to share what she learned.

凯瑟琳，欢迎来到节目。

Catherine, welcome to the show.

很高兴你能来。

It's great to have you.

你好，

Hi,

那我们先从你的报道说起吧。

So let's start with your reporting.

AWS内部到底发生了什么？

What is going on inside AWS?

对。

Yeah.

工程师们已经被告知，他们需要尽量限制自己使用的计算资源。

So engineers have been told that they need to limit, you know, do what they can to limit the compute they're using.

这涉及到各种不同类型的EC2实例，这算是AWS提供云计算的基礎服务。

And that is for all different types of something called EC two instances, which is sort of the basic service, AWS uses for cloud compute.

而且我说的所有类型，也包括在CPU上运行的计算，还有在GPU和AI芯片上运行的计算。

And when I say all types, you know, that includes, compute that runs on CPUs in addition to compute that runs on, you know, GPUs and AI chips.

所以，从五月份开始工作的工程师们，都在努力提升效率，更加注意自己在软件开发过程中使用的计算资源。

So, you know, engineers that started in May are trying to sort of be more efficient, be more careful with, you know, the compute they're using as they work on software engineering.

但是，我采访的一位工程师说，他们仍然在获取所需计算资源方面遇到一些问题。

But, you know, know, one engineer I talked to said that they there are still sort of, issues getting the compute they need.

比如，我采访的那位工程师需要获取计算容量，现在可能要等好几天，而以前只需要几个小时。

Like, you know, it could take days to get the the the capacity that, you know, one engineer I spoke with needed, whereas before, it would only take a couple hours.

所以，AWS内部确实存在一些限制，这些限制影响了CPU。

So, there are some constraints happening within AWS, and they're affecting CPUs.

是的。

Yeah.

我们在节目里经常讨论，AI需求导致了GPU短缺。

And you know, we have talked a lot on the show about how AI demand has created a shortage of GPUs.

你现在说的是另一种计算资源——CPU，它存在的时间要长得多。

And you're talking about a different type of compute CPUs, has been around for a lot longer.

为什么它们现在也开始成为故事的一部分了？

Why are they now kind of becoming part of the story?

是的，你知道，原因有很多。

Yeah, you know, there are a lot of reasons for it.

我觉得还有一点很重要，就是工程师们被告知，嘿，你要小心使用这个容量资源。

I think it's also important to say that like, you know, the engineers were told like, hey, you know, be careful how you use this resource of capacity.

这不一定意味着只有CPU在短缺。

That doesn't necessarily mean that it's only the CPUs in play that are in shortage.

比如CPU也需要和内存芯片配合工作。

Like CPUs also work with memory chips.

所以，我们知道内存供应紧张，内存短缺可能也是原因之一。

So, you know, we know memory is in short supply, so shortages in memory could be contributing to this.

物理数据中心空间不足也可能导致这个问题。

Shortages in physical data center space could be contributing to this.

我们只知道它影响到了运行在CPU上的服务。

We just know it's hitting a service that runs on the CPUs.

所以，你知道，这未必只是因为现在CPU需求增加才影响到的，我们也不太确定到底哪个部件最紧缺。

So, know, you it's not necessarily only the fact that CPUs are in more demand now that is affecting this, and we're not exactly sure what what component is the most, in short supply.

话虽这么说，你知道，现在CPU的需求确实增加了。

Having said that, you know, CPUs are in more demand now.

它们现在被大量使用，尤其是在智能体工作负载中。

They are being used, especially in agentic workloads.

很多时候，如果你试图运行一个智能体来写代码或构建软件，你会在云端用CPU来运行，或者在一台可能使用CPU的电脑上运行。

A lot of times, you know, if you are trying to run an agent to do coding or to build software, you are running that, on a cloud in the cloud, using CPUs or on a on a computer, that might use CPUs.

所以，我认为这就是我们看到需求增加得那么多的一部分原因。

So, I think that is part of the reason, that, you know, we're seeing a lot more demand.

而且，实际上，像英特尔和其他大型芯片公司的高管都说过，AI中GPU和CPU的比例正在下降。

And, actually, executives at, like, Intel and, you know, some of the other big chip companies have said, you know, the ratio, between GPUs and CPUs and AI is sort of coming down.

你知道，我们看到它在不同的工作负载中接近持平。

You know, we're seeing it be closer to parity, in in in different workloads.

嗯。

Mhmm.

抱歉。

And sorry.

为什么CPU适合运行代理，还是说这更多是成本的问题？

Is there a reason why CPUs are good for running agents, or is it more of like a cost thing?

对。

Yeah.

我觉得这是个好问题。

I mean, that's a good question.

我想，其实你并不真的需要GPU来运行代理。

I think that, I mean, yeah, I I guess you don't really I my interest like, you know, people are talking to say, you don't really need the GPUs to often run the agent.

所以我猜这更多是成本的问题。

So I imagine it is something of a cost.

是的，GPU很贵，但可能也不是必需的。

I mean, yeah, GPUs are expensive, but it's also maybe just like not necessary.

也许有些工程师觉得用CPU更容易。

Think, I mean, maybe some some engineers find it easier.