Trump Admin Nears AI Rules Framework, Microsoft’s Mythos Alternative, Why Claude Code is King

欢迎大家收听Information的TITV节目。

Welcome everyone to the Information's TITV.

我是Akash Pasricha。

My name is Akash Pasricha.

今天是7月28日，星期二。

It is Tuesday, July 28.

在开始之前，我想先提几条新闻。

There are a couple of stories that I want to flag for you before we get started.

The Information独家报道，Kimi k3的开发商Moonshot AI正在研发下一代模型Kimi k4。

The information exclusively reported that Kimi k three maker Moonshot AI is now working on its next model, Kimi k four.

为了做到这一点，公司正想方设法获取更多英伟达Blackwell芯片，尽管美国对中国的出口限制依然存在。

And to do that, the company is looking to get its hands on more Nvidia Blackwell chips despite US export restrictions to China.

The Information还独家报道，中国一家国有背景的公司已经开始生产芯片制造设备，这些设备有望与芯片设备巨头ASML竞争。

The information also has exclusive reporting that a state backed company in China has started manufacturing chip making equipment that would be competitive to chip equipment giant ASML.

报道发布后，ASML的股价下跌了。

Shares of ASML fell after the report was published.

这两条报道都在我们的网站上。

Both of those stories are on our website.

今天节目里，我们会详细讲讲美国政府的人工智能框架即将完成的情况，以及里面具体包含哪些内容。

On today's show, we have the details on how the government's AI framework is nearly finished coming together in The US and what exactly is in it.

接着我们会讨论微软针对Mythos推出的新安全项目。

We'll then discuss Microsoft's answer to Mythos with a new security project.

关于Claude、Codex和Cursor之间的竞争，我们有一个让人意外的更新。

We have a surprising update for you on the Claude versus Codex versus Cursor rivalry.

最后，我们来看看英伟达在Vera Rubin上会不会遇到和Blackwell一样的发布难题。

And finally, will Nvidia have the same rollout challenges with Vera Rubin as it did with Blackwell?

我们负责英伟达的记者会带来一些很棒的报道，来回答这个问题。

We will have the answer to that from our Nvidia reporter who has some great reporting on that question.

今天的节目会很精彩，那我们就直接开始吧。

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

我的同事施瓦茨，他是我们负责AI和政策的记者，带来新报道说，特朗普政府即将敲定针对AI公司的自愿框架。

My colleague, Schwartz, our AI and policy reporter, has some new reporting that the Trump administration is close to finalizing its voluntary framework for AI companies.

我想请Leo来和我们分享他所知道的情况。

I want to bring on Leo to share with us what he knows.

Leo，欢迎回到节目。

Leo, welcome back to the show.

很高兴你能来。

It's great to have you here.

很高兴来到这里。

Good to be here.

好的。

Okay.

那么关于政府的自愿性AI框架，我们知道些什么？

So what do we know about the government's voluntary AI framework?

进展到哪一步了？

Where is it at?

对。

Yeah.

我很高兴能来这儿。

I'm excited to come here.

我觉得终于有好消息了。

I feel like finally with some good news.

通常都是没有明确的信息。

It's usually there's no clarity.

关于正在发生的事情有很多困惑。

There's a lot of confusion about what's happening.

对于那些试图了解政府如何对待前沿模型的AI公司来说，终于可能看到隧道尽头的光明了。

There could be finally a light at the end of the tunnel for AI companies who are trying to have some clarity around how the government is treating frontier models.

所以如果你还记得五月底六月初的时候，白宫发布了一项行政命令，基本上是说，我们要建立一个自愿框架，让公司在发布模型之前与政府分享。

So if you remember way back to late May, early June, there was this executive order out of the White House that basically said, we're going to create a voluntary framework for companies to share their models with the government before release.

这项行政命令还规定，政府有六十天的时间来制定这个框架。

The EO, the executive order also said the government had sixty days to come up with that framework.

不知不觉，那六十天已经过去了。

Somehow those sixty days have already passed.

我们快到8月1日的截止日期了，我们的报道显示政府即将发布这个框架。

We're nearing that deadline of August 1, and our reporting shows that the government is close to releasing it.

他们一直在和三大实验室传阅一份草案，我们可能终于能了解这个流程会是什么样子了。

They've been circulating a draft with the top three labs, and we might finally have an understanding of what this process will look like.

好的。

Okay.

那里面有什么内容？

And so, what's in it?

我们目前知道些什么？

What do we know about that?

这个问题还没有定论。

So, that remains an open question.

哎呀，Leo。

Oh, man, Leo.

你带来了好消息，但我们还是有更多疑问。

You come bearing good news and then we still have more questions.

但每个重大新闻故事往往都是这样，问题比答案多。

But that is the case with every great news story is that more questions than answers often.

所以我们现在知道的是，这个框架会带来两个重要内容。

So what we know is that there's two big things that will come out of this framework.

第一个是，什么是受监管的前沿模型？

The first is what is the definition of a covered frontier model?

基本上就是，一个模型要达到什么标准才能有资格参与这个项目？

So basically, what does a model have to reach in order for it to actually qualify for this program?

你可能觉得这很明显，就是看它的能力，看模型的大小。

You might think that's obvious, its capabilities, it's the size of the model.

这正是各家公司还在努力与白宫关键机构敲定的问题，其中包括国家网络总监办公室。

That's the question that the companies are still trying to hammer out with these key White House agencies, which include the Office of the National Cyber Director.

看起来他们正在围绕能力这个想法进行讨论。

It seems like they're circling around this idea of capabilities.

但正如我们在文章中报道的那样，小公司仍然担心自己会被排除在这个过程之外，因为这个过程可能只针对那三家大公司来制定。

But as we report in our article, there are still concerns from smaller companies that they might be frozen out of this process that will only be written with the top three companies in mind.

那另一个大问题是，一旦你成为受覆盖的前沿模型，这到底意味着什么？

Then the other big question is, okay, once you are a covered frontier model, what does it actually mean?

政府审查你的模型的过程是怎样的？

What is this process for having your model vetted by the government?

文章里有提到，这可能会结合国家安全局和凯西，也就是商务部下面那个负责模型审查和测试的小机构来共同完成。

We do have details in the article that it will likely be done with a combination of the National Security Agency and Casey, the small agency under commerce that does model review and testing.

但最终还是要等他们给出答案，决定哪些模型有资格参与这个项目，以及实际的审查流程是怎么运作的。

But still, what we're going to have to wait for is that final answer on how it will decide which models are eligible for the program and then how the actual review process works.

所以就是三大实验室。

So big three labs.

我们说的是OpenAI、Anthropic，还有谷歌。

So we're talking OpenAI, we're talking Anthropic, and then Google.

这三家公司正在和政府紧密合作。

These are the three companies that are working closely with the government on this.

他们对此大体上满意吗？

Are they largely happy about this?

他们有没有什么顾虑？

Do they have concerns?

他们的立场是什么？

Where do they stand?

嗯，OpenAI和Anthropic显然都对政府对待他们的方式有些不满。

Well, OpenAI and Anthropic obviously both have gripes with the government over how it's treated.

Anthropic新模型的推出，特别是Mythos和Fable，情况更严重一些。

The rollout of new models with Anthropic is a little more severe with Mythos and Fable.

当然，政府因为OpenAI明显存在的安全漏洞而实施了出口管制。

Of course, the government imposed export controls over apparent security vulnerabilities with OpenAI.

特朗普政府要求他们分批发布最新模型GPT-5.6，Sam Altman对此很不高兴，并且公开表达了不满。

The Trump administration ordered them to do a staggered rollout of their their latest model, g p t 5.6, which Sam Altman was not pleased about and said as much publicly.

所以我觉得，只要有任何实际的政策出台，他们就已经很高兴了。

So I think just having any sort of actual real policy out there, they're they're thrilled about.

同时，我们也知道，他们一直在推动关于什么构成前沿模型以及测试如何进行的各种考虑。

At the same time, we know that they've been pushing for different considerations around these questions about what will constitute a frontier model and how the testing will actually be done.

再说，如果你还记得很久以前，也就是两个月前的报道，当时有个担忧，就是公司向政府提交模型的时间会不会太长、太繁琐。

Again, if you remember back to reporting from way long ago, two months ago, there was this concern about whether the period of which companies would have to submit their models to the government would be too long and too onerous.

前AI和加密货币主管大卫·萨克斯，大家都知道，他在最后一刻成功介入，阻止了特朗普签署行政令的早期版本。

Former AI and crypto czar David Sachs famously was able to stage a last minute intervention to prevent Trump from from signing an earlier version of the of the EO.

所以他们想确保这个框架能落地，有具体的政策，但也不会给他们带来不必要的负担，因为他们正忙着发布新模型，还要在这场AI竞赛中赢过中国。

So they wanna make sure that this will be in place, that it will have concrete policy, but also that it won't create any unnecessary burdens for them as they're trying to release their new models and and win this AI race against China.

在自愿框架里，有没有关于开源的内容？

Is there anything about open source in the voluntary framework as we know it?

这又是一个我们等着看的问题。

Yet another question that we're waiting to see.

我们知道，而且我们之前也报道过，在这篇文章里也提到了，开源公司一直在推动在行政令或者框架里，能有潜在的豁免条款。

We know and we've reported previously and in this article that open source companies have been pushing for considerations in the executive order or in the framework, I should say, both that there could be potential exemptions.

基本上，他们想确保自己不必非得走这个流程，因为那可能又是一种监管负担，而他们正急着创新，想挑战目前领先的中国同行。

Basically, they wanna make sure they don't have to necessarily go through this process, which could represent another regulatory burden as they're trying to race ahead with innovation and be able to challenge their Chinese counterparts, which are further ahead right now.

但我也听到一些担忧，有些开源公司想确保，就像我之前说的，这个框架不能只针对GPT-5.6和Mythos来制定。

But I've also heard concerns that some open source companies wanna make sure, as I said before, that the framework isn't written only with GBT 5.6 and Mythos in mind.

换句话说，他们希望自己的模型能够真正通过这个审查流程，并获得政府的重要批准，这将为他们打开行政命令中规定的各种不同条款，比如基本上能够与关键基础设施合作，无论是银行还是国内不同的团体，以便进行网络安全、安全以及其他类型的项目审查。

In other words, they wanna make sure that their models can actually go through this vetting process and can be given this essential stamp of approval from the government, which will open it up to all sorts of different provisions laid out in the executive order, like basically being able to work with, critical infrastructure, whether that's banks or or different groups in the country to be able to vet for cybersecurity, safety, and and other and other types of programs.

所以，基本上，如果你确实通过了这个项目，可能会有一些随之而来的好处。

So, basically, if you do go through the program, there probably will be some advantages that go with it.

太好了。

Great.

好的，Leo，非常感谢你来做客。

Well, Leo, I want to thank you for coming on.

这位是Leo Schwartz，我们The Information的AI和政治记者。

That is Leo Schwartz, our AI and politics reporter here at The Information.

微软正在推出一个新的以网络安全为重点的项目，其模型旨在与Mythos竞争。

Microsoft is launching a new cybersecurity focused project with models that are meant to be competitive with Mythos.

我们实际上在几周前就报道过这件事会发生，现在确实发生了。

We actually reported this was going happen a couple of weeks ago, and it happened.

所以我想请我们的微软记者Aaron Holmes来为我们解读微软在这方面的策略。

So I want to bring on Aaron Holmes, our Microsoft reporter, to walk us through Microsoft strategy with this play.

Aaron，欢迎回到节目。

Aaron, welcome back to the show.

很高兴你能来。

It's great to have you here.

很高兴来到这里。

Happy to be here.

所以这是个好消息，对吧？

So it's great news, right?

我的意思是，你三周前来的时候，告诉我们“感知计划”，我记得是叫这个名字。

I mean, you come on three weeks ago, you tell us Project Perception, I think it's called.

它要实现了。

It's going to happen.

现在公告也出来了。

Here we have the announcement.

具体是什么？

What is it?

他们最终宣布了什么？

What do they end up announcing?

对。

Yeah.

基本上，这是微软推出的一款新产品，定位是Anthropic Mythos的更便宜替代品。这两款工具都是企业可以用AI模型来测试自己环境漏洞的，理论上AI能比人类网络安全测试员更快地发现漏洞。

So basically, this is a new product that Microsoft is pitching as a cheaper alternative to Anthropic Mythos, and that's basically you know, both of them are tools that companies can use powered by AI models to, you know, test their own environments for vulnerabilities, and the AI can theoretically find them a lot faster than human cybersecurity testers previously could.

具体来说，微软表示这款产品由自家的网络安全模型以及Anthropic和OpenAI的其他模型共同驱动。

And specifically, you know, Microsoft is saying this is powered by its own cybersecurity models as well as a mix of other models from Anthropic and OpenAI.

因此，它比Mythos更便宜，也更容易让企业使用。

And as a result, it's, cheaper to use and easier for companies to get access to than to Mythos.

至于实际效果是否如此，我们还得拭目以待，但这是他们的宣传卖点。

Whether that ends up being true in practice, we still have to wait and see, but that's the pitch that they're making.

所以它应该和Mythos竞争。

So it's supposed to be competitive with Mythos.

那它需要经过同样的政府审查流程吗？

Does that mean that it had to go through any kind of the same government review process at all?

据我所知，政府没有对微软施加像对Mythos和Fable那样的出口管制，部分原因可能是因为在特朗普政府的眼中，微软不像Anthropic那样有争议。

As far as I know, the government hasn't put any of the same export controls that they originally put on Mythos and Fable, you know, partly just because I don't think Microsoft has as much of a controversial place in the in the Trump administration's outlook right now as Anthropic does.

不过，他们发布这个产品的范围比Anthropic更广，或者说更广泛。

But they, you know, they they have released this broadly or or more broadly than Anthropic has.

也就是说，理论上，这个产品会对所有微软客户开放，而不是像Anthropic推出Mythos和OpenAI推出DeepRake那样，只对少数几家公司开放。

Meaning, you know, in theory, this is something that will be available to all of Microsoft's customers rather than just a select few companies, which is the way that Anthropic has been rolling out its Mythos and OpenAI has been rolling out its DeepRake model so far.

但你今天文章里提到的一个有趣的点是，微软实际上仍然依赖OpenAI和Anthropic的模型来提供部分功能，对吧？

But one of the interesting parts of what you talked about in your article today was that they actually are Microsoft is still relying on OpenAI and Anthropic models for part of this offering, right?

对。

Yeah.

他们在这个产品以及Copilot和其他AI产品上的做法是，基本上会选择最适合不同任务的模型。

Their approach with this and also with Copilot and their other AI products is that they will essentially choose which model is best suited to to different tasks.

所以，在某些情况下，Anthropic的一个非常强大的模型可能会处理测试中编写代码的部分，但之后会把任务交给微软的模型或一个较小的模型来处理其他部分。

So, you know, in some cases, a very capable model from Anthropic might handle writing some code for part of a test, but then it'll hand things off to one of the Microsoft models or a smaller model for other parts.

因此，他们说这样可以降低成本，让客户以更低的价格获得基本相同的结果。

As a result, they say that they can drive down the cost and get customers essentially the same results for a lower dollar amount.

总的来说，微软的网络安全业务做得怎么样？

How is Microsoft cybersecurity business doing broadly speaking?

你在这方面写过很多文章。

You have written extensively on this.

他们现在也有了一位新的负责人。

They have a new leader at the top as well.

这个业务有多大？

How big is the business?

增长有多快？

How fast is this growing?

这能反映出他们整体的战略是什么？

What does this tell us about the strategy overall?

微软的网络安全业务规模超过200亿美元。

So Microsoft cybersecurity business is over $20,000,000,000.

这个数字其实是2023年的，那是微软最后一次公布其业务规模。

That number is actually from 2023, which is the last time that Microsoft told us how big its business is.

我知道从那以后它又增长了，但我没有确切的数字。

I know that it's grown since then, but I don't have the exact number.

但我听说的是，微软基本上想再次加速其网络安全收入的增长。

But what I have heard is that Microsoft basically wants to accelerate its cybersecurity revenue growth again.

具体来说，今年早些时候成为网络安全部门新负责人的Hyatt Gallet，基本上在彻底改革他们的战略，试图更专注于这些AI驱动的工具，以及能帮助公司抵御那些同样被AI加速的新威胁的软件。

And specifically Hyatt Gallet, who's the the new leader of its cybersecurity unit as of earlier this year, has been basically overhauling their strategy and trying to focus more on these AI powered tools and on software that can help companies protect against, you know, newer threats that are also being supercharged by AI.

所以我们看到她基本上在微软内部进行了大清洗，彻底改革了整个安全业务。

And so we've seen her essentially clean house and kind of overhaul the entire security business at Microsoft.

我认为这次产品发布是她所指引方向的首批重大信号之一。

And I think that this product release is one of the first big signs of the direction that she's taking.

所以，艾伦，很快说一下，我知道你得走了，但本周我们有财报。

So very quickly, Aaron, I know you gotta go, but we have earnings this week.

关于微软的财报，我们应该关注什么？

What should we be watching for with respect to Microsoft earnings?

大问题是，微软能否从它在AI上的空前投入中获得可观的回报。

The big question is just whether Microsoft is going to be able to get a meaningful return from the unprecedented amount that it's spending on AI.

你知道，这家公司已经一次又一次地上调了今年的资本支出指引，而且我们可能会看到它甚至超过他们为截至六月的财年预测的1900亿美元。

You know, the company has already been revising its CapEx guidance for the year up and up and up, and we'll we might see it, you know, go even higher than the 190,000,000,000 that they forecast for the year ending in June.

同时，投资者和股东真的很想看到Azure收入增长的显著加速，尤其是因为目前谷歌云的增速远快于微软的Azure。

At the same time, you know, investors and shareholders really wanna see meaningful acceleration in its Azure revenue growth, especially because Google Cloud is growing at a much faster rate than Microsoft's Azure at this point.

而且，大家还想看到Copilot的更多销售。

And, you know, people wanna see more sales of Copilot.

我们知道有2000万付费Copilot席位，这只是微软4.5亿总付费Office用户中的一小部分。

We know that there's 20,000,000 paid Copilot seats, which is a really small fraction of the 450,000,000 total paying office users that Microsoft has.

所以，如果他们这个季度不能展示出上述所有这些方面都在获得进展，并且AI支出有回报，那可能会给股票带来更大压力，而股票今年已经大幅下跌了。

So if if they can't show, you know, kind of all of the above that they're getting traction and ROI from their AI spending this quarter, it's probably gonna, you know, put even more pressure on the stock, which has already fallen significantly this year.

谢谢。

Grateful.

Aaron，感谢你抽出时间接受我们的采访。

Aaron, I want to thank you for making time for us.

那是Aaron Holmes，我们The Information的微软记者。

That is Aaron Holmes, our Microsoft reporter, here at The Information.

Claude Code、Codex 和 Cursor 之间的竞争非常激烈。

The rivalry between Claude Code, Codex, and Cursor is fierce as can be.

为了了解更多谁在胜出的情况，我想请来劳拉·布拉顿，她是我们的《应用AI》通讯的作者，这周她发现了一些关于这个话题的有趣新数据。

For more on who is winning, I want to bring on Laura Bratton, author of our Applied AI newsletter, who uncovered some interesting new data on that story this week.

劳拉，欢迎回到节目。

Laura, welcome back to the show.

很高兴你能来。

It's great to have you here.

很高兴来到这里。

Good to be here.

周二快乐。

Happy Tuesday.

周二快乐。

Happy Tuesday.

那我们聊聊你收集的这些数据吧。

So let's talk about the data here that you collected.

你这里有一些有趣的数据。

You had some interesting numbers here.

Claude Code，它获得的关注度，我们之前聊过Codex以及它对这个故事构成的威胁。

Claude Code, the traction that it's getting, we kind of have talked about Codex and the threat that it offers to that story.

实际情况到底怎么样？

What is the reality on the ground?

我看到的实际情况是，很多公司仍然坚持使用Claude Code，因为它是他们企业里第一个使用的工具。

The reality on the ground that I've seen is companies are still sticking with Claude Code in many cases because it's the first tool that was in their enterprise.

我发现的一些数据显示，在Anthropic的约200个客户中，他们都在使用Claude Code。

And some of the data that I found was that among around 200 customers of Anthropics, they use Claude Code.

即使因为Anthropic转向按使用量定价导致每个用户的成本上升，给公司的账单带来了压力，他们组织内部的使用量仍在持续增长。

Even as cost per user rose because of Anthropic shift to usage based pricing, hitting companies' bills, they were still continuing to see usage within their organizations grow.

那这是为什么呢？

And why is that?

我的意思是，我们可以稍后讨论具体数字，但关键问题是，如果成本在上升，他们为什么不直接改用Codex或Cursor呢？

I mean, we can talk about the magnitude in a second, but the headline, why is if costs are going up, why wouldn't they just go to Codex or to Cursor?

对。

Yeah.

我觉得事情没那么简单。

I think it's just not that simple.

公司既想要Anthropic的专有模型，也想要OpenAI的前沿工具。

Companies want access to proprietary models and frontier tools from both Anthropic and OpenAI.

如果他们用这些模型，他们就想看看它们能做什么。

If they're using those models, they want to see what they can do.

很多情况下，公司只是先采用了Claude Code。

In many cases, companies just adopted Claude Code first.

他们还在让工程师使用这些工具的早期阶段。

They're still in the early days of getting engineers to use these tools.

工程师们更喜欢Claude Code。

Engineers have a preference for Claude Code.

即使有些人可能开始用Codex，或者尝试开源工具，但Claude Code在用户数量上仍然占主导地位，这些用户目前主要用Claude Code做编码工作。

Even you know, some may be beginning to adopt Codex, they might be experimenting with open source tools, Claude Code is still dominant in terms of the number of users who are primarily doing their coding work there today.

你收集到的那些数据，能跟我们讲讲大概的规模吗？成本涨了多少，但还有多少人继续用Claude Code？

And the data that you collected and got your hands on, I mean, just walk us through some of the magnitude here of how much costs did increase and yet how many more people were still using Claude Code?

因为我觉得这些数据本身就挺能说明问题的。

Because I think the data itself was pretty telling.

对。

Yeah.

数据来自一家叫Weave的AI初创公司，他们提供模型路由服务。

So the data was from an AI startup, Weave, which provides model routing services.

他们显示，尽管他们的客户中有些是大企业，但一半是初创公司，还有大约15家是中型公司。

And they showed that, you know, even though among their customers, which in some cases were large enterprises, half were about startups, and then about 15 were medium sized companies.

数据显示，即使Claude Code的成本大幅上升，使用量也在增加。

And it just showed that, you know, even as Claude Code costs rose substantially, you know, usage also increased.

特别是对于Cursor，数据显示，即使每个用户的平均成本上升，使用量实际上下降了17%。

And then for Cursor in particular, it showed that even as the median cost per user climbed, you know, usage actually declined 17%.

所以，我觉得这显示了像Cursor和Claude Code这样的工具在定价权上的差异。

So it just I think it shows the difference in the pricing power between a tool like Cursor and Claude Code.

好的。

Okay.

那我们接着聊聊其他几个名字以C开头的编程巨头吧。

So let's go to the other coding giants that also have decided to start their names with C.

现在有Claude、Codex和Cursor之间的竞争。

We've got the Claude versus Codex versus Cursor rivalry.

根据你找到的数据，Codex和Cursor现在的情况怎么样？

What's going on with Codex and Cursor according to the data that you found?

在我和企业交流的过程中，有些企业说Codex的使用量增长得比Claude Code还快。

So in my conversations with enterprises, some said that usage of Codex is actually growing faster than Claude Code.

特别是T-Mobile和数据库公司Neo4j，他们说工程师们对Codex很感兴趣。

In particular, T Mobile and a database provider, Neo4j, said that, you know, their engineers are getting excited about Codex.

有些情况下，Neo4j告诉我他们觉得从Codex那里得到的回报更好，我觉得这挺有意思的。

In some cases, you know, Neo4j told me that they feel they're getting a better return from Codex, which I thought was interesting.

这些都是基于主观感受的。

This was all vibes based.

他们并没有给我提供具体的数据来支持这个说法。

They didn't necessarily have data they provided to me to back this up.

但我觉得这确实是个有意思的点，我们听到一些工程团队在这么说。

But I think it is an interesting point that we're seeing from some engineering teams.

所以Codex增长很快。

So Codex is growing fast.

而且，虽然我听到很多公司说他们从Cursor转到了Claude Code，或者之前已经转过，但我发现Cursor其实还是很有市场的。

And then even though I'm hearing a lot companies say that they're switching from Cursor to Claude Code or they previously switched from Cursor to Claude Code, one of my sources, I found out that Cursor is still seeing a lot of traction.

今年到目前为止，他们的企业收入增长了大约七倍，而且他们自有的编程模型的采用率也在上升。

Year to date, their revenue from enterprises grew about sevenfold and adoption of their proprietary coding models is growing.

所以，他们的Composer和Grok模型，采用率或者使用量今年到现在增长了11倍。

So, you know, their composer and grok models adoption or usage of those models grew 11 fold year to date.

那这两家公司，他们用的是按使用量计费的模式吗？

So it's Are those two companies are they using usage based pricing?

他们的价格在上涨吗？

Are their prices going up?

他们现在的价格是不是基本保持不变？

Are they staying relatively stagnant right now?

Codex 的采用还很新，而且是按使用量计费的，但我觉得这种按使用量计费的变化，对那些客户的冲击，跟那些先采用了一个工具、然后看到计费模式改变的客户，是不一样的。

So Codex adoption is so new and it is usage based, but I don't think a shift I don't think a shift toward usage based pricing is hitting those customers in the same way it's hitting customers that adopted one tool and then saw the pricing model change.

Cursor 在去年年中把计费模式改成了按使用量计费，但这只是刚刚开始影响到我采访的一些客户，因为他们的合同今年才到期续约。

Cursor changed its pricing model to usage based in the middle of last year, but that's only just beginning to hit some of the customers I talked to's contracts as they're up for renewal this year.

对。

Right.

你知道，我现在在想，Cursor 不可避免的，看起来它将成为 SpaceX AI 生态系统的一部分。

You know, I'm sort of thinking now that Cursor inevitably will it's looking like it's going to be part of the SpaceX AI ecosystem.

不是，我是说，我不知道交易是否正式完成，但我们基本知道 Cursor 现在是 SpaceX AI 家族的一员了。

Not, mean, don't know the the details on whether or the deal has officially closed or not, but we pretty much know that Cursor is now part of the SpaceX AI family.

我在想，Cursor 在这里会采取什么策略，因为你可以想象一个世界，埃隆说，嘿，让 Cursor 成为亏损产品。

I sort of wonder what the strategy will be here with Cursor because you could imagine a world where Elon says, Hey, let's make Cursor the loss leader.

让我们快速抢占市场份额。

Let's get market share quickly.

反正我们也在烧钱。

We're burning cash anyway.

所以这真的无所谓。

It doesn't really matter.

我们这几年本来就不会盈利。

We're not going to turn a profit for years on this.

我也不太确定。

I I don't know.

看看这个策略怎么发展，可能会很有意思。

It could be interesting to see how that strategy plays out.

对。

Yeah.

我觉得这绝对会发生。

I think that's absolutely what's going to happen.

我的预测是，通过SpaceX的收购，Cursor能获得更多GPU资源，他们自己的编程模型会变得越来越好。

We're going to see my prediction would be that we're going to see Cursor's own coding models get better and better as they have greater access to GPUs through the SpaceX acquisition.

我觉得这会大大提升他们的地位。

I think it's going you know, improve their position a lot.

而且那些作曲模型只会越来越好。

And the composer models are just gonna get better.

我觉得很多公司都特别希望在这个领域看到另一个主导者，除了Anthropic和OpenAI，还有那些开源玩家之外。

I think that companies are really eager to see another dominant player in this space other than Anthropic and OpenAI in addition to the open source players.

所以我觉得Cursor是有空间的。

So I think that there's space for Cursor.

我觉得公司不会想通过Cursor来用Anthropic或OpenAI的模型。

I don't think that companies are going to want to use Anthropic or OpenAI's models through Cursor.

而这正是我们看到很多Cursor客户流失到Anthropic和OpenAI的原因。

And that's where I think we've seen a lot of the leakage from Cursor customers to Anthropic and OpenAI.

就是那些公司发现，如果Cursor要涨价，那还不如直接用Claude Code或者Codex。

It's companies that are finding that, you know, if Cursor is going to raise prices, we might as well just go to the Claude Code or Codex.

用那种为Anthropic和OpenAI模型优化的工具，也就是Claude Code或Codex，而不是通过Cursor。

Use, you know, a harness, AKA Claude Code or Codex that are optimized for Anthropic and OpenAI's models rather than going through Cursor.

对。

Right.

而且我觉得，你知道，我刚才描述的那个策略，真的得靠Grok和那个模型变得先进很多，而且Cursor至少得有自己的某种优势，但根据我们的报道，Grok现在也遇到了一些阻力。

And I guess, you know, the strategy that I just sort of described, I guess that would really have to rely on Grok and that model becoming a lot more advanced, and Cursor having at least some kind of its own advantage there, which as we know from our reporting, Grok itself is facing a little bit of headwinds right now.

是的。

Yeah.

想回到Both

Want to go back Both

那两个模型，我觉得，肯定会对Cursor的策略非常重要。

of those models, I think, will be really important to Cursor's strategy for sure.

对。

Right.

对。

Right.

所以，你提到了你和T-Mobile高管的对话，这挺有意思的，因为你在文章里指出，T-Mobile允许员工使用所有三种编码代理。

So, you mentioned the conversation you had with a T Mobile executive, and this was kind of interesting because you pointed out in your article that T Mobile is a company that they are allowing their employees to use all three coding agents.

我读到那篇文章时就在想，这种情况普遍吗，还是只有T-Mobile这样？

I read that thinking, is that common or is this unique to T Mobile?

对大多数企业来说，实际情况是怎样的？

What's the reality for most of these businesses?

是的，这绝对是我从几乎所有谈过的CIO和CTO那里听到的普遍说法。

Yeah, it's absolutely a commonplace narrative that I'm hearing from nearly every CIO and CTO that I've talked to.

公司希望让工程师们能使用所有前沿工具，如果可能的话，也包括开源工具。

Companies want to give their engineers access to all the frontier tools to open source tools if they can.

即使不是每个工具都开放，他们至少也会提供不止一个，因为他们真的很想让工程师们来推动采用。

Even in cases where they're not giving access to every tool, they're at least giving more than one because they really wanna see their engineers drive adoption.

而且我觉得他们希望如果某个AI实验室涨价、成本变动，或者政府政策变化影响公司获取这些模型的方式，他们能轻松地在不同供应商之间切换。

And I think they wanna be able to replace one tool with another if one of the AI labs jacks up prices, shifts costs, if government policies change and that affects how companies are able to access these models, they wanna be able to switch between providers easily.

好的。

Okay.

这种情况有时确实会发生。

And this is what happens sometimes.

你有个问题，就把那篇文章给忘了。

You have a question, you forget the essay.

那开源这部分在这一切中又扮演什么角色呢？

Where does the open source equation fit into all this?

对。

Yeah.

我注意到，就在几周前，甚至几个月前，我聊的时候听到的是，公司不想用开源模型来写代码。当时有个假设，就是如果你用一个小一点、便宜一点的模型，或者开源模型，结果肯定不如用最新的Anthropic模型那么好。

One thing I noticed is that in my conversations a few even just a few weeks ago, if not, you know, a couple months ago, what I heard was that companies didn't want to use open source models for coding There was this assumption that if you use a smaller, cheaper model or an open source model, you're not going to get as great of a result as if you use the latest and greatest Anthropic model.

但现在情况真的变了，我们看到很多公司开始采用像Kilo Code、OpenCode、PyCline、Hermes这些开源工具，还有一堆其他服务，基本上能让你在写代码时轻松切换专有模型和开源模型。

But that has really changed as we've seen companies begin to adopt open source harnesses like Kilo Code, OpenCode, PyCline, Hermes, another a number of services that basically allow you to switch easily between a proprietary model and an open source model for coding work.

而且越来越明显的是，写代码不一定非得用最新最强的模型。

And what's becoming clearer is that you don't necessarily need the latest and greatest model to do coding work.

其实更重要的是有一个工具或者软件层，能让模型像编程代理一样工作，让这个工具根据情况选择最适合写代码的模型。

It's actually more important to have a harness or a software layer that lets a model act as a coding agent to have that harness basically leverage whatever model is best for coding work.

就像我们看到公司用模型路由器来处理其他AI任务一样。

Same way we've seen companies use model routers for other kinds of AI work.

对。

Right.

所以我想问的是，回到编码代理模型的多样性，比如Claude Code、Codex和Cursor，你觉得这最终会是一个赢家通吃的局面，还是说最终企业和客户会从可能演变成价格战的竞争中受益？

And so what I was going to ask you was then going back to the diversity of the models of the coding agents here with Claude Code, Codex, and Cursor, do you see this as a winner takes all story in the end, or is this a story where ultimately it's the enterprises and the customers that will benefit from what inevitably could shake out to be a pricing war in the end?

我觉得我们会越来越多地看到专有模型提供商，比如Anthropic和OpenAI，与用于编码工作的开源模型之间的紧张关系。

I think more and more, we're going to see a tension between the proprietary model providers in Anthropic and OpenAI and open source models that are used for coding works.

所以我们会看到Anthropic和OpenAI之间的竞争，以及像OpenCode、KiloCode、Klein这样的开源工具。

So we'll see a sort of rivalry between Anthropic OpenAI and then these open source harnesses such as OpenCode, KiloCode, Klein.

我们会发现，对于编码工作来说，工具本身会比模型本身重要得多。

And we're going to see that the harness becomes a lot more important than the actual model for coding work.

工具指的是模型周围的软件，基本上就是应用程序，对。

The harness meaning the, again, the software around the the model itself, basically the application that Yeah.

人们每天都在用的。

People are using every

你可以把工具想象成手臂和腿，它们使用模型这个大脑来完成编码工作。

So think of the harness as kind of the arms and legs that use the model, which is the brain to do coding work.

所以在某些情况下，你不需要大脑百分之百运转就能完成编码工作。

So in some cases, you don't need your brain to be working at 100% capacity to do coding work.

你只需要你的手臂和手，往正确的方向伸，然后拿起正确的东西，就这样。

You just need your arms to, you know, and your hands to reach in the right directions and pick the right thing with, you know, whatever.

对。

Yeah.

不过，虽然现在有这些悄悄话式的操作，我的意思是，我们现在连手臂和腿都不需要了。

Although although with all this whispering that's going on, I mean, now we don't even need arms and legs.

你只需要说出你需要什么就行。

You just you just dictate what you need.

而且，你知道，他反正……我需要一个那种麦克风。

And, you know, he's anyway, I need one of those microphones.

对。

Yeah.

我们会想出一个更好的比喻。

We're going to come up with a better analogy.

我们得想个比喻。

We're going to come up with an analogy.

不不。

No, no.

我觉得，我一直搞不明白的就是，harness和应用本身到底有什么区别。

I think look, it is certainly the thing that I can never figure out is what the difference is between a harness and then the application itself.

我猜harness更像是应用和模型之间的中间层基础设施。

And I guess harness is more of like the middle layer infrastructure between the application and model.

对吗？

Is that right?

对。

Yes.

我也是这么理解的。

That is my understanding.

是的。

Yeah.

好的。

Okay.

劳拉，非常感谢你来做客，帮我们理清这个问题。

Well, Laura, I want to thank you for coming on and helping us make sense of it.

这位是劳拉·布拉顿，她是《The Information》的Applied AI通讯作者。

That is Laura Bratton, author of our Applied AI newsletter here at The Information.

英伟达的客户开始陆续拿到公司最新的Vera Rubin机架了。

Nvidia customers are starting to get their hands on the company's latest Vera Rubin racks.

我们的英伟达记者Phoebe Liu有一些关于客户对这些机架安装难度初步评价的新报道。

Our Nvidia reporter, Phoebe Liu, has some new reporting on the initial reviews those customers have on how difficult they are to install.

我想请她来聊聊她的发现。

I want to bring her on to talk about what she found.

Phoebe，欢迎回到节目。

Phoebe, welcome back to the show.

很高兴你能来。

It's great to have you here.

很高兴来到这里。

Great to be here.

好的。

Okay.

那么，英伟达正在向客户推出其Vera Rubin机架。

So Nvidia is rolling out its Vera Rubin Racks to customers.

这次推出的初步反馈怎么样？

What are the initial reviews of that rollout?

进展如何？

How's it going?

是的。

Yeah.

这里最大的问题是，现在说还为时过早，而且我接触过的大多数客户基本上只有测试机架。

So the big, big asterisk here is that it's still very early to tell, and most of the customers that I've spoken to only have basically test racks.

他们只是拿几台来玩玩，确保能顺利上线运行，然后再接收更大的订单，比如成百上千台机架。

So a handful that they're kind of playing with to make sure they can get them online and running smoothly before getting larger shipments of, I don't know, hundreds and thousands of racks.

但目前来看，我觉得大家都持谨慎乐观的态度，至少几乎所有人都是这样。

But right now, I think everyone is cautiously optimistic or at least almost everyone.

他们让测试机架上线比上次布莱克韦尔机架时容易多了。我同事之前报道过，那次因为和上一代差别太大，过程非常非常艰难。

Getting the test racks online has been easier for them than at least last time around with the Blackwell racks, which my colleagues reported on was basically very, very excruciatingly difficult because of how different it was from the previous generation.

那具体有哪些问题呢？

And what exactly were were some of the issues?

给我们回忆一下。

Remind us.

上次发生这种情况时，出现了哪些问题？

Last time this happened, what were the problems that came up?

对。

Yeah.

这里我要借用我优秀同事的报道，因为那是在我加入The Information之前的事。

So cribbing on my excellent colleagues reporting here since this was before I joined the information.

几个最大的问题。

A few of the biggest issues.

首先，Blackwell机架和之前的Hopper芯片一个很大的不同是，它是个完整的机架级解决方案，他们喜欢这么叫，把72个GPU和36个CPU连在一起，再加上存储、网络等等。

One, a big difference between the Blackwell racks and the previous Hopper chips was that it was a full kind of rack scale solution, they like to call it, that connected 72 GPUs and 36 CPUs together plus storage and networking and all of that.

你想想看，要把这么多芯片连起来，得用多少线缆。

And that's just imagine the cabling that it would take to connect that many chips together.

对。

Yeah.

而且还要让它们之间通信，速度要达到毫秒级。

And and make it talk to each other in, like, speeds at the order of, like, milliseconds.

这挺难的。

That's pretty hard.

我觉得这是个大问题，比如一根线缆坏了，怎么排查问题？

And I think that was one of the big things where, like, if one cable failed, it was like, how do we figure out, like, what's going on?

是不是整个机架都瘫痪了？

Did it take down the entire rack?

影响的不只是一个芯片。

Like, it's impacting more than just one chip.

那是挺难搞的一件事。

That was one of the difficult things.

这次我采访了一个人，他实际上在维护一些大型集群，他说排查网络故障就像是在玩黑魔法一样。

Someone I talked to this time around who actually works on some of these large clusters described finding where networking failures occurred like black magic.

我无法想象那有多难，但那是

I can't imagine how hard that was, but that

就是说，你有一个巨大的数据中心，或者可能是一整排巨大的机架。

was what Meaning going you have this giant data center or maybe this giant sort of row of racks.

而且这些机架基本上就是芯片叠着芯片。

And again, the racks are literally chips on top of each other.

就像整个服务器挂了，或者数据中心里这一整排都挂了。

It's like this whole server is down, or this whole row in this data center is down.

基本上就是找出故障点在哪里。

Where is the fault line, basically?

有点像大海捞针，

Kind of like finding a needle goes in through,

对，如果我没记错的话，每个机架都有好几公里的线缆。

yeah, like multiple kilometers of cabling per rack, if I remember correctly.

所以，确实，那真的很难。

So, yeah, that that's pretty hard.

而且我的同事们还提到，早期有个设计缺陷导致了一些延误，还有散热的问题，因为机架功率越大，发热也越厉害，这是功率的必然结果。

And then also my colleagues reported that there was, like, an early design flaw that caused delays and then also issues with kind of heat because as racks get more powerful, they also get hot because that's how power works.

是的。

Yeah.

所以相关的工程难度也相当大。

So the engineering involved in that was also pretty difficult.

那么，解决方案是什么呢？我是说，薇拉·鲁宾在设计上有什么本质上的不同吗？

So is the solution to that I mean, is that is there inherently something different about the way that they have architected Vera Rubin?

他们有没有让客户更容易安装这些东西？

Have they made it easier for customers to install these things?

还是说，只是客户在数据中心安装这些巨型机架时，条件更好了？

Or is it just that the customers are sort of better positioned to install these giant racks in in data centers?

是的。

Yes.

有一点

It's a little bit

两者都有。

of both.

我采访过的几位云服务高管基本上都表示，他们认为这是第三代产品，因为这是英伟达发布的第三款整机柜产品。

A couple of the cloud executives I spoke to for this basically said they see this as a third generation product because it's the third kind of full rack product that Nvidia has released.

第一款是初代Blackwell，第二款是Blackwell的GB300系列。

The first being the first Blackwells and then the second Blackwells, the g b three hundreds.

现在有了Vera Rubin，也有人叫它VR200。

And now there's Vera Rubin or v r 200, some people are calling them.

因为这是这种配置的第三次迭代，72个GPU和36个CPU，所以他们觉得这次会稍微容易一些，毕竟之前已经做过类似的东西了。

So because it's kind of the third iteration of this configuration with 72 GPUs and 36 CPUs, they're thinking that this is a little bit easier because they've done something relatively similar before.

另外，几周前我在圣克拉拉和英伟达的一些高管聊过，他们说Vera Rubin的计算托盘——就是那个上面有四个GPU的托盘——自动化程度很高，而且比Blackwell机柜里的线缆少得多。

And then the other thing is that when I was talking to some Nvidia executives down in Santa Clara a couple weeks ago, I wanna say, they were saying that because so much of the Vera Rubin compute tray, so that's one tray with four GPUs on it, is so automated and has fewer cables than in the Blackwell rack.

第一次就能把装配做对的难度就低多了。

It's easier to get that assembly right on the first try.

我记得有位高管说，第一次就能成功运行的概率大概是95%，而最初的Blackwell机架只有大约20%。

I think one of the executives said it was, like, a 95% chance that it would work on the first try versus about 20 for the initial Blackwell trace.

哇。

Wow.

对。

Yeah.

我不太清楚这个数据在多大规模上经过实际测试，但这是他引用的数字。

I don't know how much that's really been tested at scale, but that was the statistic that he cited.

是的。

Yeah.

我是说，我们之前聊过英伟达在推出Blackwell芯片时遇到的挑战，那是什么时候来着？

I mean, when we talked about these challenges that Nvidia had rolling out the Blackwell chips couple what is it?

差不多一年前了吧，我想。

Nearly a year ago now, I think.

你知道，我们和编辑部以及来过的嘉宾讨论过的一件事是，这往往是客户第一次进行这种规模的数据中心建设，还要安装芯片之类的。

You know, one of the things that we did talk about with our newsroom and also with guests that came on is that this is oftentimes the first time that customers have embarked on a data center build out of this scale and installing chips and stuff like that.

所以我觉得这挺有意思的。

And so I do think that it's kind of interesting.

我想客户们自己可能也很高兴。

I think the customers themselves are probably also very happy.

第二次创业总是更容易一些。

It's easier to found a company the second time around.

第二次建数据中心也更容易。

It's easier to build a data center the second time around.

所以英伟达算是开了条路，顺便说一句，AMD也发布了它的Helios RAC系统，也会部署到数据中心里。

So Nvidia was sort of paving the path here, which by the way, we should say, I mean, AMD also has its Helios RAC system that it's announced and it will put in data centers.

所以就像总得有人当第一个——我是家里三个孩子中的老大，总得有人给大家开路。

So it's sort of like somebody had to be I'm the first of three children, and somebody has to blaze the trail for everyone else.

AMD的RAC系统很有意思。

The AMD RAC is very interesting.