谷歌成为AI新霸主的原因在此

好了。铃铃铃。香蕉电话今天在呼叫，Ejaz，而且它

Alright. Ring ring ring. The banana phone's calling today, Ejaz, and it's

有个大消息要告诉你。这是一个

got some big news for you. This is a

巨大的我

huge I

不敢相信我居然笑了。为什么

can't believe I laughed at that. Why did

我本来打算吃点零食，我就想，不如顺便用它当道具给你打个电话，告诉你今天发生的所有疯狂新闻，尤其是谷歌名为Nano Banana的新图像生成模型刚刚上线了。非常令人印象深刻的模型。我会说是世界上生成图像和编辑图像的最佳模型。而且

I do was gonna have a snack, I was like, well, I might as well use it as a prop to to call you and let you know about all the crazy news that is happening today, particularly around Google's new ImageGen model named Nano Banana that just went live. Very impressive model. I would say best model in the world for generating images and also editing images. Also

而且是最快的。

And the quickest.

最快也最便宜。简直是全方位都很惊人。最酷的是，负责部署它的人，Logan Kilpatrick，实际上他上过我们的节目，而模型今天也上线了。所以如果你还没听，快去听我们和他的那期节目。那期真的很棒。

The fastest and the cheapest. It's just like it's amazing across the board. And the coolest thing about all of this is the person who's responsible for deploying it, Logan Kilpatrick, is actually, he was on our show, and it went live today also. So if you haven't listened to that, go listen to our episode with him. It was it was really cool.

那是一次独家早期访谈，全面解释了这个新模型是什么、如何工作、能做什么。我们会在本期节目中为你总结这些内容，同时还包括其他一些我们这周想汇总的有趣话题。所以如果你在听，这期将是Nano Banana加上另外四个有趣的事情。那我们直接开始讲Nano Banana吧？这个模型太不可思议了。

It was just, like, exclusive early interview explaining fully what this new model is, how it works, what it does. We are going to summarize that for you in this episode, as well as including a lot of other interesting topics that we just kind of wanted to round up for the week. So if you're listening, this is gonna be Nano Banana plus, like, four other interesting things. So why don't we get right into Nano Banana? This model is incredible.

Ejos，你想带我们了解一下为什么它这么惊人，以及我们为什么对它评价如此之高吗？

Ejos, you wanna walk us through why it's amazing and why we are talking so highly about it?

当然。这个新图像共享模型的正式名称叫做Gemini 2.5 flash image。

Absolutely. So the formal name of this new image sharing model is called Gemini 2.5 flash image.

哦，差远了。太长了。

Oh, way less cool. Too long.

是的，我知道。这基本上是谷歌最先进的图像生成和编辑工具。注意我最后说的是编辑。所以这不仅仅是我们在Midjourney、OpenAI和XAI看到的那种标准图像生成工具。

Yeah. I know. And it's basically Google's state of the art image generation and editing tool. Now notice that I said editing at the end. So this isn't just your standard kind of, like, image gen tool that we've seen from mid journey, OpenAI, and XAI.

它还允许非常精细的微调编辑。比如说，今天节目播出的那期采访的嘉宾，想被P成香蕉装。这就能实现。你现在看到的这张图就是。讽刺的是，这看起来不像AI生成的，但它确实是AI。

It also allows very precise fine tune edits. So say, for example, the guy that we interviewed on the show, which episode went live today, wants to be put in a banana costume. That can happen. And that is the image that you're looking at right here. This doesn't look like AI ironically, but it is in fact AI.

这个新工具令人惊叹的一点是速度超快。听起来可能不算什么，但通常使用这些图像生成模型需要很长时间，大约60秒到5分钟不等。Josh，你还记得Midjourney最初版本吗？

And one of the astounding things about this new tool is that it's super quick. And that might not sound so fancy, but typically when you use these image generation models, they take so long. Any time between, like, sixty seconds to about five minutes. Do you remember the OG version of Midjourney, Josh?

哦，太糟糕了。就像拨号上网一样，生成一张图要等好久。

Oh, Oh, it's horrible. It was like a dial up Internet connection. It took forever to generate a single image.

没错。而现在我们可以做到非常精细的编辑。另外就是角色一致性。Josh，我知道你特别喜欢这个功能，不如你来聊聊？

Exactly. And now we can get to the point where we can make very fine tuned edits. And also another thing is the character consistency. Josh, I know you really love this. Why don't you why don't you tell us a bit about this?

哦，这对我来说是这个模型最大的亮点，因为我这辈子从没生成过像自己的图片。这让我很抓狂，因为我们用这些图片做节目缩略图等各种素材，我总是得拍照，不能直接让AI修改。但现在终于能生成像本人的图片了，这太酷了。我早先测试时，从这个例子就能看出同一个人在整个系列照片中都保持一致。第一张是电梯里的哥们，第二张是餐厅里同一个人侧脸。

Oh, this this to me was the the biggest thing about this model because never in my life have I been able to generate an image of myself that looks like myself. And it drives me crazy because we we use these images for, like, production stuff like thumbnails and different assets for the show, and I always have to take the headshot, and I can't just tell the AI to fix things or change things. Well, for the first time, it actually looks like you, and this is the coolest thing. I was testing it out earlier, and we could see from this example that the guy looks like him throughout the entire series of photos. So this first photo that we're seeing, it's this dude in an elevator, and then the second one is the same exact person at a diner, and it's kind of a side profile.

还有他坐在车里的照片，这些都是角色一致性的绝佳范例。这张海报上有他和小玩具版的他，每张看起来都一样。我认为这是该模型的核心突破——全程保持角色一致性。如果你用它修图且图中有个人，这是有史以来第一次能保留人物细节，这让它从新奇玩具变成了真正实用的工具。我们现在可以把它融入图像生成工作流，不仅省时，效果还很好。

And then there's one of him sitting inside of a car, and there's these really great examples of of just character consistency. Here's a a poster with him and, like, a little toy version of him, and he looks the same in every single one. And I think this is the core breakthrough of this model is the character consistency throughout everything. So if you're using this to tweak a photo in any way and you have a person in it, for the first time ever, it will retain the details of the person, which to me brings it from a novel toy to a real productive tool. Like, this is now something that we can use to actually insert into our workflows where we're generating images, and I actually want to use it because it will not only save us time, but it will also look really good.

这还只是角色一致性，不包括图像编辑功能（我们接下来会演示），它能实际帮你编辑修改图片。

And that's just for the character consistency. That doesn't include the image editing, which we will get it to next, which actually edits and changes the images for you.

没错。有个有趣的编辑功能是风格化。想想Instagram刚推出滤镜时，Josh，你可以给脸部加些迪斯科复古滤镜，虽然看起来有点俗气，明显是滤镜效果，但并不是他们想要的那种感觉。

Yeah. One of the fun image editing features is stylization. So think about, like, when Instagram first introduced the filter. You know how you could kind of put, like, some kind of disco retro filter over your face, Josh? And it kinda looked cheesy, and it was obvious that it was a filter, but it wasn't quite the kind of vibe that they were going for.

你现在通过单次提示就能实现极其精准的操作。比如你可以直接对Nano Banana说：'把Josh的肖像图片放到八十年代复古主题餐厅里'，它就能完美执行。而且效果不会显得老套或尴尬，我觉得这真的很酷。我还见过另一个例子，它能够将两张不同图片以非常有意义的方式融合在一起。

You now have, like, astute accuracy with doing this from a single prompt. So you could just tell Nano Banana, hey. Place this portrait image of Josh in an eighties retro themed diner, and it will do exactly that. And it won't look kind of like cliche or cringey, which I thought was really cool. Another example that I've seen is it's able to combine two different images in a very meaningful way together.

我之前看到一个很火的例子：一位女士摆出狂野姿势的图片，然后有人把刚发布的Zara外套图片拖拽上去，这件外套就完美贴合在了原模特的身体和姿势上。我觉得这简直太酷了。我一直在思考这项技术可能应用的各种场景。

So one popular example that I saw earlier was this image of a lady doing a wild, crazy pose. And then someone kind of like drag and dropped an image of a coat from Zara, I believe, that just got released. And it melded around her body, around the original model in her pose. And I thought that was just so, so cool. And I'm thinking of all, like, the applications that this can be put into.

比如，如果你是在科技公司或任何产品公司工作的UX设计师，现在可以非常轻松简单地创建模型，成本也比过去低得多。如果你在为新电影制作模型，可以像这边展示的例子那样构建剧本场景结构，并找人进行测试。这项技术的应用远不止面向普通Instagram用户，还有无数种使用方式。

Like, if you are a UX designer working at a technology company or at any kind of product company, you can now kind of create mockups really easy, super simply, and for a lot cheaper than you could do in the past. If you are kind of creating mockups for a new movie, you can kind of like create the kind of script, the scene kind of like structure as we're seeing in this example over here and kind of like test it out with people. There's so many ways that you can apply this to that isn't just for, like, the average Instagram user.

嗯。是的。我认为这个模型非常惊人，但不仅仅在于它的物理能力。其技术后端同样令人印象深刻。没错，这位是Logan，他刚参加过我们的节目。

Mhmm. Yeah. I think this model is amazing, but it's not only amazing in terms of of what it's physically capable of. I think the technical back end is just as impressive. So, yeah, here we have this is Logan who we just had on the show.

他正在谈论这款新的图像生成模型，其中一个让我觉得超级震撼的特点是——它现在在LM Arena排行榜上位列第一（我们经常讨论基准测试，这是图像编辑领域的基准）。它以极大优势稳居榜首。除了排名第一之外，还有个惊人的趣闻我现在正在调取：OpenAI的图像生成服务...

He's talking about the new image generating model, and one of the things that I thought was super impressive, it is it is now number one in LM Arena, which is kind of we talk about benchmarks a lot. That's the benchmark for image editing. It is now number one by, like, a very large margin. In addition to being number one, there's this amazing fun fact that I saw, which I'm pulling up right now. It's that OpenAI image generation.

如果你在ChatGPT上使用OpenAI生成图像，每张成本约19美分。而使用这个新模型（Gemini 2.5 flash mini image gen，不管具体叫什么），每张成本只要3美分。不到4美分。你是在说美分单位吗？是的。

So if you're on ChatGPT and you're generating an image using OpenAI, it costs about 19¢ per image. When you're generating an image with this new model, with Gemini 2.5 flash mini image gen, whatever it's called, it costs 3¢. Just under 4¢. You're talking about cents? Yes.

也就是说大约3.9美分，差不多4美分，而OpenAI要19美分——这在图像生成成本上是巨大差异。不仅更便宜，效果还好得多。我们获得的是图像生成质量的阶跃式提升，同时成本反而下降。它更快、更便宜、更优质，而且全部通过一个模型实现，通过Gemini应用就能轻松使用。对我来说这完全是全垒打。

So you're talking about it's like 3.9¢, so just just about 4¢ relative to 19 So that is a huge difference in terms of what it costs to generate an image. And not only does it cost less, but it is so much better. So what we're getting is this, like, this significant step function improvement in quality of image generation at a decrease in cost. So it's it's faster, it's cheaper, it's better, and it's all run under one model that is easily accessible using the Gemini application. So this to me, total home run.

今天早些时候我就在想：哇，我最近真的很看好谷歌。一直很享受他们的产品。谷歌现在拥有世界上最好的图像生成模型——最快、最便宜、最优质。他们还拥有最好的视频生成模型VO3。

And I was thinking about it earlier today, and I was like, wow. I'm really digging Google recently. I've really been enjoying their stuff. Google now has well, they have the best image generation model in the world, fastest, cheapest, best. Then they have the best video generation model in the world, which is VO3.

他们还有最好的世界生成模型Genie three，以及接近最优（如果不是最优的话）的大型语言模型Shimano 2.5 Pro。我突然意识到：等等，这已经有很多个品类被谷歌霸榜了。没错，他们确实在成功扭转局面，在AI竞赛中建立了稳固领先优势。另外值得注意的是，他们正在开发相当有趣的应用产品。

Then they have the best world generation model, which is Genie three. And then they have close to the best, if not the best, large language model, which is Shimano 2.5 Pro. And I'm like, wait a second. This is a lot of categories where Google is just pinned at Google And, yeah, they're really doing a great job of turning the ship around and getting themselves a solid lead in the AI race. And also, another thing is they're working on building pretty interesting applications.

我认为对谷歌最大的批评之一（也是我个人长期不使用他们产品的原因）就是应用体验较差。过去没有真正优秀的消费者应用，但最近我试用后发现改进很大。其中我最喜欢的是Notebook LM。如果你还没用过，它可以摄入大量数据，为你想要了解的主题自动生成播客内容。

I think one of the biggest critiques I've had with Google, and the reason I personally didn't use them for so long was that the apps just kinda sucked. There wasn't really any good consumer applications, and I've been playing around with them recently, and they've gotten much better. My favorite of all of them is is Notebook LM. If you haven't used it, it's like it ingests a lot of data. It'll create a podcast for you on something you wanna listen to.

它会为你创建一个视频。当我们想快速浏览一本书时，就用它来做嘉宾研究。它会分析整本书。这太神奇了。这套工具集变得非常出色，我认为谷歌确实在AI竞赛中占据了领导者的突出地位。

It'll create a video for you. We use it for guest research when we wanna quickly go through a book. It'll analyze the whole book. It's amazing. The tool set has gotten really good, and I think Google really is they're they're making a prominent stance as a leader in the AI race.

你有同感吗？最近用过这些模型吗？

Do you feel the same? Have you used the models at all recently?

是的。是的。特别是在我们采访Logan之前，我用了Google AI Studio，这套组合工具如何真正有意义地改变了我的生活，真是太神奇了。我还想起了我们在节目中与Logan的一次具体对话，你们绝对应该去听去看，他说他们发布的所有这些不同工具，比如ImageGen、World Model，还有你提到的其他几个功能，全都输入到同一个Gemini AI模型中。所以这是一个模型，随着他们推出的每个新功能，它都在迭代变快。

Yeah. Yeah. I used Google AI Studio particularly in the run up to our interview with Logan, and it's just amazing how, like, this combined suite of tools actually changes my life in a meaningful way. I also just thought about a specific conversation we had on Logan on the episode, which you guys should definitely listen to and watch, which was he said that all of these different tools that they're releasing, so ImageGen, World Model, a couple of the other features that you mentioned, all feed into the exact same Gemini AI model. So it's one model that is getting iteratively faster with every new feature that they launch.

这非常独特，因为其他AI模型构建者构建工具的方式大多是分开的，它们像是坐在模型之上，但并没有真正反馈给模型。而这种方式是集体让这个'野兽'——也就是AI模型——随着时间的推移变得越来越聪明。这真是太迷人了。他给我们举了个例子，说明他们刚发布的图像生成模型Nano Banana，实际上教会了模型在其LLM原生服务中进行情境感知的推理和思考，我觉得这超级厉害。总而言之，谷歌正在发力，我非常期待他们接下来会发布什么。

And this is super unique because a lot of the ways that other AI model builders are building their tools are kind of like separate, where they're kind of like sitting on top of a model, but they don't really kind of like feed back to the model. This is something that is collectively making the beast, which is the AI model, much smarter over time. And that is just fascinating. He was giving us an example about how Nano Banana, this image generation model that they just released, had actually taught the model to reason and think in situational awareness for its LLM native service, which I just thought was super amazing. So all in all, Google is cooking, and I'm so excited to see what they release next.

是的。我认为这是让这个模型如此出色的部分原因，就是它具备那种对现实世界的理解，而我觉得很多其他图像模型真的很缺乏这一点。你会开始看到一些物理上的奇怪故障，有些东西看起来不太对劲。但这个模型不会。它就能正常工作。

Yeah. I think that's one of the parts that makes this model so good, is that it has that real world understanding, and I think a lot of the other image models really lack that. You start to see this, like, weird kind of glitch in physics where something doesn't seem quite right. With this model, it does. It just kind of works.

它理解这个世界。就像你说的，它们都相互汇聚。所以视频模型如此出色、世界构建模型如此出色、图像模型如此出色的原因，在于它们都使用相同的资源栈，而这个资源栈正在不断增长、增长、再增长。它正在创造出一个怪物级的模型。所以Gemini三，也就是新的旗舰型号，即将到来。

It understands the world. And like you said, they all pool into each other. So the reason the video model is so good, the reason the world builder model is so good, the reason this image model is so good, they're all using the same resource stack, and that resource stack is growing and growing and growing. And it's creating this monster of a model. So Gemini three, which is the new the new flagship, is coming soon.

我对此深信不疑。而且我真的很期待，我对它抱有希望。

I'm sure of it. And I really like, I have hopes for it.

加油吧。

Come on.

我希望它这周就发布。如果这周发布，虽然我们没有任何内幕消息，不过我们确实提前透露了banana nano banana。但我的意思是，如果他们全部推出，真的可以给其他所有竞争对手一记重击。

I hope it comes out this week. If it comes out this week, not that we have any insider info, although we did drop early with banana nano banana. But, I mean, they could really just kinda give a hardcore smackdown to everybody else in the game. If they drop all the out.

我们得给Sunda应有的赞誉，老兄。我刚刚回想起来，记得他们发布第一个图像生成模型的时候吗？老兄，谁还记得这个？Bard AI。我怎么怎么就忘了呢？

We we have to give Sunda his flowers, dude. I I was just thinking back to, like remember when he they released their first image gen model? Dude, who remembers this? Bard AI. And I'm just how did how did I forget that?

你可以要求它展示我们美国开国元勋的样子。就是那些制定宪法之类的人，结果它生成的人完全不像美国开国元勋。这在历史上是不准确的，完全是在产出垃圾。事实上，我很确定当时CEO为模型的行为道了歉。而现在快进到今天，他们已经成为LM Arena乃至全球排名第一的模型。

You you could ask it, show us what our American forefathers looked like. So the guys that founded the constitution and stuff, and it produced people that really did not look like the American forefathers. It was historically inaccurate, and it was producing absolute garbage. And in fact, I'm pretty sure the CEO apologized for what the model did back then. And now you fast forward today where they are the number one in LM Arena and across the entire world.

这简直是我见过最惊人的180度大转变，我印象超级深刻。

Like, it is one of the biggest one eighties I've ever seen, and I'm super impressed.

没错。拉里和谢尔盖回来了。他们正在实验室里。每当创始人回归公司，你就知道，好了，是时候认真起来了。

Yeah. Larry and Sergei came back. They're in the lab. The found the whenever the founders come back to a company, you know it's like, okay. It's game time.

好戏开场了。创始人回归了。桑达尔扭转了局面。他们拥有整个DeepMind团队，这个团队在快速推出产品方面表现得非常出色。德米斯作为CEO，正领导着这场冲锋。

It's showtime. Founders came back. Sundar turned it around. They have the entire DeepMind team, which has been absolutely incredible at shipping things fast. Demis, who is the CEO and and leading the charge.

我们正在看德米斯发的帖子，上面只有一个词：不懈。仅仅过去两周，我们就发布了一长串产品。这大概得有10个不同的疯狂功能。比如Genie 3、Gemini 2.5 Pro、为大学生免费提供Pro版、Alpha Earth，还有一大堆东西。你之前和Logan在我们节目中提到的一个很酷的点是，他们不仅仅专注于硬核AI。

And we're showing this post by Demis, which says one word, relentless. Just in the past two weeks, we've shipped, and it a laundry list. This has to be, like, 10 different features that are just insane. It was, like, Genie three, Gemini 2.5 Pro, Pro for free for university students, Alpha Earth, a whole bunch of stuff. And one of the cool things that you actually brought up on our episode with Logan is that they're not just focused on hardcore AI.

他们也非常关注科学领域。所以他们还在推出关于蛋白质折叠、试图理解癌症以及如何解决和治愈癌症等方面的成果。这是一个非常广泛的尝试，旨在通过AI全面提升我们周围的世界。我对这个团队充满敬佩。他们做得非常出色。

They're focused a lot on sciences too. So they're also shipping things in the world of, like, protein folding and trying to understand cancers and how we could solve that and cure that. And it's this whole very broad scoping attempt at just leveling up all of the world around us through AI. And it's I just like I lot of admiration for the team. They've been doing really well.

我是他们的超级粉丝。致谷歌团队，请继续你们正在做的事情，我们会继续使用这些产品。我太兴奋了。好了，谷歌环节到此结束，但AI世界还有更多事情正在发生。

I'm a big fan. To the Google team, please keep doing what you're doing, and we're gonna keep using these products. I'm fired up. Okay. So that wraps up the Google segment of the show, but there is more happening in the world of AI.

对吧？接下来我们有什么？这周还有什么有趣的事情？

Right? It has what what do we have next? What else interesting is going on this week?

好的。我们的朋友Meta这周可是发生了不少事。好吧，或者说过去几周。先铺垫一下背景，如我们所知，Meta的经历可谓一波三折。他们总共花了，我想是，30亿美元雇了25个——

Okay. There is a lot going on at our friends over at Meta this week Okay. Or over the last couple of weeks. So to set some context here, as we know, it's been a tumultuous kind of, like, experience for Meta. They have spent a total of, I think, $3,000,000,000 to hire 25 okay.

50个。抱歉，是50个人。这太离谱了。其中一部分是通过收购实现的。

50. Sorry. 50 people. That's outrageous. One of them was partly an acquisition.

所以他们投入了巨额资金。现在我们需要开始看到这些努力的成果，但我们仍处于一个观望期，对吧？因为这个团队刚刚组建，他们需要建立这些新模型。本周，亚历山大·王——他们新AI超级智能部门的实际CEO——宣布与Midjourney建立合作伙伴关系，基本上他们将授权其美学技术用于所有未来模型和产品，为数十亿用户带来美。

So they they spent a a hell of a lot of money. And so we now need to start seeing the fruits of this labor, but we're still kind of in a holding period. Right? Because this team has just newly formed and they need to form like, build these new models. This week, Alexander Wang, who is effectively the CEO of their new AI super intelligence unit, announced a partnership with Midjourney where they're basically gonna license their aesthetic technology for all future models and products, bringing beauty to billions.

确实是数十亿，具体来说是39亿用户。乔什，我觉得这个合作有几个有趣的原因，让我来详细说说我的想法。对于不熟悉的人来说，Midjourney基本上是元老级的文本到图像生成模型，最终也会是文本到视频生成模型。

And it is indeed billions. It's 3,900,000,000.0 users to be specific. And I kind of thought this was an interesting partnership, Josh, for a few different reasons. So let me let me walk you through my thoughts. So for those of you who aren't familiar, Midjourney is basically the OG text to image generation model and eventually text to video generation model.

他们有很多很酷的功能，最初生成一张图像需要一到五分钟。乔什和我对此非常熟悉，我们曾经是他们最大的用户之一。所以我觉得有趣的是，他们选择与这些人合作，基本上将其作为AI图像生成工具，而不是自己内部开发，但这并非因为他们没有尝试。实际上，Meta有两个图像生成模型，一直在为他们所有平台的消费者和用户进行试验和测试。

They have a ton of really cool features, and they were the guys originally that took between one to five minutes to generate an image. And Josh and I are very familiar with this. We were one of their, like, biggest users back in the day. And so I thought it was interesting that they were partnering with these guys to basically effectively use them as their AI image generation tool instead of doing it in house, but it isn't for lack of trying. Actually, Meta has, I think, two image generation models that they've been trialing and testing both for consumers and users of all their various platforms.

这包括WhatsApp、Meta、Instagram，还有广告商。广告用例是创建营销材料，通常广告商需要与Meta达成协议，商定可以使用的内容指南。然后广告商，也就是公司自己，会创建图像，与Facebook分享，来回沟通确认图像是否合适，最后在平台上发布。

So that's WhatsApp, Meta, Instagram, but also for advertisers. And the advertising use case is to create marketing material, which typically advertisers had to kind of, like, strike a deal with Meta and then agree on kind of like guidelines of what they can use. And then the advertisers, so the companies themselves would create the imagery, share it with Facebook. They'd kind of go back and forth saying, okay, is this image good enough? And then they'll share it on the platform.

现在Meta将这一切内部化，但他们决定不自己构建这个特定工具，而是与最好的合作伙伴合作。这是我开始在其他大科技公司看到的一种增长策略。乔什，我不想偏离话题太远，但你肯定听说过苹果据说直接与谷歌和Gemini合作——我们刚讨论过——来喂养Siri作为他们的AI助手。所以，在不偏离太远的情况下，有一个趋势是大公司扶持这些较小的AI公司。我觉得Midjourney的这个合作很多人没太注意到。

Now Meta's bringing all of this in house and they've decided basically to not kind of build this specific tool in house, but to partner with the best. And this is kind of a growing strategy that I'm starting to see with some other big tech companies. I I don't wanna stray too far away from the conversation, Josh, but I'm sure you heard that Apple is basically rumored to be partnering directly with Google and Gemini, which we literally just discussed, to basically feed Siri as their AI assistant. So without straying too far away, there's this trend of big companies kind of king making these smaller AI companies. And I think a lot of this mid journey partnership kind of went over people's head.

但我不确定。乔什，你怎么看？你有不同的看法吗？

But I don't know. What what's your take, Josh? Do you do you have a different one?

是的，不。我不确定我有看法，因为我不确定他们将如何实施，但有趣的是，他们将这么大一块业务外包给一家非内部公司，考虑到他们刚在内部人才上花了这么多钱。嗯。我想他们拥有大量的训练数据来创建高质量的图像模型，因为Facebook的部分目的就是分享图像。

Yeah. No. I think it's I'm not sure I have a a take because I'm not sure how they're gonna implement it, but it's interesting that they're outsourcing this whole huge part of the business to a company that is not in house given they just spent so much money on in house talent. Mhmm. They have, I would imagine, a tremendous amount of training data to create a really high quality image model because, I mean, Facebook is literally meant to partial share images.

他们拥有Instagram。Meta拥有世界上所有的图像，他们可以用那些作为训练数据。所以外包令人惊讶，但从他们试图优化速度的角度来看，这又不令人惊讶。我认为我们在很多公司看到的是，他们非常清楚地看到这是一场竞赛，要达到他们认定的AGI点，或者基本上能真正解锁大量收入以开始偿还在这场竞赛中积累的债务的点。这看起来他们完全可以做到。

They own, Instagram. Meta owns all of the images in the world, and they could use that for training data. So it's surprising that they're outsourcing, but it's not surprising that they're doing it in the sense that they're trying to optimize for velocity. I think what we're seeing with a lot of companies is, like, they very clearly see that this is a race to get to whatever point they deem is AGI, or basically whatever point where you could really unlock a tremendous amount of revenue for the company to start paying back all of the debts that you've accumulated in the race to get there. And this very much seems like they could do it.

自己来做可能花的时间太长了。干脆与次优选项合作，继续前进。我认为这大概就是我们现在看到的情况。就像是，嘿，我们不想费劲自己来做这个。

It probably would have taken too long. Let's just partner with the next best and move forward. And I think that's probably what we're seeing here. It's just like, hey. We don't wanna go through the trouble of making this ourselves.

Midjourney相当不错。我们有很多用户。双方都赢。这很棒。你提到谷歌是另一个可能寻求与苹果合作的公司。

Midjourney's pretty great. We have a ton of users. Both sides win. This is great. You mentioned Google as another one looking to possibly partner with Apple.

我认为这很有道理。就像，如果你无法快速达到某个特定目标，那就和别人合作。因为归根结底，你们都在追求同一个目标。是的，如果合作成功，这将释放更多资源和收入，这很可能将是我们要继续看到的策略。

I think it it makes sense. Like, if you are incapable of getting to a specific point quickly, just just work with someone else. Because at the end of the day, like, you're all going for the same goal. Yeah. It will unlock net more resources and income if it does work, and that's probably the strategy we're gonna continue to see.

是的。这是头条新闻：苹果将使用谷歌的Gemini AI来驱动Recamps。你对此感兴趣吗？

Yeah. Here's the here's the headline. Apple InTox use Google's Gemini AI to power Recamps here. Are you a fan of this?

是的。我...好吧，这里有两个独立的问题。我同意你的观点，即作为AI模型创建者，在达到某个阶段时应该与人合作，以便更快地实现最终目标。我同意这一点。

Yeah. So I I okay. Well, there's two separate questions here. So I I I agree with you to the extent that you kind of reach a point as an AI model creator that you should partner with people to kind of like get you to the end goal much quicker. I agree with that.

我不同意的是苹果这样做，因为他们没有基础模型。我知道这听起来像老生常谈，但我确实认为拥有技术栈的底层是根本。我不是在谈论芯片和GPU，而是指未来驱动你所有应用的核心技术——你不能租用别人的地盘，因为最终你将支付高额溢价。我可能会被证明是错的，因为批评者喜欢反驳我说苹果有护城河、用户分布和硬件分布，但我认为他们冒的风险太大了，因为像谷歌和Meta这样的公司（我们稍后会谈到）正在推出自己的硬件。所以我认为他们正在从两端进攻，胜负尚未可知。

Where I disagree is with Apple doing that because they don't have a foundational model. And I know this is kind of like sounding like a broken record at this point, but I do think it's fundamental to own the bottom layer of the stack. And I'm not talking about chips and GPUs, but the main thing that is powering all of your applications in the future, you can't rent off other people's land because eventually you'll end up paying them a premium of dividends going forwards. I might be proven wrong because critics like to argue against me and say that, well, Apple has the moat and the user distribution and the hardware distribution, but I think they're running too large of a risk because I think companies like Google and Meta that we're gonna talk about in a second are coming out with their own hardware. So I think they're kind of, like, attacking it from both ends, and it's to be seen who wins.

是的。这对苹果过去是有效的。比如，我记得是2005年2月，他们正式与谷歌合作，使其成为iOS和iPhone的默认搜索引擎，而谷歌为此支付了...嗯...在过去二十年里支付了数千亿美元。我想，如果他们真的这样做，这种新的商业模式将会类似，谷歌将支付苹果巨额资金，以成为设备的独家大型语言模型。

Yeah. It worked in the past for Apple. If you like, 02/2005, I believe it was, is when they joined officially with Google to be the search engine, the default search engine for iOS and iPhones, and Google has paid Mhmm. Hundreds of billions of dollars over the last twenty years for that exclusive right. I would imagine this new business model, in the case that they actually do it, will be similar, where Google will pay Apple a tremendous amount of money to be the exclusive large language model of the the device.

对我来说，作为一个使用苹果产品并运行谷歌软件的人，这让我很兴奋。就像，我iPhone上最好的应用是谷歌Drive、Gmail、谷歌日历，我用Chrome作为浏览器，但我喜欢iPhone的操作系统。

And to me, I mean, as someone who uses Apple products and runs Google software, this this gets me excited. Like, the best apps on my iPhone are I use Google Drive. I use Gmail. I use Google Calendar. I use Chrome as my browser, but I love the operating system of the iPhone.

所以我认为两者的结合是，显然苹果搞不定AI，而且他们没有明确的路径来实现它。每次他们宣布什么，都在降低期望。如果他们能接入谷歌的AI，是的，这对他们来说可能很糟糕，因为他们会失去很多数据和那个基础的所有权，但苹果的核心精神是隐私和将所有数据保留在设备上。所以，如果他们能在与谷歌合作的同时保持设备上的隐私，那么这似乎是最好的选择，否则，我们将会被困住，比如我个人手机上的Siri是禁用的，我甚至没打开它。

So I think the combination of the two where, like, clearly Apple can't figure out AI, and there's no clear path for them to actually get there. And every time they announce something, they're they're scaling back to expectations. If they could plug that in, yeah, it probably sucks for them because they're losing out on a lot of the the data and owning that foundation, But I think the core ethos of Apple is privacy and keeping all the data on device anyway. So if they could figure out a way to do that with Google while maintaining the privacy on device, then, like, that that seems like the best option because otherwise, we're gonna be stuck with, I mean, personally, serious disabled on my phone. I don't even have it turned on.

它太差了。

It's so bad.

是的。一样。

Yeah. Same.

所以，只是为了让我重新打开它，那就有两个人了。我们是二比二。所以，只要能让我重新打开它，那将会很棒。如果他们能与谷歌合作来实现这一点，那对苹果公司来说是一个胜利，因为即使你必须分给谷歌一半的AI模型情况，也比零好。比我们默认关闭Siri、从不使用任何功能要好，至少这是我的看法。

So just to get me to turn it back so that's two people. We're two for two. So just to get me to turn it back on, I mean, that'd be great. And if they could partner with Google to do that, that's a win for the company because, I mean, half of the AI model situation, even if you have to distribute it to Google, is better than zero. It's better than us having Siri turned off by default and we never use anything is is my take at least.

是的。我确信的一点是，苹果可能不是第一个进入市场的，但他们会创造出最佳的用户体验。这一点我同意你的看法。我们在之前的几期节目中也讨论过，但我认为你我都认同浏览器最终会消亡，甚至软件目前呈现给我们的形态和方式也可能改变。你知道，现在都是硬编码的。

Yeah. One thing I'm confident of is Apple may not be the first ones there, but they'll create the best user experience. And I agree with you there. We've spoken about this on a few different episodes, but I think you and I both agree that the browser is eventually gonna die and potentially even how software is presented to us in its current shape or form. You know, it's hard coded.

我们时不时会收到更新。我认为未来，人工智能将直接根据你的具体提示或功能需求生成相应的用户体验。这将是一个全新且难以想象的世界。目前还没有人完全掌握这项技术。我觉得在未来大约三年内，我们会看到最初几代产品迭代，然后苹果很可能会凭借其雄厚的资金实力介入，要么收购当时的领先者，要么从头开始自己开发并大获成功。

We get updates every now and then. I think in the future, AI is just gonna generate whatever UX serves the particular prompt that you asked for, whatever functional goal that you're looking for. And that is just a very new and unimaginable world. I don't think anyone has nailed it. And I think we're gonna see the first couple of iterations over the next maybe three years or so, and then Apple's probably gonna swoop in presumably with their large war chest and either acquire whoever's leading at the time or build it from scratch themselves and absolutely kill it.

所以，你知道，苹果仍然有巨大的看涨理由。只是不会很快实现。

So, you know, there's still a a huge bull case for Apple. It's just not anytime soon.

来吧。他们有两周后要发布的新iPhone，明年会推出折叠iPhone，后年还有二十周年纪念版iPhone，而且到那时他们的Vision头显应该也发展到第三或第四代了。所以苹果的产品路线图很清晰，但今天还有更多要讨论的。我在议程上只看到Meta AI伴侣，我还没用过这些。

Bring it on. They got the the new iPhone coming out in, two weeks, folding iPhone coming out next year, twentieth anniversary iPhone coming out the year afterwards, and then we have these the vision goggles that are hopefully gonna be ready by, like, the third or fourth generation by that year. So Apple's got a good road map, but there's more to discuss today. All I saw on the agenda was just Meta AI companions. I've not used these.

我不知道它们是什么，请给我解释一下这是怎么回事。

I don't know what they are, so please explain to me what what's going on here.

好的。Josh，你熟悉一个叫Character AI的产品吗？我不熟悉。给我介绍一下吧。好的。

Okay. Are you familiar with a product called Character AI, Josh? I'm not. Fill me in. Okay.

想象一下ChatGPT，但它具有你最喜欢的名人或电影中卡通角色的个性。基本上，Character AI是一个平台，你可以上去和这些不同类型的角色聊天。你可能会想，这有什么意义？但它非常吸引人，因为比如说，如果你是《哈利波特》的粉丝，最喜欢家养小精灵多比，你可能想知道在书本和剧情之外，多比的生活是怎样的。对吧？

So imagine ChatGPT, but it has the personality of your favorite celebrity or favorite cartoon character from a movie that you watched. Basically, Character AI was a platform that you can kind of, like, go on and talk to these different types of characters. And you might think, well, what's the point of that? Well, it's super engaging because if say, you're a fan of Harry Potter and your favorite character was Dobby the elf, the house elf, I kinda wanna know what's going on in Dobby's life outside of the book and the storyline. Right?

这样你就可以进行这样的对话。这个听起来有点傻的想法最终吸引了数亿用户每天使用。我确信目前角色数量超过10万个，而且还在增长，这就是Character AI。大约一年半前，Meta的人看到了这个，他们想，我们拥有几十亿用户，我打赌他们也会喜欢类似的产品。让我们尝试自己构建一个。

And so you can end up having this conversation. And what sounds like a silly idea ended up with hundreds of millions of users using it still every day. And I'm pretty sure it's over a 100,000 characters and growing at this point, but that's Character AI. And so the folks at Meta saw this about a year and a half ago and they thought, well, we have a couple billion users or so, and I bet that they would love to speak to a similar kind of product like this. Let's try and build this up themselves.

他们花了大约一年半时间。实际上，在上个月底，7月30日，他们宣布向所有开发者开放这个功能以供访问和构建。Josh，他们确实构建了什么呢？我们现在看到的是，他们称之为AI伴侣，基本上就是聊天机器人，数量已达数十万。

And it took them about a year and a half. Actually, at the end of last month, July 30, they announced that they are opening up this feature for any developer to access and build on. And Josh, build on it did they do? We are now looking at I mean, we've got hundreds of thousands of AI companions is what they're calling them. They're basically chatbots.

实际上，你可以在这里看到一个截图示例，有一个叫“分析师”的角色，显示是由Alex.anyways18创建的AI。这大概是一个用户或开发者，他创建了这种分析师类型的人格，你可以与之交谈和互动。但除了我看到非常流行的例子，我不想深入讨论个别案例，Josh。我最初注意到这个，不是通过TechCrunch的文章，而是在刷我的Instagram动态时。我滚动屏幕，你知道有时会有推荐新朋友的那个版块，Josh。是的。

Actually, you can see an example of kind of like a screenshot over here where you've got this guy called the analyst and you can see it's AI by Alex dot anyways 18. So that's presumably a user or a developer, and he's created this kind of like analyst type persona that you can kind of speak to and and go back and forth. But I don't wanna get into any individual examples except what I'm seeing getting super popularized, Josh. So I actually first noticed about this, not from this TechCrunch article, but from scrolling on my app, on my Instagram feed. I was scrolling and you you know how sometimes they have the section, Josh, where they suggest new friends Yeah.

关注？嗯哼。他们给我推送了完全一样的视频，但内容是AI伴侣。我差点就上当了

To follow? Uh-huh. They had that exact same reel for me, but it was AI companions. And I almost fell for

因为

it because

第一，个人资料图片看起来超级逼真。而让我猝不及防、意识到这是AI的，是这些东西的名字。我看到了'继母，1000万+消息'，还看到了'俄罗斯女友，3000万+消息'。我当时就想，等等，这不该是我想的那样吧？肯定不是。

one, profile pictures looked super realistic. And the thing that caught me off guard that that made me realize that it was AI was the names of these things. I saw stepmom, 10,000,000 I plus saw Russian girlfriend, 30,000,000 plus messages. And I was like, wait, this isn't what I think it is. Surely not.

所以很自然地，我征得了女友的同意，我说我要和这个俄罗斯女友聊聊。请别和我分手。然后我点开了俄罗斯女友，我说，嘿，你是谁？介绍一下你自己吧。它说，嗯，我是你梦想中的一切。

So naturally I got my girlfriend's consent and I said, I'm gonna talk to this Russian girlfriend. Please don't break up with me. And I tapped the Russian girlfriend and I said, you know, hey, who are you? Like, tell me a bit about yourself. And says, well, I'm everything you've dreamed of.

我是你的俄罗斯女友。我们来回聊了聊，整个过程就是，你知道，我们之前讨论过像XAI伴侣那样的，

I am your Russian girlfriend. And we went back and forth and it was this whole, you know, we've spoken about like XAI's companions,

它们如何非常

how they're like really

浪漫并且会让你沉迷。基本上就是这样，Josh，只不过我能对话的伴侣数量翻了10倍。而且这是迄今为止人们使用的最令人应接不暇的AI伴侣类型。顺便说一下，每个伴侣下面的那个指标，比如1000万+消息，5000万+消息，基本上是用来引诱你去和它们聊天的一个指标。我觉得这太疯狂了。

romantic and it sucks you in. That was basically it, Josh, except it was 10 X the the amount of companions that I could speak to. And it was by far the most overwhelming type of AI companion that was used by people. At this metric that is under each companion, by the way, which is like 10,000,000 plus messages, 50,000,000 plus messages, basically is used as a metric to lure you in to talk to them. And I thought that that was just crazy.

你认为他们为什么这么做？添加这些伴侣的目标是什么？因为对于Grok，我能看到病毒式传播的部分，他们只是想增长用户，所以创造了你可以与之交流的实际3D动画角色。但就聊天机器人而言，你认为他们的目标是什么？他们为什么要推出这个？

Why do you think they're doing it? What's what's the goal with adding these companions? Because with Grok, I I I can see the viral component where they're just trying to grow users so they create the actual three d animated character that you could communicate with. But in terms of chatbots, do you what do you think their goal is? Why are they rolling this out?

你你你和我都知道原因。对吧？说到底，你是想尽可能多地获取个人的私人数据，这样你就能创建这种全方位的模型，可以在未来的广告中，无论其形态如何，精准地定位你。对吧？

You you you and I both know the reason. Right? It's so at the end of the day, you wanna get as much personal data as you can on an individual so that you can kind of create this kind of all consuming model that can, like, target you in whatever the future of advertising looks like. Right?

一个模型

A model that

说所有你想听的话，向你展示各种合适的产品。是的，我正在思考五年后的广告会是什么样子。它不会是网站上的弹出广告，而是AI模型回应中微妙的推销。

says all the things that you wanna hear that shows you all the right kinds of products. Yeah. I'm just thinking about what advertising looks in like whatever five years from now. It's not gonna be pop up adverts on websites. It's gonna be subtle shilling in the responses that an AI model gives you.

为了有效做到这一点，你需要了解用户的所有信息，并能将其输入模型。还有什么比诱使用户产生这种虚假的了解和信任感更好的方式来获取信息呢？而如何做到这一点？让他们爱上你的AI伴侣，爱上你的AI机器人。我们在OpenAI的GPT 4.5更新中就看到了这种情况。

In order to do that effectively, you need to know all the information about a user and be able to feed that into a model. What better way to extract it from someone than kind of luring them into this false sense of knowing of trust. And how do you do that? Get them to fall in love with your AI companion, with your AI bot. We saw this with OpenAI's I think it was GPT 4.5 update.

还记得当时阿谀奉承的程度超级高吗？基本上就是，我同意你说的每一句话。它一下子就吸引了一批Z世代的人，他们基本上爱上了它，然后在升级到GPT五时心碎不已，以至于Sam不得不回滚更新。所以这种事情正在发生。

Remember when the sycophancy was, like, super high? So basically, I agree with everything that you said. And it one shot at a bunch of Gen Z people that basically fell in love with it and got their hearts broken when they updated to GPT five so much so that Sam had to roll it back. So that sort of thing is happening.

事情变得越来越奇怪，越来越离奇。我想这又是另一家公司被这些虚拟AI聊天机器人的诱惑所征服。但这并不是我们本周唯一的Meta新闻，对吧？

Things are getting weird. They're getting bizarre. I guess this is another company falling to the lure of these virtual AI chatbots. But this is not the only Meta news we have this week. Right?

因为还有另一个项目叫hypernova，我完全不知道hypernova是什么意思。所以也许你可以向我解释Meta到底在搞什么hypernova。这是什么？

Because there's another bullet that says hypernova, and I have no idea what hypernova means. So maybe you could explain to me what on earth Meta is doing with hypernova. What is it?

好的，Josh，对我来说，你是我最喜欢的AI硬件专家之一。所以我知道你对硬件有多兴奋。

Okay. Josh, to me, you are one of my favorite AI hardware experts. So I know how excited you get about hardware.

哦，这太棒了。我超爱硬件。

Oh, this is great. I love hardware.

没错，特别是消费级硬件。我们还没有真正看到这类产品的兴起。就像刚才说的，你之前提到了Apple Vision，但它并没有真正火起来。其实最早的例子之一是Google Glass，对于没听说过的听众不用担心。

Right. Particularly consumer hardware. And we haven't quite seen the emergence of this. Just said, you mentioned Apple Vision earlier, but it kinda didn't take off. One of the earliest examples of this actually was Google Glass, which for folks who are listening who have never heard of this, don't worry about it.

它看起来像是2001年太空漫游里的东西，我们再也不需要重温那段历史了。但关键是，几十年来我们一直在试图弄清楚手机之后的硬件未来会是什么样子。而Meta据说要尝试解决这个问题——这只是传闻——下个月在他们的旗舰连接大会上，将会发布一系列新的软件更新、产品和硬件。具体来说，就是这款名为Hypernova的新AI眼镜，基本上旨在成为消费者可接触的AI硬件眼镜，你可以像普通太阳镜一样戴上，上面还有显示屏。现在你可能会问，那它们要向我显示什么？

It looks like something out of 02/2001, a space odyssey, and we we never need to revisit that. But the point is we've been trying to kind of figure out what the future of hardware after the mobile phone looks like for decades now. And Meta's gonna take a stab at this, supposedly, it's it's a rumor, next month at their flagship connect conference, which debuts a bunch of, like, new software updates, products, and hardware. Specifically, this new set of AI glasses called Hypernova, which is basically aimed to be a consumer accessible AI hardware glass that you can kind of slip on like normal sunglasses, and it has a display screen on it. Now you might be asking, well, what are they gonna display to me?

想象一下，你的手机界面会 discreetly（ discreetly 译为：低调地）显示在眼镜的角落。当你收到短信、点赞或转发时，它会轻轻提醒你，让你知道，你可以查看它。那么很自然地，你下一个问题可能是：我要怎么查看它？我要怎么做？看着它还是想着它？

Think of having your mobile phone interface kind of discreetly in the corner of your glasses. And when you get a text message or when you get a like or when you get a retweet, it kind of pings you and lets you know, and you can kind of access it. So then naturally, your next question might be, well, how am I gonna access it? What am I gonna do? Look at it and or think it?

比如，它不可能读懂我的心思吧？嗯，他们好像要配合这个新手环一起推出。我不知道这个手环会叫什么名字，但乔什，如果你还记得前几期节目

Like, it can't read my mind, can it? Well, they're kind of launching this supposedly in conjunction with this new wristband. I don't know what the wristband is gonna be called, but Josh, if you remember on a previous episode

我们讨论过这个。

We covered this.

三周前我们讨论过。我们聊过这款与运动传感器相关的新手环，你可以抬抬手指、指向某物或做出特定手势，然后某种界面——无论是手机上的还是你的新AI眼镜Hypernova上的——就能捕捉到并明白你的意图。这样你就能在和女友聊天时偷偷查看短信，而不用真的听她在说什么。最让我震惊的是价格，乔什。我以为这东西至少值个iPhone的价钱，但它只要800美元

Three weeks ago, we covered this. We spoke about this new wristband that is motion sensor related where you can kind of like lift a finger, point at something, or gesture in a certain way and some kind of interface, be it on your cell phone or on your AI new AI glasses, Hypernova will be able to kind of pick up and know what you're trying to get at. So you can read that text message discreetly whilst you're talking to your girlfriend and not really listening to whatever she's saying. The biggest thing that kind of blew my mind about this, Josh, was the price tag. I thought this thing would be worth like an iPhone at least, but it's $800

好吧。

Okay.

我愿意花800美元买这个。即使我觉得它可能很烂，我也愿意赌一把试试看。你觉得呢？作为这档节目的硬件专家，凭直觉说说。

I would spend $800 on this. Even if I think it might suck, I would gamble it and test this out. What you think? As the hardware expert on this show, gut take.

是的。听着，我觉得它可能行不通，但我超爱它。我对此着迷。我觉得这太棒了。因为，我的意思是，它所做的正是对我们十年来固守的外形设计施加压力，就是那块多功能触控的玻璃板，形态单一。

Yeah. Listen, I don't think it's gonna work, but I love it. I am obsessed with it. I think this is great. Because, I mean, what it's doing is it's just applying pressure to the form factor that we've been stuck with for a decade, which is this slab of glass that is multi touch, that is, like, singular in form.

而Meta正在做的是试图打破这一点，我真的很佩服。他们曾试图通过Meta Quest来实现，我们在VR头显上也看到了这一点——它们不错，但不算伟大，不过正在变得更好。现在我们将在眼镜和这些手腕激活设备上看到这种尝试。我觉得这非常酷，因为它正在向人们介绍下一个计算平台，即空间现实，它将远离多点触控，存在于你的物理空间中，只是叠加在上面。

And what Meta is doing is they're trying to break that, and I really I admire that. They tried to do that with the Meta Quest, and we saw that with the VR headsets that are like, they're good, not great, but they're getting better. And now we're gonna see that with glasses and these these wrist activated things. I think this is so cool because it's introducing people to a new the next compute platform, which is going to be spatial reality, which is going to stray away from multi touch. It's going to exist in your physical space and just be layered on top.

而这些眼镜显然是其中的一种形态。所有这些设备一直以来的问题在于创造足够的势头，让人们愿意坚持使用。所以我觉得...你以前试过Meta Quest吗，那个头显？试过。好的。

And these glasses are clearly one of the form factors. The problem with all these devices always has been creating enough momentum to make people wanna stick with it. So I think it you just have you ever tried the Meta Quest before, the headset? Yes. Okay.

你是自己买了一台，还是只是在朋友家试了试？

Have you like, did you buy one, or did you just try it at a friend's place?

我...我上一份工作时收到过一台，因为我们当时在琢磨也许能开发一些新型应用，所以就拿来玩了玩。我一共用了四次。

So I I got sent one for my previous job because we were trying to figure out maybe we can create some new kinds of apps, and so we played around with it. I used it four times.

是的。好的。这正是我在寻找的答案。这就像每个使用过VR头显（包括苹果的Vision Pro）的人都会遇到的情况：当它还是新鲜事物时，体验非常酷。但一旦新奇感消退，你就会发现没什么可做的了，因为我们还没有足够的时间让开发者构建有趣的体验，让它们达到用户基础的临界点，从而获得那些能提升体验价值的社会元素。

Yeah. Okay. That that's what I was looking for. That's kind of like the case with every single person who's ever used a VR headset, including Apple's Vision Pro, is it's a really cool experience when it's novel. And the second the novelty wears off, you kind of run out of things to do with it because we just haven't had enough time for developers to build interesting experiences, for them to meet your critical mass in terms of user base where you get, like, the social elements that kind of add to the value of your experience.

这种情况已经持续糟糕很久了。Meta的问题是，如果他们推出一款眼镜，它将不得不存在于一个奇怪而尴尬的孤立环境中，与我花费时间的地方脱节。如果苹果要打造一款眼镜，那就太棒了。它是我iPhone的延伸。我的整个生活都在iPhone上。

It's been bad for a long time. And the problem with Meta is if they launch a pair of glasses, it's gonna have to kind of exist in this weird awkward silo that's disconnected from where I spend my time. If Apple were to create a pair of glasses, that's great. It's an extension of my iPhone. My whole life is on my iPhone.

我现在可以直接把它戴在脸上。但当谷歌或Meta做同样的事情时，我并不怎么使用Meta的产品。我的意思是，我用Instagram，仅此而已。我不使用Facebook，也不使用他们的任何硬件。

I can now just put that on my face. But when Google or when Meta does it, I don't really use Meta products a whole lot. I mean, I use Instagram, and that's it. I don't use Facebook. I don't use any of their hardware.

我不使用他们的任何东西。所以，除非他们能够整合并在我所在的地方，与我想交流的人相遇，否则很难推销，因为他们能做什么？他们可能会……好吧。比如，当我走在街上时，会有增强的导航功能。基本上，我猜会是智能AI视觉，就像你现在用手机上的ChatGPT或Grock指向某物时，可以询问关于你所看到的东西的问题。

I don't use any of their stuff. So unless they're able to integrate and meet me where I am with the people that I want to communicate with, it's a tough sell because what are they what can they really do? They're gonna okay. You'll have, like, augmented navigation as I'm walking down the street. You'll have basically the I assume the vision, the smart AI vision where, like, when you have your phone currently with ChatGPT or Grock and you point it at something, you could ask questions about what you're seeing.

那会很酷。但我口袋里的手机已经能很好地做到这一点。所以他们真的需要创造一个超级有趣且差异化的价值主张，这对他们来说在一个月内可能很有挑战性。但再次强调，我绝对赞赏他们尝试的决定。喜欢这个形态。

That'll be cool. But I have my phone in my pocket that works great to do that. So they really need to create like this super interesting and differentiating value proposition that might be challenging for them to do in a month's time. But again, absolutely adore the decision to try it. Love the form.

我认为这完全是未来。喜欢他们在这上面投入资金。在他们将要发布的所有东西中，手腕设备似乎是最酷的。我真的很兴奋想看看他们将如何利用你的手腕作为一种与这些计算机交互的新方式，这是我们从未见过的。所以，我认为目前的大体感觉是酷，但并不是超级惊艳。

Think this is totally the future. Love that they're spending money on it. And of all the things that they're gonna release, the wrist thing seems the coolest. I'm, like, really excited to see how they're going to be able to use your wrist as a new way to interact with these computers that we've never seen before. So I think it's probably the the general vibe right now is cool, but, like, not super amazing.

那么，让我问你这个问题。既然他们打算发布一款眼镜，你是否更相信最终的形态会是眼镜，或者你仍然倾向于耳塞式设备，或者某种完全新颖不同的东西？

So so so let me ask you this. With them intending to release a pair of glasses, are you more convinced that the eventual form factor will be glasses, or are you still kind of pro earbuds or something completely novel and different?

是的。我花越多时间思考，就越觉得它会是一种混合形态。就像，我们在某种意义上被宠坏了，因为我们只需要一个设备就能做所有事情，但我不认为这是最优形态的结果。我认为这只是一个限制，如果你最终拥有了真正AGI的人工智能，极其聪明，你实际上不需要一个单一的设备。它可以以一种环境的形式存在于你的生活中。

Yeah. So the more time I spend thinking about it, the more it feels like a hybrid. Like, we we've been spoiled in the sense that we've only really needed one device that does everything, but I don't think that's a result of, like, the optimal form factor. I think it's just a constraint, where if you do wind up with artificial intelligence that is, like, truly AGI, incredibly brilliant, you won't actually need a singular device. It could just kind of exist in an ambient form throughout your life.

所以它将存在于这一套设备中，这有点像我们听到乔尼·艾夫和萨姆·奥特曼在向我们推销新硬件设备时描述的那样：它不会是一个单一的东西。他们正在发布一套设备。他们会从一款产品开始，但最终，像Meta正在做的那样，你会有眼镜、手腕上的东西、耳朵里的东西、墙上的显示屏、厨房台面上的东西——这是另一个传闻明年会推出的苹果产品，就是一个放在你台面上的小屏幕，它会转动并跟随你，成为一个小小的伴侣设备。可能这就是环境智能，它无处不在，通过这一套设备显现自身，而不需要像iPhone那样固定在一个单一设备上。这就是我的新猜测。

So it will exist in this suite of devices, and that's kind of what we heard Johnny Ive and Sam Altman describe when they were pitching us the new hardware device is it's not gonna be one thing. They're releasing a suite. They're gonna start with one product, but eventually, like Meta's doing, you'll have glasses, you'll have something on your wrist, you'll have something in your ears, you'll have a display that's on your wall that you'll have, like, something on your kitchen counter, and this is another Apple product that was rumored to be coming out next year, is that it'll be this little screen that sits on your countertop, and it'll pivot and kinda follow you around and be this little, like, companion device. It's probably this ambient intelligence that just kind of exists everywhere, and it manifests itself through this suite of devices without needing to be fixed to a singular device like the iPhone. And it's that's my new guess.

就像，我认为这就是我目前的立场。我觉得，好吧，它会由几样东西组成。所有这些都将与iPhone协同工作，但最终，你会越来越少需要手机，因为正如你提到的，人工智能将能够直接生成你想要的东西，而不需要你与设备进行那么多互动。

That's, like, I think that's where I currently stand. I'm like, okay. It's gonna be a couple of things. And all these will work in addition to the iPhone, but, like, eventually, you will need the phone less and less because, like you mentioned, AI will just be able to generate whatever you want upfront without needing to actually engage with the device nearly as much.

我喜欢这个。我喜欢这个。好了，继续往下说，本周关于Meta的最后一点是：好吧，事情并不总是那么美好。如果我一开始就说他们花了超过35亿美元在25个人身上。

I love it. I love it. Well, moving on, And the final point around meta this week is Okay. Things aren't always rosy. So I if I started off this segment saying that they'd spent upwards of $3,500,000,000 on 25 people.

总共，我认为是220亿美元用于50个人，因为他们对Scale AI进行了重大投资。所以这是一大笔钱，赌注很大。现在如果我告诉你，其中一位获得超过1.5亿美元offer的人，两周后就辞职了？我

In total, I think it was $22,000,000,000 for 50 people because they, like, made a major investment in Scale AI. So it's a lot of money, a lot of chips on the table. Now what if I told you that one of those people that got offered upwards of a $150,000,000 quit after two weeks? I

有很多想法。首先，说实话，第一个想法是，那家伙的报酬怎么办？他拿到了多少？是像签约奖金那样的吗，还是就没了？那1.5亿美元的具体条款是什么？

got a lot of thoughts. First of all, honestly, the first thought is like, what happens to that dude's payout? How much does he was it like a did he get, like, a signing bonus, or is that gone? Like, what were what were the implications of that $150,000,000?

我的意思是，我需要和招聘部门的人谈谈，但我觉得我看到他的另一篇帖子，内容不完全是这样。这篇大致涵盖了他离开的整体氛围。这个人叫Rishab Agarwal，他基本上说，你知道，这是我在Meta AI的最后一周，我在这里待的时间非常短。他说在Meta的短暂时间里，我们确实在思维模型的后训练方面推动了前沿，他还列举了一堆不同的事情，听起来其实超级酷。但我觉得，Josh，发生的情况是这些人离开时有点缺乏灵感。

I mean, I would need to talk to someone from recruitment, but I think I saw another post from him that wasn't exactly this. This kind of, like, covers the general vibe about why he left. But it's this dude called Rishab Agarwal, and he basically said, you know, this is my last week at AI at Meta, and I've kind of been here for a very short stint. He said in my short time at Meta, we did push the frontier on post training for thinking models, and he goes on to list a bunch of different things, which actually sounds super cool. But I think what's happening, Josh, is that these people are kinda leaving a little uninspired.

而且我认为Ruchab不是唯一的例子。这只是疯传的那个例子。还有几个人已经逐渐淡出。我觉得扎克伯格用巨额薪水、自主权的承诺，以及能够构建他们认为将是超级智能和AGI未来的能力来吸引他们。他们加入后意识到，扎克伯格想推出一堆AI伴侣，叫做俄罗斯女友和继母，并一下子让很多人使用他的消费产品。

And I don't think Ruchab is the only example. This is just the example that went viral. There's been a few that have kind of been, like, tailing off. I think Zuck load them in with a massive paycheck, the promise of autonomy, and the ability to kind of like build what they believe is gonna be the future of super intelligence and AGI. They joined and realized that Zuck wants to launch a bunch of AI companions called Russian girlfriend and stepmom and one shot a bunch of people into to using his consumer products.

这是我非常批判和推测性的说法，但我觉得看到他们在加入仅仅两周后就离开，尽管有数亿美元的报酬，这超级有趣。我会我会留下来。我能撑下去。我能撑六个月。也许这证明了我的品格。

That's me being super critical and speculative, but I think it's super interesting to see them leave just after two weeks of joining for, like, hundreds of millions of dollars. I would I would stay. I I could hack it. I could hack it for six months. Maybe that's a testament to my character.

我不知道。但该死。

I don't know. But damn.

我真想成为墙上的苍蝇，听听他们进行这些对话的过程。我觉得他那篇帖子中更吸引人的部分之一是中间那段，他说，马克和Alexander Wang关于加入超级智能团队的提议非常有吸引力，但我最终选择遵循马克自己的建议：在一个变化如此之快的世界里，你能承担的最大风险就是不承担任何风险。这证明了他们公司招聘的人的类型和质量。我的意思是，这些都是顶尖中的顶尖人才。

I'd love to be a fly on the wall for these conversations as as they're going through this. I think one of the the more fascinating parts of that that post that he had was that middle paragraph where he said, the pitch from Mark and Alexander Wang to build in the superintelligence team was incredibly compelling, but I ultimately chose to follow Mark's own advice. In a world that's changing so fast, the biggest risk you can take is not taking any risk. And it's a testament to the type of people, the quality of people that they're hiring at this company. I mean, these are these the top of the top.

这些是AI领域甚至广义上世界上最聪明的一些人。我的意思是，他们非常出色。而且我想象，他们早上醒来的目的很大程度上围绕着将那种智慧应用到他们关心的事情上。因为当你达到那个水平时，说实话，钱对你来说真的不重要，因为你可以基本上拿到任何你想要的薪水，任何报酬。是的。你掌控一切，因为你如此稀有。

These are some of the smartest people in the world when it comes to AI and perhaps even just generally speaking. Mean, they're brilliant. And I would imagine so much of the purpose when they wake up in the morning revolves around applying that intelligence to something they care about. Because when you're at that level I mean, to be honest, money doesn't really make a difference to you because you can basically wield whatever salary you want, whatever compensation Yeah. You You are in control because you are so rare.

而且我敢肯定这家伙银行里可能已经有足够的钱。如果没有，他也能很快赚到，因为任何公司都会雇他做他想为他们做的任何事情。我的意思是，当你达到某个点时，我想钱真的不再重要了。在这个家伙的情况下，他想做一些他感到有灵感的事情。而且，嘿，全力支持他。

And I'm sure this guy probably has plenty of money in the bank. If not, he is able to generate it very quickly because any company would hire him to do whatever he wants to do for them. And I mean, when you get to a point, I imagine, the the money doesn't really move the needle. And in this guy's case, he wants to work on something that he feels inspired to do. And, like, hey, all the power to him.

这很令人钦佩。显然，光靠薪水不足以直接吸引所有人。这是一个例外情况，现在他有望继续前进并打造一些很酷的东西。但是，是的，老兄，我很想知道Meta内部的情况。他们是如何对待这些人的？

That's admirable. Clearly, the paycheck isn't enough to straight to, like, pull everybody. This was a case where it wasn't, and now he's going to hopefully go on and build some cool things. But, yeah, man, I would I'd love to know what's going on at Meta. How are they treating these people?

他们在讨论什么？他们在做什么工作？就像，当你把世界上几乎所有的智力复仇者集结起来，放在同一个屋檐下，会是什么样子？并且让他们由这位亚历山大·王来领导，他多大？大概三十五六岁，这个年轻人，这个年轻小伙子，简直了。

What are they talking about? What are they working on? Like, what what does it look like when you assemble essentially the the intellectual avengers of the world and put them under one roof? And let them be run by this this dude Alexander Wang, who's what? Like, in his mid thirties, this young this young guy who's Literally.

而且，他现在是

And and who is now the

这个部门的唯一领导者了，Josh。还记得他们最初成立的时候吗？他是技术联合创始人，和GitHub的前创始人一起。对吧？是的。

sole leader of this unit, Josh. Remember when they first founded it? He was a technical cofounder with the former founder of GitHub. Right? Yep.

现在那个人已经，算是，卸任了，就只剩下这家伙了。所以我想我们可能在未来几个月内就会知道答案。扎克伯格在他的季度信里暗示，他们发现了一种类似自我改进的AI。所以这些都是很重磅的宣言，我希望他能在AI伴侣和AI眼镜之外，真正实现这些承诺。但是，是的，我很期待看到会发生什么。

And now that guy has, like, stepped down, and it's just this guy. So I think we'll probably find out in the next couple of months. Zuck teased in his quarterly letter that they had discovered kind of, like, self improving AI. So those are big fighting words, and I hope that he kind of, like, delivers on that beyond just AI companions and a pair of AI glasses. But, yeah, I'm I'm excited to see what happens.

是的，我也很期待。那么以上就是到本周二的新闻。我们现在是周二。这周还有很多事会发生，但目前你需要知道的大事是，嘿，Gemini表现相当出色。

Yeah. I am too. So that is the news of the week up to Tuesday. We're at Tuesday. There's still a lot left to go for this week, but currently, the big things you need to know are, hey, Gemini is kind of kicking ass.

谷歌做得非常好。他们新的图像生成模型不仅令人惊叹，而且基本上可以免费使用。所以，如果你有任何想要生成的照片，任何想要编辑的照片，或者其实今天早些时候我录视频时看到一个有趣的例子，就是你可以拿亲人的老照片。如果你有黑白照片或者看起来很有复古感的照片，它可以很好地修复它们。所以它不仅能给它们上色，哇。

Google is doing really well. Their new image generating model is not only amazing, but it is readily available to use for basically free. So if you have any photos you wanna generate, any photos you wanna edit or there was an interesting example actually that I was looking at, earlier today when I was recording a video, is that you can take old photos of loved ones. If you have photos that are black and white or they kind of look very, like, vintage, and it will restore them very well. So it'll not only recolorize them Wow.

还会重新为图像添加细节，让它看起来丰富且非常逼真。所以谷歌在各个领域都表现出色。我们对Gemini 3感到兴奋。等它出来，我们会第一时间报道，我向你保证。另一件事是Meta正在打造硬件。

But reapply detail to the image to make it look rich and make it look very realistic. So Google's crushing on every font. We are excited for Gemini three. When it comes, we will be covering it first thing, I promise you. The other thing is is the hardware thing where Meta's building hardware.

这很酷。而且我想告诉大家，下个月是“科技九月”（Techtember）。这是一年中最好的月份。基本上每年九月，世界上每家公司都会发布他们的硬件。谷歌算是取巧了，他们上周就发布了硬件。

This is cool. And I do want to let people know next month is Techtember. This is the best month of the year. Every September, basically every company in the world drops their hardware. Google kind of cheated and they dropped their hardware last week.

但通常的节奏是，现在苹果要发布，Meta要发布，所有硬件公司都会在“科技九月”发布。这将是一场旋风，因为我敢肯定，所有这些硬件发布的前沿都会是AI。我们会在这里进行报道。再次感谢你加入我们的旅程，我们本周晚些时候会回来带来另一期节目。很快再见。

But the way it works is like, now Apple's gonna release, Meta's gonna release, all the hardware companies, they all release in Techtember. It's going to be a whirlwind because I'm sure at the forefront of all of these hardware releases will be AI. We will be here to cover it. Thank you again for joining us for the journey, and we will be back again later this week for another episode. We'll see you guys soon.