Meta CTO 安德鲁·博斯沃思：我们的前沿 AI 之路、租用模型与消费级 AI 的困境

Meta 首席技术官 Andrew Bosworth 加入我们，聊聊公司的 AI 努力以及为什么要打造自己的新 AI 眼镜。马上开始。欢迎来到 Big Technology Podcast，一档冷静而细致地探讨科技世界及其之外的节目。今天我们请来了 Meta 首席技术官 Andrew Bosworth，他将和我们聊聊公司的 AI 努力、新的 AI 眼镜、公司文化，最后还有一些宏大的思考。Bos，很高兴见到你。欢迎回到节目。

Meta Chief Technology Officer Andrew Bosworth joins us to talk about the company's AI efforts and why it's building its own new AI glasses. That's coming up right after this. Welcome to Big Technology Podcast, a show for coolheaded and nuanced conversation about the tech world and beyond. We have a great show for you today. We're joined today by Meta Chief Technology Officer Andrew Bosworth, who's going to talk to us all about the company's AI efforts, its new AI glasses, the company's culture, and some big thoughts at the end. Bos, great to see you. Welcome back to the show.

嗯，谢谢邀请我。

Well, thanks for having me.

我们刚开始录制前还在聊，科技世界现在是个多么疯狂的时刻。据我所知，我们从未见过这样的进步。核心部分是 AI 模型。AI 模型支撑着一切。没有可用的 AI 模型或领先的 AI 模型，就很难构建。长期以来的理论是，要构建一个伟大的 AI 模型，你需要大量的算力和优秀的研究人员来研究算法。Meta 有大量的算力和最优秀的研究团队来研究算法，但领先的 AI 模型尚未出现。那么，你能谈谈你在这方面学到了什么，以及关于打造伟大 AI 模型所需条件的核心假设是否错了？

We were just talking before we started rolling about what a crazy moment it is in the tech world. We haven't seen progress like this as far as I can remember. The core part of it is the AI model. The AI model underpins everything. Without a working AI model or a leading AI model, it's tough to build. The theory for a long time was that to build a great AI model, you needed a ton of compute and great researchers to work on the algorithm. Meta has a ton of compute and a team of the best researchers to work on the algorithm, but the leading AI model hasn't materialized yet. So, can you talk a little bit about what you've learned there and whether that core assumption about what it takes to make great AI models is wrong?

嗯，我唯一想补充的另一个要素是优质数据。你们有，我想我们也有。所以，是的，这里有两个故事。第一个是，你知道，我们回溯到 Llama 1、Llama 2、Llama 3。我们确实处于前沿，推动着进展，你当然知道这一点。Facebook 的基础 AI 研究团队可以再往前追溯十年。我其实最早接触 AI 是在 AI 消息机器人出现在我的信息流里，然后我遇到了 Yann，开始认识 FAIR 的人，当时就觉得这项技术进展得非常快，所以 Meta 很早就介入了。而真正的差距，我认为已经相当公开了，是我们当时没有真正提出的，是当我们整合 Llama 4 时，抱歉，是当我们整合 Llama 3 时，我们真的把所有研究、所有我们拥有的每一分力量都投入了进去，无意中扼杀了后续管线。研究人员，你知道，运作方式是，你建立一个基础，有人开拓基础的增量版本，有人在外探索全新的策略。而当时我们不知道，这也说明我们没有足够专注，Llama 3 是一个很棒的模型，广受好评，为了达到那个模型，他们把所有未来的赌注都提前拉进了那个模型。这意味着当 Llama 4 到来时，我们没有其他实验室仍在进行的探索。所以，现在你在推理上落后了。现在你在专家混合上落后了。现在你在一系列用于保持进步步伐的关键技术上落后了。这是一年前我们相当公开的失望，导致 Mark 从“好吧，AI 不是我们的赌注之一”——那是我们当时的想法，AI 只是我们众多赌注之一——转变为 AI 是对整个公司至关重要的赌注，所以我们要改变思考方式。这很老套，但我找不到更好的词。幽灵创始人模式，他真的切换到了一个独特且专属于 Mark 的模式，他变得如此专注于为我们获取所需的全部算力、所需的所有人才、我们签下的研究人员，你说你知道，他们大约一年前真正到位了。我想 Alexander Wang 刚满一年元宇宙周年，我很喜欢和他合作，已经从他身上学到了很多，我们正在看到成果。所以如果你看 Muse Spark，那是你们最新的模型，不是我们的前沿模型，但它是我们发布的最新模型，非常非常受欢迎，根据基准测试，它在我们最关心、我们认为对我们产品独特的事情上表现很好。所以，是的，你完全正确，就公众对我们在模型方面的看法而言，我们已经建立了团队，我真的相信，我们拥有所需的所有算力和数据，所以我非常有信心我们会达到需要的位置。我要补充第二点，我认为这在战略上非常重要，那就是模型是现成的。你可以去租一个模型。你可以用 Anthropic 的。你可以用 OpenAI 的。你可以用谷歌的。它们都是很棒的模型。你可以去获取它们。你可以使用它们。这很棒。我们在世界上创造的真正价值是产品。而我们为个人超级智能所设想的愿景，我认为我们特别适合去实现。不仅仅是因为我们有数据。那很酷。我们实际上比几乎任何人都更有机会理解你、你想做什么、你在世界中的身份以及什么对你重要。所以，拥有模型是一部分，你希望从战略上拥有它，这样你就不会依赖别人，但你主要是想用它来控制自己的命运。模型本身不是价值。我认为我们很快就会进入一个消费者不在乎的世界。他们不想指定使用的模型。他们不在乎是 4.7 还是 4.8，就像你不在乎是用 Oracle 还是 SQL 数据库一样，你只想要功能，你只想要东西好用，我认为这就是我们所有人都将被要求达到的标准。所以今天讨论的是模型，这至少对我来说表明我们在用户侧以及人类如何受益方面有点投入不足。所以我认为这是我们需要讲述的故事，除了展示我们在技术上所做的工作，我们还需要真正向消费者展示价值。

Well, the only other ingredient I would add is great data. And you have that and we do have that I think as well. So, yeah, there's two stories here. The first one is, you know, I think, you know, we go back to Llama 1, Llama 2, Llama 3. We really were, you know, kind of at the forefront and advancing things and and you of course know this. The Facebook fundamental AI research group goes back a decade more. I've been I mean that's where I actually first got queued into what was going on with AI is when the AI messaging bot popped up in my feed and then I met Yan and started to meet the FAIR people and was like oh this technology is progressing really fast so Meta was on it very early and and so the real gap which I think has been pretty public was what we didn't really raise at the time was when we were pulling Llama 4 together sorry when we were pulling Llama 3 together we had really pulled in all the research all the every we pull out every single stop we had and unwittingly kind of killed the pipeline. So researchers, you know, the way that it works is you build a base and you've got people pioneering an incremental version of the base and you got people out there pathfinding entirely new strategies. And kind of unbeknownst to us at the time and kind of speaks to the fact that we weren't focused enough on it, Llama 3, which was a great model and was well-received to get to that model, they had kind of pulled forward all the future bets into that to deliver that model. Well, that meant when it came time for Llama 4, we didn't have any of the pathfinding the other labs still had going. So, that makes you now now you're behind on reasoning. Now you're behind on a mixture of experts. Now you're behind on a bunch of these critical technologies that have been used to continue continue the pace of progress. This is a pretty public you know disappointment I think a year ago for us and led to Mark shifting from okay AI isn't one of our bets which is how we thought of it up to that point. AI was just one of the many bets we had. AI is a bet that's foundational to the entire company and so we're going to change how we're thinking about this. And this is such a cliche but I don't have a better word for it. Ghost founder mode like he really did flip into a mode that is like unique and reserved for Mark that where he just became so focused on getting us all the compute we needed getting us all the talent that we needed the researchers that we've signed that you said you know and they really landed about a year ago I think I think Alexander Wang just hit his one-year metaversary and I have loved working with him and I've learned so much from him already and we are seeing the fruit of that so if you look at Muse Spark which is your latest model which not our frontier model but it's the latest model that we've released to focus very very well received model and depending on the benchmark it does really well on things that we care the most about that we think are unique to our products and so yeah you're absolutely right on where we are in terms of what the public you know perception of it is model wise we've built the team I really believe in uh we've got all the compute and the data that we need so I'm very confident that we're going to be where we need to be I'll add a second piece to this which I think is strategically very important though which is that you So models are available. Like you can go rent a model. You can go use Anthropic's. You can use OpenAI's. You can use Google's. They're great models. You can go get them. You can use them. And that's pretty great. The real value we're going to create in the world is the product. And the products that we the vision that we have for personal super intelligence, I think, is a vision that we're uniquely suited to deliver. It's not just that we have data. That's cool. we actually have a better chance of understanding you and what you're trying to do and who you are in the world and what matters to you than I think almost anybody else does. So, having the model is one piece and you want to have that strategically so you don't have a dependency on somebody else, but you mostly want to be able to control your destiny with that. The model itself isn't the value. And I think we're going to get to a world very soon where consumers, they don't care. They don't want to specify the model they're using. They don't want to they don't care if it's 4.7 or 4.8 like you don't care what Oracle if I'm using Oracle or SQL databases like you just want the functionality you want the thing to work well that's the standard to which I think we're all going to be held so today the discussion is about models which suggests to me at least that we're a little underindexed on the user side of it and how humans are going to benefit so I think that's the story that we need to tell in addition to showing the work that we've done technically we need to actually demonstrate the value to consumers

所以我想从科学角度谈一下。我一开始提到的是，你可以通过蛮力方式打造一个有竞争力的模型。

So I just want to talk about this scientifically for a moment. The thing that I brought up in the beginning was this idea that you could kind of brute force your way to a competitive model.

我听出来的答案是：不再是了，因为现在有了像混合专家和推理这样的新技术，你实际上需要对那个基础预训练做一定程度的精炼，才能构建出我们今天看到的顶级模型。而这正是 Meta 现在正在努力解决的问题。

I think the answer that I'm hearing from you is not anymore, because there are new techniques like mixture of experts and reasoning that you actually need some level of refinement of that base pre-training in order to be able to build the models that were the top tier models that we're seeing today. And that's what Meta is working through right now.

是的。还不止如此。顺便说一句，这是整个行业的情况。单体模型的时代大约在 Llama 3 发布时就终结了。那种“只有一个模型，我们只要测试它有多聪明，就能知道它在很多事情上有多好”的想法已经过时了。我们现在所处的世界是，当你使用这些工具链时，无论是 Open Code、Claude Code、Codex，使用这些工具链，它们会在底层根据任务去调用许多不同的模型。所以可能会去调用一个多模态模型。如果你用的是 Gemini，当它要生成图像时，会把任务分派给 Nano Banana。所以我们已经真正超越了那个“一个模型统治一切”的世界。你真正想要的是一个运行成本极高的智能模型，你可以用各种有趣的方式和场景把它蒸馏下来，只在必要时才使用它那极致的智能，因为运行那些模型非常昂贵。而在其他所有你并不需要天才级智力的地方，就用更便宜、更快、延迟更低的模型。因为如果你想想人类的任务，我真的很相信缩放定律，所以你会看到随着算力规模扩大，这种持续的增长曲线向右上方延伸，原始智能、模型本身会不断扩展。但人类任务并没有无限的智能需求。有很多人类任务你用常规水平的智能就能完成。所以我确实认为会出现分层，不再是“好吧，哪个模型能统治一切？”而是“好吧，把哪些模型集合在一起，能以性能、价格和价值的恰当平衡来解决这些问题？”

Yeah. It's not just that. By the way, this is the whole industry. The era of the monolithic model kind of died around the Llama 3 launch. The idea that there's one model and let's just test how smart this model is, and that was how good it's going to be at lots of things. We're now in a world where when you're using these harnesses, whether it's open code, Claude Code, Codex, using these harnesses, they're shopping underneath to lots of different models depending on the task. So that might be going to a multimodal model. If you're using Gemini, it'll farm tasks out to Nano Banana if it's trying to do image generation. So we've really moved past this world where there is just one model that rules everything. What you really want to have is a very expensive-to-run intelligent model that you can distill down in all these interesting ways and places, and use it for its exquisite intelligence only when necessary, because it's very expensive to run those models. And otherwise have models that are cheaper and faster and have lower latency in all of these other places where it turns out you don't need to have a genius-level intellect. Because if you think about human tasks, I really believe in scaling laws, so you're going to see this continued growth up and to the right as compute scales up, that the raw intelligence, the model scales up. But human tasks don't have infinite intelligence demands. There are a lot of human tasks that you can do with conventional levels of intelligence. And so I do think there's going to be stratification, where it's not just, okay, cool, what's the one model that rules them all? It's, cool, what is the collection of models that are brought together in such a way that they solve these problems with the right balance of performance and price and value?

是的，你说了几个有意思的点。首先，重要的是产品。我同意你的看法。而且为了自主可控，拥有自己的模型很重要。那我们就来聊聊这个。我相信你看到了苹果的做法，他们和谷歌达成协议，蒸馏 Gemini 或者做 Gemini 的某个分支，从早期报道来看，Siri 用那项技术表现得相当不错。那你有没有考虑过和谷歌做类似的交易，同时并行构建自己的模型来实现自主可控，但至少短期内能够尽可能快地推进你的产品？

Yeah, you said a couple interesting things. First of all, it's the product that matters. I would agree with you. And that it's important to have your own model for self-reliance. So let's talk about that. I'm sure you saw what Apple did, where they made a deal with Google to distill Gemini or do some fork of Gemini, and it looks like from the early reports Siri is working pretty well with that technology. So have you considered doing a similar deal with Google and then building your own in parallel for that self-reliance, but at least being able in the near term to advance your products as fast as you can?

嗯，这有两个部分。我们今天使用很多不同的模型，我认为你再次想要为消费者提供最适合他们的最佳模型。所以显然这里有价格和性能的考量，也有延迟的考量。但拥有自己的模型让你不仅能掌控自己的命运，在你想达成某些交易时，你还有强得多的谈判条件，确保你能为消费者拿到最好的答案。

Well, there's two parts. So we use lots of different models today, and I think again you want to provide consumers the best model that's going to work for them. And so there's obviously a price and a performance that matters here, and there's a latency that matters here. But having your own model gives you the ability to not just control your destiny, you also have much stronger negotiating terms when you're trying to figure out the types of deals that you want to make, to make sure that you're getting the consumers the best available answer.

花那么多钱。好像是给了谷歌十亿美元。

Spend that much money. It was like a billion dollars to Google.

现在说还为时过早。我还不知道体验会怎样。我还没有机会用上它。所以我们会看到的。另外，至少对我们来说，我们谈的是个人超级智能。我们希望能够发挥出巨大的特定能力，不只是通用智能，而是为我们打造的产品发挥出特定的能力。这对我们来说非常重要。我们不把这看作是对现有系统的增值。我们把它看作人们与计算机交互的全新方式。这确实可以追溯到我们在 Reality Labs 做了很久的很多工作。我们一直试图以施乐帕克研究中心、斯坦福研究所或贝尔实验室这样的先驱为榜样，我们试图思考的是，我们如何把信息从大脑传入机器，因此我们做神经接口，因此我们做所有这些事情。以及如何把信息从机器传回大脑，因此我们做增强现实和虚拟现实。AI 可能是我们见过的把信息从大脑传入机器的最佳工具，尤其是如果它能观察我们周围的许多事物。这些是我们试图发挥的独特能力。这不仅仅是模型，而是模型处理所有这些新颖输入并从中创建闭环系统的能力。

And it's too early to tell. I don't know what the experience is going to be yet. I don't have access to it. So we'll find out. I also, for us at least, we're talking about personal superintelligence. The ability we want to be able to have to bring a tremendous specific capability to bear, not just a general intelligence, but a specific capability to bear for the products that we build. That really matters to us a lot. We're not seeing this as a value add for an existing system. We're seeing this as an entirely new way that people are going to interact with their computers. It does go back to a lot of the work we've done in Reality Labs for a long time. We've always tried to model ourselves after pioneers like Xerox PARC or Stanford Research Institute or Bell Labs, where we're trying to think about what is the way that we get information from our brains into the machine, and that's hence our work on neural interfaces, hence our work on all these things. And what's a way to get the information from the machine back into our brains, hence our work on augmented reality and virtual reality. AI is potentially the best tool we've ever seen to get information from our brains into the machine, especially if it's able to observe a lot of things around us. Those are unique capabilities that I think we're trying to bring to bear. It's not just the model, it's what's the model's ability to work with all these novel inputs and create a closed-loop system out of it.

所以我认为我们正在努力打造出色的模型，我对我们组建的团队非常有信心。我的意思只是这还不够，至于苹果仅仅租用那个模型是否足够，我不知道他们是否有更宏大的愿景，关于它如何融入人们的生活。

So I think that we are working on having incredible models and I'm very confident in the team that we've assembled to do that. My point is just that it's not enough, and whether it's enough for Apple to just go rent that model, I don't know if they have a broader vision for how it integrates with people's lives.

好。所以你不会租用模型？

Okay. So you wouldn't rent the model?

不，我们确实租用模型。就像我说的，我们使用，我们是——

No, we do rent models. Like I said, we use, we're—

从哪里？

From where?

我们没有理由——我们，你知道，当我们在内部做开发时，我们确实有很多开发是在我们自己的模型上进行的。也有一些开发领域我们用的是从谷歌、Anthropic 或 OpenAI 那里使用的模型。能够做到模型无关并且在经济上合理，实际上有点取决于你拥有一个有竞争力的模型，对吧？

There's no reason for us—we, you know, when we're doing development internally, we do have a lot of development happening on our own models. There's also some areas of development that we do on models that we use from Google or from Anthropic or from OpenAI. The ability to be model agnostic and have that be economically sensible actually kind of hinges on you having a competitive model, right?

一个你可以在需要时退回去用的模型。而且它形成了一个真正的底线，限制别人能在此基础上向你收取多少租金。

That you can go back to if you need to. And it creates a real backstop on how much rent somebody can try to charge you on top of that.

但同样值得注意的是，无论是——我说的是公司内部的开发者，还是消费者——我不希望他们随着时间推移还要操心模型。今天他们不得不操心。今天一切都紧密地绑在一起，但随着时间的推移，他们只是有一个想要达成的目标，那才是他们应该关注的主要焦点。

But it's also worth noting, whether it's—I'm talking about a developer inside of the company or I'm talking about a consumer—I don't want them to worry about the model over time. Today they have to. Today it's all very tightly tied together, but over time they just have a goal they're trying to accomplish and that's the major focus that they should have.

所以有这样一个战略构想：拥有一个模型，并且让它成为绝对领先的最先进模型，这超级重要。但并不是说你有了它，就突然赢了。你还需要把很多环节连接到它上面，包括产品、分发和消费者体验。

So there's this strategic construct of having a model and having it be an absolute leading state-of-the-art model and that's super important. But it's not like when you have that suddenly you win. There's a bunch of pieces that you have to connect that to in product and in distribution and in the consumer experience.

我认为，正是这四样东西的集合，构成了我们相对于竞争对手的超能力，而大多数竞争对手——无论是 Apple、Anthropic、OpenAI 还是 Google——只拥有其中一样。

And I think it is the collection of all four of those things that we see as our superpower relative to the competitors, most of whom—whether it's Apple or Anthropic, OpenAI or Google—only have one of those things.

好。我待会儿会更深入地聊产品。但首先，上次我们聊天时，你说过你不会与 AI 融合，但你现在描述的方式是，你用技术把想法从你的大脑传到电脑，再从电脑传回你的大脑。听起来很像那么回事。你改变主意了吗？

Yeah. I'm going to get into product deeper in a moment. But first, last time we spoke you told me you wouldn't merge with AI, but the way you're talking about this is you use technology to get your thoughts from your mind to a computer and then from a computer back to your mind. Sounds a lot like that. Have you changed your mind?

没有。我不认为这是与 AI 融合。我仍然希望事物之间有非常清晰的界限。

No. I don't see this as merging with AI. I still want to have a very clear separation between things.

下次我们聊天时我还会再问你。

I'm going to ask you again next time we talk.

我知道。我们会继续聊下去。这是一个持续的趋势，实际上是趋势的延续和加速，即我们与机器之间、机器与我们之间的比特率随时间不断提高。我们已经做过一些有趣的版本。自动更正。自动更正就像一个小 AI，坐在你和电脑之间，帮助改进、减少损失，并有效提高你和机器之间的比特率。我们一直在使用所有这些小工具来加速这个循环。二维码，我最喜欢的一个，二维码。它就像是一种方式，说：酷，我想输入一个 URL，但我绝对不想手动输入 URL，因为错误率会太高，它不会带我到正确的网站，我还得找。所以，我们使用二维码。我认为，如果 AI，如果你有一个 AI，能够以非常人性化的方式，用人类语言理解事物，那将可能极大地提升我们利用已有算力的能力，即使只是在输入方面。现在，你把它与 AI 更有效地综合信息并反馈给我们的能力结合起来，你就真正极大地提高了比特率。这就是 Doug Engelbart，当他离开 NASA 去创办斯坦福研究所时的想法。他的想法是，人类问题变难的速度比人类能力提升的速度更快。他想创造这种人机共生。他说，唯一的方法就是让团队以某种方式与计算机融合，让它去做——这就是为什么他领导了第一次视频通话、第一次联合文档编辑、鼠标，所有这些都源于想要提高比特率。我认为 AI 正是这种东西。

I know. We'll keep it going. It's a continuous—it's really a continuation of a trend, an acceleration of a trend where the bit rate between us and machines and machines back to us goes up over time. And like there's funny versions of this that we've already been doing. Autocorrect. Autocorrect is like a little AI that sits between you and the computer that helps improve, reduce the loss and effectively improve the bit rate between you and the machine. And there's all these little tools that we use all the time to accelerate the loop. QR codes, one of my favorite ones, QR codes. It's like a way of being like, cool, I want to enter a URL, but I definitely don't want to type a URL because the error rate is going to be too high and it won't take me to the right website and I'll have to look. So, we use QR codes. I think if AI, if you have an AI that's really able to in very human terms, in human language terms, understand things, that is a potentially profound improvement of our ability to take advantage of the compute we already have, even if it was just on the input side. Now, you combine that with the AI's ability to synthesize information more effectively to get back to us, you've really tremendously improved the bit rate. This is that Doug Engelbart when he left NASA to start Stanford Research Institute. His idea was that human problems were getting harder at a steeper rate than human capability was improving. And he wanted to create this human-computer symbiosis. And he said that the only way he could do it is if teams of people could merge with computers in some way to make it do—and that's why he led, you know, the first ever video call, the first ever joint document editing, the mouse, like all these things came from wanting to increase the bit rate. I think AI is exactly that kind of thing.

好。所以它的表现形式可能是这个个人助理，对吧，它了解你的上下文。是的。出去为你办事。它可能通过手机或电脑上的聊天界面发生，或者通过 Meta 正在做的那种眼镜。所以从产品角度，我想你已经预览了一些，但想和你多聊一点。难道所有产品最终不会趋同吗？难道所有 AI 产品最终不会趋同于这个个人助理用例吗？如果你想想 OpenAI 是什么，我们刚请了 Greg Brockman 上节目。OpenAI 试图做的是创造这个超级应用，为你办事，理解你，真正帮助你，当你和它说话时，它会出去在世界上为你做事。Anthropic 也一样，Meta 类似。Apple 也有类似的愿景，尽管我们得等它实际推出后看看。那么你如何区分，你同意一切都趋同于这个中心助理用例吗？

Okay. And so the way that it manifests could be in this personal assistant, right, that knows your context. Yeah. Goes out and gets things done for you. It could happen via a chat interface on a phone or a computer or through glasses like the type that Meta is making. And so from a product standpoint, and I think you've already previewed a little bit of this, but would like to talk to you a little bit about it a little bit more. Don't all products end up converging? Don't all AI products end up converging on this personal assistant use case. If you think about what OpenAI is, we just had Greg Brockman on the show. And what OpenAI is trying to do is trying to create this, you know, super app that will get things done for you and understand you and really help you out, you know, as you talk to it, it will go out and do things in the world for you. Same thing with Anthropic, similar with Meta. And Apple has again a similar vision, although we'll wait to see what it looks like when it's in the wild. So how do you differentiate and do you agree that everything sort of converges on this central assistant use case?

是的。嗯，我认为每个人都在做令人兴奋的工作，我们处于最前沿。所以很难说。我会说，今天的工作，Anthropic 正在做的业务，以及 OpenAI 似乎越来越追求的，在 Greg 领导下集中精力的事情，是一个企业业务，他们在构建这些框架。那是赚钱的地方，我理解他们需要钱。所以这是一个重要的起点，实际上非常依附于企业。实际上，他们所有的收入都在那里。我明白，大公司，钱集中在一个地方，所以你只需要做较少的销售，就能获得较大的资本，而他们玩的是一场资本密集型的游戏。我认为他们的主要焦点肯定是在这些工作用例上，我认为这些非常有价值。显然，我们也在专业工作中利用它们。但这不是我们的主要焦点，我们的主要焦点 100% 在于这将如何帮助消费者生活。我认为真正的问题——我不确定 AI 会变得彼此无法区分。我认为有一个真正的问题——实际上你自己也说了，这些就像是个人助理，它们可以访问关于你的信息，你肯定不希望这些信息广泛传播。但它对那个个人助理是可用的。它是一个可信赖的助理。

Yeah. Well, I think everyone's doing exciting work and we're on the very forefront of it. So it's hard to say. I would, you know, the work today, the business that Anthropic is doing and that OpenAI appears to be increasingly pursuing, concentrating things under Greg, is an enterprise business where they're building these harnesses. That and that's where the money is and I understand that's—they need money. So it's an important place to start where it's like it's actually very much attached to the enterprise. That's like where all their revenue is as a practical matter. And I get that that's like that's—you know big companies there's a lot of money in one place so you have a smaller number of sales that you have to make and you can get like larger amounts of capital and this is a capital intensive game that they're playing. I think their major focus is definitely on these work use cases and I think those are super valuable. Obviously we take advantage of them as well in terms of our professional work. That's not our major focus like our major focus is 100% on how this is going to help consumers in their lives. And I think the real question—I don't know that the AIs become indistinguishable from one another at all. I think there's a real question of—actually you framed it yourself like you—these are kind of like a personal assistant and they have access to information about you that you certainly wouldn't want broadly distributed. It's available to that personal assistant. It's a trusted assistant.

嗯，如果你曾经有过个人助理并雇了一个新的，会有一个适应期。所以如果你有这个个人助理，它实际上已经深入你的生活并且做得很好。我认为这创造了一种真正的联系，需要其他竞争对手提供大量价值才能取代。

Well, if you've ever had a personal assistant and hired a new one, there's like a ramp up period that involves that. So like if you have this personal assistant that's actually quite embedded in your life and is doing well. I think that creates a real connection that you have that's requires a lot of value from some other competitor to go replace.

你认为为什么消费者 AI 起飞如此缓慢？我的意思是，已经有一些尝试。有像 Character AI、Replika 这样的。但你看 OpenAI，你说得对，他们肯定从赚钱的角度转向了。他们确实有一些消费者应用，比如营养、健康，对吧？这些是消费者可能涉足我们更广泛行业的东西。但是，这个想法，你会想象消费者 AI 对人们非常有吸引力，从娱乐角度、陪伴角度，以及帮助你完成生活中的事情，而你在商业角度不会那么做。但它一直很慢。

Why do you think consumer AI has been so slow to take off? I mean there have been some attempts. There have been like the character AIs the replicas. Um but you saw with open eye you're right they definitely pivoted from from a money standpoint. They do have some consumer applications that they want like nutrition, health, right? these are consumer thing that might tap into some of our, you know, some of our broader industries. But, um, this idea, you would imagine that like consumer AI would be very appealing to people, um, from an entertainment standpoint, a a companionship standpoint, and helping you, I guess, get done things in your life in a way that you wouldn't, you know, call on when you're doing it from a business standpoint. But it's been slow.

是的。嗯，我想，你知道，我不知道为什么我们认为这个会免疫，但炒作周期是一个永恒的概念，我们的行业不断陷入其中。并不是说人们经常误解炒作周期。

Yeah. Well, I think, you know, I don't know why we thought this one was going to be immune, but the hype cycle is an evergreen concept that our industry continues to fall for. And it's not that people often misunderstand the hype cycle.

他们考虑的是炒作周期——如果你不了解的话，就是先有一个炒作的顶峰，然后跌入失望的低谷，最后才达到最终的产品市场契合。炒作周期的重点不是说这项技术是假的，而是说愿意费尽周折去使用它的人只占人口中相对较小的一部分。

They think about the hype cycle — for those who don't know, there's a peak of hype, then the valley of discontent, and then the ultimate eventual product-market fit. The point of the hype cycle isn't that the technology is fake. It's just that people willing to go through a bunch of hoops to make it work are a relatively small percentage of the population.

而把它带给所有人这件事实际上是很艰苦的工作。这种艰苦工作不仅仅是你解决了一个困难的技术问题。你还得让用户界面变得可用，让它容易上手。人们要能理解它的价值，因为人们正在过自己的生活，没有这个工具他们也过得很好。你是在要求他们改变习惯，要求他们以一种相当剧烈的方式改变与电脑打交道的方式。

And the work of bringing it to everybody is actually hard work. And it's hard work that is not just a matter of, great, you've done this hard technology problem. It's also you've made the user interface workable. You've made it easy to use. People understand the value because people are living their lives. They're having great success living their lives without this tool. You're asking them to change their habits. You're asking them to change how they deal with computers in a pretty dramatic way.

他们大多不喜欢这样。

They mostly don't like it.

这不是——你必须以价值为先。我们在做什么？我们会为你做哪些具体的事情，让你的生活变得更好？

It's not — you have to lead with value. What are we doing? What are the specific things that we're going to do for you that are going to make your life better?

也许我最喜欢的例子就是智能体方面的工作。所以，就像我们行业里许多其他人一样，我在 12 月很早就开始用 Pi，然后是 Myclaw，使用、构建、把玩这些智能体框架。我发现它们非常强大，但它们并不太用户友好。它们很难构建、很难维护，而且会随时间漂移。所以当我想到——嘿，我给我妻子和我自己建了一个，我把它放在一个 WhatsApp 聊天里，她可以用。她从来不用。我一直在用。她不用。它很难融入工作流程。她只是让我去做事情。我就是那个智能体，然后我们就从那里开始。

Maybe my favorite example of this is the agentic work. So like many other people in our industry, I was very early on in December with Pi and then with Myclaw, using and building and playing with these agentic frameworks. And I find them very powerful, but they're not very user friendly. They're very hard to build, to maintain. They have drift over time. And so when I think about — hey, I built one for my wife and I, and I put it on a WhatsApp chat and she could use it. She never uses it. I use it all the time. She doesn't use it. It's just hard to integrate into a workflow. She just asks me to do things. I'm the agent, and then I like, you know, we go from there.

然后你委托出去。

And you delegate.

是的。然后我再去找智能体。所以这就是那个中转。其实挺合理的。对她来说运行得很好。我不怪她。如果我成功了，我其实会担心——如果我做出一个智能体，成功地把我从那个循环里解放出来。所以我并不那么渴望那个。所以我的意思是，我们还没有让这些东西变得容易使用。我认为我们在处理搜索用例和研究用例方面做得很好。我认为人们理解那些。我认为人们理解生成式 AI 用于内容，比如我想做一张好玩的图片。我认为有一些用例，人们现在理解了我们的能力，他们想去用那些，但我们还没有做工作，让它成为人们想融入日常生活的东西。它还不够容易使用。它没有创造足够的价值。它太繁琐了。所以这就是要解决的问题。这是要解决的产品问题。

Yeah. Then I go to the agent. So that's the pass-through. It's actually pretty reasonable. It's working well for her. I don't blame her. If I succeed, I'm actually worried if I make an agent that successfully gets me out of that loop. So I'm not that eager for that. So my point is like we have not made these things easy to use yet. I think we've done a great job of handling search use cases and research use cases. I think people understand those. I think people understand generative AI for content like I want to make this funny image. I think there's a few use cases that people understand our capabilities now and they want to go use those, but we have not done the work to make it something that people want to integrate into their daily life yet. It's not easy enough to use. It doesn't create enough value. It's too fussy. And so that is the problem to tackle. It's the product problem to tackle.

你需要出色的模型来做到这一点，但出色的模型还不够。

You need great models to do it, but great models are not enough.

对。

Right.

你对 AI 伴侣怎么看？因为说到人们会依赖的助手，有一种信念是：你构建功能，然后人们就会来用它。另一面是：你构建一个头像，一个让人们觉得像朋友的 AI 头像，这就是你差异化的方式。我的意思是，我们知道个性很重要。

Where do you stand on AI companions? Because when it comes to what will be an assistant that people rely on, there is this belief that you build the functionality and then people will come to it. The other side of it is you build an avatar, an AI avatar that people feel like they're friends with, and that is the way that you differentiate. I mean, we know personality matters a lot.

我要说的是，我们学到的一件事——我当然认为 Anthropic 在 Claude 的各个世代中也学到了——作为人类，我们非常在意自然语言是否吸引我们。所以个性对这些模型很重要。话虽如此，我认为你会发现人群中分布非常广。我认为有些人绝对希望这个 AI 是具身的，有个性，有张脸。事实上，有些人在智能体世界里，想创建 20 个不同的智能体，每个都有不同的个性，用于生活的不同部分——一个教练、一个营养师、一个医生助手，以及所有这些不同类型的东西。我不是那种人。我其实喜欢——不，我只想让我的 AI 极其可靠、值得信赖，我不介意它是一个无形的实体。它不需要有人类结构，我才会关心它。我当然不想应付 20 个。我只想应付一个，让它做所有我需要的事情。所以我认为我们——现在非常早。现在下结论还为时过早。我认为你会看到人们想要如何接触这项技术以及什么让他们感到舒适，会有很大的范围。因此，我预计市场会提供这些。

So I will say that one thing we've learned — and I certainly think Anthropic has learned over the various generations of Claude — we care a lot as humans about the way natural language appeals to us or doesn't appeal to us. And so personality matters for these models. Having said that, I think what you're going to find is a very big distribution among the population. I think some people absolutely would like this AI to be embodied and have a personality and have a face. In fact, there's been some people who, in the agentic world, they want to go create 20 different agents that each have a different personality for different parts of their lives — a trainer and a nutritionist and a doctor's assistant and all these different types of things. I'm not one of those people. I actually like — nope, I just want my AI to be extremely reliable and trustworthy, and I'm fine with it being an amorphous entity. It doesn't have to have a human structure for me to care about it. And I certainly don't want to deal with 20 of them. I just want to deal with one of them and have it do all the things I need. So I think that what we're — it's very early. It's too early to say for sure. I think you're going to see a big range of how people want to engage this technology and what makes them comfortable with it. And as a consequence, I would expect the market to deliver that.

你知道，有一种未来是这些 AI 伴侣成为——说得直白点——新的社交媒体，对吧？社交媒体是一个你去看看朋友们在做什么并与之互动的地方。它可以包罗万象，最好的情况下是令人满足的，而花费的时间是一个相当重要的指标，尽管你花时间之后的感受也很重要。

You know, there is a future where these AI companions become — this is a blunt way to put it — but the new social media, right? Social media is a place where you go to see what's going on with your friends and you engage with it. It can be all-encompassing and, in its best case, fulfilling, and time spent is a pretty important metric, although how you feel after you spend that time is also important.

时间花得值。

Time well spent.

时间花得值。也许那会被人们花时间与——我的意思是，最终就像你如何在电脑上与某物互动？也许那会被人们花时间与某个非常关心他们的 AI 实体相处所取代。

Time well spent. And maybe that gets replaced by people spending time with — I mean, ultimately it's like how do you engage with something on your computer? Maybe that gets replaced with people spending time with some AI entity that cares a lot about them.

是的。我的意思是，我尽量不评判人们选择的方式。

Yeah. I mean, I try not to judge the way people choose.

不，我同意。是的。

No, I agree. Yeah.

对于技术，我的直觉是，对绝大多数人来说， AI 的主要好处将是增加与他们在乎的人、他们爱的人进行人际接触的时间。你知道，我在增强现实的语境下经常谈到这一点。甚至只是我们有的相机眼镜——当我和孩子们在一起时，我既能记录一些东西并与我妻子分享，这对我们很有意义，同时又能完全在场，我和他们之间没有手机。这对我来说很重要。我谈过，如果你能更高效地工作，你就有更多时间不用通勤，更多时间不用离开家人，离开你爱的人。我个人的感觉是，对绝大多数人来说，真实人际连接的价值只会随时间上升，不会随时间下降。我认为我们从人们早期对 AI 的反应中已经看到了一点。我认为人们担心它是一种替代技术。我自己不这么认为，因为——我是它的重度用户。

With technology, my instinct is that for the overwhelming majority of people, the major benefit of AI is going to be increased time for human contact with people that they care about, people they love. And you know, I talked to this a lot in the context of augmented reality, for example. Even just the camera glasses that we have — when I'm with the kids, I'm able to both record something and share it with my wife, which is meaningful to us, and also be fully present, and I don't have a phone between me and them. And that's an important piece for me. I've talked about if you were able to be more effective with your work, that's more time that you're not spending commuting, that's more time that you're not spending away from your families, from the ones that you love. My personal sense is that for the overwhelming majority of people, the value of authentic human connection only goes up over time. It doesn't go down over time. And I think we're seeing that a little bit in how people's reactions to AI early on have been. I think people are worried that it's a replacement technology. I don't find it that way myself, having — I'm an avid user of it.

实际上，多亏了它，我大部分时间反而不需要待在电脑前，而不是相反。所以我认为这就是我的预测：绝大多数人会如何与它互动，以及它会如何影响他们与媒体和亲人的关系。我认为这会凸显真实连接和真实人类时刻的价值。但我确信整个分布都会存在。

Actually, mostly I'm spending more time not having to be at my computer thanks to it, not the opposite. So I think that's my prediction on how the overwhelming majority of people will interact with it and how it will affect their relationship to media and to their loved ones, which I think puts a premium on authentic connection and authentic human moments. But I'm sure the entire distribution will exist.

是的。当然，AI 眼镜是这一愿景的核心。

Yep. And of course, the AI glasses are kind of core to that vision.

没错。

Yeah, that's right.

所以，我们稍后就聊这个。

So, we'll talk about that right after this.

大家好，我是 Alex Kantrowitz。我想向大家介绍一部我与 Gravity 合作制作的纪录片，探讨 AI 智能体安全的未来。为了了解我们是否真正为自主智能体做好了准备，我采访了麻省理工学院教授 Ramsh Roskar、前白宫 CIO Theresa Payton、米其林集团首席数据与 AI 官 Ambika Roger Gopal，以及阿里巴巴前高管 Sharon Guy。他们各自对这一不断演变的领域提供了独特见解。最后，我们与 Gravity 首席执行官 Rory Blundell 讨论了前进之路。在 Gravity 的引领下，加入我们的旅程。你可以在节目笔记中的链接观看完整纪录片。

Hi everyone, Alex Kantrowitz here. I want to tell you about a documentary I've made with Gravity to explore the future of AI agent security. To find out if we're truly ready for autonomous agents, I sat down with MIT professor Ramsh Roskar, former White House CIO Theresa Payton, Michelin's group chief data and AI officer Ambika Roger Gopal, and Sharon Guy, a former executive at Alibaba. They each offer unique insights into this evolving landscape. We conclude with Rory Blundell, CEO of Gravity, to discuss the path forward. With Gravity leading the way, join us on this journey. You can watch the full documentary at the link in the show notes.

我们回到 Big Technology Podcast，嘉宾是 Andrew Bosworth，Boz，Meta 的首席技术官。Boz，很高兴再次见到你。感谢你抽出时间与我交谈。如果我们切到广角镜头，可以看到我们在纽约，此时你和你的团队正在发布三款 Meta 设计的眼镜。这是我们在节目中一直在争论的话题：手机是 AI 设备，还是可穿戴设备？我们又遇到了这个时刻，回到 Apple，他们似乎准备发布一个真正能用的 Apple Intelligence 版本，能在一定程度上了解你的上下文，并可能为你完成任务。然后我们看到反面是 Snapchat 眼镜的发布，让很多人说也许我们不需要——我的意思是那些太糟糕了，人们只是——你不必评论。我来说吧。

And we're back here on Big Technology Podcast with Andrew Bosworth, Boz, the CTO of Meta. Boz, great to see you again. Thank you for taking the time to speak with me. If we go to the wide shot, we can see we're here in New York at a moment where you and your team are releasing three new pairs of Meta designed glasses. It's something we've been debating on the show: is your phone the AI device or is it a wearable? And we've had this moment again going back to Apple where it looks like they're preparing to release a version of Apple Intelligence that actually works, that knows your context to a degree and might be able to get things done for you. And then we see the opposite side is the Snapchat specs release which got a lot of people saying maybe we don't—I mean those were so bad that people were just—you don't have to comment. I'll say it.

我不能评论。我还没见过。我自己还没见过。

I can't comment. I haven't seen them. I haven't seen them myself yet.

我们就说——我只是——我的评论反映了市场的反应。Evan Spiegel 戴着它们出席某个发布会。我想 Snap 股价立即下跌了约 6%。事情就是这样。

Let's just say—I'll just—my comment reflects what the market did. Evan Spiegel wore them out to some presentation. I think Snap stock went down like 6% immediately. It's just what happened.

好吧，这将是我戴着我们新眼镜的第一段视频。我们看看会发生什么。让市场来决定吧。

Well, this will be the first video of me wearing our new glasses. We'll see what happens. We'll let the market decide.

是的。但我想听听你的想法——显然 Meta 在这方面投入了很多。你相信这是一个引人注目的用例。如果我说也许我们不需要 AR 或 AI 眼镜，我们可以直接用手机。你会怎么说，让你觉得这个赌注的另一面？

Yeah. But I'd love to hear your thoughts on—obviously Meta has invested a lot in this. You believe it's a compelling use case. If I were to say maybe we don't need AR or AI glasses, we can just use our phone. What would you say that makes you feel the other side of that bet?

是的，手机很棒。我的意思是，我喜欢手机。我有两部。我认为它们是极好的设备。从一开始，眼镜——我们问自己的正是这个问题。我们说，‘好吧，手机很棒。有什么东西是你希望能访问的，它在你的手机上，但不需要把手机从口袋里拿出来？’我们想到了相机和音频。很简单。就像，‘酷，如果我能做到这一点。’AI 一直是一股巨大的顺风，实际上它随着时间的推移解锁了比手机通过蓝牙连接所能做的更广泛的能力。所以是的，现在它比两三年前看起来更有前景。两三年前，这看起来像是，嘿，在某个时候你必须在这上面加一个显示屏，它必须成为一个独立系统，并且必须附上所有这些配件。现在，实际上看起来市场上有足够的空间容纳各种各样的可穿戴设备。当然是眼镜，可能不只是眼镜，可能还有很多其他东西。人们不想戴眼镜，他们想戴不同的东西。其中一些设备只是你手机的输入和输出。那很酷。就像你的手机很棒，如果它只是让你的生活更高效，在输入和输出方面，那太棒了。有些会更完整。例如，我们的带显示屏眼镜，我们刚刚为它推出了一个 vibe coded 平台，所以任何想的人都可以去为眼镜构建任何你想要的应用程序。现在，你构建应用程序，然后把它放到眼镜上。但在未来，没有理由不能是你戴着眼镜，实时告诉眼镜你现在想要什么应用程序，并让它即时为你构建那个应用程序。

Yeah, phones are great. I mean I love phones. I have two of them. I think they're wonderful devices. The glasses from the very beginning—the question we asked ourselves was this exact question. And we said, 'Okay, phones are great. What is something that you wish you could get access to that's on your phone without having to take your phone out of your pocket?' And we came up with camera and audio. It's just very simple. It's like, 'Cool, if I could just do that.' The AI has been this tremendous tailwind where actually it unlocks a much larger swath of potential capability over time than what the phone can do just through, you know, Bluetooth connections. And so yeah, it's much more promising now than it looked two years ago or three years ago. Two or three years ago, this looked like, hey, at some point you have to put a display on this and it has to become a standalone system and it has to have all this, you know, kind of accessories attached to it. Now, it actually looks like there's totally enough room in the market for a big range of wearable devices. Glasses certainly, probably not just glasses, probably a lot of other things. People don't want to wear glasses, they want to wear different things. And some of those devices are just going to be input and output to your phone. That's cool. Like your phone's great and if it's just making your life more efficient in terms of how it's doing input and output, that's awesome. Some of them will be more complete. So for our band display glasses, for example, we just launched a vibe coded platform for it and so anybody who wants to can go literally just build whatever app you want for the glasses. Now right now you kind of build the app and you like put them on the glasses. But in the future, there's no reason that couldn't just be you wearing the glasses in real time, telling the glasses what app you want right now and having it on the fly build that app for you.

有趣。

Interesting.

你明白我的意思吗？所以我认为我们正走向一个非常酷的区域，它不太像特定应用花园。你仍然会有这些内容家园。内容继续是一个常青且重要的东西，就像它在电视上、在社交媒体上、在所有地方一样。所以仍然会有你想要接触的媒体所在的地方，那些看起来有点像应用程序或频道或随便什么，没有更好的词。但有一个长尾的东西，比如为什么我的烤面包机需要一个应用程序？让我认真问你这个问题。就像我的烤面包机有一个应用程序。

You know what I'm saying? And so I think we are headed towards a very cool zone where it's a little less like app garden specific. You're still going to have these content homes. Content continues to be an evergreen and important thing as it has been on TV, as it has been on social media, as it has been everywhere. So there's still going to be places where media that you want to reach lives and those look kind of like apps or like channels or whatever, lack of a better term. But there's a long tail of things like why does my toaster need an app? Let me ask you this in seriousness. Like my toaster has an app.

我不认为它需要。

I don't think it needs one.

我不想要那个，对吧？

I don't want that, right?

我只想告诉我的 AI 智能体，给我我想要的吐司。这是我每天吃的同样的吐司。就给我拿来。我不想去做那些事。你的烤面包机应用程序是——它让你——

I just want to tell my AI agent, get me the toast that I want. It's the same toast I have every day. Just get it for me. I don't want to have to go do whatever thing is. What does your toaster app—is it—does it let you—

老实说——我拒绝安装它。我拒绝安装它。

I honestly—I refuse to install it. I refuse to install it.

我尊重这一点。

I respect that.

我拒绝。我绝对不会这么做。所以——

I refuse. I absolutely won't do it. And so—

你总得坚持点什么。

You have to stand up for something.

是的。听着，有一条线——有一条线，没人——你知道——

Yeah. Listen, there's a line—there's a line that nobody—you know—

我认为你实际上可以——这是一个——我不得不承认有时候，你可以有一个特定的应用程序来控制事物的每一个方面，这太酷了，我尊重这一点，我是个技术人，对吧？所以我喜欢它的那种可摆弄的特性。但就像字面上，在这一点上它有点失控了，当我真的只想告诉一个智能系统，嘿，给我我想要的东西，它就能为我做到。

I think you can actually—it's a—I have to admit sometimes like it's so cool that you can have a specific app to control every aspect of the thing and I respect that and I'm a tech guy, right? So I like the fidgety nature of it. But it's like literally at this point it's kind of gotten out of hand when I really just wanted to tell an intelligent system, hey, get me the thing that I want and it can do that for me.

我们看到了一个早期形式，你知道，我们与 Spotify 的合作，你让眼镜播放音乐，它会——如果你链接了 Spotify 账户，它会去获取你想要的音乐，就像，是的，这很棒。这就是我想要的。我不想经历一堆步骤来做这个。

And we see an early form of, you know, our partnership with Spotify, you ask the glasses to play music, it goes—if you have a Spotify account linked, it goes and gets the music you want and it's like, yeah, this is great. This is what I wanted. I didn't want to have to go through a bunch of steps to do this.

至少对我来说，我的思考方式并不是手机很棒，它们会继续很棒。我不认为 App 那种模式会是未来的样子。我认为未来会是为你提供的有价值的服务，你以你想要的方式、在你需要的时候获得这些服务，并向提供这些有价值服务的人付钱，所有这一切都可以提前协商或按需协商。

For me at least, the way I'm thinking about this is not that phones are great and they're going to continue to be great. I don't think the app thing is the way the future's going to look. I think the future is going to be valuable services that are provided to you, and you getting access to those services the way that you want, when you need it, and paying money to the people who provide those valuable services, all negotiated either in advance or on demand.

是的，我真的很相信这一点。我看到你有——我今天用了 Meta AI 应用，看到眼镜有一个 Garmin 连接器。对我来说，当我在训练时，我很希望能说，我正在为这个半程马拉松做准备。Meta AI，在我所在区域、在这个时间窗口内帮我找一个 5 公里赛，并帮我报名。

Yeah, I really believe in this. I saw you had the—I was on the Meta AI app today and I saw there's a Garmin connector to the glasses. And for me, as I'm training, I'd love to be able to say, well, I'm building up to this half marathon. Meta AI, find me a 5K in my area in this window and sign me up.

完全没问题。

Totally.

而且是在我跑步的时候就能做到。

And to do that as I'm on a run.

是的。

Yeah.

这样我就不用花一个小时自己去弄明白。

So I don't need to spend an hour figuring it out on my own.

完全同意。再往更高层次说，你的 Meta AI 理想情况下会已经知道你在训练，你有一个想要达到的目标，并且它与所有重要的部分相连——你的营养，你的——这就是我们想让这个东西发展的方向。从现在到那时还有很多步骤，但那就是我们要去的地方。

Agree completely. And taking it to a higher level, your Meta AI ideally would already know that you're training and you have a goal that you're trying to reach, and it's tied into all the pieces that matter—your nutrition and your—it's like that's the direction we want to get this thing. There's a lot of steps between now and then, but that is where we're going.

Orion 眼镜，我们上次谈过。它们现在进展如何？那是完整的 AR 全眼镜体验。

The Orion glasses, we talked about those last time. Where do those stand? Those are the full AR full glasses experience.

是的。Orion 对我们来说是一个非常重要的时刻。我们拥有这个 AR 愿景已经很久了，它终于给了我们一个可以用来开始开发软件的设备。尽管我们无法把价格做到我们认为可以作为消费产品推出的水平，但当我们设计和开发它时，它就是消费级的设计和意图。所以产品本身相当可穿戴、相当可用。我家里有一副。我们用它来测试软件。所以我们继续在软件上迭代，我们在软件上取得了更多进展，不仅因为 AI 变得更好了——这对那个软件是什么产生了巨大影响——还因为你有 Orion 可以用来开发，这带来了很大不同。所以是的，我们继续非常专注于整个光谱。我们在这里暗示过，除了显示眼镜和摄像头眼镜，还有一整个价格区间可能低于那个的眼镜系列。嗯，也可能有。我仍然相信全 AR 是这个领域的未来。我认为我们会继续采取到目前为止的同样方法，以及我们没有推出 Orion 的同样原因。仅仅具备所有这些功能是不够的。它必须看起来很棒。必须足够舒适，让你想戴它。必须在一个合理的人会说“是的，这很值”的价格点上。

Yeah. So Orion was such an important moment for us. Having had this AR vision for such a long time, it finally gave us the device that we could use to start to play with the software on. And even though we couldn't get the price to be one that we felt comfortable launching as a consumer product, when we designed it and developed it, it was a consumer design and intention. And so the product itself is quite wearable, quite workable. I have a pair at home. We use it to test the software. So we've continued to iterate on the software and we've made so much more progress in the software, not just because AI has gotten better—that makes a huge difference to what that software is—but also because you have Orion to develop on, which makes a big difference. So yeah, we continue to be very focused on the entire spectrum. We've hinted here that in addition to display glasses and camera glasses, there's a whole range of glasses that may be below that in the price range. Well, there also may be there. I really still believe in full AR as a future for the space. I think we're going to continue to take the same approach we have so far, and the same reason we didn't launch Orion. It's not just enough that it does all this functionality. It has to look great. Has to be comfortable enough that you want to wear it. Has to be at a price point that a reasonable person would say, yeah, this is a good value.

那还有多远？

How far away is that?

我不会说一个确切的数字。我会说我喜欢我们正在取得的进展。

I'm not going to say an exact number. I will say I like the progress we're making.

以年还是月来衡量？

Measured in years or months?

我不会回答那个。

I'm not gonna answer that.

好吧，这很公平。

All right, that's fair.

我很欣赏你的努力。

I appreciate the hustle.

不得不问。

Have to ask.

我知道你不得不问。这是我有些保留的原因。在我们这样的公司待过的人都知道这一点。我们不断审视各种载体，问自己，是这个吗？准备好了吗？是这个吗？天哪，我们正在进入状态。这相当令人兴奋。

I know you do. It's some of my reticence. People who have been in companies like ours know this. We're constantly looking at vehicles and asking ourselves, is this the one? Is it ready yet? Is this the one? And man, we're getting into the zone. It's pretty exciting.

好的，酷。我们聊一下 Meta 的文化。你在运营这个应用 AI 部门。没错。这已经成为一些报道的主题。

Okay, cool. Let's talk about Meta culture for a moment. You're running this applied AI division. That's right. Which has been the subject of some reporting.

我运营智能体转型加速器。其中的一个团队是 AI 团队。

I run the agentic transformation accelerator. One of the groups in that is the AI team.

好的。我就读一下《连线》的引述。一名员工告诉《连线》：“这简直就是 goolog。你突然之间人生毫无目标。你几乎不与人互动。你每周就只有这些任务。”显然是在说那里的员工被安排了一些类似 AI 谜题的任务，他们必须尝试完成，以帮助训练 AI。那里发生了什么？

Okay. I'm just going to read the quote from Wired. One employee told Wired, "It's literally the goolog. You have zero purpose in life all of a sudden. You barely interact with anyone. You just have these tasks every week." Apparently talking about how the employees there have been put on some like AI puzzles that they have to try to accomplish that helps train the AI. What's going on there?

我不确定这个人有没有谷歌过 goolog 是什么样的，以及它与硅谷六位数软件工作有多相似或不相似。

I'm not sure this person's ever googled what a goolog was like and how similar or not it is to a six-figure software job in Silicon Valley.

看起来不像，但他们会这么说——

Doesn't seem like it, but the fact that they would say that—

撇开夸张不谈。好的。我们内部在这上面花了很多时间。这对我们来说是一个极其重要的话题。你报道我们很久了，所以你知道——这是一家会进入封锁状态的公司。当我们面前有一个紧迫的机会时，我们就会这样做。我们在移动端这样做过，在视频上这样做过，在 Stories 上这样做过。我们做过。并不是每一件事都会让整个公司转向，但有些时刻我们会想，等等，如果我们现在在某件事上投入极致的努力，我们认为市场上有巨大的机会。在这种情况下，我们看到了——我们真的觉得，当我们推出 Muse Spark 时——我想谨慎一点，Muse Spark 是一个很棒的模型，我们真的很兴奋——而且编程本来不是我们在模型上的重点，但它实际上在开箱即用的编程方面比我们预期的要好。

Setting aside the hyperbole. Okay. So we've been spending a lot of time on this internally. It's a hugely important topic for us. You've been covering us a long time so you know this—this is a company that goes into lockdowns. When we have an urgent opportunity ahead of us, we do this. We did it with mobile, we did it with video, we did it with stories. We've done it. And it's not that every one of these things pivots the entire company, but there are moments we're like, wait, if we put exquisite effort on something right now, we think there's a tremendous opportunity for us in the market. And in this case we saw that—we really feel like when we came out with Muse Spark—and I want to be careful, Muse Spark is a great model and we're really excited about it—and it's what coding had not been a focus for us on the model, but it actually was better out of the box at coding than we had expected it to be.

我们早期通过实验发现，只要给它相对少量经过训练、专家指导的示例，我们就可以对模型进行后训练，并大幅提高其竞争力。所以当你开始算数字、算这笔账时，你会想，哦，这对我们来说是一个不可思议的机会，可以构建一个编程模型，不仅让我们在运营公司的方式上拥有独立性，而且我们认为它在内部也很有价值——如果你给用户能够编程的 AI，这显然是过去一年在这些 AI 系统中变得非常普遍的强大工具之一。然后也让我们能够随着时间推移让模型本身更广泛地可用。所以我们基本上看到了这个巨大的机会，如此之大，以至于我们几乎在瞬间转向，把公司里很多人，数千人，调到这个 AI 组织来做这些专家轨迹。我们绝对需要他们的专业知识。如果做得不好，就行不通。事实证明，如果你用一段糟糕的代码来训练模型，你会对它造成一些损害。

And we found early on through experiments that actually giving it just a relatively modest number of trained, expertly guided examples, we could post-train the model and dramatically improve its competitiveness. And so when you start to run the numbers and the math on this, you're like, oh, this is an incredible opportunity for us to build a coding model that not only allows us to have independence in how we operate the company, but also something that we think is going to be valuable both inside—if you give users AI that's able to code, that's obviously one of the very powerful tools that's become very common in these AI systems over the last year. And then also for us to be able to make the model itself more widely available over time. So we basically saw this huge opportunity, such a big opportunity that we pivoted kind of on a dime and brought a lot of people across the company, thousands of people, out into this AI organization to do these expert traces. We absolutely need their expertise. It doesn't work if you do a bad job. It turns out if you use a bad piece of coding to train the model, you do some damage to it.

你不想强化失败。

You don't want to reinforce failure.

它们必须做得很好。

They have to be well done.

他们必须得到专业的引导。我们做得非常快，结果就是它没有太多结构，也没有很好的沟通。我曾公开说过……其实不对，我没有公开说过。我是被泄露了，说它糟透了。

They have to be expertly guided. Now, we did it very quickly, and as a consequence, it did not have a lot of structure. It did not have great communication around it. I've been on record... Actually, that's not true. I wasn't on record. I was leaked calling it atrocious.

你说过，也许不是这 20 年来最糟糕的，但也差不多了。绝对排得上号。

You said maybe not the worst it's ever been in 20 years here, but it's up there. It's definitely up there.

那其实不是我说的。我不知道那是从哪儿……

That actually was not a quote from me. And I don't know where that...

你没说过那句话。

You didn't say that.

我没说过。

I didn't say that.

好吧。好吧。

Okay. Okay.

但我说过类似的话。我不介意。但问题是，作为一家大公司，我们看到了这个紧迫的机会，并做出了我认为战略上绝对正确的改变，但我们没有做足功课，没有走到每个人面前说：让我跟你谈谈这是什么、为什么我们需要它、为什么它重要。

But I've said things like it. I'm fine with it. But so, the degree to which it's a big company, the degree to which we saw this urgent opportunity and made the change that I think strategically was absolutely the right change, but did not do the work to kind of go to each person and be like, let me talk to you about what this is and why we need it and why it's important.

是的。

Yeah.

明知道他们本来有自己兴奋的其他工作，却要暂停那些工作来做这件事。但这就是我们公司在看到那些稍纵即逝的、难以置信的机会时会做的事。所以，是的，我们正在应对这个行业里正在发生的变化，它也在每家公司内部发生。这就像我们从未见过的事情。你一开始就说：这就像我们职业生涯中从未见过的事情。我认为这让人们停下来思考，所以这对我和其他领导者提出了更高的要求，要比我们过去做得更好，去沟通正在发生什么、为什么会发生、它如何影响你、我们如何看待它的长期发展。确保他们明白，他们扮演的角色是我们认为非常关键、非常重要的，否则我们显然不会做出那个改变。

Knowing that they had other work that they were excited about that they were putting on pause to come do this work. But this is something our company does when we feel like we see these unbelievable opportunities that exist in moments of time. And so yeah, we are navigating this change that's happening in the industry, is happening inside every company as well. And it's like nothing we've ever seen. You said you led out with this: it's like nothing we've ever seen before in our careers. And I think that is giving people pause, and so it raises the bar on me and other leaders to do a much better job than we have done communicating what's going on, why is it happening, how does it affect you, how do we see it playing out long term. Make sure they understand that the role they're playing is one that we consider very critical, very important, otherwise we wouldn't have made that change, obviously.

我们能简单谈谈追踪这件事吗？其实，如果我是员工，我觉得我不会喜欢它，但我最近在我们的节目上实际上为你可能这么做的理由做了辩护。现在我们就坐在一起，那就聊聊吧，因为基本上报道说 Meta 已经开始追踪一些按键和员工的打字方式，并基本上用它来训练模型。我的看法是，随着模型训练进入强化学习——我记得 Scale AI 的 Alexandr Wang 说过，他们现在大部分训练都是强化学习，而不是我们之前谈过的预训练。随着技术进入强化学习，让这些模型学习如何在通常称为 gym 或不同区域中完成任务非常有价值，就像真实世界活动的模拟，它们进去尝试完成。所以，我是不是可以认为这个项目基本上就是那种做法的大规模扩展版，模型观察员工完成任务，然后学习自己完成任务？

Can we talk about the tracking briefly? I actually, you know, if I was an employee, I don't think I'd be a fan of it, but I actually sort of made the case for why you might be doing it on our show recently. And now that we're sitting next to each other, let's talk about it because basically the reports have been that Meta has started to track some keystrokes and the way that employees type and basically use that as a way to train models. And my perspective on this was as model training moves into reinforcement learning, where I think Scale AI, where Alexandr Wang came from, said most of their training is reinforcement learning now, as opposed to pre-training which we talked about previously. As the technology moves into reinforcement learning, it's very valuable for these models to learn how to accomplish tasks in what's typically called gyms or like different areas that different like simulations of real world activity that they go in and try to accomplish. And so am I right in thinking that this program is basically just a massively scaled up version of that where the models watch employees work through their tasks and then learn how to accomplish tasks on their own?

是的。嗯，这有两个部分。第一部分是你完全正确。强化学习在今天的人工智能中扮演的角色比人们两三年前可能预测的要大得多。但不仅仅是这个。还有人类知识和行为的长尾非常长。而且大部分，就像互联网上的全部文本语料一样，我们知道的大部分东西仍然不在互联网上。它在我们的头脑里。它是经验。它是随时间积累的。它是我们习以为常的行为。所以这个系统在某些方面我认为相当天才。员工不需要改变他们日常做事的方式，可以像往常一样工作，同时产生这个独特的数据语料。在这种情况下，设计和人类如何使用计算机？人工智能实际上仍然非常奇怪地不擅长使用计算机。就像一个令人惊讶的难题，还没有很好地解决。这就是所有精力都投入计算机使用和智能体式的地方，那都是计算机。

Yeah. Well, there's two parts to this. The first one is you're absolutely right. Reinforcement learning is playing a much bigger role in today's AI than people had maybe predicted two or three years ago that it would. It's not just that though. There's also the long tail of human knowledge and behavior is very long. And most of it, as much as for all the text, for the entire corpus of text on the internet, most of the stuff that we know is still not on the internet. It's like in our heads. It's experience. It's built up over time. It's behaviors that are second nature to us. And so this system was in some ways I thought quite genius. You've got employees who need to change nothing about how they go about their day, can go about it as they always have, and in doing so produce this corpus of unique data. In this case, design and how do humans use computers? AIs are actually still really weirdly bad at just using computers. Like it's a surprisingly hard problem that is not well solved. And that's where all the energy is going with computer use and agentic, that's all computer.

你当然可以在前端提升智能，然后尝试从中蒸馏。

And you can ramp up the intelligence in the front end for sure and then try to distill down from that.

但我们确实认为拥有这些数据有可能让人们的生活更轻松。这甚至不是关于内容。沟通上的挑战，再次说我们做得不好，是这甚至不是关于你所做的事情的内容。而是关于计算机如何能够理解这个数字界面内正在发生的事情，这是我们今天访问许多工具的方式。第二件事是，所以我认为这个数据集很有趣，但我们还不知道。它是一个长期运行的数据集。所以第二部分是，对于长尾专家训练，你最好做像我们应用 AI 团队所做的工作，那个 AI 团队，那是相对少量、记录非常完善的任务，可以训练、后训练模型。这是不同的东西。这就像非常长期运行。一旦我们有了一年的数据，你就有了一些可能有趣的东西可以带入模型。我还想补充，自推出以来我们也对这个项目做了一些改变。我们增加了……

But we do think having this data has the potential of making people's lives easier. It's not even about the content. The thing that was a challenge to communicate, and again we did a poor job, was it's not even about the content of the thing that you're doing. It's about how is the computer able to understand what's happening inside this digital interface, which is the way we access a lot of our tools in the world today. The second thing is, and so I think this data set is interesting, but we won't know. It's a long-running data set. So the second part of this is you're still for long-tail expert training, you're better off doing work like we are doing with our applied AI team, the AI team, like that is a relatively small number of really well-documented tasks that can train, post-train a model. This is a different thing. This is like very long-running. Once we have like a year of data, you have something that's potentially interesting to bring to bear in the model. I do want to add we've also made a bunch of changes to the program since the launch. We've added...

休息 30 分钟。

Take a 30-minute break.

无限暂停，人们可以因为各种原因选择退出。所以，我们为有顾虑的人对这个项目做了很多改变。

Unlimited pausing, people can opt out for a bunch of reasons. So like we've made a bunch of changes to the program for people who had concerns about it.

所以你把很多旧博客文章发到 Substack 上，我一直在邮件里收到并阅读它们。最近我读了一篇非常有趣的，讲的是你在做生物学研究时，医生说疼痛就是康复。你需要那种疼痛才能愈合。你写道，在某个时刻，你必须能够拥抱疼痛才能取得真正的进步。给定两个其他方面相同的故事，人类会记住唤起更强烈情感的那个。情感是我们大脑分类记忆的方式。有时它必须痛，你的大脑才会优先处理它。

So you are posting a lot of your old blog posts to Substack and I've been getting them in my email and reading them and there was a very interesting one that I read recently talking about how you were doing some biology research and the doctor said the pain is rehab. You need that pain in order to be able to heal. You write as at some point you have to be able to embrace the pain to make real progress. Given two otherwise equal stories, humans remember the story that evoked stronger emotion. Emotion is how our brain triages memories. Sometimes it has to hurt for your brain to prioritize it.

向 BS80 致敬，那是我在哈佛的神经生物学课。

Shout out to BS80, a class, my neurobio class at Harvard.

AI 是进化生物学。AI 正在带走很多痛苦，对吧？就像人类现在用 AI 做的大部分事情，就是把我们工作中很多痛苦的部分交给 AI。

AI is evolutionary bio. AI is taking away a lot of the pain, right? Like big part of what humanity is doing with AI right now is a lot of the painful parts of our work we're giving it to AI.

如果那个目标实现了，痛苦又在哪里呢？

If that goal is accomplished, where do we find the pain?

这个我很喜欢。顺便说一句，我做的一件事是让我的智能体把我的博客文章搬到 Substack 上，这样将来我可以两边都发。直到最近我才意识到，它不会处理项目符号列表，直接就把格式剥掉了。我的智能体根本不懂项目符号列表。所以，我们在智能体方面还有很长的路要走，这是第一阶段。痛苦就是康复过程。对，对。我们当时在研究一个问题，就是药物戒断期间发生的神经生物学变化。一个学生问，嘿，我们有这么多症状，为什么不干脆给人们用止痛药呢？教授说，你不明白，痛苦本身就是药。比如，体验对毒品的渴望、寻药行为，然后让它变得极其痛苦，这就是你重新训练大脑克服寻药行为的方式。如果你消除了痛苦，那个人就永远改不掉。所以，这是一种有生产力的痛苦。顺便说一句，我认为 AI，所有这些发生在 Meta 以及每家公司里的‘过氧化物’（指混乱、错误），就是我在说的痛苦。那种痛苦是，除了穿过它别无出路，你必须找到穿过的路径，弄清楚什么行得通、什么行不通，而且就是很艰难。我们社会中确实有很多其他类型的痛苦，与创造真正价值无关。这在教育中经常出现，是个好例子。我记得在学校时有人告诉我——我相信你也遇到过——嘿，考试不能用计算器。你在现实世界中过日子时不会随身带计算器。我身上随时至少带着三个计算器。更不用说，我直接问我的眼镜数学题就行。我计算器多得很。结果发现，不用计算器做数学题、做数学考试，是一种痛苦，但不是特别有用的那种。用计算器做更难的、需要批判性思维的数学考试，可能才是更有价值的做法，对吧？所以，我确实认为，重要的是让我们经历的痛苦与我们在世界上试图创造的价值对齐。我认为，就像学习整合 AI，你可以逃避那种痛苦。直接跳过，不做。你我都知道，这会让你面临真正的风险。你会落后于那些能够使用 AI 并且想做同样工作的人。你会在经济上或在你提供的产品上落后于其他已经整合 AI 的公司。你知道，谢丽尔·桑德伯格有句名言，她说公司通常不是设定艰难目标然后失败，而是设定容易目标然后一路达成。所以，我觉得你今天很容易就能逃避痛苦，只要说，对，我们就是不做，我们就让它发生，以后再想办法。所以，我认为有生产性的痛苦和非生产性的痛苦，也许需要一点判断力来区分哪个是哪个。

So, I love this. And, very small aside, one of the things I did is I assigned my agent the task of bringing my blog posts over to Substack so at some point I could do both. I didn't realize until very recently that it wasn't any bulleted list. It would just strip out. My agent did not understand bulleted lists. So we have a long ways to go on agents, is my okay phase one. The pain is the rehab. Yeah, yeah. There was a question we were studying, the neurobiology that would occur during withdrawal from drug use, and a student asked, hey, we have all these symptoms, why don't we just give people the pain a pain medicine? And the professor was like, you don't understand, the pain is the medicine. Like experiencing the desire to pursue drugs, drug-seeking behavior, and then having it be immensely painful is the way you reprogram your brain to overcome the drug-seeking behavior. And if you get rid of the pain, then the person is never going to do it. There is, so this is a productive form of pain. By the way, I would argue AI, all these peroxisms happening not just at Meta but at every company, is the pain I'm talking about. That is the pain that there is no way out but through, and you have to figure out the path through it to figure out what works and what doesn't work, and it's just gritty. We do have lots of other types of pain in our society that have nothing to do with real value being created. This comes up a lot in education, is a good example. I remember being told, I'm sure you were when I was in school, hey, you can't use a calculator on this test. You will not have a calculator with you as you go about your day in the real world. I have at least three calculators on my person at all times. Not to mention, I can just ask my glasses math problems. I'm filthy with calculators. It turns out doing a math problem, doing a math test without a calculator is a certain kind of pain, not a particularly useful kind. Doing a harder math test that requires critical thinking with a calculator is probably the more valuable way to do that thing, right? So, I do think it's important to align the pain that we're experiencing with the value we're trying to create in the world. I think like learning to integrate AI, you could avoid that pain. You just skip it. You don't do it. You and I both know that it puts you at real risk. You're gonna fall behind people who are able to do AI and want to do the same job as you. You're going to fall behind other companies that have integrated AI either economically or in the products that you offer. You know, there's this Cheryl always had this, Cheryl Samber has this great quote, which is that companies don't usually fail by setting tough goals and missing them. They fail by setting easy goals and hitting them all the way down. And so like I think you could easily avoid the pain today by just being like, yeah, we're just not going to do it. We're just going to let it happen and then we'll figure it out later on. So I think there is productive pain and unproductive pain and maybe a little bit of judgment to know which one's which.

Bos，和你交谈总是很愉快。非常感谢你来参加。

Bos, it's really always a pleasure to speak with you. Thanks so much for coming on.

谢谢邀请我。

Thanks for having me.

好了，各位。非常感谢收听和观看，我们下次在 Big Technology Podcast 再见。

All right, everybody. Thanks so much for listening and watching and we'll see you next time on Big Technology Podcast.