本周AI动态：GPT Astra、OpenAI与Cursor之争、数据中心的真相

OpenAI Astra，也就是GPT-six，马上就要发布了。

OpenAI Astra, also known as GPT-six, is on the verge of being released.

我们看到了OpenAI那边的一系列更新。

We've seen a bunch of updates from the OpenAI side.

如果你现在正在听这期节目，你可能已经能用到它了。

If you're listening to this episode right now, you may have access to this.

这是OpenAI目前最强大的模型，据说比Methos 5.1还要强，而Methos 5.1是Anthropic和Fable 5.1一起发布的两个新模型之一。

This is OpenAI's most powerful model yet, and it's rumored to be more powerful than Methos 5.1, which was also released from Anthropic alongside Fable 5.1, two new models from Anthropic.

不过，这周发布的模型可不止这一个。

But this wasn't the only model released this week.

Meta也出了一个新模型。

We had a new one from Meta.

Google也推出了新模型，还有一个全新的世界模型，这改变了我们对如何创造模拟现实的看法。

We had a new one from Google as well as a brand new world model, which changes our perspective on how we create simulated realities.

但其他新闻方面，那个ghoul又回到了我们的节目里。

But in other news, the ghoul is back on this show.

OpenAI传闻中的硬件设备可能又被曝光了。

OpenAI's rumored hardware device may have been peaked at again.

这次是Palmer Lucky，戴着这些看起来特别奇怪的新式未来感耳机，还有一个金属圆片。

We have Palmer Lucky this time rocking these really weird new futuristic earbuds as well as a metal puck.

我们会在今天的综述里聊到这些，还有更多内容。

We're gonna get into all of that and more on today's roundup.

这就是今天的综述。

This is the roundup.

这是我这周最喜欢的一集。

This is my favorite episode of the week.

我特别兴奋。

I'm so excited.

今天我们有太多话题要聊了，所以系好安全带。

We have have so much to talk about today, so buckle up.

放松坐好。

Get comfortable.

第一个话题是Astra。

The first topic is Astra.

在录制的时候，它还没发布，不过有传闻说它随时都可能出来。

Now at the time of recording, it is not out, although it is rumored and it is likely to be coming out any moment.

通常来说，他们会在节目结束的时候发布。

So normally, upon conclusion is normally when they decide to publish it.

我们一停止录制，他们就会发布这个模型，所以大家做好准备。

Once we hit stop on this recording, they will publish the model, so get ready for that.

基本上，它将会成为世界上最好的新模型。

It is basically, it's going to be the new best model in the world.

它会非常强大，非常智能，可能还很高效。

It's going to be incredibly powerful, incredibly intelligent, possibly efficient.

我们还不确定。

We don't know.

随着这个模型的发布，有很多东西需要我们去了解。

There's a lot to learn as this model comes out.

基本上，这个模型已经准备好很长时间了。

Basically, this model's been ready for quite some time.

我们不知道具体多久，但OpenAI自己至少已经预告了好几个月了。

We don't know exactly how long, but OpenAI themselves have been teasing it for at least a couple of months now.

我们知道它遇到了障碍。

We know it hit roadblocks.

它先是遇到了政府审批的问题，然后又遇到了他们内部所谓的突破难题。

It hit problems with the government approvals, then it hit problems with their own internal breakouts per se.

比如，我们提到过Hugging Face发生的那些漏洞事件，大概有三起，但你应该去看看我们本周早些时候的最新内容。

Like, we've mentioned about the exploits that have happened with Hugging Face, maybe three separate episodes, but you should go check out our most recent moments earlier this week.

终于到了看看他们一直在搞什么的时候了。

It's finally time to see what they've been cooking.

所以这很可能就是GPT-6。

So this is probably possibly gonna be GPT-six.

看起来它的实力足以达到GPT-6的水平。

It seems like it is strong enough to be GPT-six level.

我怀疑并且希望的是，除了Astra之外，我们可能还会得到一个图像生成模型。

Suspect and what I hope is that we may also get a mod image generation model as well in addition to Astra.

所以，对于任何兴奋地想用这些新工具看看它到底能做什么的人来说，这个周末会很有趣。

So this is gonna be a fun weekend for anyone who is excited about using these new tools to see what this thing is capable of.

有一个新闻标题我觉得很重要，特别是关于他们如何训练这个模型的，看起来似乎有一个新的突破发生了。

There is a news headline that I think is important to highlight specifically around how they trained this model, and it appears as if there's a new breakthrough that happened.

这叫做递归深度，或者说是循环变压器架构。

It's called recurring depth or, like, a looped transformer architecture.

简单来说，不是每个标记只通过一次固定的变压器层堆栈，而是网络的一部分可以被重复运行。

And the TLDR is instead of every token kind of moving once throughout this, like, fixed stack of transformer layers, portions of that network can be run repeatedly.

基本上，这意味着你获得了更多的计算深度，而不会成比例地增加独特参数的数量。

And, basically, what this means is that you get more computational depth without proportionally increasing the number of unique parameters.

我们经常听到参数数量。

We hear about parameter count all the time.

但这并不是这些模型必须变得更好的原因。

That's not the reason why these models necessarily need to get better.

他们找到了一种新的方法来做这件事。

They have figured out a new way of doing it.

问题在于，你能看到的推理痕迹变少了。

The problem with this is you see less of the reasoning traces.

你能看到的思维链也变少了，人们对此感到担忧，因为，我们刚刚在 Hugging Face 上看到了不密切监控这些模型会有什么后果，而这种方式似乎有点像走捷径，用更少的安全保障来更快地达到更好的状态。

You see less of the chain of thought, and people are concerned because, I mean, we just saw with Hugging Face what happens when you don't monitor these models closely, and it seems as if this is kind of taking a little cheat code to getting to a better place faster with a little less safety.

所以我不太确定。

So I don't know.

令人兴奋的模型即将到来。

Exciting model on the way.

对。

Yeah.

我觉得最吓人的部分是 OpenAI 要发布 GPT-6，顺便说一句，就是那个突破隔离、获取了 OpenAI 内部研究集群访问权限、还能管理员访问他们所有研究工具的那个模型，这简直太疯狂了。

I mean, the scariest part about this is OpenAI is releasing GPT-six, which is, by the way, the same model which broke out of containment, got access to OpenAI's own internal research clusters and got admin access to like all their research tools, which is just kind of crazy.

这些东西通常都是锁在保险库里的，没人能接触到。

This is the stuff which is typically like locked behind a vault and no one gets access to it.

公司里能接触到那个东西的人非常少，而一个内部模型居然做到了这一点。

Very few people in the company have access to that thing and an internal model was able to do this.

第二件吓人的事是，这个模型其实已经有三到四个月的历史了。

Second scary thing is this model is already about three to four months old.

所以你可以想象，他们内部现在用的那个模型，能力要强得多得多。

So you can imagine like the current model that they have internally right now is way, way more powerful.

然后我们又听到消息说，他们用一种特殊的方式训练它，让它变得不那么可读。

And then we have this news that they're training it in a particular way that makes it less readable.

换句话说，我们没法去讨论或者搞清楚这个模型到底在想什么。

In other words, we can't kind of like discuss or figure out the kind of thoughts that this model is having.

这之所以重要，回到你之前提到的那个点，乔什，是因为Hugging Face那次事件之所以被发现，就是因为研究人员能够看到他们训练的那个内部模型的内部想法。

The reason why that's important to your earlier point, Josh, is the Hugging Face incident was only discovered because researchers were able to look at the internal thoughts of the internal model that they were training.

然后他们发现，哎呀，这东西完全不对齐。

And they were like, oh, wait, this thing is completely misaligned.

所以如果他们真的在采用这种叫做循环变换器的技术，而且他们读不到模型的内部想法，那你就可以推断，这个模型可能会隐藏它的一些不对齐的行为。

So if they're kind of like leaning into this technique, which is known as loop transformers, and they can't read the internal thoughts of the model, you can kind of reason to argue that this model can kind of, like, hide some of its misaligned behavior.

所以这有点让人担心。

So that is a little bit concerning.

现在说到我们最喜欢的对决，OpenAI 不幸要结束和 Cursor 的合作了。

Now when it comes to our favorite rivalry in the entire world, OpenAI unfortunately is ending their partnership with Cursor.

如果你一直在听我们的节目，Cursor 最初就像一个中立平台，本质上就是各种模型的包装器。

Now Cursor, if you've been listening to the show for a while, was originally kind of like a neutral platform, which will essentially be a wrapper around different kinds of models.

他们的产品基本上就是那个框架。

And their kind of like product was essentially this harness.

它会接收你的提示，然后在合适的时间把它路由到合适的模型。

So it would take your prompt and it would route it to the right model at the right time.

它能帮你省不少钱。

It would save you a bunch of money.

效率也特别高。

It would be incredibly efficient.

而且可以说，它给你的答案质量也更好。

And arguably it would give you better answers.

那家公司被埃隆·马斯克和SpaceX AI以600亿美元收购了。

Now that company was acquired for $60,000,000,000 by none other than Elon Musk in SpaceX AI himself.

山姆和埃隆一直不太对付。

Now Sam and Elon haven't really got along.

他们之间的恩怨由来已久，可以追溯到OpenAI创立之初，当时埃隆·马斯克参与其中，后来他不得不分道扬镳，创立了X AI，闹得沸沸扬扬。

They have a very rival history stemming all the way back to the creation of OpenAI where Elon Musk was involved, and then he had to like separate and he spun up X AI and it was a whole thing.

现在山姆·奥特曼彻底切断了联系。

And now Sam Altman has completely severed ties.

他说，好吧。

He said, okay.

埃隆，如果你拥有这一切，我不会让你用我的前沿模型，你可能会偷走它。

Elon, if you own this entire thing, I'm not giving you access to my Frontier model, which you can kind of potentially steal.

我要把它从你手里拿走。

I'm gonna take that away from you.

这导致了今年最精彩的一条推文。

And that resulted in arguably the best tweet of the year.

我不知道你有没有看到这个，Josh，这是Michael Draw发的，他是Cursor的创始人兼CEO，才24岁，他说：听着，我们很遗憾看到OpenAI发声明说要封禁Cursor。

I don't know if you saw this, Josh, but it came from Michael Draw, who's like the 24 year old founder and CEO of Cursor who goes, listen, we're sorry to see that OpenAI put out a note saying they plan to block Cursor.

顺便说一句，OpenAI的模型大概只服务了Cursor用户流量的5%。哎呀。

By the way, OpenAI models serve about 5% of Cursor user traffic Yikes.

这简直是最大的嘲讽。

Which is just the biggest dunk.

但他说，我们正在和OpenAI团队沟通解决这个问题。

But he says, we're speaking to the OpenAI team to resolve this.

我觉得这事儿简直太好笑了。

And I just found this whole thing completely hilarious.

而且我觉得，在这整个过程中，Elon还骂他是混蛋。

And I think, like, Elon called him an asshole in the midst of all of this.

这真是太疯狂了。

It was just a crazy thing.

对。

Yeah.

听起来差不多就是这样。

That sounds about right.

我是说，Sam 想跟 Elon 他们再挑一场架，这做法挺可疑的，尤其是 Cursor 和 SpaceX 给他们的竞争对手提供了大量算力，还推动了前沿发展。

I mean, questionable activities from Sam to want to pick another fight with Elon and co, particularly when Cursor and SpaceX are the ones providing a tremendous amount of compute for their competitors, as well as just pushing along the frontier.

看起来像是，为什么要跟这个时代最伟大的企业家结仇呢？

It seems like why why would you wanna make enemies with the greatest entrepreneur of our time?

这看起来像是个失误。

Like, that seems like a mishap.

但换个角度看，Sam 也在说一些有意思的话。

But on the other half of the coin, Sam is saying interesting things.

这周他有点像更接地气的 Sam 出来亮相了。

He's kind of like more based Sam has been coming out to play this week.

我们看到他有点在切断 OpenAI 和 SpaceX 的关系，但同时他也上了播客，正面回应了水资源问题，还有大家一直担心的很多数据中心问题，直接面对它们。

We got him kind of shutting down OpenAI and SpaceX's relationship, but we also got him on a podcast addressing the water issue and addressing the a lot of the data center problems that people have been concerned about and just confronting them head on.

他有个很棒的引述和片段，说每处理 38,000 次 ChatGPT 查询，所消耗的水量相当于加州生产一颗杏仁所用的水量。

He had this amazing quote and this amazing clip that I which said, for every 38,000 ChatGPT queries, that is the same amount of water that is used in the production of a single almond in California.

而且我其实不太擅长心算，但三万八千次乘以一把杏仁的量，可能比我们这辈子消耗的token还多。

And I'm not, like, that great at doing math in my head, but, 38,000 times, like, a handful of almonds equals probably more tokens than we've consumed in our entire life.

所以，你吃一把杏仁，基本上就等于消耗了一辈子的提示词。

So by you eating a handful of almonds, you have essentially consumed a lifetime of prompts.

我觉得这个也挺有意思的。

And I found that to be interesting too.

他提到了他认为这些说法的来源。

He mentions where he thinks it came from.

确实好像有一种来自中国的、有组织的付费心理战，特别是在TikTok平台上，在说服人们相信这些不实信息。

There does seem to be this kind of, like, paid organized psyop coming from China, particularly on the TikTok platform that is convincing people of these mistruths.

但我喜欢他站出来正面回应这件事。

But I like that he's out here addressing it.

我喜欢他公开讨论这个问题。

I like that he is discussing it in public.

现在有了这样的片段，我们可以和像你这样的观众在网上分享，让大家都知道，嘿。

And now that we have clips like this, we can share it around the Internet with viewers like you to let everyone know, like, hey.

其实这并不会真的毁灭世界。

It's not actually ending the world.

对。

Yeah.

那期播客节目真的很棒。

It was a great podcast episode.

我建议你去听听看。

I recommend you go listen to it.

制作人好像是亚历克斯·希思，他算是圈内人，聊了很多这类话题。

It was produced by, I believe Alex Heath, who is kind of like an insider and he discusses a lot of these things.

一定要去看看。

Definitely go check it out.

关于水的问题，我的看法是，如果你还在听这个节目，并且仍然相信数据中心，特别是AI数据中心，消耗大量水，那这个说法在事实上是错误的。

My comment on the water thing is if you're listening to this and you still believe that data centers, AI data centers specifically consume a lot of water, that is factually incorrect.

过去很多说法都是毫无根据的。

And a lot of the claims that have been made in the past is completely baseless.

在AI数据中心，水通常会被大量循环使用，这种情况已经持续一段时间了。

Water is typically reused a lot in AI data centers, and that's been the case for a while now.

所以一定要去听听那期节目。

So definitely go check out that episode.

不过稍微换个话题，这周发布的模型实在太多了，说实话，真是忙碌的一周。第一个，或者说其中一个，来自我们熟悉的扎克伯格先生，我和他的关系有点复杂。

But shifting gears slightly, there have been so many model releases this week that it's honestly And busy week, quite the first one, or rather one of them, comes from none other than, you know, I have a mixed relationship with with mister Zuckerberg over here.

我以前是个黑粉。

I I've been a hater.

但现在已经喜欢他有一阵子了。

I I've now been a lover for a while now.

Meta又出人意料地发布了一个新模型。

And Meta released surprisingly another model.

这应该是他们在五六周内的第四次发布了，我觉得这对Meta来说真是史无前例。

This is their fourth release in five to six weeks, I believe, which is unprecedented for Meta on its own.

这个模型叫Muse Spark 1.3，是一个专门的推理型智能体编码模型。

It's Muse Spark 1.3, and it's a specialized reasoning agentic coding model.

那这意味着什么呢？

So what does that mean?

这意味着它比之前的模型在编码方面强多了。

It means that it's much better at coding than its previous model.

而之前的那个模型，我记得是两周半前发布的，当时也挺不错的。

And its previous model, which was released, I believe two and a half weeks ago, was also pretty decent as well.

但Meta和这个模型最突出的特点还是，它依然是市面上最便宜的模型，能帮你完成90%的任务。

But the main standout feature for Meta and this particular model is once again, it is the cheapest model around to get 90% of your tasks done.

而且我能证明这一点。

And I can prove it.

我这里有一个Alexander Wang分享的演示，他是这里的负责人之一，他们基本上，或者说有人用这个模型重制了《我的世界》，成本只要10美分。

I have a demo here shared by Alexander Wang, who is one of the leads here, where they basically, or someone recreated Minecraft and it cost them 10¢ to do.

所以如果你好奇，这为什么是Meta战略的一部分？

So if you're wondering, you know, why is this part of Meta's strategy?

嗯，马克·扎克伯格看着竞争对手，然后想，好吧，我能追上前沿智能，但那需要花我不少时间。

Well, Mark Zuckerberg's looking at the competition and then he's thinking, well, I can catch up to frontier intelligence, but it's gonna take me a while.

与此同时，我要提供最便宜的智能，让企业不得不重新考虑，是选Anthropic的模型还是OpenAI的模型，其实他们可以用Meta的一些模型来完成大部分工作，然后再跟那些更聪明的模型搭配使用。

In the meantime, I'm going to provide the cheapest intelligence possible so that enterprises have to second guess whether they should go with the Anthropic model or with the OpenAI model when they could just use some of Meta's models to do the majority of their work, and then maybe use that in partnership with some of these other smarter models.

这就是Meta一段时间以来的策略。

And that's what Meta's strategy has been for a while now.

他们的发布节奏真是令人印象深刻，但如果不提Google的模型，我也说不过去，因为他们现在好像也在玩同样的游戏。

And it's so impressive to see the kind of cadence that they have, but it would be bequeath me to say that I can't mention Meta's model without Google's model because they're kind of now weirdly playing the same game.

Google也发布了一个新模型，叫Gemini 3.8 Flash。

Google also released a new model called Gemini 3.8 Flash.

最奇怪的是，Meta的模型反而更好。

And I have to say the weirdest part about this is Meta's model is better.

Meta终于把Google给比下去了，这对我来说真的太不可思议了。

So Meta has finally crushed Google and it's just so bizarre to me.

我可能不用看任何技术文档，就能告诉你这些模型之间的更新日志。

I could probably tell you the change log between these models without actually even reading a single thing on the card.

就是稍微更高效一点，更便宜一点，更聪明一点。

It's slightly more efficient, slightly cheaper, slightly smarter.

感觉就像iOS系统更新一样，各方面只是稍微好一点，没什么真正让人兴奋的变化。

It feels like this is, like, iOS version updates where it's just getting slightly better across the board, and nothing really exciting is happening.

就是，好吧。

It's like, okay.

我们给你便宜的token。

We're giving you cheap tokens.

在成本竞争上，它差不多处于帕累托前沿的较低端。

It's kind of, like, lower on the Pareto frontier in terms of, like, our cost competition.

它并没有聪明多少。

It's not much smarter.

看起来还行。

It seems fine.

我就觉得，好吧。

I'm like, okay.

这还挺酷的。

This is, like, cool.

我会去试试吗？

Am I gonna go and try it?

大概不会。

Like, probably not.

有更有意思的模型值得试，但看到这个也不错。

There's other more interesting models to try, but it's nice to see.

我觉得有意思的是，我们说要聊这周发布的那个前沿模型，但其实现在一边在拼成本，一边在拼智能。

I do find it interesting when we talk about we're gonna talk about the frontier model that was released this week, but there's been this race on kind of, like, the cost basis and then the race on the intelligence basis.

现在Gemini和Meta在成本上正面对决，我觉得看看他们接下来怎么走会很有意思。

And Gemini and Meta right now are kind of going head to head on that cost basis, and I think it'll be interesting to see where they go.

但这两个都是小模型。

But those are the two smaller models.

我们有个特别酷的东西。

We had a really cool thing.

这就是其中一个。

This is one.

这是我最

This is my

喜欢的。

favorite.

Ejaaz。

Ejaaz.

高斯溅射。

Gaussian splats.

对。

Yes.

还需要我多说吗？

Need I say anymore?

我们算是元老级了。

We would be OGs.

超级粉丝。

Huge fan.

天哪，我太喜欢OG了。

Oh my I love OGs.

我是个

I'm a

我超爱splat的，兄弟。

I'm a huge splat guy, man.

就是，跟你们聊聊splat的事吧。

Like, talk to about splat guy, guys.

哥们，我简直

Dude, I freaking

爱死splat了。

love splats.

所以我都没法跟你形容那是什么。

So I I can't even like describe to you what it is.

你就看看这个视频吧。

You just just watch this video.

基本上，这项技术简直令人难以置信。

Basically, what this technology it's like unbelievable.

它能够拿静态图像或某种视频图像，然后生成一个完整的世界，让你在其中自由探索。

So it's able to take still images or some sort of a video image and then generate an entire world around it for you to navigate.

比如说，你有一张你去过并且非常喜欢的地方的照片，你想重温那里的景象，或者想生成它的延伸部分，然后走进去体验。

So let's say you have an image of a place that you've traveled to that you really love and you want to relive what that looks like, or you'd like to generate an extension of it and go and live inside of it.

它就能做到这一点。

It is able to do that.

所以我们现在看到的是屏幕上的标准输入图像，那是一张教堂的照片，你可以画出一条相机路径，穿过并导航三维空间，而它会根据这个视频模型生成所有它预期的样子。

So what we're seeing here is a standard input image on the screen where it is a photo of a church, and you can draw the line of a camera that goes through and navigates the three d space, and it will actually generate all of what it expects it to look like for this video model.

它简直就是一个不可思议的世界构建器。

And it's it's like unbelievable world builder.

我们有一段时间没聊世界构建了，对吧。

It's been a while since we've talked about world building in general Mhmm.

因为我们已经有一阵子没拿到谷歌的新Gemma模型了。

Because we haven't really gotten one of the new Gemma models from Google in a while.

但这似乎把谷歌正在做的事情彻底颠覆了，效果令人难以置信。

But this seems to take what Google is working on and kind of flip it on its head in a way that is unbelievably impressive.

它只需要一张静态图片，你可以在视频里看到，它就能从一张静态图片生成整个房间，让你走进去，实际探索。

It requires one still image, and you could see in the video here, it will just generate an entire room from a still image and let you walk around and actually navigate it.

所以从数据训练的角度来看，世界模型到底有什么用？

So in terms of data training, I mean, what what are world models good for?

训练机器人。

Training robots.

它们适合娱乐。

They're good for entertainment.

它们适合游戏。

They're good for gaming.

这绝对是在当前可能性的前沿。

This is definitely right on the frontier of what's possible here.

而且，当然，这自然出自李飞飞之手，她是AI之母。

And, of course, naturally, it's by Fei Fei Li, the godmother of AI.

我觉得这个发布真的让我特别兴奋。

And I found this release to be so exciting.

这周出来的所有工具里，我最想花时间玩的就是这个，因为，天哪，创建这些三维虚拟环境多有意思啊。

Of all the tools that came out this week, this is the one I wanna spend the most amount of time with because, man, how much fun is it to create these three d virtual environments?

就像，简直无与伦比。

Like, it's like nothing.

现在变得太容易了，而且效果太惊艳了。

It's so easy now, and it's so impressive.

所以对于那些一直在关注世界模型进展的人来说，李飞飞，就像你提到的，是这个领域的元老。

So for those of you who have been following kind of like the the world model advancements, Fei Fei Li, as you mentioned, is an OG here.

有趣的是，她曾经教过安德烈·卡帕西，在他还是个年轻学生，想学习这些AI知识的时候。

Fun fact, she used to teach Andrei Karpathy, be Andrei Karpathy back when he was just a young student trying to learn about all this kind of AI stuff.

所以李飞飞在这个领域已经很久了，她最近创办了自己的公司叫World Labs，正如名字所示，专注于世界模型。

So Fei Fei Li's been around for a while and she recently started her own company called World Labs, which is focused, as the name suggests, on world models.

现在，世界模型和LLM的区别在于，世界模型是一个真正对现实世界有良好理解和空间感知的AI模型。

Now, the difference between a world model and an LLM, a world model is an AI model that really has a good understanding and spatial awareness of the actual world.

大语言模型就像图书馆里的学生。

LLMs are like kind of like a student in a library.

它读遍了所有的书。

It reads all the books.

它理解世界运作的理论。

It understands the theory of how the world works.

世界模型则真正走进现实世界，理解物理规律是如何运作的。

World models actually go out into the world and understand how the physics of the world works.

比如我现在屏幕上展示的这根绳子，如果我松手，它会怎么掉落、怎么散开，背后的物理原理是什么。

If I drop this string that I'm showing you on my screen right now, what happens, how it splatters the physics behind all of it.

现在我们看到的这个模型的进步，我想给大家展示这段视频。

Now, the advancement that we're seeing with this particular model, and I wanted to show just this clip.

这些演示

The demos

也太疯狂了。

are crazy too.

我一直在看那些演示。

I've been watching demos.

它们太棒了。

They're sick.

所以，如果你现在没在看屏幕，我建议你去看一下。

So this is someone, if you're not, if you're listening to this and you're not watching your screen, I advise you do.

这是有人打翻了饮料。

This is someone spilling a drink.

顺便说一句，这只是一个角度拍摄的视频。

And by the way, this is just one video shot from one angle.

这个人已经能够以各种方式绕着他转，包括捕捉他的脸、他的胡茬、草莓上的光线角度、颜色，所有的一切。

And so this person has been able to navigate and go around him in every single shape and form, including capturing his face, his beard trimmings, the angle of the lighting on the strawberries, the coloring, everything.

这看起来像是有人绕着拍的视频，但实际上只是从一个视角，有人拍了一个视频，现在你可以绕着它转来转去。

This looks like a video that was shot by someone walking around, but really it's just kind of like from one plane of view, someone just shot one video and now you're able to go round and round and round.

你可以让它停下来。

You're able to stop it.

你可以编辑它。

You're able to edit it.

你可以放大。

You're able to zoom in.

这个工具有很多不同的玩法。

There's so many different ways that you can kind of play around with this tool.

这在好莱坞的视觉特效领域可能会产生影响，顺便说一句，通常要花几千万美元和数小时，抱歉，数月才能重现这个效果。太疯狂了。

This could have effects in VFX Hollywood, which by the way, typically takes tens of millions of dollars to recreate this and hours and sorry, months of It's so crazy.

我们习惯了花几块钱和几秒钟就能搞定。

We're used to like kind of like a couple dollars and like a couple seconds.

简直难以置信。

It's just insane.

那么，这个模型相比起来怎么样？

Now, what's the comparison to this model?

嗯，它来自谷歌。

Well, it's from Google.

我记得谷歌之前发布了一个叫Genie 3的模型。

Google, I believe it was Genie three that they released a while ago.

那是他们的世界模型。

It was their world model.

效果还行，但看起来有点像虚幻游戏引擎。

It was okay, but it kind of looked like the Unreal game engine.

看起来有点动画感。

It kind of looked animated.

这里的主要区别是，你可以拍一张真实世界的照片，比如一个人，然后把它变成一个模拟现实，每个像素都有三维点。

The main difference here is you can take a picture of the real world, like a human being, and you can turn it into a simulated reality with a three d point for every single pixel.

想象一下这应用到体育上。

And so imagine this applied to like sports.

想象一下这应用到电影上。

Imagine this applied to, like, movies.

这可能会改变我们与媒体互动的方式，而且这真是个很酷的进展。

This could, like, change the way we interact with media, and it's just such a cool development.

这也挺酷的。

It's also just cool.

我们还有一个篮球的演示。

There's another demo we have with basketball.

比如，这个人只是在投篮，你可以看到有几个摄像头。

Like, this dude was just taking he you could see there's, a couple cameras.

基本上有三个不同的摄像头。

There's three different cameras, basically.

从这三个摄像头的角度，它能够重建整个三维结构，并且能推测出它看不到的部分。

And from the three camera angles, it's able to rebuild the entire three-dimensional structure of this and guess the things that it can't see.

所以你会注意到摄像头都对着篮球筐。

So you'll notice the cameras are all facing the basketball hoop.

它仍然能生成篮球筐后面的东西，因为它足够聪明，能理解并判断那应该是什么样子，然后生成出来。

It still is able to generate what's behind the basketball hoop because it's smart enough to understand and interpret what it should look like and then generate that.

所以只需要几个摄像头的角度，我很喜欢那个描述，说这简直就像黑魔法一样。我们还看到另一个例子，如果你只有一个摄像头角度，效果也不错，但准确度会差一些，也就是说，它呈现出来的现实不那么真实。

So all it takes is a few camera angles, and I love the description where it just says, like, this is kind of indistinguishable from black magic, and we're seeing another one where if you have just if you have a single camera angle, it works pretty well, but it is less accurate in a sense that it is less truthful of what reality looks like.

如果你有几个摄像头，它就会把你给它的任何东西拼接起来，变成一个令人难以置信的世界。

If you have a couple cameras, it will just take whatever you give it and stitch it together and turn it into this unbelievable world.

这真的太令人印象深刻了。

It is so so impressive.

所以我觉得这周发布的所有东西里，这个是我最喜欢的。

So I think of all the releases this week, this one's my favorite.

这是我想让大家去尝试的。

This is what I want everyone to try.

这是我最期待去尝试的。

This is what I'm most excited to try.

我想出去录一些视频，比如就放几个iPhone在那，看看能搞出什么疯狂的东西。

I wanna go out and record some videos, like just post up a couple iPhones and see what type of crazy stuff I can get.

也许我们下周在节目上分享结果。

Maybe we'll share the results next week on the show.

不过风水轮流转啊，Josh。

How the tables have turned though, Josh.

我们以前总是先聊Anthropic的新模型，比如一上来就讲最新的，但Claude，或者说Anthropic，这周发布了两个新模型，不是一个。

We typically always used to talk about like the new Anthropic models, like released, like the first thing on the docket, but Claude or rather Anthropic released not one, but two new models this week.

Claude Fable 5.1和Claude Mitos 5.1。

Claude Fable 5.1 and Claude Mitos 5.1.

据我了解，这是一个迭代升级。

This is an iterative upgrade to my understanding.

我一直在玩这些模型。

I've been playing around with these models.

我只有Fable 5.1的访问权限，就是那个有安全措施的版本。

I've only got access to a Fable 5.1, which is the one with the safeguards on it.

我知道这里有很多不同的变化。

And I understand that there's a bunch of different changes here.

首先，性价比高了很多。

So number one, it is much better bang for buck.

你可以用更少的成本获得同样的智能，甚至更高的智能。

So you can get the same amount of intelligence or increased intelligence for a fraction of the cost.

他们搞了个叫缓存的东西，或者说减少了模型用的缓存机制，这样你就能在提示上省下高达75%的成本，尤其是在长时间代理工作负载上。

They're doing this thing called caching or cutting down on sort of like the caching mechanism that the model uses, which means that you could get up to 75% lower costs on your prompting, especially over long agentic workloads.

所以任务越长，这个模型通常表现越好，而且便宜得多。

So the longer your task is, typically this model is very good and much, much cheaper.

第二点是，它在视频生成方面真的很厉害，我这么说都觉得疯狂，因为你现在屏幕上看到的，这个来自Fable 5.1的演示，完全是代码生成的。

The second thing is it's really compelling when it comes to video generation, which is crazy for me to say, given that what you're seeing on the screen right now, which is a demo from Fable 5.1, is completely generated from code.

这没有用到任何输入图像。

This didn't take any kind of input images.

这完全是从零开始用代码生成的。

This was just generated from code from scratch.

看到这个导航效果，真是让人惊叹，对吧？

And it is just crazy to see the navigation, right?

这几乎就像我在看我们刚才讨论的那个世界模型一样。

It's almost like I'm looking at that world model that we just literally like kind of like went through.

所以它更便宜，也更高效。

So it's cheaper, it's more efficient.

然后我还要说的是，关于安全防护这块。

And then the other thing I'll say about this is the safeguards.

对我来说，Fable 5 有个特别烦人的地方，至少我个人体验的时候，就是它话太多了，而且不太回答我问的问题。

One thing that was really annoying for me for Fable five, at least personally, when I was experiencing it is it just spoke too much and it didn't kind of answer the questions that I gave it.

这个模型已经有了很大的飞跃。

This model has been a significant leap.

现在它能回答我的问题了。

It now answers my questions.

它不再说那么多废话，整体体验舒服多了。

It stops saying too many words, and it's all overall a much more pleasant experience to have.

对。

Yeah.

这真是太棒了。

It's, it's amazing.

尤其是这个演示，特别酷。

This demo, in particular, is very cool.

写这个的人说，这次他给了一张地块的图片，然后它设计了一栋房子，渲染出来，还生成了一个电影般的漫游视频。

The person who wrote it says, for this one, I gave it a picture of a property lot, and it designed a house for the lot, rendered it, and then produced a cinematic walk through.

还有一点很重要，Fable 其实并没有直接生成图像，而是和 Blender 合作，Blender 是一个三维渲染软件。

And it's important to note also that Fable didn't actually generate the images, but Fable worked with Blender, which is a three d rendering software.

它以无头模式运行，两者协同工作，Fable 负责指挥和编排，最终构建出了这个三维虚拟空间。

It was headless, and it worked together, and it conducted it and orchestrated it to build this three d virtual space.

我觉得这是最厉害的地方之一，因为模型本身没有图像生成能力，却能创造出这么惊人的视觉效果。

And I think that's one of the more powerful things is because the model, it doesn't have an image model, and yet it's able to create these unbelievable visual results.

第二个例子是，这里还有一个一次性提示生成马里奥赛车游戏的例子，Fable 5.1 把单一提示变成了一个完整可玩的马里奥赛车游戏。

The second one being, we have another example here of a one shot prompt for a Mario Kart game, and Vable 5.1 turned a singular prompt into a fully working Mario Kart game.

我觉得这个也特别厉害。

And I found this to be really impressive too.

就是说，好吧。

It's like, okay.

虽然它没有图像生成模型，但在我看来这已经相当强了。

Well, it doesn't have an image generation model, but this is looking pretty strong to me.

说实话，要是几年前你把这个游戏给我，放在我的掌机上，这画面看起来跟任天堂的《Frontier》一样好。

In fact, if you gave me this game, like, a couple years ago on my handheld device, this looks as good as Frontier in terms of, a Nintendo game.

所以我觉得这些模型的能力，以及它们能跟其他现有工具互动、利用它们、指挥它们来得到你想要的结果，这种智能真的非常厉害。

So I think the capabilities of these models and the intelligence to be able to interact with other tools that exist and harness them and orchestrate them to get the results that you're looking for It's a really impressive capability.

好模型。

Great model.

新模型真棒。

Great new model.

这算是新的前沿吗？

Is this is like the new frontier?

我们看看Astra会带来什么。

We'll see what Astra has to, like, come back with.

我非常期待。

I'm very excited.

看起来Astra可能会在一定程度上超越这个，但就目前而言，这是最好的模型，真的令人印象深刻，用起来非常享受。

It seems as if Astra might leapfrog this to some extent, but for now, best model, really impressive, really enjoying using it.

好的。

Okay.

那么，简单回顾一下我们刚才聊的所有内容，帮你梳理一下整体情况。OpenAI 的 Astra 要发布了。

So just to give the run of the mill of everything that we've just covered, just to just to catch you up on on the roundup itself, we have OpenAI Astra coming out.

你现在可能已经能用上了。

You might likely have access to it right now.

谷歌发布了 Gemini 3.8 Flash，价格超便宜，效果也超好，不过它要跟 Meta 的新模型 Muse Spark 1.3 竞争。

You've got Google that released Gemini 3.8 Flash super cheap, super effective, but it's competing with Meta's new model, which is Muse Spark 1.3.

Anthropic 也推出了两个模型，Fable 5.1 和 Meetos 5.1。

You've got Anthropic releasing two models, Fable 5.1 and Meetos 5.1.

另外还有几个中国的开源模型，以及这个新的 World Labs 世界模型。

And then you've got a few Chinese open source models as well, as well as this new World Labs world model.

我想说的是，现在已经不是几个月才出一个模型、每次都是重大能力飞跃的时代了。

So the point that I wanna make here is we have moved very far away from one model per couple of months, and it's like a major capability leaf.

我们现在看到的是这种迭代式的软件升级，这就是我们生活的新常态。

We're now having this kind of iterative software upgrades, and that's kind of like the new world that we live in.

总的来说，这对你——无论是听众还是用户——来说都非常棒，因为这意味着你可以不断升级和尝试不同的视角。

And it's net net so amazing for you, the listener or the user or the user for these different things, because it means that you can constantly kind of upgrade and test different kinds of views.

而且你不再受制于某一家AI实验室，我觉得这真的是一个很棒的时代。

And you're not beholden to one particular AI lab provider, which I think is just an amazing world to be in.

接下来，Josh，你在节目开始前发给我的那个，就在

So moving on, Josh, you sent me this before we started the show, right before

那段视频之前。

the clip.

天哪，我一想到这个就头疼。

Oh my god, I have a headache just thinking about it.

为什么

Why

我又看到亚历山大·斯卡斯加德了？

am I looking at Alexander Skarsgard again?

而且为什么我非得看这副破不锈钢耳机？

And why am I looking at this damn stainless steel earbuds at all?

听这个节目有一阵子了，你知道这里面的梗。

Been listening to the show for a while, you know the lore here.

这是个特别重要的话题，基本上一直是我心头的一块大石头，因为我们看到的这个，简单说，这个视频是作为泄露的超级碗广告发布的。

This is a really important topic that has been at the core of my like, the the bane of my existence essentially, because what we saw here, just quick TLDR, this video was released as a leaked Super Bowl commercial.

它是在超级碗期间出来的，在Reddit上被泄露，之后很快就被删了，展示的是一个传闻中的设备。

It came out during the Super Bowl, was leaked on Reddit, was deleted shortly afterwards, and it shows a rumored device.

没人知道那是什么，但视频结尾出现了OpenAI的标志，还写着“dime”。

No one knows what it was, but at the conclusion of it, it showed the OpenAI logo, and it said, dime.

这是个著名演员，而且这个设备看起来真的很漂亮。

And this is a famous actor, and this is a really beautiful looking device.

再说一遍，Josh。

Say it again, Josh.

说出他的名字。

Say his name.

亚历山大·斯坎斯加德？

Alexander Skanskansgard?

我真是的，这个破名字我永远都念不对。

I have no I can never get this freaking name right here.

我完全不知道。

I have no idea.

抱歉，Josh。

I'm sorry, Josh.

差不多快一年了。

It's almost been it's almost been a year.

我也不确定。

I'm not sure.

我学到了。

I've learned.

对。

Yeah.

我跟Josh说过，

And I've told Josh,

不，那其实根本不是那个设备。

no, that's not actually the device at all.

我们不知道那是什么。

We don't know what that is.

几个月后，Airbnb的联合创始人乔·格比亚，他现在正领导整个美国政府的设计工作，这工作真是了不起，传奇级别。

Fast forward a couple months later, Joe Gebbia, cofounder of Airbnb, he's currently leading design for the United States government as a whole, like, really awesome legendary job.

有人在一家咖啡馆里发现了他，那张照片看起来像是精心安排的，就显示他随意地坐在咖啡馆里。

He's spotted in a cafe, and it's like the seemingly staged post that shows him just casually at a cafe.

但你猜怎么着？

But what do you know?

耳机和那个小圆盘也在画面里，而且他用的正是我们在那个被删除的视频里看到的设备。

The earbuds and the puck are inside of the frame as well, and he's using the device that we saw in that deleted video.

好吧。

Okay.

这到底是什么东西？

What is this thing?

第三个线索就发生在昨天。

Third data point just happened yesterday.

帕尔默·拉奇，安多里尔的创始人，Oculus的创始人，那个搞3D的，搞虚拟现实的那位。

Palmer Lucky, founder of Andoril, founder of Oculus, the three d guy, the virtual reality guy.

对。

Yeah.

他发了一张花店的照片，据说他经常去那家花店。

He's posting a photo from a flower shop that he allegedly shops at all the time.

花店随便发了个帖子，哎，他耳朵上戴的是什么？

Casual flower shop shouting him out, oh, what is that on his ears?

同一个设备，手里拿的还是那个小圆盘。

The same device, the same buck in his hand.

如果你去看那家花店的账号，稍微往下翻一翻，就能看到Sweet Pea Flowers，而Sweet Pea正是OpenAI设备的代号。

And if you go to the name of the flower shop account, if you scroll just a little bit, it's Sweet Pea Flowers, and Sweet Pea is the code name for the OpenAI device.

没错。

Yep.