AI 编程代理与开源模型：马克的愿景

好的，马克，感谢再次来到播客。

All right Mark, thanks for coming on the podcast again.

嗯，很高兴来。见到你真好。

Yeah, happy to do it. Good to see you.

我也是。上次你来的时候，你刚发布了 Llama 3。现在你发布了 Llama 4——好吧，第一个版本。没错。有什么新东西？有什么激动人心的？有什么变化？

You too. Last time you were here, you had launched Llama 3. Now you've launched Llama 4. Well, the first version. That's right. What's new? What's exciting? What's changed?

哦，嗯，整个领域变化太快了。我觉得自从上次我们聊过之后，发生了很多变化。Meta AI 现在每月有近 10 亿用户在使用。这相当惊人。而且我认为今年会是这一切非常重要的一年，因为一旦你开始启动个性化循环——我们现在才刚刚开始真正构建——既利用所有算法对你兴趣内容的了解、你的个人资料信息、社交图谱信息，也利用你与 AI 交互的内容，我认为这将是下一个非常令人兴奋的方向。所以我对这个非常看重。建模方面也在持续取得令人瞩目的进展，你也知道。Llama 4 方面，我对第一批发布很满意。我们宣布了四个模型，发布了前两个：Scout 和 Maverick，它们属于中等规模模型，中到小型。实际上，最受欢迎的 Llama 3 模型是 80 亿参数的版本。所以我们在 Llama 4 系列中也会推出一个类似的。我们内部代号叫它“小 Llama”。但那个可能在未来几个月内发布。而 Scout 和 Maverick，它们确实很好。它们是市面上性价比最高的模型之一。原生多模态，非常高效，可以在单台主机上运行，专为我们内部构建的许多用例而设计，追求高效和低延迟。这就是我们的一贯做法：我们基本上构建我们想要的东西，然后开源，这样其他人也能使用。所以我对此很兴奋。我也对即将推出的 Behemoth 模型感到兴奋。那将是我们第一个处于前沿的模型。它有两万亿以上的参数。所以，正如其名，它非常大。我们正在想办法让它对人们有用。它太大了，我们不得不构建大量基础设施才能自己进行后训练。我们也在思考，普通开发者如何能使用这样的东西，以及我们如何让它能够蒸馏成合理大小的模型来运行，因为你显然不会想在消费级模型中运行那样的东西。但嗯，还有很多工作要做。就像你去年看到的 Llama 3 那样，最初的 Llama 3 发布令人兴奋，然后我们全年都在此基础上不断改进。3.1 版本我们发布了 4050 亿参数的模型。3.2 版本我们加入了所有多模态功能。所以今年我们也有类似的路线图。所以有很多事情在进行。

Oh, well, I mean, the whole field's so dynamic. So, I feel like a ton has changed since the last time that we talked. Meta AI has almost a billion people using it now monthly. So that's pretty wild. And I think that this is going to be a really big year on all of this because, especially once you start getting the personalization loop going, which we're just starting to build in now really, from both the context that all the algorithms have about what you're interested in feed and all your profile information, all the social graph information, but also just what you're interacting with the AI about, I think that's just going to be kind of the next thing that's going to be super exciting. So really big on that. The modeling stuff continues to make really impressive advances too, as you know. The Llama 4 stuff, I'm pretty happy with the first set of releases. We announced four models and we released the first two, the Scout and Maverick ones, which are kind of like the midsize models, midsize to small. It's not like you actually the most popular Llama 3 model was the 8 billion parameter model. So we've got one of those coming in the Llama 4 series too. Our internal code name for it is Little Llama. But that's coming probably over the next over the coming months. But the Scout and Maverick ones, and I mean they're good. They're some of the highest intelligence per cost that you can get of any model that's out there. Natively multimodal, very efficient, run on one host, designed to just be very efficient and low latency for a lot of the use cases that we're building for internally. And that's our all thing. We basically build what we want and then we open source it so other people can use it too. So I'm excited about that. I'm also excited about the Behemoth model which is coming up. That's going to be our first model that is sort of at the frontier. I mean it's like more than two trillion parameters. So it is, as the name says, it's quite big. So we're kind of trying to figure out how we make that useful for people. It's so big that we've had to build a bunch of infrastructure just to be able to post train it ourselves. And we're kind of trying to wrap our head around how the average developer out there, how are they going to be able to use something like this and how do we make it so it can be useful for distilling into models that are of reasonable size to run because you're obviously not going to want to run something like that in a consumer model. But yeah, I mean there's a lot to go. As you saw with the Llama 3 stuff last year, the initial Llama 3 launch was exciting and then we just kind of built on that over the year. 3.1 was when we released the 405 billion model. 3.2 was when we got all the multimodal stuff in. So we basically have a road map like that for this year too. So lot going on.

我很想了解更多。有一种印象是，过去一年里，最好的闭源模型和最好的开源模型之间的差距变大了。我知道 Llama 4 全系列还没发布，但 Llama 4 Maverick 在 Chatbot Arena 上是 35 分，而在许多主要基准测试上，o4 mini 或 GP Gemini 2.5 flash 似乎击败了同级别的 Maverick。你怎么看这种印象？

I'm interested to hear more about it. There's this impression that the gap between the best closed source and the best open source models has increased over the last year where I know the full family of Llama 4 models isn't out yet but Llama 4 Maverick is 35 on Chatbot Arena and then on a bunch of major benchmarks it seems like o4 mini or GP Gemini 2.5 flash are beating Maverick which is in the same class. What do you make of that impression?

嗯，好吧，有几件事。实际上，我认为今年对开源整体来说是很好的一年。对吧？回顾去年，我们做的 Llama 几乎是唯一真正超级创新的开源模型。现在领域里有很多这样的模型了。而且我认为，总体而言，关于今年开源将普遍超越闭源、成为最常用模型的预测，大致上是正确的。我觉得一个有趣的惊喜——在某些方面是积极的，在其他方面是消极的，但总体是好的——是不仅仅有 Llama。市面上有很多好模型。所以我认为这很好。然后是你提到的推理现象，你谈到了 o3、o4 和其他一些模型。我确实认为正在发生一种专业化：如果你想要一个在数学问题或编程等方面最好的模型，那么这些推理模型——它们能够消耗更多的测试时间或推理时算力来提供更多智能——确实是一个非常有吸引力的范式。但对于我们关心的很多应用来说，延迟和良好的性价比实际上是更重要的产品属性。如果你主要设计消费级产品，人们不一定愿意等半分钟让模型思考答案。如果你能在半秒内提供一个同样相当好的答案，那就很棒，这是一个很好的权衡。所以我认为这两个方向最终都会很重要。我对随着时间的推移将推理模型与核心语言模型整合持乐观态度。

Yeah, well okay there there's a few things. I actually think that this has been a very good year for open source overall. Right? If you go back to the like where we were last year, what we were doing with Llama was like the only real super innovative open-source model. Now you have a bunch of them in the field. And I think in general the prediction that this would be the year where open-source generally overtakes closed sources, the most used model models out there, I think is generally on track to be true. I think the thing that's been sort of an interesting surprise, I think positive in some ways, negative in others, but overall good, is that it's not just Llama. There are a lot of good ones out there. So I think that's quite good. Then there's the reasoning phenomenon which you are alluding to with talking about o3 and o4 and some of the other models. And I do think that there is this specialization that is happening where if you want a model that is sort of the best at math problems or coding or different things like that. I do think that these reasoning models with a lot of the ability to just consume more test time or inference time compute in order to provide more intelligence is a really compelling paradigm. But for a lot of the applications that we care about, latency and good intelligence per cost are actually much more important product attributes. If you're primarily designing for a consumer product, people don't necessarily want to wait like half a minute to go think through the answer. If you can provide an answer that's generally quite good too in like half a second then that's great and that's a good trade-off. So I think that both of these are going to end up being important directions. I am optimistic about integrating the reasoning models with the core language models over time.

我们正在尝试构建一个能推进 Llama 研究的编码智能体。我猜测，在未来 12 到 18 个月的某个时候，我们将达到这样一个点：大部分用于这些工作的代码将由 AI 编写。

We're trying to build a coding agent that advances Llama research. I would guess that like sometime in the next 12 to 18 months we'll reach the point where like most of the code that's going towards these efforts is written by AI.

我倾向于认为，至少在可预见的未来，这会导致对人力工作的需求增加，而不是减少。如果你把提供某项服务的成本降低到原来的十分之一，那么现在去做这件事可能就变得合理了。

I tend to think that for at least the foreseeable future this is going to lead towards more demand for people doing work not less. If you've gotten the cost of providing that service down to one-tenth of what it would have otherwise been, maybe now that actually makes sense to go do.

我猜测，如果你认为某人正在做的事情是坏的，而他们却认为它非常有价值，那么世界会变得更有趣、更奇怪。根据我的经验，大多数时候他们是对的，你是错的。我担心我们正在消除所有摩擦，从而完全被我们的技术所奖励劫持。

I would guess that the world is going to get a lot more like a lot funnier and like weirder if you think that something someone is doing is bad and they think it's really valuable. Most of the time in my experience they're right and you're wrong. I am worried that we're just removing all the friction between getting totally reward hacked by our technology.

我认为这是一个合理的担忧。我们需要在设计这些系统时深思熟虑。

I think that's a valid concern. We need to be thoughtful about how we design these systems.

好的，马克，感谢再次来到播客。

All right Mark, thanks for coming on the podcast again.

嗯，很高兴来。见到你真好。

Yeah, happy to do it. Good to see you.

我也是。

You too.

我觉得这基本上就是 Google 在最近一些 Gemini 模型上走的方向，我认为这很有前景。但我觉得还会有很多不同的事情发生。你还提到了整个聊天机器人竞技场的事情，我觉得这很有意思，它指向了如何做基准测试这个挑战，对吧？基本上就是如何知道哪些模型擅长哪些事情。过去一年我们通常做的一件事就是，把更多模型锚定在我们的 Meta 产品北极星用例上，因为无论是开源基准测试还是像 LM Arena 这样的东西，问题在于它们往往偏向于非常特定的用例集，而这些通常不是任何普通人在你的产品中实际做的事情。它们衡量的组合往往与人们在特定产品中关心的东西不同。正因为如此，我们发现过度优化这些东西常常让我们误入歧途，实际上并没有带来最高质量的产品、最多的使用量和 Meta 内部用户使用我们产品时最好的反馈。所以我们试图将北极星锚定在用户向我们报告的产品价值、他们说的想要的东西以及他们的显示偏好上，利用我们拥有的经验。所以有时候我觉得这些东西并不完全一致，而且我认为其中很多都相当容易被操纵，对吧？我的意思是，在竞技场上，你会看到像 Sonnet 3.7 这样的东西。它是个很棒的模型，对吧？但它并不在顶部。而我们的团队相对容易地调出了一个版本的 Llama 4 Maverick，基本上就排在了顶部，而我们发布的那个纯模型根本没有针对这个做任何调优，所以它排名更靠后。我觉得你只需要对某些基准测试保持谨慎，我们将主要根据产品来定指标。

And I think that's sort of the direction that Google has gone in with some of the more recent Gemini models. And I think that's really promising. But I think there's just going to be a bunch of different stuff that goes on. You also mentioned the whole chatbot arena thing, which I think is interesting and it goes to this challenge around how do you do the benchmarking, right? And basically how do you know what models are good for which things? And one of the things that we've generally tried to do over the last year is anchor more of our models in our Meta product northstar use cases, because the issue with both open-source benchmarks and any given thing like the LM Arena stuff is that they're often skewed for a very specific set of use cases, which are often not actually what any normal person does in your product. They are often weighted, the portfolio of things that they're trying to measure is different from what people care about in any given product. And because of that, we've found that trying to optimize too much for that stuff has often led us astray and actually not led towards the highest quality products and the most usage and best feedback within Meta as people use our stuff. So we're trying to anchor our northstar in basically the product value that people report to us and what they say they want and their revealed preferences, using the experiences that we have. So sometimes I think these things just don't quite line up, and I think a lot of them are quite easily gameable, right? I mean, on the arena, you'll see stuff like Sonnet 3.7. It's a great model, right? And it's not near the top. And it was relatively easy for our team to tune a version of Llama 4 Maverick that basically was way at the top, whereas the one we released that's the pure model actually has no tuning for that at all, so it's further down. I think you just need to be careful with some of the benchmarks, and we're going to index primarily on the products.

你是否觉得存在某个基准测试，能够捕捉到你认为的对用户价值的北极星，可以在不同模型之间进行某种客观衡量，然后你会说“我需要 Llama 4 在这个上面排第一”？

Do you feel like there is some benchmark which captures what you see as a northstar of value to the user, which can be sort of objectively measured between different models, and you're like, 'I need Llama 4 to come out on top on this'?

嗯，我们的基准测试基本上就是 Meta AI 中的用户价值，对吧？所以它不能用来比较其他模型。嗯，我们也许可以，因为我们可能能在那个环境中运行其他模型并做出判断，我认为这是开源的优势之一：基本上你有一个很好的社区，他们可以挑刺，“你的模型哪里不好，哪里好”。但我认为现实是，目前所有这些模型都针对略微不同的组合进行了优化。我的意思是，每个人都在朝着同一个方向努力。我认为所有领先的实验室都在试图创造通用智能，对吧，以及超级智能，不管你怎么称呼它，基本上就是能够引领一个富足世界的 AI，让每个人都拥有这些超人工具来创造他们想要的任何东西，从而极大地赋能人们并创造所有这些经济效益。我认为无论你怎么定义，这基本上就是很多实验室的目标。但毫无疑问，不同的人针对不同的事情进行了优化。我认为 Anthropic 的人真的专注于编码和围绕编码的智能体。OpenAI 的人，我认为最近更偏向推理。而且我认为有一个领域，如果我猜的话，最终可能会成为最常用的领域，那就是快速、交互非常自然、原生多模态，能够融入你一天中想要与之交互的方式。我想你有机会试用我们正在发布的新 Meta AI 应用。我们在里面放的一个有趣的东西是全双工语音的演示。它还很早期，对吧？我的意思是，我们还没有把它设为应用中的默认语音模型是有原因的，但它那种自然的对话感我觉得真的很有趣、很吸引人。我认为能够将它与正确的个性化结合，将会带来一种产品体验，你可以想象，再过几年，我们基本上会整天和 AI 聊天，聊我们好奇的各种事情。你会拿着手机，在手机上说话，在浏览信息流应用时和它说话。它会给你提供关于不同事物的背景信息。它会在你与人们在消息应用中互动时帮助你。最终，我认为我们会在日常生活中戴上眼镜或其他类型的 AI 设备，整天都能无缝地与之交互。所以我认为这就是北极星，无论什么基准测试能够让人们觉得质量好到他们愿意与之交互，我认为那才是最终对我们最重要的事情。

Well, our benchmark is basically user value in Meta AI, right? So it's but you can't compare other models. Well, we might be able to because we might be able to run other models in that and be able to tell, and I think that's one of the advantages of open source: basically you have a good community of folks who can poke holes at 'okay where is your model not good and where is it good.' But I think the reality at this point is that all these models are optimized for slightly different mixes of things. I mean, everyone is trying to go towards the same thing. I think all the leading labs are trying to create general intelligence, right, and superintelligence, whatever you call it, basically AI that can lead towards a world of abundance where everyone has these superhuman tools to create whatever they want, and that leads to dramatically empowering people and creating all these economic benefits. I think that's sort of however you define that, I think that's kind of what a lot of the labs are going for. But there's no doubt that different folks have optimized towards different things. I think the Anthropic folks have really focused on coding and agents around that. The OpenAI folks, I think, have gone a little more towards reasoning recently. And I think there is a space which, if I had to guess, will end up probably being the most used one, which is quick, very natural to interact with, very natively multimodal, that fits into throughout your day the ways that you want to interact with it. And I think you got a chance to play around with the new Meta AI app that we're releasing. One of the fun things we put in there is the demo for the full duplex voice. It's early, right? I mean, there's a reason why we haven't made that the default voice model in the app, but there's something about how naturally conversational it is that I think is just really fun and compelling. And I think being able to mix that in with the right personalization is going to lead towards a product experience where, you know, I would basically just guess that you go forward a few years, we're just going to be talking to AI throughout the day about different things that we're wondering. You'll have your phone, you'll talk on your phone, you'll talk to it while you're browsing your feed apps. It'll give you context about different stuff. It'll help you as you're interacting with people in messaging apps. Eventually, I think we'll walk through our daily lives and we'll either have glasses or other kinds of AI devices and just be able to seamlessly interact with it all day long. So I think that is kind of the north star, and whatever the benchmarks are that lead towards people feeling like the quality is that they want to interact with that, I think is actually the thing that is ultimately going to matter the most to us.

我有机会试用了两者，还有 Meta 应用，语音模式非常流畅，令人印象深刻。关于不同实验室在优化什么的问题，为了公正地表述他们的观点：我认为很多人认为，一旦你完全自动化了软件工程和 AI 研究，你就可以引发一场智能爆炸，拥有数百万个这些软件工程师的副本，复制从 Llama 1 到 Llama 4 之间发生的研究。那种规模的改进在几周或几个月内就能实现，而不是几年。所以关键在于闭环软件工程师，然后你就能成为第一个达到 ASI 的人。你怎么看？

I got a chance to play around with both and also the Meta app, and the voice mode was super smooth. It was quite impressive. On the point of what the different labs are optimizing for, to steelman their view: I think a lot of them think that once you fully automate software engineering and AI research, then you can kick off an intelligence explosion where you have millions of copies of these software engineers replicating the research that happened between Llama 1 and Llama 4. That scale of improvement again in the matter of weeks or months rather than years. And so it really matters to just close the loop on the software engineer, and then you can be the first to ASI. What do you make of that?

嗯，我个人认为这很有说服力。这也是为什么我们也有一个大型的编码项目。我们在 Meta 内部正在开发多个编码智能体。你知道，因为我们并不是一家企业软件公司。我们主要是为自己构建。所以，我们再次针对具体目标。我们并不是要构建一个通用的开发者工具。

Well, I personally think that's pretty compelling. And that's why we have a big coding effort, too. We're working on a number of coding agents inside Meta. You know, because we're not really an enterprise software company. We're primarily building it for ourselves. So again, we go for the specific goal. We're not trying to build a general developer tool.

我们正在尝试构建一个编码智能体和一个 AI 研究智能体，专门推进 Llama 的研究，并且完全接入我们的工具链。我认为这很重要，最终会成为完成这些工作的关键部分。我猜测在未来的 12 到 18 个月内，我们会达到这样一个点：大部分用于这些工作的代码将由 AI 编写。我不是指自动补全。现在你有很好的自动补全，但我说的是给它一个目标，它能运行测试、改进代码、发现问题。它已经能写出比团队中平均水平很高的工程师更高质量的代码。我认为这肯定会成为其中非常重要的一部分，但我不确定这是否就是全部。我认为这将是一个巨大的行业，也是 AI 发展的重要部分。但还有其他人——我的意思是，一种思考方式是，这是一个巨大的空间。我不认为只会有一家公司用一个优化函数来尽可能好地服务所有人。有很多不同的实验室在针对不同领域做领先的工作。有些更偏向企业或编码，有些更偏向生产力，有些更偏向社交或娱乐。在助手领域，有些会更偏向信息或生产力，有些更偏向陪伴。空间非常巨大。有趣的一部分是走向 AGI 的未来。需要发明的东西有一些共同线索，但最终需要创造很多东西。我想你会开始看到团队之间出现更多专业化分工。

We are trying to build a coding agent and an AI research agent that advances Llama research specifically and is fully plugged into our tool chain. I think that's important and will end up being an important part of how this stuff gets done. I would guess that sometime in the next 12 to 18 months, we will reach the point where most of the code going towards these efforts is written by AI. I don't mean autocomplete. Right now you have good autocomplete, but I'm talking about giving it a goal, it can run tests, improve things, find issues. It already writes higher quality code than the average very good person on the team. I think that's going to be a really important part of this, but I don't know if that's the whole game. I think that's going to be a big industry and an important part of how AI gets developed. But there are still guys—I mean, look, one way to think about this is this is a massive space. I don't think there's just going to be one company with one optimization function that serves everyone as best as possible. There are a bunch of different labs doing leading work towards different domains. Some are more enterprise or coding focused, some more productivity focused, some more social or entertainment focused. Within the assistant space, some will be more informational or productivity, some more companion focused. There's a huge amount of space. Part of what's fun is going towards this AGI future. There are common threads for what needs to be invented, but a lot of things need to be created. I think you'll start to see a little more specialization between the groups.

我觉得很有意思，你基本上同意会有智能爆炸，最终出现类似超级智能的东西。但如果真是这样——别让我误解你的意思——如果真是这样，为什么还要费心做个人助手之类的东西？为什么不先达到超人智能，然后再处理其他一切？

It's really interesting to me that you basically agree with the premise that there will be an intelligence explosion and something like superintelligence on the other end. But if that's the case, don't tell me I'm misunderstanding you. If that's the case, why even bother with personal assistants and whatever? Why not just get to superhuman intelligence first and then deal with everything else?

我认为这只是飞轮的一个方面。我通常不同意快速起飞论的一点是，建设物理基础设施需要时间。如果你想建一个吉瓦级的算力集群，那需要时间。Nvidia 需要时间来稳定新一代系统，然后你需要搞定网络、建楼、拿许可、获取能源——无论是燃气轮机还是绿色能源——整个供应链都需要时间。上次我来播客时我们聊过很多。其中一些是物理世界的人类时间问题。当你在堆栈的一个部分获得更多智能时，你会遇到另一组瓶颈。工程学总是这样：解决一个瓶颈，又出现另一个。另一个瓶颈是人们逐渐习惯、学习并与系统形成反馈循环。这些系统通常不会凭空出现，然后人们神奇地知道如何使用。存在一种共同进化：人们学习如何最好地使用 AI 助手，AI 助手学习人们关心什么，开发者让助手变得更好。你还会积累一个上下文基础。一两年后，AI 助手可以引用你几年前谈论过的事情。这很酷，但如果你第一天就推出完美的东西，这是做不到的。所以我的观点是：智能在巨大增长，人们与 AI 助手互动的采纳曲线和学习反馈及数据飞轮在快速上升，同时还有供应链、基础设施和监管框架的建设，以支持物理基础设施的扩展。所有这些都不可或缺，不仅仅是编码部分。

I think that's just one aspect of the flywheel. Part of what I generally disagree with on the fast takeoff thing is it takes time to build out physical infrastructure. If you want to build a gigawatt cluster of compute, that just takes time. It takes Nvidia time to stabilize their new generation of systems, then you need to figure out networking, build the building, get permitting, get energy—whether gas turbines or green energy—there's a whole supply chain. We talked about this last time I was on the podcast. Some of these are physical world human time things. As you get more intelligence in one part of the stack, you'll run into a different set of bottlenecks. That's how engineering always works: you solve one bottleneck, you get another. Another bottleneck is people getting used to and learning and having a feedback loop with the system. These systems don't tend to show up fully formed and then people magically know how to use them. There's a co-evolution where people learn how to best use AI assistants, the AI assistants learn what people care about, and developers make the assistants better. You also build up a base of context. Now you wake up a year or two into it, and the AI assistant can reference things you talked about a couple years ago. That's pretty cool, but you couldn't do that if you launched the perfect thing on day one. So my view is there's huge intelligence growth, a rapid curve on uptake of people interacting with AI assistants and the learning feedback and data flywheel, and also the buildout of supply chains, infrastructure, and regulatory frameworks to enable scaling of physical infrastructure. All of those are necessary, not just the coding piece.

嗯。

Mhm.

我想举一个具体的例子，我觉得很有意思：甚至几年前，我们在广告团队有一个项目，要自动化排名实验。这是一个相当受限的环境——不是开放式的编码。基本上是查看公司整个历史，每个工程师在广告系统中做过的每个实验，看哪些有效、哪些无效、结果如何，然后为可能提升性能的测试制定新假设。我们发现，我们被算力瓶颈卡住了，无法运行足够多的测试。事实证明，即使只有广告团队现有的工程师，我们已经有太多好想法要测试，但算力或测试人群不够。即使有 35 亿人使用你的产品，每个测试也需要统计显著性，所以需要几十万或几百万人。

I guess one specific example of this that I think is interesting: even a few years ago, we had a project on our ads team to automate ranking experiments. It's a pretty constrained environment—not open-ended code. Basically, look at the whole history of the company, every experiment any engineer has ever done in the ad system, look at what worked, what didn't, the results, and formulate new hypotheses for tests that could improve performance. What we found was we were bottlenecked on compute to run tests based on the number of hypotheses. It turns out even with just the humans we have on the ads team, we already have more good ideas to test than we have compute or cohorts of people to test them with. Even with three and a half billion people using your products, each test needs to be statistically significant, so it needs hundreds of thousands or millions of people.

通过测试能获得的吞吐量是有限的。所以即使只有我们现有的这些人，我们已经无法真正测试所有想测的东西了。因此，仅仅能测试更多东西并不一定能带来增量。我们需要达到这样一个点：AI 生成假设的平均质量，要高于我们实际能测试的那些、团队中最优秀的人类所能做到的水平，这样它才会变得稍微有用。我认为我们很快会达到那个点。但这并不是说，哦，这东西能写代码了，一切就突然大幅改善了。现实中有这些约束：首先它需要能做得还不错，然后你需要有算力和人力去测试，然后随着质量慢慢提升。我不知道，五年或十年后，会不会没有任何人类能生成比 AI 系统更好的假设？也许吧。在那个世界里，显然所有价值都将由此创造，但那不是第一步。

There's only so much throughput that you can get on testing through that. So we're already at the point, even with just the people we have, that we already can't really test everything that we want. So now just being able to test more things is not necessarily going to be additive to that. We need to get to the point where the average quality of the hypotheses that the AI is generating is better than what all the things above the line that we're actually able to test, that like the best humans on the team have been able to do, before it'll even be marginally useful for it. So I think we'll get there pretty quickly. But it's not like, okay cool, the thing can write code, all of a sudden everything is just improving massively. There are these real world constraints that basically it needs to first be able to do a reasonable job, then you need to have the compute and the people to test, and then over time as the quality creeps up. I don't know, are we here in like five or ten years and it's like no set of people can generate a hypothesis as good as the AI system? I don't know, maybe. In that world, obviously that's going to be how all the value is created, but that's not the first step.

公开可用的数据正在枯竭。因此，像 Meta、Google DeepMind 和 OpenAI 这样的主要 AI 实验室都与 Scale 合作，以突破可能的边界。通过 Scale 的数据工厂，主要实验室可以获得高质量数据来推动后训练，包括高级推理能力。Scale 的研究团队 Seal 正在通过实用的 AI 安全框架以及围绕安全和对齐的公开排行榜，为将高级 AI 融入社会奠定基础。他们最新的排行榜包括 humanity's last exam、Enigma、Eval、Multi-Challenge 和 Vista，这些测试涵盖从专家级推理到多模态谜题解决再到多轮对话表现等一系列能力。Scale 还刚刚发布了 Scale Evaluation，帮助诊断模型局限性。领先的前沿模型开发者依赖 Scale Evaluation 来改进其最佳模型的推理能力。如果你是一位 AI 研究员或工程师，想了解更多关于 Scale 的数据工厂和研究实验室如何帮助你超越当前能力前沿，请访问 scale.com/thwarkcash。

Publicly available data is running out. So major AI labs like Meta, Google DeepMind, and OpenAI all partner with Scale to push the boundaries of what's possible. Through Scale's data foundry, major labs get access to high-quality data to fuel post training, including advanced reasoning capabilities. Scale's research team, Seal, is creating the foundations for integrating advanced AI into society through practical AI safety frameworks and public leaderboards around safety and alignment. Their latest leaderboards include humanity's last exam, Enigma, Eval, Multi-Challenge, and Vista, which test a range of capabilities from expert level reasoning to multimodal puzzle solving to performance on multi-turn conversations. Scale also just released scale evaluation, which helps diagnose model limitations. Leading frontier model developers rely on scale evaluation to improve the reasoning capabilities of their best models. If you're an AI researcher or engineer and you want to learn more about how Scale's data foundry and research lab can help you go beyond the current frontier of capabilities, go to scale.com/thwarkcash.

那么，如果你认同这就是智能发展的方向，看好 Meta 的理由显然是你拥有所有这些分发渠道，你也可以利用它们来学习更多对训练有用的东西。你提到 Meta 应用现在有十亿活跃用户。不是那个应用，不是那个应用。那个应用是我们刚刚推出的独立产品。我觉得想用的人会觉得它有趣，体验很酷。我们可以稍微聊聊这个。我们在里面尝试一些新想法，我认为很新颖，值得讨论。但我主要说的是我们的应用。Meta AI 实际上在 WhatsApp 中使用最多。明白了。所以 WhatsApp 主要在美国以外使用。我们在美国刚刚突破一亿用户，但它不是美国主要的通讯系统，iMessage 才是。所以我认为美国人可能有些低估 Meta AI 的使用量，但这也是独立应用如此重要的部分原因——美国出于很多原因是最重要的国家之一。而 WhatsApp 是人们使用 Meta AI 的主要方式，但它不是美国主要的通讯系统，这意味着我们需要另一种方式来构建一个面向用户的一流体验。我想把问题说完，看空的观点是：如果 AI 的未来不仅仅是回答问题，而更像是虚拟同事，那么 Meta AI 在 WhatsApp 内部如何为你提供相关训练数据来打造一个完全自主的程序员远程工作者，这一点并不清楚。所以在这种情况下，现在谁拥有更多 LLM 分发渠道还那么重要吗？

So if you buy this view that this is where intelligence is headed, the reason to be bullish on Meta is obviously that you have all this distribution, which you can also use to learn more things that can be useful for training. You mentioned the Meta app now has a billion active users. Not the app, not the app. The app is a standalone thing that we're just launching now. It'll be fun for people who want to use it. It's a cool experience. We could talk about that for a bit. We're kind of experimenting with some new ideas in there that I think are novel and worth talking through. But I'm talking mostly about our apps. Meta AI is actually most used in WhatsApp. Got it. So it's, and WhatsApp is mostly used outside of the US. We just passed like 100 million people in the US, but it's not the primary messaging system in the US, iMessage. So I think people in the US probably tend to underestimate the Meta AI use somewhat, but it's also part of the reason why the standalone app is going to be so important is the US is, for a lot of reasons, one of the most important countries. And the fact that WhatsApp is the main way that people are using Meta AI and that's not the main messaging system in the US means that we need another way to build a first class experience that's in front of people. And I guess to finish the question, the bearish case would be that if the future of AI is less about just answering your questions and more so just being a virtual co-worker, it's not clear how Meta AI inside of WhatsApp gives you the relevant training data to make a fully autonomous programmer remote worker. So in that case, does it not matter that much who has more distribution right now with LLMs?

嗯，再说一次，我只是认为会有不同的东西。就像如果你身处互联网发展的初期，你会想，互联网的主要东西会是什么？会是知识工作，还是像大规模消费者应用？我不知道，你两者都会得到。你不必只选一个。现在世界很大很复杂，一家公司能建造所有这些东西吗？我认为通常答案是否定的。但针对你的问题，人们大部分情况下不会在 WhatsApp 里写代码，我也不预见人们开始在 WhatsApp 里写代码会成为主要用例，尽管我确实认为人们会让 AI 做很多事情，结果 AI 在编码，而他们可能并不知情。所以那是另一回事。但我们在 Meta 有很多人在写代码，他们使用 Meta AI。我们有一个内部工具叫 Metamate，以及围绕它构建的许多不同的编码和 AI 研究智能体，这有它自己的反馈循环，我认为可以很好地加速这些努力。但再次，我只是认为会有很多事情。我认为 AI 几乎肯定会解锁知识工作和代码方面的巨大革命。我也认为它会成为下一代搜索，以及人们获取信息和执行更复杂信息任务的方式。我还认为它会很有趣。我认为人们会用它来娱乐。互联网上很多东西都是表情包和幽默。我们手头有这项惊人的技术。想想看，人类有多少精力花在娱乐自己、设计、推动文化前进以及用幽默的方式解释我们观察到的文化现象上，这真是令人惊叹又有点好笑。我认为未来几乎肯定也会是这样。如果你看看 Instagram 和 Facebook 这类东西的演变，回到 10 年、15 年、20 年前，那时是文字。然后我们都有了带摄像头的手机，大部分内容变成了照片。然后移动网络变得足够好，如果你想在手机上观看视频，它不会一直缓冲。

Well, again, I just think that there are going to be different things. It's like if you were sitting at the beginning of the development of the internet and it's like, well, what's going to be the main internet thing, is it going to be knowledge work or is it going to be like massive consumer apps? It's like, I don't know, you get both. You don't have to choose one. And now the world is big and complicated and does one company build all that stuff? I think normally the answer is no. But to your question, people do not code in WhatsApp for the most part, and I don't foresee that people starting to write code in WhatsApp is going to be a major use case, although I do think that people are going to ask AI to do a lot of things that result in the AI coding without them necessarily knowing it. So that's a separate thing. But we do have a lot of people who are writing code at Meta and they use Meta AI. We have this internal thing that we call Metamate, and a number of different coding and AI research agents that we're building around that, and that has its own feedback loop and I think can get good for accelerating those efforts. But again, I just think that there are going to be a bunch of things. I think AI is almost certainly going to unlock this massive revolution in knowledge work and code. I also think it's going to be kind of the next generation of search and how people get information and do more complex information tasks. I also think it's going to be fun. I think people are going to use it to be entertained. And a lot of the internet is like memes and humor. And we have this amazing technology at our fingertips. And it is sort of amazing and kind of funny when you think about it how much of human energy just goes towards entertaining ourselves and design and pushing culture forward and finding humorous ways to explain cultural phenomenon that we observe. And I think that's almost certainly going to be the case in the future. If you look at the evolution of things like Instagram and Facebook, if you go back 10, 15, 20 years ago, it was like text. Then we all got phones with cameras. Most of the content became photos. Then the mobile networks got good enough that if you wanted to watch a video on your phone, it wasn't just like buffering.

所以这变得很有意思。过去大概 10 年里，大部分内容都已经转向视频了。现在人们在 Facebook 和 Instagram 上花的时间大部分都是看视频。但我不确定，你觉得 5 年后我们还会只是坐在信息流里消费视频媒体吗？我觉得不会，它会变成交互式的，对吧？就像你刷信息流时，里面的内容一开始可能看起来像一条 Reel，但你可以跟它说话、跟它互动，它会回应你，或者改变它正在做的事情，你甚至可以像进入游戏一样跳进去跟它互动，而这一切都将由 AI 驱动，对吧？所以我的意思是，有这么多不同的事情，而我们很有野心，所以我们在同时做其中的很多项。但我不认为有哪一家公司能把所有事情都做完。

So that got good. So over the last like 10 years, most of the content has moved, basically towards video at this point. Most of the time spent in Facebook and Instagram is video. But I don't know, do you think in 5 years we're just going to be sitting in our feed and consuming media that's video? It's like, no, it's going to be interactive, right? It's like you'll be scrolling through your feed and there will be content that is basically, I don't know, maybe it looks like a reel to start, but then you talk to it or you interact with it and it talks back or it changes what it's doing or you can jump into it like a game and interact with it, and that's all going to be like AI, right? So I guess my point is there's just all these different things and I guess we're ambitious so we're working on a bunch of them. But I don't think any one company is going to do all of it.

好的。那么在 AI 生成内容或 AI 交互这一点上，人们已经与 AI 治疗师、AI 朋友（甚至可能更多）建立了有意义的关系。随着这些 AI 变得更具独特性、更有人情味、更聪明、更自然和有趣，这种情况只会愈演愈烈。我们如何确保人们与 AI 建立的关系是健康的？

Okay. So on this point of AI generated content or AI interactions, already people have meaningful relationships with AI therapists, AI friends, you know, maybe more. And this is just going to get more intense as these AIs become more unique and more personable, more intelligent, more spontaneous and funny and so forth. How do we make sure people are going to have relationships with AI? How do we make sure that these are healthy relationships?

嗯，我认为有很多问题只有当你开始看到实际行为时才能真正回答。所以可能最重要的事就是从一开始就提出这个问题，并在每一步都关心它。但我也认为，如果过早地规定太多，说我们认为这些事情不好，往往会扼杀价值，对吧？因为人们会使用对他们有价值的东西。我设计产品的核心指导原则之一就是：人是聪明的，对吧？他们知道什么对自己的生活有价值。偶尔产品中会出现不好的事情，你当然希望把产品设计好来尽量减少这种情况。但如果你认为某个人在做的事情是坏的，而他们觉得那很有价值，根据我的经验，大多数时候他们是对的，你是错的，你只是还没有找到一个框架来理解为什么他们正在做的事情在他们的生活中是有价值和有帮助的。所以这基本上是我思考这个问题的主要方式。

Well, I think there are a lot of questions that you only really can answer as you start seeing the behaviors. So probably the most important upfront thing is just ask that question and care about it at each step along the way. But I think also being too prescriptive upfront and saying we think these things are not good often cuts off value, right? Because people use stuff that's valuable for them. One of my core guiding principles in designing products is that people are smart, right? They know what is valuable in their lives. Every once in a while something bad can happen in a product and you want to make sure that you design your products well to minimize that. But if you think that something someone is doing is bad and they think it's really valuable, most of the time in my experience they're right and you're wrong, and you just haven't come up with the framework yet for understanding why the thing that you're doing is valuable and helpful in their life. So that's kind of the main way that I think about it.

我确实认为人们已经在用 AI 来做很多这类社交任务了。我们看到人们使用 Meta AI 的主要用途之一，就是演练他们需要在生活中与他人进行的困难对话。比如，“我跟女朋友之间出了这个问题，帮我进行这场对话”，或者“我需要跟老板进行一次艰难的谈话，我该怎么开口？”这非常有用。然后我认为，随着个性化循环的启动，AI 开始越来越了解你，这会变得非常有吸引力。

I do think that people are going to use AI for a lot of these social tasks already. One of the main things that we see people using Meta AI for is kind of talking through difficult conversations that they need to have with people in their life. It's like, okay, I'm having this issue with my girlfriend or whatever, help me have this conversation, or I need to have this hard conversation with my boss at work, how do I have that conversation? That's pretty helpful. And then I think as the personalization loop kicks in and the AI just starts to get to know you better and better, I think that will just be really compelling.

你知道，在社交媒体领域工作久了，有一个统计数据我一直觉得挺疯狂的。我认为普通美国人拥有的朋友数量少于三个——他们视为朋友的人——而普通人对朋友的需求却要大得多，大概是 15 个左右，对吧？当然可能到某个点你会说，“好吧，我太忙了，应付不了更多人了。”但普通人想要的连接比他们实际拥有的要多。所以很多人会问，这会不会取代面对面的连接或现实生活中的连接？我的默认答案是，可能不会。我认为当你能够拥有物理连接时，它在很多方面都更好。但现实是，人们就是缺乏连接，而且很多时候他们感到比想要的更孤独。所以我认为，今天这些可能还带有一点污名化的事情，随着时间的推移，我们作为社会会找到合适的词汇来表达为什么它们有价值，为什么做这些事情的人是理性的，以及它们如何为生活增加价值。但我也认为这个领域还非常早期。

You know, one thing just from working on social media for a long time is there's a stat that I always think is crazy. The average American I think has fewer than three friends — three people that they'd consider friends — and the average person has demand for meaningfully more. I think it's like 15 friends or something, right? I guess there's probably some point where you're like, "All right, I'm just too busy. I can't deal with more people." But the average person wants more connectivity, more connection than they have. So there's a lot of questions that people ask of stuff like, okay, is this going to replace in-person connections or real life connections? And my default is that the answer to that is probably no. I think there are all these things that are better about physical connections when you can have them. But the reality is that people just don't have the connection and they feel more alone a lot of the time than they would like. So I think that a lot of these things that today there might be a little bit of a stigma around, I would guess that over time we will find the vocabulary as a society to be able to articulate why that is valuable and why the people who are doing these things are rational for doing it and how it is adding value for their lives. But also I think that the field is very early.

我的意思是，有一些公司在做虚拟治疗师，还有类似虚拟女友之类的东西，但都还非常早期，对吧。这些产品的具身化程度还很弱。很多产品你打开后，只是一个治疗师或你正在交谈的人的形象。有时会有一些非常粗糙的动画，但算不上具身化。你见过我们在 Reality Labs 里做的 Codec Avatars，感觉就像真人一样。我认为这就是未来的方向。你基本上可以拥有一个始终在线的视频聊天，而且 AI 还能——你知道，手势也很重要。在实际对话中，超过一半的沟通不是你所说的话语，而是所有非语言的东西。

I mean, there are a handful of companies and stuff who are doing virtual therapists and there's like virtual girlfriend type stuff, but it's very early, right. The embodiment in the things is pretty weak. A lot of them, you open it up and it's just an image of the therapist or the person you're talking to or whatever. I mean, sometimes there's some very rough animation, but it's not like an embodiment. I mean, you've seen the stuff that we're working on in Reality Labs where you have the Codec Avatars and it feels like it's a real person. I think that's kind of where it's going. You'll be able to basically have an always-on video chat where it's like, oh, and also the AI will be able to, you know, the gestures are important, too. More than half of communication when you're actually having a conversation is not the words that you speak. It's all the non-verbal stuff.

我前几天确实有机会体验了 Orion，我觉得它超级令人印象深刻。我对这项技术总体上是乐观的，因为就像你提到的，我基本上是自由意志主义的——如果人们在做什么事，可能认为那对他们有好处。不过我不确定，如果有人在用 TikTok，他们会不会说对自己花在 TikTok 上的时间感到满意。所以我乐观的另一个原因是，如果我们未来要生活在 AGI 的世界里，为了跟上它，人类也需要用这样的工具来升级自己的能力。而且，如果你到处都能看到吉卜力风格的东西，世界总体上会变得更美。但我担心的是，你的团队给我展示的一个旗舰用例是：我坐在早餐桌前，视野边缘有一堆 Reel 在滚动。也许未来我的 AI 女友就在屏幕的另一边之类的。所以我担心我们正在移除所有摩擦，让自己完全被技术劫持了奖励机制。

I did get a chance to check out Orion the other day and I thought it was super impressive, and I'm mostly optimistic about the technology just because generally I'm, as you mentioned, libertarian about if people are doing something, probably think it's good for them. Although I actually don't know if it's the case that if somebody is using TikTok they would say that they're happy with how much time they're spending on TikTok or something. So I'm mostly optimistic about it also in the sense that if we're going to be living in this future world of AGI, we need to, in order to keep up with it, humans need to be upgrading our capabilities as well with tools like this. And just generally there's going to be more beauty in the world if you can see Studio Ghibli everywhere or something. I was worried that one of the flagship use cases that your team showed me was I'm sitting at the breakfast table and on the periphery of my vision is just a bunch of reels that are scrolling by. Maybe in the future my AI girlfriend is on the other side of the screen or something. And so I am worried that we're just removing all the friction between getting totally reward hacked by our technology.

嗯，我们如何确保五年后不会变成这样？

Um yeah, how do we make sure like I don't know this is not what ends up happening in 5 years?

我觉得人们很清楚自己想要什么。你看到的那个演示只是为了展示多任务和全息影像，对吧？我同意，未来不会是你视野角落里总有东西在争夺你的注意力。我觉得人们不会喜欢那样。所以我们在设计这些眼镜时，最关注的一点就是：眼镜首先要做到不碍事，成为一副好眼镜。顺便说一句，我认为这也是 Rayban Meta 产品成功的原因之一——它很适合听音乐、打电话、拍照和录像，AI 在你需要时出现，不需要时它就是一副好看又受欢迎的眼镜，几乎不打扰你。我想这会是增强现实未来一个非常重要的设计原则。我在这里看到的主要问题是，数字世界在我们生活中如此重要，但我们只能通过物理的数字屏幕来访问它，这有点疯狂。比如你有手机、电脑，可以放一台大电视，都是巨大的物理设备。现在技术似乎已经到了物理世界和数字世界应该完全融合的阶段，而全息叠加就能实现这一点。但我同意，这方面的设计原则很大一部分是：你会与人互动，并能将数字物件无缝带入这些互动中。比如我想给你看个东西，这里有个屏幕，我展示给你，你可以与之互动，它可以是 3D 的，我们可以一起玩。你想玩纸牌游戏之类的，这里就有一副牌，我们两个真人在这里，还有一个全息投影的朋友，他也能参与。但我认为在那个世界里，就像你不希望物理空间杂乱一样，数字化的物理空间也不应该让人有心理负担。这更多是审美和规范的问题，需要慢慢解决，但我们会搞定的。

I mean again I think I think people have a good sense of what they want. I mean that experience that you saw was a demo just to show multitasking and holograms, right? So I mean I agree that I don't think that the future is like you have stuff that's trying to compete for your attention in the corner of your vision all the time. I don't think people would like that too much. Um, so it's actually one of the things as we're designing these glasses that we're really mindful of is probably the number one thing that glasses need to do is get out of the way and be good glasses, right? And um, as an aside, I think that's part of the reason why the Rayban Meta product has done so well is like all right, it's great for listening to music and taking phone calls and taking photos and videos and the AI is there when you want it, but when you don't, it's a great, good-looking pair of glasses that people like and it kind of gets out of the way. Well, um, I would guess that that's going to be a very important design principle for the augmented reality future, right? The main thing that I see here is, you know, I think it's kind of crazy that for how important the digital world is in all of our lives, the only way we can access it is through these physical, you know, digital screens, right? It's like you have a phone, you have your computer, you can put a big TV. It's like this huge physical thing. Um, it just seems like we're at the point with technology where the physical and the digital world should really be fully blended and that's what the holographic overlays allow you to do. Um, but I agree. I think a big part of the design principles around that are going to be okay, you'll be interacting with people and you'll be able to bring digital artifacts into those interactions and be able to do cool things very seamlessly, right? It's like if I want to show you something here, like here's a screen. Okay, here it is. I can show you. You can interact with it. It can be 3D. Um, we can kind of play with it. Um, you want to, you know, like play a card game or whatever. It's like all right, here's a deck of cards. We can play with it. It's like two of us are here physically like you have a third friend who's just hologramming in, right? And that they can participate too. Um, but I think that in that world people are going to be you know just like you don't want your physical space to be cluttered. It's sort of like it just kind of has a it wears on you psychologically. I don't think people are going to want the digital kind of physical space to feel that way either. So, I don't know that that's more of an aesthetic and one of these norms that I think will have to get worked out. But, um, I think we'll figure that out.

回到 AI 的话题，你提到物理基础设施可能是一个巨大的瓶颈。关于其他开源模型，比如 DeepSeek。DeepSeek 目前的算力比 Meta 这样的实验室少，但可以说它与 Llama 模型有竞争力。如果中国在物理基础设施、工业规模化、上线更多电力和数据中心方面更擅长，你有多担心他们可能会在这方面击败我们？

Going back to the AI conversation, you're mentioning how big of a bottleneck the physical infrastructure can be. Related to other open source models like Deepseek and so forth. Deepseek right now has less compute than a lab like Meta and you could argue that it's competitive with the llama models. Um, if China is better at physical infrastructure, industrial scaleups, getting more power and more data centers online, how worried are you that this will they might beat us here?

我认为这是一场真正的竞争。工业政策正在发挥作用。中国正在上线更多电力，因此美国真的需要专注于简化数据中心建设和能源生产的流程，否则我们将处于显著劣势。同时，我认为对芯片等产品的出口管制显然在起作用。关于 DeepSeek，大家都在讨论他们做了非常令人印象深刻的底层优化。他们确实做了，这很了不起，但问题是为什么美国实验室没有这样做？因为他们使用的是被阉割的芯片——由于出口管制，英伟达只能在中国销售这种芯片。所以 DeepSeek 不得不花费大量精力进行底层基础设施优化，而美国实验室不需要。他们在文本上取得了不错的结果，但 DeepSeek 只支持文本。基础设施令人印象深刻，文本结果也不错。但现在每个新的大模型都是多模态的，支持图像、语音，而他们的不是。为什么？我不认为他们没能力做，而是因为他们必须把精力花在基础设施优化上，以克服出口管制。但当你比较 Llama 4 和 DeepSeek 时，我们的推理模型还没发布，所以 R1 的比较还不清楚。不过，在技术方面我们基本处于同一水平，但我们的模型更小，因此每单位智能的成本更低。在文本方面，Llama 更高效，而在多模态方面我们实际上领先，他们的产品甚至没有这些功能。所以我认为 Llama 4 模型与他们的相比是优秀的，人们会更倾向于使用 Llama 4。但有趣的是，那边显然有一个优秀的团队，你关于电力、算力和芯片可及性的问题问得很好，因为不同实验室的工作在一定程度上取决于这些因素。

I mean, I think it's a real competition. I mean, I think that you're seeing the industrial policies really play out. Um, where yeah I mean I think China's bringing online more power and because of that I think that the US really needs to focus on streamlining the ability to build data centers and build and produce energy or I think we will be at a significant disadvantage. Um, at the same time, I think some of the export controls on things like chips, I think you can see how they're clearly working in a way because, you know, there was all the conversation with DeepSeek about, oh, they did all these like very impressive low-level optimizations. And the reality is they did, and that is impressive, but then you ask why did they have to do that when none of the like American labs did it? And it's like, well, because they're using like partially nerfed chips that are the only thing that Nvidia is allowed to sell in China because of the export controls. So, DeepSeek basically had to go spend a bunch of their calories in time doing low-level infrastructure optimizations that the American labs didn't have to do. Now, they produced a good result on text, right? It's like I mean, DeepSeek is text only. Um, so the infrastructure is impressive, the text result is impressive. Um, but every new major model that comes out now is multimodal, right? It's image, um, it's voice and theirs isn't. And now the question is why is that the case? I don't think it's because they're not capable of doing it. I think that they basically had to spend their calories on doing these infrastructure optimizations to overcome the fact that there were these export controls. Um, but when you compare like Llama 4 with DeepSeek, I mean, our reasoning model isn't out yet. So, I think that the kind of R1 comparison isn't clear yet, but um, but we're basically like effectively same ballpark on all the tech stuff is what DeepSeek is doing, but with a smaller model. So it's much more kind of efficient per the kind of cost per intelligence is lower with what we're doing for llama on text and then all the multimodal stuff we're effectively leading at and it just doesn't even exist in their stuff. So um so I think that the llama 4 models when you compare them to what they're doing are good and I think generally people are going to prefer to use the llama 4 models. Um, but I think that there is this interesting contour where like it's clearly a good team that's doing stuff over there and I think you're right to ask about the accessibility of power, the accessibility of compute and chips and things like that. Um, because I think what the kind of work that you're seeing the different labs do and play out I think is somewhat downstream of that.

高端产品会吸引大量虚假账户注册、机器人流量和免费层滥用。现在 AI 太强了，在注册页面上放六个歪歪扭扭的数字验证码基本没用。以 Cursor 为例，人们为了利用 Cursor 的免费额度不择手段，创建并删除数千个账户、共享登录信息，甚至通过 Reddit 协调。所有这些都在推理算力和 LLM API 调用上花费了 Cursor 大量资金。然后他们接入了 Work OS Radar。Radar 能够区分人类和机器人。

Premium products attract a ton of fake account signups, bot traffic, and free tier abuse. And AI is so good now that it's basically useless to just have a capture of six squiggly numbers on your signup page. Take Cursor. People were going to insane lengths to take advantage of Cursor's free credits, creating and deleting thousands of accounts, sharing login, even coordinating through Reddit. And all this was costing Cursor a ton of money in terms of inference compute and LLM API calls. Then they plugged in work OS Radar. Radar distinguishes humans from bots.

它从你的 IP 地址、浏览器，甚至电脑上安装的字体等 80 多种信号进行检测，确保只有真实用户能通过。Radar 目前每周运行数百万次检查。当你把 Radar 接入自己的产品时，你立刻就能受益于 Radar 从其他顶级公司那里已经见过的数百万个训练样本。以前，只有大公司才能内部搭建这种级别的先进防护。但现在有了 WorkOS Radar，高级安全只需一个 API 调用。了解更多请访问 workos.com/radar。好了，回到 Zuck。

It looks at over 80 different signals from your IP address to your browser to even the fonts installed on your computer to ensure that only real users can get through. Radar currently runs millions of checks per week. And when you plug Radar into your own product, you immediately benefit from the millions of training examples that Radar has already seen through other top companies. Previously, building this level of advanced protection in-house was only possible for huge companies. But now with WorkOS Radar, advanced security is just an API call away. Learn more at workos.com/radar. All right, back to Zuck.

Sam Altman 最近发推说 OpenAI 要发布一个开源推理模型。推文里有一部分是说，他们不会做那种傻事，比如规定只有用户数少于 7 亿才能用。DeepSeek 用的是 MIT 许可证。而 Llama，我记得许可证里有几个附加条件，要求在使用它的应用上标注“built with Llama”，或者任何用 Llama 训练的模型名字必须以“Llama”开头。你怎么看这个许可证？对开发者来说，应该更宽松一些吗？

So Sam Altman recently tweeted that OpenAI is going to release an open-source reasoning model. I think part of the tweet was that we will not do anything silly like say that you can only use it if you have less than 700 million users. DeepSeek has the MIT license. Whereas Llama, I think a couple of the contingencies in the Llama license require you to say "built with Llama" on applications using it, or any model that you train using Llama has to begin with the word "Llama". What do you think about the license? Should it be less onerous for developers?

你看，我们基本上是开源大语言模型的开创者。所以我不认为这个许可证很苛刻。当我们开始推动开源时，行业里有一场大辩论：这样做到底合不合理？开源能做到安全可信吗？开源模型能足够有竞争力以至于有人在乎吗？在回答这些问题时，Meta 团队做了很多艰苦的工作——虽然行业里也有其他人——但实际上是 Llama 模型在很大程度上打开了整个开源 AI 的局面。我们非常关注：好吧，如果我们投入这么多精力，那么至少，如果像微软、亚马逊、谷歌这样的大型云公司要转售我们的模型，我们至少应该能在他们这么做之前和他们谈一谈，讨论一下我们应该有什么样的商业安排。但我们的许可证目标并不是阻止人们使用模型。我们只是觉得：好吧，如果你是那些公司之一，或者你是苹果，那就来和我们谈谈你想做什么，我们一起找到一种有成效的方式。所以我认为这总体上是没问题的。现在，如果整个开源领域发展到有很多其他优秀选择的方向，并且如果许可证最终成为人们不想用 Llama 的原因，那么我们就得重新评估策略。但我认为还没到那一步。实际上我们并没有看到公司来找我们说不想用这个，因为你的许可证规定如果达到 7 亿用户就必须来和我们谈。到目前为止，这更多是来自开源纯粹主义者的声音：这个模型的开源程度是否像你希望的那样纯粹？这种争论从开源诞生之初就存在，比如 GPL 许可证和其他许可证的争论。是不是任何接触开源的东西都必须开源，还是人们可以拿来以不同方式使用？我确信关于这个的争论还会继续。但如果你花了几十亿美元训练这些模型，我认为要求其他那些同样巨大、完全有能力与我们建立关系的大公司在使用之前先和我们谈谈，这似乎相当合理。

I mean, look, we've basically pioneered the open-source LLM thing. So I don't consider the license to be onerous. When we started pushing on open source, it was a big debate in the industry: is this even reasonable? Can you do something safe and trustworthy with open source? Will open source ever be competitive enough that anyone will care? When we were answering those questions, a lot of the hard work by the teams at Meta—though there are other folks in the industry—but really the Llama models broke open this whole open-source AI thing in a huge way. We were very focused on: okay, if we're going to put all this energy into it, then at a minimum, if you're going to have these large cloud companies like Microsoft, Amazon, and Google turn around and sell our model, we should at least be able to have a conversation with them before they do that, around what kind of business arrangement we should have. But our goal with the license isn't to stop people from using the model. We just think: okay, if you're like one of those companies or if you're Apple, just come talk to us about what you want to do, and let's find a productive way to do it together. So I think that's generally been fine. Now, if the whole open-source part of the industry evolves in a direction where there are a lot of other great options, and if the license ends up being a reason why people don't want to use Llama, then we'll have to reevaluate the strategy. But I just don't think we're there. That's not in practice a thing we've seen—companies coming to us saying we don't want to use this because your license says that if you reach 700 million people you have to come talk to us. So far, it's more something we've heard from open-source purists: is this as clean of an open-source model as you'd like it to be? That debate has existed since the beginning of open source with GPL license stuff versus other things. Does it need to be that anything that touches open source has to be open source, or can people just take it and use it in different ways? I'm sure there will continue being debates around this. But if you're spending many billions of dollars training these models, I think asking other huge companies that can easily afford to have a relationship with us to talk to us before they use it seems pretty reasonable.

如果事实证明其他模型也不错——有很多优秀的开源模型，所以你那一部分使命已经完成了，而且也许其他模型在编程方面更好。有没有一种情况，你会说：看，开源生态系统很健康，竞争很充分，我们很乐意就用其他模型——无论是用于 Meta 内部的软件工程，还是部署到我们的应用——我们不一定非要用 Llama 来构建？

If it turns out that other models are also good—there are a bunch of good open-source models so that part of your mission is fulfilled, and maybe other models are better at coding. Is there a world where you just say: look, the open-source ecosystem is healthy, there's plenty of competition, we're happy to just use some other model—whether for internal software engineering at Meta or deploying to our apps—we don't necessarily need to build with Llama?

嗯，再说一次，我们做很多事情，所以有可能。让我们退一步。我们构建自己大模型的原因是我们希望能够精确地构建我们想要的东西。世界上没有其他模型完全符合我们的需求。如果它们是开源的，你可以拿来用不同方式微调，但你仍然要处理模型架构，它们会做出不同的大小权衡，影响延迟和推理成本。在我们运营的规模上，这些东西真的很重要。我们把 Llama Scout 和 Maverick 模型做成特定大小是有特定原因的——因为它们适合放在一台主机上，而且我们想要特定的延迟，尤其是我们正在开发的语音模型，我们希望它渗透到我们做的所有事情中，从眼镜到所有应用再到 Meta AI 应用等等。所以只有当你自己构建时，你才能对自己的命运有一定程度的掌控。话虽如此，AI 将被用于每家公司做的每一件事。当我们构建一个大模型时，我们也需要选择内部要优化哪些用例。那么这是否意味着对于某些事情，我们不会认为也许 Claude 更适合构建这个团队正在使用的特定开发工具？好吧，那就用那个。没问题，很好。我认为我们不想自缚手脚。我们做了很多不同的事情。你还问到：会不会因为其他人在做开源，所以这件事就不重要了？我对此有点担心，因为我认为你必须问问那些现在出现并做开源的人——既然我们已经做了。

Well, again, we do a lot of things, so it's possible. Let's take a step back. The reason we're building our own big models is because we want to be able to build exactly what we want. None of the other models in the world are exactly what we want. If they're open source, you can take them and fine-tune them in different ways, but you still have to deal with the model architectures, and they make different size trade-offs that affect latency and inference cost. At the scale we operate, that stuff really matters. We made the Llama Scout and Maverick models certain sizes for a specific reason—because they fit on a host and we wanted certain latency, especially for the voice models we're working on, which we want to pervade everything we're doing, from the glasses to all our apps to the Meta AI app and all this stuff. So there's a level of control of your own destiny that you only get when you build the stuff yourself. That said, there are a lot of things where AI is going to be used in every single thing every company does. When we build a big model, we also need to choose which use cases internally we're going to optimize for. So does that mean for certain things, we're not going to think that maybe Claude is better for building this specific development tool that this team is using? All right, cool. Then use that. Fine, great. I don't think we want to fight with one hand tied behind our back. We're doing a lot of different stuff. You also asked: would it maybe not be important because other people are doing open source? I'm a little more worried on this, because I think you have to ask for anyone who shows up now and is doing open source now that we have done it.

有一个问题：如果我们不做开源，他们还会做吗？

There's a question which is would they still be doing open source if we weren't doing it?

我认为有一部分人看到了越来越多的开发转向开源的趋势，他们会想，糟了，我们得赶上这趟车，否则就会落后。我们有一些闭源模型 API，但越来越多的开发者不想要这个。所以你会看到很多其他玩家开始做一些开源工作，但尚不清楚这对他们来说是浅尝辄止，还是像对我们一样是根本性的。一个很好的例子是 Android 的情况，对吧？Android 最初是开源的，现在基本上没有真正的开源替代品了。我觉得随着时间的推移，Android 变得越来越封闭了。所以如果我们是你们，我们就会担心，如果我们停止推动行业朝这个方向发展，其他这些人可能只是因为他们要和我们竞争、和我们推动的方向竞争才这么做。他们已经表露过，如果开源不存在，他们会构建什么。所以我只是觉得，我们需要谨慎地依赖他们持续的行为来支撑我们公司未来要构建的技术。

I think that there are a handful of folks who see the trend that more and more development is going towards open source and they're like, crap, we kind of need to be on this train or else we're going to lose. We have some closed model API and increasingly a lot of developers don't want that. So I think you're seeing a bunch of the other players start to do some work in open source, but it's just unclear if it's dabbling or fundamental for them in the way that it has been for us. A good example is what's going on with Android, right? Android started off as open source. There's not really any open source alternative. I think over time Android has just been getting more and more closed. So if you're us, you'd kind of need to worry that if we stopped pushing the industry in this direction, all these other people maybe are only really doing it because they're trying to compete with us and the direction that we're pushing things. They already have their revealed preference for what they would build if open source didn't exist. So I just think we need to be careful about relying on that continued behavior for the future of the technology that we're going to build at the company.

我还听你提到过，标准应该围绕像 Llama 这样的美国模型来构建，这一点很重要。我想理解你的逻辑，因为对于某些类型的网络，苹果 App Store 确实有很大的偶然性。但似乎如果你为 DeepSeek 构建了某种脚手架，你也不能轻易地把它切换到 Llama 4，尤其是因为不同代际之间，比如 Llama 3 不是，Llama 4 是。所以模型代际之间也在变化。为什么认为事情会以这种偶然的方式围绕特定标准构建呢？

Another thing I've heard you mention is that it's important that the standard gets built around American models like Llama. I wanted to understand your logic there because it seems like with certain kinds of networks it is the case that the Apple App Store just has a big contingency around what it's built around. But it doesn't seem like if you build some sort of scaffold for DeepSeek, you couldn't have easily just switched it over to Llama 4, especially since between generations, like Llama 3 wasn't, Llama 4 is. So things are changing between generations of models as well. What's the reason for thinking things will get built out in this contingent way on a specific standard?

我不确定你说的“偶然”是什么意思？

I'm not sure what do you mean by contingent?

哦，意思是人们为 Llama 构建而不是为通用大语言模型构建很重要，因为这会决定标准是什么。

Oh, as in like it's important that people are building for Llama rather than for LLMs in general because that will determine what the standard is for sure.

嗯，你看，我认为这些模型编码了价值观和思考世界的方式。我们早期有一个有趣的经历：我们拿了一个早期版本的 Llama，把它翻译成了法语或其他语言。法国人的反馈是，这听起来像一个学了法语的美国人，不像一个法国人。我们问，什么意思？它法语说得不好吗？不，法语说得很好，只是它思考世界的方式有点美国化。所以我认为这些微妙的东西被嵌入了模型。随着时间的推移，模型越来越复杂，它们应该能够体现世界各地不同的价值观。所以这可能不是一个特别复杂的例子，但我认为它说明了问题。我们在测试一些模型时看到的一些东西，尤其是来自中国的模型，它们编码了某些价值观，而且不是简单的微调就能让它们变成你想要的样子。现在情况不同了，对吧？所以我认为语言模型，那些嵌入了世界模型的东西，有更多的价值观。推理模型，我觉得，推理也有价值观或思考方式，但推理模型的一个好处是它们在可验证的问题上训练。所以如果你的模型在做数学题，你需要担心文化偏见吗？可能不需要。一个其他地方构建的推理模型通过以一种阴险的方式解数学题来影响你的可能性很低。我认为围绕编码有一整套不同的问题，这是另一个可验证的领域。你需要担心有一天醒来，一个与某个政府有联系的模型在代码中嵌入了各种漏洞，然后与该政府相关的情报机构可以利用这些漏洞。所以在未来的某个版本中，如果你使用来自另一个国家的模型来保护或构建我们的许多系统，然后突然有一天你醒来，发现一切都以那个国家知道但你不知道的方式变得脆弱，或者某个漏洞在某个时刻被触发。这些都是真实的问题。

Well, look, I think these models encode values and ways of thinking about the world. We had this interesting experience early on where we took an early version of Llama and we translated it. I think it might have been into French or some other language. And the feedback that we got from French people was, this sounds like an American who learned to speak French. It doesn't sound like a French person. It's like, well, what do you mean? Does it not speak French? No, it speaks French fine. It's just the way that it thinks about the world seems slightly American. So I think there are these subtle things that get built into it. Over time, as the models get more sophisticated, they should be able to embody different value sets across the world. So maybe that's a not particularly sophisticated example, but I think it illustrates the point. Some of the stuff that we've seen in testing some of the models, especially coming out of China, is they sort of have certain values encoded in them, and it's not just a light fine-tune to get that to feel the way that you want. Now the stuff is different, right? So I think language models, something that has a kind of world model embedded into it, have more values. Reasoning I think is, I mean, there are kind of values or ways to think about reasoning, but one of the things that's nice about the reasoning models is they're trained on verifiable problems. So do you need to be worried about cultural bias if your model is doing math? Probably not. The chance that some reasoning model that was built elsewhere is going to incept you by solving a math problem in a way that's devious seems low. There's a whole set of different issues I think around coding, which is the other verifiable domain. You kind of need to be worried about waking up one day and does a model that has some tie to another government embed all kinds of different vulnerabilities in code that then the intelligence organizations associated with that government can go exploit. So in some future version where you have some model from another country that we're using to secure or build out a lot of our systems, and then all of a sudden you wake up and everything is just vulnerable in a way that that country knows about but you don't, or it turns on a vulnerability at some point. Those are real issues.

所以我们基本上发现的是，我对研究这个非常感兴趣，因为我认为开源的一个主要有趣之处在于蒸馏模型的能力。大多数人，主要价值不仅仅是拿一个现成的模型说，好的，Meta 构建了这个版本的 Llama，我要拿它来在我的应用中原样运行。不，如果你只是运行我们的东西，你的应用并没有做任何不同的事情。你至少会微调它，或者尝试将它蒸馏成不同的模型。当我们谈到像巨兽模型这样的东西时，它的全部价值在于能够将这种非常高的智能蒸馏成一个更小的模型，你实际上会想运行它。但这就是蒸馏的美妙之处。我认为这是自我们上次坐下来以来，在过去一年中真正成为一种非常强大技术的东西之一。而且我认为它的效果比大多数人预测的要好。你基本上可以拿一个更大的模型，提取它大约 90% 或 95% 的智能，然后在只有其 10% 大小的模型中运行它。你能得到 100% 的智能吗？不能。但以 10% 的成本获得 95% 的智能，对很多事情来说已经非常好了。另一个有趣的事情是，现在有了更多样化的开源社区，不仅仅是 Llama，你还有其他模型，你可以从多个来源进行蒸馏。

So what we've basically found is, I'm very interested in studying this because I think one of the main things that's interesting about open source is the ability to distill models. Most people, the primary value isn't just taking a model off the shelf and saying, okay, Meta built this version of Llama, I'm going to take it and run it exactly in my application. No, your application isn't doing anything different if you're just running our thing. You're at least going to fine-tune it or try to distill it into a different model. And when we get to stuff like the behemoth model, the whole value in that is being able to basically take this very high amount of intelligence and distill it down into a smaller model that you're actually going to want to run. But this is the beauty of distillation. It's one of the things that I think has really emerged as a very powerful technique in the last year since the last time we sat down. And I think it's worked better than most people would predict. You can basically take a model that is much bigger and take probably 90 or 95% of its intelligence and run it in something that's 10% the size. Now do you get 100% of the intelligence? No. But 95% of the intelligence at 10% of the cost is pretty good for a lot of things. The other thing that's interesting is now with this more varied open-source community where it's not just Llama, you have other models, you have the ability to distill from multiple sources.

所以现在你基本上可以说，好吧，Llama 在这方面真的很强，也许架构本身就很好，因为它本质上是多模态的，而且本质上更有利于推理、更高效。但假设另一个模型在编程上更好。没问题，你可以从两者中蒸馏，然后针对自己的用例构建一个比两者都更好的东西。这很酷，但你需要解决安全问题，确保蒸馏过程安全可靠。这是我们一直在研究并投入大量时间的事情。我们基本得出的结论是：任何涉及语言的东西都相当棘手，因为其中嵌入了很多价值观。所以除非你不在乎从模型中获得的价值观，否则你可能不想直接蒸馏一个纯粹的语言世界模型。在推理方面，我认为可以通过限制在可验证领域、运行代码清洁度和安全过滤器（比如我们开源的 Llama Guard 或 Code Shield）来取得很大进展，这些工具可以让你将不同的输入整合到模型中，并确保输入和输出都是安全的。然后还需要大量的红队测试，让专家来审视：这个模型在蒸馏后有没有做我不希望它做的事？我认为结合这些技术，你很可能在可验证领域安全地进行推理蒸馏。我对此相当有信心，我们也做了大量研究。但这是一个非常大的问题：如何做好蒸馏？因为其中蕴含着巨大的价值，但同时……

So now you can basically say, okay, Llama's really good at this, like maybe the architecture is really good because it's fundamentally multimodal and fundamentally more inference friendly and more efficient. But like let's say this other model is better at coding. Okay, well just you can distill from both of them and then build something that's better than either of them for your own use case. So that's cool, but you do need to solve the security problem of knowing that you can distill it in a way that is safe and secure. And so this is something that we've been researching and have put a lot of time into. And what we've basically come to is like look, anything that's kind of like language is quite fraught because there's a lot of values embedded in that. So unless you don't care about having the values from whatever the model is that you got, you probably don't want to distill a straight like language world model. On reasoning, I think you can get a lot of the way there by limiting it to verifiable domains, running kind of code cleanliness and security filters like whether it's the Llama Guard open source or the Code Shield open source things that we've done that basically allow you to incorporate different input into your models and make sure that both the input and the output are secure, and then just a lot of red teaming to make sure that you have people who are experts who are looking at this, like is this model doing anything that isn't what I want after distilling from something? And I think with a combination of those techniques, you can probably distill on the reasoning side for verifiable domains quite securely. That's something I'm pretty confident about and it's something that we've done a lot of research around. But I think this is a very big question: how do you do good distillation? Because there's just so much value to be unlocked, but at the same time...

你是否认为不同模型存在某种根本性的偏见？说到价值释放，你认为 AI 的正确变现方式是什么？因为数字广告显然利润丰厚，但相对于总 GDP，它占比很小，远不及所有远程工作。即使你能提高生产力而不取代工作，那也价值数十万亿美元。那么，广告可能不是出路吗？你怎么看？

Do you just think that there is some fundamental bias in the different models? Speaking of value to be unlocked, what do you think the right way to monetize AI will be? Because obviously digital ads are quite lucrative, but as a fraction of total GDP, it's small in comparison to like all remote work. Even if you can increase productivity and not replace work, that's still worth tens of trillions of dollars. So, is it possible that ads might not be the way? How do you think about this?

就像我们之前讨论的，会有各种各样的应用，不同的应用倾向于不同的模式。广告在你想提供免费服务时很棒，对吧？因为是免费的，你需要某种方式覆盖成本。广告解决了这个问题：用户无需付费就能获得很棒的东西。而且，顺便说一句，在现代广告系统中，很多时候人们认为如果做得好，广告本身就能增加价值。你需要擅长排序，也需要有足够的广告库存流动性。这样，如果系统中只有五个广告商，无论你排序多好，可能都无法向用户展示他们感兴趣的内容。但如果有一百万个广告商，并且你擅长从海量信息中找出用户可能感兴趣的，那么你很可能找到非常吸引人的内容。所以我认为广告肯定有它的位置，但显然也会有其他商业模式，包括那些成本更高、以至于免费提供不合理的模式。顺便说一句，这种商业模式一直存在。社交媒体免费且由广告支持是有原因的，但如果你想看 Netflix 或 ESPN，就需要付费。这没问题，因为其中的内容需要制作，制作成本很高，而且服务中可能无法插入足够多的广告来覆盖内容制作成本。所以基本上，你需要付费才能访问。代价是使用的人更少，对吧？是数亿人而不是数十亿人。这里存在一个价值转换。我认为类似地，不是每个人都想要一个软件工程师或一千个软件工程智能体，但如果你需要，你可能愿意支付数千、数万甚至数十万美元。所以我认为这恰恰说明了需要创造的事物的多样性：在频谱的每个点上都会有商业模式。在 Meta，对于消费者部分，我们肯定希望提供免费产品，我相信最终会由广告支持。但我也认为我们需要一种商业模式，支持人们使用任意数量的算力去做比免费服务所能提供的更令人惊叹的事情，为此我们最终会推出高级服务。但我认为我们的基本价值观是服务尽可能多的人。

I mean like we were talking about before, there's going to be all these different applications and different applications tend towards different things. Ads is great when you want to offer people a free service, right? Because it's free. You need to cover it somehow. Ads solves this problem of a person does not need to pay for something and they can get something that is amazing for free. And also, by the way, with modern ad systems, a lot of the time people think that the ads add value to the thing if you do it well, right? You need to be good at ranking and you need to be good at having enough liquidity of advertising inventory. So that way, if you only have five advertisers in the system, no matter how good you are at ranking, you may not be able to show something to someone that they're interested in. But if you have a million advertisers in the system, then you're probably going to be able to find something pretty compelling if you're good at picking out the different needles in the haystack that that person's going to be interested in. So, I think that definitely has its place, but there are also clearly going to be other business models as well, including ones that just have higher costs, so it doesn't even make sense to offer them for free. Which by the way, there have always been business models like this. There's a reason why social media is free and ad supported, but then if you want to watch Netflix or like ESPN or something, you need to pay for that. It's okay because the content that's going into that needs to be produced and it's very expensive for them to produce and they probably could not have enough ads in the service in order to make up for the cost of producing the content. So basically, you just need to pay to access it. Then the trade-off is fewer people do it, right? They're talking about hundreds of millions of people using those instead of billions. So, there's kind of a value switch there. I think similar here, not everyone is going to want like a software engineer or a thousand software engineering agents or whatever it is, but if you do, that's something that you are probably going to be willing to pay thousands or tens of thousands or hundreds of thousands of dollars for. So, I think that this just speaks to the diversity of different things that need to get created: there are going to be business models at each point along the spectrum. And at Meta, for the consumer piece, we definitely want to have a free thing and I'm sure that will end up being ad supported. But I also think we're going to want to have a business model that supports people using arbitrary amounts of compute to do really even more amazing things than what it would make sense to be able to offer with a free service, and for that I'm sure we'll end up having a premium service. But I think our basic values on this are we want to serve as many people in the world.

Lambda 是面向 AI 开发者的云服务。他们拥有超过 50,000 块 Nvidia GPU，可供初创公司、企业和超大规模用户使用。不过算力似乎是一种商品，那为什么要选择 Lambda 而不是其他厂商呢？与其他云提供商不同，Lambda 只专注于 AI。这意味着他们的 GPU 实例和按需集群预装了 AI 开发者所需的所有工具，无需手动安装 CUDA 驱动或管理 Kubernetes。如果你只需要 GPU 算力，可以省去通用云架构的开销，从而节省大量资金。Lambda 甚至提供合同，允许企业使用其产品组合中的任何类型的 GPU，并轻松升级到下一代。对于所有想用 Llama 4 构建的人，Lambda 提供了无服务器 API，没有速率限制。它专为快速扩展而设计，用户可以将推理消耗提升 1000 倍，而无需申请配额或与任何人沟通。前往 lambda.ai/dark AI/DLCH 免费试用他们的推理 API，该 API 提供 DeepSeek 和 Llama 等最佳开源模型，价格行业最低。好了，回到 Zach。

Lambda is the cloud for AI developers. They have over 50,000 Nvidia GPUs ready to go for startups, enterprises, and hyperscalers. Compute seems like a commodity though, so why use Lambda over anybody else? Well, unlike other cloud providers, Lambda's only focus is AI. This means their GPU instances and on-demand clusters have all the tools that AI developers need pre-installed. No need to manually install CUDA drivers or manage Kubernetes. And if you only need GPU compute, you can save a ton of money by not paying for the overhead of general purpose cloud architectures. Lambda even has contracts that let enterprises use any type of GPU in their portfolio and easily upgrade to the next generation. For all of you wanting to build with Llama 4, Lambda has a serverless API without rate limits. It's built with rapid scaling in mind. Users have 1000x their inference consumption without ever having to apply for a quota or even speak to a human. Head to lambda.ai/dark AI/DLCH for a free trial of their inference API featuring the best open source models like DeepSeek and Llama at the lowest prices in the industry. All right, back to Zach.

你是如何跟进所有这些不同项目的？其中一些我们今天讨论过，我肯定还有很多我不知道的。

How do you keep track of all these different projects, some of which we've talked about today? I'm sure there's many I don't even know about.

作为统管一切的 CEO，从直接去 Llama 团队说“这是你应该用的超参数”，到只给出“去把 AI 做得更好”这样的指令，这中间有很大的跨度。项目众多，你如何思考自己最能发挥价值、并统筹所有这些事情的方式？

As the CEO overseeing everything, there's a big spectrum between going to the Llama team and saying 'here are the hyperparameters you should use' to just giving a mandate like 'go make the AI better.' There are many different projects. How do you think about the way in which you can best deliver your value add and oversee all these things?

嗯，我花大量时间做的事情之一，就是努力把优秀的人才招到团队里。这是其一。然后还有跨团队的事情，比如“你做了 Meta AI，想把它放进 WhatsApp 或 Instagram”，那我就需要让这些团队互相沟通。还有一大堆问题，比如：Meta AI 在 WhatsApp 里的对话线程，是应该像其他 WhatsApp 线程那样，还是应该像其他 AI 聊天体验那样？两者有不同的范式。所以我认为，关于这些东西如何融入我们正在做的一切，有很多有趣的问题需要回答。然后还有另一部分工作，基本上就是推动基础设施。如果你想搭建一个千兆瓦级的集群，首先这会对我们建设基础设施的方式产生很多影响。它对你与建设这些设施的各个州的互动方式有政治层面的影响。它对公司有财务层面的影响——世界上有很多经济不确定性。我们现在要不要加倍投入基础设施？如果要，那我们在公司内部要做出哪些其他取舍？这些是其他人很难真正做出的决策。然后我认为还有品味和质量的问题，就是什么时候一个东西足够好到可以发布？我总体上觉得我是公司在这方面把关的人，不过我们也有很多其他有良好品味的人，他们也是不同事情的过滤器。但没错，我认为这些基本上就是那些领域。不过我觉得 AI 很有趣，因为与我们做的其他一些事情相比，它更偏向研究和模型驱动，而不是真正的产品驱动。你不能先设计你想要的产品，然后试图构建模型来适配它。你真的需要先设计模型和你想要的能力，然后你会得到一些涌现特性。然后你就会说，“哦，因为结果是这样，你可以构建一些不同的东西。”而且我认为归根结底，人们想用最好的模型。所以部分原因在于，当我们谈论构建最个性化的 AI、最好的语音、最好的个性化，以及非常智能且低延迟的体验时，这些就是我们基本上需要设计整个系统去构建的东西。这就是为什么我们在做全双工语音，为什么我们在做个性化——既要能从你与 AI 的互动中很好地提取记忆，又要能接入所有其他 Meta 系统，也是为什么我们设计那些具有特定大小和延迟参数的模型。

Well, a lot of what I spend my time on is trying to get awesome people onto the teams. So there's that. And then there's stuff that cuts across teams, like 'you build Meta AI and you want to get it into WhatsApp or Instagram.' Then I need to get those teams to talk together. And then there are a bunch of questions like, do you want the thread for Meta AI in WhatsApp to feel like other WhatsApp threads, or do you want it to feel like other AI chat experiences? There are different idioms for those. So I think there are all these interesting questions that need to get answered around how this stuff basically fits into all of what we're doing. Then there's a whole other part of what we're doing, which is basically pushing on the infrastructure. If you want to stand up a gigawatt cluster, first of all that has a lot of implications for the way that we're doing infrastructure buildouts. It has sort of political implications for how you engage with the different states where you're building that stuff. It has financial implications for the company in terms of, there's a lot of economic uncertainty in the world. Do we go double down on infrastructure right now? And if so, what other trade-offs do we want to make around the company? Those are things that it's tough for other people to really make those kind of decisions. And then I think there's this question around taste and quality, which is when is something good enough that we want to ship it? I do feel like in general I'm the steward of that for the company, although we have a lot of other people who have good taste as well, who are also filters for different things. But yeah, I think those are basically the areas. But I think AI is interesting because more than some of the other stuff that we do, it is more research and model-led than really product-led. You can't just design the product that you want and then try to build the model to fit into it. You really need to design the model first and the capabilities that you want, and then you get some emergent properties. Then it's like, 'oh, you can build some different stuff because this turned out in this way.' And I think at the end of the day, people want to use the best model. So that's partially why, when we're talking about building the most personal AI, the best voice, the best personalization, and also a very smart experience with very low latency. Those are the things that we basically need to design the whole system to build, which is why we're working on full duplex voice, which is why we're working on the personalization to both have good memory extraction from your interaction with AI but also be able to plug into all the other Meta systems, and why we design the specific models that we design to have the kind of size and latency parameters that they do.

说到政治，有一种看法认为一些科技领袖一直在与特朗普结盟。你和其他人向他的就职活动捐了款，我们还和他同台过，而且我记得你和解了一起诉讼，结果他们拿到了 2500 万美元。我想知道这是怎么回事。这是与政府打交道的成本吗？最好的思考方式是什么？

Speaking of politics, there's been this perception that some tech leaders have been aligning with Trump. You and others have donated to his inaugural event, and we were on stage with him, and I think you settled a lawsuit which resulted in them getting $25 million. I wonder what's going on here. Is it the cost of doing business with an administration? What's the best way to think about this?

我的看法是，他是美国总统。作为一家美国公司，我们的默认态度应该是努力与任何执政者建立富有成效的关系。我会这样做。我们也曾尝试支持过之前的政府。我之前相当公开地表达过对前一届政府的一些不满，他们基本上不与我们或更广泛的商界互动，而我认为坦率地说，要在这些事情上取得进展，这种互动是必要的。如果没有对话，他们也不优先考虑做这些事情，我们就无法建立起所需的能量水平。所以，从根本上说，我认为很多人想写一个关于“人们往哪个方向走”的故事。我只是觉得，我们是在努力做出伟大的东西。我们想与人合作并建立富有成效的关系，这就是我的看法。我想大多数其他人也是这么看的。但显然我不能代表他们。

My view on this is, he's the president of the United States. Our default as an American company should be to try to have a productive relationship with whoever is running the government. I would do this. We've tried to offer to support previous administrations as well. I've been pretty public with some of my frustrations with the previous administration, how they basically did not engage with us or the business community more broadly, which I think frankly is going to be necessary to make progress on some of these things. Like we're not going to be able to build the level of energy we need if you don't have a dialogue and they're not prioritizing trying to do those things. So, fundamentally, I think a lot of people want to write the story about what direction are people going? I just think it's like we're trying to build great stuff. We want to work with and have a productive relationship with people, and that's sort of how I see it. And it is also how I would guess most others see it. But obviously I can't speak for them.

你曾公开谈到，你重新思考了过去在内容审核方面与政府互动和遵从政府的一些方式。你现在如何看待 AI 治理？因为如果 AI 像我们认为的那样强大，政府就会想介入。在这方面最富有成效的方法是什么，政府应该考虑什么？

You've spoken out about how you've rethought some of the ways in which you engage and defer to the government in terms of moderation stuff in the past. How are you thinking about AI governance? Because if AI is as powerful as we think it might be, the government will want to get involved. What is the most productive approach to take there, and what should the government be thinking about here?

是的，我想在过去我可能只是……我的意思是，我发表的大部分言论我认为都是在内容审核的背景下，对吧？过去十年在这方面是一段有趣的历程。关于在线内容审核，出现了一些新颖的问题。其中一些问题，我认为促成了富有成效的新系统的建立，比如我们的 AI 系统能够检测国家行为体试图干涉他国选举。我认为我们会继续建设这些东西，这总体上是积极的。我认为其他一些事情我们走了一些弯路。比如，我觉得事实核查这件事不如社区注释有效，因为它不是一个互联网规模的解决方案。没有足够的事实核查员，而且人们不信任那些特定的事实核查员。他们想要一个更强大的系统。所以我认为我们在社区注释上得出的结论是正确的。但我在这方面的观点更多是，我认为历史上我可能有点过于遵从媒体及其批评，或者遵从政府在他们实际上没有权威的事情上，仅仅因为他们是一个中心人物。

Yeah, I guess in the past I probably just... I mean most of the comments that I made I think were in the context of content moderation, right? Where it's been an interesting journey over the last 10 years on this. There have been novel questions raised about online content moderation. Some of those have led to, I think, productive new systems getting built, like our AI systems to be able to detect nation states trying to interfere in each other's elections. I think we'll continue building that stuff out, and that has been net positive. I think other stuff we went down some bad paths. Like I just think the fact-checking thing was not as effective as community notes because it's not an internet-scale solution. There weren't enough fact checkers, and people didn't trust the specific fact checkers. They wanted a more robust system. So I think where we got with community notes is the right one on that. But my point on this was more that I think historically I probably deferred a little bit too much to either the media and their critiques or the government on things that they did not really have authority over, but just as a central figure.

我认为我们曾试图构建系统，让自己不必做出所有内容审核决策。我猜过去 10 年的成长过程就是：好吧，我们是一家有影响力的公司。我们需要为自己必须做出的决定负责。我们应该听取人们的反馈，但不应过度依赖那些实际上没有决定权的人，因为最终是我们坐在这个位置上，需要为自己的决定负责。所以我认为这是一个成熟的过程，有些方面很痛苦，但我认为我们可能因此成为了一家更好的公司。

I think we tried to build systems where maybe we could not have to make all of the content moderation decisions ourselves. I guess part of the growth process over the last 10 years is just okay, we're a meaningful company. We need to own the decisions we need to make. We should listen to feedback from people but shouldn't defer too much to people who do not actually have authority over this, because at the end of the day we're in the seat and we need to own the decisions we make. So I think it's been a maturation process, in some ways painful, but I think we're probably a better company for it.

关税是否会增加在美国建设数据中心的成本，并将建设转移到欧洲和亚洲？

Will tariffs increase the cost of building data centers in the US and shift buildouts to Europe and Asia?

很难预测结果会如何。我认为我们可能还处于早期阶段，很难说。

It is really hard to know how that plays out. I think we're probably in the early innings on that, and it's very hard to know.

你每周最高杠杆率的一个小时是什么？你在那一小时里做什么？

What is your single highest leverage hour in a week? What are you doing in that hour?

我不知道。每周都有点不同。很可能你每周做的最具杠杆效应的事情并不相同，否则按定义你应该每周花超过一小时做那件事。但确实，我不知道。这也是这份工作的乐趣之一，而且行业如此动态，事情总是在变化。现在的世界与年初、与去年年中相比已经大不相同。我认为自我们上次坐下来谈话（大约一年前）以来，很多事情都有了显著进展，很多牌已经翻开了。

I don't know. Every week is a little bit different. It's probably the case that the most leveraged thing you do in a week is not the same thing each week, or else by definition you should probably spend more than one hour doing that thing every week. But yeah, I don't know. It's part of the fun of this job, but also the industry being so dynamic is that things really move around. The world is very different now than it was at the beginning of the year, than it was six months into the middle of last year. I think a lot has advanced meaningfully and a lot of cards have been turned over since the last time we sat down about a year ago.

你之前提到招聘人员是你做的超级高杠杆的事情。

You were saying earlier that recruiting people is a super high leverage thing you do.

这确实是高杠杆。

It's very high leverage.

如果软件生产力在两年内提高 100 倍，可能实现什么？我们能建造哪些现在无法建造的东西？

What would be possible if software productivity increased like 100x in two years? What kinds of things could we build that we can't build right now?

这是个有趣的问题。我认为这次对话的一个主题是，即将释放的创造力将是巨大的。如果回顾人类社会和经济在过去 100 到 150 年的整体轨迹，基本上是从人们主要从事农业、大部分精力用于养活自己，到这一比例越来越小。满足基本物质需求所消耗的人类精力比例越来越小，这带来了两个影响：一是更多人从事创造性和文化追求，二是总体上更多人花更少时间工作、更多时间在娱乐和文化上。我认为这几乎肯定会继续下去。这不是一两年内拥有一个超级强大的软件工程师会发生什么，而是随着时间的推移，如果每个人都拥有这些超人工具来创造大量不同的事物，你将看到令人难以置信的多样性。一部分将是解决难题，比如疾病或科学相关的问题，或者只是让生活更美好的技术。但我猜测，很多最终会变成文化、社交追求和娱乐。世界会变得更有趣、更奇怪、更古怪，就像过去 10 年互联网上的迷因一样。这增加了某种丰富性和深度，而且以有趣的方式实际上帮助你更好地与人连接。现在我整天都在网上找有趣的东西，然后发到群聊里给我在乎的人，他们也会觉得好笑。今天人们能够制作出表达非常细微特定文化想法的媒体，这很酷，我认为这会继续发展，并以多种方式推动社会进步，即使不是像治愈疾病那样的硬科学方式。从元社交媒体的世界观来看，我认为未来人们会花更多时间做这些事情，而且会变得更好，帮助你连接，因为它有助于表达不同的想法。世界会变得更复杂，但我们用来表达这些非常复杂事物的文化技术——比如一个非常有趣的短视频——会变得更好。

It's an interesting question. I think one theme of this conversation is that the amount of creativity that's going to be unlocked is going to be massive. If you look at the overall arc of human society and the economy over 100 or 150 years, it's basically people going from being primarily agrarian and most of human energy going towards just feeding ourselves to that becoming a smaller and smaller percent. The things that take care of our basic physical needs are a smaller and smaller percentage of human energy, which has led to two impacts. One is more people are doing creative and cultural pursuits, and two is that more people in general spend less time working and more time on entertainment and culture. I think that is almost certainly going to continue. This isn't a one to two year thing of what happens when you have a super powerful software engineer, but over time, if everyone is going to have these superhuman tools to be able to create a ton of different stuff, you're going to get this incredible diversity. Part of it is going to be solving hard problems like diseases or different things around science, or just different technology that makes our lives better. But I would guess that a lot of it is going to end up being cultural and social pursuit and entertainment. The world is going to get a lot funnier and weirder and quirkier in a way that the memes on the internet have over the last 10 years. That adds a certain richness and depth, and in funny ways it actually helps you connect better with people. Now I just find interesting stuff on the internet and send it in group chats to the people I care about who I think are going to find it funny. The media that people can produce today to express very nuanced specific cultural ideas is cool, and I think that'll continue to be built out and advance society in a bunch of ways even if it's not the hard science way of curing a disease. From the meta social media view of the world, I think people are going to spend a lot more time doing that stuff in the future, and it's going to be a lot better and help you connect because it helps express different ideas. The world is going to get more complicated, but our cultural technology to express these very complicated things in a very funny little clip is going to get so much better.

我倾向于认为，至少在可预见的未来，这将导致对人们工作的需求增加，而不是减少。人们可以选择花多少时间工作。我给你举一个有趣的例子。我们每天有近 35 亿人使用我们的服务。我们一直纠结的一个问题是如何提供客户支持。今天你可以写邮件，但我们从未认真考虑过提供语音支持，让人们可以打电话进来。这可能是免费服务的一个副产品。人均收入不够高，无法建立让人们打电话的经济模型。而且每天有 35 亿人使用你的服务，你会拥有大量人员，就像世界上最大的呼叫中心。每年要花 100 亿或 200 亿美元来配备人员，这太荒谬了。所以我们从未真正认真考虑过，因为一直觉得这根本行不通。

I tend to think that for at least the foreseeable future, this is going to lead towards more demand for people doing work, not less. People have a choice of how much time they want to spend working. Let me give you one interesting example. We have almost three and a half billion people use our services every day. One question we've struggled with forever is how do we provide customer support. Today you can write an email, but we've never seriously been able to contemplate having voice support where someone can just call in. That's maybe one of the artifacts of having a free service. The revenue per person is not so high that you can have an economic model where people can call in. Also with three and a half billion people using your service every day, you would have a massive number of people, like the biggest call center in the world. It would be 10 or 20 billion something ridiculous a year to staff that. So we've never really thought too seriously about it because it was always just no way that makes sense.

但随着 AI 越来越好，你会到达一个地方，AI 可以处理很多人的问题，但不是全部，对吧？因为也许 10 年后它能处理所有问题。但当我们考虑 3 到 5 年的时间范围时，它能处理很多。有点像自动驾驶汽车可以处理很多地形，但总的来说，在大多数情况下它们还不能完全自主完成全程，对吧？就像人们以为卡车司机的工作会消失一样。实际上，现在卡车司机的工作比我们大约 20 年前开始谈论自动驾驶汽车时还要多。

But now as the AI gets better, you're going to get to this place where the AI can handle a bunch of people's issues, not all of them, right? Because maybe 10 years from now or something it can handle all of them. But when we're thinking about a 3 to 5 year time horizon, it'll be able to handle a bunch. Kind of like self-driving cars can handle a bunch of terrain, but in general, they're not doing the whole route by themselves yet in most cases, right? It's like people thought truck driving jobs were going to go away. There's actually more truck driving jobs now than there were when we started talking about self-driving cars almost 20 years ago.

回到客户支持这件事，就像，好吧，我们不可能为每个人安排人工客服。但假设 AI 能处理 90% 的问题，如果它处理不了，就转给人工。现在，如果你把提供服务的成本降到原来的十分之一，那么也许这样做就合理了，那会很酷。所以，最终结果是我认为我们可能会雇佣更多的客服人员，对吧？就像人们普遍认为的那样，哦，这显然会自动化工作，所有工作都会消失。但实际上，技术的历史并不是这样运作的。你可以创造出消除 90% 工作量的东西，但这会让你想要更多的人，而不是更少。

And I think for going back to this customer support thing, it's like, all right, it wouldn't make sense for us to staff out calling for everyone. But let's say the AI can handle 90% of that, then if it can't handle it, it kicks it off to a person. Now, if you've gotten the cost of providing that service down to one-tenth of what it would have otherwise been, then maybe that actually makes sense to go do and that would be kind of cool. So, the net result is I actually think we're probably going to go hire more customer support people, right? It's like the common belief that people have is that, oh, this is clearly just going to automate jobs and all these jobs are going to go away. I actually just that has not really been how the history of technology has worked. It's been you can create things that take away 90% of the work and that leads you to want more people not less.

最后一个问题。当今世界上，你最常向谁寻求建议？

Final question. Who is the one person in the world today who you most seek out for advice?

哦，天哪。我觉得我的风格之一就是喜欢有广泛的顾问。所以，不只是一个人，但我们有一个很棒的团队。我的意思是，公司里有人，董事会里有人。行业里也有很多人在做新东西。没有某一个人。但这很有趣，而且当世界充满变化时，有一个理由和你喜欢的人一起做酷的事情。对我来说，这就是生活的意义。

Oh man. Well, I feel like it's part of my style is I like having a breadth of advisors. So, it's not just one person, but we've got a great team. I mean, there are people at the company, people on our board. And there's a lot of people in the industry who are doing new stuff. There's not a single person. But it's fun and also when the world is dynamic, just having a reason to work with people you like on cool stuff. To me, that's what life is about.

是的。好的，很好的结尾。太棒了。谢谢你接受采访。

Yep. All right, great note to close on. Awesome. Thanks for doing this.

嗯，谢谢。

Yeah, thank you.

希望你喜欢这一集。如果喜欢，最有帮助的事情就是把它分享给你认为可能喜欢的人。发给你的朋友、群聊、Twitter 或其他地方。让消息传开。除此之外，如果你能在 YouTube 上订阅并在 Apple Podcast 和 Spotify 上留下五星好评，那会非常有帮助。查看下方描述中的赞助商。如果你想赞助未来的节目，请访问 dwarcash.com/advertise。感谢收听。下期再见。

I hope you enjoyed this episode. If you did, the most helpful thing you can do is just share it with other people who you think might enjoy it. Send it to your friends, your group chats, Twitter, wherever else. Just let the word go forth. Other than that, super helpful if you can subscribe on YouTube and leave a five-star review on Apple Podcast and Spotify. Check out the sponsors in the description below. If you want to sponsor a future episode, go to dwarcash.com/advertise. Thank you for tuning in. I'll see you on the next one.