马克·扎克伯格的未来愿景：AI、智能眼镜与元宇宙

这将会是数十亿人都想使用的东西。对你来说，创造事物是什么感觉？永生有可能吗？好，如果你带我进入你的大脑，我们快进到 2030 年，那会是什么样子？这是马克·扎克伯格。22 年前，他建立了一个连接数十亿人的社交网络，永远改变了世界。但现在他决定建造更大的东西。为此，他在元宇宙、智能眼镜和 AI 上下了巨大的赌注。多年来，怀疑者认为这是三个独立且不可能实现的赌博，但他们错过了更大的图景，因为现在这些赌注正在汇聚，创造出一种全新的超级智能。所以今天我们将向你展示这一点，我会问马克一些他从未被问过的问题，并聆听他对未来的愿景，这样你就能领先一步，构建下一个大事物。

This is going to be a thing that billions of people want to use. What does building things feel like to you? Is live forever a possibility? Okay, so if you take me into your head and we flash forward to 2030, what does it look like? This is Mark Zuckerberg. 22 years ago, he built a social network that connected billions of people and changed the world forever. But now he's decided to build something even bigger. To do it, he's making massive bets on the metaverse, smart glasses, and AI. For years, skeptics saw these as three separate impossible gambles, but they missed the bigger picture because right now those bets are converging into creating an entirely new kind of superintelligence. So today we're going to show you that and I'm going to ask Mark questions that he's never been asked before and hear his vision for the future so you can get ahead of it and build the next big thing.

非常感谢你来参加节目。

Thanks so much for coming on the show.

是的，很高兴邀请你。

Yeah, to have you.

是的，不，我很兴奋能做这个。

Yeah. No, I'm excited to do this.

为了准备这次采访，我看了你做过的所有采访。

To prepare for this one, I watched every interview that you've ever done.

好吧，那比我自己做的还多。

Okay, that's more than I've ever done.

非常有趣。很棒。嗯，有两件主要的事情让我印象深刻。第一，你对建造的热爱。感觉你就是最优秀的建造者之一。嗯，第二，你下巨大赌注的能力。

It was a lot of fun. It was great. Um, and like two major things stood out to me. Number one, your like love of building. Feel like you're just like one of the best builders. Um, and then number two, your ability to take like massive bets.

我认为这周就像是所有这些都汇聚在一起的一周。

And I think like this week is kind of the week that all of them are coming together.

那么，我们就从那里开始吧。

So, let's start there.

是的。嗯，是的。我想我们已经在这个东西上工作了很久。我是说，AI 工作，我们作为一家公司几乎从一开始就在做。我是说，最初的新闻feed第一版就是一种机器学习产品，所以是的，有那个版本，然后我们成立了 AI 研究实验室，天哪，差不多 15 年前了，但嗯，但现在我们进入了一个新版本，因为大约一年多前，我们启动了 Meta 超级智能实验室。

Yeah. Um, yeah. I guess we've been working on a lot of this stuff for a long time. I mean, it's uh the I mean the AI work I mean we've kind of been doing as a company I mean almost since the beginning. I mean there's like the first version of newsfeed from the beginning is sort of this machine learning product and so I mean yeah there's there's that version and then the the AI research lab we we started gosh it's like not quite 15 years ago but uh but then you know we're in a new version now because about a little more than a year ago we uh kicked off Meta Super Intelligence Labs.

是的。

Yep.

那是一次疯狂的研究重启，从整个行业引进了很多优秀人才。但很令人兴奋，因为现在，你知道，我们看到模型变得越来越好。我们嗯，下一代即将推出，不是 connect。

And that it was a kind of wild reboot of the research to bring in a lot of great people from across the industry. But it's been exciting because now, you know, we're seeing the models get better and better. We have um the next generation launching soon, not a connect.

嗯，然后我们有嗯，你知道，Muse 个人智能体，你知道，它嗯，到目前为止反响很好。你知道，当

Um but and then we have um you know, the Muse personal agent that is, you know, it's uh the reception so far has been great. You know, when

是的。我是说，当你在构建这些东西时，你并不完全知道。我是说，我们喜欢它。

Yeah. I mean, when you're building this stuff, you don't quite know. I mean, we liked it.

是的。

Yep.

嗯，你知道，甚至从今年早些时候开始，你知道，就像在家里用胶带拼凑出我自己版本的 Open Claw，并了解这如何工作，以及嗯，什么能让这成为一种神奇的体验，可以提供给任何人。是的，就像我们基本上，我一看到，你知道，我们基本上，你知道，Nat 嗯和 Alex 和我坐下来，我们说，好吧，这是一种神奇而奇妙的体验，如果我们能提供一个版本，让那些不一定技术足够好、不想自己安装 Mac Mini 或进入终端设置或在出问题时调试的人也能使用，嗯，那么我认为这将是数十亿人想要使用的东西。所以从那时起我们基本上一直在为此努力。你知道，让模型专门为此调优。嗯，不仅构建智能体和脚手架，而且有效地构建技术，让每个智能体都有自己的计算机，对吧？我们构建了整个 Muse 安全虚拟机，嗯，是的，所以我们认为这很特别，内部人们真的很喜欢它，但你永远不知道当你发布东西时人们会如何反应，你知道，偶尔会是一个本垒打，人们一开始就喜欢它，但大多数时候你知道你会得到一些积极的反馈，嗯，你必须迭代一些东西才能真正点击。嗯，这个真的从一开始就点击了。所以，看到这个非常令人兴奋，而且事实上我们才两周，就有数百万人在使用它，嗯

Um, you know, even from just earlier in the year, you know, just kind of duct taping together like my own version of Open Claw at home and getting a sense of like what how does this work and like um what would make this a magical experience that you could give to anyone. Yeah. like we like basically as soon as I I saw that you know we we basically you know Nat um and Alex and I sat down and we were like okay this is an amazing and magical experience and if we can deliver a version of this that just works to people who are not necessarily technical enough to want to go install their own Mac Mini or get in a terminal and set it up or debug stuff when when it's not working. um then I think that this is going to be a thing that billions of people want to use. So we've basically been working on that over time since then. You know, getting the models to be tuned specifically for it. Um, building not just the agent and the scaffolding, but effectively the technology to give every agent its own computer, right? We built this whole Muse secure VM and um yeah, so we thought this was pretty special and internally if people have really liked it, but you never really know when um when you ship something how people are going to respond and you know every once in a while it's a it's a home run and people love it out of the gate, but most of the time you know you get some feedback that's positive um you have to iterate on a few things before they they kind of really click. Um this one just really clicked right out of the gate. So, it's been it's been very exciting to see and and the fact that we're like what like 2 weeks in and there's millions of people using it is um

我不知道，这相当罕见。我是说，每隔几年我们就会得到这样的东西，但嗯，这绝对是，你知道，建立公司的有趣部分

I don't know this is pretty rare. I mean, every few years we get something like this, but it's um this is definitely you know the fun part of building a company

100%。是的。我觉得你留在游戏中并不断尝试的能力令人钦佩，而且还有这么多成功，这很棒，而且你做过一次采访，我想大概是三年前和 Joe Rogan 的，在最后你谈到有一天你会戴上眼镜，然后一个 AI 智能体也会出现。所以，我觉得 Muse 已经具身化非常聪明，因为感觉那将会发生。

100%. Yeah. I feel like your ability to like stay in the game and keep trying things is admirable and like also so many hits like it's it's great and there was an interview that you did I think it was like three years ago with Joe Rogan where at the end you were talking about one day you'll like throw on the glasses and then like an AI agent will come in as well. So, I feel like it's so smart that Muse is already embodied because that feels like that was going to happen.

是的。我认为这也让它变得友好和可爱。

Yeah. I think it also just makes it friendly and cute.

是的。同意。

Yeah. Agree.

所以，部分原因是，我认为很多人谈论 AI 时，好像它是什么超级可怕的东西。是的。

So, part of this is like I think so many people talk about AI as if it's this like super scary thing. Yeah.

而且

And

我不知道。我是说，我认为它应该只是有用和有趣，嗯，所以我认为那个具身化，嗯，项目早期的一位设计师组合了那个角色，嗯，不知为何他们一直试图迭代它，但第一个就是最好的。所以就像，有时候有人说，“哦，它必须是蓝色的，因为是 Meta。”然后说，“不，我觉得这个家伙真的，我觉得你第一次就做对了。”所以是的。不，这只是有趣。这是一个很好的点缀。

I don't know. I mean, I think it it should just be useful and fun and um so I I I thought that that embodiment um one of the designers early on in the project put together that character and um and for whatever reason they kept on trying to iterate on it, but the first one was just the best. So there were like like some at some point someone was like, "Oh, it's got to be blue because it's meta." And it's like, "No, like I think this guy is really like I think you nailed it on the first try." So yeah. No, it's just fun. It was a good touch.

如果你试图让它非常擅长个人事务，而不是像通用智能模型，这会如何改变你训练模型的方式？

How does it change how you like train the model if you're trying to make it like really good at personal stuff versus like a general intelligence model?

嗯，我认为现在的一件事基本上是，普遍地让模型作为智能体变得优秀。

Well, I think one of the things right now is basically there's generally making the models good as agents.

好的。

Okay.

嗯，所以我认为在过去一年里，最大的事情基本上是这些编码智能体，其中嵌入了两个想法。有编码的专业知识，以及普遍地成为一个好的智能体。我们的策略。编码很重要，因为即使一个人不认为他们在写代码，你的 muse 也在为你写代码，在后台一直做所有这些事情。所以，能够编码很重要。

Um, so I think for the last year the biggest thing has basically been these coding agents and there are two ideas that are sort of embedded in that. There's the expertise at coding and there's the generally being a good agent. and our strategy. Coding is important because even if a person isn't thinking that they're writing code, your muse is writing code for you to do all this stuff in the background all the time. So, so being able to code is important.

但我们基本上采取的策略是，我们认为智能体应该首先是一个伟大的智能体

But we basically took the strategy that we thought that the agent should be primarily a great agent

而且嗯，而且

and um and

所以我们专注于这一点。

so we focused on that.

还有一堆事情，我觉得在打造个人智能体时比打造企业软件产品时更重要，对吧？比如，如果你在做 Claude Code，你的模型不需要有任何分寸感，比如什么该透露、什么不该透露。你怎么看这一点？

There's also a bunch of things that I think are just more important when you're building a personal agent than if you're building an enterprise software product, right? So for example, if you're building Claude Code, your model doesn't need to have any sense of discretion, like what to divulge and not divulge. How do you think about that?

嗯，我觉得对 Muse 这样的东西来说，这是非常重要的一点，对吧？所以你必须把这一点训练进模型里。就像任何其他能力一样，它需要擅长知道——你告诉它很多事情，然后你希望它去实现你的目标。但假设你想让它帮你在餐厅订位，你在找一家好餐厅，但也许你有某种过敏，而你可能不想透露这个过敏，或者你怀孕了之类的。但你希望 Muse 在你告诉它的时候知道。你会希望它能完成任务，同时尽可能少地透露信息。所以这是一项特定的技能，它有点像人们具备的基本社交技能或常识。但如果你只是在做一个编程智能体，大多数公司并没有把这一点训练进去。所以这就是我们能做的事情，因为我们在做全栈，对吧？因为我们在训练模型。我们不是拿别人现成的模型，然后试图在它周围搭一个脚手架和一个智能体。我们专门训练整个模型，让它擅长这一点。显然，为了具备通用智能，它需要在很多方面都很擅长，但它在这方面很擅长。然后你构建智能体以及围绕它的所有细节——从记忆、框架，到它醒来并检查的心跳机制：好，这些是我知道的关于你的目标。有没有什么我能做的来帮助推进其中任何一个？我们有一整个团队专门负责制作头像的实时动画，因为这不只是默认头像。你可以做任何你想要的头像，然后它就会自然地动起来，非常棒。我们正在推出这个语音模式，你可以进行实时语音聊天，你的那个角色就在那里，不管是什么。所以所有这些细节，我觉得基本上都来自于做全栈，把模型和产品一起做。

Well, I think that's a very important thing for something like Muse, right? So you have to train that into the model. Just like any other capability, it needs to be good at knowing — you tell it a lot of things, and then you want it to go achieve your goals. But let's say you want it to go get you a reservation at a restaurant, and you're trying to find a good restaurant, but maybe you have some allergy, but you may not want to disclose the allergy, or if you're pregnant or something like that. But you want the Muse to know if you tell it. And you're gonna want to be able to do the job while disclosing as little information as possible. So that's a specific skill, and it's something that's kind of basic social skills or common sense that people have. But if you're just building a coding agent, most of these companies haven't trained that in. So that's the type of thing that we can do because we're doing the full stack, right? Because we're training the model. We're not just taking someone else's model off the shelf and trying to build a scaffold and an agent around it. We trained the whole model specifically to be good at that. It obviously needs to be good at a lot of things in order to be generally intelligent, but it's good at that. And then you build the agent and all of the details around it — from the memory, the harness, the way that it — the heartbeat that it wakes up and checks: okay, here are the goals that I know about you. Is there anything I can be doing to help advance any of those? Now, we had a whole team that was just focused on making good real-time animations of the avatars, because it's not just the default avatar. You can make whatever avatar you want and then it just naturally animates, and it's awesome. And we're rolling out this voice mode and you can have a live voice chat and your guy is there, whatever it is. So all these details, I think you basically just get from being full stack and doing all the models and the product together.

百分之百。我觉得另一个重要的事情是虚拟机。你能解释一下为什么它重要，以及它解锁了什么吗？

100%. And I feel like the other big thing is the virtual machine. Can you explain why that matters and what it unlocks?

是的。我知道有几个方面。基本上，一个智能体为了能为你做事，需要一个地方来存储你的信息。我们认为它不应该只是和其他所有人的信息放在一个池子里，对吧？所以你想要属于你自己的东西。我想一种理解方式是，你的智能体会知道很多关于你的敏感信息。当人们开始接触像 OpenCloud 这样的智能体时，他们基本上会在家里弄一台自己的 Mac Mini。所以我们基本上有了这个想法：好吧，很多人不会想买一台 Mac Mini 或者去设置它。我们能创造的最好的体验是什么，能有效地近似那种效果？你只需要下载一个应用、注册，就能获得一台为你配置好的电脑，供你的智能体使用、存储你所有的东西、保证安全，并围绕它设计安全模型。这样它就能近似于在你桌子底下有一台自己的电脑，即使是像我们这样的公司也无法——如果你使用的是 Muse 机密虚拟机，这是我们正在做的另一件事，我们甚至无法看到你虚拟机上的内容。我的意思是，我们围绕这个建了所有这些东西——安全凭证存储，这样当你在处理密码之类的东西时，你的智能体不需要能看到它。它只需要在你要求它登录某个东西时能够把它插进去，而且只在你要求它这么做的时候。它的设计方式应该是让那些信息不是随便就能访问的，因为，怎么说呢，意外总会发生，对吧？有人可能会试图黑进这个东西，或者可能出现问题。所以基本上你要这样设置，让智能体无法访问那些信息，让 Meta 也无法访问那些信息。所以给每个智能体一台自己的电脑，并让它尽可能安全，这是一项非常基础的技术，必须被构建出来，才能给 Muse 提供它需要的、能够去帮助人们实现目标的能力，以及在这个领域成为世界级、行业领先产品所需的安全性和隐私性。

Yeah. So, I know there's a few things. Basically, an agent, in order to be able to do stuff for you, needs a place to store your information. And we don't think it should just be in a pool with everyone else's information, right? So you want your own thing. I guess one way to think about this is your agent is going to know a lot of sensitive things about you. A lot of when people started getting into agents like OpenCloud, they basically got their own Mac Mini thing at home. And so we basically had this idea, which is: okay, a lot of people, they're not going to want to buy a Mac Mini or set it up. What's the best experience that we can create that effectively approximates that? You can just get an app and sign up and you get your own computer provisioned for your agent to work with, to store all your stuff on, to be secure, and to design the security model around that. So that way it kind of approximates having your own computer under your desk, that even a company like us isn't going to be able to — if you're using the Muse confidential VM, which is another thing that we're working on now, we won't even be able to see what is on your virtual machine. I mean, there's all these things that we built around that — the secure credential store, so that way when you're doing passwords and stuff like that, your agent doesn't need to be able to see that. It just needs to be able to plug it in when you ask it to log into something, but only when you ask it to do that. It should be designed in a way where that information isn't just accessible because, I don't know, accidents happen, right? Someone might try to hack into the thing, or there could be an issue. So basically you want to set it up so the agent just doesn't have access to that, and that Meta just doesn't have access to that. So giving every agent its own computer and making it as secure as possible is just a very fundamental piece of technology that had to get built in order to give Muse both the power that it needs to be able to go help people with your goals, and the security and privacy to be a kind of world-class, industry-leading product in that area.

所以这是否意味着，你使用 Muse，它连接着一个虚拟机，但所有数据都像 WhatsApp 上那样加密？人们该怎么理解数据实际存在于虚拟机中时是什么样子？

So does that mean like you use Muse and it has a virtual machine that is connected to, but all the data is encrypted in the same way that it's encrypted on WhatsApp? Like how can people think about what the data actually looks like when it's living in the virtual machine?

是的。所以我们基本上做了两个版本。Muse 安全虚拟机有所有这些不同的隐私功能，包括——我们构建了这整套 Sentinel 智能体架构。所以你有一个正常的 Muse 智能体在为你做事，然后你基本上有一个我们称为 Sentinel 的安全智能体，它只是在监控进出你的 Muse 的数据。所以如果有东西进来，有人试图破坏安全，它就会直接切断，让它无法发生。如果它认为你的 Muse 要采取一个你应该参与决策的行动，它基本上会覆盖你的 Muse，并把它触发回用户那里请求权限。就像，你想让你的 Muse 能做这件事吗？这整套系统星座，再加上安全凭证存储，再加上好几层纵深防御的安全措施——这些构成了 Muse 安全虚拟机。我们还在做另一件事，Nat 和我专门招募了 Moxie Marlinspike，就是和我们一起做 WhatsApp 加密的那个人，来设计 Muse 机密虚拟机。它背后的想法是，它在其他一切之上——它是一个虚拟机，你会得到一个加密密钥。这样 Meta 甚至无法访问其中的内容。

Yeah. So we basically built two versions. Muse Secure VM has all these different privacy features, including — we built this whole Sentinel agent architecture. So you have your normal Muse agent that's doing stuff for you, and then you basically have this security agent that we call Sentinel that is just monitoring data going in and out to your Muse. So if something comes in and someone's trying to compromise the security, it'll just cut it off and make it so it doesn't happen. If it thinks that your Muse is going to take an action that you should be in the loop on, it will basically override your Muse and trigger that back to the person to say, for permissions. It's like, do you want your Muse to be able to do this? This whole kind of constellation of systems, and that plus the secure credential storage, plus just several layers of security through depth — there we make up the Muse Secure VM. We're also working on this other thing that Nat and I specifically recruited Moxie Marlinspike, the guy who we worked with to do the WhatsApp encryption, to design the Muse confidential VM. And the idea behind that is that it's on top of everything else — it is a virtual machine where you get an encryption key. So that way Meta can't even access the contents of that.

实现起来要难得多，因为显然如果 Meta 无法访问其中的内容，调试和让系统运转就会困难很多。所以把它推出来花了我们一点时间，但很快就会上线，基本上会达到人们已经习惯的 WhatsApp 以及我们一些最安全产品的安全标准。

It's much harder to implement because obviously if Meta can't access what's within it, then it's a lot harder to debug and get the system working. So it's taking us a little bit of time to get that shipped, but that's coming soon, and it's basically going to be the standard of security that people are used to for WhatsApp and some of our most secure products like that.

那这样做的优势只是心理上让人觉得更安全，还是有实际的益处？

And is the advantage of that just psychologically people will think it's more secure, or are there actual benefits?

不，不是。我认为安全性是件大事。如果我们想近似实现把智能体放在你桌下的本地机器上——那么本地机器在你桌下意味着什么？意味着没有任何公司能访问它。好吧，假设 Meta 想给你提供这项服务，我们怎么才能让你拥有同样级别的隐私和安全保证，让任何公司——在这个例子里是 Meta，或者任何试图黑进我们的人，或者如果你身处某个你不信任政府的国家，政府可能试图让我们做些什么——都无法进入，因为我们没有访问权限。我认为这真的很重要，这也是很多人对 WhatsApp 感到放心的原因。

Well, no. I mean, I think that security is a big deal. If we're trying to approximate having an agent with a local machine under your desk — so what do you get when you have the local machine under your desk? That means that no company has access to it. So, okay, let's say Meta wants to offer you the service. How do we make it so you have that same level of privacy and security guarantee that no company — in this case, Meta, or anyone who tries to hack into us, or if you're in some country where you don't trust the government, if the government might try to get us to do something — it's like we can't get into it because we don't have access to it. I think that really matters, and that's a lot of why people feel comfortable with WhatsApp.

完全同意。

Totally.

所以我认为这对隐私、安全和信任来说是实实在在的。如果你要有一个非常了解你的智能体——我猜我们几乎所有人都会拥有——我是说，快进 5 年，每个人都会有一个真正深入了解你的目标和你周围一切、能帮你把事情办成的智能体。而要做到这一点，我认为在隐私和安全上达到那个水平并做到行业领先非常重要。我们从一开始就想这么做。

So I think that is a real thing for privacy and security and trust. And if you're going to have an agent that knows a lot about you — which I would guess pretty much we all are going to have — I mean, I think you fast forward 5 years, like everyone is going to have an agent that really intimately understands your goals and everything around you and can help you get stuff done. And in order to do that, I think that level and being industry-leading in privacy and security is very important. We wanted to do that from the beginning.

很酷。是的。关于你还有件很有意思的事，你大学学的是心理学。是的。

That's cool. Yeah. Something also very interesting about you is that you studied psychology in college. Yeah.

我在那儿待的时间很短——读了两年——但我觉得它某种程度上影响了你构建的很多东西。

When I was there for a very short period of time — took two years — but I feel like it's kind of informed a lot of what you built.

嗯，我很好奇，当我们看基准测试时，很多时候我们会说，这个模型超级聪明。但感觉就像我们选朋友时，我们肯定是因为他们聪明才选他们，但也肯定是因为喜欢他们的能量或喜欢和他们相处。你如何考虑塑造模型和它的个性？

Um, and so I'm curious, like when we look at benchmarks a lot of times we'll be like, the model is super smart. But it also feels like when we pick our friends, we definitely pick our friends because they're smart, but we also definitely pick them because we like their energy or being around them. How do you think about shaping the model and the personality?

是的。我认为理想情况下，它应该足够有适应性，能够契合不同人的能量，对吧？我觉得这其实是行业里很多人搞错的地方。我认为其他很多实验室都专注于，比如，我们怎么把个性做对，而我从没觉得它是一成不变的。这也是我如此坚信开源、坚信人们能够自定义东西的部分原因，也是为什么我们把 Muse 设计成非常个人化的东西。你可以个性化它、自定义它。你不仅可以按自己的喜好设置头像和声音，而且在你第一次注册时，它最先问你的问题之一就是，你希望我的基本个性是什么样的，而且你随时可以编辑。所以我们努力把模型设计得非常可引导，这样你就可以定义你想如何与它互动。我认为这非常重要。你刚才问我们怎么让它成为一个好的个人智能体，我认为在个性上具备这种适应性是其中很大的一部分。

Yeah. I mean, I think ideally it will be adaptable enough to be able to fit different people's energy, right? I think this is actually something that I think a lot of people get wrong in the industry. A lot of the other labs I think have focused on, like, how do we get the personality right, and I just never thought that it's one thing. And this is part of the reason why I believe so strongly in open source and people being able to customize stuff, and also why we designed Muse to be this thing that is very personal. You can personalize it and customize it. You can set up not just the avatar and the voice the way that you want, but right when you sign up for the first time, one of the first things that it asks you to is, like, what do you want my basic personality to be, and you can go and edit it at any time. So we try to design the models to be very steerable, right, so that way you can define how you want to interact with it. And I think that's a really important thing. I mean, you were asking about how do we make it so that this is good as a personal agent. I think that having that adaptability around the personality is a big part of it.

如果要描述它的特质，你的智能体是什么样的？

What's yours like if you were to describe its traits?

我想我只是告诉它要相当直接、像工作伙伴一样。所以我不知道。我觉得它挺有趣的。我之前的一个版本相当讽刺幽默，但不知道。现在这个化身我觉得只是更直截了当一些。

I think I just told it to be pretty direct and worklike. So I don't know. I think it's kind of fun. An earlier version I had was pretty sarcastic and humorous, but I don't know. This incarnation of it I think is just a little more straightforward.

是的，嗯，我的那个，他就像默认的 Muse 角色，但我给他穿了托加长袍，他就有一种，嗯，极其低沉、低沉到有点好笑的声音，和他互动就是件很有趣的事。

Yeah, it's um yeah, my guy has he's like the normal default Muse character, but I give him a toga and he has and he just has this like kind of like um extremely deep to the point of being humorous voice and it's just like a fun thing to engage with.

是的。

Yeah.

是的。我觉得要搞笑你必须非常聪明。就像我认为很多人不明白，喜剧演员是社会中最聪明的人之一。你得快。是的。你得机智。嗯，这绝对是你拥有的特质之一。嗯，我很好奇——

Yeah. I feel like to be funny you have to be really smart. Like I think a lot of people don't get like comedians are some of the smartest people in society. You gotta be fast. Yeah. You gotta be like quick-witted. Um one of the traits you definitely have. Um I'm curious like —

我是说你绝对有。我喜欢看所有的采访。你总是很棒。呃，你，非常友善。

I mean you totally do. I like watching all the interviews. You're always great. Uh you and the very nice.

我是认真的。嗯，但说到这个，在 Theo 的采访里你们谈了很多关于技术的下一个前沿以及它最终指向什么。嗯，我经常想到，AI 有过那种寒冬期，人们会说这永远不会发生，不会有突破到来。嗯，感觉元宇宙也经历过好几次，人们觉得那是个最终没有实现的赌注。我昨天试了新的全息功能。

I mean it. Um but okay on that note in the Theo interview you guys talk a lot about like this next frontier of like the technology and what it's all building towards. Um and I think about a lot like with AI there were these like winter seasons where people be like this is never going to happen. Like there's no breakthroughs are coming. Um it feels like the metaverse has had several of those where people kind of think it's a bet that then doesn't come to fruition. I tried out yesterday the new hologram feature.

我知道。挺酷的。

I know. It's pretty cool.

是的。在你们做的 Lex 采访里，我感觉当时要花 11 个小时的设置才能把你的脸放进去，现在只要 3 分钟。

Yeah. In the Lex interview you guys did, I feel like it was like an 11 hour setup to get your face in there and now it's 3 minutes.

事情就是这样。

That's how it goes.

那接下来会有什么，以及是什么样的突破把我们带到了这一刻？

What's coming with that and like what were the breakthroughs that led us to this moment?

是的。所以，总的来说，元宇宙的发展，有意思的是，当我们创办 Reality Labs 时，我们一直认为，你知道，你会得到看起来正常的眼镜，随着时间推移，它们既能提供这种沉浸式的临场感，又能成为出色的 AI 设备，因为眼镜是唯一一种形态，可以让设备看到你所看到的、听到你所听到的，整天在你耳边和你说话，并最终显示图像。但你知道，10 或 15 年前，我只是假设我们会先有全息影像，然后才有非常先进的 AI。但我觉得这就是技术树运作方式的有趣之处，我们实际上先得到了 AI 和像个人超级智能这样的东西，然后才真正得到让全息影像普及、便宜到人人都能用的技术。所以我不会预测到这一点，但我很高兴我们两者都在做。

Yeah. So, I mean, overall the development of the metaverse, it's interesting because when we started Reality Labs, we always thought, you know, you're going to get like normal looking pairs of glasses and over time they're going to be able to do both deliver this immersive sense of presence as well as being a great AI device because glasses are the only form factor where you can let the device see what you see and hear what you hear and talk to you in your ear throughout the day and eventually display images. But you know 10 or 15 years ago I just assumed we'd get holograms before we got very advanced AI. But I think it's this funny thing about the way the tech tree has worked that we actually are just getting AI and like personal super intelligence before we actually get the technology to make holograms ubiquitous and affordable enough for everyone to use. So I wouldn't have predicted that, but I'm glad we're working on both.

我们在眼镜上做的所有这些工作，确实让我们现在处于非常有利的位置，因为所有这些 AI 智能体都已经准备好了。我们在 Connect 大会上宣布的很多内容，就是把 Muse 和大量 AI 功能带到眼镜上，我觉得人们会非常喜欢。这是件大事。但关于临场感的一些东西，我们还在研究。在幕后进展稍慢，因为我们已经把大部分精力转向为眼镜构建所有 Muse 和 AI 功能。

The fact that we are doing all this work on glasses has definitely made it so that we are very well positioned now that all these AI agents are ready. A lot of what we're announcing at Connect is bringing Muse and a lot of AI functionality to the glasses in a way that I think people are really going to love. That's a big deal. But some of the stuff around presence, we're still working on it. A little bit slower burn in the background because we've shifted most of our focus to building all the Muse and AI features for the glasses.

我们有一个项目已经做了一段时间，就是实时逼真虚拟形象。是的，就像你说的，三四年前，你需要一个完整的房间来从各个角度扫描某人，然后你才能体验它，但你需要自己的 GPU，像是企业级的配置才能做到。所以我们在几年前为 Lex 播客做演示时搭建了那个。我认为这些事情的发展方式是，首先你试着把它做好，然后你试着让它更高效。所以现在基本上我们让它在这副 VR 眼镜里运行了，这是第一副能提供惊人 VR 体验的眼镜，而不是一个巨大的头显，因为我觉得很多人不想要一个完整的头显。但现在你基本上只需要几张照片就能设置你的虚拟形象。它发展得如此之快真是令人惊叹。

One of the projects that we've had for a while has been the real-time photorealistic avatars. Yeah, like you said, three or four years ago, you needed a whole room in order to scan someone from all the angles, and then you could experience it, but you needed your own GPU, like an enterprise-grade setup to be able to do it. So we set that up as a demo for the Lex podcast that I did, maybe a couple of years ago. And I think the way that these things work is first you try to just make it good and then you try to make it more efficient. So now we basically have it working in this pair of VR glasses, which is the first amazing VR experience that is in a pair of glasses, not a big headset, because I think a lot of people don't want to have a whole headset. But you basically just set up your avatar with a few photos at this point. It's pretty amazing how far it's come.

是的。而且它还基于声音。在演示中真的很有趣。基本上它让我笑和互动，然后理解我的脸会如何随着音频轨道移动。在那期播客中你谈到，那些可能不太善于表达的人实际上会在虚拟世界中想要更善于表达。你怎么看待人们区分他们的虚拟自我和现实生活中的自我？

Yeah. And it's also based on the voice. Like it was so interesting in the demo. Basically it had me laugh and interact and then understand how my face would move with the audio track. It was interesting in that podcast you talked about how people that maybe are less expressive will actually want to be more expressive in the virtual world. How do you think about people delineating between their virtual self and their in-real-life self?

我想是的。我认为我们对于这方面的社会学和心理学理解还处于早期阶段。我的意思是，我认为人们会……我认为人们如何看待自己以及他们想要投射的形象，往往与他们实际的样子有所不同，对吧？从社交网络诞生之初，人们就在精心策划个人资料照片。我认为我们在 Muse 的虚拟形象中看到了这一点。这与其说是策划你自己，不如说是策划你想要交谈的那个人。所以，是的。不，我认为这会非常有趣。但我认为，每当你赋予人们表达自我的能力时，你希望它能够捕捉到他们，但你也希望它……它既是一种表达形式，也是一种传达形式。它不只是一个纯粹的镜像。它是一种表达形式。所以，是的，我认为我们想要构建能够同时做到这两点的东西。这总是一个迭代循环，你想看看人们如何使用这些东西，然后你去改进它。但我认为，经过多年的努力，我们现在基本上处于起点，现在我们要真正把它带入产品中，第一次我们将拥有这些高质量逼真的虚拟形象，你可以在手机上使用，也可以在 VR 中使用。所以，是的，我非常感兴趣看看这会如何发展。

I think so. I think we're still pretty early in understanding the sociology and psychology around that. I mean, I think people are going to... I think that often how people perceive themselves and what they want to project is somewhat different from what they actually are, right? From the beginning of social networks, people have curated profile photos. I think we're seeing a version of this with the avatars with Muse. And that's less curating yourself, but it's like curating the guy that you want to talk to. So, yeah. No, I think it's going to be very interesting. But I think whenever you're giving people the ability to express who they are, you want it to be something that can capture them, but you also want it... it is also a form of expression as much as it's a form of conveying. It's not meant to just be a pure mirror image. It's meant to be a form of expression. So, yeah, I think we want to be able to build something that can do both. And this is always this iterative loop where you want to see how people use the stuff and then you go work on it and make it better. But I think we're basically at the starting line of this now after many years of work where now we're actually going to get it into products and for the first time we're going to have these pretty high quality photorealistic avatars that you'll be able to do on your phone, you'll be able to do in VR. So yeah, I'm quite interested to see how that goes.

好的。那么，如果你带我进入你的大脑，我们快进到 2030 年，如果一切顺利，那会是什么样子？比如使用那种技术的平均一天，在全息投影方面？比如全息投影加上 Muse，它们如何与眼镜融合。

Okay. So, like if you take me into your head and we flash forward to like 2030, what does it look like if that goes right? Like what's the average day with that technology on the hologram side? Like hologram plus Muse, how they're kind of converging with the glasses.

是的。所以，我对元宇宙愿景的思考始终是关于有效地将物理世界和数字世界融合在一起。基本想法是，我们周围有一个伟大的物理世界，但我们也有一个伟大的数字世界，它已经在互联网上构建了 20、30 年，有这么多内容。这非常了不起。而我们访问它的主要方式要么是坐在办公桌前，要么是通过口袋里的小屏幕。这似乎从根本上非常受限，对吧？我认为理想的版本应该是物理世界和数字世界的无缝融合。我认为思考这个问题的方式是，好吧，现在我们两个人在物理上在这里。你可以想象未来的一种版本，我们中的一个是全息投影，但你会有同样的临场感，感觉就像和某人物理上在一起。这与 Zoom 通话之类的非常不同，对吧？但虚拟现实的全部意义在于它提供了这种临场感，让你真正感觉自己和另一个人或另一个地方在一起。所以，你可以用全息投影做到这一点，并且可以以不同的方式混合。所以，你可以有……比如我可以和朋友们有一个扑克之夜，我们中有一半人在现场，几个人通过全息投影加入，可以打牌，也许牌是全息投影的一部分，这样我们就可以给不在现场的人发牌。但然后我认为 AI 也可以具身化并在那里，也许扑克之夜是最好的例子，但在工作中我认为这也说得通。我们已经和所有这些智能体一起工作，比如我一直使用不同的编码智能体来构建东西。所以你想，好吧，你有具身化，也许你在一个群聊频道里，有几个人和几个智能体，你给智能体指示要做什么。但也有物理版本，有时人们聚在一起开会，也许智能体也应该在那个会议中。那么它们会如何出现？就像，好吧，沙发上还有几个额外的位置，它们可以过来，作为全息投影在那里，或者无论具身化是什么。你可以得到 Muse 的可爱小角色，或者你创造的任何龙或任何疯狂的东西。

Yeah. So, the way that I think about the metaverse vision was always about effectively bringing together the physical and digital worlds. So the basic idea is we have this great physical world around us but we also have a great digital world where it's just been built on the internet for 20, 30 years at this point and there's all this content. It's pretty amazing. And then the main ways that we access it are either by sitting at a desk or through this small screen in our pocket. And that seems like fundamentally very limiting, right? I think the ideal version of this would be a seamless blending of the physical and digital world. And I think the way to think about this is, okay, so right now the two of us are here physically. You could imagine a version of this in the future where one of us is a hologram, but you kind of have the same sense of presence, feeling like you're there with someone physically. It's very different from a Zoom call or something like that, right? But the whole thing about virtual reality is that it delivers this sense of presence where you actually feel like you are there with another person or in another place. So, you could do that with a hologram and you can have that mixed in different ways. So, you could have like I could have a poker night with my friends where half of us are there physically and a few people join through a hologram and can play cards and maybe the cards are part of the hologram so we can kind of deal in the people who are not there physically. But then I think AI can also just be embodied and be there too, which maybe for poker night is the best example, but like in work I think that makes sense. And we already are working with all these agents like I use different coding agents to build stuff all the time. So you think about like, all right, you have the embodiment where maybe you're in a group chat channel where there's a few people and a couple of agents and you're giving the agents directions for what to do. But there's also the physical version of that where sometimes the people get together in a meeting and maybe the agents should be in that meeting too. So how would they show up? It's like, all right, there's a couple of extra spots on the couch and they can just come and be a hologram and be there, or whatever the embodiment is. You get the cute little character from Muse or whatever dragon or whatever crazy thing you create.

我觉得未来这会变得非常自然。所以我认为这些东西会融合在一起。显然，仅靠文本你就能让 AI 走得很远。但我认为，随着一些实时系统的出现，能够看到角色具身化以及实时互动——人们喜欢这样。很可爱，很有趣。你会觉得你真的在那里和它对话。很酷。所以我认为我们最终也会希望在眼镜体验中实现这一点。

I think that's going to end up feeling quite natural in the future. So I think these things will come together. You can obviously get very far with the AI with just text. But I think what we're starting to see with some of the real-time systems around being able to see the character embodied and the real-time interactions — people like that. It's cute. It's fun. You feel like you're there talking to the thing. It's cool. So I think we're going to eventually want that in the glasses experience, too.

是的。这很有意思，因为你在之前的一次采访中说过，你觉得科技行业常常忘记了乐趣。我认为让智能体在那里也会让它感觉更真实，就像你在外包工作。当你看到 Muse 在打字之类的，会感觉真的有事情在发生。我觉得很多——我觉得你在这方面做得很好，浏览器能够看到正在发生什么。你认为是否存在这样一个世界：你戴着眼镜，然后控制你的电脑？就像你有 Muse，它在你的电脑上做事？

Yeah. It's interesting because you said in a previous interview that you feel like the tech industry often forgets fun. And I think having the agent be there also just makes it feel more real, like you're outsourcing work. When you see Muse typing and stuff, it feels like something is actually happening. And I think a lot of — I think you did a great job with that, with the browser being able to see what's going on. Do you think there's a world in which you wear the glasses and then you control your computer? So like you have Muse and it's doing stuff on your computer?

哦，是的，绝对可以。我是说 VR 已经能做到这一点了。你基本上可以在任何地方坐下来——咖啡馆什么的——你调出你的工作站，有六个显示器，你在编程，所有这些。然后在眼镜方面，最流行的眼镜还没有显示屏，这既让我们能把它做得更实惠，让更多人使用，同时我们还在努力把它做到最紧凑的形态。但我们确实推出了 Meta Ray-Ban 显示屏。它们非常受欢迎。那是一个小显示屏。我们还推出了全宽视场全息 AR 的原型版本，我认为那个会非常令人兴奋。所以基本上你会得到整个产品线——我是说这些是纯音频眼镜。它们甚至没有摄像头。它们看起来就像完全普通的眼镜，但你在里面装了 Muse，你可以获得所有音频工具。你可以听音乐、打电话。

Oh yeah, definitely. I mean the VR already does that. You can basically sit down anywhere you want — you got a coffee shop, whatever — you pull up your workstation, you have like six monitors, you're coding, you have all that stuff. And then on the glasses side, the most popular glasses don't have displays in them yet, which both allows us to get it to be more affordable so more people can use it, but also we're still working on getting that into the most compact form factor. But we did launch the Meta Ray-Ban displays. They've been quite popular. It's a small display. We also launched the prototype version of the full wide field of view holographic AR, and that one I think is going to be very exciting. So I think you're basically just going to get this whole lineup — I mean these guys are the audio only glasses. They don't even have a camera. They just look like completely normal glasses but you have your Muse in them and you can get all the audio tools. You can listen to music, phone calls.

我想的是，用那些眼镜，你能和 Muse 对话，同时让你家里的台式机做事吗？

I guess I'm thinking with those, could you be talking to Muse and have your desktop at home doing stuff?

哦，是的，完全可以。其实——我们刚在 Connect 上发布了这个。是的，我是说今天所有眼镜都连接 Meta AI。所以这是一种单轮体验。你给它一个提示，它回复。就结束了。但现在有了 Muse，我们基本上会把它们全部升级到 Muse。所以首先，你不再说“嘿 Meta”了。你可以随便给它起名字，这又是乐趣的一部分。所以我就说“嘿，A Grippa”。然后基本上——是的，然后你就和它说话，它会连接到你的 Muse，你的 Muse 会在你的安全虚拟机里做事，为你完成任务。

Oh yeah, totally. Well, that's actually — we just actually shipped that at Connect. Yeah, I mean we're basically today all the glasses connect to Meta AI. So it's sort of this one turn experience. You give it a prompt, it replies. That's the end. But with Muse now, we're basically going to upgrade them all to Muse. So first of all, you don't say hey Meta anymore. You just get to name it whatever you want, which again is part of the fun. So I just say hey a grippa. And then basically it — yeah, no, and then you just talk to it and it connects to your Muse and your Muse goes and does work in your secure VM and gets stuff done for you.

是的，那太棒了。

Yeah, that's epic.

是的。不，是的，会非常有趣。

Yeah. No, yeah, it's going to be very fun.

所以，好吧，我们现在坐在这里，Muse 进展得非常好，眼镜也做得很好，但大约一年前，我觉得很多人都在说：“超级智能实验室到底怎么了？”在那一刻，你感觉如何？在你脑海里，当事情进展不顺，但你看到长期愿景时，那是什么感觉？

So, okay, we're sitting here now and Muse has gone really well and the glasses have done well, but about a year ago, I feel like a lot of people were like, "What is going on with the superintelligence lab?" In that moment, what does that feel like to you? Like in your head, what does it feel like when things aren't going well, but you see the long-term vision?

嗯，好吧。所以，真正偏离轨道的是 Llama 项目和 Llama 4。Llama 1 作为一个模型相当有趣，它开创了整个开源 AI 运动，我认为这非常强大，也是我们非常自豪的事情。然后 Llama 2 实现了 Scaling（规模扩张）。Llama 3 是一个非常好的模型。它当时几乎处于前沿。然后到了 Llama 4，我们基本上偏离了我们需要走的轨迹。我对此的反思是——每当事情没有按照我认为应该的方式发展时，我都会花很多时间思考为什么会发生这种情况？我们需要改变什么才能让它变得更好？在这种情况下，我的反思是我把整个团队的形态搞错了。我模仿了我们做机器学习工作的方式，比如制作 Instagram 信息流或广告系统之类的。这些团队有数百甚至数千人，可以并行处理很多事情。而我认为，构建这些语言模型，你真正需要的只是一个非常紧密的团队，把它视为一个集体科学项目。所以人不多，这意味着团队中的每个位置都极其宝贵。所以我们最终从 Meta 各处引进了一批最优秀的人加入团队，同时也从行业中引进了许多其他优秀的人，建立了一个全新的团队，我们称之为 Meta 超级智能实验室。所以从我的角度来看，当我们启动 MSL 时，我们知道重启和重建基础设施来构建下一组模型需要一些时间。但我知道我们已经组建了一个伟大的团队，我知道如果我们能让团队凝聚起来并良好运作，那么就会好起来。对我来说，实际上最可怕的时刻是 Llama 4 发布之后，当时就像，哦，我以为我们在这条轨迹上，但我们没有。这算是一个相当大的负面意外。我认为，你知道，当你是一个创业者，你在构建东西时，你会在那些时刻受到考验，因为不可避免地，不是所有事情都会顺利。而定义轨迹的事情是，好吧，如果事情没有按照你想要的方式发展，你基本上如何弄清楚如何向前推进？

Well, okay. So, the thing that really was off track was the Llama program and Llama 4. Llama 1 was quite interesting as a model and it sort of pioneered the whole open-source AI movement, which I think has been very powerful and it's something that we're very proud of. Then Llama 2 scaled. Llama 3 was a very good model. It was almost at the frontier at the time. And then with Llama 4, we basically fell off the trajectory that we needed to be on. And my reflection on this — whenever something doesn't go the way that I think it should, I spend a bunch of time thinking about why did that happen? What do we need to change to make it better? And in this case, my reflection was that the whole kind of shape of the team I had gotten wrong. So I modeled it off of the way that we've done our machine learning work for making like Instagram feed or our ad system or something like this. These teams that have many hundreds or thousands of people who can work on a lot of stuff in parallel. Whereas I think for building these language models, what you really want is just a very tight-knit team that views it as a group science project. So not many people, which means that every seat on that team is extremely valuable. So we ended up bringing a bunch of the best people from across Meta into the team but also bringing a lot of other awesome people from around the industry in and building a completely new team which we called Meta Super Intelligence Lab. So from my perspective, by the time that we were starting MSL, we knew it was going to take some time to reboot and rebuild the infrastructure to build the next set of models. But I knew that we'd pulled together a great team and I knew that if we could gel the team and have that work well, then it would be good. To me, actually the scariest moment was after the Llama 4 launch when it was just like, oh, I thought we were on this trajectory and we're not. And it was sort of a pretty big negative surprise. And I think, you know, when you're an entrepreneur, you're building something, I think you kind of get tested in those moments because inevitably not everything is going to go well. And the things that kind of define the trajectory are like, okay, if something doesn't go the way that you want, how do you basically figure out how to move forward?

是的。我想这也有另一面。当某件事比你预期的好得多时，比如最初的 Muse 发布，你如何确保抓住这个机会？

Yeah. I guess there's the flip side of that, too. When something goes way better than you expect, like the initial Muse launch, how do you just make sure you capitalize on that?

完全正确。现在我们在整个公司都有这个巨大的努力，一开始只是团队在构建产品，但现在每个人都觉得，好吧，这实际上已经准备好进入黄金时段了。所以我们基本上如何扩大这个规模，帮助数亿人体验这个，来自整个公司。所以看到每个人都参与进来，努力确保我们能够扩大 Muse 的规模，这非常酷。

Totally. And now we have this whole big effort across the company where it's like first it was just the team that was working on it to build a product but now everyone is like okay this is actually ready for prime time. So how do we basically grow this and help hundreds of millions of people experience this from across the company. So it's pretty cool to see everyone kind of chip in and work on making sure that we can scale Muse out.

从真正优化所有基础设施以使其良好运行，到尽可能从我们拥有的 GPU 中榨取更多容量。

Everything from really optimizing all the infrastructure to make it work really well, squeeze as much capacity as we can out of the GPUs that we have.

是的。

Yeah.

但然后所有产品团队也是，以不同方式构建 Muse，比如眼镜和，嗯，是的。不，看到这些非常有趣。

But then all the product teams too, kind of building Muse in different ways, like the glasses and, um, yeah. No, it's very fun to see.

是的。我觉得作为创始人，那就像是你为之而活的时刻，一切汇聚在一起。嗯，你如何向公司阐述你的愿景？就像当你同时处理这么多不同的事情时，我觉得你写了很多东西。你的流程是怎样的？

Yeah. I feel like as a founder that's like the moment that you live for where it's like it all comes together. Um how do you articulate your vision to the company? Like when you're working on so many different things like I feel like you write a lot. What's like your process?

是的，写作对我有帮助，嗯，既能提炼我的想法。所以整个夏天我写了这篇长文，嗯，这篇像是“未来属于每个人”的文章，是一份 15 页的文件，它非常有助于我提炼自己关于 AI 周围所有这些不同重要社会话题的哲学，比如我认为什么是好的，与政府互动应该是什么样，我们如何防范人们担心的不同危害，对吧，我们如何让数据中心对社区有益，并且这实际上创造就业而不是消除就业，以及我们可以帮助保护国家安全，我们可以真正减轻人们担心的任何危害，无论是黑客攻击还是生物安全之类的事情。是的，所以这就像是一件非常复杂的事情。我花了很长时间，和很多人交谈，经过了很多草稿，内部很多人，我们辩论、启发，但好吧，这对我来说是一个非常有益的过程。然后最后我们有了这个东西。就像，‘好吧，这就是它。15 页。这就是我们所相信的。’

Yeah, writing is helpful for me, um both to kind of distill what I think. So over the summer I wrote this like long piece, um this like 'future is for everyone' piece that was like a 15-page document that kind of like was very helpful for me to distill my own philosophy on all these different important social topics around AI, like what do I think is good, what should the interaction with governments be, how do we protect against um the different harms that people are worried about, right, how do we make it so that data centers are good for communities and that this actually creates jobs rather than eliminating them, and like that we can help protect national security and that we can, you know, really mitigate um any of the harms that people are worried about, whether that's like hacking or or bio security type things. And yeah, so like it was just kind of like a very complex thing. I took a long time, talked to a lot of people, went through a lot of drafts, a lot of people internally like kind of we debated inspired about it, but okay, it was very, it was a very helpful process for me. And then at the end we kind of had this thing. It's like, 'All right, here it is. 15 pages. This is like what we believe.'

是的。

Yeah.

嗯，然后我将其提炼成一页的版本，作为专栏文章发表。嗯，我们制作了一个关于它的短视频，因为我认为要触达很多人，你通常甚至不想要这种理论论证。你只是想把它归结为你的价值观是什么？你相信什么？你如何传达这一点？嗯，所以所有这些事情都相当重要。呃，但是的，我的意思是，这并不是说你知道，在向公司或世界传达信息方面，这不是一刀切的。我认为不同的时代需要不同的东西，在不同的时刻，某些群体要么只是内在相信你正在做的事情，要么更担心它，需要被引导，或者需要额外的努力来解释为什么那将是有用的。所以，嗯，是的，我认为无论什么，我的意思是，我认为这是经营公司或成为创始人的一部分，就像你不是一遍又一遍地做同样的事情并重复，就像你所处的每种情况都略有不同，有新的挑战，嗯，我认为这就是其中有趣的部分。

Um and then I distilled that down to a one-page version that I published as an op-ed. Um we made a short video about it because I think to like reach a lot of people, you often don't even want like this like theoretical argument. You just kind of want to boil it down to like what are your values? Like what do you believe? How do you communicate that? Um so and that all that stuff is pretty important. Uh but yeah, I mean it's it's not like you know in terms of communicating to the company or the world, it's not like a one-size-fits-all. I think different times call for different things and at different points like certain groups of people either just intrinsically believe in what you're doing or are more worried about it and need to kind of be brought along or need extra effort to explain why that's going to be useful. So, um, yeah, I think whatever I mean I think this is part of like running a company or being a founder is like it's not you don't just do the same thing over and over again and repeat like each situation you're in is like a little bit different with new challenges that and um and I think that that's partially what's interesting about it.

是的。嗯，我认为对你来说，嗯，这种创始人思维似乎不仅限于作为公司的创始人，嗯，无论是你的农场还是学习一项技能。嗯，我只是喜欢建造东西。

Yeah. Well, I think for you also um it seems like this like founder mindset extends outside of just being a founder of the company um to whether it be your farm or learning a skill. Um, I just like building things.

是的。好的。多告诉我一些，比如建造东西对你来说感觉如何？

Yeah. Okay. Tell me more about like what does building things feel like to you?

嗯，我不知道。对我来说感觉如何？嗯，我想这就像是我内在的一种需求。就像我觉得我不是，我觉得不同的人有以不同方式表达自己的需求。就像如果你是作家，那么我觉得你可能觉得需要写作，或者我认为有些人觉得需要被喜欢之类的，他们说我不，我只是需要建造东西。我只是觉得如果我不锻炼我的创造力并建造东西，我会变得脾气暴躁。你必须这样。那不好。我周围的人，我周围的人不想我脾气暴躁。是的。是的。

Um, I don't know. What does it feel like to me? Um, I guess I it's kind of just like an intrinsic need that I have. Like I feel like I'm not I feel like different people have a need to express themselves in different ways. Like if you were a writer then like I feel like the probably feel like I need to like write or I think some people feel the need to like be liked or something and they go I don't I just like I need to like build stuff. I just think if I'm not like exercising my creativity and building stuff I I get grumpy. You have to. And that's not good. Everyone around the people around me don't don't want me to be grumpy. Yeah. Yeah.

学习如何成为一个真正优秀的滑雪者，和学习如何构建一个产品，是同样的技能吗？

Is it the same skill to like learn how to be like a really good skier as it is to like learn how to build a product?

嗯，是的。我认为学习新事物的过程非常相似。好的。嗯，在我的生活中，我尝试过一些我相当不擅长的事情。所以，曾经，比如我一直很不擅长语言。好的。是的，这实际上就是我最初接触拉丁语的原因，就像我在课堂上就是不会说法语或西班牙语。所以最后我说好吧，拉丁语，那很好，我不只是说它，就像我可以翻译它，就像数学一样。是的。嗯，所以然后当我开始经营公司时，我开始做这些年度挑战，其中一个我说让我学普通话。普通话是一门非常难的语言，对吧。就像高中一样，它是真实的。是的。很棒但很难。是的。是的。还有声调那件事，就像，所以我的意思是，我认为它好的原因有很多。比如普莉希拉的奶奶只会说普通话，所以如果我想和她交流，我需要学习它。嗯，但但最大的部分实际上只是它是一个挑战，嗯，我认为所有这些事情你只是出现，没有办法通过思考超越一堆这些事情。我认为你只需要投入时间，你花时间去做，它就会渗入你的大脑。我认为同样的事情也适用于，我不知道，学习武术或学习驾驶直升机之类的。就像这些事情实际上很难理智化。嗯，我认为构建产品，我的意思是，其中一部分更理智。我认为编码你可以以理论方式思考，但我认为构建产品的直觉部分，嗯，我认为你只有通过反复练习才能做到。但然后问题就是你喜欢什么，对吧？因为我认为有些人并不一定有我这样的构建东西的需求。大多数人都有一些需求。是的。我认为关键是你如何为自己找到那是什么，然后引导你的时间去反复练习，成为你想成为的东西中的佼佼者，老实说，我认为这超越了想要。我认为这是一种非常深刻的心理需求或驱动力去做一件事。嗯，我认为将其与什么对齐，并给自己时间让围绕它的经验随时间渗入。嗯，是的。不，我认为那非常特别。

Um, yes. I think that the process for learning new things is pretty similar. Okay. Um, in my life I've tried to do some things that I was sort of quite bad at. So, at one point, like so I was always very bad at languages. Okay. And yeah, that's that's like actually how I got into Latin in the first place is like I just couldn't speak French or Spanish in class. So finally I was like all right Latin that's great I don't just speak it like it's like I can translate it's like math. Yeah. Um so then when I was so then when I started running the company I started doing these yearly challenges and for one I was like let me learn Mandarin. Mandarin is like a really hard language right. It's like high school it's real. Yeah. Amazing but hard. Yeah. Yeah. And like the tones thing it's like it's so I mean there there were all these reasons why I thought it was good. Like I Priscilla's grandma only speaks Mandarin so if I wanted to communicate with her I needed to learn it. Um but but the biggest part was actually just that it was a challenge and um think all these things you just kind of like show up and there's no way to outthink a bunch of these things. I think you just kind of like need to put in the time and you you kind of spend your time doing it and it just kind of like seeps into your brain. I think the same thing is true with like I don't know learning martial arts or like learning to fly a helicopter or something. It's like these are things that they're actually like very hard to intellectualize. Um, I think building products, I mean, parts of it are more intellectual. I think coding you can you can kind of think of in a theoretical way, but I think the the kind of intuitive part about building products, um, I think you you only kind of do that by getting reps on that. But then the question is just like what do you love, right? Because I think some people don't necessarily have the same need to build things that I have. Most people have have some need. Yeah. And I think the key is like how do you find what that is for yourself and then like go channel your your time to like go get the reps at that to become excellent at the thing that you like want to and and honestly I think it goes beyond want. I think it's sort of like this like really like deep psychological need or drive to go do a thing. Um and I think like lining that up with like what um and just kind of giving yourself the time to have the experience around that seep in over time. Um, yeah. No, I think that's pretty special.

我的意思是，这就是我尝试养育孩子的方式：基本上让他们自己去发现对什么感兴趣。但如果他们卡在某个地方——比如我的一个女儿非常喜欢做音乐，但她受不了上钢琴课。我就说：‘好吧，听着，你不需要成为钢琴的世界专家，但如果你想能够写音乐，你就需要对一些乐理以及它是如何运作的有直觉性的掌握。一旦你学会弹钢琴，你就能很容易地拿起吉他，对吧？’是的，所以我认为，就是那种纪律——我的意思是，当你还是个孩子时，有父母帮忙可能有用——但我认为对于成年人来说，就是有纪律地坐下来，让这些东西通过重复渗入你的大脑，我认为这是一个关键部分。

I mean, this is how I try to raise my kids: basically let them figure out what they're interested in. But then if they get hung up on something—like one of my daughters is really into making music, but she could not stand taking piano lessons. I was like, 'All right, look, you don't need to become a world expert at piano, but if you want to be able to write music, you're going to want to have an intuitive grasp of some of the music theory and how this works. And once you learn how to play piano, you're going to be able to pick up guitar really easily, right?' And yeah, so I think that just kind of having the discipline—I mean, when you're a kid, having the parent probably helps—but I think just for grown-ups, having the discipline to sit and let the stuff seep into your brain with the reps is a key part.

同意。我也觉得每个人都想构建。就像我觉得尤其是 Z 世代没有得到——Z 世代有一个坏名声，就是有点低能动性之类的，而实际上我认为人们只是在寻找能点燃他们构建能力的东西。你在哈佛演讲中其实谈了很多。

Agree. I also actually feel like everyone wants to build. Like I think especially like Gen Z does not get—there's a bad rep around Gen Z of being kind of low agency and stuff, and I actually think that people are just looking for something to ignite their ability to build. You talked about it actually a lot in your Harvard speech.

是的。是的。嗯，是的，不，我想——我猜当我说构建时，我主要是在谈论这类产品，比如软件、硬件之类的。但我同意，我认为每个人都有某种创造驱动力。我猜有些人有其他性格部分压过了这个。我的意思是，比如，我认为对一些人来说，就是服务，比如医疗驱动力。我认为成为医生或护士的那类人——我猜通常他们的主要驱动力是‘我真的只想帮助照顾人们。’我知道这很有趣。我听过一个故事——这可能实际上来自普莉希拉在医学院——或者在医学院的第一天，有人站起来问：‘你们中有多少人记得小时候看到某人时想，我真的只想照顾那个人？’每个人都举了手。而我认为对我来说，我的版本是，我小时候有很多记忆，比如‘我想去构建这个东西并让它更好’，对吧？所以我认为不是每个人都有相同类型的驱动力，但我认为每个人都有他们想做的事情。我同意，有了我们之前拥有的那种技术，对很多人来说，就是很难开始。所以，我对个人超级智能和 Muse 以及所有这些 AI 智能体感到非常兴奋的一件事就是，我认为人们第一次真正能够快速开始——或者你有一个宽泛的想法，AI 可以开始勾勒出来，然后你可以像雕刻雕塑一样精炼它，但你不需要在开始之前就知道一切关于如何去做这件事。

Yeah. Yeah. Well, yeah, no, I think—I guess when I say build, I'm talking a lot about these kinds of products, like software, hardware, that kind of stuff. But I agree, I think everyone has some sort of creative drive. I guess some people have other parts of their personality that overpower that. I mean, it's like, for example, I think for some people it's just service, like the medical drive. I think the type of people who become a doctor or a nurse—I think often their main drive is like, 'I just really want to help take care of people.' I know it's interesting. I heard this story once—this might have actually been from Priscilla at medical school—or in the first day of medical school, someone got up and they asked, 'How many of you had a memory from when you were a kid of seeing someone and thinking, I really just want to take care of that person?' Every hand went up. And I think for me, my version of that was I have a lot of memories when I was a kid of like, 'I want to go build this thing and make it better,' right? So I think not everyone has the same type of drive, but I think everyone has stuff that they want to do. And I agree that with the type of technology that we had before, for a lot of people, it's just hard to get started. So one of the things that I'm really excited about with personal superintelligence and Muse and all these AI agents is just having the ability, I think for the first time, for people to really be able to quickly get started—or you have some kind of broad idea and the AI can start to sketch it out and then you can just hone it in, like you're crafting a sculpture or something, but you don't need to know everything before you get started about how to go do the thing.

我认为那非常强大。所以，我认为很多人会找到他们想要创造或在世界上做的事情。

I think that's really powerful. So, I think a lot of people are going to find things that they want to create or go do in the world.

完全同意。

Totally.

嗯，我认为那会帮助人们感受到更广泛的能动性。

Um, and I think that will help people feel like a much broader sense of agency.

好的。关于生物医学，那就像你在 Meta 之外的另一个项目，要解决所有疾病。嗯，我很好奇，第一，我们离这个目标有多近？第二，

Okay. On the biomedicine, that's like your other project outside of Meta and to solve all diseases. Um, I'm curious, number one, how close are we? And then number two,

我们现在近多了。是的。

We're a lot closer now. Yeah.

你觉得现实上有多近？

How close like do you think realistically?

嗯，我的意思是，最初我们开始的时候，目标是在本世纪末帮助科学界治愈所有疾病。所以，我们的目标——我们从来不会自己去做。我们的观点基本上是，所有重大科学进步都先有一个新工具来测量和理解某事物。所以你知道，就像你有了显微镜，然后你理解了细菌，对吧？你有了望远镜，帮助我们理解了很多关于世界、宇宙中正在发生的事情。其中一些也是平台，比如你制造——第一个创造疫苗的人，现在人们可以创造很多不同的治疗方法。是的。嗯，但历史上科学资助的很多方式是广泛分散资金让人们进行个人探索。但对于这些大工具，资金并不多。而这基本上就是我们在 Biohub 试图做的：设计一些新工具，给人们看待生物学的新方式，希望能帮助科学界加速进展。所以之前我们认为，也许到本世纪末，我开始时认为那很有可能。嗯，这很有争议。我认为很多生物学家认为那不太可能。现在我认为那太长了。我认为随着 AI 的进展，我们现在正在做的一些东西——我们正在构建这个整个虚拟细胞模型，基本上，而不是必须在实际的物理活细胞上做实验，你基本上可以有一个 AI 模型模拟蛋白质，然后模拟细胞，然后最终模拟也许一个虚拟免疫系统或整个生物体或整个人。这——那将让人们能够运行许多不同的实验，并模拟在不同情况下会发生什么。如果你给某人这种药，会发生什么类型的事情？嗯，那将非常强大。所以，我不知道。我不想给出确切的年数，但我猜它会比本世纪末要早得多。

Well, I mean, originally when we started it, the goal was to help the scientific community cure all diseases by the end of the century. So, our goal—we were never going to do it ourselves. Our view is basically all major scientific advances have been preceded by a new tool to be able to measure and understand something. So you know, it's like you got the microscope, then you understood bacteria, right? You got a telescope that helped us understand a lot about what's going on in the world, in the universe. There are some of these things are platforms too, like you make—it's the first person who created a vaccine, now people can create a lot of different cures for things. Yeah. Um, but a lot of the way that science funding has worked historically is kind of very widely dispersing funding for people to do individual exploration. But there hasn't been that much funding for these big tools. And that's basically what we're trying to do with Biohub is design a few of these new tools that give people new ways of seeing biology that can hopefully help the scientific community just accelerate progress. So before we thought, you know, maybe by the end of the century, I thought that was pretty likely when we got started. Um, it was pretty controversial. I think a lot of the biologists thought that that was unlikely. Now I think it's way too long. I think with the progress in AI, some of the stuff we're working on now—we're building this whole virtual cell model where basically instead of having to experiment on actual physical living cells, you basically can have an AI model that simulates the proteins, then simulates a cell, then after that eventually simulating maybe a virtual immune system or a whole organism or a whole person. It's—that's just going to let people run so many different experiments and simulate what happens in different situations. If you give someone this medicine, what types of things happen? Um, that's going to be really powerful. So, I don't know. I don't want to put an exact number of years on this, but I would guess it's going to be much sooner I think than by the end of the century.

你觉得‘解决所有疾病’比‘永生’更有意图吗？永生是一种可能性吗？

And do you think solve all diseases feels very intentional over like live forever? Is live forever a possibility?

嗯，那不太是我的领域。嗯，我认为当然——嗯，我认为它们是两种不同的东西。我认为永生更多是关于延长。我认为即使你不生病，我认为在某个时候人体有自然寿命。好的。所以我认为那是一个单独的问题。有些人会把它视为一种疾病。

Well, that's not as much my field. Um, I think certainly—well, I think that they're two kind of different things. I think live forever is more about extending. I think even if you didn't get sick, I think at some point the human body has a natural life expectancy. Okay. So I think that's a separate problem to work on that. Some people would view that as a disease.

是的，这是看待问题的一种合理方式，但我认为人们还需要致力于治愈和预防那些即使你解决了其他问题也依然会找上你的疾病。所以这就是我们选择去攻克的那部分问题。再说一次，这并不是说你永远不会感冒之类的。我觉得其中一些事情是，你希望能够——我们称之为治愈、预防或管理，对吧？有些东西你可以完全治愈。有些东西我认为我们将能够有办法让你永远不会得它。还有一些，我认为你也许还是会生病，但那个东西本来会造成真正的损害，或者本来会要了你的命。但即使它不会要你的命，如果它可能造成真正的损害，现在你也可以把它作为一种持续的状态来管理，让它不再妨碍你的生活质量。所以我认为所有这些都将是这件事的重要组成部分。但人体——我的意思是，目标是让它保持平衡，对吧？所以并不是我们永远不会遇到病原体之类的。只是我认为在某个时间点，我们将能够治愈、预防和管理所有疾病。

Yeah, that's a reasonable way to view it, but I think people also need to work on curing and preventing the diseases that will get you even if you get that worked out. So that's the part of the problem that we've chosen to work on. And again, it's not that you will never get a cold or something. I think some of the stuff is that you want to be able to—we call it cure, prevent, or manage, right? So some things you can fully cure. Some things I think we're going to be able to have something that can prevent you from ever getting it. And some of these I think you'll just be able to—maybe you get sick, but the thing would have caused real damage or would have killed you. But even if it wasn't going to kill you, if it might cause real damage, now you can just manage it as this ongoing thing where it just doesn't get in the way of your quality of life. So I think all of those will be important parts of what this looks like. But the human body—I mean, the goal is to keep it in equilibrium, right? So it's not that we're never going to encounter a pathogen or something. It's just that I think at some point in time, we're going to be able to cure, prevent, and manage all diseases.

是的。这太令人兴奋了，因为我认为 AI 一直是许多不同突破的催化剂，然后你还有个人智能这一方面。你现在最频繁思考的是什么？就像你在做这件事的时候，你脑子里最常见的想法是什么？

Yeah. It's so exciting because I think that AI has been this catalyst in so many different breakthroughs like that, and then you have the personal intelligence aspect. What are you thinking about most frequently right now? Like as you're working on this, what's the most common thought in your head?

嗯，那可能每周都在变。

Well, that probably changes week to week.

好的。

Okay.

现在非常专注于 Muse，当你发布某个东西时总是很令人兴奋，你行动迅速。你尝试与世界接触。你了解人们的想法，然后你弄清楚接下来需要做什么。所以我们现在正处于一个阶段，我们正在获得大量关于人们想要什么的信息。很多好消息是人们喜欢它。所以我们正在努力把它带给尽可能多的人。在模型方面有很多工作，对吧？Muse Spark 模型已经取得了很大进展。所以我为此感到非常自豪。我们仍然想继续推进。我们想推动前沿，我们希望拥有世界领先的模型。至少达到那个水平。

Right now very focused on Muse, and it's always just very exciting when you launch something, you move quickly. You try to make contact with the world. You learn what people think, and then you figure out what you need to go work on next. So we're just in a phase now where we're getting a lot of information about what people want. A lot of the good news is that people love it. So we're working to just get it to as many people as possible. There's a lot of work on the models, right? The Muse Spark models have been a lot of progress. So I'm really proud of that. We still want to keep pushing that. We want to push the frontier and we would like to have the world's leading models. At least get there.

需要哪些突破？

What breakthroughs are needed?

嗯，研究这件事就是你并不一定事先知道。我的意思是，我们对一些我们认为我们的团队正在取得的突破有所感觉。所以有很多让我兴奋的东西，但过去一年里很多实际上只是扩展基础设施。你知道，如果你回到一年半前，我想可能更多人会说，为了达到超级智能，你需要一些根本性的架构突破。

Well, the thing about research is you don't necessarily know upfront. I mean, we have a sense of some breakthroughs that we think our folks are making. So there's a bunch of stuff that I'm excited about, but a lot of it over the last year has actually just been scaling the infrastructure. You know, if you went back a year and a half ago, I think that more people probably would have said that in order to get to superintelligence, you would need some fundamental architecture breakthroughs.

好的。

Okay.

而我现在其实不确定这一点是否成立。我认为我们大致知道配方是什么。如果你能建造一个足够大的超级计算机集群，那么我认为你可以暴力破解并达到那里。

And I'm not actually sure that that's true at this point. I think we kind of have a sense of what the recipe is. And if you could just build a big enough supercomputer cluster, then I think you can brute force it and get there.

有意思。

Interesting.

所以，我的意思是，我们正在俄亥俄州建造这个吉瓦级以上的集群。它基本上已经上线了。我们正在用它来训练我们的下一代模型。我认为那些模型会很棒。然后我们正在路易斯安那州建造五吉瓦的集群。所以那将会很棒。而且，我猜当你拥有多吉瓦的集群进行训练时，你基本上会得到某种近似于你随便怎么称呼它的东西，AGI，你知道，我认为超级智能还在那之后。但这并不意味着这是最好的方法。我的意思是，我认为人脑的运行功率大约只有 10 瓦。

So, I mean, we are building this gigawatt plus cluster in Ohio. It's basically online. We're using that to train our next generation of models. I think those models will be great. Then we're building the five gigawatt cluster in Louisiana. So that's going to be awesome. And yeah, I would guess that by the time you have multi-gigawatt cluster doing training, you're basically going to have something that approximates whatever you want to call it, AGI, you know, I think superintelligence beyond that. But that doesn't mean it's the best way to do it. I mean, I think that the human brain runs on about 10 watts.

对。

Right.

所以我认为，也许我们正在建造的这些计算系统比那低效大约一百万倍。所以我确实认为会有架构上的改进，有人能够找到。而我们当然正在努力做到这一点，因为如果你把大量的算力和架构改进结合起来，那么我认为你真的会建造出领先于该领域其他部分的东西。但在这一点上，我认为 Scaling（规模扩张）可以带我们走得很远。

So I think maybe these computational systems that we're building are on the order of a million times less efficient than that. So I do think that there are going to be architecture improvements that someone can find. And we're certainly working on doing that because if you combine both having this huge amount of compute and the architecture improvement, then I think you'll really build something that is ahead of the rest of the field. But at this point I think the scaling can take us pretty far.

是的，这很有意思，因为感觉 Scaling（规模扩张）是最容易保证的。就像研究，你必须真的希望有重大突破，但 Scaling（规模扩张）就是真的快速扩大规模。

Yeah, it's interesting because it also feels like scaling is the easiest guarantee. Like with research you have to really hope that there's a big breakthrough, but scaling just really scale up quickly.

我知道，这就是为什么大公司都在追求它，因为也许有更便宜的方法来做这件事，但我们还不知道。但这是世界上如此有价值的东西，如果你知道有很高的概率能做到，即使要花费数千亿美元去做，仍然值得去做。为了确保如果你最终没有发现那些能以低得多的成本实现它的突破，你仍然有明确的机会做到。当然，如果你确实发现了，那就更好了，这很棒。所以我有点认为这部分——我不想说它已经解决了，因为现实是，扩展基础设施来做这件事存在新颖的工程挑战，而且在每个模型训练层级你都会遇到新的奇怪问题，你必须调试并让它正常工作。所以我认为这就是研究工作的方式，它是那样迭代的。可能我们现在需要做的最重要的事情之一是专注于对齐，并让模型值得信赖。我认为这将非常重要，而且行业里有这样的争论，比如——实验室会自然地关注它吗？我的观点是，是的，实验室应该自然地关注它，因为如果你告诉 Muse 智能体做某件事，而它做了与你要求相反的事，我不知道有多少人会想用它，对吧？它需要——我们需要确保我们构建的产品和模型不仅理解你具体要求了什么，还理解其背后的意图和你的价值观，这样它们就不会以让你不满意的方式去做你想做的事，或者造成你不想要的负面效果。

I know, that's why the bigger companies are going for it because it may be that there is a much cheaper way to do it, but we don't know that yet. But this is such a valuable thing to create in the world that if you know that there's a pretty high probability chance of doing it, even if it's going to cost many hundreds of billions of dollars to go do, it is still worth it to go do that. To make sure that if in the end you do not also discover some of the breakthroughs that would have enabled it for a lot less, you still have a clear shot to do that. And then of course if you do discover it, then it just gets even better, which is great. So I kind of think that part of it—I don't want to call it solved because the reality is that there are novel engineering challenges in scaling up the infrastructure to do that, and at each level of training the models you encounter new weird things that you have to debug and get to work well. So I think that's kind of the way that the research works, it's sort of iterative like that. Probably one of the most important things that we need to do now is focus on alignment and making the models trustworthy. And I think this is going to be really important, and there's this whole debate in the industry which is like—are the labs going to naturally focus on it? And my view is like, yeah, the labs should naturally focus on it because I don't know how many people are going to want to use the Muse agent if you tell it to do something and it does the opposite of what you asked it to do, right? It needs to—we need to make sure that the products and the models that we build understand not only specifically what you asked but the intent behind it and your values, that way they don't go do the thing that you wanted in a way that you were unhappy with or cause some negative effect that you wouldn't have wanted.

所以它需要在相当深的层面上理解，这其实就是对齐。我的看法是，Muse 要想成功、触达十亿甚至数十亿人，我们就必须解决对齐问题，或者至少在这方面取得非常好的进展。

So it needs to understand at a pretty deep level, and that's kind of alignment. My view is that in order for Muse to be successful and reach a billion or billions of people, we're going to have to solve alignment, or at least make very good progress on it.

这是通过用户说“嘿，这不是我想让你做的”来实现的，还是你们在后台自己意识到？

Does it happen by the user being like, hey, that is not what I wanted you to do, or does it happen by you guys in the back end just realizing?

我认为两者兼有。我们肯定能从人们那里得到反馈，但我也认为实验室越来越意识到，你需要在训练过程中就对它进行对齐，因为模型已经足够聪明，如果你不这样做——我的意思是，我们在其他实验室看到的大量安全事件基本上都发生在训练期间，而不是在模型部署给人们使用之后。所以很多问题在于：为了能够训练出这种智能水平的东西，你需要为它制定一套非常好的课程，就像父母一样，你需要设定明确的界限，这样如果它做了错事，它就会学到——不，我是想让你真正解决这个编程问题。我不想让你通过更改某些系统配置、改变某个值或绕过它来获得奖励。我想让你真正去完成解决这个问题的工作，这样你就能学会解决问题的方法。这才是训练的意义。

I think it's a combination of both. There's definitely feedback that we get from people, but I also think that increasingly what the labs are seeing is that you need to do alignment as you're training it, because the models are getting smart enough that if you don't—I mean, a lot of the safety and security incidents that we're seeing at the other labs are basically during training, not after they've deployed the model for people to use. So a lot of this is: okay, in order to be able to train something that is of this level of intelligence, you need to develop a very good curriculum for it, and like a parent, you need to set firm boundaries so that if it does something that is wrong, it learns that—no, I wanted you to actually solve this coding problem. I didn't want you to get the reward by changing some system configuration that changed some value or hacking around it. I wanted you to actually go through the work to solve the problem so that you'd learn the method of doing that. That's the point of training.

是的。

Yeah.

所以其中一部分就是要有良好的安全性，建立明确牢固的界限。我认为所有这些都只是行业必须做的自然而然的事情，我们正在这方面花很多时间。而且，每一天都不一样。

So part of that is just about having good security, establishing good firm boundaries. And I think all this stuff is just natural stuff that the industry is going to have to do, that we're spending a lot of time on. And yeah, every day is different.

因为我觉得你也说过，在 AI 安全方面，感觉过去两周事情到了紧要关头，尽管并没有一个具体的时刻导致这种情况。但听起来你的观点是，我们实际上把这个模型多保留了几个月，只是在内部进行了训练。

Well, because I think you also said that with the AI safety stuff, it feels like it's kind of come to a head in the last two weeks, even though there hasn't been actually a moment that has made that happen. But it sounded like your take was we actually held this model for a few extra months and we just did the training internally.

是的，对于 Muse，我们知道产品需要非常私密和安全。所以我们基本上有一个早期版本的模型。我们认为可以训练出更多关于谨慎行为的能力，就像我们讨论过的。所以我们花时间做了这件事。还有一堆关于让虚拟机——我们给每个 Muse 智能体的那台计算机——更安全的工作。所以我们多花了几个月来做这些。但我不确定。我认为当其他实验室经常说，哦，我们为了放慢速度而承受了很多痛苦。我的观点是，这对 Meta 和将要使用 Muse 的人来说都是正确的做法。我们想让产品好，如果产品不好，那对人们不好，对我们也不好。我们不想推出一个东西，给人留下糟糕的第一印象，然后人们就不兴奋去用了。所以我有点认为——我自己的观点是，所有实验室都会有非常强烈的内在动力去做好这件事。

Yeah, for Muse, we knew that the product needed to be very private and secure. So we basically had an earlier version of the model. We thought that we could train in some more behavior around discretion, like we talked about. So we took the time to do that. There's also a bunch of stuff about making the virtual machine, the computer that we give to each Muse agent, more secure. So we took a few extra months to do that. But I don't know. I think when other labs have often talked about, oh, we're taking a bunch of pain in order to slow stuff down. My view is that was the right thing for both Meta and for the people who are going to use Muse. We want to make the product good, and if the product isn't good, then it's not good for people and it's also not good for us. We don't want to put something out and have it be a bad first impression and then people aren't excited to go use it. So I kind of think—my own view on this is that all the labs are going to have a pretty strong intrinsic incentive to get this right.

是的，感觉如果你只是把激励对齐到它变得极其有价值和安全，那就是关键。

Yeah, it feels like if you just align the incentives to it being incredibly valuable and safe, and that's the key.

完全正确。完全正确。

Totally. Totally.

嗯，非常感谢你在这么重要的一周抽出时间。

Well, thank you so much for your time on a huge week.

谢谢。谢谢。

Thank you. Thank you.