# Meta CTO Andrew Bosworth: Our Path To Frontier AI, Renting Models, Consumer AI's Struggles

https://www.youtube.com/watch?v=qZhCeV6XATw
Translation: zh-CN

[00:00] 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.
  Meta首席技术官安德鲁·博斯沃思加入我们，谈论公司的AI努力以及为什么它要制造自己新的AI眼镜。

[00:07] That's coming up right after this.
  这将在稍后播出。

[00:10] Welcome to Big Technology Podcast, a show for Coolheaded and Nuance conversation of the tech world and beyond.
  欢迎收听《大科技播客》，这是一个关于科技世界及以外的冷静和细致对话的节目。

[00:15] We have a great show for you today.
  我们今天为您准备了一个精彩的节目。

[00:16] 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.
  今天我们邀请到了Meta首席技术官安德鲁·博斯沃思，他将与我们谈论公司的AI努力、其新的AI眼镜、公司的文化以及最后的一些重要想法。

[00:28] Bos, great to see you.
  博斯，很高兴见到你。

[00:29] Welcome back to the show.
  欢迎回到节目。

[00:30] Well, thanks for having me.
  嗯，谢谢你的邀请。

[00:32] Um, we were just talking before we started rolling about what a crazy moment it is in the tech world.
  嗯，我们在开始录制之前，刚才还在谈论科技界现在是多么疯狂的时刻。

[00:37] We haven't seen progress like this as far as I can remember.
  据我记忆所及，我们从未见过如此进步。

[00:43] Um, the core part of it is the AI model.
  嗯，它的核心部分是AI模型。

[00:45] The AI model underpins everything.
  AI模型支撑着一切。

[00:48] Without an a working AI model or a leading AI model, it's tough to build.
  没有一个有效的AI模型或领先的AI模型，就很难构建。

[00:54] Um, the theory for a long time was that to build a great AI model, you needed a ton of compute and great researchers to
  嗯，长期以来的理论是，要构建一个伟大的AI模型，你需要大量的计算能力和优秀的研究人员来

[01:01] Work on the algorithm.
  致力于算法的研究。

[01:04] Meta has a ton of compute and a team of the best researchers to work on the algorithm,
  Meta拥有大量的计算能力和一支顶尖的研究团队来致力于算法的研究，

[01:09] but the leading AI model hasn't materialized yet.
  但领先的人工智能模型尚未出现。

[01:11] 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?
  那么，您能否谈谈您在那里学到的东西，以及关于构建伟大的人工智能模型所需的那些核心假设是否是错误的？

[01:21] Well, I the only other ingredient I would add is great data.
  嗯，我唯一会补充的另一个要素是优质的数据。

[01:23] Um, and you have that and we do have that I think as well.
  嗯，你们拥有这个，而且我认为我们也拥有这个。

[01:24] So, yeah, there's two two two stories here.
  所以，是的，这里有两个两个两个故事。

[01:26] The first one is, you know, I think, you know, we go back to Llama 1, Llama 2, Llama 3.
  第一个是，你知道，我想，你知道，我们回顾一下 Llama 1、Llama 2、Llama 3。

[01:29] We really were, you know, kind of at the forefront and advancing things and and you of course know this.
  我们确实，你知道，处于最前沿并推动事物发展，而你当然知道这一点。

[01:35] We the Facebook fundamental AI research group goes back a decade more.
  我们 Facebook 的基础人工智能研究团队可以追溯到十多年前。

[01:39] I've been I mean that's where I actually first got queued into what was going on with AI is when M the AI messaging Bob 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
  我一直在，我的意思是，那是我第一次真正了解到人工智能方面正在发生什么的时候，是当 AI 消息 Bob 出现在我的信息流中，然后我遇到了 Yan 并开始认识 FAIR 的人们，然后我说哦，这项技术正在快速发展，所以 Meta 很早就开始关注它了，所以我认为一直很公开的真正差距是，我们当时并没有真正提出的是，当

[02:02] we were pulling Llama 4 together sorry
  我们正在一起拉动 Llama 4，抱歉

[02:04] when we were pulling Llama 3 together we had really pulled in all the research
  当我们一起拉动 Llama 3 时，我们确实纳入了所有的研究

[02:09] all the every we pull out every single stop we had and unwittingly kind of killed the pipeline.
  所有的，我们拉出了每一个我们拥有的停止点，不知不觉地就搞垮了流水线。

[02:14] So researchers, you know, the way that it works is you build a a base and you've got people pioneering an incremental version of the base and you got people out there pathfinding entirely new strategies.
  所以研究人员，你知道，它的工作方式是，你建立一个基础，然后有人在开创基础的增量版本，你还有人在那里探索全新的策略。

[02:29] Um, and kind of unbeknownst to us at the time and kind of speaks to the fact that we weren't focused enough on it, um, Llama 3, which was a great model and was wellreceived
  嗯，而且在当时我们是不知道的，这在一定程度上说明了我们没有足够关注它，嗯，Llama 3，这是一个很棒的模型，并且受到了好评

[02:37] to get to that model, they had kind of pulled forward all the future bets into that to deliver that model.
  为了达到那个模型，他们已经把所有的未来赌注都提前投入到那个模型中以交付那个模型。

[02:42] Well, that meant when it came time for Lava 4, we didn't have any of the pathfinding the other labs still had going.
  嗯，这意味着当 Llama 4 来临时，我们没有任何其他实验室仍在进行的探索。

[02:46] So, that makes you now now you're behind on reasoning.
  所以，这让你现在落后于推理。

[02:50] Uh, now you're behind on a mixture of experts.
  呃，现在你在专家混合方面落后了。

[02:51] Now you're behind on a bunch of these critical technologies that have been used to continue continue the pace of progress.
  现在你在许多这些关键技术方面落后了，这些技术一直被用来继续保持进步的步伐。

[02:58] Um this is a pretty public you know disappointment I think a year ago for us and led to Mark
  嗯，我认为一年前这对我们来说是一个相当公开的，你知道的，失望，并导致了马克

[03:04] um shifting from okay AI isn't one of our bets which is how we thought of it up to that point.
  嗯，转变一下，好的，人工智能不是我们的赌注之一，这就是我们直到那时对它的看法。

[03:09] AI was just one of the many bets we had.
  人工智能只是我们拥有的众多赌注之一。

[03:13] AI is a bet that's foundational to the entire company and so we're going to change how we're thinking about this.
  人工智能是一项对整个公司都至关重要的赌注，因此我们将改变我们对它的思考方式。

[03:16] And um this is such a cliche but I don't have a better word for it.
  而且，嗯，这太陈词滥调了，但我找不到更好的词来形容它。

[03:20] ghost founder mode like he really did flip into a mode that is like unique and reserved for Mark um 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 um 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 um 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 um to focus very very well received model um 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 um and so yeah you're absolutely right on where we are in
  幽灵创始人模式，就像他真的进入了一种模式，这种模式是独一无二的，并且是为马克保留的，嗯，在那里他变得如此专注于为我们获取所有我们需要的计算能力，为我们获取所有我们需要的才能，我们签下的研究人员，你说你知道，他们大约在一年前真正到位，嗯，我想，我想亚历山大·王刚刚度过他的一周年元宇宙纪念日，我喜欢和他一起工作，并且我已经从他那里学到了很多，嗯，我们正在看到它的成果，所以如果你看看 Muse Spark，这是你最新的模型，虽然不是我们的前沿模型，但它是我们发布的最新模型，嗯，专注于一个非常非常受欢迎的模型，嗯，根据基准测试，它在我们最关心的、我们认为对我们的产品独一无二的事情上表现得非常好，嗯，所以是的，你对我们所处的位置绝对是正确的。

[04:04] 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.
  就公众的认知而言，就模型而言，我们已经组建了团队，我真的相信，我们拥有所需的所有计算能力和数据，所以我非常有信心我们将达到我们所需的位置。

[04:15] I'll add a second piece to this which I think is strategically very important though which is that you So models are available.
  我将在此基础上添加第二点，我认为这在战略上非常重要，那就是，所以模型是可用的。

[04:21] Like you can go rent a model.
  比如你可以去租用一个模型。

[04:22] You can go use anthropics.
  你可以使用 Anthropic 的。

[04:23] You can use open eyes.
  你可以使用 OpenAI 的。

[04:25] You can use Google's.
  你可以使用谷歌的。

[04:25] They're great models.
  它们是很棒的模型。

[04:27] You can go get them.
  你可以去获取它们。

[04:27] You can use them.
  你可以使用它们。

[04:27] Um and that's pretty great.
  嗯，这很棒。

[04:28] The real value we're going to create in the world is is the product.
  我们将在世界上创造的真正价值是产品。

[04:34] Um 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.
  嗯，我们对个人超级智能的产品，我们所拥有的愿景，我认为是我们独特适合交付的愿景。

[04:41] It's not just that we have data.
  不仅仅是我们拥有数据。

[04:41] That's cool.
  这很酷。

[04:43] 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.
  我们实际上比几乎任何其他人都有更好的机会了解你，了解你试图做什么，你在世界上是谁，以及对你来说什么最重要。

[04:52] Um, 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.
  嗯，所以拥有模型是一方面，你希望在战略上拥有它，这样你就不会依赖别人，但你主要希望能够控制你的命运。

[04:59] The model itself isn't the value.
  模型本身不是价值。

[05:01] And I think we're going to get to a world very soon where
  而且我认为我们很快就会进入一个世界，在那里

[05:04] consumers, they don't care.
  消费者不在乎。

[05:06] They don't want to specify the model they're using.
  他们不想指定他们正在使用的模型。

[05:07] They don't want to they don't care
  他们不想，他们不在乎

[05:09] if it's 4.7 or 4.8 eight 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
  它是否是4.7或4.8，就像你不在乎Oracle，我是否在使用Oracle或SQL数据库一样，你只想要功能，你想要事情能正常工作

[05:18] that's the standard to which I think we're all going to be held um so today the discussion is about models
  我认为我们将被要求达到这个标准，所以今天的讨论是关于模型

[05:21] which suggests to me at least that we're a little underindexed um on the user side of it and how humans are going to benefit
  这至少对我来说表明，我们在用户方面有点不足，以及人类将如何受益

[05:29] so I think that's the story that we need to tell in addition to showing the work that we've done technically
  所以我认为这就是我们需要讲述的故事，除了展示我们技术上所做的工作之外

[05:33] we need to actually demonstrate the value to consumers
  我们需要实际展示对消费者的价值

[05:37] so I I just want to talk about this um the science scientifically for a moment
  所以我想就科学方面稍微谈谈

[05:41] Um the 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.
  嗯，我在开头提到的就是你可以通过蛮力的方式获得一个有竞争力的模型。

[05:47] I think the answer that I'm hearing from you is not anymore
  我认为我从你那里听到的答案是，不再是这样了

[05:51] because there are new techniques like mixture of experts and reasoning that you actually can you need some level of refinement of that base pre-train in order to be able to build the models that were the top tier models that we're seeing today.
  因为有新的技术，比如混合专家和推理，你实际上需要对基础预训练进行一定程度的精炼，才能构建出我们今天看到的顶级模型。

[06:02] And that's what Meta is
  这就是Meta正在做的

[06:06] working through right now.
  现在正在处理。

[06:07] Yeah.
  是的。

[06:07] It's not just that.
  不仅仅是这样。

[06:08] It's by the way this is the whole industry.
  顺便说一句，这是整个行业。

[06:10] Um, the era of the monolithic model kind of died around Llama 3 launch.
  嗯，单体模型的时代在大约 Llama 3 发布时就终结了。

[06:13] Like the the idea like there's one model and just like let's just test how smart this model is and that was how good it's going to be at lots of things.
  就像，就像有一个模型，然后我们只是测试这个模型有多聪明，以及它在很多事情上会有多好。

[06:20] We're now in a world where um when you're using these harnesses whether it's um you know open code, cloud code, codeex using these harnesses, they're shopping underneath to lots of different models depending on the task.
  我们现在身处一个这样的世界，嗯，当你使用这些框架时，无论是嗯，你知道的，开源代码、云代码、codeex 使用这些框架，它们会根据任务在后台选择许多不同的模型。

[06:34] So that might be going to a multimodal model.
  所以这可能是走向一个多模态模型。

[06:35] um you know if you're using Gemini it'll it'll farm tasks out to Nano Banana if it's trying to do image generation.
  嗯，你知道的，如果你使用 Gemini，它会把任务分配给 Nano Banana，如果它正在尝试进行图像生成。

[06:41] So um we've we really moved past this world where this is just one model that rules everything.
  所以，嗯，我们已经真正走出了这个只有一种模型统治一切的世界。

[06:46] Um what you really want to have is a very uh expensive to run intelligent model that you can distill down in all these interesting ways and places um and use it for its exquisite intelligence only when necessary because it's very expensive to run those models.
  嗯，你真正想要的是一个运行成本非常高昂的智能模型，你可以用所有这些有趣的方式和地方将其提炼出来，嗯，并且只在必要时利用其卓越的智能，因为运行这些模型非常昂贵。

[07:01] um and otherwise have models that are cheaper and faster and have lower latency in all
  嗯，否则就拥有更便宜、更快、延迟更低的模型，在所有

[07:07] 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 um I really believe in scaling laws so you're going to see this continued growth up and to the right of as compute scales up um that the the raw intelligence the model scales up but like human tasks don't have infinite intelligence demands like there's a lot of human tasks that like you can do with conventional levels of intelligence um and so I do think there's going to be stratification then 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 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 model your own model for self-reliance so let's talk about that um 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
  在这些其他地方，事实证明你不需要拥有天才级别的智力，因为如果你考虑人类任务，嗯，我确实相信规模法则，所以你会看到这种持续的增长，随着计算能力的提升，原始智能，模型会扩展，但人类任务没有无限的智能需求，有很多你可以用常规智能水平完成的人类任务，所以我认为将会出现分层，不再是“好吧，酷，哪个模型能统治一切”，而是“酷，哪些模型被整合在一起，以正确平衡性能、价格和价值的方式解决这些问题”。是的，你说了一些有趣的事情，首先是产品很重要，我同意你，并且拥有自己的模型对于自力更生很重要，所以我们来谈谈这个，嗯，我敢肯定你看到了苹果公司与谷歌达成的协议，用来提炼 Gemini 或做 Gemini 的某个分支，从早期报道来看，Siri 在这项技术上表现得相当不错。那么你呢

[08:09] 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.
  考虑与谷歌达成类似的交易，然后并行构建自己的交易以实现自给自足，但至少在短期内能够尽快推进您的产品。

[08:20] Well, there's two parts.
  嗯，这有两个部分。

[08:21] So we 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.
  所以我们今天使用许多不同的模型，我认为再次您想为消费者提供最适合他们的模型。

[08:27] And so there's obvious there's a per there's a price and a performance that makes that matters here and there's a latency that matters here.
  所以显而易见，这里有一个价格和性能很重要，还有一个延迟也很重要。

[08:31] Um, but like 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.
  嗯，但拥有自己的模型不仅能让你掌控自己的命运，还能让你在试图确定你想达成的交易类型时拥有更强的谈判筹码，以确保你为消费者提供最佳的可用答案。

[08:46] Spend that much money.
  花那么多钱。

[08:47] It was like a billion dollars to to Google and it's it's too early to tell.
  这就像给谷歌花了十亿美元，现在说还为时过早。

[08:50] I don't know what the experience is going to be yet.
  我还不知道体验会是什么样的。

[08:51] I don't have access to it.
  我无法访问它。

[08:53] So, we'll find out.
  所以，我们会知道的。

[08:53] Um, I also I for for us at least, we're talking about personal super intelligence.
  嗯，至少对我来说，我们谈论的是个人超级智能。

[08:57] 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 pro products that we build.
  我们想要拥有的能力是能够发挥巨大的特定能力，不仅仅是通用智能，而是为我们构建的产品发挥特定的能力。

[09:07] That really matters to us a lot.
  这对我们来说非常重要。

[09:09] Um, we're not seeing this as um
  嗯，我们不认为这是...

[09:13] like a value ad for an existing system.
  就像是对现有系统的一种价值宣传。

[09:16] We're seeing this as an entirely new way that people are going to interact with their computers.
  我们正将此视为人们与计算机交互的全新方式。

[09:22] It does go back to a lot of the work we've done in reality labs for a long time.
  这确实回溯到我们在现实实验室长期以来所做的大量工作。

[09:25] Um, you know, we've always tried to model ourselves after pioneers like Xerox Park or um, Sanford Research Institute or Bell Labs where we're 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.
  嗯，你知道，我们一直试图效仿施乐帕克或嗯，斯坦福研究所或贝尔实验室等先驱，在那里我们试图思考如何将信息从大脑传输到机器，因此我们进行了神经接口方面的工作，因此我们进行了所有这些方面的工作，以及如何将信息从机器传回我们的大脑，因此我们进行了增强现实和虚拟现实方面的工作。

[09:47] Um, 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.
  嗯，人工智能可能是我们见过的将信息从大脑传输到机器的最佳工具，特别是如果它能够观察到我们周围的许多事物。

[09:58] Um, those are unique capabilities that I think we're trying to bring to bear that don't have any, it's not just the model, it's like what's the model's ability to work with all these novel inputs and create a closed loop system out of it.
  嗯，我认为我们正试图发挥这些独特的能力，这些能力并非仅仅是模型本身，而是模型处理所有这些新颖输入并从中创建一个闭环系统的能力。

[10:10] Um, so I think that we are working on
  嗯，所以我想我们正在努力

[10:13] having incredible models and I'm very confident in the team that we've assembled to do that.
  拥有令人难以置信的模型，而且我对我们组建的团队非常有信心能够做到这一点。

[10:18] My point is just that it's not enough and whether it's enough for Apple to just go rent that model.
  我的观点是这还不够，而且对于苹果来说，仅仅去租用那个模型是否足够。

[10:22] I don't know if they have a broader vision for how it integrates with people's lives.
  我不知道他们是否有更宏大的愿景来将其融入人们的生活。

[10:26] Okay.
  好的。

[10:26] So, you wouldn't rent the model?
  所以，你不会租用模型吗？

[10:28] No, we do rent models.
  不，我们会租用模型。

[10:28] Like I said, we we use we we you know, we're from where?
  就像我说的，我们使用我们我们你知道，我们来自哪里？

[10:31] There's no reason for us.
  我们没有理由。

[10:31] We we uh you know, we when we're doing development internally, we do have a lot of uh development happening on our own models.
  我们我们呃你知道，当我们进行内部开发时，我们确实有很多呃在自己的模型上进行开发。

[10:39] There's also some areas of development that we do on models that we use from uh from Google or from Anthropic or from OpenAI.
  还有一些开发领域，我们是在使用来自呃来自谷歌或来自Anthropic或来自OpenAI的模型上进行的。

[10:47] the ability to be model agnostic and have that be economically sensible actually kind of hinges on you having a competitive model, right?
  能够不拘泥于模型并且在经济上是合理的，实际上取决于你拥有一个有竞争力的模型，对吧？

[10:55] That you can go back to if you need to.
  如果你需要，你可以回到它。

[10:56] And it creates a real backs stop on like how much rent somebody can try to charge you on top of that.
  它创造了一个真正的后盾，就像别人可以试图在你之上收取多少租金一样。

[11:01] But it's also worth noting um 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.
  但同样值得注意的是，无论我是在谈论公司内部的开发者，还是在谈论消费者，我都不希望他们随着时间的推移而担心模型。

[11:10] Today they have to.
  今天他们不得不。

[11:10] today it's like it's all
  今天就像是它全部

[11:14] very tight tied together but over time
  非常紧密地联系在一起，但随着时间的推移

[11:16] they just have a goal they're trying to accomplish and that's the major focus that they should have.
  他们只有一个目标，他们正努力去实现，而这应该是他们关注的重点。

[11:20] So there's this like strategic construct of having a model and having it be uh an absolute leading state-of-the-art model and that's super important.
  所以存在一种战略构建，即拥有一个模型，并且它是一个绝对领先的、最先进的模型，这非常重要。

[11:30] Um but it's not like when you have that suddenly you win.
  但并非当你拥有它时，你就突然赢了。

[11:33] Um there's a bunch of pieces that you have to connect that to in product and in distribution and in the consumer experience.
  你需要将它与产品、分销和消费者体验中的一堆东西联系起来。

[11:41] 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 Enthropic OpenAI or Google only have one of those things.
  我认为正是这四件事的集合，我们将其视为相对于竞争对手的超级能力，而他们中的大多数，无论是苹果、Anthropic、OpenAI还是谷歌，都只有其中之一。

[11:51] Yeah. I'm going to get into product deeper into product in in a moment.
  是的。我马上会更深入地探讨产品。

[11:55] But first um 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.
  但首先，上次我们谈话时，你告诉我你不会与人工智能合并，但你谈论的方式是，你使用技术将你的想法从你的大脑传送到电脑，然后再从电脑传回你的大脑。

[12:05] Sounds a lot like that.
  听起来很像。

[12:07] Have you changed your mind?
  你改变主意了吗？

[12:08] No. I I don't see this as merging with AI.
  不。我不认为这是与人工智能合并。

[12:10] [laughter] I still want to have a very clear separation between things.
  [笑声] 我仍然希望在事物之间保持清晰的界限。

[12:13] I'm going to ask you again next time we
  下次我们见面时我会再问你

[12:14] talk.
  谈话。

[12:14] I know.
  我知道。

[12:14] We'll keep it keep it going.
  我们将继续下去。

[12:15] It's a continuous uh 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.
  这是一个持续的，呃，它确实是一个趋势的延续，一个趋势的加速，在这个趋势中，我们与机器之间以及机器与我们之间的比特率随着时间的推移而增加。

[12:28] Um and like there's funny versions of this that we've already been doing.
  嗯，而且我们已经做过一些有趣的类似的事情。

[12:29] Autocorrect.
  自动更正。

[12:31] Autocorrect is like a little AI that sits between you and the computer that like helps improve reduce the loss and effectively improve the bit rate between you and the machine.
  自动更正就像一个小的AI，它介于你和计算机之间，可以帮助改进、减少损失，并有效地提高你和机器之间的比特率。

[12:40] Um, and there's all these like little tools that we use all the time to accelerate the loop.
  嗯，而且我们一直使用所有这些小工具来加速循环。

[12:49] QR codes, one of my favorite ones, QR codes.
  二维码，我最喜欢的之一，二维码。

[12:50] It's like a way of like being like, cool, I want to like 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.
  这就像一种方式，就像，酷，我想输入一个URL，但我绝对不想输入URL，因为错误率会太高，它不会带我到正确的网站，我将不得不查找。

[12:59] So, we use QR codes.
  所以，我们使用二维码。

[12:59] Um, 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
  嗯，我想如果AI，如果你有一个AI，它真的能够用非常人性化的术语，用人类语言的术语来理解事物，那将是对我们利用计算能力的潜在的深刻改进。

[13:15] we already have, even if it was just on the input side.
  我们已经有了，即使它只在输入端。

[13:20] 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.
  现在，你将它与人工智能更有效地综合信息的能力结合起来，反馈给我们，你已经极大地提高了比特率。

[13:27] This is that Doug Doug Engelbart when he left um NASA to start Sanford Research Institute.
  这是道格·恩格尔巴特，当他离开美国国家航空航天局去创办斯坦福研究所时。

[13:33] His idea was that human problems were getting harder at a steeper rate than human capability was improving.
  他的想法是，人类面临的问题正在以比人类能力提高更快的速度变得更加困难。

[13:38] And he wanted to create this human computer symbiosis.
  他想创造这种人机共生。

[13:39] 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 the mouse like all these things came from wanting to increase the bit rate.
  他说，他能做到这一点，只有当人们的团队以某种方式与计算机融合，才能做到这一点，这就是为什么他领导了，你知道的，有史以来第一次视频通话，第一次联合文档编辑，鼠标，所有这些都来自于想要提高比特率。

[13:55] I think AI is exactly that kind of thing.
  我认为人工智能正是这样的东西。

[13:57] Okay.
  好的。

[13:59] And so the way that it manifests could be in this personal assistant, right, that knows your context.
  因此，它的表现形式可以是这个个人助理，对吧，它了解你的背景。

[14:03] Yeah.
  是的。

[14:03] Goes out and gets things done for you.
  它会为你外出办事。

[14:06] It could happen via a chat interface on a phone or a computer or through glasses like the type that Meta is making.
  它可以通过手机或电脑上的聊天界面，或者通过像Meta正在制造的那种眼镜来实现。

[14:13] And so from a product standpoint, and I think you've already previewed a little bit of this,
  所以从产品角度来看，我认为你已经对这一点进行了少量预览，

[14:15] but would like to talk to you a little bit about it a little bit more.
  但想和你多聊一点。

[14:19] Um, don't all products uh end up converging?
  嗯，所有的产品最终不都会趋同吗？

[14:22] Don't all AI products end up converging on this personal assistant use case.
  所有的人工智能产品最终不都会趋同于这个个人助理用例吗？

[14:26] If you think about what um what OpenAI is, we just had Greg Brockman on the show.
  如果你想想 OpenAI 是什么，我们刚请了 Greg Brockman 上节目。

[14:30] 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.
  OpenAI 试图做的是创建一个超级应用程序，为你完成事情，理解你，并真正地帮助你，你知道，当你和它交谈时，它会为你到外面去办事。

[14:40] Same thing with Anthropic, similar with Meta.
  Anthropic 也是一样，Meta 也是类似。

[14:42] And Apple has again a similar vision, although we'll wait to see what what it looks like when it's in the wild.
  苹果也有类似的愿景，尽管我们要等到它真正发布时才能看到它是什么样子。

[14:49] So how do you differentiate and do you agree that everything sort of converges on this central assistant use case?
  那么你如何区分呢？你是否同意一切都趋同于这个中心化的助理用例？

[14:55] Yeah. Well, I think I think everyone's doing exciting work and we're on the very forefront of it.
  是的。嗯，我认为每个人都在做令人兴奋的工作，我们正处于最前沿。

[14:58] So it's hard to say.
  所以很难说。

[14:59] I would, you know, the work today, the the business that Anthropic is doing that and that OpenAI appears to be increasingly pursuing uh concentrating things under under Greg is an enterprise business where they're building these harnesses.
  我想，你知道，今天的工作，Anthropic 正在做的业务，以及 OpenAI 似乎越来越多地追求的，嗯，在 Greg 手下集中精力，是一家企业业务，他们正在构建这些系统。

[15:11] Um that and that's where the money is and I understand that's they
  嗯，那就是钱所在的地方，我理解他们

[15:16] Need money.
  需要钱。

[15:17] So it's it's an important place to start where it's like it's actually very much attached to the enterprise.
  所以它是一个重要的起点，它就像，它实际上与企业非常紧密相连。

[15:20] Um that's like where all their revenue is um as a practical matter.
  嗯，这就像是他们所有的收入来源，从实际角度来看。

[15:24] Um, and I get that that's like that's 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 of game that they're playing.
  嗯，我明白，你知道大公司在一个地方有很多钱，所以你需要完成的销售数量较少，你可以获得大量的资本，而这是一个资本密集型的游戏，他们正在玩。

[15:36] Um, I think their major focus is definitely on these like work use cases and I think those are super valuable.
  嗯，我认为他们的主要重点绝对是这些工作用例，我认为它们非常有价值。

[15:43] Obviously we take advantage of them as well in terms of our professional work.
  显然，我们在专业工作中也利用了它们。

[15:46] That's not our major focus like our major focus is 100% on how this is going to help consumers in their lives.
  那不是我们的主要焦点，我们的主要焦点是 100% 关注这如何能帮助消费者改善他们的生活。

[15:51] Um, and I think the real question I don't know that the AIs become indistinguishable from one another at all.
  嗯，我认为真正的问题是，我不知道人工智能是否会变得完全无法区分。

[15:57] I think there's a real question of uh actually you you framed it yourself like you this these are kind of like a personal assistant and they have access to information about you that you certainly wouldn't want broadly distributed.
  我认为有一个真正的问题是，实际上你自己也这样描述了，这些就像是个人助理，它们可以访问关于你的信息，而你肯定不希望这些信息被广泛传播。

[16:08] It's available to that personal assistant.
  这些信息对那个个人助理是可用的。

[16:09] It's a trusted assistant.
  这是一个值得信赖的助理。

[16:11] Well, if you've ever had a personal assistant and hired a new one, there's like a ramp up period um that is that
  嗯，如果你曾经有过一个私人助理并雇佣了一个新的，那么有一个适应期，嗯，那就是

[16:17] that involves that.
  这涉及到它。

[16:19] So like if you have this personal assistant that's actually quite embedded in your life and is doing well.
  所以，如果你有一个非常融入你的生活并且做得很好的个人助理。

[16:24] I think that creates a real connection that you have that's requires a lot of value from some other competitor to go replace.
  我认为这会建立一种真正的联系，你需要从其他竞争对手那里获得很多价值才能取代它。

[16:28] Why do you think consumer AI has been so slow to take off?
  你为什么认为消费者人工智能的普及如此缓慢？

[16:33] I mean there have been some attempts.
  我的意思是，已经有一些尝试了。

[16:35] There have been like the character AIs the replicas.
  有过像角色人工智能、复制品这样的东西。

[16:40] Um but you saw with open eye you're right they definitely pivoted from from a money standpoint.
  嗯，但你看到了，你说得对，他们肯定从金钱的角度进行了转变。

[16:43] They do have some consumer applications that they want like nutrition, health, right?
  他们确实有一些他们想要的消费者应用，比如营养、健康，对吧？

[16:47] these are consumer thing that might tap into some of our, you know, some of our broader industries.
  这些是消费者可能涉及我们一些，你知道的，我们更广泛的行业的东西。

[16:53] 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.
  但是，嗯，这个想法，你会想象消费者人工智能会非常吸引人，嗯，从娱乐的角度，从陪伴的角度，以及帮助你，我猜，在你生活中完成事情，以一种你不会，你知道的，从商业角度去做的方式。

[17:11] But it's been slow.
  但它一直很慢。

[17:13] Yeah.
  是的。

[17:13] Well, I think, you know, I don't know why we thought this one was going to be immune, but the hype cycle is is
  嗯，我想，你知道，我不知道为什么我们认为这个会免疫，但炒作周期是是

[17:19] an evergreen concept that our industry continues to fall for.
  一个我们行业持续着迷的永恒概念。

[17:24] Um, and it's not that people often misunderstand the hype cycle.
  嗯，并不是人们常常误解了炒作周期。

[17:26] They think how there's this the hype cycle for those who don't know, you know, there's a there's a peak um of hype, then there's the valley of discontent, and then there's the ultimate eventual product to market fit.
  他们认为，对于那些不知道的人来说，存在着炒作周期，你知道，有一个炒作的顶峰，然后是失望的低谷，最后是最终的产品与市场契合。

[17:35] And the point of the hype cycle isn't that the technology is it's fake.
  炒作周期的重点并不是技术是虚假的。

[17:38] 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.
  只是愿意付出巨大努力使其奏效的人只占人口的一小部分。

[17:45] And the work of bringing it to everybody is actually hard work.
  而将其推广给所有人的工作实际上是艰苦的。

[17:50] Um, and it's hard work that is not just a matter of great, you've done this hard technology problem.
  嗯，而且这是一项艰苦的工作，不仅仅是解决了这个技术难题。

[17:54] It's also you've made the user interface, you know, workable.
  它还包括你让用户界面，你知道，变得可用。

[17:58] You've made it easy to use.
  你让它易于使用。

[18:00] Um, people understand the value because people are living their lives.
  嗯，人们理解其价值，因为人们正在过着自己的生活。

[18:03] They're having great success living their lives without this tool.
  他们可以在没有这个工具的情况下过着成功的生活。

[18:05] You're asking them to change their habits.
  你要求他们改变习惯。

[18:08] You're asking them to change how they deal with um computers kind of in a pretty dramatic way.
  你要求他们以一种相当戏剧性的方式改变与计算机打交道的方式。

[18:12] They mostly don't like it.
  他们大多不喜欢它。

[18:13] It's not going to it's not the you have to lead with value.
  它不会，它不是你必须以价值为导向。

[18:16] What are we do? How what are the specific things that we're
  我们做什么？我们具体做些什么？

[18:20] going to do for you that are going to make your life better?
  将为您做什么，让您的生活更美好？

[18:22] Um maybe my favorite example of this is the agentic work, you know.
  嗯，也许我最喜欢的例子是代理工作，你知道的。

[18:24] So like many other people in our industry, I was very early on in December with Pi and then with Myclaw, you know, using building playing with these agentic frameworks and I find them very powerful, but they're not very user friendly.
  所以，就像我们行业里的许多其他人一样，我在十二月很早就开始使用 Pi，然后是 Myclaw，你知道的，使用构建，玩弄这些代理框架，我觉得它们非常强大，但它们并不非常用户友好。

[18:40] They're very hard to build to maintain.
  它们很难构建和维护。

[18:41] They have drift over time.
  它们会随着时间漂移。

[18:45] Um, and so when I think about, hey, um, I built one for my my wife and I, uh, and I put it like on a WhatsApp chat and she could use it.
  嗯，所以当我想到，嘿，嗯，我为我和我妻子建了一个，呃，我把它放在 WhatsApp 聊天上，她可以用。

[18:52] She never uses it.
  她从不使用它。

[18:54] I use it all the time.
  我一直使用它。

[18:54] She doesn't use it.
  她不使用它。

[18:56] Um, it's just it's hard to like integrate into a workflow.
  嗯，它只是很难融入工作流程。

[18:57] She just asks me to do things.
  她只是让我做事。

[18:58] I'm the agent and [laughter] then I like, you know, we go from there and you delegate.
  我是代理人，[笑声] 然后我喜欢，你知道的，我们从那里开始，你委托。

[19:01] Yeah.
  是的。

[19:01] Then I and then I go to the agent.
  然后我，然后我去代理人那里。

[19:02] So it's like that's that's the pass through.
  所以就像那是那个通道。

[19:03] It's actually not it's pretty reasonable.
  实际上并非如此，它相当合理。

[19:05] It's working well for her.
  对她来说效果很好。

[19:06] I don't blame her.
  我不怪她。

[19:08] Um, if I succeed, I'm actually worried if I make an agent that successfully gets me out of that loop.
  嗯，如果我成功了，我实际上担心我会创建一个代理，成功地让我摆脱那个循环。

[19:11] So I'm I'm not that eager for that.
  所以我对此并不那么热衷。

[19:13] Um, so my point is like we have not made these things easy to use yet.
  嗯，所以我的观点是，我们还没有让这些东西易于使用。

[19:15] I think we've done a great job of like handling search use cases and
  我认为我们在处理搜索用例方面做得很好，而且

[19:22] Research use cases.
  研究用例。

[19:23] I think people understand those.
  我认为人们理解那些。

[19:25] I think people understand generative AI for content like I want to make this funny image.
  我认为人们理解用于内容的生成式AI，就像我想制作这张有趣的图片一样。

[19:29] I think there's a few use cases that people understand our capabilities now and they want to go use those.
  我认为现在有一些用例，人们了解我们的能力，并且他们想去使用它们。

[19:33] But we have not done the work to make it something that people want to integrate into their daily life yet.
  但我们还没有做足够的工作，使其成为人们想要融入日常生活的东西。

[19:38] Um it's not easy enough to use.
  嗯，它不够容易使用。

[19:39] It doesn't create enough value.
  它创造的价值不够。

[19:42] It's too fussy.
  它太挑剔了。

[19:44] Um and so that is the problem to tackle.
  嗯，所以这就是要解决的问题。

[19:46] It's the product problem to tackle.
  这是要解决的产品问题。

[19:48] You need great models to do it but great models are not enough.
  你需要强大的模型来做到这一点，但强大的模型是不够的。

[19:50] Right.
  对。

[19:52] Where do you stand on AI companions?
  你对AI伴侣有什么看法？

[19:57] Because you know when it comes to what will be a assistant that people rely on, there is this belief that you build the functionality and then people will come to it.
  因为你知道，当涉及到人们会依赖的助手时，有一种信念是，你构建了功能，然后人们就会来使用它。

[20:01] The other side of it is you build a um you build a avatar, an AI avatar that people feel like they're friends with and that is the way that you differentiate.
  另一方面是，你构建一个嗯，你构建一个虚拟形象，一个AI虚拟形象，让人们感觉像朋友一样，这就是你区分的方式。

[20:12] I mean we know personality matters a lot.
  我的意思是，我们知道个性很重要。

[20:16] So I will say that one thing we've learned and I certainly you know I think Anthropic has learned over the various generations of Claude.
  所以我想说，我们学到的一件事，我当然你知道，我认为Anthropic在Claude的各个代际中都学到了。

[20:19] We certainly it
  我们当然它

[20:22] we care a lot as humans about the way

[20:25] natural language appeals to us or

[20:27] doesn't appeal to us. And so personality

[20:30] matters for these models. Having said

[20:32] that, um I think what you're going to

[20:34] find is a very big distribution among

[20:37] the population. I think some people

[20:39] absolutely would like this um this AI to

[20:43] be embodied and have you know a

[20:45] personality and have a a face. In fact,

[20:48] there's been some people who when they

[20:49] in the agentic world, they want to go

[20:51] create 20 different agents that each

[20:53] have a different personality for

[20:54] different parts of their lives, a

[20:55] trainer and a nutritionist and a

[20:57] doctor's assistant and all these

[20:59] different types of things. Um, I'm not

[21:01] one of those people. I actually like,

[21:03] nope, I just want my AI to be like

[21:04] extremely reliable and trustworthy and

[21:06] like I'm fine with it being uh an

[21:09] amorphous entity. It doesn't have to

[21:10] have a human structure for me to care

[21:12] about it. And I certainly don't want to

[21:13] deal with 20 of them. I just want to

[21:14] deal with one of them and have it do all

[21:16] the things I need. So I think that what

[21:18] we're it's very early. It's too early to

[21:20] say for sure. I think you're going to

[21:21] see a big range of how people want to

[21:24] engage this technology and how what

[21:25] makes them comfortable with it. And as a

[21:26] consequence, I would expect the market

[21:27] to deliver that.

[21:28] You know, there's there is a future

[21:30] where these AI companions become this is

[21:34] a blunt way to put it, but the new

[21:35] social media, right? Social media is a

[21:38] place where you go to see what's going

[21:39] on with your friends and you engage with

[21:41] it. It's like it's very can be, you

[21:44] know, all-encompassing and and um and in

[21:48] its best case um fulfilling and you know

[21:52] and and time spent is like a pretty

[21:54] important metric although how you feel

[21:55] after you spend that time is also

[21:57] important.

[21:57] Time well spent.

[21:58] Time well spent. And maybe you know that

[22:01] gets replaced by people spending time

[22:03] with I mean ultimately it's like how do

[22:05] you engage with something on your

[22:06] computer? Maybe that gets replaced with

[22:07] people spending time with some AI entity

[22:12] that cares a lot about them.

[22:14] Yeah. I mean, I try not to judge the way

[22:15] people choose. No, I agree. Yeah.

[22:17] With technology, my my instinct is that

[22:20] for the overwhelming majority of people,

[22:23] um, the major benefit of AI is going to

[22:25] be increased time for human contact with

[22:27] people that they care about, people they

[22:28] love.

[22:29] Um, and you know, I talked to this a lot

[22:32] in the context of of augmented reality,

[22:34] for example. you know, even just the the

[22:36] camera glasses that we have, you know,

[22:38] when I'm with the kids, I'm able to both

[22:41] record something and share it with my

[22:43] wife, which is meaningful to us, and

[22:44] also be fully present, and I don't have

[22:46] a phone between me and them. And that's

[22:48] an important piece for me. Um, I've

[22:50] talked about with if you were able to be

[22:52] more effective with your work, that's

[22:54] more time that you're not spending

[22:56] commuting, that's more time that you're

[22:57] not spending away from your families,

[22:58] from the ones that you love. My personal

[23:01] sense is that the overwhelming pe

[23:02] majority of people um the value of

[23:06] authentic human connection uh only goes

[23:08] up over time. It doesn't go down over

[23:10] time. Um and I think we're seeing that a

[23:12] little bit in how people's reactions to

[23:14] AI early on have been. I think people

[23:16] are worried that it's a replace of

[23:17] technology. I don't find it that way

[23:19] myself having I'm an avid user of it. Um

[23:21] and actually mostly I'm spending more

[23:23] time not having to be at my computer

[23:25] thanks to it, not the opposite. So, I

[23:27] think that's how the I think that's

[23:28] that's my prediction on how the

[23:31] overwhelming majority of people will

[23:32] interact with it and how it will affect

[23:33] their relationship to media and to their

[23:35] loved ones, which I think it's a a

[23:37] premium on authentic connection and and

[23:39] authentic human moments. Um, but I'm

[23:42] sure the entire distribution will exist.

[23:44] Yep. And of course, the AI glasses are

[23:46] kind of core to that vision.

[23:47] Yeah, that's right. So, we'll talk about

[23:49] that right after this.

[23:50] Hi everyone, Alex Canitz here. I want to

[23:52] tell you about a documentary I've made

[23:54] with Gravity to explore the future of AI

[23:57] agent security. [music]

[23:58] To find out if we're truly ready for

[24:00] autonomous agents, I sat down with MIT

[24:03] professor Ramsh Roskar, former White

[24:05] House CIO Terresa Payton, Michelin's

[24:08] group chief data and AI officer Ambika

[24:10] Roger Gopal, and Sharon Guy, a former

[24:13] executive at Alibaba. They each offer

[24:17] unique insights into this evolving

[24:19] landscape. We conclude with Rory

[24:21] Blundell, CEO of Gravity, to discuss the

[24:24] path forward. [music] With Gravity

[24:26] leading the way, join us on this

[24:28] journey. You can watch the full

[24:29] documentary at the link in the show

[24:31] notes.

[24:33] [music]

[24:40] And we're back here on Big Technology

[24:42] Podcast with Andrew Bosworth, Bos, the

[24:44] CTO of Meta. Bos, great to see you

[24:46] again. Thank you for taking the time to

[24:48] speak with me. Um, if we go to the wide

[24:50] shot, we can see we're here in New York

[24:52] at a moment where you and your team are

[24:54] releasing three new pairs of Meta

[24:56] designed glasses. Um, it's it's uh it's

[25:00] something we've been debating on the

[25:01] show is sort of is your phone the AI

[25:05] device or is it a wearable? And we've

[25:08] had this moment again going back to

[25:09] Apple where it looks like they're

[25:11] preparing to release a version of Apple

[25:12] intelligence that actually works um that

[25:14] knows your context to a degree and might

[25:16] be able to get things done for you. And

[25:18] then we see, you know, sort of the

[25:20] opposite side is the Snapchat specs

[25:22] release which got a lot of people saying

[25:25] um maybe we don't it I mean those were

[25:28] so bad that people were just You don't

[25:29] have to comment. I'll say it.

[25:30] I can't comment. I haven't seen them. I

[25:32] haven't seen them myself yet.

[25:33] Let's just say what I I I'll just my

[25:35] comment reflects what the market did.

[25:36] Evan Spiegel wore them out to some

[25:38] presentation. I think Snap Stock went

[25:40] down like 6% immediately. Um it's just

[25:43] what happened.

[25:44] Well, this will be the first video of me

[25:45] wearing our new glasses. We'll see what

[25:47] happen [laughter]

[25:48] we'll let the market decide.

[25:49] Yeah. Um but but I'd love to hear your

[25:51] thoughts on your obviously Meta has

[25:53] invested a lot in this. You believe it's

[25:55] a compelling use case. Um if I were to

[25:58] say maybe we don't need AR or AI

[26:02] glasses, we can just use our phone. What

[26:04] would you say that makes you feel the

[26:06] other side of that bet?

[26:06] Yeah, phones are great. I mean I love

[26:08] phones. Uh I have two of them. I think

[26:09] they're I think they're wonderful

[26:10] devices. I I the the glasses from the

[26:13] very beginning the question we asked

[26:15] ourselves was this exact question. And

[26:16] we said, "Okay, phones are great. What

[26:18] is something that you wish you could get

[26:21] access to? It's on your phone without

[26:23] having to take your phone out of your

[26:24] pocket." And we came up with camera and

[26:25] audio. It's just very simple. It's like,

[26:26] "Cool, if I could just do that." The AI

[26:28] has been this tremendous tailwind where

[26:30] actually it unlocks a much larger swath

[26:33] of potential capability over time than

[26:35] what the phone can do um just through,

[26:38] you know, Bluetooth connections. Um and

[26:40] so yeah, it's much more promising now

[26:42] than it looked two years ago or three

[26:44] years ago. Two or three years ago, this

[26:45] looked like, hey, at some point you have

[26:47] to put a display on this and it has to

[26:48] become a standalone system and it has to

[26:50] have all this, you know, kind of

[26:53] accessories attached to it. Now, it

[26:54] actually looks like there's a totally

[26:56] enough room in the market for a big

[26:58] range of of of wearable devices. Glasses

[27:02] certainly probably not just glasses,

[27:04] probably a lot of other things. People

[27:05] don't want to wear glasses, they want to

[27:06] wear different things. Um, and some of

[27:09] those devices are just going to be input

[27:11] and output to your phone. That's cool.

[27:12] like your phone's great and and if it's

[27:14] just making your life like more

[27:16] efficient in terms of how it's doing

[27:17] input and output that's awesome. Some of

[27:19] them will be more complete. So for our

[27:21] the band display glasses for example we

[27:23] just launched a vibe coded platform for

[27:24] it and so anybody who wants to can go uh

[27:28] literally just build whatever app you

[27:30] want for the glasses. Now right now you

[27:32] kind of build the the the app and you

[27:34] like put them on the glasses. But in the

[27:36] future, there's no reason that couldn't

[27:37] just be you wearing the glasses in real

[27:39] time, telling the glasses what app you

[27:42] want right now and having it on the fly

[27:45] build that app for you.

[27:46] Interesting.

[27:46] You know what I'm saying? And so I think

[27:47] we are headed towards a very cool zone

[27:51] where it's a little less like app garden

[27:54] specific. You're still going to have

[27:55] these content homes. Content continues

[27:57] to be an evergreen and important thing

[27:58] as it has been on TV, as it has been on

[28:00] social media, as it has been everywhere.

[28:02] Um, so there's still going to be places

[28:04] where media that you want to reach lives

[28:07] and those are those look kind of like

[28:09] apps or like or channels or whatever

[28:10] lack of a better term, but there's a

[28:12] long tale of things like why does my

[28:14] toaster need an app? Let me ask you this

[28:15] in seriousness. Like my toaster has an

[28:18] app.

[28:18] I don't think it needs one.

[28:19] I don't want that, right?

[28:20] I just want to tell my AI agent, get me

[28:23] the toast that I want. It's the same

[28:24] toast I have every day. Just get it for

[28:26] me. I don't want to have to go do

[28:27] whatever thing is. What does your

[28:29] toaster app um is it does it let you

[28:31] I honestly I I I refuse to install it. I

[28:34] refuse to install it.

[28:34] I respect that.

[28:35] I refuse. I absolutely won't do it. Um

[28:38] and so

[28:39] you have to stand up for something.

[28:40] Yeah. Listen, there's a line there's a

[28:42] line that nobody you know

[28:44] I think I think you can actually it's a

[28:46] I have to admit sometimes like the it's

[28:49] so cool that you can have a specific app

[28:51] to control every aspect of the thing and

[28:53] I I respect that and I'm a I'm a tech

[28:55] guy, right? So I like the fidgety nature

[28:56] of it. But it's like literally at this

[28:58] point it's kind of gotten out of hand

[29:00] when I really just wanted to to tell an

[29:01] intelligent system, hey, get me the

[29:03] thing that I want and it can do that for

[29:05] me.

[29:06] Um, and we see an early form of, you

[29:07] know, our partnership with with Spotify,

[29:09] you ask the glasses to play music, it go

[29:11] if you have a Spotify account linked, it

[29:12] goes and gets the music you want and

[29:13] it's like, yeah, this is great. This is

[29:14] what I wanted. I didn't want to have to

[29:16] go through a bunch of steps to do this.

[29:18] Um, so for me at least, the way I'm

[29:20] thinking about this is not that uh

[29:22] phones are great and they're going to

[29:24] continue to be great. Um, I don't think

[29:26] the appy thing is the way the future's

[29:28] going to look. I think the future is

[29:30] going to be valuable services, right,

[29:31] that are provided to you and you getting

[29:34] access to those services the way that

[29:35] you want when you need it and paying

[29:38] money to the people who provide those

[29:39] valuable services all negotiated either

[29:41] in advance or on on demand.

[29:43] Yeah, I really believe in this. I saw

[29:45] you had the uh I was on the Meta AI app

[29:47] today and I saw there's a Garmin

[29:48] connector to the glasses and for me you

[29:50] know as I'm training uh I'd love to be

[29:52] able to say well I'm building up to this

[29:54] like half marathon um Meta AI find me a

[29:58] 5K in my neighbor in my area that in in

[30:02] this window um and sign me up

[30:05] totally

[30:06] and and to do that as I'm on a run.

[30:08] Yeah.

[30:08] So I don't need to spend an hour

[30:10] figuring it out on my own.

[30:12] Agree completely. and and taking it a

[30:13] higher level, you know, your meta AI

[30:16] ideally would already know that you're

[30:18] training and you have a goal that you're

[30:19] trying to reach and it's tied into all

[30:21] the pieces that that matter, your

[30:22] nutrition and your, you know, it's like

[30:24] that's like that's the direction we want

[30:25] to get this thing. Um, there's a lot of

[30:27] steps between now and then, but that is

[30:29] where we're going.

[30:30] The Orion glasses, we talked about those

[30:32] last time.

[30:33] Where do those stand? Those are the a

[30:34] full AR full glasses experience.

[30:36] Yeah. So Orion was such a a important

[30:39] moment for us, you know, having had this

[30:40] AR vision for such a long time, finally

[30:43] gave us the device that we could use to

[30:44] start to play with the software on. Um,

[30:47] and even though we couldn't get the

[30:48] price to be one that we felt comfortable

[30:51] launching as a consumer product, we did

[30:53] int when we designed it and developed

[30:54] it, it was a consumer design and

[30:56] intention. And so the product itself is

[30:58] like is quite wearable, quite workable.

[30:59] Like I have a pair at home. We use it to

[31:01] test the software. Um, so we've

[31:03] continued to iterate in the software and

[31:04] we've made so much more progress in the

[31:06] software, not just because AI has gotten

[31:07] better that that makes a huge difference

[31:09] to what that software is. Uh, but also

[31:11] because you have Orion to develop on,

[31:13] which makes a big difference. Um, so

[31:15] yeah, we continue to be very focused on

[31:17] the entire spectrum. You know, we've

[31:19] hinted here, uh, that, you know, in

[31:20] addition to display glasses and camera

[31:22] glasses, you know, there's a whole range

[31:24] of glasses that may be below that in the

[31:26] price range. Well, there also may be

[31:28] there. I really still believe in full AR

[31:30] as a future for the space. Um, I think

[31:33] we're going to continue to take the same

[31:35] approach we have so far and the same

[31:36] reason we didn't launch Orion. Um, it's

[31:39] not just enough that it does all this

[31:40] functionality. It has to look great. Has

[31:43] to be comfortable enough that you want

[31:44] to wear it. Um, has to be at a price

[31:47] point that that a reasonable person

[31:48] would say, "Yeah, this is this is a good

[31:50] value." So,

[31:51] how far away is that?

[31:53] I'm not going to say exact number. I

[31:55] will say uh I like the progress we're

[31:57] making.

[31:58] Measured in years or months?

[32:00] I'm [laughter] not gonna answer that.

[32:01] All right, that's fair.

[32:02] I appreciate the hustle.

[32:04] Have to ask.

[32:04] I know you do. It's some some of my

[32:06] reticence, you know, we we people who uh

[32:09] have been in companies like ours know

[32:11] this. We're constantly looking at

[32:12] vehicles and like asking ourselves, is

[32:14] this the one? Is it ready yet? You know,

[32:16] is this the one? And uh man, we're

[32:18] getting into the zone. It's pretty

[32:19] exciting.

[32:19] Okay, cool. Uh let's talk about

[32:21] metaculture for a moment. Um, you've

[32:23] you're running this applied AI division.

[32:25] That's right. Which has been the subject

[32:26] of some reporting.

[32:27] I run the agentic transformation uh

[32:30] accelerator,

[32:30] right?

[32:31] One of the groups in that is the is the

[32:32] AI team. Yeah.

[32:33] Okay. I'm just going to read the quote

[32:34] from Wired. Uh, one employee told Wired,

[32:37] "It's literally the goolog. You have

[32:38] zero purpose in life all of a sudden.

[32:40] You barely interact with anyone. You

[32:42] just have these tasks every week."

[32:43] apparently talking about how um you know

[32:46] the employees there have been put on

[32:47] some like AI puzzles that they they have

[32:50] to try to accomplish that helps train

[32:53] the AI. What's going on there?

[32:54] I'm not sure this person's ever googled

[32:57] what a goolog was like and how similar

[32:59] or not it is to a six-f figureure

[33:01] software job in Silicon Valley

[33:03] doesn't seem like it but the fact that

[33:04] they would say that [laughter]

[33:05] setting setting aside setting aside the

[33:07] hyperbole. Okay.

[33:08] Yeah. So we've been spending a lot of

[33:09] time on this internally. It's a hugely

[33:10] important topic for us. Uh you've been

[33:12] covering us a long time so you know this

[33:14] like this is a company that goes into

[33:15] lockdowns like when we have an urgent

[33:18] opportunity ahead of us we like do this

[33:20] we did it with mobile we did it with

[33:22] video we did it with stories we've done

[33:24] it and it's not that they

[33:26] um every one of these things pivots the

[33:28] entire company but there are moments

[33:29] we're like wait if we put exquisite

[33:32] effort on something right now we think

[33:34] there's a tremendous opportunity for us

[33:36] in the market um and in this case we saw

[33:38] that we we really feel like you know the

[33:41] when we came out with Muse Spark um and

[33:44] I want to be careful like Muse Spark is

[33:45] is a great model and we're really

[33:46] excited about it um and it's what coding

[33:50] had not been a focus for us on the model

[33:53] but it actually was a better out of the

[33:54] box at coding than we had expected it to

[33:56] be [clears throat]

[33:57] and we found early on through

[33:59] experiments that like actually giving it

[34:01] just um a relatively modest number of

[34:03] trained kind of expertly guided um uh

[34:07] examples and we could post-train the

[34:09] model we could dramatically improve its

[34:12] competitiveness. Um, and so when you

[34:16] start to like run the numbers and the

[34:17] math on this, you're like, "Oh, this is

[34:18] an incredible opportunity for us to

[34:20] build a coding model that not only um

[34:23] allows us to have independence to how we

[34:25] operate the company, but also something

[34:26] that we think is going to be valuable uh

[34:28] both inside of if you, you know, give

[34:30] users AI that's able to code, that's

[34:33] obviously one of the very very powerful

[34:34] tools that's kind of become very common

[34:36] in these AI systems over the last year."

[34:38] Um and then also um for us to be able to

[34:41] make the model itself more widely

[34:43] available over time. So we basically saw

[34:45] this huge opportunity, such a big

[34:47] opportunity that we pivoted kind of on a

[34:50] dime and brought a lot of people across

[34:52] the company, thousands of people um out

[34:55] into this AI organization to to do these

[34:58] expert traces. We absolutely need their

[35:01] expertise. It doesn't work if you do a

[35:03] bad job. It turns out if you use a bad

[35:05] piece of coding to train the model, uh

[35:07] you do some damage to it. They have

[35:08] you don't want to reinforce failure.

[35:10] They have to be well done. They have to

[35:12] be expertly guided.

[35:13] Um now we did it very quickly and as a

[35:16] consequence it did not have a lot of

[35:17] structure. It did not have great

[35:18] communication around it. I've been on

[35:20] record. Actually, that's not true. I

[35:21] wasn't on record. I was leaked calling

[35:22] it atrocious.

[35:23] You said maybe uh not the worst it's

[35:25] ever been in 20 years here, but it's up

[35:27] there. It's definitely up there.

[35:28] That actually was not a quote for me.

[35:29] And I don't know where that you didn't

[35:31] say that.

[35:31] I didn't say that.

[35:32] Okay. Okay.

[35:32] But I've said things like it I'm I'm

[35:34] fine with it. And uh but so the degree

[35:38] to which it's a big company. The degree

[35:40] to which we saw this urgent opportunity

[35:43] and made the change that I think

[35:44] strategically was absolutely the right

[35:45] change but did not do the work to kind

[35:47] of go to each person and be like let let

[35:49] me talk to you about what this is and

[35:50] why we need it and why it's important.

[35:52] Yeah.

[35:53] Knowing that they had they had other

[35:55] work that they were excited about that

[35:56] they were putting on pause to come do

[35:58] this work. um you know but this is that

[36:00] is something our company does

[36:02] um when we feel like we see these like

[36:04] unbelievable opportunities that exist in

[36:06] moments of time um and so yeah we are

[36:09] like navigating this this change that's

[36:11] happening in the industry is happening

[36:13] inside every company as well um and uh

[36:18] it's like nothing we've ever seen you

[36:20] said you let out with this it's like

[36:21] nothing we've ever seen before in our

[36:22] careers and I think that is giving

[36:24] people pause and so it raises the bar on

[36:28] me and other leaders do a much better

[36:29] job than we have done communicating

[36:32] what's going on, why is it happening,

[36:34] how does it affect you, how do we see it

[36:35] playing out long term. Um, make sure

[36:37] they understand that the role they're

[36:39] playing is one that we consider very

[36:41] critical, very important, otherwise we

[36:42] wouldn't have made that change,

[36:43] obviously.

[36:43] Can we talk about the tracking briefly?

[36:45] Um, I actually, you know, I, if I was an

[36:48] employee, I don't think I'd be a fan of

[36:50] it, but I actually sort of made the case

[36:52] for why you might be doing it on our

[36:54] show recently. And now that we're

[36:56] sitting next to each other, um, let's

[36:58] talk about it because um, so basically

[37:01] the the reports have been that that Meta

[37:03] has uh, started to track some keystrokes

[37:06] and the way that employees type um, and

[37:09] and basically use that as a way to train

[37:11] models. And my perspective on this was

[37:15] as model training moves into

[37:17] reinforcement learning where I think

[37:19] scale AI where Alexander Wayne came from

[37:20] said most of their training is

[37:22] reinforcement learning now um as opposed

[37:24] to pre-training which we talked about

[37:26] previously. Um as the technology moves

[37:28] into reinforcement learning. very

[37:30] valuable for these models to learn how

[37:33] to accomplish tasks in what's typically

[37:35] called as gyms or like different areas

[37:38] that different like simulations of real

[37:40] world activity that they go in and try

[37:42] to accomplish. And so am I right in

[37:44] thinking that this program is basically

[37:47] a just massively scaled up version of

[37:50] that where the models watch employees

[37:53] work through their tasks and then learn

[37:55] how to accomplish tasks on their own.

[37:56] Yeah. Well, there's two parts to this.

[37:58] Um the first one is uh you're absolutely

[38:01] right. Reinforcement learning is playing

[38:02] a hu a much bigger role in today's kind

[38:04] of AI than people had maybe predicted

[38:06] two or three years ago that it would. Um

[38:08] it's not just that though there's also

[38:10] the long tale is long like the long tale

[38:12] of human knowledge and and behavior is

[38:15] very long.

[38:16] And most of it as much as for all the

[38:18] text for the entire corpus of text on

[38:20] the internet most of the stuff that we

[38:22] know is still not on the internet. It's

[38:24] like in our heads. It's experience. It's

[38:25] built up over time. It's behaviors that

[38:27] are second nature to us. Um, and so this

[38:29] system was in some ways I thought quite

[38:32] genius. You've got employees who need to

[38:34] change nothing about how they go about

[38:35] their day, can go about it as they

[38:38] always have, and in doing so produce

[38:41] this corpus of unique data. In this

[38:43] case, design and how do how do humans

[38:45] use computers? AIS are actually still

[38:47] really weirdly bad at just using

[38:49] computers. Like it's it's like it's like

[38:52] a it's a surprisingly hard problem that

[38:54] is not well solved. And that's where all

[38:56] the energy is going with computer use

[38:58] and aentic that's all computer

[38:59] and you can you can ramp up uh you know

[39:01] the intelligence in the front end for

[39:03] sure and then try to distill down from

[39:04] that

[39:05] but we do think having this data has the

[39:07] potential of making people's lives

[39:09] easier. Um it's not even about the

[39:11] content the the thing that was a

[39:14] challenge to communicate and again we

[39:15] did a poor job was it's not even about

[39:17] the content of the thing that you're

[39:19] doing. It's about how is the computer

[39:20] able to understand what's happening

[39:22] inside this digital interface which is

[39:24] the way we access a lot of our tools

[39:26] like in in the world today. Um the

[39:28] second thing is uh and and so it's so I

[39:31] think this data set is interesting but

[39:33] it's we won't know it's like it's a

[39:34] longunning data set. So the second part

[39:36] of this is you're still for longtail

[39:39] expert training you're better off doing

[39:41] work like we are doing with our uh

[39:43] applied AI team uh the AI team like that

[39:46] is um a relatively small number of

[39:50] really well doumented um you know tasks

[39:54] that can train post train a model um

[39:56] this is a difference thing this is like

[39:58] very longunning once we have like a year

[40:00] of data you have something that's

[40:02] potentially interesting to to bring to

[40:04] bear in the model I do want to add we've

[40:06] also made a bunch of changes to the

[40:07] program since the launch. We've added

[40:09] take a 30-minute

[40:09] break unlimited pausing people can opt

[40:12] out for a bunch of reasons. So like

[40:13] we've we've made a bunch of changes to

[40:15] the program for people who had concerns

[40:16] about it.

[40:17] So you are posting a lot of your old

[40:19] blog posts uh to Substack and I've been

[40:22] getting them in my email and reading

[40:24] them and there was very interesting one

[40:26] uh that I read recently talking about

[40:28] how um you were doing uh some biology

[40:31] research and the doctor said the pain uh

[40:33] is rehab. You need that pain in order to

[40:36] be able to heal. Um you you write as at

[40:38] some point you have to be able you have

[40:40] to embrace the pain to make real

[40:42] progress. Given uh two otherwise equal

[40:44] stories, humans remember the story that

[40:46] evoked stronger emotion. Emotion is how

[40:49] our brain triages memories. Sometime it

[40:52] has sometimes it has to hurt for your

[40:54] brain to prioritize it. Um

[40:56] shout out to BS80 a class my my neurobio

[40:58] class at Harvard.

[40:59] AI is evolutionary bio.

[41:00] AI is taking away a lot of the pain,

[41:03] right? like big part of what humanity is

[41:05] doing with AI right now is a lot of the

[41:07] painful parts of our work we're giving

[41:10] it to AI. If that goal is accomplished

[41:13] where do we find the pain?

[41:15] So the this I love this and uh very

[41:18] small aside my sub I I one of the things

[41:20] I did is I assigned my agent the task of

[41:22] bringing my sub my blog posts over to

[41:24] Substack so at some point I could do

[41:26] both. Um I didn't realize until very

[41:28] recently that it wasn't any bulleted

[41:30] list. It would just strip out. My my

[41:31] agent did not understand bulleted lists.

[41:32] So we have a long ways to go on agents

[41:33] is my okay [laughter] phase one um the

[41:36] the pain is the rehab came yeah yeah

[41:37] there was a a question we were studying

[41:39] uh the neurobiology that would occur

[41:42] during withdrawal from uh drug use and a

[41:45] student asked hey we have all these uh

[41:47] symptoms why don't we just give people

[41:49] the pain a pain medicine and the the

[41:53] professor was like you don't understand

[41:54] the pain is the medicine like

[41:56] experiencing

[41:57] uh desire to pursue drugs drug-seeking

[42:00] behavior and then having it be immensely

[42:01] painful

[42:02] is the way you reprogram your brain to

[42:04] like overcome the drug-seeking behavior.

[42:07] Um, and if you get rid of the pain, then

[42:08] the person is like never going to do it.

[42:10] There is so this is a productive form of

[42:12] pain. By the way, I would argue AI all

[42:16] these peroxisms happening not just at

[42:17] Meta but at every company is the pain

[42:19] I'm talking about. That is the pain that

[42:21] there is no way out but through and you

[42:23] have to figure out the path through it

[42:24] to figure out what works and what

[42:26] doesn't work and it's just gritty. We do

[42:28] have lots of other types of pain in our

[42:30] society that have nothing to do with

[42:31] real value being created. Uh this this

[42:34] comes up a lot in education is a good

[42:35] example. Um I remember being told, I'm

[42:38] sure you were when I was in school. Uh

[42:40] hey, you can't use a calculator on this

[42:41] test. You will not have a calculator

[42:42] with you uh in as you go about your day

[42:45] in the real world. I have at

[42:48] least three calculators on my person at

[42:49] all times. Not to mention, I can just

[42:50] ask my glasses math problems. I I'm

[42:52] filthy with calculators. It turns out

[42:55] doing a math problem, doing a math test

[42:57] without a calculator is a certain kind

[42:59] of pain, not a particularly useful kind.

[43:02] Doing a harder math test that requires

[43:04] critical thinking with a calculator is

[43:06] probably the more valuable way to do

[43:07] that thing, right? So, I do think it's

[43:10] important to align the pain that we're

[43:11] experiencing with the value we're trying

[43:13] to create in the world. Um, I think like

[43:15] learning to integrate AI, you could

[43:16] avoid that pain. You just skip it. You

[43:18] don't do it. You and I both know that it

[43:20] puts you at real risk. You're gonna fall

[43:22] behind people who, you know, are able to

[43:25] do AI and want to do the same job as

[43:27] you. You're going to fall behind other

[43:29] companies that have integrated AI either

[43:31] economically or in the products that you

[43:32] offer. Um, you know, there's this Cheryl

[43:35] always had this, Cheryl Samber has this

[43:36] great quote, which is that companies

[43:38] don't usually fail by setting tough

[43:39] goals and missing them. They fail by

[43:41] setting easy goals and hitting them all

[43:43] the way down. Um, and so like I think

[43:45] you could easily avoid the pain today by

[43:48] just being like, "Yeah, we're just not

[43:49] going to we're just not going to do it.

[43:50] we're just going to let it happen and

[43:52] then we'll figure it out later on. Um,

[43:54] so I think there is productive pain and

[43:55] unproductive pain and maybe a little bit

[43:57] of judgment to know which one's which.

[43:59] Bos, it's really always a pleasure to

[44:00] speak with you. Thanks so much for

[44:01] coming on.

[44:01] Thanks for having me.

[44:02] All right, everybody. Thanks so much for

[44:03] listening and watching and we'll see you

[44:05] next time on Big Technology Podcast.