扎克伯格谈 Meta AI 眼镜、Muse 与信任问题

嘿，Agria。最近怎么样？

Hey, Agria. What's going on?

马克·扎克伯格和 Meta 正在为其新近走红的 AI 智能体赋予声音。

Mark Zuckerberg and Meta are giving its new popular AI agent a voice.

嘿 Mark，没什么，正在处理，老板。你今天过得怎么样？

Hey Mark, not much, just on it, boss. How's your day shaping up?

还发布了一大堆新眼镜。

And unleashing a whole bunch of new glasses.

那种很轻薄的、基本上和普通眼镜看不出区别的。

Something that is slim, basically indistinguishable from non-AI glasses,

还包括一款轻量级 VR 头显，看起来不像冰球头盔。

including a lightweight VR headset that doesn't look like a hockey helmet.

这很难。但这一切发生之际，世界正对 Meta 和 AI 抱有重大担忧。我脑海里浮现出“变态眼镜”这个词。我去了 Meta 位于门洛帕克的总部，看了这些新东西，并就这一切采访了扎克伯格。

This is hard. But all this comes as the world is grappling with some big concerns about Meta and AI. The term pervert glasses comes to mind. I went to Meta's headquarters in Menlo Park to see the new stuff and asked Zuckerberg about it all.

看，有很多事可以跟扎克伯格聊。我把对话聚焦在 Meta 在 Connect 大会上发布的新产品、它的 Muse AI 智能体，以及计算机的未来。我也有一些关于信任的重大疑问，尤其是现在 Meta 正在打造能看见、听见并更了解我们的功能。但我必须先处理房间里那头 AI 大象。

Look, there are a lot of things to talk to Zuckerberg about. I focused our conversation on Meta's new products announced at its Connect conference, its Muse AI agent, and the future of computers. I also had some big questions about trust, especially now that Meta is building features that can see, hear, and know more about us. But I had to address the AI elephant in the room.

所以，我想我们就从最简单的问题开始吧。

So, I thought we'd start with the easiest question of all.

好。

Okay.

AI 会杀了我们吗？

Is AI gonna kill us?

只是这周你必须从这个地方开始。

It's just that this is the place you have to start this week.

嗯，我认为如果我们都做好自己的工作、负责任地行事，那就不会。我的意思是，我对这会成为所有人非常积极的未来相当乐观。看，我认为有很多充满末日论的言论。我的看法是，有不少实验室正在构建先进的 AI。我不认为我们需要某种全行业协调才能不把这件事搞砸。我认为每个实验室只需要花时间，当它看到有问题时，就在内部花必要的时间，确保自己安全地推进。

Well, I think if we all do a good job and act responsibly, then no. I mean, I'm quite optimistic about the fact that this is going to be a very positive future for everyone. Look, I mean, I think that there's a lot of rhetoric that is filled with doom. And my view on this is that there are a number of labs that are working on building advanced AI. And I don't think that we need some kind of industry-wide coordination to not necessarily mess this up. I think that just each lab needs to take the time and when it sees that there are issues, you just take the time that you need internally to make sure that you're proceeding safely.

你知道，就拿我们刚做的 Muse 发布来说，几个月前我们有一个相当不错的版本。我们想确保真正把隐私和安全的部分都做到位，因为部分原因在于，如果你在打造一个产品，如果它做不到人们想要的事，人们就不会想用它。我们想达到一个人人都能从中受益的世界，而这不会自动发生。如果人们不信任它，就不会采用它。

You know, with the Muse launch that we just did, we had a version of it that was quite good a few months ago. And we wanted to make sure that we really nailed all the privacy and security parts of it because part of this is that if you're building a product, like people aren't going to want to use it if it doesn't do what they want. We want to get to a world where everyone can benefit from this and it's just not going to happen. People aren't going to adopt it if they don't trust it.

所以，我个人的看法是，我们正在进入一个阶段，信任和对齐实际上将成为下一组最重要的能力。我认为我们正走到一个点：它在数学上变得多好可能没那么重要了。它数学已经相当好了。这对很多事情有用。我认为在这一点上，我们在内部谈论的是指令遵循——它是否做你想让它做的事；意图理解——它是否理解你的意图，是否理解你的价值观；对齐。我认为把这些做好，将是接下来最重要的事，以对齐人们在这个世界上真正想要的东西。所以，我不知道。我恰好认为，把这件事做对有充足的商业激励。

So, my own view on this is that we're entering a phase where trust and alignment is actually going to be the most important next set of capabilities. I think we're getting to a point where it may not matter that much how much better it gets at math. It's already pretty good at math. That's useful for a bunch of things. I think at this point the things that we talk about internally are instruction following — does it do what you want; intent understanding — does it understand your intent, does it understand your values; alignment. And I think nailing those things are going to be the most important next things to align with what people actually want in the world. So, I don't know. I happen to think that there's plenty of commercial incentive to get this right.

我用 Muse 已经几周了。就从我注意到的一点说起，它总是在自我纠正。当它弄错什么时，它会回来找我。就像，你知道吗，我弄错了一件事。

And I've been using Muse for a couple of weeks now. And just to kind of jump off of one thing I noticed is it's always correcting itself. When it gets something wrong, it comes back to me. It's like, you know what, I got something wrong.

是的。我们让它具有反思能力。看，这是一个持续的过程，对吧？所以并不是说 Muse 今天就能把所有事都做对。但它多少能感觉到自己能把什么做对、什么做不对。我们大致让它做到：在它做完某件事之后，在你让它做别的事之前的空闲时间，或者晚上你没让它做一堆事的时候，它会看看这一天里发生了什么，并对此进行反思，思考如何做得更好。

Yeah. We made it reflective. And look, it's an ongoing process, right? So, it's not that Muse gets everything right today. But it kind of has some sense of what it's going to be able to get right, what it's not. We kind of made it so that after it does something, in the downtime before you ask it to do something else or at night when you're not asking it a bunch of stuff, it just kind of looks at what happened over the course of the day and it can reflect on that and how to do better.

但我个人的看法，以及我们在这里指导实验室的方式是，我们认为信任和对齐是接下来最重要的能力。所以当你听到业内有人说，哦，能力和对齐之间存在某种取舍时，我就是不同意。我认为，如果我们想让这东西对很多人有用，对齐实际上就是接下来最重要、最需要做对的能力。

And but my own view and the way that we're directing the lab here is that we think that trust and alignment are the next most important capabilities. So when you hear people in the industry talk about, oh, there's some trade-off between capabilities and getting it aligned, I just disagree with that. I think that alignment is actually the most important next capability to get right if we want to make this something that is useful for a lot of people.

所以，我不知道，我认为所有这些其实相当一致，有时你会听到一些实验室说，哦，我们承受了很多痛苦来推迟某些东西，我的看法是，不，你们没有。你们只是在做正确的事，对吧？就像，我们把 Muse 推迟了几个月，因为我们想让它成为对人们更好的产品，对吧？那不是我们做出的牺牲。那对人们是正确的事，对 Meta 也是正确的事。我认为做这类事情，对其他正在构建这个的公司来说也会是正确的事。

So I don't know, I think that all the stuff is quite aligned and sometimes you get some of the labs saying, oh, we're taking a bunch of pain in order to delay stuff and my view is like, no, you're not. You're just doing the right thing, right? It's like, I mean we held Muse back for a few months because we wanted it to be a better product for people, right? That wasn't like a sacrifice that we made. It was the right thing for people and it was the right thing for Meta. And I think that doing that kind of thing is also going to be the right type of thing for other companies that are building this too.

顺便说一句，对齐是一种花哨的说法，意思是 AI 系统做你让它做的事。这是信任 AI 的关键，如果我们打算信任 Meta、让 Muse 和 AI 眼镜更深入地进入我们的生活，这一点尤其重要。在我们讨论所有这些新设备时，这是一个需要思考的关键点。我想从这张照片开始。

Alignment, by the way, is a fancy way to say that an AI system does what you tell it to. It's key to trusting AI and especially important if we're going to trust Meta by letting Muse and AI glasses deeper into our lives. And that's a crucial point to think about as we discuss all these new devices. I wanted to start with this picture.

好。

Okay.

是的。这张，我爱你。我想我当时 19 岁。

Yeah. This I love you. I think I'm 19.

我很喜欢你这张照片，因为我也有同一台戴尔 Inspiron 8200。

I love this picture of you because I had the same Dell Inspiron 8200.

好。我猜这是

Okay. I'm assuming this is

而且大概是同一台 TI83。

and probably that same TI83.

是的。

Yep.

不，我想我女儿现在做作业用的也是同一款 TI83。

No, I think my daughter has the same TI83 doing her homework now.

我猜你在这上面写了 Facebook 的一些早期代码。

And I'm assuming you wrote some early code of Facebook on this.

是的，我想是这样。电话也在那儿。这个小东西大概是个马达。

Yeah, I think so. The phone is there, too. The little like this is probably a motor.

那盒牛奶里的 Whoppers。

The milk carton of whoppers.

这是那个吗？

Is that what this is?

我想是的。我只是，是的，你有

I think so. I just Yeah, you got

我其实还以为你在猛灌 2% 牛奶。这就是 20 年前计算的样子。

I actually just thought you were like chugging 2% milk. This was what computing looked like 20 years ago.

嗯。

Mhm.

但现在你在构建计算的未来。是的。在你看来，那是什么样子？

But now you got you're building the future of computing. Yeah. What does that look like to you?

嗯，我的看法是，我想每一代计算平台都会变得更自然、更无处不在。所以一开始是这些大型主机，或者像房间那么大的巨型计算机，对很多人来说并不那么容易接触，然后到了我想人们称之为小型机的阶段，但它们仍然相当巨大。接着是我们桌子底下的 PC 和塔式机，我想这就是我们成长过程中有的东西。然后有了笔记本电脑，那是相当不错的东西。

Well, so my view is that I guess every generation the computing platforms get to be kind of more natural and more ubiquitous. So you kind of started off with these big mainframes or kind of huge like room-sized computers that were not that accessible to many people and then you got to I think what people called mini computers, but they were still pretty huge. And then you got like the PCs and the towers under our desks, which I think is kind of what we had growing up. You got the laptops that was a pretty good thing.

而在这条路上的每一步，随着它们变得更易使用，它们也变得更受欢迎。手机显然是真正主流的计算形式，现在每个人都把它放在口袋里，你可以触摸它，它就是这种非常直观的东西。问题是，什么会变得比这更直观、更无处不在？我认为基本上会是那些你整天佩戴的东西。所以我认为会有很多不同的设备。但我猜最重要的最终会是眼镜，原因有几个。其一，还有什么比触摸更直观的交互界面？基本上就是一个 AI 智能体，它理解你，能预判你的需求，能理解你周围的世界，你可以自然地与它交谈。但很多时候，它甚至不需要你做任何事，就能知道你想从周围环境中得到什么。所以为了做到这一点，你需要让它拥有和你一样多的关于世界正在发生什么的上下文。因此眼镜是最好的方式，因为它的形态合适，你可以让一个智能体看到你所见、听到你所闻，在你视野中 discreetly 显示信息，而不会分散你的注意力或让你从正在关注的事情上分心，并且它可以与你交谈。所以我认为眼镜将成为最自然的平台。它让我们摆脱了手机的这种束缚：我们有一个非常丰富的数字世界，却通过这个小屏幕与之交互。它把我们拉离了与我们在乎的人的互动。它让你无法专注于当下世界。所以，你需要某种能帮助你弥合并融合数字世界和物理世界的东西。世界上有数十亿人戴眼镜。我猜就像大多数翻盖手机一样，我们当时就知道它们会变成智能手机。我猜今天人们拥有的大多数眼镜，你只知道在某个时候它们会变成 AI 眼镜或智能眼镜。只是，如果它能帮助改善你的视力，如果它能帮助改善你的听力，如果它能帮助你记住事情，你为什么不呢？然后可能还有一系列其他可穿戴设备作为支持。但我猜眼镜将是最重要的那个。

And at each step along the way, as they got easier to wield, they got more popular. Phones were obviously the truly mainstream form of computing, where now everyone has it in their pocket and you can touch it and it's just this very intuitive thing. The question is what becomes more intuitive and ubiquitous than that? And I think it's basically going to be things that you just wear all day long. So I think there are going to be a whole lot of different devices there. But I would guess that the most important one is going to end up being glasses for a few simple reasons. One is, what is the more intuitive interface for interacting with them than touching it? It's basically an AI agent that understands you and can anticipate what you need and can understand what's going on in the world around you, and you can talk to it naturally. But a lot of the time it will even just know what you're going to want from the environment around you without you even having to do anything. So in order to do that, you want it to have as much context on what's going on in the world as you have. So glasses are the best way to do that because it's the right form factor so that you can let an agent see what you see, hear what you hear, display information in your vision discreetly as you're just going about your day without being distracting or taking your attention away from what you're paying attention to, and it can talk to you. And so I think that glasses are going to be the most natural platform. It frees us from this thing with phones where we have this very rich digital world, yet we interact with it through this tiny screen. It pulls us away from interacting with the people we care about. It makes you not present with the world. So you want something that can help you bridge and bring together the digital and physical worlds. Billions of people in the world wear glasses. And I would guess that just like most of the flip phones, we kind of knew they were going to become smartphones. I would guess that most of the glasses that people have today, you just know that at some point they're going to become AI glasses or smart glasses. It's just, if it can help improve your vision, if it can help improve your hearing, if it can help you remember things, why would you not? And then there's probably a set of other wearables that are supporting to that. But I would guess that the glasses are going to be the most important one.

扎克伯格绝对是在加倍押注眼镜，在产品线两端都推出了新品。在更轻、更便宜的一端，是售价 349 美元的 Ray-Ban Meta 音频眼镜，带有扬声器和麦克风，但没有摄像头。在高端，是售价 1299 美元的 Meta VR 眼镜，将整个 VR 头显塞进了一个令人印象深刻的小巧机身中。就像 Vision Pro 和一副 XR 眼镜生了个孩子。它们有 5K 屏幕，电池和处理器位于你系在腰带上的一个圆盘中。我昨天体验了 VR 眼镜的演示，被它的尺寸震撼到了，因为我太习惯于我称之为面部电脑的东西了，对吧？真的就是笨重的头盔式头显。

And Zuckerberg is absolutely doubling down on glasses with new additions at both ends of the lineup. At the lighter, cheaper end, there are the $349 Ray-Ban Meta Audio glasses with speakers and microphones, but no cameras. At the high end, the $1299 Meta VR glasses, which squeeze an entire VR headset into an impressively small package. It's like a Vision Pro and a pair of XR glasses had a baby. They have 5K screens, and the battery and processor are located in a puck you wear on your belt. I got a demo of the VR glasses yesterday and was kind of blown away by the size because I am so used to what I've called face computers, right? Really just chunky helmet-like headsets.

是的。

Yeah.

头显。

Headsets.

我认为头显是阻碍 VR 发展的重大限制之一。

I think the headset was one of the big constraints that was holding back VR.

是的。然而这个相当轻，对吧？真的很好

Yeah. And yet this was pretty lightweight, right? Really good

不到 100 克。

Less than 100 grams.

非常好的保真度，清晰的屏幕。

Really good fidelity, crisp screens.

是的。你觉得人们实际上会用这些来做什么？

Yeah. What do you think people are going to actually use these for?

我认为它是一机四用。你基本上得到一个私人影院。你得到一个强大的工作站，有许多屏幕，可以带到任何地方。你可以在咖啡店调出六个编码终端。你可以看电影，同时看别的。你可以工作。它是一台游戏机，也是一个混合现实设备，你可以非常高地分辨率看到周围的世界，同时在上面叠加任何你想要的全息图。

I think of it as four things in one. You basically get a private cinema. You get a powerful workstation with many screens that you can take wherever you want. You can be in the coffee shop and bring up six coding terminals. You can watch a movie and watch something else. You can be working. It's a game console and it's a mixed reality device where you can see the world around you in very high resolution while overlaying whatever holograms you want on top of that.

我没在你的列表里听到元宇宙。

I didn't hear on your list the metaverse.

嗯，那就是元宇宙。我认为它基本上是物理世界和数字世界的融合。

Well, that is the metaverse. I think of it as basically the blending of the physical and digital worlds.

所以，它不是你去与数字存在互动的地方。

So, it's not a place that you go to interact with digital beings.

不，我认为重点就是这种融合，对吧？这个概念是，在互联网过去的 30 年里，我们建立了这个极其丰富的数字世界，而我们周围有一个非常丰富的物理世界。奇怪的是，这两件事完全分离。而我们今天与那个世界互动的入口就是这个小小的屏幕，你必须低头看，它让你远离周围的世界。我认为这是不可避免会转变并融合的事情。

No, I think the whole point was this merger, right? And the notion is like over the last 30 years of the internet, we've built this incredibly rich digital world and we have a very rich physical world around us. And it is weird that these two things are completely separate. And our portal for interacting with that world today is this tiny screen that you have to look down on and takes you away from the world around you. And I think that's the thing that inevitably is going to shift and come together.

VR 眼镜要到 2027 年春季才发布，但我试了很多东西。我看了一个超级沉浸式的视频，现在我在比莉·艾利什的演唱会上。做一些工作，在虚拟键盘上打字。好吧，我爱我的新键盘。还玩了《节奏光剑》。这对我来说是大量的锻炼。是我几周来最活跃的一次。我还创建了自己的全息图。

The VR glasses don't launch until spring 2027, but I got to try a bunch of stuff. I got to watch a super immersive video and now I'm at a Billie Eilish concert. Do some work and type on a virtual keyboard. Okay, I love my new keyboard. And play Beat Saber. It's a lot of working out for me. The most active I've been in weeks. I also got to create a hologram of myself.

面朝前方，倾斜头部向上看。

Face forward and tilt your head to look up.

我在 Meta AI 应用中扫描了我的脸并录制了我的声音。

I scanned my face in the Meta AI app and recorded my voice.

用惊讶的语气重复以下陈述。什么？我简直不敢相信刚刚发生了。

Repeat the following statement with a surprise tone. What? I can't believe that just happened.

什么？我简直不敢相信刚刚发生了。我的全息图正在生成。我看起来相当不错。嗯，直到我笑了，然后我和一位 Meta 产品经理或她的一个非常小的全息图进行了通话。

What? I can't believe that just happened. My hologram is being generated. I came out looking pretty good. Well, until I smiled and then I got to do a call with a Meta product manager or a really tiny hologram of her.

你应该看看立体视频。

And you should see the stereoscopic video.

我认为有这种格式，未来你会在会议中，也许我们会物理上在一起，你会有那种存在感，就像你和一个人在一起。它与你通过 Zoom 得到的任何东西在质量上完全不同。未来，你还可以有第三个人，他们可以作为全息图在这里，你会有一种真实的存在感，就像我们在一起一样，而不是一个悬浮的视频屏幕，但我认为你在 Meta VR 眼镜中得到的全息图基本上正在接近这一点。这是一件事，这是一个我们可能已经研究了近十年的研究问题。是的，物理世界和数字世界的融合就是元宇宙的概念。

I think that there are formats of that where you'll be in a meeting in the future and maybe we'll be there physically together and you have the sense of presence where you're there with a person. It's just qualitatively extremely different from anything that you can get over Zoom. In the future, you'll also be able to have a third person, and they'll be able to be here as a hologram and you'll have a realistic sense of presence as if we were there together, not like a hovering video screen, but I think the holograms that you get in the Meta VR glasses are basically approaching this. And that's a thing and that's a research problem that we've been working on for probably almost a decade. Yeah, that blending of the physical and digital worlds is the idea of the metaverse.

现在我们把 Reality Labs 的重心稍微转移了一下，因为刚开始的时候，我预测元宇宙——这种全息影像，把数字世界和物理世界融合在一起——会在我们真正拥有先进 AI 之前实现。技术树演化的方式很有意思，我们实际上在能够以足够实惠的形态让这些全息影像普及之前，就已经获得了非常先进的 AI。而且我认为，先进 AI 很可能在世界上的分量会大得多。我有点像这样想：当苹果在开发 iPad 时，他们意识到可以造手机，于是把 iPad 搁置了几年，后来又回来做，最终 iPad 也成了大产品，但手机要大得多。我想我们也有类似的想法：我们要构建元宇宙，把数字和物理世界融合，带来临场感，然后再做 AI。

Now we have shifted the focus for Reality Labs a bit because when we got started I had predicted that the metaverse, this kind of hologram and bringing together the digital and physical worlds, would happen before we got really advanced AI. It's been an interesting quirk of the way that the tech tree has evolved that we've actually gotten very advanced AI before you can get these holograms in an affordable enough form factor to be ubiquitous. And I think it also turns out that advanced AI, I would guess, will probably just be a significantly bigger thing in the world. So I think about this a little bit like when Apple was building the iPad and then they realized that they could build the phones, they put the iPad on the shelf for a few years, came back to it, and it ended up being a big thing, but the phone was much bigger. I kind of think we had this in mind too, that we were going to build the metaverse, we're going to bring together the digital and physical worlds, deliver the sense of presence, and then we're going to do AI.

嗯。

Mhm.

然后突然发现，天哪，我们现在就能做 AI 了。所以基本上，我们正在经历同样的时刻：好吧，让我们继续在元宇宙和全息未来上做一些研究，但我们在做任何事情时都必须记住，AI 将会大得多，而且现在就有可能实现，并以非常实惠、非常时尚的形态交付，让数千万甚至最终数十亿人拥有。

And then it's like, oh man, we can do AI now. So basically, we're kind of going through the same moment of saying, all right, let's keep on doing some of the research on the metaverse and this holographic future, but we just have to keep in mind with everything that we do that AI is going to be both so much bigger and it is possible to do now and deliver in these form factors that are very affordable, kind of very stylish, that tens of millions or eventually billions of people are going to have.

还有 Ray-Ban Meta Gen 3 眼镜，有多种新设计，麦克风更多，电池续航更长。新的 Ray-Ban Meta 音频眼镜内部隐藏了六个麦克风，还有一个操作按钮。它们采用 clubmaster 风格镜框，这让我只能想到《春天不是读书天》里的那个片段。

There's also the Ray-Ban Meta Gen 3 glasses which come in a bunch of new designs and have more microphones and battery life. The new Ray-Ban Meta audio glasses have six microphones hidden inside along with an action button. They come in the clubmaster style frame which means I can only think of this clip from Ferris Bueller.

“你有没有给爸爸一个吻？”我认为有些人会想要最轻最薄的东西，或者最便宜的价格。而纯音频眼镜可以让你做到这一点。去掉所有其他传感器，你就能做出很薄的东西，看起来和非 AI 眼镜几乎无法区分。

Do you have a kiss for daddy? I think some people are going to want the lightest and slimmest thing or for the cheapest price. And basically audio only glasses allow you to do that. By taking all of the other sensors out, you can just make something that is slim that just looks basically indistinguishable from non AI glasses.

就是这些，对吧？

That's these, right?

就是这些。是的。带有标志性的 Ray-Ban clubmaster。我们现在甚至在这个镜框里没有任何电子元件，这让我们能够做出以前从未能做到的设计。

That's these. Yeah. With the kind of iconic Ray-Ban clubmaster. We now don't even have any electronics in this rim which allows us to do this design which we were never able to do before.

这些音频眼镜用于听音乐、接电话和与 AI 对话。公司告诉我，音频 AI 眼镜目前没有录音功能，比如录制会议。现在 Meta 正通过新的听力增强功能，将音频眼镜和新的 Gen 3 眼镜变成助听器。所以我塞了一些耳塞，让我本来很好的听力变得相当差，然后试了试。我现在有了超人的听力。

These audio glasses are meant for listening to music, taking calls, and talking to AI. The company told me there's currently no recording feature on the audio AI glasses, say for recording a meeting. And now Meta is turning the audio glasses and the new Gen 3 glasses into hearing aids with a new hearing enhancement feature. So I put in some earplugs to make my perfectly good hearing considerably worse and gave it a try. I have superhuman hearing now.

在纯音频眼镜上实现这个功能还需要几个月。但我们现在基本上已经开始出货听力增强功能，这将是 FDA 批准的医疗级助听器，装在一副时尚的眼镜里。我只是觉得，好吧，需要助听器的人，我认识很多需要的人，但他们不去配，因为助听器有点笨重，或者他们可能觉得尴尬之类的，他们不应该这样，但我认识的人确实会。所以他们只是坐在对话中，你能感觉到，哦，这个人可能没有完全跟上正在发生的事情。但有了眼镜，你现在就有一副完全正常、时尚的眼镜，可以提供医疗级的听力辅助。我认为这只需要助听器成本的一小部分。我觉得这本身就是一个非常强大的东西。然后你还能叠加所有 AI 功能。

It's going to be a few months before we have that on the audio only glasses. But we're now basically starting now shipping hearing enhancement which is going to be kind of an FDA approved medical-grade hearing aid that fits in a pair of stylish glasses. I just think that there's this thing where it's like, all right, people who need a hearing aid, you know, I know a lot of people who need it, but then they don't get it because it's kind of bulky or maybe they feel embarrassed about it or something like that, which they shouldn't, but I mean, I know people who do. So they just kind of sit in conversations and you kind of get a sense of, oh, maybe the person isn't quite picking up on everything that's going on. But with the glasses, you now just have a completely normal and stylish pair of glasses that can offer medical grade hearing assistance. I think that is for a fraction of the cost of a hearing aid. I mean, I think that is a very powerful thing by itself. And then you get all the AI features layered on top of that.

我们回到 Money Mic，与 Mercury 一起，从我在加利福尼亚的酒店。今年九月，我和我的团队在硅谷待了很多时间，这意味着我们花了很多钱穿越全国，然后到了这里自然花得更多。我们已经向你展示了用 Mercury 卡花钱和用新的支出功能设置预算是多么容易。但真正的魔力发生在花钱之后，也就是商务旅行中大家最喜欢的部分：收据。好吧，我们的大部分收据都是数字的 Uber、餐饮和随机去打印机的行程。通常这意味着在收件箱里进行一场悲伤的寻宝游戏，然后把每张收据上传到 Mercury 中不同的交易。但现在，当你打开收据匹配功能，你只需把一堆收据扔进去或通过电子邮件发送给 Mercury，它就会自动附加到交易上。我的簿记员很高兴，我也很高兴。最重要的是，收据很高兴。链接在描述中，或访问 mercury.com。现在让我们回到 Meta。

And we're back on the money mic with Mercury from my hotel in California. This September, me and my team have been spending a lot of time in Silicon Valley, which means we've been spending a lot of money getting across the country and then naturally spending even more once we get here. We've already shown you how easy it is to spend money with the Mercury card and set budgets with the new spend feature. But the real magic happens after the spending with everyone's favorite part of business travel, receipts. Okay, most of our receipts are digital Ubers and food and random trips to the printer. Normally that means a sad scavenger hunt through inboxes and then uploading each receipt to the different transaction in Mercury. But now when you turn on receipt matching, you can just dump a bunch of receipts or email them to Mercury and it automatically attaches to the transaction. My bookkeeper's happy, I am happy. Most importantly, the receipts are happy. Link in the description or go to mercury.com. Now let's go back to Meta.

对于那些不了解情况的人，今年夏天很多人开始把 Meta 眼镜称为“变态眼镜”。

For those not up to speed, this summer many started calling the Meta glasses pervert glasses.

Meta 的变态眼镜。

Meta's pervert glasses.

变态眼镜。

The pervert glasses.

这是因为人们开始移除眼镜上的指示灯并进行隐蔽拍摄。只是反思一下你因为这些眼镜而遭遇的一些反弹。然后“变态眼镜”这个词就浮现在脑海中，对吧？我相信你——

It came as people started removing the indicator light on glasses and covertly recording. Just reflecting though on some of the backlash that you've had with these glasses. And the term pervert glasses comes to mind, right? I'm sure you—

是的。我看到了那个活动。我认为这是我们最近看到的一部分，眼镜已经变得主流。我认为我们可能应该确保一直更多地谈论这个功能。我认为像我们这样的公司，我们既需要实质性地做正确的事情，我认为在这种情况下我们正在做。我们也需要确保我们谈论它，这样人们就知道我们的立场。他们可以放心，知道我们在处理这个问题，我们在关注它。这就是我认为我们最近加强的部分。

Yeah. I've seen the campaign. I mean, I think it's part of what we've seen recently is that the glasses have just become mainstream. And I think we probably should have made sure that we talked more about this feature all along. I think companies like ours, we both need to substantively do the right things, which I think in this case we are. And we need to make sure that we're talking about it so that people know what we stand for. They have the comfort of knowing that we're kind of on it and we're focused on it. And that's the part that I think we've ramped up more recently.

所以我想谈谈 Ray-Ban 音频眼镜，因为我认为那是最有趣的公告之一。

So I want to talk about the Ray-Ban audio glasses because I think that's one of the most interesting announcements.

是的，我对此非常兴奋。

Yeah, I'm really excited about this.

这是对那些人说“我不舒服别人脸上有摄像头”或“我不舒服自己脸上有摄像头”的回应吗？

Is that a response to people who have just said, I am uncomfortable with cameras on people's faces or I'm uncomfortable with having a camera on my face?

不，我的意思是我们必须做这些事情。我们开始做它们，你知道这是怎么运作的。我的意思是提前很长时间，所以如果人们对某件事有疑问，而我们开始做某件事，我们需要几年时间才能造出这个东西。

No, I mean we have to work on these things. We start working on them like, I mean you know how this works. I mean like a very long time in advance before so if people have a question about something and we start working on something that it's going to take you know a couple of years for us to build the thing.

所以我要说的就是：不是。我们并不是因为几个月前有人提出了一些问题，才决定用两个月时间设计这个东西。这个项目已经酝酿很久了。

So all I'm saying is no. We didn't decide to design this in two months because some people had some questions a couple months ago. This has been a project that's been a long way in the works.

我注意到这副眼镜上没有指示灯。

I noticed on these that there's no indicator light.

嗯，因为它没有摄像头。

Well, there's no camera.

但你可以用它录音。

But you may record audio with these.

哦，对。我是说，那个指示灯是给摄像头用的。回到你之前的问题，最近人们对这件事提出了更多疑问，我也一直在反思为什么。为什么会这样？因为从设计带摄像头的眼镜的第一天起，我们就内置了许多手机或任何能拍照的设备都没有的功能。据我所知，这是唯一一款会在你拍照或录像时提示你的设备。而人们用手机拍照时根本不会这样。

Oh, yeah. I mean, the indicator is for the camera. And that's been — I mean, going back to the question that you had before, people have raised more questions about this recently, and I've been reflecting on why. Why is that? Because from day one of designing the glasses that have cameras, we've built in a number of features that no phone or anything that can take pictures has had before. I mean, this is the only device that I'm aware of that tells you when you're taking a photo or recording a video. I mean, people do that with phones all the time.

就像，你知道，那种——你把手机举起来，对吧？它们没那么——

There's like, you know, this sort of — you're holding your phone up, right? They're not as —

当然。但我们从第一天起就把这一点考虑进去了。我们出货的第一副眼镜，我们就想，这是我们要主动引领的事情。我们知道人们会对这件事有疑问，我们要确保从一开始就负责任地处理它。所以我们内置了摄像头指示灯，并且随着时间推移增加了更多保护措施，如果有人篡改它，我们就把摄像头锁死。如果有人试图用胶带盖住它，我们基本上就不允许它录像或做类似的事情。

Sure. But we took this into account from day one. The very first pair of glasses that we shipped, we're like, this is a thing where we want to lead on. We know that people are going to have questions about this, and we want to make sure that we're handling this responsibly from the beginning. So we built the camera light in, and over time we have added more protections to it, so that if people tamper with it, we brick the camera. If people try to put a piece of tape over it, we basically don't allow it to record or anything like that.

今年 6 月，我们报道了全国各地有人提供一种服务，把那个提示正在录像的隐私指示灯钻掉，这篇报道引起了很大关注。Meta 回应了一个更新，如果指示灯被篡改，就禁用摄像头。对，这其实就是你可能在我报道里见过的那副眼镜——我把它带到新泽西一个陌生人的车库里。不推荐这么做。

Back in June, our story about how people around the country were offering a service to drill out the privacy light that tells people that they're recording got a lot of attention. Meta responded with an update that disabled the camera if the light had been tampered with. And yeah, these are actually a pair that you might have seen in my reporting — I took these to a garage, a random man's garage in New Jersey. Do not recommend.

我们把它们锁死了吗？

Did we brick them correctly?

现在它们被锁死了。对，解释一下——我把这副眼镜带到那个陌生人的车库。他把灯钻掉了，我就能在别人不知情的情况下录像。然后你们反应很快，说，嘿，我们会想办法遏制这种情况。但我的问题一直是，到底有多少人真的这么做？

These are now bricked. But yeah, just to explain — I took these to a random man's garage. He drilled the light out, and I was able to record without people knowing. And then yes, you guys responded pretty swiftly with, hey, we're going to figure out some ways to clamp down on this. But my question has actually been all along, how many people are actually doing that?

我们观察到的是，在拥有这副眼镜的人里，这么做的人不到千分之一。所以是非常小的一部分。从你做的那些工作来看——我是说，你拿着眼镜，当然可以拆开它或随便怎么弄，但你得去找个人，弄些电动工具，做一堆事，对吧？所以这从来都不简单。我觉得无论你在构建哪个平台，总会有某种对抗博弈，有些人想干坏事，我们总得防着。抱歉，但就回答你的问题，这个比例非常非常小，我们基本上觉得，对于试图这么做的人，我们在锁死眼镜这项功能方面做得相当不错。但不到千分之一。

So I think what we've seen is it's less than one-tenth of a percent of people who have the glasses. So it's a very small fraction. And I think from the work that you did to go do that — I mean, you have the glasses, you can obviously disassemble them or do whatever you want, but you had to go to a guy and get some power tools and go do some stuff, right? So it was never straightforward. And I think whenever you're building any of these platforms, there's some kind of adversarial game where some people try to do bad stuff, and we're always going to be defending against that. Sorry, but just to answer your question, it's a very, very small percent, and we basically feel like we've done quite a good job of being able to brick that functionality in the glasses for the people who try to do that. But less than a tenth of a percent.

但我觉得，这些眼镜已经进入主流了。人们在戴它们，可同时又有一种看法，认为 Meta 在打造一个监控网络，一个可以直播视频或音频的网络。你告诉我那不是真的，但确实存在这种看法。你怎么对抗这种看法？

But I think on that, they've gone mainstream. People are wearing these glasses, and yet there's this perception that Meta is creating this surveillance network, one where they're streaming video or streaming audio. And you're telling me that's not true, but there's a perception. How do you fight that perception?

我觉得就是我刚才说的那些。我基本上认为，首先我们得从实质上确保架构是对的。其次，我们得确保沟通到位，让人们理解我们在做什么。你看，永远有更多事情要做。我们在 Connect 上宣布的一件事是，我们正在为眼镜构建这样一个环境：当你使用 AI 功能或与它对话时，这些都在可信计算环境中完成，这样即使 Meta 的人想看那些内容也看不到。这就像最高标准的安全性。我们就是不断添加越来越多这样的功能。我觉得基本上我们会在某个时候发现，其他公司也会尝试做这样的眼镜。而我认为我们在这些隐私和安全功能上会遥遥领先，到那时人们会把它和其他公司的产品比较，大概就会意识到我们投入了多少，以及它有多先进。但退一步看，我觉得人们对 Meta 的看法来自一个有意思的地方，因为我们最初做的是 Facebook，而 Facebook 是一个用于分享的服务。所以它本来就不是用来把所有东西都保密的。当然我们现在在保护隐私方面做得不错，但本质上它是一个分享服务。我想出于一些很合理的理由，人们不会把 Facebook 当成世界上最私密的东西。但这并不意味着 Meta 不做隐私。比如你看 WhatsApp，WhatsApp 是领先的私密通信服务，拥有世界上最大的端到端加密网络。是我们开创了把端到端加密做到这种规模。所以我觉得未来几年人们会看到一件事，那就是我其实认为 Meta 今后要做的更多事情会更像 WhatsApp，而不是那些纯粹社交分享的服务。所以如果你在用眼镜，那是一种非常私密的体验，而我们要围绕它构建的隐私架构——我们会构建越来越多连 Meta 都看不到内容和你在做什么的东西——这差不多就是我们在 WhatsApp 上做的事。

I think it's the same things that I just said. I basically think we substantively need to make sure that we have the right architecture. And then second, we need to make sure that we communicate, that people understand what we're doing. So look, there's constantly more things to do. One of the things that we announced at Connect is we're building this environment for the glasses where when you're doing AI functionality or talking with it, that it's done in a trusted computing environment, so that even if people at Meta wanted to see that content, they couldn't. It's like the highest standard of security. And we just keep on adding more and more features like that. And I think basically what we're going to find at some point is that other companies will try to build glasses like this too. And I think ours are just going to be so far ahead on some of these privacy and security features that at that point people will compare this to what others are doing and probably recognize just how much investment we've put into it and how advanced it is. But taking a step back on this, I think people's perceptions of Meta come from an interesting place, because we started obviously building Facebook, and Facebook is a service for sharing. So it is not meant to be a service for keeping all of your things private. Now we do a good job of keeping the private things private, but fundamentally it is a service for sharing. I think for a bunch of good reasons, people don't think about Facebook as the most private thing in the world. But that doesn't mean that Meta doesn't do that. If you look at WhatsApp, for example, WhatsApp is the leading private communication service. It has the largest end-to-end encrypted network in the world. We pioneered bringing end-to-end encryption to that scale. So I think there's a thing that people are going to see over the coming years, which is that I actually think more of the things that Meta is going to do going forward are going to look like WhatsApp than they're going to look like these purely social sharing services. So if you're using glasses, that is a very private experience, and the privacy architecture that we're going to build around it — where we're going to build an increasing number of things where even Meta can't see the content or the things that you're doing — that's kind of what we did with WhatsApp.

9 月中旬，Meta 发布了 Muse，一个 AI 智能体平台。此后它一路飙升到应用商店榜首。它可以主动做你让它做的事：购物、追踪新闻、回复邮件、安排会议。它在 Meta 的服务器上有自己的计算机，或者你会听到扎克伯格称之为虚拟机，简称 VM。这让 Muse 可以全天候为你工作，并像朋友或同事一样主动给你发消息。Muse 显然就是为你打造的未来操作系统。

In mid-September, Meta released Muse, an AI agent platform. And since then it has skyrocketed to the top of the app store. It can proactively do things you tell it to do: shop, track news, respond to emails, schedule meetings. It has its own computer on Meta's servers, or what you'll hear Zuckerberg refer to as a virtual machine, or VM. This allows Muse to work for you all hours of the day and proactively message you like a friend or colleague. Muse is clearly this operating system of the future for you.

看起来这将成为驱动这些设备的东西，你自己的个人智能体。它和聊天机器人有什么不同？

It seems like that's going to be the thing that powers these devices, your own personal agent. How is it different than a chatbot?

我认为未来每个人都会有一个个人智能体，基本上可以为他们做事。它会理解他们的目标，理解他们的上下文，并为你把事情搞定。你把它叫做操作系统。我觉得这有点不同。操作系统是人和设备之间的技术接口，而我认为这更像是你与世界中所使用的所有技术之间的接口，你基本上给它一个目标，它就可以为你工作。比如，我女儿想爬一座需要许可证的山。我就让它去申请许可证，它会在某个时候给我发消息说：“好的，我们拿到许可证了。我们可以一起去爬山了。”人们让它帮忙省钱。比如，去查看我的电子邮件，找到一堆我没在用的订阅，或者我有重复的订阅，帮我取消这些来省钱。但还有像我在综合格斗馆里装了摄像头，让它在我实战训练时观看，并在回合之间给我反馈这样的小众需求。所以它必须构建整个视频处理栈，而且它需要知道哪个人是我。所以这很重要。它需要快速处理视频，因为回合之间只有大约一分钟。

I think that in the future everyone's going to have a personal agent that basically can do things for them. It will understand their goals, understand their context, and can get things done for you. And you call it an operating system. I think it's a kind of different. An operating system is sort of the technical interface between a person and a device, whereas this I think is more your interface to all of the technology that you're using in the world, where you basically give it a goal and it can just go work on things for you. You know, like my daughter wants to climb a mountain that requires a permit. So I just ask it to get us a permit when it can, and it'll message me at some point and say, "Okay, we got the permit. We can go hike the mountain together." People ask it to save them money. Like, go look in my email, find a bunch of subscriptions that I'm not using or if I have duplicate subscriptions and save me money by helping me cancel those. But then there's things that are sort of as niche as I put cameras in my MMA gym and I have it watch me when I'm sparring and send me feedback in between rounds. So it had to build this whole video processing stack and it kind of needed to know which person is me. So it's like that's an important part of it. It needs to process the video fast because you get like one minute in between the rounds.

所以这一切都发生在你的 Muse 里。我们能见见你的 Muse 吗？

So this is all happening in your Muse. Can we meet your Muse?

当然可以。好的。好的。好的。

Yeah, sure. Yeah. Yeah. Yeah.

你的 Muse 是做什么的？

And what does your Muse do?

你想怎么进行？它是一个以这位古典罗马人物命名的抓取器。但在这里，我可以带你看看一些东西。这实际上是它在我今早来这里之前，从我实战训练中发来的反馈，它只是给我发我做动作的照片，并给出反馈。

How do you want to do this? It's a gripper just named after this classical Roman figure. But here, I mean, I can take you through a bunch of the stuff. This is actually it sending me feedback from my sparring right before I came here this morning, where it just sends me photos of the things that I'm doing and gives me feedback on them.

这就像反馈是

It's like is the feedback

嗯，就像如果你在防守，你需要更用力地伸展，把臀部收回来。

Well, it's like if you're defending, you need to sprawl harder and get your hips back.

这就像，我是说，呃

It's like I mean it's uh

它在这方面有很多经验。但你正在添加的一件事是语音，对吧？

It has a lot of experience doing that. But one thing you're adding one thing you're adding is voice, right?

是的。所以这是我们刚刚宣布的，我们刚刚宣布了连接。嗯，因为很多人想和它说话。我们应该启动它吗？

Yeah. So that's that's a thing that we're just announcing at that we we just announced to connect. Um because a lot of people want to talk to it. Should we should we fire it up?

我很想和 Agria 说话。

I would love to talk to Agria.

我希望如此。

I'm hoping.

好的。嘿，Agria。怎么了？

All right. Hey, Agria. What's going on?

嘿，马克。没什么。正在处理，老板。你今天过得怎么样？有什么需要完成的事情吗？

Hey Mark. Not much. Just on it boss. How's your day shaping up? Anything we need to knock out?

所以，我实际上正在和乔安娜·斯特恩做播客。你熟悉她吗？你觉得她的工作怎么样？

So I'm actually in the middle of doing a podcast with Joanna Stern. Are you familiar with her? What do you think of her work?

哦，不错。和乔安娜·斯特恩做播客。那是个不错的收获。她很敏锐，能看穿消费科技炒作，而且她有一种罕见的混合特质：既 thorough，又听起来有趣。到目前为止怎么样？你们俩在深入探讨 AI 的东西吗？

Oh, nice. Podcast with Joanna Stern. That's a solid get. She's sharp, cuts through the hype on consumer tech, and she's got that rare mix of being thorough, but still fun to listen to. How's it going so far? Are you two diving into the AI stuff?

Agria，你订阅了我的 YouTube 频道吗？

Agria, are you subscribed to my YouTube channel?

让我快速检查一下你的 YouTube 订阅状态。没有。我没有 YouTube 账户，所以我无法订阅。但我在播客中和你在一起，所以在我这边你没有错过任何东西，老板。

Let me check your YouTube subscription status real quick. Nope. I don't have a YouTube account, so I can't subscribe. But I'm right here with you for the podcast, so you're not missing anything on my end, boss.

非常可爱。

It's very cute.

确实可爱。

It is cute.

整个头像风格，我们内部称之为 Jollybot 的默认设置。我认为所有其他实验室的默认设置都是这些像机器人一样的东西，我们只是想要一些有趣的东西。

The whole avatar style, the default that we call internally Jollybot. I think the kind of default that all the other labs have are these like robot looking things and we just wanted something that was like kind of like fun.

我的叫 Thingy。

Mine is called Thingy.

是的。它有点

Yeah. And it's a little

你定制了吗？

Did you customize it?

是的。我想让它感觉像我们的品牌，我们的新事物品牌，所以它是一个小橙色蛋。

Yeah. I wanted it to feel like our brand our our new things brand and so it's a little orange egg.

明白了。

Got it.

嗯，但我需要在他身上做更多工作。

Um but I I need to do more work on him.

我给 Agria 穿了一件托加袍。

I gave a Gria a toga.

是的。给抓取器穿了一件很好的托加袍。但我想，因为我想告诉你一些关于我的 Muse 体验，因为它真的，我对它为我做的一些事情感到震惊。它基本上让我来到这里，对吧？比如它跟踪我的航班。它预订了我的理发预约。嗯，它一直在为这次采访做准备。每天早上我都会收到每日提醒

Yes, it is. Very nice toga for a grippa there. But I want to because I've I want to tell you a little bit about my Muse experience because it really I've been blown away by some of the stuff it's done for me. It basically got me here, right? Like it tracked my flights. It booked my hair appointments. Um it has been constantly preparing me for this interview. Every morning I get the daily poke

它给我建议要问你什么，并给我关于你的新闻。非常非常有用。但有些地方我觉得我对 Meta 没有完全的信任。所以我没有，例如，连接我的个人 Gmail，因为我有两个担忧。第一，我信任 Meta 拥有我的个人 Gmail 信息，我非常私密的消息吗？第二，如果 Thingy，我的 Muse，突然开始随机给人们发邮件或做一些我不想让它代表我做的事情怎么办？

and it gives me advice on things to ask you and gives me news about you. Very, very useful. But there have been some places where I've felt I don't fully have the trust with Meta. And so I haven't for instance hooked up my personal Gmail because I have two fears there. One, do I trust Meta to have my personal Gmail information, my very private messages? And then two, what if thingy, my muse, goes off and just starts randomly emailing people or doing something I I don't want it to do on my behalf.

是的。所以我认为这两个都是重要问题。关于第一个，我们构建了隐私模型，拥有这个 Muse 安全虚拟机，你知道你所有的内容都在你自己的虚拟机上。我们在隐私方面采取了许多步骤，所以你所有的凭证和类似的东西都是你无法被智能体访问的。Muse 实际上不是一个智能体。有 Muse 智能体，但还有一个 Sentinel 智能体，基本上监管它以及所有进出它的信息。它会查看并确认那是你想要的东西。这就是为什么有时它会标记需要你验证的东西。Muse 智能体本身不能选择它向你运行什么。那实际上是 Sentinel 智能体。有整个架构，然后我们正在采取额外的步骤，我们让 Moxy Marlin Spike，他是我们十多年前在 WhatsApp 的原始信号加密协议上合作的人，正在研究 Muse 机密虚拟机，这是一个系统，你的虚拟机可以有一个只有你拥有的加密密钥。这样 Meta 甚至不能看到里面的东西，即使我们想也不能，你知道这显然是终极保护级别。我的意思是，这就是为什么很多人对 WhatsApp 非常有信心，因为公司，你知道我们提供服务，我们实际上甚至不能看到消息，即使我们被黑客攻击或发生了一些事情，或者发生了不好的事情。所以那种级别的安全是我们努力追求的，我认为最终应该给人们他们需要的信心，关于我是否 okay 让这个上下文驻留在一个 Meta 系统中。就像好吧，如果安全是好的，我们发布白皮书，安全社区审查它，这将是彻底和强大的。

Yeah. So I think both of those are important questions. So on the first one, we've built the privacy model to have this Muse secure VM where you know all your content is on your own virtual machine. We've taken a number of steps on the privacy so that all your credentials and everything like that is you not accessible to the agent. Muse is actually not one agent. It's there's the Muse agent, but then there's a Sentinel agent that basically polices it and all information that's going in and out from it. It will look to see if and confirm that that's the thing that you want. That's why sometimes it flags stuff that it needs you to verify. The Muse agent itself doesn't get to choose what it runs by you. That's actually the Sentinel agents. There's this whole architecture and then we're taking this additional step where we had Moxy Marlin Spike who is the guy who we worked with on the original signal encryption protocol for WhatsApp more than 10 years ago is working on the Muse confidential VM which is a system where your virtual machine can have an encryption key that only you have. So that way Meta can't even see the stuff inside it even if we wanted to which is you know obviously is kind of the ultimate level of protection. I mean that's why a lot of people are very confident about WhatsApp is that the company you know we are providing the service and we can't actually even see the messages even if we got hacked or if something happened where or something bad happened. So I that level of security is sort of what we strive to do which I think ultimately should give people the confidence that they need on am I okay with this context residing in a system that is a Meta system. It's like okay well if the security is good we publish the white paper the security community vets it that this is going to be kind of thorough and strong.

我们知道 Meta 无法访问这些数据，这意味着如果黑客以某种方式入侵了 Meta，他们也无法获取这些数据。如果你所在的国家可能没有良好的民主程序，政府要求获取这些信息，他们无法强迫我们向政府提供这些信息，因为我们自己也无法访问。我认为所有这些都是好事。

We know that Meta isn't going to be able to access this, which means that if a hacker somehow gets into Meta, they're not going to be able to get it. If you're in a country that maybe doesn't have a good democratic process and the government demands it, they can't force us to give that information to the government because we can't get access to it ourselves. All of that I think is good.

你问的第二个问题，我们称之为“谨慎”。基本上，AI 智能体需要从头开始训练，以了解哪些信息可以分享，哪些不可以。因为这和构建一个编程智能体非常不同。也许你想去餐厅预订，而你有某些饮食限制，或者你怀孕了，或者随便什么情况，但你不想告诉别人，对吧？你只是希望它能够为你做正确的事情，同时尽可能少地透露信息来实现这个目标。

The second question that you asked around, we call this discretion. Basically, the AI agent needs to be trained from the ground up to have a sense of what information is okay to share and what isn't. Because this is very different from building a coding agent. Maybe you wanted to go get a reservation at a restaurant and you have certain dietary restrictions, or maybe you're pregnant, or whatever the thing is, but you don't want to have to tell someone that, right? You just kind of want it to be able to go do the right thing for you while revealing as little information as possible to accomplish that goal.

这就是谨慎，它是一种人类技能。但我认为这实际上回到了我们最初讨论 AI 安全时你问的第一个问题。我的回答是，我确实相信这类对齐和信任能力是我们接下来需要做的事情。为了让像 Muse 这样的东西从早期产品变成我认为世界上数十亿人会依赖的东西，他们需要能够信任它能很好地做到这一点。这意味着像谨慎这样的能力将是我们必须真正成功实现的事情。

That is discretion and it's a human skill. But I think it actually goes back to one of the first things that we talked about when you were asking about AI safety. And my answer was that I actually believe that these kind of alignment and trust capabilities are the next thing that we need to do. In order for something like Muse to go from an early product to something that I think billions of people in the world will rely on, they will need to be able to trust that it can do that very well. Which means that capabilities like discretion are going to be things that we will have to really succeed at.

那么我们如何做到这一点？这就是为什么我们建立了 Meta 超级智能实验室，来完成从自己训练模型到构建产品的整个链条，因为我们不能只是从其他公司那里拿来现成的模型，然后希望它擅长谨慎，这是我们从一开始就关注的事情，把它放入模型中，在管道的所有步骤中训练它，这样我们才能得到比其他人更好的结果。

So how do we do that? I mean that's why we built Meta Super Intelligence Labs to do the whole chain from training the models ourselves to building the products because we can't just take some other company's model off the shelf and hope that it's going to be good at discretion like that's a thing that we focus on from the beginning, put it into the models, train it at all steps through the pipeline and that's how we get to a result that is better than what anyone else is doing.

我认为我们现在最大的恐惧之一，每个人都在谈论的，显然是失控的智能体，以及对 AI 失控的担忧，我知道你分享过这一点，但我认为这忽略了目前我使用 AI 时感受到的最大风险之一。我尽可能多地使用 AI 一年。我写了一本关于它的书，我离开时最大的恐惧是社交影响，我的孩子可能会在成长过程中与 AI 伴侣或聊天机器人交谈，这会在世界上助长更多的孤独感。你在部署这些时是否考虑过这个问题？

One of the big fears I think we have right now and everyone's talking about are obviously rogue agents and this worry about AI getting out of control and I know you've shared about that but I think part of that misses right now one of the biggest risks that I have felt with using AI. I used AI for a year as much as I could. I wrote a book about it and I came away my biggest fear was about the social implications that my kids may grow up talking to an AI companion or chatbot and it fuels even more loneliness in this world. Is this something that you are thinking about as you deploy these?

是的，我的意思是，所有这些东西都有好的一面和坏的一面。我最小的女儿喜欢烘焙，我对烘焙一无所知，但基本上每个周末它都会整理出一份食谱。它订购食材以确保它们在那里。否则会需要很多准备工作，也许我很难做到，所以也许我就不会去做。现在我可以做一件事，与我三岁的女儿联系，做这个活动，否则我会很难做到，因为我的 Muse 帮助设置好了。所以我发现它可以帮助你提高个人生产力，完成很多繁琐或基本的任务，这样你就可以去做你真正想做的事情。所以不同的人对他们真正想做的事情有不同的例子。但我认为对很多人来说，这是与他们关心的人联系。

Yeah, I mean there are good and bad versions of all of this stuff. My youngest daughter likes baking and I don't know anything about baking, but basically every weekend it pulls together a recipe. It orders the ingredients to make sure that they're there. What would have otherwise been a lot of prep and maybe difficult for me to get to, so maybe I just wouldn't do it. Now I can do a thing and connect with my three-year-old daughter and do this activity that would have been a lot harder for me to have done otherwise because my muse is helping get that set up. So I found it to be that it can kind of help you with personal productivity and doing a lot of the kind of drudgery or basic tasks so that way you can get to the stuff that you really want to do. So different people have different examples of what they really want to do. But I think for a lot of people it's connecting with the people they care about.

但你担心孩子们会这样吗？我的意思是，你的孩子有 AI 伴侣吗？

But do you worry about that with kids? I mean do your kids have AI companions?

还没有。我的意思是，但我的孩子都 10 岁及以下。所以，我认为那还不是一回事。我确实和我的女儿们一起使用 AI 编程智能体来构建东西，她们有一定的自由度自己去做，即使她们不是坐在我旁边做。我认为这也只是一个育儿问题。我的意思是，我们设定了界限，我们确保监督孩子们在做什么。我认为这只是我为自己的家庭所相信的事情。我不认为这会是一个有价值的工具。我认为它会帮助孩子们，就像，你知道，孩子们学习。我的意思是，教育将是一个非常有价值的事情，但这不是 Meta 正在直接做的事情，但我认为这将是 AI 的一个有价值的用途。

Not yet. I mean but my kids are 10 and under. So, I think that that's not quite a thing yet. I do work with my daughters to use AI coding agents to build stuff and they kind of have some leeway to go do that on their own even when they're not like sitting with me doing it. I think that this is kind of also just a parenting question. I mean, we set our boundaries and we make sure that we oversee what the kids are doing. And I think that that's just something that I believe in for my own family. I don't I think this is going to be a valuable tool. I think it's going to help kids and like, you know, kids learn. I mean, education is going to be a really valuable thing over time, but that's not a thing that Meta is working on super directly on any of this stuff, but I think it will be a valuable use for AI.

全息图、更小的面部计算机、看起来像普通眼镜的科技眼镜、在你睡觉时为你工作的 AI 助手。这是一个非常雄心勃勃的赌注，很明显扎克伯格正在再次重塑公司，这次没有改变它的名字。问题不仅仅是你是否会购买我看到的那些令人印象深刻的新玩意儿。而是你是否会接受扎克伯格对未来的愿景，并足够信任这家公司来驱动这一切核心的 AI。

Holograms, smaller face computers, tech glasses that look like regular glasses, AI assistants that work for you when you sleep. It's one really ambitious bet, and it's really clear Zuckerberg is reinventing the company again, this time without changing its name. The question isn't just whether you'll buy all those new impressive gadgets I saw. It's whether you'll buy into Zuckerberg's vision of the future and trust the company enough to power the AI at the heart of it all.

我认为你基本上会拥有一个极其能干的助手，它能够理解你关心的事情，并帮助处理许多你不想处理的琐事，这样你就可以专注于你想做的事情。它全天候陪伴你的一部分方式是人们会拥有眼镜或其他一系列设备，可以随身携带。这样他们的智能体既能理解周围发生的事情，又能以一种非常轻量级的方式，不让你脱离当下，不分散注意力，为你提供全天所需的背景信息。

I think you're basically going to have an extremely capable assistant who can understand the things that you care about and help get a lot of the kind of basic things that you don't want to deal with out of the way so that way you can focus on the things that you want. Part of the way that it's going to be there with you throughout the day is people will have glasses or some other set of constellation of devices that they can have with them. So that way both their agent can understand what is going on around them and can in a very lightweight way without taking you away from the moment without distracting can give you the context that you need throughout the day.

当然，信任一个穿着托加袍的小快乐机器人总是更容易。

Of course it's always easier to trust a little jollybot in a toga.

我给一个握把一件托加袍。

I give a grip a toga.

是的，给那里的格利亚一件非常好的托加袍。

Yes is very nice toga for a gria there.