本周AI动态：Grok机器人（及4.6版本）、AI水印、Deepseek战略、湖人队出售

市面上出现了一个新的前沿模型，而且它来自一个不太可能让人想到的公司——SpaceX AI。

There's a new frontier model in town, and it comes from a not so suspecting company in SpaceX AI.

他们刚刚发布了Grok 4.6。

They just released Grok 4.6.

这现在可能是目前最好的模型，具体取决于你在意什么。

This is now the possibly the best model that there is out there depending on what you care about.

看这个基准测试卡，Grok 4.6 high在某些任务上的得分甚至超过了Fable 5。

As we look at this benchmark card, it shows that Grok 4.6 high is scoring higher than even Fable five on some tasks.

这真的很让人震惊，因为，我是说，就在上周，如果你问我们Grok在AI竞赛中的位置，我们会说，嗯，他们大概排在第三，跟Meta和Gemini竞争。

And this is shocking because, I mean, just last week, if you would have asked us where Grok stood in the AI race, we would say, well, they're like a distant third competing probably with Meta and Gemini.

实际上，有了这个新模型，他们现在直接冲到了最前面，而且不仅如此，新模型Grok 4.7预计会在未来三到四周内发布，而且会比这个更好。

The reality is that with this new model, they are now right at the front, and not only that, but the new model, Grok 4.7, is expected to come out in the next three to four weeks and be even better than this one.

这背后是Grok发布了一个叫GrokBot的新产品，我们稍后会详细讲，还有这周其他很多有趣的新闻。

This comes on the back of Grok releasing a new product called GrokBot, which we'll get into, as well as a lot of other interesting news from this week.

这是我们的周报，每周五发布，涵盖所有关于新闻和AI的内容，比如新云厂商的财报，一直到你最喜欢或最不喜欢的体育球队被收购的消息。

This is the roundup, the weekly roundup every Friday that we publish, where we cover all things about news and AI, things like covering earnings from neoclouds all the way over to the new acquisition of your favorite or least favorite sports team.

但我们得从Grok 4.6开始说起。

But we have to start here with Grok four point six.

他们回来了。

They're back.

他们现在就在前沿。

They're right here at the frontier.

这是个了不起的模型。

This is an amazing model.

确实回来了。

So back.

对。

Yeah.

不。

No.

很长一段时间里，SpaceX有点失去了前沿竞赛的领先地位。

So for the longest time, SpaceX has kind of lost the frontier race.

马斯克的工程团队之前犯了一些错误，结果落后于Anthropic、OpenAI和谷歌这些公司。现在说这话有点疯狂，因为谷歌后来也落后了不少，但他们现在正式回归了。

Elon's engineering team had made a few wrong moves and they kind of fell behind the likes of Anthropic OpenAI and Google, which is crazy to say because Google's fallen behind a lot since then, but they're officially back.

Grok 4.6在Grok 4.5发布大约三周后就推出了，而且是个更大更好的模型。

Grok 4.6 comes, I think three weeks after Grok 4.5 got released, and it is a bigger and better model.

现在具体来看，如果比较几个不同的基准测试，有个叫AA智能指数的东西。

And now specifically, if we look at a few different benchmark comparisons, there's this thing called the AA intelligence index.

这个指数是测试不同模型整体智能水平的一个很好的方法。

It's a really good way of testing just the general level of intelligence across different models.

在屏幕上，你可以看到它直接和GPT 5.6 Sol Max对比。

And on your screen here, you'll see it compared directly to GPT 5.6 Sol Max.

那是OpenAI最高端、最难的版本，还有Fable 5 Max。

So that's like the highest, hardest version from OpenAI, as well as Fable five Max.

你会看到它和GPT 5.6 Sol相当，只比Fable 5低一分。

And you'll see it is the equivalent to GBT 5.6 Sol and just a point off Fable five.

所以可以说它的整体智能水平已经达到那个级别了，但最棒的是，它的成本只有那些模型的一小部分。

So you could argue that its general intelligence level is right up there, but the best part is it comes at a fraction of the cost.

所以它的输出成本比Sol便宜5倍，比Fable便宜8.5倍。

So it's 5X cheaper output than Sol, 8.5X cheaper output than Fable.

总的来说，你大概能比这两个模型便宜3到5倍，这对那些用前沿模型的公司来说是个大事，他们想方设法把成本从几千万美元降到几亿美元。

Overall, you're kind of ranging between three to 5X cheaper than these two models, which is a big thing if you're like a company that's using these frontier models and want to kind of figure out how to cut your costs down from tens to hundreds of millions of dollars.

现在，他们能实现这种前沿智能的主要方式，其实有好几种，但最主要的是他们启动了一堆智能体。

Now, the main way that they're able to achieve this frontier intelligence, well, there's a few ways, but the main one is they spin up a bunch of agents.

所以Grok 4.6用多个智能体来帮你解决问题。

So Grok 4.6 uses multiple agents to kind of like solve your query.

这其实不是什么新鲜事。

Now this isn't something that's relatively new.

Meta在他们的较小模型里也在这么做。

This is something that meta's doing in their smaller models.

很多中国的开源模型也在做类似的事情。

It's something that a lot of Chinese open source models are also doing.

这是其中一方面，但第二点，我得强调，这是SpaceX最大的护城河。

So that's one part, but two, and this is like, I have to emphasize, this is SpaceX's like biggest moat.

他们拥有大量数据，很多有用的数据，这让这些模型变得极其聪明。

They have a lot of data, a lot of useful data that makes these models insanely smart.

训练的时候，他们可以直接把这些数据扔进去。

They can just bung it into it when they're training it.

还有海量的算力，埃隆在过去一年里一直在非常有策略地积累这些算力。

And also a hell of a lot of compute, which Elon has been very strategically amassing over the last year.

他拥有的算力比这个领域里任何人都多。

He has like the most compute out of anyone in the game.

我们开始看到这一点，就像你提到的，4.7版本几周后就要发布了，它会比这个更大更好。

And we're starting to see this, like you mentioned, like 4.7 is coming out in a few weeks time it's gonna be even bigger and better than this.

我觉得它会是一个3到5万亿参数的模型。

Think it's gonna be like a 3 to 5,000,000,000,000 parameter model.

所以你可以想象，随着成本优势和智能水平的提升，SpaceX现在正处于一个全新的境界。

So you can only imagine that, like, you know, with the cost benefits and with the intelligence starts increasing, like SpaceX is in a new world right now.

这非常令人印象深刻。

It's very impressive.

这就是事情开始变得真正有趣的地方，因为之前主要是能不能的问题，而不是什么时候的问题。

This is where, yeah, this is where things are gonna start to get really interesting because it was mostly a matter of if, not when.

现在感觉就是什么时候的问题了，他们什么时候真正达到前沿并超越它，原因有几个。

And now it feels like it's a matter of when, when they actually get to the frontier and eclipse it for a few reasons.

首先，正如我们一直说的，SpaceX团队是世界上最擅长制造实体东西、移动原子、快速搭建数据中心的团队。

The first of which is that, I mean, as we mentioned all the time, the SpaceX team is the best team in the world at building physical things, at moving atoms around, at compiling data centers quickly.

这跟造火箭比起来，简直是小菜一碟，就是把几台服务器放到地上的架子上，插上电就行了。

It is nothing compared to building a rocket ship to just put a couple servers on a rack on the ground and plug them into electricity.

所以他们一直做得又快又高效。

So they have been doing this quickly and efficiently.

在昨天的节目里，我们聊到了MicroHard，它是MacroHard的小型版本，但在集群中的GPU数量上，占地面积却完全相同。

In the episode yesterday, we talked about microhard, which is the small version of macrohard, yet it is the same exact footprint in terms of GPUs in a cluster.

所以基本上，他们已经找到了办法，把20万块协同工作的GPU装进只有MacroHard三分之一大小的空间里。

So basically, they've able to they've been able to figure out how to get 200,000 coherent GPUs into a package the size that's about one third of MacroHard.

所以他们在硬件上有着惊人的迭代过程，在软件上也有惊人的迭代过程，他们计划每月发布一次模型，这些模型会越来越强大。

So it's this amazing iterative process that they have on hardware, and then they have this amazing iterative process on software where they're planning to publish these models once a month that are incrementally stronger and stronger.

你刚才说，Ejaaz，我觉得我们现在是三到五。

You said now, Ejaaz, I think we're at three to five.

我相当确定他们有计划，参数数量大概在八万亿到十万亿。

I'm pretty sure there's plans for, like, eight to 10,000,000,000,000 parameters in terms of parameter count.

他们现在确实正在进行一个十万亿参数模型的训练。

They actually have that training run going right now for a 10 Tron in parameter model.

现在正在训练中。

It's in the oven right now.

所以等它完成训练后的对齐阶段，我们就能拿到它。

So we are going to get that after it's done with the post training alignment phase.

但关键是，这一切都会来的。

But the idea is that it's it's all coming.

一切都在发生。

It's all happening.

最难的部分基本已经完成了。

The hard part is mostly done.

他们从零做到了有。

They've gotten from zero to one.

他们已经重新达到了前沿水平。

They've reached it back to frontier.

现在就看他们能推多远、推多快了。

Now it's just a matter of seeing how far they could push it, how quickly.

我猜会很快。

I'm gonna guess, like, pretty quick.

我不知道。

I don't know.

说到其他优势，比如他们的数据集，你提到过，他们能接触到所有专有的太空火箭科学数据，这在科学和情报方面似乎很有价值。

When I think about other advantages like their dataset, like you mentioned, having access to all the proprietary space rocket science data that seems valuable in terms of science and intelligence.

基本上，他们有一个开放的语料库，是专有的，没有任何模型在上面训练过。

Basically, they have this open corpus of data that no model has ever been trained on that's proprietary.

对于那些不熟悉的人来说，星舰，就是要去火星的那艘巨型火箭，完全是人类建造的。

And for those that aren't familiar, like Starship, the giant rocket that's going to Mars, that was built a 100% by humans.

整个过程都没有用到AI。

There was no AI in the process.

所以所有的数学、科学、硬件图纸，这些都非常有价值，而Grok 4.6正是在这些数据上训练的。

So all of the math, all of the science, all of the hardware schematics, that's incredibly valuable, and that's what Grok 4.6 is being trained on.

所以这个模型非常出色，我很期待看到Grok接下来的发展。

So this model is an amazing model, and I'm very excited to see where Grok goes from here.

而且别忘了，马斯克已经说了，到今年年底，Grok 5就会发布。

Not to mention that Elon Musk has specified that by the end of the year, Grok five will be released.

我相信那会是一个拥有10万亿参数的模型。

I believe that will be the 10,000,000,000,000 parameter model.

所以这算是接近AGI级别的模型了，但它会是世界上最聪明的工程AI模型。

So we're talking about methos grade level model, but it'll be the smartest engineering AI model in the world.

他这么说的原因是，这个模型将完全用SpaceX的工程数据来训练。

And the reason why he says that is it'll be trained entirely on SpaceX's engineering data.

世界上没有比SpaceX更优秀的物理工程团队了。

And there's no better physical engineering team in the world that isn't at SpaceX.

所以他不仅在构建更优秀的通用模型，还在打造高度专业化的模型，希望能推动科学和材料科学的前沿发展。

So not only is he building better generalized models, he is building hyper specific ones that are gonna hopefully push the frontiers of science, material sciences forward.

现在，那些批评者，你知道，他们合理地指出埃隆经常许下很多大承诺，但并没有真正兑现。

Now, those haters out there that have, you know, validly pointed out that Elon makes a lot of big promises, but doesn't really fill his shoes.

但这次他确实做到了。

This is one instance where he definitely is doing this.

而且我觉得直到几天前，市场才把这一点计入价格。

And I don't think it was until a few days ago priced into the market.

但现在你看这只股票，过去五天已经上涨了26%。

But like right now, when you look at the stock, it is up 26% in the last five days.

我记得上周我们做节目的时候，股票还跌了，大概35%吧。

I remember when we made our episode last week, the stock was down, I don't know, like 35%.

我们当时确实说过，你知道，有很多利好消息要出来。

And we literally said that like, you know, like there's a lot of unlocks coming out.

我们理解这里有很多供应限制和问题，但如果你是中长期持有者，SpaceX显然有执行力的记录，能完成这样的壮举，而人们之前就是没有把这一点计入这家公司的股价。

We understand that there's a lot of supply constraints and issues here, but if you are somewhat a mid to long term holder, SpaceX so obviously has a track record for execution, being able to pull feats like this off and people just weren't pricing into the air company.

记住，你买SpaceX的股票，不只是买一家火箭公司，对吧？

Remember when you buy SpaceX stock, you aren't just buying like a rocket company, right?

你买的是一家卫星通信公司。

You're buying a satellite connectivity company.

你买的是进入太空、发射自己算力的最便宜的方式。

You are buying the cheapest way to get into space and launch your own compute.

如果你想利用这一点，你也是在买一家火箭公司。

If you wanted to harvest that, you're buying a rocket company as well.

你同时也在买一家人工智能公司。

You're And also buying an AI company.

哦，还有全球的社交媒体公司，它会分发很多这些东西，包括这个刚推出的新产品，我们一直在玩的那个叫GrokBot的东西。

Oh, also the social media company of the world that's gonna distribute a bunch of these things, including this new product that has just come out, which we have been playing around with called GrokBot.

据我了解，这基本上就是他们版的OpenClaw，只不过它真的能用。

It's my understanding that this is basically their version of OpenClaw, except it actually works.

我说得对吗，Josh？

Do I have that right, Josh?

这真是个了不起的产品。

It's a remarkable product.

我对GrokBot简直着迷了。

I am obsessed with GrokBot.

感觉SpaceX的AI团队直接跳到了智能体产品的终极形态。

It feels like the SpaceX AI team is jumping to endgame of what a agentic product could actually look like.

就像你提到的，它和OpenClaw很相似，但感觉它比OpenClaw改进和精炼得多，因为它完全自主。

Like you mentioned, it's very similar in sense to OpenClaw, but it feels like it's a much more improved and refined version of OpenClaw in the sense that it's fully agentic.

也许我可以解释一下它的工作原理。

And maybe I could explain the way it works.

基本上，你加入这个应用，它会给你一个类似iMessage风格的文本框。

So basically, you join this application, and it just gives you a text box similar in style to iMessage.

你可以想象一下发短信的界面，但你可以启动一个智能体。

You could picture, like, where you send your text messages, except you could spin up an agent.

你可以给它分配任务。

You can give them a task.

你可以给它们设定一个性格。

You can give them a personality.

你给它们起个名字，然后它们就会为你扮演那个人。

You give them a name, and then they will be that person for you.

比如，如果你喜欢锻炼，你去健身房。

So for example, if you like working out, you go to the gym.

你创建一个聊天机器人，然后说，嘿，你是我的健身伙伴。

You create a little chatbot, and you say, hey, you're my gym workout companion buddy.

你是世界级的健身专家。

You are a world class expert at fitness.

你会根据我对自己身体的这10个了解，给我制定一个完整的锻炼计划，我想每天和你一起努力，变得更健康、更健壮。

You are going to give me an entire workout program based on these 10 things that I know about my body, and I'd like to work with you every single day to actually get more fit and more healthy.

每天提醒我去锻炼，如果我没去，你就跟我确认，然后一遍又一遍地提醒我，直到我真的去健身房锻炼为止。

And remind me every day to go and work out, and if I don't, check-in with you, then remind me again and again until I actually get my ass to the gym and work out.

这只是它能力的一个很基础的例子，但我觉得它真的很有趣，不像OpenClaw那样，因为OpenClaw需要一些技术知识。

And this is a very basic version of what it's capable of, but I think it's really interesting in a way that OpenCLaw wasn't because OpenCLaw required a bit of technical expertise.

你得自己搭建服务器。

You need to host your own server.

你得在服务器上下载软件。

You need to download the software on the server.

你得自己维护。

You need to run maintenance.

你得管理记忆系统。

You need to manage the memory system.

你得处理各种细节和复杂问题。

You need to manage all the nuance and complexities.

用这个程序，你只需要登录，然后直接跟它对话就行了。

With this program, all you have to do is just log in and you just start talking to it.

它托管在他们的服务器上。

It hosted on their servers.

这个应用界面很漂亮。

The application is beautiful.

他们做了一个我非常喜欢的很酷的设计，就是

They've done this really cool thing that I love, which is the design of the

那些智能体的设计。

actual agents.

他们有一个非常可爱的动画。

They have this very cute animation.

它有点消除了你在跟那种有点吓人、有点诡异的AI打交道的感觉，把它变成了一种非常有趣的体验。

It kind of disarms the feeling that you're working with this, like, kind of scary, spooky AI and just turns it into a really fun experience.

而且它在手机上也能用。

And it's available on mobile.

在电脑上也能用。

It's available on desktop.

我真的强烈建议大家去试试。

And I, like, cannot encourage trying it out enough.

你甚至都不会意识到你是在跟一个类似Open Cloud的系统对话，因为设置起来太简单了。

And you don't even realize that you're talking to a system similar to Open Cloud because it's so easy to set up.

这些都在后台处理好了。

It's all handed on the background.

我觉得Grok这个产品真的很棒。

I think amazing product from Grok.

如果你正在听这个，心里在想，好吧。

If you're listening to this and you're wondering, okay.

这是一个新的AI模型吗？

Is this a new AI model?

我应该用这个吗？

Should I be using this?

我试过那个OpenClaw的东西，简直太糟糕了。

I tried that OpenClaw thing out, and it was terrible.

它把我银行账户里的钱全偷走了。

It stole all my money from a bank account.

让我试着解释一下这个产品可能面向哪些用户，好吗？

Let me try and explain like who this product might be pitched at, right?

所以如果你是个硬核软件工程师，每天都要写代码，那它并不是理想的选择。

So it is not the ideal product to use if you are a hardcore software engineer, if you want to code every day.

它的目标用户是那些花大量时间在软件或桌面任务上的知识工作者，他们想自动化一些比较琐碎的事情，比如邮件管理、日程安排或客户支持之类的。

This is pitched at sort of knowledge workers who spend a lot of time doing software or desktop based tasks, but want to automate some of the more, menial things like email admin or calendar scheduling or customer support and stuff like that.

这些事情真的很耗时。

These types of things take up a lot of time.

如果你把所有这些在各个行业里的经济价值加起来，那可是一笔不小的数目。

And if you kind of like sum up the economic value of all of it across like every single sector, it is quite a large sum.

我觉得要理解为什么GrokBot不只是个玩具，得看看埃隆·马斯克大概六个月前说的话，他说XAI，现在是SpaceXAI的使命，就是要替代大量人类基于软件的任务，不是要取代人们的工作，而是让人们不用浪费时间在那些没必要浪费的事情上。

And I think to understand why GrokBot isn't just a toy is to understand what Elon Musk said, I think about six months ago, which is it is very much XAI and now SpaceXAI's mission to replace a lot of human software based tasks, not so that it replaces people's jobs, but to give them the freedom to not waste their time on stuff that doesn't need to be wasted.

他们可以把工作中那些让人厌倦的枯燥重复劳动自动化掉。

They can kind of like automate away all the boring manual labor that you get tired of doing at work.

这样你就能腾出时间去做更有创意、更有影响力的工作。

And it frees you up to do more creative, impactful things in your actual job.

他把它称为“人类模拟器”。

He references it as a human emulator.

所以基本上，你可以和这些智能体互动，甚至都不知道对方是另一个人。

So it's basically, you can interact with these agents and you don't even know that it's another human.

这才是真正的考验，是另一端的基准测试。

That's like the real test, the benchmark on the other end.

这是迈向那个目标的第一步。

And this is the first step towards that.

那么，现在这个就是那个吗？

Now, is this that?

不是。

No.

但正如你提到的，乔什，这是一个非常精致且易于使用的版本，对吧？

But is this a really refined and easy to use version like you mentioned, Josh?

是的。

Yes.

你注册后，他们会在云端启动一台虚拟机，你的智能体就像撒欢一样运行，它们可以同时接管你的电脑，但这一切都以非常安全的方式进行，所以它们不会做出任何疯狂的事情。

You sign up, they spin up a virtual machine in the cloud and your agents just kind of like run riot and they can like simultaneously take over your computer, but it's done in a very safe mannered way so that they can't do anything crazy.

我读完这个产品的所有演示和反馈后，脑子里冒出来的第一个想法是，这直接针对的不是OpenAI或Anthropic，而是苹果，因为苹果几个月后就要推出Siri AI了，而且它是要完全在手机上运行和操作的。

And the thing that pops to the top of my head when I read through and watched all the demos of this product and read the feedback is this is a direct attack against not OpenAI or Anthropic, but Apple who is releasing Siri AI in a few months time now and is meant to function and perform completely on your phone.

它要跟你的所有应用和各种东西互动。

It's meant to interact with all your apps and all that kinds of stuff.

现在，GrokBot还做不到这些，但它在手机上就能用。

Now, GrokBot can't do that right now, but it is accessible and available on your phone.

你想什么时候给它发消息都行。

You can text it whenever you want.

而且我猜，最终它会获得某些手机的访问权限，可能先是安卓，因为苹果的生态系统太封闭了，但我认为这会是苹果的直接竞争对手。

And I can only imagine that eventually this will get access to some types of phones, maybe Android first, because the Apple ecosystem is so closed, but I see this as a direct competitor there.

而且我很高兴看到埃隆做出了一个能用的产品，而不只是一个模型，因为那是他最擅长的。

And I'm just so happy to see Elon make a functional product that isn't just a model because that's where he's best at.

对。

Yeah.

而且至少，你手机上能装上这个应用。

And at the very least, at least you get the app on your phone.

而且新的Siri其实大概一个月后就要上线了，所以很快就来了。

And and the new Siri is actually launching in about a month from today, so it's coming very soon.

但这感觉就像我当初用OpenClaw时的那种魔力，只不过更精致、更高效。

But this feels like the magic that I felt with OpenClaw, except a more refined and far more productive version.

目前它内置了g p c 4.6，我们刚说过，这个模型性能出色，价格却非常低。

Currently, it's baked in with g p c 4.6, which we just said is a remarkable model for an incredibly low price.

随着这些模型不断进步，把它们集成到这种智能代理体验中，感觉真的很棒。

As these models continue to improve, having those baked into this agentic experience feels pretty incredible.

就像你从口袋里掏出手机，就有一群聪明的超级计算机助手，能帮你做世界上任何事。

It's like you open up your phone out of your pocket, and you have this series of brilliant supercomputer helpers that can do anything in the world for you.

它们可以协同工作。

They could work together.

它们能获取你的上下文信息。

They can gain access to your context.

这感觉就像窥见了未来。

It feels like a peek into the future.

这感觉就像是OpenClaw 2.0。

This feels like OpenClaw two point o.

所以这真的是一个非常令人兴奋的产品。

So it's a really exciting product.

我觉得我们会密切关注这个，看看其他公司对这样的产品会有什么反应。

I think we're gonna follow this one closely and see what the other company's reactions to a product like this are going to be.

这就是Grok的更新。

But that is the Grok update.

恭喜你获得了新的前沿位置。

Congratulations on your new frontier slot.

期待下一个。

Looking forward to the next one.

下一个话题，我们还有一个前沿模型，不过这次是开源那边的。

For the next topic, we have another, I mean, kind of frontier model except this one on the open source side.

DeepSeek也带着新版本回来了。

DeepSeek is back with another release as well.

DeepSeek V4 Pro。

DeepSeek V four Pro.

而且价格底线又往下移了。

And the price floor has moved once again.

这可是个大新闻。

This is a pretty big deal.

对。

Yeah.

DeepSeek，如果我们回想一下，大概一年半前，其实也不算太久，它在AI领域引起了巨大的震动，因为那时候算力就是王道，Sam Altman如日中天。

DeepSeek, if we kind of hearken back to, what is it like a year and a half now, which is not that long ago at all, caused a massive disruption in AI field because back then compute was king, Sam Altman was on top.

只要你能买到足够的算力，你就能训练出最聪明的模型。

And if you could buy as much compute as you can, you could train the smartest models.

然后DeepSeek出现了，我记得它好像是从一家对冲基金里孵化出来的，梁文锋。

DeepSeek came along, which is, I believe, birthed out of a hedge fund, Liangwongfeng.

对，他是中国的一位对冲基金经理，基本上就是说，嘿，他让一群量化交易员去搞一个能辅助金融交易的模型，结果搞出了一个开创性的模型，而且不需要那么多算力。

Yeah, he's a hedge fund manager in China and basically said, hey, yeah, ran a bunch of quants and asked them to make a model that would help it with financial trading and ended up creating a groundbreaking model that didn't require as much compute.

这简直把市场搅得天翻地覆。

And it kind of like tore the market apart.

从那以后，他们大概又发布了一个模型，现在带着他们的重磅模型回来了。

Well, since then they've released, I think one more model and now they're back with like their big boy model.

不过这不是官方发布。

Now this isn't an official release.

一直有传闻说它很快会发布，但我们确实看到了关于这个模型DeepSeek V4 Pro基准测试的官方泄露。

This has been rumored to be released soon, but we had an official leak about the benchmarks for this model, DeepSeek V4 Pro.

没错，一如既往，它的成本表又被重新评估了。

Now, yes, it is rerated the cost sheet as always.

它一直是所有AI实验室中单位智能成本最低的。

It is always the cheapest cost per unit intelligence for an AI lab out there.

但让我震惊的是泄露出来的那些基准测试成绩。

But what struck me was some of these benchmarks that were leaked.

你看Terminal Bench 2.1，这是个很流行的编程基准测试，它得了87.9分，对吧？

Now, if you look at Terminal Bench 2.1, which is a popular coding paradigm or benchmark, it comes in at 87.9, right?

现在，如果你直接拿它跟Fable五比，那是什么水平？

Now, when you compare that directly to say Fable five, it is what is that?

在这个基准测试里，它的编码能力只有Fable五的十分之一。

Point one of Fable five's level of coding capability measured at this benchmark.

再看看Opus 4.8、KimiK三和GLM Fiber二，它们绝对碾压它。

If you look at, Opus 4.8, KimiK three, and GLM Fiber two absolutely crushes it.

但另一个让人担心的地方是这个。

But the other worrying part is this one over here.

就是CyberGym基准测试。

It's the CyberGym benchmark.

在网络能力上，它不仅不输Fable五，还稍微强一点。

It is not only rivaled, but it's slightly better than Fable five at cyber capabilities.

我担心的是，大概一个月前，Dario发了一篇很好的文章，说大约六个月后，我们会有一个开源版本的Methos级别的网络能力。

Now the reason why this worries me is Dario had a really good post, I think about a month ago where he said, in about six months time, we are going to have an open source version of Methos' level of cyber capability.

那会是个问题，因为那东西没有任何防护措施。

And that's gonna be an issue because it's gonna have no safeguards around it.

所以，Methos之所以没能给世界带来混乱，是因为它周围有防护措施。

So the reason why Methos hasn't been able to wreak havoc on the world is because they have safeguards around it.

DeepSeek如果以开源方式发布这个，会有很多很棒的效果，比如你可以在家运行它。

DeepSeek, if they release this in an open source fashion, is gonna have, you know, a bunch of amazing effects where you can run it at home.

更便宜、更私密，等等等等。

It's cheaper, private, blah, blah, blah.

但如果落到恶意行为者手里，它可能会造成混乱。

But also if it was put in the hands of a malicious actor, it could end up causing havoc.

所以我就想指出这一点。

So I just wanted to point that out.

虽然这还不是正式发布，但我会密切关注。

Like, it isn't an official release just yet, but it's something that I'm I'm keeping my eye on.

它快来了。

It's coming.

嗯，我们确实有正式发布，而且是以财报发布的形式出现的。

Well, we did have an official release, and it came in the form of an earnings release.

实际上，我们有两个这样的发布。

In fact, we had two of them.

一个来自CoreWeave，另一个来自Nebius。

One from CoreWeave and the other from Nebius.

对于不熟悉的人来说，这些是新型云服务提供商。

For those not familiar, these are neocloud providers.

新型云服务提供商，他们基本上是与GPU供应商合作，收集GPU。

Neocloud providers, they're basically they they work with the GPU providers to collect GPUs.

他们建立自己的基础设施，然后把这些基础设施租给那些想训练AI模型或用于推理等用途的人。

They build their own infrastructure, and then they rent out that infrastructure to people who want to train their AI models or use them for inference or whatever it may be.

所以你可以把它看作是提供AI训练或AI推理服务。

So you could think of it as, like, AI training or AI inference as a service they offer to people.

这两家公司这个季度的表现都非常惊人。

They had a pretty unbelievable quarter, both of these companies.

我们看看Nebius，当天涨了多少？

When we look at nevius, how much was it up on the day?

我在图表上看不到

I'm not seeing on

但我知道它涨了很多。

the chart, but I know it was up, like, a huge amount.

我记得它涨了

I think it was up

当天涨了25%。

25% on the day.

对。

Yeah.

而且我相信CoreWeave这周至少涨了31%。

And CoreBeat, I believe, was up 31% at least this week.

是的。

Yes.

是的。

Yes.

这两家公司的业绩都非常亮眼，主要原因是，就像我们上期节目提到的，GPU每小时的租赁价格正在飙升。

Both of these companies showed blowout numbers mostly because, as we mentioned in the previous episode, the price per rental hour of GPUs is just going through the roof.

每个人在最初评估这些新型云服务提供商时，都假设GPU会随着时间折旧。

Everyone, when they initially evaluated these neocloud providers, they were baking in depreciation of these GPUs over time.

因为当你拥有一块GPU时，你会认为它是折旧资产，每年都会降低其价值，直到达到一个稳定点。

Because when you have a GPU, you assume it's a depreciating asset, you mark it down year over year until it reaches a settling point.

但现在发生的情况恰恰相反。

What's happening is the inverse.

价格在上涨，所以他们赚的收入也在增加。

The number is going up, so therefore the amount of revenue that they're making is going up.

在Core Web的财报电话会议上，这里有一句话，说我们最近签了一份a 100合同。

And from the Core Web Earnings call, have a quote here, it says, we recently signed an a 100 contract.

再说一次，那些是2020年的GPU，合同延续到2029年，价格很有吸引力。

Again, those are the twenty twenty GPUs that extends into twenty twenty nine at an attractive price.

所以人们正在签三年期的GPU合同，而这些GPU是——

So people are signing contracts for three years out on GPUs that are

高出25%。

25% higher.

顺便说一下，比两年前租出去的价格还高。

Than they rented out two years ago, by the way.

太不可思议了。

That's incredible.

价格还在上涨。

Prices are still going out.

我是说，

I mean,

那看起来很不错。

that looks that looks great.

我们觉得，好吧。

We're like, okay.

这是一家好公司。

This is a good company.

但另外，我确实有

But then also, I do have

我得说，

to I do have to say,

我研究了一下收益，特别是CoreWeave的，因为我对它更熟悉，然后我发现他们其实有点麻烦。

I was looking into the the earnings, particularly of CoreWeave, which I'm more familiar with, and I realized that they actually have a bit of an issue on their hands.

他们在花很多他们没有的钱。

They're spending a lot of money that they don't have.

所以当我们看超大规模云服务商，比如谷歌和亚马逊，我们关注的一点是，他们花很多钱，但用的是自由现金流。

So one of the things that we look at when we look at the hyperscalers in particular, like Google and Amazon, is they're spending a lot of money out of their free cash flow.

他们不会负债。

They're not going into debt.

而Corbiv的情况是，他们有65亿美元的资本支出，但自由现金流是负的，接近60亿美元，这意味着他们在亏钱。

What's happening with Corbiv at least is they have $6,500,000,000 of CapEx, but they have a negative free cash flow of just under $6,000,000,000, which means they're losing money.

他们的债务大约有350亿美元，当我看到这些时，我就想，哦，等等。

They have debt around $35,000,000,000 of it, and this is a company when I see that, I'm like, oh, wait.

这就是我们一直在关注的问题。

That's the thing we've been looking out for.

超大规模云服务商没有这个问题，因为他们的收入足以抵消开支。

The hyperscalers don't have that because they're making enough revenue in order to offset that.

但这家公司，我就不确定了。

But this company, I don't know.

我刚刚注意到这一点，然后我就想，好吧。

I just noticed that, and I was like, okay.

心里稍微敲了一下警钟。

A little little flag went off.

我就想，好吧。

I'm like, okay.

财报很漂亮。

Great earnings.

股价涨了，但我们现在的支出有点超出承受能力了。

Number went up, but we're spending a little bit more than we can afford right now.

但他们还有1040亿美元的营收积压，而且刚刚签下了250亿美元的客户承诺，其中大概有50%到60%是预付的，光第三季度就这么多。

But they also have a revenue backlog of $104,000,000,000 and they just signed $25,000,000,000 in customer commitments, which I think about 50 to 60% of that is paid upfront, just for Q3 alone.

所以你指出债务积压的问题，绝对是对的。

So you're definitely right to call out the debt backlog.

我看到我们的金融分析师马丁·施韦勒，那个爱挑刺的家伙，提到他们负债累累。

I saw Martin Schwerle, our financial analyst provocateur, make the point around, that they're in a lot of debt.

但我认为这就是未来几年的游戏规则，很多超大规模云服务商，尤其是那些新云厂商，要想跟微软、谷歌和AWS这些大玩家竞争，就必须在AI资本支出上加倍投入，否则根本没戏。

But I think this is just the name of the game over the next couple of years where a lot of hyperscalers, especially these neo clouds who have to compete with the big boys like Microsoft and Google and AWS need to double down on AI CapEx if they want any shot of winning.

顺便说一句，如果你看一下新云的细分数据，我们应该简单聊聊营收数字。

And if you look over at the nebious kind of breakdown, by the way, we should get into the revenue numbers just very briefly.

营收同比增长了454%。

Revenue is up 454% year upon year.

这业务。

Is business.

而且环比也差不多，我觉得他们每个季度大概复合增长50%，这简直太疯狂了。

And that is not too indifferent quarter to quarter as well, where I think they're compounding by like 50%, which is just absolutely insane.

但关于Nebius，我要说的是，这算是CoreWeave的一个更成熟的版本，而且他们手头有大约80亿美元的现金。

But the point around nebious is this is a more, shall I say grown up version of Core Weave and they're sitting on about $8,000,000,000 in cash.

所以我觉得CoreWeave就像个年轻气盛的小伙子，特别急切，可能会飞得太靠近太阳，但他们确实有NVIDIA的支持。

So I think of like Core Weave being a younger gun that's like really eager and might end up flying too close to the sun, but they do have the backing of NVIDIA.

他们确实和Anthropic有合同，我记得，这基本上支撑了我之前提到的那个收入积压。

They do have contracts with Anthropic, I believe, which essentially prop up that revenue backlog that I mentioned earlier on.

我想解释一下，这些新云和那些超大规模云服务商之间的区别。

And I just want to explain something that separates what these neoclouds are versus like some of these hyperscalers.

你看谷歌、AWS、微软，他们在云服务领域已经深耕多年了。

When you look at Google, when you look at AWS, when you look at Microsoft, they've been in the cloud game for a while.

所以当AI这波热潮来的时候，他们基本上看了一眼，就说，哦，不错。

So when this AI stuff came along, they basically looked at it and said, oh, cool.

我在普通云服务上能赚30%到40%的利润。

I'm squeezing 30 to 40% margins on normal cloud.

AI应该也能做到同样，对吧？

I should be able to do the same with AI, right?

所以他们去找黄仁勋，买了所有这些GPU，然后发现，哦，黄仁勋拿走了70%的利润。

So they go to Jensen and they buy all these GPUs and they realize, oh, Jensen's taking 70% margin.

所以我在这上面能赚的百分比就少了。

So I have to take less percentage on this.

所以他们变得有点难打交道。

So they become a little more difficult to work with.

当你是一家想从这些超大规模云服务商那里租用计算资源的公司时，他们不会为你做所有事情。

When you're a company that wants to least compute for some of these hyperscalers, they don't do anything and everything for you.

相比之下，Nebius和CoreWeave会专门为你构建一个适合你特定工作负载的集群。

Compare that to nebius and Coreweave, they will go out to build a specific cluster that works for your specific workload.

这就是这两家公司或这两类公司之间的区别，也是我认为Nebius和CoreWeave未来会表现不错的原因。

And that's the difference between these two companies or these two types of companies and why I think nebius and Coreweave are actually gonna do pretty well in the future.

对。

Yeah.

希望如此。

I hope so.

走着瞧吧。

We'll see.

另外，我们还听到了同领域另一家公司的消息，就是英特尔。

We also had news from another company similar in the space with Intel.

我特别喜欢这家公司。

I love this company.

他们就像在说，嘿。

They they're like, hey.

其实，我们有点需要更多资金。

Actually, we kinda need some more money.

能不能给我们150亿美元？

Can we get $15,000,000,000?

他们开口要150亿美元，市场对此反应很负面。

And they asked for $15,000,000,000, and the market kind of hated it.

消息公布后，股价下跌了5%，因为他们宣布将发行150亿美元的普通股，来资助人工智能基础设施建设和增长机会。

The stock fell 5% after announcing that they are doing a $15,000,000,000 common stock issuance to help fund the AI build out and growth opportunities.

市场反应是，哦，我们不喜欢这个。

Market was like, oh, we don't like that.

又回到资本开支的问题了。

Again, to the CapEx thing.

不知道为什么，大家一提到资本开支就特别敏感。

Everyone's allergic to CapEx for some reason.

他们觉得这些公司不擅长分配资金。

They think that these companies are not good at allocating money.

显然他们错了，因为，我是说，就在昨天，英特尔已经反弹了。

Clearly, they're wrong because, I mean, in the last day, Intel has now bounced back.

他们表现还不错。

They're doing fairly well.

但这次反应背后的故事是什么呢？以及那种最初让他们有点不满的膝跳反应。

But what is the story behind this reaction and the kind of knee jerk reaction in which initially they were a little upset.

但今天看来，市场似乎开始清醒过来，意识到，等等。

But then it seems like today, the market's kind of coming to its senses and figuring, like, wait a second.

也许我们确实喜欢资本支出。

We maybe we do like CapEx.

嗯，人们显然喜欢，但特定的人，也就是英特尔的投资者。

Well, people obviously do like it, but specific people, which are the investors in Intel.

他们最初要求150亿美元，但最终需求达到了1000亿美元。

So they originally asked for $15,000,000,000, but they ended up getting demand for a $100,000,000,000.

他们想，好吧，我们不会接受1000亿美元的股权出售，但我们会额外接受50亿美元。

And they thought, okay, well, we won't take a $100,000,000,000 worth of equity sell, but we'll take an extra $5,000,000,000.

所以他们最终在不到24小时内筹集了200亿美元，顺便说一句。

So they ended up raising $20,000,000,000 in under twenty four hours, might I add.

现在你可能想知道他们到底为什么需要更多钱？

Now you might be wondering why on earth do they need more money?

他们为什么要出售股权？

Why are they selling equity?

他们为什么要稀释股东权益？

Why are they diluting shareholders?

首先，稀释其实没那么严重。

The well, number one, the dilution isn't actually that much.

谁在乎呢？

Who cares?

但如果你听了他们最近的财报电话会议，CEO李·布坦，他一直在大力改革英特尔，还有CFO，他们说了一些很具体的话，我引用一下。

But if you listen to their last earnings call, Lee Butan, who is the CEO, who has been changing things around a lot for Intel, as well as the CFO said something very specific and I'll quote it.

他说，事情进展得很顺利，我们有将近200到300亿美元的现金余额，可以花在任何事情上。

He goes, things are going really well and we have almost 20 to $30,000,000,000 cash balance that we can spend on anything and everything.

不过，如果事情进展顺利，我们可能会筹集外部资金。

However, if things are going really well, we may raise external capital.

他指的是，如果他们有很多已签署的合同和即将到来的大客户，想用他们的代工服务，有个叫英特尔代工的东西，英特尔基本上在建造一个芯片工厂，要和英伟达以及AMD这样的公司竞争，那么他们就会筹集外部资金。

And what he was referring to there is if they have a lot of signed contracts and upcoming customers, big customers that want to use their foundry, there's something called Intel Foundry where Intel is basically building a chip fab that is going to go up against NVIDIA and the likes of AMD as well, then they will raise external capital.

这是他们拥有这些客户的最明确信号。

This is the clearest signal that they have those customers.

所以我感觉在接下来的几个月里，我们会看到英特尔发布公告，说他们签了一个大客户，比如苹果、Anthropic或OpenAI，基本上就是说，是的，我们要为他们设计定制芯片，我觉得这是件大事。

So I have a feeling over the next couple of months, we're gonna see an announcement from Intel saying that they've signed a major client like, I don't know, an Apple or an Anthropic or OpenAI basically saying, yeah, we're gonna be designing custom chips them, which I think is a big deal.

所以我很高兴看到英特尔赢，我是它的忠实粉丝。

So I'm happy to see Intel winning, and I'm a pretty big fan.

对。

Yeah.

嗯，他们需要扩建的部分原因是因为，我们之前聊过，推理的需求量一直在增加。

Well, part of the reason why they're needing this build out is because, I mean, again, we talked about the increased amount of inference that has been happening.

因为智能体的存在，需要的推理令牌太多了。

There's just so many inference tokens required because of agents.

如果你在用智能体开发，我们有一个来自我们的朋友和赞助商Ledger的消息要告诉你。

And if you are building with agents, we have a message for you from our friends and sponsors over at Ledger.

因为很多时候，如果你在用智能体开发，你可能会担心安全问题。

Because oftentimes if you're building with agents, you're probably worried about security.

我们最近在这个领域看到了很多突破。

We've seen all these breakouts that have happened recently in this space.

有点吓人。

It's kinda scary.

Ledger 有办法解决这个问题。

Ledger has a solution to this.

这个方案分为三步：代理提出建议，人类批准，然后 Ledger 签名者执行。

It comes in the form of three parts where the agents propose, the humans approve, and then the Ledger signers enforce.

这三步流程能确保你在使用代理时是安全的。

This is the three step process to ensure that you are safe when you're working with agents.

Ledger 有一个叫 agent stack 的东西。

Ledger has this thing called an agent stack.

它完全是开源的。

It's totally open source.

它自带一些工具，能帮你管理你和代理之间的协作方式。

It comes equipped with tools to help you manage the way that you work with your agents.

它支持 Cloud Code、Codex、Cursor，基本上你所有和 AI 交互的地方都能用，而且完全开源。

It works with Cloud Code, Codex, Cursor, anywhere you really interface with AI, and it's totally open source.

你可以在下方描述里的链接中查看。

So you could check it out in the link in the description below.