Meta CTO 安德鲁·博斯沃思谈 AI 转型、智能眼镜与员工追踪

这个行业之所以如此混乱——不仅在 Meta，而是每家公司——并且弥漫着巨大的不安，原因之一是我们对哪怕相对近的未来都缺乏清晰的愿景，这在很长一段时间里都是最模糊的。我想两年前我们开始意识到这比我们想象的更大。去年它变得如此严肃，以至于对我们很多人来说几乎成了一场存在主义式的对话。大家好，我是 Nicholas Thompson，这里是《AI 中最有趣的事》，由 Rethink 制作，大西洋月刊的创意营销工作室出品。如果你到了一定年纪，你可能还记得一个非常不同的 Facebook。它有点像你朋友的通讯录。你做不了太多事，只能评论、点赞，还有个叫“戳一下”的功能。但在 2006 年，Facebook 做了一件有争议的事。他们加入了动态消息。突然之间，每个人的 Facebook 互动都被广播给了整个网络。用户说这侵犯了他们的隐私。这可能是公司第一起隐私丑闻。Facebook 短暂道歉，然后继续推进。你知道吗？它成了 Facebook 有史以来最重要、最成功的产品之一。这背后的推手是一个叫 Andrew Bosworth 的人，人称 Boz。他是 Facebook 非常早期的员工，一直待到现在。他常常处在风暴中心，推动那些用户不满但最终成为公司业务核心的事情。将 Facebook 更名为 Meta 的决定部分涉及 Reality Labs、他们的 VR 眼镜、AR 眼镜，以及现在有些人喜欢、有些人说是在引入大规模间谍软件的眼镜。他还部分负责公司的 AI 业务，这当然是一个巨大的赌注。那里也有点争议，Facebook 曾表示要记录员工的击键来改进 AI。在员工愤怒后他们撤回了，但谁知道接下来会发生什么。所以，以下是我与 Andrew Bosworth 的对话。Andrew Bosworth，欢迎来到《AI 中最有趣的事》。

One of the reasons it is such a chaotic time in the industry, not just at Meta but across every company, and there's a tremendous amount of unease, is that this is the least clear vision we've had of even the relatively near future in a long time. I think two years ago we started to get a sense that this is bigger than we thought. Last year it got to a point of seriousness that it became an almost existential conversation for a lot of us. Hello, I'm Nicholas Thompson and this is the most interesting thing in AI, a podcast produced by Rethink, the Atlantic's creative marketing studio. If you're above a certain age, you might remember a very different version of Facebook. It was sort of a directory of your friends. You couldn't do much. You would just comment and like, and there's a thing called poke. But in 2006, Facebook did something controversial. They added News Feed. Suddenly, everyone's Facebook interactions were blasted out to the whole network. Users said it violated their privacy. It may have been the company's first privacy scandal. Facebook briefly apologized and then just pressed forward. You know what? It became one of the most important, successful products that Facebook ever made. The man behind it was a guy named Andrew Bosworth, known as Boz. He's been a very early Facebook employee and he has been there ever since. Often he's in the middle of the storm pushing things forward that users are upset about, but that end up being really core to the company's business. Part of the decision to rename Facebook Meta was a charge of their Reality Labs, their VR glasses, their AR glasses, their current glasses that some people love, some people say are introducing mass spyware. He's partly in charge of their AI operations, which of course is a massive, massive bet for the company. Also a little controversy there when Facebook said it was going to record the keystrokes of their employees in order to make their AI better. They then rolled that back amid employee outrage, but who knows what's going to come next. So, here's my conversation with Andrew Bosworth. Andrew Bosworth, welcome to the most interesting thing in AI.

谢谢邀请。

Thanks for having me.

你是为眼镜来的。请解释一下我们在眼镜开发上处于什么阶段，因为要让它们成为大众产品并取代手机，你需要一个庞大的开发者生态系统。但要获得庞大的开发者生态系统，你需要眼镜被大规模采用。要让眼镜被大规模采用，你需要一个庞大的开发者生态系统。你在这个飞轮中处于什么位置？

You're here for the glasses. Explain where we are in the development of the glasses because in order to have them become a mass product and to displace the phone, you need a big developer ecosystem. But to get a big developer ecosystem, you need mass adoption of the glasses. To get mass adoption of glasses, you need a big developer ecosystem. Where are you in this flywheel?

是的，我们正在取得巨大进展。我告诉你，因为事实证明眼镜不需要开发者生态系统就能很棒。我们在推出时就知道了：这些眼镜必须本身就很棒，人们才会想用。有趣的是，即使没有 AI 功能——我认为这是它们非常令人兴奋的部分——它们也是很好的眼镜。人们一直在用。他们用音频功能，用视频和免手录制。在有小小孩的父母中非常受欢迎，这样他们可以捕捉那一刻、分享它，而不用从手机上分心。在运动员中也很受欢迎，他们长跑或骑行时想免手捕捉或同时听音频。所以眼镜凭实力卖得很好。我们当初造眼镜时甚至没有设想今天这样的 AI。现在我们有了 AI，它们就更好了，更能帮助人们。至于显示眼镜，我们刚为显示屏推出了一个 vibe coding 平台。我们看到开发者热情高涨，因为这不是把世界上每个应用都搬到眼镜上，而是你需要为你的眼镜专门做什么，你可以去获取。我确实认为我们前进的方向是远离应用甚至静态开发的概念。我认为我们正走向非常动态的 AI 生成应用，你只需在眼镜上或任何地方向 AI 提出你需要什么，它就会即时生成并交付给你。所有服务提供商都因其创造的价值获得报酬。一切都在幕后代表你协商，但你在需要时得到你需要的东西。

Yeah, we're making great progress. I tell you that because the glasses turns out don't need a developer ecosystem to be great. We knew that was going to be the case when we launched them: these have to be great on their own terms for people to want to use them. And actually, what's funny is that even without the AI features, which I think are such an exciting part of them, they're great glasses. People are using them all the time. They're using them for the audio. They're using them for the video and the hands-free recording. Very popular with parents with young children so they can capture that moment, share it, but not be separate from it from their phone. Very popular with athletes who are going on long runs or going on rides and want to be able to do hands-free capture or have audio going at the same time. So the glasses are selling very well on the merits. We didn't even conceive of AI as it exists today when we built the glasses that we have. Now that we have AI, they're even better. So they're even more capable of helping people. And then with the display glasses, we just launched a vibe coding platform for the displays. And we're seeing tremendous enthusiasm from developers because it's not about taking every app in the world and bringing it to the glasses. It's about what is the thing that you need for your glasses specifically and you can go get that. I do think where we're headed over time is away from this concept of apps or even static development. I think we're headed towards very dynamic AI generated applications where you just ask your AI on your glasses or wherever for what you need and it generates it on the fly and delivers it to you. And all the service providers get paid for the value they're creating. It's all negotiated on your behalf behind the scenes, but you get what you need when you need it.

我最喜欢的眼镜用例是戴上它们，像和 AI 对话一样，比如我在做早餐时想了解今天的一些东西。今天早上我在给孩子们做早餐时问它关于 Muse Spark 的组件，只是为了更跟上进度。对我来说，我认为对世界来说，杀手级应用是如果它们能调整视力，对吧？如果你能输入你的处方，或者能转个旋钮，那什么时候能实现？

My favorite use case for the glasses is I put them on and I just talk to them like talk to AI like I'm making breakfast and I want to learn about something today. I was asking this morning while I was making my kids breakfast. I was asking it about the components in Muse Spark just so I would be a little more up to speed on that. The killer app for me and I think for the world would be if you could adjust vision with them, right? And if you could enter your prescription or if you could turn a knob and when will that be possible?

是的。那些叫液体透镜。它们有很多不同的形式。所以现在我要说清楚，眼镜可以矫正你的视力。现在

Yeah. So those are called liquid lenses. They exist in a lot of different formats. So right now I say we want to be clear the glasses can correct your vision. Right now

你只是配处方镜片。

you just get prescription lenses.

现在它们有处方眼镜。

Right now they have prescription glasses.

拿这些，对吧？我会，你知道，把这些镜片拿出来，放到这里，然后我就能完美地看到你。对。但

Take these, right? And I'll like, you know, take these lenses out, put them in here, and then I can see you perfectly. Right. But

所以它们确实有处方镜片。没错。

so they certainly do prescription lenses. That's right.

现在我看不清你。

Now I can't see you.

是的。所以我想说清楚，眼镜确实能处理处方镜片。你问的是什么时候能有自动调节的液体透镜

Yeah. So I want to be clear that the glasses do handle prescription lenses. You're asking like when can you have autotuned liquid lenses

或者能调整它的软件，不管解决方案是什么，因为我觉得这是这些东西的杀手级应用。

or software that adjusts it like whatever the solution is cuz I feel like this is the killer app for these things.

软件调整真的非常非常难。没有技术路径。但可能存在的技术路径是，你可以有由两层膜组成的镜片，在压力或张力下变形，内部有某种液体，这在一两个轴向上相当可控。这是一种可能性。我认为那项技术存在。我们不这么做的主要原因是那些镜片更脆弱，而获得处方镜片并分发给人们的稳健性已经随着时间变得好多了。

Software adjustment is really really hard. There's not a technology path for that. But the technology path that probably does exist is you can have lenses that are two membranes that deform under pressure, under tension, that have some kind of a liquid inside and that's quite controllable on either one or two axes. That is a possibility. I think that technology exists. The main reason we don't do it is that those are a little more fragile and the robustness of getting prescription lenses and getting those distributed to people has gotten quite a lot better over time.

第二个最具争议且更难的 app 是人脸识别，对吧？它显然会极其有用，但显然会让人不安。

The second app that is most controversial and harder is face recognition, right? It'll obviously be like supremely useful, but obviously it weirds people out.

是的。

Yeah.

有争议。

Controversial.

我们刚刚向美国每一位盲人退伍军人捐赠了 18 万副 Ray-Ban Meta 眼镜。为什么这么做？这款眼镜在盲人群体中非常受欢迎。

We just donated 180,000 pairs of Ray-Ban Metas to every blind veteran in America. Why did we do that? The glasses are incredibly popular with the blind community.

你们早期有一位盲人设计师参与，对吧？其中一位产品负责人不就是盲人吗？

You had a blind designer on them early, right? Wasn't one of the product leads blind here?

是的。实际上我问过她，我说，你知道，为什么这些眼镜对盲人这么重要？为什么它们对盲人帮助这么大？她说，你知道吗？人们对失明有误解。我能到餐厅。我有谷歌地图。我有可听设备。我有手杖。我有导盲犬。我能到餐厅。她说，你知道我做不到什么吗？找不到门。谷歌地图告诉我我在餐厅了。好。门在哪儿？我看不了菜单。所以他们失去了独立性。于是他们打电话给像 Be My Eyes 这样的服务。Mike Buckley，CEO，其实是前 Meta 员工。他们和看着摄像头的人通话，对方告诉他们：“好，门在你左边。好，就在那儿。你找到了。”现在，他们仍然可以通过我们与眼镜的合作做到这一点，但他们也可以直接问 AI，AI 可以为他们回答。所以，他们获得了独立性。如果你问盲人群体，我们从他们那里收到的第一大请求是什么，那就是：谁在房间里？告诉我谁在这里。他们不是在要求一个全球中央人脸数据库。他们说：“我认识的人，比如谁，如果我认识的人在这里呢？”有趣的是，我上周在 Oscar Mike。他们来到园区。这是一个退伍军人团体，处理创伤性脑损伤的人。所以他们经常有记忆问题。他们有认知问题。这个问题出现了，这又是他们的首要请求。嘿，我只想知道我在和谁说话，我怎么认识他们的，我和这个人有什么历史？他们不是在要求一个中央人物数据库。他们要求的是，这是我认识的人。

Yeah. And actually I asked her, I was like, you know, why is it important to blind people to have these glasses? Why are they helping blind people so much? And she's like, you know what? People get blindness wrong. I can get to the restaurant. I've got Google Maps. I've got hearables. I've got a cane. I've got a dog. I can get to the restaurant. She's like, you know what I can't do? Can't find the door. The Google Maps told me I'm at the restaurant. Cool. Where's the door? I can't read the menu. So, they lose independence. So, they call services like Be My Eyes. Mike Buckley, CEO, former Meta employee, actually. And they talk to somebody who's looking at the camera and tells them, "Okay, the door is to your left. Okay, there it is. You got it." Now, they can still do that with a partnership that we have with glasses, but they can also just ask the AI, and the AI can answer it for them. So, they're gaining independence. If you ask the blind community what the number one request that we get from them is, it's who's in the room with me? Tell me who's here. They're not asking for a global central database of faces. They're like, "People I know like who what if people what people that I know are here?" It's funny. I was at Oscar Mike last week. They came to campus. This is a veterans group that deals with people who had traumatic brain injuries. So, they often struggle with memory. They struggle with cognition. And this came up and this is again a top ask for them. Hey, like I just want to know who I'm talking to and like what how do I know them and what's my history with this person? They're not asking for a central database of people. They're asking for like this is someone I know.

等等，所以这个被报道的项目叫 Name Tag。它的理念不是：这是一个你从未见过的人。这是他们是谁。这是他们的 LinkedIn。而是：这是一个你见过的人，或者在你的 Facebook 社交图谱里的人

Wait, so the idea the program that has been written about is called Name Tag. The idea of that is not here's a person you haven't seen before. Here's who they are. Here's their LinkedIn. It's that this is a person you've met or who's in your Facebook graph

本地加密到你的设备。你戴着眼镜亲自见过的人，他们自我介绍过，或者你说：“好，这是 David。记住这个人”，只在你戴着眼镜时对你可用。这是你以前见过的人。这是他们的名字。他们就在你面前。这就是我在说的。我们称之为名牌功能，

encrypted locally to your device. Somebody that you met in person with your glasses on who introduced themselves or you said, "Okay, this is David. Remember this person" only available to you when you're wearing your glasses. This is a person you've met before. Here's their name. They're right in front of you. That's what I'm talking about. That's we call a name tags feature,

对吧？

right?

所以我认为人们会困惑，因为他们听到人脸识别，就以为：“哦，有一个中央人脸数据库，每个人都在不断被扫描进去。”不是那样的。但你说得对。它仍然有争议。我认为这是，你知道，在某些州根据州法律你甚至不能做的一件事。你知道，我们在伊利诺伊州有 BIPA，在德克萨斯州有 CUBI 等等。所以，我认为这会是一个很棒的功能。顺便说一句，我通过盲人低视力社区或通过 Oscar Mike 讲过这个故事。你会用它。我会用它。比如我去派对，我肯定以前见过某人。你是记者。你一定经常遇到人，脑子里不断想，我认识这个人。名字、脸，我们上次谈话是什么时候？这里的背景是什么？我们称之为鸡尾酒会问题。这是一个非常普遍的人类问题。几乎每个和我交谈的人，无论他们的视力或记忆状况如何，都理解这是一个能帮助他们在社交场合更自在的问题。如果他们拥有它，但必须以让人们感到舒适的方式来做，

And it's so I think people get confused because they hear face recognition and they think, "Oh, there's a central face database that everyone's being scanned constantly into." It's not that. But you're right. It's still it's still controversial. I think it's, you know, one of those things that you can't even do in some states based on the state laws. You know, we've got BIPA in Illinois and CUBI in Texas and these things. So, it's a thing that I think it would be a great feature. By the way, and I've told the story through the blind low vision community or told the story through Oscar Mike. You'd use it. I'd use it. Like I go to parties where I like definitely have met somebody before. You're a journalist. You must meet people constantly and constantly in your head be like, I know this person. Name, face, what was the last time we talked? What's the context here? It's we call the cocktail party problem. It's a very universally human problem. Nearly every human I talk to no matter what their status of their vision or their memory is, understands that this is a problem that would help them feel more comfortable in social situations. If they had it, but it has to be done in a way that people feel comfortable,

对吧？这绝对是让人毛骨悚然的东西，毫无疑问，每个人都能看到它的实用性。比如如果你去过高中同学聚会，天哪，你会用到这个实用性，但也很明显让人感到不安。作为 Meta，你的优势之一或优势之一是每个人都在用它做消费者用例。你能够把反馈回去。解释一下你如何在你的员工身上训练。这最近显然因为模型能力计划而引发争议。你想从员工身上学到什么？正如媒体报道的那样，你在追踪他们输入的内容，从他们的智慧中学习。我相信扎克伯格说过：“我们的员工比普通人更聪明。我们想看看他们如何使用机器，以便我们能让 AI 更好。”解释一下。

right? And it's definitely the thing that creeps, no question, everybody can see the utility. Like if you've been to a high school reunion, my lord, you could use the utility, but also quite clearly um creeps people out. The advantage that or one of the advantages that you have as meta is that everybody's using it for consumer use cases. You're able to feed that back in. Explain how you're training on your own employees. This has obviously been controversial recently with the model capability initiative. What are you trying to learn from your employees? As has been reported in the press, you're, you know, tracking what they type in, learning from their intelligence. I believe it was Zuckerberg said, "Our employees are more intelligent than the average person. We want to see how they use machines so that we can make our AI better." Explain that.

是的。不过那个项目现在已经暂停了。所以

Yeah. Although that's that program is now paused. So

整个项目只是暂时暂停

the whole program is paused just for now

因为数据问题。

because of the data issues.

嗯，我们正在努力确保我们对这个项目有完整的交代。让我，但让我解释一下它的理论。理论其实非常直接。你有一群员工已经在每天做这项工作。而且不需要他们额外付出努力，你就有可能改进你的模型。

Well, we're trying to make sure we we make sure we have a full accounting for for the program. Let me but let me explain that the theory of it. The theory is actually very straightforward. You've got a bunch of employees that are already doing this work every day. And for no incremental effort on their part, you can potentially improve your models.

是的。

Yep.

这不仅有利于你的员工回去工作时，也有利于消费者，想使用 AI 的人。现在，这项研究的范围，它确实是研究。值得注意的是，这是一项我们预计 18 个月内不会有成果的研究，最早也是你实际看到这在模型内部会是什么样子的最早时间。它需要非常长期的数据， literally 就像如何使用电脑这样简单的事情，而 AI 作为计算机本身却出奇地不擅长。就像字面意义上，你如何导航 UI？你接下来点击什么？我们的很多工作流程都在这些数字孤岛中。它们会在那里存在很长时间。比如一年、一年半后，仍然会是这种情况，我们每天在工作中经历的流程。所以，你知道，最基本的前提是，嘿，我们能教 AI 使用电脑吗？这是一个实际上不存在大规模的数据集。你可以做这个的急性版本。

And that benefits not only your employees when they go back and try to do their jobs, but also consumers, people who want to use AI. Now, the scope of this research and it really was research. It's worth noting this is research that we didn't expect to yield for like 18 months, like at the earliest would be the earliest that you would actually have a look at what this would look like inside of a model. It needs very long-running data was literally things as simple as how to use a computer, a thing that AIs are surprisingly bad at for being themselves computers. Like literally how do you navigate UIs? Like what do you click next? And a lot of our workflows are in these like digital silos. They're going to be there for a long time. Like a year, a year and a half from now, it's still going to be the case that these that the flows that we're going through every day as we go about our work. So the, you know, the most basic premise was like, hey, can we teach AI to use computers? It's a data set that doesn't really exist at scale. And you can do acute versions of this.

你当然可以通过 Scale AI 或其他地方雇人，让他们给你快速演示一下，但我们想看看在几个月的时间里，人们究竟是如何用电脑推进工作的。所以这就是这项研究的前提。

You can certainly hire people through Scale AI or other places to give you a quick walkthrough, but we wanted to see over the course of months how people make progress on work literally using a computer. So that was the premise of this research.

你们录制和追踪的是什么？

What were you recording and tracking?

对，就是字面意义上的鼠标移动、你点了什么、你在哪里、你激活了哪个窗口。我们试图在这里建立一个足够大的数据产业。

Yeah, it was literally your mouse movements, what you're clicking, where you are, what window you're activating. And we tried to create enough of a data industry here.

是所有员工吗？你追踪自己了吗？

And was it all employees? Were you tracking yourself?

不过在美国，有少数员工被排除在外。一个很好的例子就是像我这样的人，以我所处的职位，我们受到各种法律保留义务的约束。也就是说，任何以我名义创建的信息都会依法自动存储和保存。这让我成为这项工作中一个相对不合适的目标。所以有一批岗位，那些经常处理敏感数据的人，都被排除了。我们排除了一些团队，也排除了一些个人，还有一整套流程让人们可以选择退出。

In the US, though, there were a few employees excluded. A good example is if you look at someone like me, in the position that I'm in, we're under any number of legal holds. Where it's like, any information that's created on my behalf is automatically stored and preserved by law. That makes me a relatively poor target for this work. So there's a bunch of jobs, people who regularly worked on sensitive data, that were excluded. So we excluded a bunch of teams, we excluded individuals, there's a whole process by which people can opt out of this stuff.

不，那会是一个绝妙的数据集。那些受法律保留义务约束的公司高管是怎么用电脑的？

No, it would make for an amazing data set. How do executives at companies dealing with legal holds handle computers?

那是一个你可以想象具有价值的专有数据集。

That's a proprietary data set that you can imagine having value.

是的。退一步说，我认为我们正处于一个非常有趣的位置。如果你看了那场庭审，开场时 Greg Brockman 写给自己的日记笔记被曝光了。而且现在我认为至少有一个法律先例表明，向 LLM 提问，即使你问的是法律问题，也不能受到保护，对吧？

Yeah. Taking a step back, I think we're at a very interesting place where if you saw the trial, the opening trial with Greg Brockman's writing a note to himself as a diary and that's being exposed. And they've now, I think there is at least one legal precedent that shows, hey, asking questions to an LLM, you can't have that be protected even if you're asking a legal question, right?

我确实认为有一个有趣的问题，我们需要在不久的将来提出来，关于保密特权，以及在一个普通人——一个甚至可能请不起律师的人——能够可信地获得法律建议的世界里，什么应该享有特权、什么不应该。先不说公司，没人关心公司，它们有足够的钱雇所有律师。一个请不起律师的人，却能从 AI 那里获得专家级的法律建议。鉴于我们制度的前提是人们应该获得法律代理，那这该是什么样子？所以我认为这是个有趣的问题，但我会把它留给法院。

I do think there's an interesting question that we need to raise at some point relatively soon on privilege and what should be privileged and not privileged in a world where you can credibly get a consumer, somebody who couldn't even afford a lawyer necessarily, could get credible legal advice. Set aside, you know, nobody cares about corporations, they got enough money to hire all the lawyers. Somebody who can't afford a lawyer but could actually get expert legal advice from an AI. Given that the premise of our system is that people should have legal representation, what does it look like? So I think there's an interesting question but I'll leave that for the courts.

我们把它留给法院和立法者吧。好，我们回到正题。所以你和其它处理敏感信息的人被排除在这个项目之外，但大多数 Meta 员工是——

We'll leave that for the court legislators. All right, let's go back. So you and other people who work on sensitive information are excluded from this program but the majority of Meta employees are—

美国——

Of American—

美国 Meta 员工。

American Meta employees.

对。对。

Yeah. Yeah.

对。欧洲绝不会让你这么做，对吧？

Right. You like Europe would never let you do this, right?

我们甚至都没问。

We didn't even ask.

甚至都没问。对。好。所以大多数人都被纳入了。你们在追踪鼠标移动。你们学到了什么？它有多大价值，你们为什么停止？

Didn't even ask. Right. Okay. So majority of them are included. You're kind of keeping track of mouse movements. What did you learn? How valuable was it and why did you stop?

对，所以我们在几周后看了看数据，发现重复比我们预期的多得多。有趣的一点是，当你想到海量数据时，通常说的是预训练；而当你谈到精良数据，也就是有非常详细的专家解释时，那往往是中训练或后训练。所以这些数据本意是长期运行的，但你需要方差。方差远比数量重要。大量相同的东西基本上会被压缩成一个例子。

Yeah, so we took a look at the data a couple of weeks in and there was a lot more duplication than we expected. One of the things interesting about this is when you think about high volumes of data you're usually talking about pre-training, and when you're talking about exquisite data where you've got a really detailed expert explanation of something, that's a lot of times mid-training or post-training. So this data was meant to be long-running but you need variance. Variance is far more important than volume. A high volume of the same thing gets collapsed into one example basically.

是的。

Yeah.

所以我们最先看到的一件事是，方差比我们预期的少得多。所以这就是为什么，我想在我们最初启动几周后，我们增加了——扩大了退出选项，给那些不想参与的人。一个暂停，无限暂停，只要你不想要就按暂停。你想按多少次都行。所以我们加了一堆东西。我们说，好吧，这里有很多重复。所以我们可能超出了所需的数量。然后我们经历了那个过程，好吧，看了那个并减少了方差，现在我们加上了这些工具。然后上周，我们出了个问题：我们存储的数据相当安全，只有相对少数人能访问。一个下游处理该数据的研究员——这里没有发生泄露——但把它放在了一个不该放的地方。所以我们现在把整个事情锁起来，直到我们能——

And so one of the first things we saw was actually there was a lot less variance than we expected. So that was why, I think a couple weeks after we initially launched it, we added—expanded opt-outs for people who didn't want to do it. A pause, like infinite pause, just whenever you don't want to have it just press pause. You can do it as many times as you want. So we added a bunch of things. We're like, okay, there's a lot of duplication here. So we probably overshot needing as much as we got. And then we went through that process of, okay, having looked at that and reduced the variance, now we've added these tools. And then this last week, we had an issue where the data that we were storing was quite secure, relatively small number of people had access. One of the researchers who was working downstream with that data—and there was no breach here—but had put it in a place it wasn't supposed to go. So we're kind of locking the whole thing down until we can—

数据是指——

The data meaning—

员工——就像——

The employee—there's like—

这些员工用了这些按键，以这些方式移动了手。

These employees used these keystrokes and moved their hands in these ways.

不是原始数据。那是它之上的一个转换，但它仍然落在了内部不该落的地方。仅限内部。没有理由相信有不当行为。我们正在检查每一次访问之类的事情。所以我们只是暂停了整个项目，直到我们查清此事。在此期间，我们实际上会做一个早期模拟，看看只用我们已有的这个小数据集训练会是什么样子，以了解数据的长期价值。因为这是一个研究项目。这不是说，嘿，我们保证这一定会成为改变游戏规则的东西。

Not the raw data. It was a transform on top of it, but it still landed someplace it shouldn't have landed internally. Internal only. No reason to believe foul play. We're looking at every single access kind of things. And so we've just paused the entire program until we can get to the bottom of this. While we're doing that, we're actually going to run an early simulation of, hey, what would it look like to do a train just with this small data set that we have to understand the long-term value of the data? Because this was a research project. This wasn't like, hey, we're guaranteed this is guaranteed to be the game changer.

为什么你们在自己的员工身上做，而不是雇一万个任务兔子？

Why did you do it on your own employees as opposed to hiring 10,000 task rabbits?

对。主要的是我们想要的是这些长轨迹。你想要的是数月之久的每日工作。而且我认为如果你真的——数据实际上看起来像在专业环境中长期真实工作的样子，这非常重要。所以我认为这是一个相当有吸引力的数据集，一个独特的数据集。我认为你在 AI 中必须有的一个长期概念是，数据是一切价值的源泉。所有将其转化为这些模型的惊人魔法几乎就是魔法，对吧？但如果你没有 upfront 的数据集，它最终出现在模型分布中的机会就很低。所以我预计，随着我们继续沿着这些曲线、这些缩放定律前进，AI 中会发生的是对这些数据集、这些长尾数据集的需求增加。你知道，我们认为互联网上所有人类媒体的语料库肯定就是人类知识的全部，但它差得远呢。

Yeah. The main thing is what we wanted were these long traces. You want like months-long every day work. And I think if you really—it's very important that the data actually look like what real work looks like over a long period of time in a professional kind of environment. And so I think it's a pretty appealing data set, a unique data set. I think one of the long-term concepts that you have to have with AI is that data is the font of all value. All the tremendous wizardry involved in transforming it into these models is almost magic, right? But if you don't have the data set up front, the chance that it's going to be in the distribution of the model in the back end is low. And so I expect to happen in AI as we continue to fight along these curves, these scaling laws, is an increased demand for these data sets, these long-tail data sets. You know, we think of all the corpus of human media on the internet as being surely this is all of human knowledge, but it's not even close.

我觉得这连皮毛都算不上。比如有人在处理《萨班斯-奥克斯利法案》的合规问题。这里面每一项都需要大量的判断。你可以读法规条文，但还是会做错，因为你没有接入行业内的人是怎么处理这些事的，监管者在想什么，真正的关切是什么。所以长尾非常长。

I don't think it even scratches the surface. You've got somebody who's working on Sarbanes-Oxley compliance. There's a tremendous amount of judgment that goes into each of those things. You can read the statutes and get it wrong because you're not plugged into how people in the industry are accounting for things, and how the regulators are thinking, and what the real concerns are. And so the long tail is very long.

那你最想获得、但目前拿不到的理想数据集是什么？

So what's the dream data set that you would like to have access to that you don't have access to?

嗯，我觉得所有能体现真实人类价值创造的数据集都很有意思。其中有些可能根本拿不到。比如一个人创作一个故事、一个精彩故事时，他的创作过程是什么？世上精彩的故事本来就不多，更别说还要有人在旁边观察记录他们的创作过程。而且如果你去问那些人他们的过程是什么，很可能他们自己都不知道。你知道，就像“我在洗澡时有了这个想法，然后就顺着做下去了”。所以 AI 是一个数据饥渴的载体，你能喂给它的独特数据越多，它在另一端能帮你的就越好，对吧？它作为工具就越有用。

Well, I think all data sets that represent real human value creation are interesting. And these are some that are probably impossible to get. What's the creative process somebody goes through when they're crafting a story, an amazing story? There's just not that many amazing stories out there, let alone the people to monitor while they're doing it. And if you ask those people what their process was, there's a good chance they don't even know what their process was. You know, like, I was in the shower, I had this idea, I followed it. So AI is a data-hungry vehicle, and the more unique data you can feed it, the better it's able to help you on the other side, right? And the more useful of a tool it is.

你的项目让员工感到不安的一个原因是，他们会想：“等等，这是不是用来取代我的？这会不会让裁掉我变得更容易？”你觉得，对于那些我们创造的、真正只属于我们自己的数据，界限应该划在哪里，才不该被喂进 AI？

One of the reasons why your program was uncomfortable to employees was they were like, "Hold on a second. Is this being used to replace me? Is this going to make it easier to wipe me out?" And where do you think the line should be drawn around the data that we create that really is just ours and shouldn't go into AI?

是的，我个人认为这个工具主要是为了最大化每一个个体的能力。所以当我想到今天 AI 为我做的事，它是在做那些我本来会做、也有能力做、但我觉得并不是我时间最佳用途的任务，对吧？我认为每当出现这类生产力提升——这在人类历史上可以追溯很久——我们都会担心：哦不，作为人的意义要消失了。但每一次，结果都是：不，并没有。事实证明，那并不是作为人的意义所在。并不是说我们必须手工缝制布料，设计服装才算一个有价值的职业。我们一直在这类工作中找到意义，这对我们是好事，但那并不是人类意义的专属领域。事实证明，我们会把意义赋予任何我们投入精力的事情。所以我把 AI 看作“个体的无限延展”。

Yeah, I'm personally somebody who thinks this is a tool that's mostly about maximizing each individual person. And so when I think about what AI does for me today, it's doing tasks that I otherwise would have done and am capable of doing, but I just don't think are the best use of my time, right? I think every time we find one of these productivity improvements, going back quite a long ways in human history, we worry it's like, oh no, the thing that it means to be human is going away. And every time, it's like, ah, no, it's not. It turns out that wasn't what it meant to be human. It wasn't the case that we needed to stitch the fabric manually for it to be a valuable profession to design garments for humans. We keep finding meaning in that work, and that's good for us, but that is not the exclusive domain of human meaning. It turns out we apply the meaning to whatever we put our energy into. And so I think of AI as individuals unbounded.

所以当我和别人合作时，我现在在公司的一个职责是我们所谓的“智能体转型加速器”。就是：我们如何让智能体代表员工工作，帮助他们完成任务？这百分之百就是那个框架。框架是这样的：嘿，现在每个个体度过一天时，他们的工作包含所有这些任务。如果你能拿走其中一半任务，我给你一个工具，你说“这个任务现在为我正确完成了”。那你用剩下的时间做什么？不是什么都不做。你在做有价值的事。你很可能在做只有你才能做的、更大、更独特价值的事。至少对我们来说，我认为我们是在构建帮助人们完成工作的工具，但不仅仅是他们的工作——而是做任何他们想做的事。我们拥有的愿景，不仅针对员工，而是更广泛的，这个“个人超级智能”愿景，是一个非常以人为本的愿景。它非常像：嘿，你想达成什么？你相信自己能做到什么？而今天你做不到，是因为你有大约 10 件事占用了你大量时间，这些事字面上任何人都能做，但没人会替你做。而 AI 可能可以帮忙。

So when I'm working with somebody, and one of the jobs that I have now at the company is what we call the agent transformation accelerator. It's like, how do we get agents that work on behalf of employees to help them with their work? And that is 100% the framing. The framing is like, hey, right now as every individual goes about their day, their work contains all these tasks. If you could take half those tasks and I just gave you a tool and you're like, this task is now done correctly for me. What are you doing with the remaining time? Not nothing. You're doing something of value. You're doing probably something of greater and more unique value that only you are capable of doing. For us at least, I see ourselves as building the tools that are going to help people do their work, but not just their work—do whatever they want to do. The vision that we have, not just for employees but broadly, this personal superintelligence vision, is a vision that's very human-centric. It's very much like, hey, what are you trying to accomplish? What do you believe that you're capable of? And today, you're not able to do it because you have like 10 things that take a bunch of your time that literally anybody could do, but no one's going to do it for you. But an AI could potentially help.

好。所以 AI 做到这一点，突然我一天的一半——理想情况下是我一天中枯燥的那一半——被处理掉了，然后我突然多出很多空闲时间，理想情况下我用它们做更高价值的任务。我不只是拿来刷手机之类的。那整个世界会怎样？这会如何改变你组织的结构？比如，你的工程师是更多了还是更少了？你给他们加薪还是降薪？工程师和产品之间的关系又会怎样？

All right. So AI does that, and suddenly half my day—ideally the mundane half of my day—is taken care of, and then I suddenly have a bunch more hours free, and ideally I use them doing higher value tasks. I don't just use them scrolling or whatever. What happens to the whole world? How does that change the way your organization is structured? Like, do you have more engineers? Do you have fewer engineers? Do you pay them more? Do you pay them less? What about the relationship between engineers and product?

是的，这是个很好的点。为了回答你的问题，我要展开三个分支。第一个是我已经在经历这件事了。写软件是我热爱的事。我整个职业生涯都在做，而今天那些曾经花我大量时间的事几乎不花时间了。这相当疯狂。但我发现，因此我生产软件的速度更快了，但没有我想象的那么快。事实证明，我写代码时做的很多工作，同时也是在思考我在构建什么、让它成形。所以如果不小心，我最终会很快得到一个执行得很好但很糟糕的产品。一个完全构建出来的原型，但东西不好。我之前没真正意识到，我带着脑子里半个想法就跳进编码，然后在编码过程中逐渐打磨这个想法。所以，写东西也是一样。就像，哦糟糕，这是一篇完整的英文文本。读起来通顺，但不好。想法不成熟，结果语料也不成熟。所以我认为我们低估了这一点：我们以为，好吧，写这个代码花了我 10 小时。如果你把这 10 小时拿走，你并不会在零小时里得到同样的项目。因为事实证明，你在编码的同时，也在思考你在构建什么，并随着时间让它变得更好，对吧？所以产品既更好——在 10 小时结束时，如果你真的动手编码，产品更好，而且你更聪明了，因为你花了那 10 小时思考它。而如果你用 AI，你并没有变聪明，产品也不好。现在你就像处在一个奇怪的测试里。所以我现在做的是，我仍然更快地拿到产品。

Yeah, so it's a great point. I'm going to make three spin-offs to this thing to answer your question. The first one is I'm already having this experience. So writing software is something I love doing. I've done it for my whole career, and today things that used to take me a lot of time take almost no time. And it's pretty wild. I'm finding though that I'm producing software more quickly as a consequence, but not as much more quickly as I thought I would. And it turns out a lot of the work I was doing while I was writing the code was also thinking about what am I building and taking shape. And so what I end up with, if I'm not careful, is I will very quickly end up with a well-executed bad product. Okay. A really completely built prototype of a thing that's not good. I didn't really realize how much I was jumping into coding with a half an idea in my head and then while the coding was happening, sharpening the idea over time. So yeah, it's the same with writing. It's like, oh shoot, this is a full English text. It reads, but it's not good. The idea was underdeveloped, and as a consequence the corpus is underdeveloped. And so I think we underestimate the degree to which we think that, okay, oh, it took me 10 hours to code this. If you take the 10 hours away, you don't have the same project at the end in zero hours. Because it turns out while you were coding, you were also thinking about what you were building and making it better over time, right? So the product is both better—at the end of the 10 hours, if you do it and you actually code it, the product is better and you are smarter because you spent those 10 hours thinking about it. And if you use AI, you're no smarter and the product's not good. And now you're like in a weird test. So what I'm doing is I'm still getting the product faster.

我现在花更多时间在前期，比如，好吧，让我先想象我有了这个，或者我在构建很多原型。嘿，这样更好还是那样更好？我就两个都做出来，自己感受一下。所以我更快达到最终结果，但不是快 10 个小时，因为我仍然得做脑力工作，做那些硬核的工作。可以说，我用这些编程工具的一天，充满了更多硬核工作，更多硬核的思考工作。

I'm spending way more time up front being like, okay, let me just imagine that I have this, or I'm building a lot of prototypes. Hey, is it better like this or better like this? I'll just build both and I can feel for myself. So I'm getting to the end result faster, but not 10 hours faster because I'm still having to do the intellectual work, the hard work. Arguably, my day with these coding tools is more full of hard work, of hard thought work.

是的。

Yeah.

比起以前我可以关掉大脑，进入编程状态一段时间。所以我在智力上更疲惫，而且更多工作是我独特能做的。这是一点。第二点，回答你的问题，未来会怎样？我举一个销售领域的好例子。比如，我们发现，我们的销售人员大约三分之一的时间花在客户身上，对吧？三分之一的时间用于准备客户会议。三分之一的时间用于客户会议之后，想办法满足客户的所有需求。

Than when I could kind of turn my brain off and just be in a coding zone for a while. So like I'm more intellectually exhausted and it's more work that I'm uniquely capable of. That's one thing. The second thing to answer your question, like, what's the future look like? Well, I'll give you a good example from a sales discipline. For example, we had found out that hey, our salespeople were spending about a third of their time with clients, right? A third of the time is kind of preparing for those client meetings. A third of the time is kind of on the aftermath of the client meeting figuring out how to, you know, get all the clients demands met.

我们构建了一堆工具，让准备客户会议变得容易多了。所以那三分之一的时间，我们大幅改善了。现在他们三分之二的时间花在客户身上。这意味着，我们花同样的薪水，他们产生了双倍的客户影响。

And we built a bunch of tools that just made a lot easier to prepare for the client meetings. And so it's like a third of their time. We dramatically improved it. Right now they're spending two-thirds of the time with clients. Well, that means that for the same amount of money that we're spending on their salaries, they're having twice the client impact.

那么，你想雇更多人。这是倾向。比如，大多数人在招聘时，会看一条边际回报曲线，就像，每多雇一个销售，带来的收入比前一个少一点，因为他们可 targeting 的广告主池子小一点。所以到某个点，边际回报曲线变平，再雇一个销售就没意义了，因为——但如果能把他们创造的价值翻倍，那就可以雇更多。

Well, you want to hire more. That's, you know, like inclination here. Like if you're having, you know, most people when you're hiring, there's some marginal return curve that you're looking at where it's like, cool, like each incremental salesperson I hire adds a little bit less revenue than the previous person because they have a little bit less of a pool of advertisers to target. And so at some point that marginal return curve flattens out and it makes no sense to hire another salesperson because you can't well if you can double the value that they can create then you can go you can hire a lot more of them

如果广告主数量是无限的，对吧？

if the number of advertisers is infinite right

对我们来说，

which for us for

实际上就是，对吧，小企业。

which effectively is right small businesses

不，所以这是对的，所以我不是说这是——我用销售为例，因为至少对我们来说，这是一条很长的曲线。所以我认为很多工作都有这个特性：如果这个人每单位薪资产出翻倍，

no so so this is right so I'm not saying this is I'm using sales because for us at least it's a it's like a very long curve and like so I think a lot of jobs have the property of if this person was twice as productive per unit salary,

对吧？你会雇更多还是

right? Would you hire more of them or

我们会雇更少？在大多数工作中，我们会雇更多，因为大多数工作你是在这条边际回报曲线上招聘。现在，你说得对。有些情况下，比如，我们只有固定量的工作，

we would hire few of them? We would mostly hire more of in most jobs then you'd hire more of them because most jobs you're hiring down this marginal return curve. Now, you're right. There are worlds where it's like, well, we just have we only have x amount of work,

对吧？比如 QA。Facebook 的 QA，对吧？如果你的 QA 人员速度翻倍，你会雇更多还是更少？这取决于 QA 是否是代码的瓶颈。

right? Like QA. So QA at Facebook, right? Like if you had if your QA people were twice as fast, would you hire more or would you hire fewer? It depends on whether QA is a bottleneck on, you know, code.

没错。这取决于它是否是瓶颈。所以这不是一个统一的故事，但我是说，我认为这会根据你的角色呈非线性。但没错，软件工程师，如果我们认为能做到，我们有很多软件想写。你知道，我们肯定想写更多软件。是的。

That's right. It depends if it's a bottleneck. So it's it's a I'm not saying it's a uniform story, but I am saying like so I think and I think it would be nonlinear based on what role you're in, but yeah software engineers like we have a lot more software we'd like to write if we thought we could do it like you know it's like we we definitely would like to write more software. Yeah.

嗯，我们有很多想法想尝试，只是没有人手。嗯，尽管我们这么大，

Um we have a lot of ideas we'd love to try and just like don't have the people to try. Um as big as we are,

对吧？

right?

所以

So

我不认为影响是统一的，我认为对每个人的体验来说，这非常赋权。就像，是的，我能自己做更多，嗯，需要更少帮助，

I don't think it's a uniform impact and I think for the experience that each individual has I think it's very empowering. It's like yeah, I can do a lot more on my own uh with less help with with

那么你的理论是，Meta 在三年后会因为 AI 而有更多工程师，因为他们都变得更高效。

So your theory then would be that face that Meta would have more engineers in three years because of AI because they all become more efficient.

嗯，是的，但这也很大程度上取决于 AI 本身的经济性。

Well, yeah, but a lot of this depends on the economics of the AI itself too.

嗯。

Mhm.

对。比如，三年内 token 会有多贵。我认为这是原因之一，

Right. Like how how expensive are tokens in a three-year time frame. One of the reasons I think it's

对，因为工程师的成本是工程师的成本加上他们使用的 token 的成本。

right because the cost of the engineer is the cost of the engineer plus the cost of the tokens they use.

没错。这是混合成本。所以真正的问题是，我们不知道。我的意思是，我真的认为，行业如此混乱的原因之一，嗯，不仅在 Meta，而是几乎所有公司，都有巨大的不安，因为这是很长时间以来我们对相对近的未来最不清晰的愿景。

That's right. It's the blended cost. So that's the the real question is like we don't know and when I'm I mean there's a genuine I actually think one of the reasons it is such a chaotic time in the industry um not just at Meta but the across kind of every company and there's a tremendous amount of of unease is like this is the least clear vision we've had of like even the relatively near future in a long time.

是的。

Yeah.

对。这很可能是我职业生涯中最大的转变，我们不知道它会落在哪里。

Right. This is the probably the biggest shift certainly in my career and we don't know where it's going to land.

你是指过去 3 个月，过去三年，还是过去两周？

You mean the last 3 months, the last three years, the last two weeks?

可能是过去——不，可能是过去——可能是过去两年。

Probably the last No, probably the last probably the last two years.

嗯。

Mhm.

但尤其是过去一年，

But like the last year in particular,

对吧？

right?

我认为，两年前我们开始有感觉，三年前我们很兴奋。两年前我们开始觉得，哇，这比我们想的更大。我认为去年它到了严肃的程度，对我们很多人来说几乎成了存在主义对话，不仅是对公司，对行业经济，对行业个人。这是我们很长时间以来对职业长期前景面临的最大不确定性。我倾向于对这些事乐观。我真的拥抱它们。我认为总的来说，当社会创造生产力工具时，这些工具产生更多工作，你知道，更多好的体验，更多机会，而不是更少。

I think I think two years ago we started to get a sense of three years ago we were excited. Two years ago we started to get a sense like wa this is bigger than we thought. I think last year it got to a point of seriousness that it became you know an almost existential conversation for a lot of us like not just you know for the companies for the economics of the industry for the individuals in the industry and it's the most uncertainty we've faced about the long-term prospects of of our profession in a long time. I'm one who's inclined to be optimistic about these things. I really embrace them. I think generally speaking when we create productivity tools as a society those productivity tools generate more jobs you know more great experiences for people more opportunities and not fewer

嗯，我认为每个手艺人、艺术家、关心自己所做之事、相信它的人，有更大能力实现他们脑中的愿景，是好事，对消费者有益，对世界有益，对很多事有益。但从这里到那里的过程会极具破坏性，我们不知道会怎样，因为我们不知道经济会落在哪里。我们不知道模型会落在哪里。我们不知道平台期在哪里。嗯，所以现在对我们所有人都有巨大的不确定性。

u and I think every individual who's a crafts person who's an artist who cares about what they do who believes in it having greater power to deliver the vision that they have in their head is a good thing and it's going to be beneficial to consumers it's going to be beneficial to the world. It's going to be beneficial to a lot of things. But the process of going from here to there is going to be incredibly disruptive and we don't know how it's going to go because we don't know where the economics are going to land. We don't know where the models are going to land. We don't know where the plateaus are. Um so there's a tremendous amount of uncertainty right now for all of us.

有没有可能没有平台期？我的意思是，从这里到那里的概念暗示有一个稳定的那里。有没有可能它只是越来越令人困惑？

Is it possible there are no plateaus? I mean the notion that we're going to get from here to there suggests that there's a stable there. Is it possible that it's just ever more confusing?

我对此持怀疑态度。我的意思是，我认为大多数技术都符合 S 曲线。

I'm skeptical of that. I mean, I think most technology fits on an S-curve.

我之前跟你提过，即使前沿智能继续遵循缩放定律。

And I kind of mentioned this to you before, even if the frontier intelligence continues to obey scaling laws.

是的。

Yeah.

就像摩尔定律那样。你知道，屈服于人们持续而令人印象深刻的努力和投资。我想回到我之前的观点，我认为我们接近这样一个点：实际上模型已经足够智能，可以做很多有用的事情，而我们实际上可能低估了行业中的工作，就是让模型足够容易使用，去做那些有用的事情。现在你有点必须去找模型，并把它集成起来，而不是它来找你并集成到你的工作流程中。我想你看到越来越多的关注，不仅来自我们和消费者，还有微软的 Copilot 和这类东西。

In the way that Moore's laws did. You know, succumbing to constant and impressive effort and investment by people. I think to my earlier point, I think a lot of we're close to the point where actually the models are intelligent enough to do a lot of useful things and we're actually the work that we probably have under index on the in the industry is making the models easy enough to use to go do those useful things. Right now you kind have to go to the model and like integrate it as opposed to like it coming to you and integrating into your workflows. I think you're seeing more and more focus on that not just from us with consumers but Microsoft with co-pilot and these types of things.

我记得在另一个播客上听到你阐述过一个理论，就是工作要么在算法之下，要么在算法之上。对吧？如果你在算法之下，你就是在为算法做事。如果你在算法之上，你就是在指导算法。

One theory I remember I heard you expound on on a on a different podcast was the theory that jobs are either under the algorithm or above the algorithm. Right? If you're under the algorithm, you're doing things for the algorithm. If you're above it, you're instructing the algorithm.

是的。

Yeah.

你说的不就是算法会不断上移、上移、上移、上移、上移。我们中更多人会处于它之下吗？

Isn't what you're saying kind of that like the algorithm is just going to kind of move up, move up, move up, move up, move up. And more of us are going to be under it.

这不是我的框架。我想我是从 Venites Ralph 那里听来的，或者我可能错了。我们得在节目笔记里更正一下。

This wasn't my framing. I I think I got this from Venites Ralph or I could be wrong. We'll have to we'll fix it in the in the show notes.

嗯，

Um,

不管你是怎么得到的，我是在你和其他人交谈时听到的，要么在舞台上，要么在播客上。

however you got it, I heard it from you when you were talking to someone else either on stage or on a podcast.

是的。不，我读过这个东西。我当时觉得它很有说服力。那个人以 Uber 和 DoorDash 为例。他说：“看，现在有一类人就像为机器工作，这是新现象，很有趣，还有在算法之下的人，以及那些为算法编程的人。”是的，我的预测是随着时间的推移，算法会向上移动。嗯，我认为与其对此感到怨恨，我们很可能会感激它。嗯，我举一个可能有点滑稽的例子，但我想你可能能理解。所以，我有，嗯，你知道，一个工作助理，我在纽约，我有这个播客。我在做那个。我们正在做的所有这些事情。现在有一个日程安排。我是个成年人。我有日历。我有能力，我可以叫 Uber。我有能力。我确定从 A 点到 B 点。我相信如果必须的话，我也可以自己安排所有这些事情，我们可以发邮件给你，我们可以协调日程。我可以做，我可以做所有这些。我不，我工作，就像我总是在公司开玩笑说，我工作，我就像一辆小赛车，我的管理员就像，他们从我们的营销团队那里得到消息，嘿，我们需要老板来纽约做这件事，所以他们为我预订了去纽约的行程。

Yeah. No, I I had read this thing. I thought it was quite compelling at the time. And the person was talking about like Uber and and Door Dash as examples. And he's like, "Look, there's like a class of people who like work for a machine now, which is new and that's interesting, and there's under the algorithm, and there's people who are working to program the algorithm." And yeah, like my prediction over time is that the algorithm moves up. Um, and like I think rather than being resentful for it, there's a good chance we'll be grateful for it. Um, I'll give you kind of a funny a silly example perhaps, but I think one that you might relate to. So, I have like um uh you know an assistant at work and I'm here in New York and I'm I've got this podcast. I'm doing that. All these things that we're doing. There's a schedule now. I'm a grown person. I have a calendar. I am cap and I could get an Uber. I'm capable. I am certain of going from position A to position B. And I'm sure if I had to, I could also schedule all these things myself and I could we could email you and we could coordinate schedules. I could do I could do it all. I don't I work like I always kind of joke in the company like I work for my I'm like a little race car my admins like take out like they got um you know they got the word from our marketing team hey we need boss in New York for this thing so they booked me a trip to New York for this thing.

我确实经常觉得我在为算法工作，对吧？

I do often feel like I work for an algorithm, right?

我真的是这个意思，而且这很棒。一点也不坏。我真的很喜欢。我在做我独特有能力做的事情。我不认为其他人能来做我在这里做的事情。嗯，其他人在做他们独特有能力做的事情。他们比我更擅长，而且他们有比我更多的时间去做。

I really mean that and I and it's it's great. It's not bad at all. I really quite like it. I'm doing the things that I'm uniquely capable of doing. I don't think anyone else could come do this the things that I'm doing here. Um, someone else is doing the things that they're uniquely capable of doing. They're better at me at it and they've got more time to do it than me.

所以，你生活中的一部分是为算法工作的。最重要的是

So, some of your life works for an algorithm. The most important thing

我已经有这种感觉了。

I already feel like that.

嗯，嗯，而且不仅仅是那样。我的意思是，我会更进一步，呃，我是说，认识我的人知道我是认真的，这就像在家里一样，你知道我妻子会，我们会有一个日程安排，她设定好了，我对此很兴奋。我就像，是的，我在一个我没有决定的日程中扮演一个角色。所以，我认为这没有我们说的那么陌生。是的。

Um Um, and it's not just that. I mean, I'll go even further and and uh I mean this people who know me know I mean this is love like at home like you know my wife will like we'll have a schedule and like she's got she's got it set and I'm excited about it. I'm like yes, I am playing a part in a schedule that I didn't decide. So, I think it's less foreign than we like to make it sound. Yeah.

嗯，我喜欢它的地方是它最大化每个人的能力，让他们做自己最擅长的事情。嗯，就像我不想处理我的日历。我想做我独特擅长的事情。所以，我很高兴有一组人帮我操作那个。他们确实引导我去某些地方。然后就像在进来的路上，他们会说：“让我告诉你你在哪里。让我告诉你你现在在做什么。这是简报。你读一下。我在走进去之前有五分钟读它。我跟上了。我准备好了。他们知道我能做什么。我知道我能做什么。这很棒。”所以，我认为即使对于相对高层的人来说，在算法内部工作也没有听起来那么陌生。

Um, and the thing that I like about it is it's maximizing each individual for the things that they're the most capable of delivering. Um, like I don't want to deal with my calendar. I want to do the things that I'm uniquely good at. And so, I'm very excited to have the set of people who like operate that for me. And like they do steer me to places. And then like on the way in, they're like, "Let me tell you what where you are. Let me tell you what you're doing right now. Here's the brief. Like you read it. I I've got five minutes to read it before I walk in. I'm up to speed. I'm ready to go. They know what I'm capable of. I know what I'm capable of. This is great. So, it's not as foreign, I think, as it sounds for even people at relatively high levels to be kind of working inside of an algorithm.

据我所知，你们有 Llama 一直到 Llama 4。它是美国最大的开源模型。我实际上理解它的商业逻辑，那就是我们是 Meta。我们拥有所有用户。我们会建立这个开源项目。其他人贡献。他们会帮助我们吸引人，一切都会变得更好，我们实际上不会真的喂养竞争对手，因为我们会把它放在只有我们有用户的地方。所以，这一切对我来说都说得通，然后你说，你知道吗，我们要切换。我们要从开放转向封闭。告诉我那个决定。

As I understand it, you had Llama up through Llama 4. It was America's biggest open source model. And I actually understood the business logic of it, which is we're meta. We have all the users. We'll build this open-source project. Other people contribute. who'll help us attract people, everything will get better, and we won't actually really be feeding competitors because we'll just put into this place where we have users nobody else can. So, it all made sense to me and then you said, you know what, we're going to switch. We're going to go from open to closed. Tell me about that decision.

是的，我不认为从开放到封闭的切换是最终决定。我确实认为我们发现必须采取不同的方法来构建有竞争力的模型。所以，让我从三个方面来阐述。第一件事是开源模型，一直到 Llama 3。我们竞争的所有模型都是单体。就像一个大模型，无论是 OpenAI 还是谷歌或其他谁，每个人都只是生产一个大模型，那个模型包含所有智能，它们很容易相互测试。嗯，在 Llama 3 之后，我认为它相当成功，行业转向了一个非常不同的方向。你现在有这些专家混合模型，你有推理模型，你有 multimodal 模型。很多时候，当你今天问 ChatGPT 一个问题，它内部会引用不同的模型，这些模型擅长不同部分的事情。所以，生产一个单一的开源模型来与所有这些模型集合竞争的想法不再有意义。

Yeah, I don't think the switch from open to closed is a final decision. I do think that we found that we had to take a different approach to build competitive models. So, let me lay this out in three ways. The first thing was open source models uh going through Llama 3. All the models we were competing with were monoliths. There was like one big model whether it was open AI or whether it was uh Google or whoever else they everyone was just producing one big model and that model contained all the intelligence and it was they were kind of easy to test one against another. Um after Llama 3 which I think was quite successful the industry shifted in a very different direction. You now have these mixture of experts models, you have reasoning models, you have multimodal models. very often when you ask today chat GPT a question um it's referring internally to different models that are good at different parts of things. So the idea of producing a single open source model that competed with all of these kind of collections of models stopped making sense.

为什么不直接让开源模型也这样做？

Why not just make the open source model do that?

完全正确。所以我们可能仍然会这样做，但这是第二件事，我们作为一家公司必须经历的过程，以构建有竞争力的模型，必须更加有主见。呃，构建最像，嘿，我只是在构建一个通用智能的东西，实际上并不是大多数主要实验室正在做的。

Totally. And so we may still do this but this is the the second thing is the process we have to go through as a company to build competitive models has to be more opinionated. uh building the most like, hey, I'm just building a generically intelligent thing isn't actually what most of the major labs are doing.

他们对于希望模型擅长哪些类型的任务有非常强烈的观点，并真正针对这些任务进行优化，然后他们会在那些分布外的领域尽力填补这种参差不齐。

They have a very strong point of view of the types of tasks they want their models to be good at, and they really optimize them for those things, and then they fill in the jaggedness as best they can in areas that are kind of out of distribution.

所以如果我理解正确的话，你们在运行 Llama 3，然后在 Llama 3 和 Llama 4 之间发生了一些事情，是不是因为开源的特性导致你们无法预测它们的发展方向？我的意思是，那正好是推理出现的时候，对吧？而且如果我没记错的话，Llama 4 并没有在推理上押那么重的注。难道你们从 Llama 3 到 Llama 4 就不能做出不同的押注吗？

So if I understand this correctly, you're running Llama 3 and then something happens between Llama 3 and Llama 4, and is it the nature of open source that you weren't able to kind of predict where they were going? I mean, that's right about when reasoning happens, right? And like Llama 4 doesn't bet as heavily on reasoning if I remember correctly. Like couldn't you have just made different bets as you went from Llama 3 to Llama 4?

我觉得对 Llama 4 来说可能还有可能，但 Llama 4 之后就不太可能了。因为在那之前我们基本上只押一个注，就是 Llama 这个注。现在你必须开始押很多注，必须开始有不同版本的模型。当然我们仍然有很多开源模型，比如最好的分割模型就是开源的，我们仍在产出大量开源 AI，实际上被其他实验室在很多地方使用。我不会说我们致力于完全放弃开源，但像 Llama 4 这样开源但竞争力不强，对整个行业没有帮助，对我们也没有帮助。所以首先必须做出有价值的模型，然后才能决定哪个版本适合开源。

I think it probably would have been possible maybe for Llama 4, but I think post-Llama 4 it's not super, not as possible anymore. Um, because it's not as much like up until then we were making one bet. We were just making the Llama bet. You have to start making lots of bets. You have to start having lots of different versions of models. Now we do still have a lot of open source models. Um, you know, we still have, you know, the best segmentation model is open source, like we are still producing a lot of open source AI that's actually being used by other labs in lots of different places. Uh, and I wouldn't say that we're committed to getting away from open source in general, but you know building a model like Llama 4 where it's open source but it's not particularly competitive with what's happening, you know, in the industry, that's not helping anybody, not us, not the rest of the industry. So you have to first and foremost make a valuable model and then from there you can make a decision on, okay, you know, what's the version of this that makes sense to be open source.

这是数十亿、数百亿美元的赌注，对吧？而且你们曾经完全致力于开源，然后你们改变了，这意味着有一个时刻。你们什么时候意识到要做出这个转变？

It's billions, tens of billions, hundreds of billions of dollars of bet, right? And I, you know, you guys were fully committed to open source and then you changed, which means there's a day, a week, a month. When did you realize that you were going to make this transition?

是的，我认为一切都关乎价值。当你推出 Llama 4 时，它采用率相对较低，而且在我们推出时它甚至没有真正的竞争力，你会觉得这没有为任何人带来价值。我们开源它没问题，没人说做得不好，只是它相对于当时中国的 DeepSeek、Qwen 等模型没有竞争力。所以如果你生产的东西没有竞争力，那么无论是开源还是闭源都没有意义。你的首要任务是创造价值，

Yeah, I mean I think you know it's all about value and at some point when you put out Llama 4 and it gets relatively little adoption and you know it wasn't really competitive at the time that we put it out even, you're like cool that wasn't a valuable bet for anybody. Like it's all fine and good that we open sourced it, like no one's saying bad job. It's just that it wasn't competitive relative to at that point, you know, Chinese, uh, like DeepSeek and Qwen and these other models that had come out. So there's no point in like going and driving towards an open source model or a closed source model for that matter if the thing that you're producing isn't competitive. Your first and foremost job has to be create the value,

对吧？

right?

然后你就有选择了。一旦创造了价值，你可以开源，也可以不开源，有很多不同的方式来分发价值。就像我说的，我不会说我们转向了闭源。我会说我们转向了确保首先创造有价值且能用的模型。然后我们可以决定什么适合开源。我还要提到，模型与传统开源非常不同。开放权重与开源非常不同。开源时，会形成一个社区，他们可以回馈，随着时间的推移变得自我维持和自我强化，因为在这些模型中，权重的方式是，人们可以在其下游构建架构，但无法真正反馈到模型本身。所以每次新模型出来，对开源社区、开放权重社区来说都是全新的。因此，我认为有一个故事，这些模型有一个商品层，来自中国的开源模型，就像是的，那是一个后盾，每个想为他们的模型收取高额费用的人都必须大幅超越那些模型，而那些模型最多落后最先进技术六到八个月，也许 12 个月。

From then you have options. Like once you've created the value, you can open source, you can not open source, you can do there's lots of different ways to distribute that value. Um, and like I said, we're not, I wouldn't describe us as like having switched to closed source. I would switch us to having like, hey, let's make sure that our first and foremost we're creating models that are valuable that work. Um, and from there we can decide what makes sense to open source or not. I'll also mention that models are very different than conventional open source. You know, this open weights is very different than open source. When you have open source, there's an entire community that forms around it and they can contribute back to it and it becomes self-sustaining over time and self-reinforcing because the way weights are in these models, people can build architecture downstream of them, but they can't really feed back into the model itself. And so every time a new model comes out like it's a completely fresh, you know, take for the open source community, open weights community. Um, and so I do think there's a story the models, there's a commodity layer of these models, the open source models that are coming out of China where it's like yep that is a backstop and everybody who wants to charge a lot for the models that they've built has to beat those models by a lot and those models are like six to eight maybe 12 months behind at most uh kind of the state-of-the-art.

好吧，这就引出了我问题背后的一个点，就是 Yann LeCun 和其他人提出的一个理论，即最终智能会商品化，前沿模型和最好的开源模型之间的差距会很小，如果差距很小，那还不如用开源模型。

Well okay so this gets to one of the things that's underlying my question which is there was this theory that Yan Lun and others put out which is that eventually intelligence would kind of be commoditized right that the gap between the frontier models and the best open source models it would be quite narrow and if it's quite narrow then might as well have the open source models

对。

right

现在差距是 6 到 12 个月，3 年后差距会是多少？

right now the gap be say 6 to 12 months what will the gap be in 3 years

是的，我们不断了解到的一件事是，前沿模型的性能一直非常稳定，缩放定律一直成立，随着我们不断投入更多算力，我们克服了每一个我们认为会遇到的障碍，比如智能曲线趋于平缓。我们一直在克服它们。我们以为会遇到数据障碍，然后我们使用合成数据，效果很好。所以我们正在突破这一点，缩放定律成立，这意味着前沿模型随着时间的推移会越来越智能。话虽如此，

yeah well we keep you know um one of the things that's the the industry keeps learning about AI is that the frontier has been very consistent in its performance and the scaling laws have held and as we keep throwing more and more compute at these things we've kind of gotten over each barrier that we thought would hit it um in terms of seeing the intelligence curve flatten out. We've just keep overcoming them. We thought we'd run into this data barrier and then we're using synthetic data and it's been quite effective. So, we're kind of powering through that and so the scaling laws are holding which means the frontier continues to get smarter and smarter over time. Having said that,

完成工作所需的智能看起来并不是这样。它更像是一种分布，如果不是正态分布，至少是随时间平坦的分布。所以我认为，越来越智能的模型正在缩小在超级困难的事情上的差距，比如编码非常复杂的任务，这很重要，但有很多工作不需要前沿级别的智能就能完成。所以你未来几年的假设是，模型的智能继续以同样的速度增长，但智能增长的边际价值有所降低，因为大多数任务不需要那种智能。

the like intelligence required to do jobs doesn't look like this. Like it has a a much more if not a normal distribution at the very least a flat distribution over time. Um, and so I think you know the smarter and smarter models are closing the gap on super hard things coding really complex tasks which is which is important but there is a lot of like there's a lot of jobs that don't require frontier level intelligence to accomplish. So your hypothesis for the next couple years is that the models continue to grow in intelligence at the same rate, but the marginal value of that growth of intelligence is somewhat less because most tasks don't require that intelligence.

我们还没到那一步。我认为在某个时候这会发生。是的。

We're not quite there yet. I think at some point this will happen. Yes. Um there's

我以前没听过这个理论。

I haven't heard that theory before.

是的。嗯，我认为很多时候，如果我考虑大多数工作，我觉得我无法去做那个人的工作，不是因为我不够聪明，而是因为我不知道，我没有判断力，没有经验。我认为智能是一个超级重要的因素，我认为那些几乎虔诚地信奉递归自我改进和 AGI 起飞的人相信智能会弥合所有差距，就像智能和算力会弥合所有差距。我认为这超级重要。

Yeah. Well, you think it's like I think a lot of times we if I like, you know, think about most jobs, I don't think I could go do that person's job and it's not because I'm not smart enough to do it. It's because like I don't know. I don't have the judgment. I don't have the experience. I think like intelligence is is a super important factor and I think the people who are the really almost religious adherence to recursive self-improvement and and the AGI takeoff believe that intelligence will just close all gaps like intelligence and compute will close all gaps. I think it's super important.

我认为智能最重要的原因是，它能让这些模型长时间运行，而错误不会随时间累积，对吧？还有长时间运行的任务。这很重要，对吧？可以说，比拥有原始智能更重要的是，有一个不会内部累积错误的长时间运行任务，对吧？

The reason I think intelligence is the most important thing is that it allows these models to run for a long time and not have errors compound over time, right? And long-running tasks. That is important, right? Arguably the more important thing than having raw intelligence is having a long-running task that doesn't have compounding error inside of it, right?

所以我非常相信缩放定律，我亲眼看到它们不断超出预期，也看到我们不断克服那些差距。这不是凭空发生的。这让我想起摩尔定律。摩尔定律，即硅片上晶体管数量每 18 个月翻一番，我们谈论它时仿佛它是自然规律。是吗？不，不，那是经过一代又一代的巨大努力，投入数十亿美元在极紫外纳米光刻等技术上，才使其持续下去。AI 中的缩放定律也是如此。它的发生不是免费的。它需要大量硬件。

And so I really am a huge believer in scaling laws and I've seen them continue to defy expectations and I've seen us continue to overcome those gaps. It's not happening for free. It kind of reminds me of Moore's law. Moore's law, the doubling of transistors on silicon every 18 months, was a law that we talked about as if it was the state of nature. Was that that? No, no, it was an incredible amount of effort successively over generations with huge and billions of dollars invested on things like extreme ultraviolet nanolithography to cause that to continue. The same thing is happening with scaling laws in AI. It's not for free that it's happening. It's happening with a lot of hardware.

但摩尔定律的惊人之处在于，这些翻倍的价值回报从未递减，对吧？随着芯片翻倍，一切变得更强大，从未出现某个时刻它不再重要。每次翻倍似乎都同样重要。

But the amazing thing about Moore's law is that there's never a decreasing return in value to those doublings, right? As the chips double and everything becomes more powerful, it's never that we reached a moment where it didn't matter. Like each doubling seemed to matter as much.

是的。现在作为一个行业，我们某种程度上已经脱离了摩尔定律，我真的感受到这对我们预测算力发展的方式有多么不同，我们不得不在问题的另一面，而不是晶体管那一面，变得更加聪明。

Yeah. And now that we're kind of off of Moore's law as an industry, I really do feel how different that is for us in terms of how we project compute forward and we're having to be more clever on the other side of the problem as opposed to the transistor side.

好吧，我们谈谈这个。你提到你在 Llama 4 上没有做出正确的赌注，或者从结构上讲，在开源中很难做出正确的赌注。Muse Spark 的赌注是什么？比如你认为 Muse Spark 在哪里能击败 Claude，能击败 ChatGPT，能击败 Gemini？

Well let's get there. So you mentioned that you hadn't made the right bets at Llama 4 or structurally it was very hard in open source to make the right bets. What are the bets in Muse Spark? Like where do you think that Muse Spark can beat Claude, can beat ChatGPT, can beat Gemini?

是的，很有趣。有几件事。我昨天在录播客时，使用它的人是一个 Meta AI 的超级粉丝。他们经常使用眼镜，我实际上印象很深刻。他们抓住了我认为我们真正擅长的三件事中的两件。一是本地。我认为这很合理，因为我们在这方面投资了很长时间，不仅在 Facebook 和 Meta，而且也为眼镜投资，知道我们会在这些眼镜中使用它。Meta AI 在本地推荐方面绝对是最好的。如果你想找附近的咖啡店，它真的非常擅长。它在商业、购物、帮助人们做决定方面也非常擅长。再次，如果你想想我们如何在构建的产品中部署 Meta AI，这就说得通了。这对我们来说是一个非常重要的用例。它在健康方面也很擅长，帮助你了解如何在医疗系统中导航等等。那是因为我们在 Meta 超级智能实验室有一组研究人员对此非常热情，并竭尽全力确保模型具备这种能力。

Yeah, it's funny. There's a couple things. I was at a podcast recording yesterday and the person who was using it was a huge Meta AI fan. They've used the glasses a lot and I was actually quite impressed. They had zeroed in on two of the three things I think we're really really good at. One is local. I think this makes sense because we've been investing in this for a long time, not just at Facebook and Meta, but then also for the glasses knowing that we were going to have these glasses in context. Meta AI is absolutely the best at local recommendations. If you want to find a coffee shop nearby, it's actually really really good at it. It's really really good at commerce, at shopping, helping people make decisions there. Again, makes sense if you think about how we deploy Meta AI inside of the products that we build. That's like a really important use case for us. It's also really good at health, helping you understand how to navigate healthcare systems and that kind of thing. And that's just because we had a set of researchers in Meta Super Intelligence Labs who were very passionate about that and went overboard in making sure that this was something that the model was capable of.

我认为我们低估了所有模型都有这些优势和劣势领域，这反映了构建它们的研究人员的热情领域。有趣。它们并不像我们描述的那样统一，尽管它们确实弥补了这一点。

And I think we underestimate the degree to which all the models have these areas of strength and weakness which reflect the passionate areas of the researchers who build them. Funny. And they're not as uniform as we kind of make them out to be, although they do make up for it.

但你在训练方式上的优势是什么？比如你认为你找到了更好的多模态方法吗？你找到了更好的推理方法吗？你有更多数据，因为你能从 WhatsApp 输入东西，对吧？你在构建方式上的优势在哪里？

But what are the strengths in how you train it? Like do you think you figured out a better way of doing multimodality? Have you figured out a better way of doing reasoning? Do you have more data because you're able to feed things in from WhatsApp, right? What is it that you—where do you have the edge in how you're building it?

是的。

Yeah.

那你在哪里没有优势？

And where do you not have the edge?

所以，如果你是像 Anthropic 或 OpenAI 这样的公司，尤其是有来自企业客户的这些查询流，你知道，这就成了一个循环，就像我们长期以来谈论产品的任何产品循环一样。产品循环就像一旦我有产品上市，我就能从中学习，我现在比没有产品上市的人进步得更快。

So, if you're someone like Anthropic or OpenAI and you have these query streams coming in from enterprise customers especially, you know, that becomes a loop that you're in and just like any product loop that we've been talking about products for a long time. Like a product loop is like once I have a product in market, I'm able to learn from that and I'm now getting better faster than somebody who has no product in market.

这非常有趣。对。所以你们全是消费者业务，对吧？你们没有任何企业客户。

That's super interesting. Right. So you're all consumer, right? You don't have any business customer.

我们非常专注于消费者。现在我认为我们实际上——我要说的一个例外是，我们实际上有一个很棒的小企业套件。比如我们为广告商提供的小企业智能体绝对是同类最佳。这是另一个我认为我们可能不完全——我们有点低调的领域。如果你是美国的小企业，实际上在世界各地，Meta 是生命线。就像在 Meta 上做广告是大多数这些企业触达客户的唯一可扩展渠道，肯定比通过印刷品或视频或电视更有效率。所以我们现在能够给每个人一个超级智能的 AI，帮助他们业务增长，这很棒。这不仅会帮助他们的业务，还会帮助我们的消费者获得更好的广告。就像一切都是一场胜利。所以我们在做——我们确实为小企业有一个很好的循环。

We're very consumer focused. Now I think we actually—the one exception I'll say is like we actually have a great small business set. Like our small business agent that we have for our advertisers is absolutely best-in-class. It's another area where I think we're probably not entirely—we're a little under the radar. If you're a small business in America, actually around the world, Meta is lifeblood. Like the advertising being done on Meta is the only scalable channel that most of these businesses have to reach their customers, certainly much more efficiently than they could through print or through video or TV. And so our ability now to give each of those people a super intelligent AI that's going to help their business grow is great. It's not only going to help their businesses, it's going to help our consumers get better ads. Like everything about it is a win. So we are doing a—we do have a great loop for small businesses.

等等，那就是循环，我的意思是我们在 Meta 上为《大西洋月刊》买了很多广告。那就是循环，我们把故事 feed 输入你的算法，你为那个故事找到受众，然后我们获得订阅。你的理论是这非常有效。

Wait and that's that is the loop where I mean we buy lots of ads for the Atlantic on Meta. That is the loop where we input a story feed into your algorithm, you find the audience for that story, then we get subscriptions. And your theory is that that is super effective.

非常有效。我认为它只会变得更加有效。对于很多小企业来说，它甚至更进一步。如果你有库存，就像直接跟 AI 分享，比如嘿，这是我想要卖的东西，你能帮我吗，它能给你定价策略吗，能给你折扣策略吗，所以这不仅仅是嘿我们可以为你投放更好的广告，我们可能能够最大化你的全部成果。所以我们有一堆我们专注的领域，我认为这些领域相对独特。现在公平地说，消费者业务在某些方面是有争议的。外面有很多消费者行动。但即使 Anthropic 在过去一年在企业客户方面取得了成功，你真的看到行业正朝着那个方向移动，试图追求那些美元。

Super effective. And I think it's only going to get a lot more so. And for a lot of small businesses it goes even further. If you're somebody who has inventory, literally like sharing to the AI like hey here's what I'm trying to sell, like can you help me, can it give you pricing strategies, can it give you discounting strategies, so it's not just like hey we can run you better ads, we might be able to maximize the full outcomes that you have. So we have a bunch of areas that we're focused on that I think are relatively unique to us. Now to be fair, consumers is in some ways contentious. There's lots of consumer action out there. But even with the success Anthropic has had over the last year with enterprise customers, you really see the industry moving in that direction to try to pursue those dollars.

对我们来说，我们确实继续把消费者作为主要客户，对吧？所以 Anthropic 和 OpenAI 都大幅转向企业市场，因为钱在那里。你没法真的去那里。那不是你的品牌。不是你的公司。不是你的专长。我的意思是，你卖广告是，但你不会为一家大型科技公司提供编程工具。

It does feel like for us we continue to be focused on consumers as our primary customer, right? So Anthropic and OpenAI have both very much pivoted towards enterprise because that's where the money is. You can't really go there. It's not your brand. It's not your company. It's not your expertise. I mean, you obviously selling ads is, but you're not going to provide a coding tool for a big tech company.

是的，我们不反对，但那不是我们有动力的地方。我们公司是关于消费者的。当你的工程师写代码时，他们用 Claude Code 还是用 Muse Spark Code？

Yeah, we're not opposed to it, but it's not the place that we're motivated. Like our company is about consumers. When your engineers write code, do they use Claude Code or do they use Muse Spark Code?

我们用，我们内部有一个叫 MetaCode 的工具，它基于一个叫 Open Code 的开源平台，我们只是在内部调整它来使用我们自己的模型。我们用那个，我们用 Codeex，我们用 Claude，我们用整个谱系，这真的取决于任务，哪个对每个任务最好。你知道，我认为不只是我们，我认为现在很多公司，你不想在你的核心工作流中严重依赖第三方公司。就像如果一个第三方

We use we have a tool called MetaCode internally which is based on an open source platform called Open Code that we just have adapted internally to use our own models. We use that, we use Codeex, we use Claude, we use the whole spectrum and it really depends on the task which one is going to be best for each one. You know, I think for not just us, I think a lot of companies right now, you don't want to have in your core workflow a major dependency on a third party company. Like if a third party

所以我听到

So I hear

可能轻率地决定你不再被允许使用我们的工具

could be flippant and decide you're not allowed to use our tool anymore

或者美国政府可能决定你不被允许使用这个工具

or the US government could decide you're not allowed to use the tool

如果那样，你无法在交付客户期望的价值方面取得进展。那将是一个非常具有挑战性的结果，对吧？所以对我们来说，我们非常专注于确保，但我们也想确保我们有最好的工具可用。所以这些编程平台在编程领域实际上都有优点和缺点。它们更擅长的事情和更不擅长的事情。

if then and you were not able to make progress in delivering the value that your customers expect. That would be a very challenging outcome, right? And so for us we're very focused on making sure that but we also want to make sure that we have the best tools available. So each of these coding platforms actually do have strengths and weaknesses inside the space of coding. Things that they're better at and things that they're worse at.

是的。

Yeah.

所以我们在构建，我们有自己的平台，我们用自己的模型使用。我们继续迭代和扩展它，特别是因为我们有很多代码是分布外的。你知道，我们有很多，因为很多这些编程模型主要基于开源代码训练。如果你使用开源代码，它们非常非常擅长解决你的问题。如果你使用的代码像我们一直在构建的代码，写了 20 年代码，其中一些代码从未见过天日。所以模型从未见过它。所以在我们自己的代码上训练我们自己的模型实际上给了我们优势，使那个代码更有效。所以它真的变化。我认为只要你随时在模型之间切换，你就处于一个相当安全的地方。我认为如果你变得过于依赖单一模型，那就是危险区，

So we're building we have our own platform that we use with our own model. And we're kind of continuing to iterate on that and expand it especially because we have a lot of code that is out of distribution. You know we have a lot of there's a lot of in so far as these coding models are largely based on trained on open source code. If you are using open source code they're really really good at solving your problems. If you're using code like we've been building code, writing for code for 20 years, some of that code has never seen the light of day. So the models have never seen it before. So training our own model on our own code actually gives us advantages on making that code you make more effective on that code. So it really varies. And I think as long as you're swapping readily between models, you're in a pretty safe place. I think if you become too dependent on a single model, that's the danger zone,

对吧？然后他们突然要么对你有了杠杆，要么你有了以前没有的风险因素。在最坏的情况下，他们有定价权。我想我应该是正确的。在最坏的情况下，有实际的禁令，他们不能做事情，或者他们只是不得不拔掉插头，但就像定价权。而且经济上不可持续。但我认为反过来也是真的。最好的情况实际上是如果你自己构建，你有更大的行动能力，比任何人能提供的都大，因为这个模型是在你自己的代码和你自己的工作系统上训练的。

right? And then you they suddenly have either they have leverage on you or you have a risk factor that you didn't have before. In the worst case, they have pricing power. I think I was supposed to be right. In the worst case, there's like an actual injunction that they're not able to do things or they they just had to pull the plug, but like pricing power. Um, and it's like economically not sustainable. Um, but I think the flip side is true. The best case actually is if you build your own, you have a greater ability to move than anybody else can provide for you because this model is trained on your own code and your own systems of working.

最后两个问题。你认为世界对 AI 最大的误解是什么？现在，

Two last questions. What do you think the world gets most wrong about AI? Right now,

我们忘记了炒作周期。就像所有技术都遵循一个非常一致的模式，你知道，巨大的炒作，然后是不满的低谷，然后是通往长期成果的艰难道路。我发现自己有时在 AI 领域没有朋友，因为我对这项技术在未来五到十年内将把我们社会带向何方非常乐观。我认为它会令人难以置信。我认为我们会非常感激。我认为它会像互联网或手机一样大，或者甚至像计算本身一样，那种级别的美好。一件我们无法想象没有它、不想使用的东西。它会在很多方面让我们变得超级强大。它会真的，这很老套，它会真的治愈疾病。它会真的为我们解决我们认为无法解决的真正问题，这太令人兴奋了。而且，我认为在未来五年内要达到那里会是一场真正的苦战，会有很多颠覆、变化、冲突、浪费的钱和花得好的钱。就像我认为它会，你知道，现在有很多不确定性。所以我发现自己经常对 AI 爱好者提出警告，对 AI 怀疑者试图鼓励他们。

we forget about the hype cycle. Like all technologies follow a very consistent pattern of you know tremendous hype then the valley of discontent and then a grinding path towards long-term outcome. I find myself sometimes with no friends in AI because I am tremendously positive on where this technology is going to take us as a society in like a five or 10 year time frame. I think it's going to be incredible. I think we're going to be so grateful. I think it's going to be a thing of the size of the internet or mobile phones like or maybe even computing itself like that level of good. a thing that we won't can't imagine not having with us and not wanting to use. And it'll make us superpowered in so many ways. And it'll like it really will, it's so cliche, it really will cure diseases. It really will solve real problems for us that like we thought were unsolvable, which is so exciting. And also, I think it's going to like be a real grind over the next like 5 years to get there with like a lot of just disruption and change and and and and strife and like misspent money and and then well spent money. It's like I think it's going to be, you know, there's a lot of uncertainty right now around it. And so I find myself often with AI enthusiasts cautioning them and with AI skeptics trying to encourage them.

我在 AI 阵营里有点没有朋友。

I kind of have no friends in the AI camp.

如果你有无限的预算，你可以花在任何事情上，你有点这样做，但不完全，你会花在哪里？这是我问所有嘉宾的问题。我们对这个问题有各种各样的疯狂答案。

And if you had an unlimited budget and you could spend it on anything, which you kind of do, but not completely, where would you spend it? That's a question I ask all my guests. We have had a whole range of crazy answers to this question.

能源。我想我只是认为这一切都是能源的下游。就像我们把能源变成所有这些事物。我们把能源变成智能。我们把阳光变成智能，我们把沙子变成思考机器，这太不可思议了。我们真的把能源变成智能。我只是认为能源是所有这些部分中最上游的向量。而且我认为我们对自己造成了巨大的伤害，你知道，通过过度监管核能，阻止了它变得安全和廉价。如果我们真的深思熟虑，我们会有安全和廉价的电力。我是一个热心的环保主义者，我认为讽刺的是，环保主义者对环境造成了如此大的损害，创造了我们现在的燃煤现状，通过切断一条更有前途的技术道路。顺便说一句，我有点担心同样的事情发生在 AI 上，我担心我们把它放在核能的桶里，就像酷，让我们在它变得有用之前杀死它，而不是让它变得有用，然后帮助我们，你知道，治愈疾病和其他事情。

Energy. I think I think I just think I think it's all a downstream of energy. Like we're turning energy into all these things. We're turning energy into intelligence. We're turning sunlight into int we turn sand into thinking machines which is incredible. We're turning energy really into intelligence. I just think energy is the the most upstream vector of all these pieces. And I think we did ourselves a tremendous disservice, you know, by overregulating nuclear that prevented it from being safe and cheap. And if we had been really thoughtful about it, we'd have safe and cheap power. I'm an avid environmentalist and I think ironically the environmentalists did so much damage to the environment creating this like coal burning present that we have by by cutting off a much more promising technological path. I worry a little bit about the same thing happening to AI by the way like I I worry about us like putting it in the nuclear bucket where it's like cool let's kill it before it becomes anything useful as opposed to letting it become something that's useful and then helps us you know cure diseases and these other things.

你认为这个的原因是因为你看到了欧洲和加州的立法，还是因为最近 Anthropic 发生的事情，政府在一个周五下午关闭了 Fable？

Is the reason you think that because you've seen legislation in Europe and California or because of what happened to Anthropic recently where the government sort of shut off Fable on a Friday afternoon?

不，我的意思是，对我来说，回到你的问题，真的回到核能，如果你有一个选择，这真的是核委员会在 60 年代面临的选择，对吧？就像嘿，你可以继续这个进步，或者你可以制定这个规定，保证更安全。

No, I mean I'm really for me that coming back to your question like really goes come back to like nuclear power where if you have a choice and this is really the choice that the nuclear commissions were faced with in the 60s, right? Was like hey you can either continue to this progress or you can like make this regulation that makes it guaranteed more safe.

听起来通过监管来保证更安全总是正确的做法，但到了某个节点，你实际上把限制设得太大，以至于扼杀了这件事本身的经济性，最终你会落得一个烧煤 60 年的世界。

It always sounds like the right thing to do is to pass the regulation that makes it guaranteed more safe, but at some point you've actually made the restrictions so great that you killed the economics of the thing itself, and you end up in a world where you're burning coal for 60 years.

嗯，这就是为什么 AI 会在法国腾飞，就像核电一样。

Well, this is why AI is going to take off in France, just like nuclear power.

你可是最早在这里听到的。

You heard it here first.

好的。非常感谢你，Andrew Bosworth。

All right. Thank you very much, Andrew Bosworth.

谢谢邀请我。

Thanks for having me.

刚才这位是 Meta 的 Andrew Bosworth。如果你喜欢这期节目，请订阅本频道以获取更多内容。如果你是第一次观看，请留言介绍一下你自己。告诉我们你这周是如何使用 AI 的，或者为什么没有用。感谢观看。

That was Andrew Bosworth of Meta. If you enjoyed this episode, please subscribe to this channel for more. And if this is your first time watching, please leave us a comment and introduce yourself. Tell us how you've used AI this week or why you haven't. Thanks for watching.