Mythos-class Model Claude Fable 5 Early Reviews, How Nasdaq Landed SpaceX's Mega IPO

欢迎各位收看《The Information》的TITV节目。

Welcome everyone to the Information's TITV.

我是阿卡什·帕什里卡。

My name is Akash Pasricha.

今天是6月10日，星期三。

It is Wednesday, June 10.

在正式开始节目之前，今天太平洋时间中午12点，东部时间下午3点，我会和我们新闻团队里负责报道SpaceX的记者们坐下来聊聊。

Before we get to the show, today at 12PM Pacific, 3PM Eastern, I'm going to be sitting down with the reporters in our newsroom who have been leading the information SpaceX coverage.

他们分别是西奥·韦特、弗莉塔·鲍威尔和科里·温伯格。这场活动和网络研讨会仅限订阅用户参加。

Theo Wait, Vlita Powell, and Cory Weinberg, the event and the webinar is exclusive to subscribers.

现在就可以去theinformation.com/events报名预约。

You can RSVP right now at theinformation.com/events.

期待在活动现场见到你们。

Look forward to seeing you there.

今天节目中要聊的是，Anthropic推出了新一代的Mythos系列模型，叫Claude Fable。

Today on the show, Anthropic has a new version of a Mythos class model called Claude Fable.

我们会详细拆解这个模型跟最初发布用于测试的版本有哪些不同。

We'll break down how this model is different from the one initially released for testing.

我们还独家报道了OpenAI正在洽谈租赁一座位于俄亥俄州、容量达10吉瓦的数据中心，这笔交易可能还包括来自英伟达的财务支持。

We also have exclusive reporting on how OpenAI is in talks to lease a 10 gigawatt data center in Ohio that is part of a deal that could include financial backing from Nvidia.

接下来我们会深入看看纳斯达克是如何促成SpaceX这场超级IPO的，最后我们会跟Nebius聊聊他们如何看待英伟达Vera Rubin平台的推出。

We'll then take a closer look at how the Nasdaq landed SpaceX's mega IPO, and we'll close out the show with a conversation with Nebius on how it is looking ahead to Nvidia's Vera Rubin rollout.

这期节目非常精彩，咱们这就直接进入正题。

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

Anthropic几周前发布Mythos的消息，引发了大家对危险AI模型可能带来的威胁的诸多担忧。

Anthropic's Mythos announcement a few weeks ago sparked a number concerns about how dangerous AI models could inevitably become.

昨天，该公司宣布向公众发布其关联模型Claude Fable 5，并增加了新的安全措施。

Yesterday, the company announced the release to the public of an affiliated model, Claude Fable 5, with new safeguards.

《信息》杂志的AI记者Stephanie Palazzolo一直密切关注Mythos的整个发布过程，我想请她来跟我们分享一下目前的情况。

Information's AI reporter Stephanie Palazzolo has been following the entire Mythos rollout, and I want to bring her on to share with us what the situation is on the ground.

Stephanie，欢迎回到节目。

Stephanie, welcome back to the show.

很高兴你来参加节目。

Great to have you here.

谢谢。

Thanks.

能来我也很高兴。

Great to be here.

好的。

Okay.

之前大家都在说 Mythos 多么危险。

So there was all this hoopla about how dangerous mythos is.

他们说，我们只给少数人开放。

They said, We're only releasing it to a couple people.

接着又说，哦，Project Glasswing，再加一百五十个人。

Then they said, Ah, Project Glasswing, we'll add another 150 people.

现在他们却说，这个模型已经公开了。

Now they say we've got this model out there.

Claude Fable 5到底是什么？有多危险？

What is Claude Fable 5and how dangerous is it?

是的。

Yeah.

所以我觉得，Claude Fable 5其实就是 Mythos 模型的一个削弱版。

So Claude Fable 5is, I think to your point, a, I guess, neutered version of the mythos model.

正如你提到的，之前大家对这个模型在网络安全方面的危险性议论很多。

So as you mentioned, there has been a lot of talk leading up to this about how dangerous the model is for cybersecurity.

它以前能找出那些用了几十年的软件的漏洞，进行攻击。

It was able to find ways to hack software out there, software that's been around for decades.

但昨天发布 Claude Fable 5 时，Anthropic 已经让它无法回答任何与网络安全相关的问题。

But with the release of the model yesterday, Claude Fable 5, Anthropic actually made it so that it's unable to answer questions related to cybersecurity.

这类问题会被转到一个能力更弱的模型上去处理。

And instead those types of questions are routed to a weaker model.

这算是 Anthropic 的一个应对方案，用来缓解大家对发布后可能引发大规模黑客攻击的担忧。

So that was Anthropic's kind of workaround, you know, for some of the dangers and fears that people had around when it would be released, that there might be a giant increase in hacks, for instance.

对。

Right.

好吧。

Okay.

所以Anthropic说它完全安全。

So Anthropic says that it's fully safe.

我猜这家公司想怎么说都行，但实际效果还得看时间久了会怎样。

I imagine this is the type of thing that, I mean, the company can say whatever it wants, and we might see in practice how that plays out over time.

至于这个模型到底有多强，我相信你已经跟用过它的人聊过了。

Insofar as how good the model is, I trust you've been talking to people who might have been using it.

我们一会儿要跟Replit聊聊他们的使用体验，但目前大家对这个模型的效果普遍怎么看？

We're going to speak with Replit shortly about their experience, but what is the consensus about the effectiveness of the model?

是的。

Yeah.

我接触的一些早期测试者和开发者都说，老实讲，这个模型真的非常出色。

So early testers and developers I spoke with said that honestly, the model is really, really good.

我觉得之前大家可能有些担心，怕这又是那种吹得厉害但实际效果一般的情况。

I think there was maybe some fears leading up to this that this could be another example of kind of over promising and under delivering.

但目前为止，大家对这款模型印象都很深刻。他们说，在处理简单任务或者像查资料、问常识这类类似搜索引擎的问题时，可能不太容易看出它和以前模型的区别。

But so far, are very impressed with the model and say that while it might be hard to tell a difference between Fable and past models when it comes to simpler tasks or questions you might ask it like, you know, Google search esque questions you ask ChatGPT or Claude.

但它真正出彩、表现特别好的地方，在于那些非常复杂、耗时较长的任务。

Where it really stands out and does really well is with these super complex, you know, long running tasks.

我和一些开发者聊过，他们说他们只用一个提示词就做出了完整的游戏，而且

I spoke with developers who said that they were able to build entire video games with just a single prompt and, you

知道，

know,

非常少量的反馈和修复。

very minimal kind of feedback and fixes.

有一些例子显示，模型能够在几个小时内迁移代码行，你知道的，这可能需要一支人类工程师团队花几天时间才能完成。

There have been examples floating around about how the model was able to migrate, you know, of lines of code within a couple hours, you know, something that might have taken a team of human engineers, you know, days to do.

所以，看起来，你知道的，我认为实际上很多人提到的一个有趣的事情是，Claude Fable模型似乎有很好的品味。

And so, it does seem like, you know, I think actually one interesting thing that a lot of people mentioned is the fact that the Claude Fable model seems to have really good taste.

我跟一位开发者聊过，他想把自己感兴趣的几本书变成一个能实时朗读的播客，让他上下班路上听。

So I spoke with a developer who wanted to, you know, take some books that he was interested in and basically create like a live podcast that would read the books to him.

他就跟模型说：‘你能帮我弄个播客体验吗？我想在走路去上班的时候听这些书。’

And so he just told the model like, Hey, can you make kind of like a podcast experience so I can listen to these books while I'm walking to work?

结果Fable自己加了功能，比如支持0.5倍、1.5倍和2倍速播放。

And then Fable actually made its own choices around, I want to add features that let you listen to the podcast at one or 1.5 or two times speed.

它还自动加了字词高亮功能，让你听的时候能跟着文字走。

I want to add something that will highlight the words as you're listening to it that will let you like follow along while you're hearing the podcast.

这些功能开发者根本没提过，但模型自己就做出了这些决定，说明它真的很有品味，懂怎么设计好用的东西。

So all these sorts of things that, you know, this developer never explicitly told the model, but the model kind of has like good, you know, engineering taste basically or design taste.

整个过程完全没靠开发者额外提示，模型自己就搞定了。

And it was able to make those choices by itself without any prompting from the developer.

那这会不会成为今后的首选模型？

So is this going to be the go to model now?

还是说其他所有模型都过时了？

Or is every other model obsolete?

那情况到底是怎样的？

Or what's the situation?

也不一定。

Not necessarily.

你知道吗，我们在上线前好不容易挤出几个小时做了测试，Fable 的成本其实不低。

You know, as we actually were able to scoop a couple hours before the launch, Fable is not a cheap model.

它的费用比 Anthropic 之前最贵的 Opus 模型还要高出两倍。

It's actually two times more expensive than, you know, prior to this Anthropic's most expensive Opus model.

所以对开发者来说，这确实是个不小的障碍，大家只愿意在最重要或最复杂的任务上用 Fable。

And so that is definitely a deterrent for developers in terms of, you know, only really wanting to use Fable for their most important or complex tasks.

而且，我跟不少开发者聊过，他们都提到，像 GBT 5.5 和 Opus 4.8 这样的模型，其实已经能应付很多比较简单的需求了。

And, you know, again, like a lot of developers I talked to did say that you can use GBT 5.5 and Opus 4.8 for a lot of tasks that are maybe more simple.

所以我觉得，大家不会把 Fable 用在所有任务上，而是会留着它，专门解决那些最难啃的问题。

And so that I think will make it so that it's not like people are going to only use this new Fable model for every single type of task, But you know, it's kind something where you save it maybe for your hardest problems that you want to work on.

但你今天在文章里说，这个模型在 token 使用上很高效。

But you said in your column today that the model is token efficient.

你这话是什么意思？

What do you mean by that?

对。

Yeah.

再说一遍，评估一个模型的总成本时，得考虑两个因素：一个是每个令牌的成本，另一个是模型处理或输出的令牌数量。

So again, you know, when you think about the overall cost of a model, you have to kind of take into consideration two things, which is the cost per token and then the number of tokens that the model is, you know, processing or outputting.

这两点构成了成本公式的组成部分。

So those are kind of the two parts of the formula.

虽然Fable每个令牌的费用涨了不少，但开发者告诉我，它在生成更少令牌或需要处理的令牌更少方面效率更高，这反而降低了整体成本。

So with Fable, even though the cost per token has gone up quite a bit, developers are telling me that it's more efficient at producing fewer tokens or having to process fewer tokens, which brings the cost down.

原因之一是，这个模型聪明得多，比如你交给它一个工程任务时，出错的次数会少很多。

And so part of that is like, because the model is so much smarter, you can imagine that when you give it a engineering task, it doesn't make as many mistakes.

它不需要反复试错，也就不会产生大量额外的令牌。

It doesn't need to go through as much trial and error, which produces a lot of extra tokens.

所以出错少了，生成的令牌就少，整体成本也就降下来了，实际上可能还没到Opus模型的两倍贵。

And so if it's not making as many mistakes, it's producing fewer tokens that can actually drive the cost down, making it, you know, less than two times as expensive as using the OPIS model, for instance.

这确实是个不小的突破，对吧？

Well, and that strikes me as a pretty significant innovation, right?

因为每个token的价格是公司自己定的，他们能控制得了。

Because price per token, that is something that the company ultimately sets and they can control that.

用了多少token，直接反映的是模型背后工程做得好不好。

The number of tokens that are used, that is a direct reflection of how good the engineering is behind the model.

这个他们没法直接设定。

That's not something that they can set.

这完完全全就是看它有多厉害。

I mean, that's literally a reflection of how good it is.

这么说来，模型未来的发展方向还挺有希望的。

So maybe that's promising in terms of where the models are headed.

不过我想回到我们一开始聊的话题，Stephanie。

I want to go back to where we started this discussion though, Stephanie.

当初Anthropic刚推出Mythos的时候，大肆宣传它有多危险。

So when Anthropic initially came out with Mythos, they made all this noise about how dangerous it is.

你接触的那些开发者，觉得当初那些警告有道理吗？

The developers that you're talking to, do they feel like those warnings were warranted?

我的意思是，如果没有这些安全措施，这种模型真的可能带来危险吗？

I mean, is this the type of model that they could see having some dangerous implications if those safeguards were not put in place?

说实话，很难判断，毕竟能测试的范围很有限，尤其是像网络安全这类能力。

You know, again, is hard to tell just because you're very limited in how much you can test Fables like cybersecurity capabilities.

我和同事接触过一些最早拿到完整版Mythos模型的人，他们确实说这模型在这方面的能力很强，让人挺担心的。

My colleagues and I have talked to people that had early access to the kind of full blown mythos model and did say that it was very capable in this area and something that they were concerned about.

但还是很难下定论。

But again, it's hard to tell.

我觉得Anthropic在设置这些安全措施上做了不少有意思的决定。

And I think Anthropic has made some very interesting choices in terms of adding these safeguards.

他们还特别加了一项防护，就是防止这个模型被用来做AI开发。

Another area that they added safeguards in is actually preventing the model from being used for AI development.

这其实是个普遍现象，很多AI实验室都在用Anthropic或者OpenAI的模型数据来提升自己的模型。

And so this has been a common theme where you see AI labs basically using data from Anthropic or OpenAI models to improve their own models.

所以在这个最新案例中，Anthropic 添加了这些限制，防止用户进行这类操作。

And so essentially with this latest example, Anthropic added in these safeguards that prevent users from being able to do that.

这确实是个很有意思的决定，我觉得 Anthropic 是第一家这么做的公司。

So that's something that is an interesting choice that actually, I think Anthropic is the first company to do such a thing.

这样一来，人们就更难真正看清这个模型在 AI 研究或网络安全等方面的全部能力了。

And so it's making it harder for people to kind of really see the full capabilities of the model when it comes to using AI to do AI research or for cyber, for instance.

对。

Right.

OpenAI 对这个模型有什么回应吗？

Does OpenAI have an answer to this model yet?

据我了解，他们好像也有类似的措施。

So from what I'm hearing, it sounds like they do.

我接触过一位测试员，他能提前用到 OpenAI 和 Anthropic 的模型。

I talked to a tester who tests, gets early access to OpenAI and Anthropic models.

我问他这个问题时，他说 Fable 确实是个很棒的模型，但并不是我目前测试过最先进的一款。

And when I asked them about this, they said, you know, Fable is a really great model, but it's not the most advanced model that I'm currently testing.

而且他明显暗示，他们说的这个模型是OpenAI的，我们还得再等等看。

And you know, strongly implied that the model they were talking about is an OpenAI one that we'll have to wait and see.

再说一遍，这只是一个测试者的看法，但我认为这说明OpenAI可能正藏着什么大招，未来几周或几个月内或许会推出能对抗Fable的版本。

Again, this is just one person's opinion, but I think that's a positive sign that OpenAI could have something up its sleeve and perhaps an answer to Fable in the upcoming weeks and months.

好的。

Okay.

最后一个问题，Stephanie。

And last question, Stephanie.

我们最常聊的这个Fable 5，之前有没有出现过Fable 4、3、2、1？

The thing that we love to talk about the most here, the name Fable 5, was there Fable 4three 20 '1?

就是说，我们到底知道些什么？

Like, do we know anything?

难道我们其实错过了第五代，而这个才是真正的第五版？

Is this really the fifth iteration that we just missed

是的，第一个版本确实没有。

Yeah, the first no.

命名方式总是特别有意思，让人忍不住去琢磨。

The naming conventions are always super fun to kind of figure out.

我觉得这次Anthropic搞了一套挺特别的命名，跟诗歌有关。

I think in this case, you know, Anthropic has done this very interesting kind of, I guess, like poetry related naming conventions.

很明显，这个模型叫Haiku，还有Sonnet、Opus这些名字。

Obviously this model is haiku and then there's sonnet, there's opus.

所以我觉得Fable这个名字是想表明，这是一款我们前所未见的全新级别模型。

And so I think Fable is meant to kind of designate that this is like a whole new kind of level of model that we've never seen before.

那这个五，是不是因为之前4.7之后就跳过了？

And I guess the five is maybe the 4.7 was the So last

对，不过还挺有意思的。

yeah, but it's kind of funny.

到现在这个阶段，大家都会留意：你是打算叫4.9，还是直接叫5？

It's, you know, at this point, people pay attention to like, okay, are you going to call it 4.9 or are you going to call it five?

要是直接叫5，那就意味着它会强出一大截。

Like, and if it's five, that means that it's going to be so much better.

所以我觉得挺有意思的是，大家总爱对这些名字琢磨来琢磨去。

And so it's just kind of funny, I think, people read into all the different names and things like that.

我敢说，Anthropic的市场团队肯定在绞尽脑汁，认真琢磨这些命名和新功能。

I'm sure Anthropic's marketing department is thinking a lot about this and getting really deep with these naming, with these new features of the models.

至少比‘神话’这名字听着没那么吓人。

Well, it's less scary than mythos.

所以，感谢Stephanie来参加我们的节目。

So that sort of Stephanie, fits want to thank you for coming on the show.

这位是Stephanie Palazzolo，我们《The Information》的AI记者。

That is Stephanie Palazzolo, our AI reporter here at The Information.

想了解更多关于Anthropic新推出的Fable模型，以及开发者们实际使用时的反馈，我们请来了Replit的总裁兼AI负责人Michele Catasta，听听他的看法。

For more coverage on Anthropic's new fable model and how developers are actually feeling about it on the ground, I want to bring on Michele Catasta, president and head of AI at Replit, his thoughts on it all.

Michele，欢迎再次做客我们的节目。

Michele, welcome back to the show.

很高兴你今天能来。

It's great to have you here.

谢谢。

Thank you.

早上好。

Good morning.

感谢邀请我来。

Thanks for having me.

好的。

Okay.

你之前已经提前体验过 Fable 了。

So you have had early access to Fable.

你们团队一直在试用它。

Your team has been playing with it.

你第一印象怎么样？

What is your initial reaction?

整体来说表现非常出色。

It's been pretty exceptional across the board.

某种程度上，它让我们觉得正面对着一类全新的模型。

And in a sense, it gave us the feeling that we're in front of a new class of models.

我刚才听你和斯蒂芬妮聊命名的事。

And I heard you talk to Stephanie right before about the naming convention.

我觉得没必要为了性能提升就起个新名字——我们不仅在产品里感受到这种提升，尤其是Replit的开发者们，用这些前沿模型时，生产力的飞跃是实实在在的。

It doesn't make sense to me that, you know, we come up with a brand new name because the performance, you know, we're experiencing both in product, but especially our developers at Replit, you know, wearing some of our new class of what you can the productivity you can obtain as a developer, you know, using frontier models.

你们公司是怎么衡量这些模型的性能的？又是怎么把它们互相比较的？

Tell me about how you measure the performance of these models for your own purposes at your company and how you stack them up against each other.

上次我们提到，马上要推出ByteBench，一个端到端的字节编码评测基准。

So last time we talked about the fact that we are about to launch ByteBench, our end to end byte coding benchmark.

Fable在我们设计的某些任务类别里，表现几乎已经到顶了。

And what Fable has accomplished in certain classes of the batch for them we created is basically almost saturated already.

这简直是一次巨大的飞跃。

So it has been like a huge leap.

尤其是对于从零到一的任务，这种进步特别明显。

And this is happening especially for zero to one tasks.

我还看了社区的评论，很多人都支持这个方向。

And I read also the comments from the community that are very much agree with this direction.

你可以给出一个非常详细、很长的指令，从零开始描述你想构建的东西，而Fable已经达到了足够的自主性，让你能深入规划该做什么，然后在长周期的任务中执行，真正从零打造一个应用。

So you can give a very detailed long prompt of something you want to create from the ground up, and Fable has accomplished a level of autonomy as such that, you know, you can really plan in-depth what should be done and then execute task on our, you know, long horizon, you know, execution for creating an app zero to one.

在这方面，我们看到了很多任务完成得非常出色。

So on that side, you know, we're seeing a lot of task performance.

但它的实际能力在我们自家产品中的表现，还有待观察。

I would say it still remains to be seen how powerful it would be in the context of our own product.

正因如此，我们可能会以一种和市面上其他AI产品不太一样的方式推出它。

So for that reason, we probably launch it in a slightly different way compared to a lot of other AI products out there.

我们在Replit里新增了一个高投入模式，会保留判断权：什么时候该把任务升级到Fable，什么时候用Anthropic提供的更小模型做高级推理就够了。

So we have a new high effort mode in Replit, and we retaining the right of deciding why it makes sense to escalate all the way to Fable versus, you know, when we can just use higher level of reasoning with the, you know, smaller models, that Anthropic offers.

我们这么做是因为不想让快速使用的用户浪费掉那些宝贵的token。

And we're trying to do that because I don't want to have like regrettable tokens used by rapid users.

我觉得应该更看重token本身的内在价值，而不是只盯着Sting的成本。

I think we need to think of what's the intrinsic value of tokens rather than just the cost for Sting.

即使说是从零到一的突破，很多人在Rapid上用性能稍弱的模型，也能以低得多的成本完成95%的效果。

So even though, say it's exceptional zero to one, it might be the case for a lot of the apps that people create on Rapid running on a less powerful model accomplishes 95% of the results without much lower price point.

随着时间推移，我们会更清楚地知道，Fable在非技术用户身上到底哪里最突出，又在哪些地方其实有点大材小用了——毕竟这技术现在发展到这个地步，已经相当厉害了。

So with time, I think we're gonna understand exactly where Fable excels for non technical users and when it's getting almost overkill because that's how hard this technology has become at this point.

没错。

Right.

在你使用它的这段时间里，有没有发现它在哪些方面表现不够好，或者没达到你的预期？

Are there areas that you've seen it doesn't excel or ways in which the model has fallen short of your expectations in the short time that you've used it?

倒也不这么觉得。

I wouldn't say so.

总的来说，我们看到它的表现更好了，效率也更高了。

I mean, across the board, we are seeing better performance, higher efficiency.

它是一个更有目的性、更精准的模型。

It's a much more intentional model.

在那些

In those

这些情况主要发生在编程场景中。

instances And this is for I mean, this is for coding.

无论是聊天互动，还是智能代理功能，都是如此。

This is for just chat interactions, this is for agentic features.

到目前为止，你主要用它来做什么？

What use cases have you used it for so far?

是的。

Yeah.

我们90%以上的工作，尤其是在探索和研究新模型方面，都集中在编程上。

I would say 90 plus percent of our efforts in terms of discovery and researching new models goes on coding.

我们也会做一些纯粹的研究性代码实验，用来探索如何进一步提升它的能力。

We also run some pure research work codes, you know, to understand of how we can push it.

比如，我们有一个内部批量任务，能根据我们在Replit上的研究自动生成一篇学术论文。

For instance, we have an internal batch run that generates a scientific paper based on some of the research that we do on Replit.

到目前为止，这个模型在从最初的想法出发，扫描我们的代码库，结合研究和文献，最终生成一篇高质量的完整论文方面，表现得最为出色。

And by far, this has been the model that goes from the initial idea, scanning to our code base, applying, you know, the research and study web in mind and rather than the entire scientific paper in the best possible way.

这是一篇综述性论文。

It's a clearing paper.

你读完之后会觉得，这真是博士生写出来的。

You can read through that and you believe that's been written by a PhD student.

没错。

Right.

对。

Right.

我想问一下，你们应该没参与过Glasswing项目吧？如果我记错了请纠正我。

Now, I wonder, I don't think you guys were involved in Project Glasswing, although correct me if I'm wrong here at all.

但我一直在试着了解，别人觉得Mythos到底有多危险，以及Fable在多大程度上能应对这些担忧。

But I have been trying to assess or get an understanding from people how dangerous they really thought mythos was and the extent to which Fable will protect from those concerns that they raised.

基于你对它的有限了解，以及你目前看到的情况，你觉得Mythos真的那么危险吗？

Based on your, again, limited experience with it, but just what you're seeing, I mean, do you get the sense that Mythos was truly quite dangerous?

那你对Fable能避免这些问题，有多大的信心呢？

And are you confident in what you're seeing that Fable will not succumb to those threats, I guess?

但我们仍然面临这种风险，对吧？

Well, we're still about exposed to that risk, right?

因为Mythos还没像Fable那样完全向公众开放。

Because mythos hasn't been open completely to the public like Fable.

事后很难判断当初的策略到底对不对。

I think that it's hard post facto to assess if the strategy was correct or not.

不过我觉得，在这种情况下谨慎一点总没坏处。

At the same time, I do feel that being cautious in this context never hurts.

我们现在面对的是前沿模型，很难分辨一份公关稿到底是资深工程师写的，还是强大的AI生成的。

We're in front of a level of frontier models that it's hard to distinguish at this point if a PR has been written by a very experienced engineer or by, you know, a powerful AI model.

所以让每个写代码的人都能用上Fable这样的技术，是有道理的。

So it's it makes sense that we have this technology as in Fable in the hands of everyone who writes code.

我们可以花点时间，慢慢琢磨出最稳妥的安全功能发布方案。

I think we can take hours with time to figure out what is the best possible rollout for the security features.

要真正搞清楚我们发布的每一条安全建议对市场造成的影响，真的需要很长时间。

It takes a lot of time, you know, to really isolate all the impact that all the different security advisories we generate on the market.

不过我觉得，就在我们说话的时候，情况已经在变了。你去看看过去几个月Firefox修复了多少漏洞，就能发现，很多开源项目的数据都显示，最近几个月解决了大量安全问题。

I do think though that something is changing as we speak, you know, because if you go and check like how many bugs Firefox has been, you know, addressing the last couple of months, you know, there are a lot of statistics from open source models that show this from open source project that show how many security issues have been solved in the last couple of months.

所以我确实觉得，保持谨慎是有道理的。

So I do think it makes sense to be cautious.

也许我们目前面临的并不是什么生存性风险。

I don't, maybe this is not an existential risk what we're going through.

但我宁愿你选择谨慎，也不愿你一味地急着推进。

But if I prefer you not to be on the side of cautious than on the side of just rushing it.

总的来说，我觉得Anthropic这种稳扎稳打的策略是对的。

And overall, I think I believe the strategy of from Anthropic to take it easy.

我相信整个行业都在这么做。

And I believe the rest of the industry is doing exactly the same.

比如OpenAI，他们的做法也差不多。

Like OpenAI is also a very similar program.

所以我觉得，大家现在都达成共识了：花时间是值得的。

So I think everyone is aligned that it does make sense to take time.

你之前稍微提过这一点，但我还是想听听你对使用模型成本的看法。

And you touched on this a little bit earlier, but I do want to get to your observations on the cost of using the model.

我们之前请了斯蒂芬妮来，她谈到过每个令牌的价格，以及模型实际使用的令牌数量。

We had Stephanie on earlier, and she was talking about, again, price per token and then the number of tokens that the model actually uses.

Anthropic 称这个模型非常节省令牌。

Anthropic is calling this a very token efficient model.

根据你目前的观察，这个说法站得住脚吗？

Does that ring true based on what you're seeing, initially?

是的。

Yeah.

我们上线还不到24小时，但已经分析了大量早期用户的使用数据和调用记录。

So we we launched literally less than twenty four hours ago, and we have been already analyzing all the traces and the usage from from rapid users.

它在输入和输出令牌上都减少了两位数百分比的用量。

It does use double digit percent, fewer input tokens, fewer output tokens.

它还能更高效地并行调用工具。

It runs more parallel tool calls.

整体上，代理的运行轨迹也更短。

It's it it runs overall shorter traces of the agent.

所以在各个方面，效率都更高了。

So there is higher efficiency in any possible way.

至于最终的性价比，还得看完成同样任务时，Opus 和 Fable 5 的实际花费对比。

It remains to be seen dollar to dollar with this duration between accomplishing the same tasks, say with Opus versus, Fable 5.

不过话说回来，价格确实在持续上涨。

That being said, the the price point of course keeps going higher.

这其实跟我们一年前说的那套说法矛盾了——那时候我们还说智能的成本会降到零。

So it kind of defeats the narrative that we were talking about a year ago, where we claimed that the cost of intelligence would have dropped to zero.

事实并非如此。

It is not the case.

我能理解背后的原因。

I can understand why.

这些模型体量巨大，需要投入大量的研究和算力才能运行。

These are giant models that require all the research and computational power to play it.

不过我觉得，这对用户来说并不会带来明显的额外成本负担。

I don't think though that the impact that this will have to the user will be literally like a twist in terms of additional cost.

大概会多出一半左右的费用吧，不会更多也不会更少。

It's probably gonna be somewhere in between, say it would be like a 50% additional cost on top of it.

这就要看AI产品怎么设计了，得让用户用上更强的模型时，完全不会觉得后悔。

And it will be up to AI products to decide what is the best product experience you can give to your users such that they never regret using more powerful tokens.

我们正在快速推进，投入大量精力来做这件事。

And we do this on rapidly with high efforts.

其他公司则更倾向于依赖现代的路由方案。

Other companies are betting more on, you know, modern routers.

整个行业都在往这个方向走，毕竟用Fable来做所有事情，本质上就是大材小用，还太贵了。

The entire industry is moving in that direction because using Fable to do literally everything will be, by definition, very much overkill and way too expensive.

是的。

Right.

很好。

Great.

米凯莱，非常感谢你今天来做客。

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

这位是米凯莱·卡塔斯塔，Replit公司的总裁兼AI负责人，欢迎收看TITV。

That is Michele Catasta, President and Head of AI at Replit here on TITV.

OpenAI正在大力争取数据中心的算力资源，我的同事、负责云与计算的记者阿妮萨·加尔迪齐独家报道了该公司即将达成的一项重大交易。

OpenAI is making another big effort to secure data center capacity, and my colleague, Anissa Gardizy, our cloud and compute reporter, published exclusive reporting on a major deal the company is closing in on.

我想请阿妮萨来跟我们详细说说她了解到的情况。

I want to bring on Anissa to share with us more about what she learned.

阿妮萨，欢迎再次回到节目。

Anissa, welcome back to the show.

关于这项新的数据中心交易，我们现在知道些什么？

What do we know about this new data center deal?

OpenAI目前正在就租赁俄亥俄州一座10吉瓦的数据中心进行最后阶段的谈判。

OpenAI is in late stage negotiations on being you know leasing a 10 gigawatt data center in Ohio.

但非常重要的是，这座数据中心还不存在，目前只是个早期项目。

Now it's very important to know that this data center hasn't been built yet so it's still an early stage project.

不过，OpenAI 正在洽谈竞标这块地。如果他们能中标，并且这个数据中心真的建起来，那它毫无疑问将成为 OpenAI 迄今为止最大的 AI 数据中心设施。

But OpenAI is in talks to bid on the site And if it were to win the bid and this data center were to be built out, it would 100% be OpenAI's largest AI data center facility.

这个地点位于俄亥俄州南部，靠近一座正在建设中的新发电厂。

This location is in Southern Ohio and it's near a power plant a new power plant that's being built.

所以，是的，这就是我们报道的内容。

So, yeah, that's what we reported So

我们来详细拆解一下。

let's unpack it.

10吉瓦，这就是星门项目对吧？

So 10 gigawatts, that's Stargate, right?

我是说，星门当时宣布说是个联合项目。

I mean, Stargate was the announcement said that it was going to be a joint effort.

对吧？

Right?

孙正义在，特朗普在，拉里·埃里森在，萨姆·阿尔特曼也在。

It was Masayoshi Son was there, Trump was there, Larry Ellison was there, Sam Alt was there.

原本这会是一次大规模的联合项目。

That was going be a big joint effort.

现在是OpenAI说，我们要自己干，单独租下这个设施。

This is OpenAI saying, we'll just do it on our own and we'll lease it independently.

实际上，你刚才提到的那项与甲骨文和软银合作的Stargate计划，依然在进行中。

Essentially, I would say the original Stargate that you just mentioned with Oracle and SoftBank has kind of still gone on.

OpenAI目前正和甲骨文合作推进约六吉瓦的项目，属于这十吉瓦计划的一部分，另外还和软银能源合作了一个小项目，加起来差不多快到十吉瓦了。

So OpenAI is doing around six gigawatts with Oracle under that 10 gigawatt umbrella and then they've counted another small project with SoftBank Energy getting them kind of close to that 10 gigawatts.

所以这个是全新的、完全独立的十吉瓦项目。

So this is a completely net new 10 gigawatts.

没错，你说得对。

And yeah, you're right.

除此之外，还有这个。

In addition to that.

另外还有

In addition

再加上这一点。

to that.

在这种情况下，OpenAI 租用了一个数据中心，如果最终落实，这将是他们迄今为止最接近自主掌控基础设施的一步。

In this case, it is OpenAI leasing a data center which if they actually end up going through with it, be as close as they've come so far to sort of controlling more of their own infrastructure.

所以，这基本上就是 Stargate 2.0 了。

So yeah, they essentially are this is Stargate maybe two point zero.

明白了。

Okay.

在你的描述里，你提到这个项目，我的意思是，要运营或建设这么大规模的算力，投入高达五千亿美元。

And so now in your story, you outlined that this effort I mean, it's a $500,000,000,000 effort to operate this much compute or construct it.

帮我们理清楚一下，这个数字具体指的是什么？

Help us understand what that figure refers to.

这并不是 OpenAI 自己要掏的钱，对吧？

That's not the amount that OpenAI will be paying itself, right?

是的，这里的数字确实非常庞大。

Yeah, there are really big numbers here.

我可以帮你一步步理清楚这些意味着什么。

So I can sort of walk you through what this all means.

5000亿美元是按照当前价格建设10吉瓦数据中心容量的总成本。

So $500,000,000,000 is the number on the total cost of building out 10 gigawatts of data center capacity based on today's prices.

OpenAI并不会自己建造这些数据中心。

And so OpenAI isn't going to be building the data center.

全部由软银能源负责建设。

That is all on SoftBank Energy.

我们之前报道过，他们可能需要通过项目融资来为OpenAI建设这个10吉瓦的站点。

And we've reported that they're going to have to potentially you know borrow project financing in order to build out this 10 gigawatt site for OpenAI.

如果交易达成，他们还将签订一份长达二十年的租赁协议。

Now on the side, they're going to be on this twenty year lease if the deal goes through.

我们估算，仅租赁这组数据中心的费用，二十年内就将达到数百亿美元。

So we estimate that would be in tens of billions of dollars over the twenty year period just to lease this data center capacity.

而OpenAI还需要承担的另一项最大开支，是数据中心里所用的芯片。

And then the other thing that OpenAI would be unhooked for is actually the largest line item here which is the chips that go inside of the data center.

所以基于这五千亿美元的数字，硬件和其他芯片设备，比如网络设备之类的采购成本大约是三千五百亿美元。这些数字简直大得惊人，我们预计，如果OpenAI真走到这一步，他们很可能会找一个融资伙伴来帮他们采购这些芯片，把数据中心填满。

And so based on that $500,000,000,000 number, the you know amount of hardware and other types of chip gear that you might have to buy like networking, that would be around $350,000,000,000 And so these are just astronomical figures and we expect that if OpenAI progressed to that stage that they would likely seek a financing partner to help them fill this data center with chips.

而且这些设施显然也会分阶段建设。

It'll obviously get built out in phases as well.

对。

Right.

那现在，Nvidia就登场了，对吧？

And so now this is where Nvidia comes into the picture, right?

因为你们之前报道过，Nvidia正在为这个租赁安排提供某种财务担保。

Because you reported that Nvidia is putting itself on the line to be a financial backstop of sorts for this leasing arrangement.

那Nvidia到底可能要承担多少责任？

So what would Nvidia conceivably be on the hook for?

他们什么时候需要真金白银地掏钱？

When would they need to give money?

是的。

Yeah.

所以，Nvidia 的角色是，他们正在洽谈担任数据中心圈子里所说的财务担保人。

So the Nvidia angle here is that they are in talks to be what is known in the data center community as like a financial backstop.

我觉得这些担保安排之所以出现，往往是因为想租用算力的一方——比如 OpenAI——信用评级不够高。

And so it doesn't I don't think you know, a lot of these backstops happen because the party trying to access capacity, in this case OpenAI, is not investment grade.

举个例子，如果 SoftBank Energy 说他们的租户是 OpenAI，那他们很难说服别人借钱给他们。

So for example, it'd be very hard for SoftBank Energy to get people to lend them money if they say that their tenant is OpenAI.

尽管很多人喜欢 OpenAI，也觉得这个数据中心肯定会成功，但贷款方根本不关心这些。

You know, as much as people might like OpenAI and think that this data center is going to be successful, lenders don't really care.

他们只在乎租户是不是有投资级信用评级。

They just care about whether the tenant is investment grade.

所谓投资级信用评级，是指像上市公司那样，债券可以自由交易的那种吗？

So when Meaning investment grade, meaning if it's a public company where bonds can sort of be freely traded?

对，就是有信用评级的。而 OpenAI 作为私营公司就没有，Anthropic 也没有，他们情况一样。

Yes, and has a credit rating and that's something that OpenAI as a private company doesn't have, Anthropic doesn't have it, they're in the same boat.

所以 Nvidia 提议的是：我们来为 OpenAI 的租金支付做担保，这样一来，贷款方就会更愿意为这个项目提供资金。

So what Nvidia is proposing to do is say, you know what, if, you know, we will backstop OpenAI's lease payments, and that is something that makes lenders feel a lot more comfortable in lending money for a project.

所以这就是英伟达扮演的角色。

So that's Nvidia's role.

而且数据中心里肯定全是英伟达的芯片，所以这也有可能成为英伟达最大的芯片数据中心之一。

And obviously the data center would be filled with Nvidia chips so this could potentially be one of Nvidia's largest data center sites as well for its chips.

好的。

Okay.

这里再澄清一个细节。

Just to clarify one other detail here.

你刚才提到OpenAI的租赁费用，我们还不清楚具体数字。

So you mentioned that the OpenAI's lease payments, we don't know the figure.

金额大概会达到数百亿美元。

It'll probably come to the tune of tens of billions of dollars.

你之前说需要大约3500亿美元来确保芯片供应，这笔钱由谁来出呢？

The $350,000,000,000 ish that it would take to secure the chips that you said, who would be paying that money?

是软银能源来支付这笔钱，用于在数据中心部署芯片吗？

Would that be SoftBank Energy to put putting the chips in the data centers?

不是。

No.

根据我们目前对谈判的了解，应该是OpenAI自己采购这些芯片。

As we understand the deal talks today, it would be OpenAI doing the chips.

所以问题来了，他们到底怎么支付这笔钱呢？

And so, yeah, the question is, or how are they gonna be able to pay for that?

这肯定是他们在讨论的事情，不过还没到必须马上解决的阶段。

This is definitely something that I think they're discussing but is a little bit down the line in terms of when they actually have to figure that out.

但有一点可能成立，那就是Nvidia可以给OpenAI提供供应商融资。

But but we you know, one thing that could be plausible is that Nvidia does a vendor financing with OpenAI.

这显然是笔规模巨大的供应商融资。

Obviously, that'd be a very large vendor financing.

所以，如果将来真看到这样的安排，我也不会感到意外。

But yeah, I wouldn't be surprised if maybe that's what we see later down the line.

数据中心的第一期工程预计将在2028年完成。

The first phase of the data center is expected to be done in '28.

所以OpenAI还有点时间去想办法解决这笔资金的问题。

So there's a little bit of time for OpenAI to figure out how to get that March.

对。

Right.

根据你在这段故事里提到的，SoftBank Energy（SB Energy）似乎是这块俄亥俄土地的主要开发商，是吧？

Now we should say SoftBank Energy, SB Energy, as you refer to them in the story, they are the lead developer on this Ohio plot of land, sounds like.

SoftBank当然是OpenAI的重要投资者。

SoftBank is, of course, a big investor in OpenAI.

我们有没有迹象表明，这种关系可能会给OpenAI带来什么便利？

Do we have any sense for how that relationship maybe could make things easier for OpenAI?

比如，租赁条件会不会更优惠？或者有什么特别的安排？

Like, are these going to be, I don't know, favorable rates on the lease or maybe Interesting.

是的。

Yeah.

我不太清楚。

I don't know

就是这个答案。

the answer.

我只是在想，这里头所有事情都紧密关联着，对吧？

I'm just thinking, you know, like everything is so interconnected here, right?

是的。

Yeah.

关于这个具体地点，我们了解到，还没提到能源部和商务部在其中扮演的角色，还有美国政府……

So this is, what we learned about this specific site is that we haven't even talked about the connection to the Department of Energy and the Department of Commerce and how the U.

政府

S.

在推动这个数据中心落地方面，政府其实起了关键作用。

Government is kind of playing a role in getting this data center off the ground.

而且，为这个数据中心新建的发电厂，就建在能源部的土地上。

And so the this new power plant that's going be built for the data center is being built on Department of Energy land.

能源部曾要求软银能源，要公开透明地开展招标，挑选数据中心的运营商。

And one thing that the Department of Energy asked SoftBank Energy was to run a fair process to find bidders for the data center.

据我们所知，有不少公司都考察过这个地点，如果OpenAI目前进展最快，可能是他们出价更高。

And so as far as we know, you know, multiple companies looked at this site and if OpenAI is the one in advanced discussions maybe they offered a higher price than other people.

我们也不太确定。

We don't exactly know.

但据我听说，政府并不想因为SoftBank Energy和SoftBank有关系，就直接把整个项目全交给OpenAI。

But I think it was a it was from what I heard, the government was interested in not just saying, you know, because SoftBank Energy which has a relation to SoftBank is doing the site, it's 100% going to go to OpenAI.

没错。

Right.

这一点也很关键，因为在这种循环融资的环境下，你可以想象，如果政府不参与，事情可能会更容易办成。

And that's a key aspect to this too, because in this world of circular financing, you could imagine a scenario where if the government wasn't evolved, maybe they make things easier.

不过你说得对。

But that's a great point.

我现在想问问你关于这个容量数字，10吉瓦。

Now, I just want to ask you about the capacity number here, 10 gigawatts.

我们知道，OpenAI显然是一家快速增长的公司。

I mean, we know that OpenAI is obviously a fast growing company.

10吉瓦的容量可不小。

10 gigawatts is a lot of capacity.

你有没有听周围的人提到，OpenAI现在真需要这么多算力，还是说这是为未来准备的？

Did you get any understanding from people around, Does OpenAI need all of this today, or is this capacity that they're building for the future?

能不能给我们讲讲这10吉瓦到底意味着什么？

Just put the 10GW into context for us.

这绝对是为OpenAI的未来布局的算力。

This is definitely future capacity for OpenAI.

你知道，这个在俄亥俄州的项目第一阶段要到2028年才能投入使用。

You know, the first phase of this project in Ohio isn't set to be ready until the 2028 timeframe.

以现在的标准来看，公司可能觉得这是近期的算力需求，但我们说的这10吉瓦，真正用上估计还得好多年以后。

So in today's world, you know, the companies might view that as near term compute but the full 10 gigawatts that we're talking about here is likely many years away.

所以OpenAI现在就急着启动这么大的项目，是想提前占位——万一到2032年算力不够，至少他们已经有个租期二十年、算力灵活的大数据中心了。

So this is definitely a bet on OpenAI's part that they need to get the ball rolling on a large site like this early because you know if it comes to be 2032 and they don't have enough compute then they at least have this data center where they have a twenty year lease and a lot of flexibility around compute.

再说回这10吉瓦，这可是目前美国在建的最大数据中心。

And then just to put the 10 gigawatts into context, this is the largest data center under development in The US.

而且我觉得，我们对此也得保持一点怀疑态度，因为很多号称要颠覆格局的大算力公告，最后都没能实现。

And you know, I think we should all that also means that we should take it with a dose of skepticism because a lot of opening eyes, big compute announcements don't pan out.

即使他们现在谈得热火朝天，甚至签了租赁合同，我们还是得看下去——投资方会不会真掏钱？最终能不能达到这10吉瓦？

So even though they're talking about this and even might sign the lease, we'll definitely have to see, you know, do lenders buy in and is it able to reach that 10 gigawatts?

也许最后实际建成的只有四到五吉瓦，但即便如此，规模也已经非常惊人了。

Or maybe it ends up somewhere around four or five which is still so large.

尽管他们快敲定第一笔交易了，但很多事情还是悬而未决。

But there's a lot up in the air here even though they are nearing this initial deal.

很好。

Great.

安妮莎，非常感谢你今天来做客。

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

这位是安妮莎·加迪兹，我们《信息》杂志负责云与算力的记者。

That is Anissa Gardizy, our cloud and compute reporter here at The Information.

SpaceX的IPO不仅会波及整个股市，也会对那些运营股票市场的公司产生影响。

SpaceX's IPO will have ripple effects on the broader stock market, and it also has implications for the companies who are in the business of operating the stock market.