AI 的加速：从预训练到递归自我改进

还有像亚历克斯这样的创新者，他们处于这项技术的前沿，我认为他们需要参与其中，尤其是因为像你们这样的企业有资源进行投资。简单说一下背景，今年大型 AI 公司在开发这项技术上的投入将是美国曼哈顿计划的 20 倍。你能想到政府动员起来做某件事的规模是今年的 20 倍的关键时刻吗？所以我认为，我们必须以不同的方式来做这件事，不像 G7 和 G20 那样。我们把产业界、创新者和政府聚集在一起，这就是我们几年前在布莱切利园所做的。峰会系列去了巴黎、韩国，现在又精彩地来到了印度，一个拥有巨大数字雄心和能力的国家。但我想在这些年里，亚历克斯，让我们从这里开始。我认为实际上关于 AI 的辩论已经从技术转向了战略。从这些工具能做什么，到政府选择用它们做什么。我想也许就从这里开始。如果你回顾过去几年，你是处于这项技术前沿的人。你知道我们过去几年走过的历程。帮助我们理解我们走到了哪里，此刻什么令人兴奋，让我们了解现状，这样我们就可以继续讨论关于我们如何利用它的一些问题。

And innovators like Alex himself who were at the frontier of developing this technology, and I thought they needed to be in the room, not least because it's businesses like yours that have the resources to invest in this. Just to give you a bit of context, this year the large AI companies are going to spend 20 times more on developing this technology than the US did for the Manhattan Project. Can you think about seminal moments where the government mobilized to do something 20 times more just this year alone? So I thought, look, it's vital that we do this in a different way, not like the G7 and G20. We bring industry, innovators, and government together, and that's what we did at Bletchley Park a few years ago. The summit series went to Paris, South Korea, and now brilliantly in India, a country with huge digital ambitions and indeed capacity. But I think over those years, and Alex, let's start with this. I think actually the debate on AI has shifted from kind of technology to strategy. You know, from what these tools can do to what governments are choosing to do with them. And I wanted to maybe just start there. If you think about the last few years, you are someone who is at the frontier of developing this technology. You know the journey that we've been on in the last few years. Help us understand where have we come, what is it that is exciting at this moment, and just ground us in where we are so that we can get on to some of these questions about what we do with it.

是的，我认为我们现在正处于这项技术极其激动人心的时刻。我认为我们正处于技术真正加速的开端，从很多方面来看，直到现在这种加速都非常明显，但我认为 2026 年将在很多方面成为一个转折点。为了说明过去几年的情况，我认为从 2024 年之前，实际上是一个长达 6 年的时期，我们处于预训练时代。我们沿着一条非常大的指数曲线来提高模型的性能。这条曲线非常可预测，具有明确的特征：投入的资源越多，获得的结果就越多。当时有很多关于这条曲线是否会继续产生回报的问题，我认为这是最近一个 AI 怀疑论的时代。然后在 2024 年，接近 2024 年底，开始了强化学习时代。这是叠加在前一波浪潮之上的下一波主要浪潮，这时我们看到模型开始学习推理。一直有个梗叫‘伊利亚看到了什么？’嗯，伊利亚看到了模型的推理能力，这显然是技术的下一波巨浪。即使是强化学习也有一些局限性，关于这个时代会持续多久以及是否存在局限性的问题层出不穷。然后在 2025 年底，就在几个月前，我认为我们真正开始了一个全新的技术范式或时代，即递归自我改进时代。我认为我们真的开始在我们的模型开发中看到这一点。大多数主要实验室都在他们的模型开发中看到了这一点，即模型本身现在已成为加速生产下一代 AI 过程的关键工具。从外部看，你看到的是整体开发速度的加快，新模型发布速度的加快。在内部，我们看到的是戏剧性的提速。单个研究人员的生产力大幅增长，我们预计这不会停止。事实上，我们预计它会平稳地继续加速，随着技术的进步，单个研究人员、单个工程师的产出本身将继续呈指数级增长。与递归自我改进这一技术开发新范式同时出现的另一波巨浪，实际上是智能体的真正到来。老实说，从 2023 年开始，这个话题就备受关注。很长一段时间，它只是一个从未真正达到预期的炒作热词。我认为也是在 2025 年底，大约 2025 年中后期，我们开始看到智能体真正发挥作用。这些是模型可以真正开始采取行动、自动化整个工作流程，并作为独立实体而非助手大幅提高生产力的地方。这个时代始于编码智能体。我们开始看到它渗透到个人智能体。这是我们在 Meta 所坚信的重要组成部分。我认为在 2026 年，我们将在经济许多领域、世界许多地区看到大规模的智能体部署。可以说，AI 的 GDP 将呈指数级增长。

Yeah, I think we are sitting here right now at such an incredibly exciting moment for the technology. I think we are at the beginning of a true acceleration in the technology, and in many ways it's felt like this very clear acceleration even until now, but I think we're going to see 2026 as an inflection point in many ways. To contextualize this over the past few years, I think from through 2024 really, which is a good 6-year period, we were in the era of pre-training. So we were riding one very large exponential curve towards improving the performance of the models. And that curve was very predictable and had clear characteristics where the more resources you put in, the more results you would get out. There were a lot of questions at the time about whether or not this curve would continue returns, and I think this was one of the more recent eras of AI skepticism. Then in 2024, towards the end of 2024, began the era of reinforcement learning. So this was the next major wave that stacked on top of the previous wave, and this is when we saw the models begin to learn to reason. There's always this meme of 'what did Ilya see?' Well, Ilya saw reasoning from the models, and it was very clearly this next mega wave of the technology. Even reinforcement learning had some limitations, and there were all these questions around how long that era was going to last and maybe there were limitations. Then at the end of 2025, just a few months ago, I think we've really begun an entirely new paradigm or era of the technology, which is really beginning the era of recursive self-improvement. I think we're really starting to see this in the development of our models. Most of the major labs are seeing this in the development of their models, which is that the models themselves have now become instrumental in accelerating the process of producing the next AIs. From the outside, what you see is maybe an acceleration of overall development velocity, an acceleration in the speed at which new models are being released. Internally, what we see is just dramatic speed ups. The productivity of an individual researcher has grown dramatically, and we don't expect that to stop. In fact, we expect it to kind of smoothly continue accelerating, where the output of an individual researcher, an individual engineer, will itself continue growing exponentially as the technology improves. The other mega wave that is somewhat coincident with this new paradigm within the development of the technology of recursive self-improvement is really the arrival of agents in earnest. This is something that has been talked about since I honestly want to say 2023 with a lot of fanfare. For a long time, it was one of these hype buzzwords that never actually lived up to the expectations. I think also at the end of 2025, toward middle to end of 2025, we started to see agents actually work. These are where the models can actually start taking actions and start automating entire workflows and being dramatically more productive as actual sort of entities unto themselves versus assistants. This era started with coding agents. We're starting to see it really percolate through personal agents. This is a big part of what we believe in at Meta. I think over the course of 2026, we will see large-scale agent deployments in many areas of the economy, many areas of the world. The GDP of AI, so to speak, is going to grow exponentially.

嗯，这非常令人高兴。我认为你提到了重要的一点，因为我记得当我刚开始做这份工作时，对话都是关于这些缩放定律以及它们是否会继续。你提到了怀疑论。我的意思是，它们不仅继续了，而且继续带来了指数级增长。感觉过去几个月确实出现了加速。这就是你提到的。这也是 Dario 和 Demis 公开对我说过的，由递归自我改进驱动，模型本身现在能够帮助我们更快地开发新模型。所以，我们正处于一个非常激动人心但也有些焦虑的时刻，因为开发时间表在加速。所以，我想谈谈信任。我给你讲个故事。我有两个小女儿，当她们使用聊天机器人之类的东西时，她们总是会说‘请，某某某’和‘谢谢’，当她们写完请求或得到答案时。到处都是请和谢谢。我觉得这很有趣。我对她们说，‘看，你不需要那样做。那不是一个人。没关系的。’

Well, that is very welcome to hear. I think you hit on something that is important because I remember when I first started doing this in the job, the conversations were all around these scaling laws and would they continue? And you mentioned the skepticism. I mean, not only have they continued and continued to deliver this exponential growth. It feels very much like the last few months have seen an acceleration. That's what you touched on. It's what Dario and Demis have also said to me and publicly, driven by this recursive self-improvement that the models themselves are now able to help us develop the new models quicker. So, we're at this quite exciting but also a moment of some anxiety because of that accelerating development time scale. So, I wanted to just then talk about trust. Because I'll give you a story. I have two young girls, and as they're using their chatbots and things, they're always like, 'Please, so-and-so and so-and-so.' And 'Thank you.' when they finish writing their request or they get the answer back. There's lots of pleases and thank yous everywhere. I thought that was interesting. I said to them, 'Look, you don't need to do that. It's not a person. It doesn't matter.'

顺便说一句，当我们为此进行算力计算时，成本高昂。他们对我说：‘你知道吗，爸爸？如果 AI 统治世界，我们希望曾经对 AI 友善过。’我觉得这是个不错的保险策略。但这触及了这种焦虑。

And by the way, it costs a fortune when we have to do the compute for that. And they said to me, 'You know what, Daddy? If AI takes over the world, we want to have been kind to the AI.' And I thought it was a good insurance policy. But it touches on this anxiety.

非常明智。

Very wise.

是的。

Yeah.

也许是这样。在全球范围内，我们看到对 AI 的不同态度。在印度这样的国家，有巨大的乐观和信任。在西方国家，焦虑仍然是主要情绪。我认为缩小信心差距既是政策任务，也是技术任务。我想听听你的想法。你在一家面向消费者的公司，每天处理数十亿用户。你如何看待这个信任问题？我们如何在不信任的地方建立信任，为了那些能带来巨大好处的事物？

In that maybe. Across the world, we're seeing different attitudes towards AI. In countries like India, there's enormous optimism and trust. In Western countries, anxiety is still the dominant feeling. I think closing that confidence gap is as much a policy task as a technical one. I'd love to get your thoughts. You're at a B2C company dealing with billions of consumers every day. How do you think about this trust question? How do we get trust up in places where it is not for something that can do untold good?

这是最重要的问题之一。甚至可能比技术发展本身更能决定技术扩散的成功程度。我们对此深思熟虑，因为我们的愿景是部署个人智能体，它们比大多数人更了解你，了解你的健康、目标、人际关系、家人和朋友，并能帮助你更有效地生活，更健康，完成项目，建立事业，探索兴趣。要构建这样的产品，需要来自消费者、政府和所有组织的巨大信任。这是一个有趣的时刻，因为我们想要的技术形式如此亲密，与消费者生活的未来紧密相连。找到与治理机构、非营利组织和其他组织合作开发技术以赢得信任的方法非常重要。我们在 Meta 的 WhatsApp 上看到了这一点。WhatsApp 因其对隐私和信任的承诺而变得无处不在。我们常常为了信任而优化，而不是业务目标。AI 产品也需要类似的旅程。我们已经看到 AI 实验室之间关于其行为是否真正建立信任的各种争论。你在布莱切利园发起的关于全球安全的主题比以往任何时候都更重要。模型正在达到的能力让我们对国家安全和网络安全风险有所警惕。正确测试并构建系统以安全监控和部署模型至关重要。

This is one of the most important questions. Perhaps even more than the technology development, it will govern how successfully the technology diffuses. We think deeply about this because our vision is to deploy personal agents that know you better than most people, about your health, goals, relationships, family, and friends, and can help you live your life more effectively, be healthier, accomplish projects, build businesses, explore interests. To build that requires immense trust from consumers, governments, and all organizations. This is an interesting moment because the form of technology we want is so intimate and tied to the future of consumer life. Figuring out ways to develop the technology in partnership with governing bodies, nonprofits, and other organizations to garner trust is very important. We've seen this at Meta with WhatsApp. WhatsApp became incredibly ubiquitous due to its commitment to privacy and trust. Often, we optimized for trust over business objectives. A similar journey will be necessary for AI products. We've already seen various kerfuffles between AI labs about whether their actions truly build trust. The topic you kicked off at Bletchley around global safety is more important than ever. Models are achieving capabilities that cause pause regarding national security and cybersecurity risks. Properly testing and building systems to monitor and deploy models safely is paramount.

你和我多年来一直在谈论同一个话题。这对我们在布莱切利园至关重要：在英国建立 AI 安全研究所来做这项工作。存在一个差距，即政府缺乏评估风险的技术能力。我们英国如何能带头，并通过与你们这样的公司合作进行部署前测试来提供全球公共产品？这不是监管，但它给公民信心。人们不会采用他们不信任的技术。听你这么说，公司有责任设计能建立信任的 AI 产品，比如 WhatsApp 的隐私例子。政府也有责任进行技术评估。我们是否需要政府之间更多的合作以及与实验室的互动？目前运作得是否理想？

You and I have been talking about the same topic for years. That was key to us at Bletchley: establishing the AI Security Institute in the UK to do exactly that work. There was a gap where governments lacked technical capability to evaluate risks. How could we in the UK take a lead and provide a global public good by working with companies like yours to do pre-deployment testing? That's not regulation, but it gives citizens confidence. People don't adopt technology they don't trust. Hearing you, there's a role for companies to design AI products that engender trust, like WhatsApp's privacy example. And a role for governments to do technical evaluation. Do we need more cooperation between governments and interaction with labs? Is it working as well as it should?

互动运作良好，但我们总能更周到。特别是当我们进入 AI 进展急剧加速的阶段时，我们需要不断思考如何做得更好。信任辩论的胜负还将在公共部门展开。公民体验到医疗保健、更高效的政府服务、更快捷的政府办事效率。当你展示实际好处时，这场辩论就从抽象变为现实。

The interaction is working well, but we can always be more thoughtful. Especially as we enter a phase of dramatically increased AI progress, we need to constantly think about how to do better. Another area where the debate on trust will be won or lost is in the public sector. Citizens experience healthcare, more efficient government services, a state that is quicker at dealing with you. This debate goes from abstract to real when you demonstrate real benefits.

你是否了解公共部门中最明显的 AI 用例，这些用例应该让我们兴奋，并且政府应该考虑？

Do you have a sense of the most obvious use cases in the public sector for AI that we should be excited about and that governments should be thinking about?

是的。如果你把政府视为公民的服务提供者，那么就有巨大的机会更有效地提供服务。我们在印度的 WhatsApp 上已经看到了这一点。在某些邦，他们直接通过 WhatsApp 提供绝大多数政府服务。对于交通，一个国家实际上能够在印度使用 WhatsApp。

I do. If you think of governments as service providers to citizens, there's a huge opportunity to deliver services more effectively. We're already seeing that in India on top of WhatsApp. In certain states, they deliver the vast majority of government services directly through WhatsApp. And for transit, a nation actually able to use WhatsApp in India.

去年，某种程度上，未来令人兴奋的机会实际上是部署 AI，以极大地提高政府服务部署和提供给公民的便捷性和速度。我认为有一个我一直使用的术语：魔法象限。这实际上关乎如何重新构想政府？我们需要摆脱大规模官僚服务和组织，思考政府如何尽可能快地行动，利用技术和人类的结合，为国家或州的公民提供最快的公共服务。我认为这贯穿了公共服务、医疗保健，当然在安全领域也扮演着重要角色。所以我认为这是一个相当重要的问题。

Last year, in some ways, the future opportunity that is quite exciting is actually deployment of AI to greatly increase the ease and speed at which government services are deployed and provided to citizens. I think there's one term I've always used: the magic quadrant. It really is about how should governments be reimagined? We need to move away from large-scale bureaucratic services and organizations, and think about how the government can move as quickly as possible to deliver, using a combination of technology and humans, the fastest possible public services for citizens in a country or a state. I think this cuts through civil services, healthcare, and certainly has a large role to play in security. So I think it's quite a problem here.

是的，我想你被中国政府联系过。我需要另一个麦克风。我想接着你刚才说的：离任的好处之一是你有更多时间在办公桌前。我离任时，有人给了我一本杰弗里·丁教授写的很棒的书，叫《技术与大国》。书中他提出了一个关于技术扩散的有力论点，回顾了从印刷术开始的通用技术历史，认为你不必是发明技术的国家，也能成为从中受益最多的国家。印度已经认识到，技术领导力不仅仅取决于发明技术，还在于你如何有效地部署它。他们专注于大规模广泛采用，并得到几个关键属性的支持：非常深厚的 AI 人才库、数字公共基础设施如身份系统、UPI 支付基础设施，以及最近的 Ayushman Bharat 健康账户，这些提供了将应用分发给超过十亿人的基础设施轨道。他们还有支持这一点的民众。专注于这一点使印度通过追求部署和采用战略，将自己定位为领导者。但考虑到这一点，如果我们想将你关于敏捷政府为公民服务的愿景变为现实，政府必须做对哪些关键属性？

Yeah, I think you were contacted by the Chinese government. I need one more mic. I wanted to suggest I pick up on something you talked about: one of the benefits of leaving office is you have more time at your desk. When I left, someone gave me a great book by Professor Jeffrey Ding called Technology and Great Powers. In it, he makes a compelling argument about technology diffusion, looking at the history of general-purpose technologies going back to the printing press, arguing that you don't need to be the country that invents the technology to be the country that gets the most benefit from it. India has recognized that technology leadership doesn't depend just on inventing the technology; it's about how effectively you deploy it. They are focused on mass broad adoption, supported by a few critical attributes: a very deep AI talent pool, digital public infrastructure like ID, UPI payments infrastructure, and most recently Ayushman Bharat health accounts, which provide infrastructure rails to distribute apps to over a billion people. They also have a population supportive of that. Focusing on that has allowed India to position itself as a leader by pursuing a deployment and adoption strategy. But thinking about that, what are the critical attributes that government has to get right if we want to turn your aspiration into a reality about an agile government delivering for our citizens?

政府确实需要以不同的方式思考问题，并需要围绕部署和扩散构建架构，无论是技能、算力还是监管。但关键可能还包括数据，这是你比大多数人更了解的。所以也许我们从那里开始，然后扩展。如果我是一位政府领导人，就像在座许多人一样，我试图弄清楚我需要做对什么。正确交付这件事的基石是什么？让我们从数据开始，你的建议是什么，然后我们再稍微扩展一下。

Government does need to think about things differently and needs to build the architecture around deployment and diffusion, whether it's skills, compute, or regulation. But crucially probably also data, which is something you know better than most. So maybe if we start there and then broaden it out. If I'm sitting as a government leader, as many people in the audience are, I'm trying to figure out what I need to get right. What are the building blocks to deliver this thing properly? Let's start with data and what would your advice be there, and then we'll broaden it out a little bit.

数据非常重要。过去几年，许多人已经意识到这一点，当我们审视这些模型的力量并揭开帷幕时，发现这些模型的所有力量和美感实际上都是由数据赋予的。数据在某种程度上就像新石油，是一种对 AI 未来和这些模型发展至关重要的基础材料。政府应该思考他们的数据储备、数据资产是什么，并在许多方面制定战略，使他们对某些数据拥有主权，或者能够构建某些数据集，从而使国家表现最佳。最明显的两个领域是医疗保健和国家安全。许多国家已经在做，其他国家也需要思考国内医疗保健数据的有效机制，以便在此基础上构建 AI 模型和 AI 智能体，为公民提供更好的医疗服务。对于国家安全，这相当不言自明，但各国需要深入思考核心数据资产。然后，围绕数据构建，下一层是基础设施：电力、计算基础设施。我不认为每个国家都需要建造大型 AI 数据中心，但他们需要制定战略，随着时间的推移如何获得这些基础设施，无论是通过联盟还是与大型云提供商等企业实体的合作。然后是创新生态系统。我们正处于重大技术变革的时刻，政府思考如何构建支持这一变革的生态系统至关重要。印度是一个很好的正面案例，很大程度上归功于人才库。昨晚我与印度创始人和风险投资家共进晚餐，数据显示印度消费 AI 公司的数量比美国还多。所以这些生态系统的惊人发展有一些光辉的例子。最后是政府服务，我们之前讨论过。所以我们几乎有了一个行动手册的要素：你需要正确有效地使用数据，以及一定程度的算力。

Data is so important. Over the past few years, many people have realized this as we look at the power of these models and look behind the curtains, realizing that all the power and beauty of these models are actually imbued by data. Data is in some ways like the new oil, a base material incredibly important for the future of AI and the development of these models. Governments should think about what are their data reserves, their data assets, and in many ways build strategies where they have sovereignty over some set of data or are able to build certain pockets of datasets that enable the country to perform the best. Two areas that are clearest for this are healthcare and national security. Many countries already are, and other countries need to be thinking about very effective mechanisms for healthcare data within the country to enable AI models and AI agents built on top of that to deliver better health services to citizens. For national security, it's pretty self-explanatory, but countries need to think deeply about core data assets. Then, building around data, the next layer is infrastructure: power, computational infrastructure. I don't think every country needs to build large-scale AI data centers, but they need a strategy for how they will access that infrastructure over time, whether through alliances or partnerships with corporate entities like large cloud providers. Then comes the innovation ecosystem. We're in a moment of momentous technological change, and it's incredibly important for governments to think through how they can build ecosystems that support that. India is a great positive case study, in large part due to the talent pool. I was at a dinner with Indian founders and venture capitalists last night, and the stat was there are more consumer AI companies in India than in the United States. So there are shining examples of incredible development of these ecosystems. Lastly, government services, which we talked about previously. So we have the elements of almost a playbook emerging: you need the right data used effectively, some degree of compute.

正如我们讨论过的，要建立信任，需要有适当的监管。然后还有一些公共部门的使用案例以及创新生态系统。这些就是你可以开始将雄心转化为现实并对人们产生实际影响的要素。我继续这个话题，你在一个了不起的公司工作，也与其他企业合作。我和很多 CEO 聊过，他们思考如何在业务中部署 AI，以及这与政府应该做的事情之间的相似之处，让我印象深刻。我认为有明显的相似之处。我很感兴趣。我认为 CEO 们一直在思考 AI。几乎所有大公司的 CEO 都在花时间琢磨：我会被 AI 颠覆或取代吗？我如何利用它保持领先？在哪里用它来提高效率和增长？他们自上而下地推动项目。他们说：‘看，这不是一个只能由 IT 部门处理的话题。这必须由 CEO 推动，我们选择用例，我们承担责任，我们开始试点并快速扩大规模。’我认为我们在政府中还没有看到同样的转变，但我觉得这是一个很好的类比：这必须由最高层的人推动。归根结底，如果你是首相，你只能亲自推动少数几件事，而这真的必须是其中之一。你对 CEO 和政府如何推动这件事有什么想法或建议？要执行好需要什么？你在 Meta 自己也做过，所以你应该有体会。

You've got appropriate regulation as we've talked about to build trust. And then we've got some public sector use cases as well and an innovation ecosystem. So, these are the kind of ingredients where you can start to turn these ambitions into reality and actual impact on people. I just staying on that theme, you know, you work at an incredible company. You work with other enterprises. I talked to a lot of CEOs and I'm struck by how they're thinking about deploying AI in their businesses and thinking about the parallel between that and what governments should be doing. I think there is a clear parallel. I'm intrigued. I think CEOs are constantly thinking about AI. Almost any large company CEO is spending their time figuring out, am I going to get disrupted, disintermediated by AI, how do I use it to stay ahead, where do I use it to get more efficiency and more growth. They are driving projects from the top down. They are saying, 'Look, this is something that is not a topic that could just stay with the IT department. This is something that has to be driven by the CEO, where we pick the use cases, we have accountability, and we start piloting these and scaling them up quite quickly.' I think we haven't quite seen that same transition in governments yet, but I thought it's a good analogy: this has to be driven by the person at the top. At the end of the day, if you're a prime minister, you can only do a few things that you drive personally from the top, and this really has to be one of them. What are your thoughts or advice for both CEOs and governments about how to drive the machine? What does it take to get execution on this? You've had to do that at Meta yourself, so you'll have a sense from that.

是的，我认为是两件事的结合。首先，在最高层面，你需要非常明确地表示，组织将是一个 AI 优先的组织。这对每个组织来说含义略有不同，但我真正思考的方式是，组织的成败完全取决于它能否恰当地拥抱 AI。所以这是自上而下的部分。需要更精心地设计一个策略。老实说，你现在可能可以请 AI 帮你设计那个策略，它可能会做得相当不错。然后你需要一个自下而上的采用和自下而上的成功案例的基础。对我们 Meta 来说，例子是公司各地的工程师能够利用 AI 将生产力提高 10 到 100 倍。这些都是非常了不起的例子，突然间组织中的其他人环顾四周，意识到，哇，这真的是必须认真对待的事情。这不仅仅是 CEO 或高管们在谈论的、脱离现实的东西。它在基层非常真实，影响巨大。所以我认为你需要这种结合：有哪些成功案例让每个人都意识到这是一次多么大的转变，同时还要确保有一个持续非常明确的自上而下的方向，即这种转变是强制性的。

Yeah, I think it's a mix of two things. First, at the highest level, you need to make it very clear that the organization will be an AI-first organization. What that means for every organization is slightly different, but the way I truly think about it is that the success or failure of the organization entirely rests on its ability to properly embrace AI. So that's the top-down component. There's a strategy that must be designed more thoughtfully. Honestly, you can probably ask AI to help you design that strategy now, and it'll probably do a pretty good job. Then you need a bed of bottoms-up adoption and bottoms-up success stories. For us at Meta, the example of this is we have engineers throughout the company who are able to use AI to be 10 to 100 times more productive. Those are such incredible examples that all of a sudden everyone else in the organization looks around and realizes, wow, this is really something that must be taken very seriously. It's not just something the CEO or C-level individuals are talking about, detached from reality. It's very real on the ground in a very big way. So I think you need this mix: what are the success stories that cause everyone to realize how big of a shift it is, and then also ensure there's a continual very clear top-down direction that this transition is mandatory.

强制性的。非常有帮助的建议。我们时间快到了。所以最后一个要点来总结，你展望未来，在这项技术的发展中，你最兴奋的是什么？你提到了个人智能体和其他科学突破。这走向何方，政府现在应该做什么才能做好准备而不是追赶？

It's mandatory. Very helpful advice. We're almost out of time. So maybe the last point to close on is, you're looking ahead, what are you most excited about in the development of this technology? You've touched on personal agents, other scientific breakthroughs. Where is this going, and what should governments be doing now so that they are prepared and not playing catch-up?

我认为你可以依赖的是：技术发展将继续加速。这个递归自我改进的时代会让人觉得一切都在加速，技术将继续快速发展。我还认为技术的扩散将会加速。过去很多年一直有个问题：每个人都在谈论 AI，但 AI 在我的日常生活中在哪里？我认为这真的会加速。也许不会快 100 倍，但可能在接下来的很多年里每年快两到三倍。我真正兴奋的是 Meta 内部的个人超级智能这个概念：每个人如何拥有一个智能体来帮助他们完成他们极其兴奋的事情？它如何帮助他们探索激情、创业、发现、发明，实现自己生活的最大版本？我认为这甚至超越了人际超级智能的概念。我们在人类社会中已经看到，有时大事发生所需的只是让正确的人以正确的方式组织起来。你在公司、非营利组织、大型运动甚至政府重组时都能看到这一点。我认为 AI 有一个不可思议的机会，可以让人类以比今天更好的方式组织起来，突然间我们也许能够以社区、团体或新公司的形式组织起来，解决我们许多最大的挑战。这听起来有点高调，但我真的非常兴奋。

I think the things you can depend on: the piece of technology development will continue accelerating. This era of recursive self-improvement will feel like everything is speeding up, and the technology will continue developing very quickly. I also think the diffusion of the technology is going to accelerate. There's been this question for the past many years: everyone's talking about AI, but where is AI in my daily life? I think that really is going to accelerate. Maybe it won't be 100 times faster, but maybe it'll be two or three times faster every year for the next many years. The thing I'm really excited about is this idea of personal super intelligence within Meta: how does each person have an agent that helps them accomplish the things they're extremely excited about? How can that help them explore their passions, build a business, make a discovery, invent a thing, achieve the most maximal version of their own lives? I think where this is going even beyond is this concept of interpersonal super intelligence. We already see in human society that sometimes the thing necessary for big things to happen is just for the right set of people to be organized in the right way. You see this with companies, nonprofits, large movements, even with governments when they reorganize. I think there's an incredible opportunity for AI to enable humans to organize in ways that are even better than we do today, and all of a sudden maybe we'll be able to organize as communities or groups or new companies to solve many of our greatest challenges. It sounds a bit highfalutin, but it's something I'm genuinely quite excited about.

是的，我能看到，这是一个非常引人注目的愿景。看，这是一次精彩的对话。很明显，技术不仅没有停滞，反而在加速。很明显，讨论已经从仅仅谈论技术转向了，如你所说，它如何改变人们的生活并被部署。我的一个担忧，也许我们就此结束，是有些政治领袖，不是全部，认为 AI 将是明天的问题，而我认为他们需要认识到这是一个今天就要行动的问题。

Yeah, I can see that and it's a very compelling vision. Look, this has been a fascinating conversation. It's clear that the technology is not just keeping pace, it's accelerating. It's clear that the debate is moving from just talking about the technology to, as you said, how it's making a difference in people's lives and being deployed. One of my concerns, and maybe we'll finish on this, is that I think some, not all, but some political leaders think that AI is going to be tomorrow's issue when I think they need to recognize that it's an action this day issue.

这需要从——你知道的——AI 需要从一个专业话题转变为政府的核心责任。它需要具备同样的紧迫感和行动力，这样我们才能实现 Alex 所说的所有好处。但这是一次很棒的讨论。感谢各位观众的建议和提示，让我们了解了未来的方向。非常感谢。

And this needs to go from, you know, AI needs to go from being a kind of specialized topic to a centralized responsibility of government. And it needs to have that same urgency and action attached to it so we can realize all the benefits that Alex is talking about. But this has been a great discussion. Thanks for all your advice and tips everyone in the audience and giving us an insight into where things are going. Thanks very much. Thank you so much.

非常感谢我们的演讲嘉宾。之前，请让我们再次为 Rishi Sunak 和 Alexander Wang 鼓掌好吗？

Thank you so much to our speakers. Earlier, can we have another round of applause, please, for Rishi Sunak and Alexander Wang?

谢谢。

Thank you.

当我成为首相时，我对 AI 非常着迷，因为我在加州的经历。你在唐宁街就职后，最早参加的会议之一就是与外交事务团队的会面。他们坐下来和你谈——我看到观众中有人会记得这个——他们会梳理你接下来 12 个月的国际日程。他们想弄清楚你应该做什么。结果发现，基本上你能想到的每一件事都有一个国际峰会。大概有 150 个。而英国外交部，作为英国外交部，他们希望我参加所有会议。虽然我确实说过，作为英国首相，我认为我可以在联合国沙尘暴高级别会议上划清界限。那是我们同意保留分歧的地方。但我注意到，没有一个会议能把人们聚集在一起，讨论我认为我们这个时代最具变革性的技术。我认为我们都很幸运，生活在 AI 成为一种通用技术的时代。这种技术并不常见，当它们出现时，过去有机会，未来也将有机会改变我们经济、社会和生活的方方面面。这就是为什么我认为领导人需要定期聚在一起讨论这个问题。但我也认为需要以稍微不同的方式进行。与以往的技术周期不同，是像 Alex 的公司和像 Alex 这样的创新者处于开发这项技术的前沿。我认为他们需要参与其中。尤其是因为像你们这样的企业有资源进行投资。给你一点背景。今年，大型 AI 公司在开发这项技术上的支出将是美国曼哈顿计划的 20 倍。如果你想想那些政府动员起来做某事的开创性时刻，仅今年一年就多了 20 倍。所以我想，‘看，我们必须以不同的方式来做这件事，不像 G7 和 G20。我们把行业、创新者和政府聚集在一起。’这就是我们几年前在布莱切利园所做的。峰会系列去了巴黎、韩国，现在精彩地来到了印度，一个拥有巨大数字雄心和能力的国家。但我认为，这些年来，Alex 会从这一点开始，实际上关于 AI 的辩论已经从技术转向了战略。你知道，从这些工具能做什么，到政府选择用它们做什么。我想也许就从这里开始。如果你想想过去几年，你是处于开发这项技术前沿的人。你知道，我们过去几年的旅程，帮助我们理解这一点。我们走到了哪里？此刻令人兴奋的是什么？让我们立足于现在，这样我们就可以继续讨论关于如何利用它的一些问题。

When I became Prime Minister, it's been an exact AI obsessed because of my time in California. You get the job in Downing Street and one of the earliest meetings you have is with the Foreign Affairs team. And they sit down with you. I can see someone who will remember this in the audience. And they go over your international calendar for the next 12 months. And they want to kind of figure out, you know, what you should be doing. And it turns out there's an international summit for basically every single thing that you can imagine. Turns out there were about 150. And the British Foreign Office being the British Foreign Office, they wanted me to go to all of them. And although I did say I'm going to as a British Prime Minister, I think I can draw the line at the United Nations high-level meeting on sandstorms. That's where we agreed to disagree. But I noticed that there was no gathering to bring together people to talk about what I thought was the most transformative technology of our time. And I think we're all fortunate to be living at a time where AI is a general-purpose technology. These don't come along all that often and when they do, they had the opportunity in the past and will in the future to change every aspect of our economies, of our societies, of our lives. And that's why I thought it was necessary that leaders came together to discuss this on a regular basis. But, I also thought it needed to be done slightly differently. In contrast to previous technology cycles, it's companies like Alex's and innovators like Alex himself who were at the frontier of developing this technology. And I thought they needed to be in the room. Not least because it's businesses like yours that have the resources to invest in this. Just to give you a bit of context. This year, you know, the large AI companies are going to spend 20 times more on developing this technology than the US did for the Manhattan Project. If you think about seminal moments where the government mobilized to do something, 20 times more just this year alone. So, I thought, 'Look, it's vital that we do this in a different way, not like the G7 and G20. We bring industry, innovators, and government together.' And that's what we did at Bletchley Park a few years ago. The summit series went to Paris, South Korea, and now brilliantly in India, a country with huge digital ambitions and indeed capacity. But, I think over those years, and Alex will start with this, I think actually the debate on AI has shifted from kind of technology to strategy. You know, from what these tools can do to what governments are choosing to do with them. And I wanted to maybe just start there. If you think about the last few years, you are someone who is at the frontier of developing this technology. You know, the journey that we've been on in the last few years, you know, help us understand that. Where have we come? What is it that is exciting at this moment? And just ground us in where we are so that we can get on to some of these questions about what we do with it.

是的，我认为我们现在正处于一个对技术来说极其激动人心的时刻。我认为我们正处于技术真正加速的开端。从很多方面来看，直到现在这种加速都非常明显，但我认为 2026 年将在很多方面成为一个转折点。为了将过去几年置于背景中，从到 2024 年为止，这实际上是一个为期 6 年的时期，我们处于预训练时代。所以，我们沿着一条非常大的指数曲线前进，以提高模型的性能。那条曲线非常可预测，并且有明确的特点：投入的资源越多，得到的结果就越多。当时有很多问题关于这条曲线是否会继续。即使它继续了，边际收益是否会递减？我认为这是最近一个 AI 怀疑论的时代。然后在 2024 年，接近 2024 年底，开始了强化学习的时代。所以，这是叠加在前一波之上的下一波主要浪潮。这时我们看到模型开始学习推理。一直有一个梗：Ilya 看到了什么？嗯，Ilya 看到了模型的推理能力。这显然是技术的下一波巨浪。即使是强化学习也有一些局限性。当时有很多问题关于它何时会、那个时代会持续多久，也许那个时代存在局限性。然后在 2025 年底，也就是几个月前，我认为我们真正开始了一个全新的技术范式或时代，即递归自我改进的时代。我认为我们真的开始在模型的开发中看到这一点。我认为大多数主要实验室都在其模型的开发中看到了这一点，即模型本身现在已成为加速生产下一代 AI 过程的关键工具。所以，从外部看，你可能会看到整体开发速度的加快，新模型发布速度的加快。在内部，我们看到的是急剧的加速。我认为单个研究人员的生产力已经大幅增长。我们预计这不会停止。事实上，我们预计它会平稳地继续加速，单个研究人员、单个工程师的产出本身将随着技术的进步而继续呈指数级增长。与递归自我改进这一新范式大致同时出现的另一波巨浪是智能体的真正到来。这从 2023 年就开始被大肆讨论，老实说。

Yeah, I think we are sitting here right now at such an incredibly exciting moment for the technology. I think we are at the beginning of a true acceleration in the technology. And in many ways it's felt like this very clear acceleration even until now, but I think we're going to see 2026 as an inflection point in many ways. I think to contextualize this over the past few years, from through 2024 really, which is a good 6-year period, we were in the era of pre-training. So, we were riding one very large exponential curve towards improving the performance of the models. And that curve was very predictable and it sort of had clear characteristics where the more resources you put in, the more results you would get out. And there was a lot of questions at the time of whether or not this curve would continue. And even to the degree it did continue, were there diminishing marginal returns? And I think this was sort of one of the more recent eras of AI skepticism. And then in 2024, towards the end of 2024, began the era of reinforcement learning. So, this was the next major wave that stacked on top of the previous wave. And this is when we saw the models begin to learn to reason. And there's always this meme of what did Ilya see? Well, Ilya saw reasoning from the models. And it was very clearly this next mega wave of the technology. And even reinforcement learning had some limitations. And there were all these questions around when that was going to, how long that era was going to last and maybe there were limitations in that era. And then at the end of 2025, so just a few months ago, I think we really begun an entirely new paradigm or era of the technology, which is really beginning the era of recursive self-improvement. And I think we're really starting to see this in the development of our models. I think most of the major labs are seeing this in the development of their models, which is that the models themselves have now become instrumental in accelerating the process of producing the next AIs. And so, from the outside what you see is maybe an acceleration of overall development velocity, an acceleration in the speed at which the new models are being released. Internally, what we see is just dramatic speedups. I think the productivity of an individual researcher has grown dramatically. And we don't expect that to stop. In fact, we expect it to kind of smoothly continue accelerating where the output of an individual researcher, an individual engineer, well, I think that itself will continue growing exponentially as the technology improves. And the other mega wave that is somewhat coincident with this new paradigm within the development of the technology of recursive self-improvement is really the arrival of agents in earnest. And this is something that has been talked about since I honestly want to say 2023 with a lot of fanfare.

在很长一段时间里，这只是一个从未兑现的炒作热词。但在 2025 年中后期，我们开始看到智能体真正发挥作用。这些模型能够采取行动、自动化整个工作流程，并且作为独立的实体而非助手，效率大幅提升。从编程智能体开始，现在正在渗透到个人智能体领域。这是我们在 Meta 坚信的重要方向。到 2026 年，我们将在经济和世界的许多领域看到大规模的智能体部署。可以说，AI 的 GDP 将呈指数级增长。

And for a long time, it was one of these hype buzzwords that never actually lived up to expectations. But toward the middle to end of 2025, we started to see agents actually work. These are models that can start taking actions, automate entire workflows, and be dramatically more productive as entities in themselves versus assistants. It started with coding agents, and we're seeing it percolate through personal agents. This is a big part of what we believe in at Meta. Over the course of 2026, we will see large-scale agent deployments in many areas of the economy and the world. The GDP of AI, so to speak, is going to grow exponentially.

听到这些非常令人振奋。你说到了关键点。我刚开始做这份工作时，大家都在讨论缩放定律以及它们是否还会持续。结果不仅持续带来了指数级增长，而且最近几个月还在加速。正如你提到的，Dario 和 Demis 也公开对我说过：通过递归自我改进，模型本身正在帮助我们更快地开发新模型。因此，我们处在一个既令人兴奋又令人焦虑的时刻，因为开发时间表在加速。我想谈谈信任问题。我有两个小女儿，她们用聊天机器人时总是说“请”和“谢谢”。我告诉她们没必要这样——那不是真人，而且还要消耗算力。她们说：“爸爸，如果 AI 统治世界，我们想对 AI 友善一点。”我觉得这是个不错的保险策略，但这触及了焦虑。

That is very welcome to hear. You hit on something important. When I first started in this job, conversations were all about scaling laws and whether they would continue. Not only have they continued to deliver exponential growth, but the last few months have seen an acceleration. It's what you touched on, and what Dario and Demis have also said to me publicly: driven by recursive self-improvement, the models themselves are now helping us develop new models faster. So we're at an exciting but also anxious moment because of that accelerating development timescale. I wanted to talk about trust. I have two young girls, and when they use chatbots, they always say please and thank you. I told them they don't need to do that—it's not a person, and it costs compute. They said, 'Daddy, if AI takes over the world, we want to have been kind to the AI.' I thought that was a good insurance policy, but it touches on this anxiety.

很明智。

Wise.

是的。世界各地对 AI 的态度各不相同。在印度这样的国家，人们非常乐观和信任。而在西方国家，焦虑仍然是主流情绪。缩小这种信心差距既是政策任务，也是技术任务。我很想听听你的想法。你在一家面向消费者的公司，每天服务数十亿用户。你如何看待这个信任问题？在信任度低的地方，我们如何提高对一项能带来巨大好处的技术的信任？

Yeah. Across the world, we see different attitudes toward AI. In countries like India, there's enormous optimism and trust. In Western countries, anxiety is still the dominant feeling. Closing that confidence gap is as much a policy task as a technical one. I'd love to get your thoughts. You're at a B2C company dealing with billions of consumers every day. How do you think about this trust question? How do we get trust up in places where it's low, for something that can do untold good?

这是最重要的问题之一，甚至可能比技术开发本身更重要。它将决定技术扩散的成功程度。我们对此深思熟虑，因为我们的愿景是部署个人智能体，它们比大多数人更了解你——了解你的健康、目标、人际关系、家人和朋友——并能帮助你更有效地生活、更健康、完成项目、创业和探索兴趣。要正确构建这一点，需要消费者、政府和所有组织的巨大信任。因为我们想要的技术形式非常私密，并且与消费者生活的未来紧密相连，所以找到与治理机构、非营利组织和其他组织合作开发技术以赢得信任的方法至关重要。我们在 Meta 的 WhatsApp 上看到了这一点。WhatsApp 之所以变得无处不在，很大程度上是因为它对隐私的承诺和我们建立的信任，我们常常为了信任而优化，而非商业目标。AI 产品也需要类似的历程。我们已经看到 AI 产品和实验室之间关于其行为是否真正建立信任的各种争论。你在布莱切利园发起的关于全球安全的议题比以往任何时候都更重要。模型正在达到的能力引发了关于国家安全和网络安全风险的担忧。正确测试并构建系统来安全监控和部署这些模型至关重要。

This is one of the most important questions, perhaps even more so than the technology development itself. It will govern how successfully the technology diffuses. We think deeply about this because our vision is to deploy personal agents that know you better than most people know you—about your health, goals, relationships, family, and friends—and can help you live your life more effectively, be healthier, accomplish projects, build businesses, and explore interests. To properly build that requires immense trust from consumers, governments, and all organizations. Because the form of the technology we want is so intimate and tied to the future of consumer life, figuring out ways to develop it in partnership with governing bodies, nonprofits, and other organizations to garner that trust is very important. We've seen this at Meta with WhatsApp. WhatsApp became ubiquitous in large part due to its commitment to privacy and the trust we built, often optimizing for trust over business objectives. A similar journey will be necessary for AI products. We've already seen various kerfuffles between AI products and labs about whether their actions truly build trust. The topic you kicked off at Bletchley around global safety is more important than ever. Models are achieving capabilities that cause pause around national security and cybersecurity risks. Properly testing and building systems to monitor and deploy these models safely is of paramount importance.

你我多年来一直在讨论同一个话题。这正是我们在布莱切利园的关键：在英国建立 AI 安全研究所来专门从事这项工作。当时的想法是，存在一个缺口，即政府缺乏评估这些风险的技术能力。那么，英国如何能够带头，同时在某种意义上提供全球公共产品，与贵公司和其他前沿实验室合作，对这些模型进行部署前测试？

And you and I have been talking about the same topic for some years. That was key to us at Bletchley: establishing the AI Security Institute in the UK to do exactly that work. The thought was there was a gap where governments lacked the technical capability to evaluate these risks. So how could the UK take a lead, but in a sense provide a global public good, working in partnership with companies like yours and other frontier labs to do pre-deployment testing on these models?

我一直认为，这样做虽然不算是监管，但能让公民对技术安全有信心，因为最终人们不会采用他们不信任的技术。我从你这里听到的是，企业有责任——我觉得 WhatsApp 的隐私例子很好——他们如何设计能建立信任的 AI 产品？然后，听起来你也同意，政府也有责任进行一些技术评估。你觉得我们是否需要政府之间加强合作，更多地与实验室互动？目前这方面运作得是否理想？

And my thought was always that by doing that, and that's not regulation, but it gives citizens confidence that the technology is safe, because ultimately people don't adopt a technology that they don't trust. And it's what I'm hearing from you: there's a role for the companies. I thought the privacy example with WhatsApp is a good one—how do they design AI products that will engender trust? And then there's a role, it sounds like you agree, for governments to do some of that technical evaluation. And is your sense that we need more cooperation between governments on that, more interaction with the labs? Is that working as well as it should at the moment?

我认为互动是良好的，但这也是一个我们总能更深入思考的领域，尤其是在我们进入 AI 进展急剧加速的阶段。我们需要不断思考如何做得更好。我认为信任之争的另一个关键战场是公共部门。如果你体验到更好的医疗、更高效的政府服务、一个能更快处理你事务的政府，那么这场辩论就会从抽象变得具体，有了实际展示益处的方式。

I think the interaction is working well, but I think this is one area where we always can be more thoughtful, and especially as we enter this phase of dramatically increased AI progress. I think we constantly need to be thinking about how we can do better. I think one of the other areas where this debate on trust will be won or lost is in the public sector. If you experience better healthcare, more efficient government services, a state that is quicker at dealing with you, I think this debate goes from abstract to having a real way to demonstrate benefit.

从你的角度看，公共部门中哪些 AI 用例或部署案例最明显、最令人兴奋，政府应该关注？

Do you have a sense from where you sit about where the most obvious use cases or deployment cases are in the public sector for AI that we should be excited about and that governments should be thinking about?

是的。首先，如果你把政府视为公民的服务提供者，那么所有政府都有巨大机会更有效地提供服务。我们在印度已经看到，在 WhatsApp 之上，某些州直接通过 WhatsApp 提供绝大多数政府服务。对于交通，一个国家能够……去年印度有 1 亿……这本身不是 AI，但未来令人兴奋的机会在于部署 AI，大幅提升这些政府服务的便捷性和速度。我一直喜欢一个概念：主动型政府。它真正关乎的是如何构想国家？我们需要摆脱大规模官僚服务机构，思考政府如何利用计算机和人员快速、及时地提供服务，让公民能够轻松验证和获取政府服务，从而引导国家开发解决方案。我认为这贯穿了公共服务、你提到的医疗，当然也在安全领域和政府内部发挥重要作用。所以我认为这是一个相当广泛的机遇。

I do. I think that one of the first, if you think of governments as service providers to their citizens, there's a huge opportunity for all governments to deliver their services more effectively. We're already seeing that in India on top of WhatsApp, where in certain states they deliver the vast majority of government services directly through WhatsApp. And for transit, a nation is able to... In India last year, 100 million... and that's not AI per se, but the future opportunity that is quite exciting is the deployment of AI to greatly increase the ease and speed at which these government services are deployed and assistance provided. I think there's one term I've always used: proactive government. It really is about how should the state be imagined? We need to move away from large-scale bureaucratic services and support organizations and think about how the government can deliver quickly and promptly using computers and people, making it easy to verify and access government services as fast as possible, so that citizens can steer the state to develop solutions. I think this cuts through civil services, healthcare as you mentioned, and certainly has a large role to play in security and within the government. So I think it's quite a broad opportunity.

我想你刚才提到了喷气技术这类概念……我还想再提一点。离任的好处之一是你有更多时间。我离任时，有人送了我一本杰弗里·丁教授写的《技术与大国》，我相信你知道这本书。书中他提出了一个关于技术扩散的有力论点，回顾了从印刷术开始的通用技术历史，认为你不需要是发明技术的国家，也能成为从中受益最多的国家。你提到了印度，印度已经认识到技术领导力不仅仅取决于发明技术，更在于如何有效部署。他们专注于大规模广泛采用，并得到几个关键属性的支持：深厚的人才库、分发渠道——数字公共基础设施，如身份系统、UPI、支付基础设施，以及最近的健康账户——这些基础设施渠道将应用分发给超过十亿人。他们还有支持这一点的民众。专注于这一点，使印度通过部署和采用战略定位为领导者。但考虑到这一点，如果我们想把你关于智能体式政府、为公民服务的愿景变为现实，政府必须做对哪些关键属性？政府需要以不同方式思考。它需要围绕部署和扩散构建架构，无论是技能、算力还是监管。但关键可能还包括数据，这是你比大多数人更了解的领域。所以也许我们从数据开始，然后扩展。如果我是一位政府领导人，就像在座许多人一样，试图弄清楚我需要做对什么，正确交付这件事的基石是什么？让我们从数据开始，你对此有什么建议？然后我们再扩展一下。

I think you were talking about the concept of jet technology and that sort... I need one more like... I want to suggest something. One of the benefits of leaving office is you have a lot more time. When I left, someone gave me a great book by a guy called Jeffrey Ding, Professor Ding, 'Technology and Great Powers'. I'm sure you know the book. In it, he makes a compelling argument about technology diffusion, looking at the history of general-purpose technologies going all the way back to the printing press, and arguing that you don't need to be the country that invents the technology to be the country that gets the most benefit from it. You touched on India, and India has recognized that technology leadership does not depend on just inventing the technology; it's about how effectively you deploy it. They have focused on mass broad adoption, supported by a few critical attributes: a very deep talent pool, the rails of distribution—digital public infrastructure like ID, UPI, payments infrastructure, and most recently health accounts—providing infrastructure rails to distribute apps to over a billion people. They also have a population that's supportive of that. Focusing on that has allowed India to position itself as a leader by pursuing the deployment and adoption strategy. But thinking about that, what are those critical attributes that government has to get right if we want to turn your aspiration into a reality about agentic government, about delivering for our citizens? Government needs to think about things differently. It needs to build the architecture around deployment and diffusion, whether it's skills, compute, or regulation. But crucially probably also data, which is something you know better than most. So maybe if we start there and then broaden it out. If I'm sitting as a government leader, as many people in the audience are, trying to figure out what I need to get right, what are the building blocks to deliver this thing properly? Let's start with data and what would your advice be there, and then we'll broaden it out a little bit.

数据非常重要。我认为过去几年，许多人已经意识到这一点，当我们看到这些模型的力量并揭开帷幕，发现模型的所有力量和美感实际上都源自数据。如果你把数据视为新石油，是 AI 未来和模型发展的基础材料，那么政府应该思考它们的数据储备是什么，它们的数据资产是什么，并在很多方面制定战略，使其对某些数据拥有主权，或者能够构建某些数据集，从而使国家表现最佳。我认为最清晰的两个领域是医疗和国家安全。许多国家已经在做，其他国家也需要思考如何在国内建立有效的医疗数据机制，以便最终随着时间的推移，在此基础上构建 AI 模型和 AI 智能体，为公民提供更好的医疗服务。第二个是国家安全。

Data is so important. I think over the past few years many people have realized this as we look at the power of these models and look behind the curtains, realizing that all the power and beauty of these models are actually imbued by data. If you think about data as the new oil, the base material that is incredibly important for the future of AI and the development of these models, then governments should think about what are their data reserves, what are their data assets, and in many ways build strategies where they have sovereignty over some set of data or are able to build certain pockets of datasets that enable the country to perform the best. I think the two areas that are clearest for this are healthcare and national security. Many countries already are, and other countries need to be thinking about very effective mechanisms for healthcare data within the country to enable ultimately over time there to be AI models and AI agents built on top of that to deliver better health services to their citizens. And the second is national security.

我认为这在很多方面不言自明，但各国需要深入思考自己的核心数据资产是什么。然后，如果你围绕数据构建并看接下来的几层，下一层是基础设施。所以我们考虑电力、计算基础设施，以及这背后的故事。我不认为每个国家都需要在本国建设大规模 AI 数据中心，但他们确实需要有一个长期战略，来获得这种基础设施的访问权。无论是通过多种形式的联盟，还是与大型云提供商等企业实体合作，我认为这都是需要仔细考虑的重要因素。接下来是创新生态系统。我确实认为我们正处于一个重大技术变革的时刻，这对于我认为近乎无限的创业和经济机会至关重要，我认为政府和国家的责任是思考如何构建支持这些机会的生态系统。我认为印度实际上在很多方面都是一个很好的正面案例，很大程度上得益于其人才库。但昨晚我和一些印度创始人和风险投资家共进晚餐，有一个数据是，印度的消费类 AI 公司，消费类 AI 初创公司，比美国还多。所以确实有一些这样的杰出例子，展示了这些生态系统的惊人发展。然后，最后是政府服务，我们之前谈过这个。所以我们几乎有了一个正在形成的行动手册，对吧？你需要有正确的数据，并确保其有效利用。你需要一定程度的算力。你需要适当的监管，正如我们讨论过的，以建立信任。然后我们还有一些公共部门用例，以及一个创新生态系统。所以这些就是你可以开始将雄心转化为现实并对人们产生实际影响的要素。

And I think that is, you know, in many ways pretty self-explanatory, but I think countries need to be thinking about that quite deeply as to what are their sort of core data assets. Then, if you build around data and look at the next few layers, the next one is infrastructure. So, we think about the power, the computational infrastructure, and what does that story look like? I don't think every country needs to build large-scale AI data centers in their country, but they do need to have a strategy of over time how they are going to have access to that infrastructure. So whether this is in the form of alliances of various formats or partnerships even with corporate entities, the large cloud providers, I think that's an important element to think through. Then comes the innovation ecosystem. I do think that we're in a moment of momentous technological change, and it is incredibly important for what I think is near boundless entrepreneurial and economic opportunity, and I think it is the responsibility of governments and countries to think through how they can build ecosystems that support that. I think India is actually, in many ways, a great positive case study, in large part due to the talent pool. But I was at a dinner with a number of Indian founders and venture capitalists last night, and the stat was, I think, there are more consumer AI companies in India, consumer AI startups in India, than in the United States. So there really are some of these shining examples of incredible development of these ecosystems. And then, lastly, I think are the government services, which we talked about previously. So we've got the elements of almost a playbook emerging, right? You need to have the right data and make sure it's used effectively. You need some degree of compute. You've got appropriate regulation, as we've talked about, to build trust. And then we've got some public sector use cases, as well, and an innovation ecosystem. So these are the kind of ingredients where you can start to turn these ambitions into reality and actual impact on people.

我继续沿着这个主题，你在一个了不起的公司工作，你与其他企业合作。我和很多 CEO 交谈过，他们思考如何在业务中部署 AI 的方式让我印象深刻，并且我思考了这与政府应该做的事情之间的相似之处。我认为有明显的相似之处。我很感兴趣。我认为 CEO 们一直在思考 AI。几乎任何大公司的 CEO 都在花时间琢磨：我会被 AI 颠覆或去中介化吗？我如何利用它保持领先？我在哪里用它来提高效率和增长？他们正在自上而下地推动项目。他们说，看，这不是一个可以只交给 IT 部门的话题，对吧？这是必须由 CEO 推动的事情，我们选择用例，我们承担责任，我们开始试点并快速扩大规模。我的感觉是，我们在政府中还没有看到同样的转变，但我认为这是一个很好的类比：这必须由最高层的人推动。归根结底，如果你是总理，我相信托尼也会这么说，而且确实说过，你只能亲自推动少数几件事，而这真的必须是其中之一。你对 CEO 和政府有什么想法或建议，关于如何推动这个机器？要执行好这件事需要什么？我的意思是，你自己在 Meta 必须这样做，对吧？所以你有这方面的经验。

I just staying on that theme, you know, you work at an incredible company. You work with other enterprises. I talk to a lot of CEOs, and I'm struck by how they're thinking about deploying AI in their businesses, and thinking about the parallel between that and what governments should be doing. And I think there is a kind of clear parallel. I'm intrigued. I think CEOs are constantly thinking about AI. I think almost any large company CEO is spending their time figuring out, am I going to get disrupted, disintermediated by AI, how do I use it to stay ahead, where do I use it to get more efficiency and more growth. And they are driving projects from the top down. And they are saying, look, this is something that is not a topic that can just stay with the IT department, right? This is something that has to be driven by the CEO, where we pick the use cases, we have accountability, and we start piloting these and scaling them up quite quickly. My sense is we haven't quite seen that same transition in governments yet, but I thought it's a good analogy in that this has to be driven by the person at the top. At the end of the day, if you're a prime minister, and I'm sure Tony would say the same thing, and indeed has, you can only do a few things that you drive personally from the top, and this really has to be one of them. What are your kind of thoughts or advice for, I guess, both CEOs and governments about how to kind of drive the machine? What does it take to get execution on this? And I mean, you've had to do that at Meta yourself, right? So you'll have a sense from that.

是的，我认为这是两件事的结合。首先，在最高层面，你需要非常明确地表明，这个组织将是一个 AI 优先的组织。这对每个组织的含义略有不同，但我真正思考的方式是，组织的成功或失败实际上完全取决于其正确拥抱 AI 的能力。所以这是自上而下的部分。而且必须更精心地设计一个战略。实际上，老实说，你现在可能可以请 AI 帮你设计那个战略，它可能会做得相当不错。然后你需要一个自下而上的采用和成功案例的基础。对我们 Meta 来说，例子就是公司里那些能够利用 AI 将生产力提高 10 到 100 倍的工程师。这些例子如此惊人，以至于突然间组织中的其他人环顾四周，意识到：‘哇，这真的是必须非常认真对待的事情。这不仅仅是 CEO 或高管们在谈论的、脱离现实的东西。’而是‘不，它在基层非常真实，影响巨大。’所以我认为你需要这种混合：有哪些成功案例能让每个人都意识到这实际上是一个多么大的转变。同时还要确保有一个非常明确的自上而下的方向，即这种转变是强制性的。

Yeah, I think it's a mix of two things. I think first, at the highest level, you need to make it very clear that the organization will be an AI-first organization. And what that means for every organization is slightly different, but the way I truly think about it is that the success or failure of the organization actually entirely rests on its ability to properly embrace AI. So that's sort of the top-down component of it. And there's sort of a strategy that must be designed more thoughtfully. And actually, honestly, you can probably ask AI to help you design that strategy now, and it'll probably do a pretty good job. And then you need a bed of bottoms-up adoption and bottoms-up success stories. For us at Meta, the example of this are these engineers throughout the company who are able to use AI to be 10 to 100 times more productive. And those are just such incredible examples that all of a sudden everyone else in the organization looks around and realizes, 'Wow, this is really something that must be taken very seriously. It's not just something that the CEO or the C-level individuals are talking about and that they're so detached from reality.' It's like, 'No, it's very real on the ground in a very big way.' And so, I think you need this sort of mix of what are the success stories that cause everyone to realize how big of a shift it actually is. And then also ensure that there's a very clear top-down direction that this transition is mandatory.

非常有帮助的建议。我们快没时间了。所以最后一个收尾的点是，你展望未来，在这项技术的发展中你最兴奋的是什么？你提到了个人智能体、其他科学突破。它会走向何方，政府现在应该做什么才能做好准备而不是追赶？

Very helpful advice. Look, we're almost out of time. So maybe a last final point to close on is, you're looking ahead, what are you most excited about in the development of this technology? You've touched on personal agents, other scientific breakthroughs. Where is this going and what should governments be doing now so that they are prepared and not playing catch-up?

我认为你可以依赖的事情是：技术发展的步伐将继续加速。我认为这个递归自我改进的时代会让人感觉一切都在加速，而且技术将继续快速发展。我还认为技术的扩散将会加速。过去很多年一直有一个问题：每个人都在谈论 AI，但 AI 在我的日常生活中在哪里？我认为这真的会加速。也许不是快 100 倍，但在未来很多年里，可能每年都会快两到三倍。

I think the things that you can depend on: the pace of technology development will continue accelerating. I think this era of recursive self-improvement will feel like everything is speeding up and I think that the technology will continue developing very quickly. I also think that the diffusion of the technology is going to accelerate. I think there's been this question for the past many years, which is, everyone's talking about AI, but where is AI in my daily life? I think that really is going to accelerate. Maybe not 100 times faster, but maybe it'll be two or three times faster every year for the next many years.

而我认为真正让我感到兴奋的方向是，我们在 Meta 内部讨论的‘个人超级智能’概念——每个人如何拥有一个智能体，帮助他们完成自己极其热衷的事情。如何帮助他们探索激情、创业、发现、发明，实现自己人生的最大版本。我认为这甚至超越了‘人际超级智能’的概念，我们在人类社会中已经看到，有时促成大事的关键就是让合适的人以正确的方式组织起来。公司在这样做，非营利组织在这样做，大规模运动在这样做，甚至政府重组时也是如此。我认为 AI 有一个不可思议的机会，能让人类以比今天更好的方式组织起来。突然之间，我们或许能够以社区、团体、新公司或任何形式组织起来，解决许多最大的挑战。这听起来有点高调，但这是我真正感到兴奋的事情。

And I think the thing where I really am excited about this going is, we talk about this idea of personal superintelligence within Meta, which is how each person has an agent that helps them accomplish the things they are extremely excited about. How can they help them explore their passions, build a business, make a discovery, invent a thing, achieve the most maximal version of their own lives. And I think where this is going even beyond is this concept of interpersonal superintelligence, which is I think we already see in human society sometimes the thing that is necessary for big things to happen is just for the right set of people to be organized in the right way. You see this with companies, nonprofits, large movements, even with governments when they reorganize. And I think there is an incredible opportunity for AI to enable humans to organize in ways that are even better than we do today. And all of a sudden maybe we will be able to organize as communities or groups or new companies or who knows what it will look like to solve many of our greatest challenges. So it sounds a bit highfalutin, but it is something I am genuinely quite excited about.

不，我能理解。这是一个非常引人注目的愿景。但这是一场精彩的对话。很明显，技术不仅在跟上步伐，而且在加速。很明显，讨论已经从仅仅谈论技术转向了，正如你所说，它如何改变人们的生活并被部署。我的一个担忧，也许我们就此结束，是有些政治领袖认为 AI 是明天的问题，而我认为他们需要认识到这是一个今天就要行动的问题。AI 需要从一种专业话题转变为政府的集中责任。它需要同样的紧迫感和行动，这样我们才能实现 Alex 所说的所有好处。但这是一场很好的讨论。感谢你给所有观众的建议和提示，让我们了解了事情的进展。非常感谢。

No, I can see that. It is a very compelling vision. But this has been a fascinating conversation. It is clear that the technology is not just keeping pace, it is accelerating. It is clear that the debate is moving from just talking about the technology to, as you said, how it is making a difference in people's lives and being deployed. And one of my concerns, and maybe we will finish on this, is that I think some, not all, political leaders think that AI is going to be tomorrow's issue, when I think they need to recognize that it is an action this day issue. And AI needs to go from being a kind of specialized topic to a centralized responsibility of government. It needs to have that same urgency and action attached to it so we can realize all the benefits that Alex is talking about. But this has been a great discussion. Thanks for all your advice and tips for everyone in the audience and giving us an insight into how things are going. Thanks very much. Thank you so much.

非常感谢我们的演讲者。请再次为 Rishi Sunak 和 Alexander Wang 鼓掌。当我成为首相时，我痴迷于 AI，因为我在加州的经历。你在唐宁街任职，最早的一次会议是与外交事务团队。他们坐下来，和你一起回顾接下来 12 个月的国际日程。他们想弄清楚你应该做什么。结果发现，几乎每件事都有一个国际峰会。大约有 150 个。英国外交部，作为英国外交部，他们希望我参加所有会议。虽然我确实说过，作为英国首相，我认为我可以在联合国沙尘暴高级别会议上划清界限。这是我们同意保留分歧的地方。但我注意到，没有一场会议能把人们聚集在一起，讨论我认为我们这个时代最具变革性的技术。我认为我们都很幸运生活在 AI 成为一种通用技术的时代。这种技术并不常见，当它们出现时，过去和未来都有机会改变我们经济、社会和生活的方方面面。这就是为什么我认为领导人需要定期聚在一起讨论这个问题。但我也认为需要以稍微不同的方式进行。与之前的技术周期不同，像 Alex 的公司和 Alex 这样的创新者正处于开发这项技术的前沿。我认为他们需要在场。尤其是因为像你们这样的企业有资源进行投资。给你一些背景信息，今年，大型 AI 公司在开发这项技术上的支出将是美国曼哈顿计划的 20 倍。如果你想想政府动员起来做事的那些开创性时刻，仅今年一年就是 20 倍。所以我认为以不同的方式来做这件事至关重要，不像 G7 和 G20。我们把行业、创新者和政府聚集在一起，这就是我们几年前在布莱切利公园所做的。峰会系列去了巴黎、韩国，现在，精彩地，在印度，一个拥有巨大数字雄心和能力的国家。但我认为这些年来，关于 AI 的讨论已经从技术转向了战略。从这些工具能做什么，到政府选择用它们做什么。我想也许就从这里开始。如果你想想过去几年，你是处于这项技术开发前沿的人。过去几年的旅程，帮助我们理解这一点。我们走到了哪里？此刻令人兴奋的是什么？让我们立足于现状，这样我们才能继续讨论如何利用它的问题。

Thank you so much to our speakers. Earlier, can we have another round of applause, please, for Rishi Sunak and Alexander Wang. When I became Prime Minister, I am AI obsessed because of my time in California. You get the job in Downing Street and one of the earliest meetings you have is with the Foreign Affairs team. They sit down with you and go over your international calendar for the next 12 months. They want to figure out what you should be doing. It turns out there is an international summit for basically every single thing you can imagine. There were about 150. The British Foreign Office being the British Foreign Office, they wanted me to go to all of them. Although I did say as a British Prime Minister, I think I can draw the line at the United Nations high-level meeting on sandstorms. That is where we agreed to disagree. But I noticed that there was no gathering to bring together people to talk about what I thought was the most transformative technology of our time. I think we are all fortunate to be living at a time where AI is a general purpose technology. These do not come along all that often, and when they do, they have the opportunity in the past and will in the future to change every aspect of our economies, societies, and lives. That is why I thought it was necessary that leaders came together to discuss this on a regular basis. But I also thought it needed to be done slightly differently. In contrast to previous technology cycles, it is companies like Alex's and innovators like Alex himself who are at the frontier of developing this technology. I thought they needed to be in the room. Not least because it is businesses like yours that have the resources to invest in this. Just to give you a bit of context, this year, the large AI companies are going to spend 20 times more on developing this technology than the US did for the Manhattan Project. If you think about seminal moments where the government mobilized to do something, 20 times more just this year alone. So I thought it is vital that we do this in a different way, not like the G7 and G20. We bring industry, innovators, and government together, and that is what we did at Bletchley Park a few years ago. The summit series went to Paris, South Korea, and now, brilliantly, in India, a country with huge digital ambitions and capacity. But I think over those years, the debate on AI has shifted from technology to strategy. From what these tools can do to what governments are choosing to do with them. I wanted to maybe just start there. If you think about the last few years, you are someone who is at the frontier of developing this technology. The journey we have been on in the last few years, help us understand that. Where have we come? What is it that is exciting at this moment? Just ground us in where we are so we can get on to some of these questions about what we do with it.

是的，我认为我们现在正处在一个对技术来说极其激动人心的时刻。我认为我们正处于技术真正加速的开端。从很多方面来看，直到现在，这种加速都非常明显，但我认为 2026 年将在很多方面成为一个转折点。为了将过去几年放在背景中，从到 2024 年，这实际上是一个很好的 6 年时期，我们处于预训练时代。所以我们沿着一条非常大的指数曲线前进，以提高模型的性能。那条曲线非常可预测，具有明显的特征：投入的资源越多，得到的结果就越多。当时有很多问题，关于这条曲线是否会继续，即使它继续了，边际收益是否会递减？我认为这是最近一个 AI 怀疑论的时代。然后在 2024 年，接近 2024 年底，开始了强化学习的时代。

Yeah, I think we are sitting here right now at such an incredibly exciting moment for the technology. I think we are at the beginning of a true acceleration in the technology. In many ways, it has felt like this very clear acceleration even until now, but I think we are going to see 2026 as an inflection point in many ways. To contextualize this over the past few years, from through 2024 really, which is a good 6-year period, we were in the era of pre-training. So we were riding one very large exponential curve towards improving the performance of the models. That curve was very predictable and had clear characteristics where the more resources you put in, the more results you would get out. There were a lot of questions at the time about whether or not this curve would continue and even to the degree it did continue, were there diminishing marginal returns? I think this was one of the more recent eras of AI skepticism. Then in 2024, towards the end of 2024, began the era of reinforcement learning.

所以，这算是叠加在前一波浪潮之上的下一个重大浪潮，我们开始看到模型学会推理。你知道，一直有个梗叫‘伊利亚看到了什么？’嗯，伊利亚看到了模型的推理能力。这显然是技术的下一个巨大浪潮。即使强化学习也有一些局限性，关于那个时代能持续多久有很多疑问。然后在 2025 年底，就在几个月前，我认为我们真正开启了一个全新的范式：递归自我改进的时代。我们开始在我们的模型开发中看到这一点。大多数主要实验室都看到了：模型本身已成为加速下一代 AI 生产的关键。从外部看，你会看到整体开发速度加快，新模型发布的速度也在加快。在内部，我们看到了巨大的加速。单个研究人员的生产力大幅提升，而且我们预计这不会停止。事实上，我们预计它会持续加速，随着技术的进步，单个研究人员或工程师的产出将呈指数级增长。与这个新范式同时出现的另一个巨大浪潮是智能体的真正到来。这个话题从 2023 年就开始被大肆宣传，但很长一段时间里，它只是一个从未兑现的炒作热词。在 2025 年底，我们开始看到智能体真正发挥作用。这些模型可以开始采取行动，自动化整个工作流程，并作为独立的实体而非助手，变得极具生产力。这个时代始于编码智能体，现在开始渗透到个人智能体。这是我们在 Meta 所坚信的重要组成部分。在 2026 年，我们将在经济和世界的许多领域看到大规模的智能体部署。可以说，AI 的 GDP 将呈指数级增长。

So, this was sort of the next major wave that stacked on top of the previous wave, and this is when we saw the models begin to learn to reason. And you know, there's always this meme of 'what did Ilya see?' Well, Ilya saw reasoning from the models. And it was very clearly this next mega wave of the technology. Even reinforcement learning had some limitations, and there were all these questions around how long that era was going to last. Then at the end of 2025, just a few months ago, I think we really began an entirely new paradigm: the era of recursive self-improvement. We're starting to see this in the development of our models. Most major labs are seeing this: the models themselves have become instrumental in accelerating the process of producing the next AIs. From the outside, you see an acceleration of overall development velocity and the speed at which new models are released. Internally, we see dramatic speed-ups. The productivity of an individual researcher has grown dramatically, and we don't expect that to stop. In fact, we expect it to continue accelerating, where the output of an individual researcher or engineer will grow exponentially as the technology improves. The other mega wave coincident with this new paradigm is the arrival of agents in earnest. This has been talked about since 2023 with a lot of fanfare, but for a long time it was a hype buzzword that never lived up to expectations. At the end of 2025, we started to see agents actually work. These are models that can start taking actions, automating entire workflows, and being dramatically more productive as entities onto themselves versus assistants. This era started with coding agents, and we're starting to see it percolate through personal agents. This is a big part of what we believe in at Meta. Over the course of 2026, we will see large-scale agent deployments in many areas of the economy and the world. The GDP of AI, so to speak, is going to grow exponentially.

嗯，这真是令人欣慰。你提到了一个重要的问题。我刚开始做这份工作时，所有的讨论都围绕着缩放定律以及它们是否会持续。你提到了怀疑论。不仅它们持续了并带来了指数级增长，而且过去几个月感觉还加速了，这得益于递归自我改进。模型本身现在能够帮助我们更快地开发新模型。所以我们正处于一个既令人兴奋又焦虑的时刻，因为开发时间表在加速。我想谈谈信任。我给你讲个故事。我有两个小女儿。她们用聊天机器人时，总是说‘请’和‘谢谢’，写完请求或得到答案后都会这样。我觉得这很有趣。我告诉她们不需要这样做——那不是真人，而且计算成本很高。她们说：‘爸爸，如果 AI 接管了世界，我们想对 AI 友善一点。’我觉得这是个不错的保险策略。但这触及了这种焦虑。在世界各地，人们对 AI 的态度不同。在印度这样的国家，有巨大的乐观和信任。在西方国家，焦虑仍然是主导情绪。缩小这种信心差距既是政策任务，也是技术任务。我很想听听你的想法。你在一家面向消费者的公司，每天服务数十亿用户。你如何看待这个信任问题？我们如何在不信任的地方提高信任，为了一个能带来巨大好处的东西？

Well, that is very welcome to hear. You hit on something important. When I first started doing this job, conversations were all around scaling laws and whether they would continue. You mentioned the skepticism. Not only have they continued and delivered exponential growth, it feels like the last few months have seen an acceleration, driven by recursive self-improvement. The models themselves are now able to help us develop new models quicker. So we're at an exciting but also anxious moment because of that accelerating development timescale. I wanted to talk about trust. Let me tell you a story. I have two young girls. As they use their chatbots, they always say 'please' and 'thank you' when they finish writing a request or get an answer. I thought that was interesting. I told them they don't need to do that—it's not a person, and it costs a fortune for the compute. They said, 'Daddy, if AI takes over the world, we want to have been kind to the AI.' I thought that was a good insurance policy. But it touches on this anxiety. Across the world, we see different attitudes towards AI. In countries like India, there's enormous optimism and trust. In Western countries, anxiety is still the dominant feeling. Closing that confidence gap is as much a policy task as a technical one. I'd love to get your thoughts. You're at a B2C company dealing with billions of consumers every day. How do you think about this trust question? How do we get trust up in places where it is low, for something that can do untold good?

是的，这是最重要的问题之一。也许甚至比技术本身更重要，它可能决定技术能否成功普及。我们对此深思熟虑，因为我们对这项技术的愿景是部署个人智能体，它们比大多数人更了解你——了解你的健康、目标、人际关系、家人和朋友——并能帮助你更有效地生活，更健康，完成你从未有时间做的项目，建立你一直想要的事业，探索你的兴趣。要正确构建这一点，需要来自消费者、政府以及所有组织的巨大信任。这是一个有趣的时刻，因为我们想要的技术形式是如此私密，与消费者生活的未来以及人与技术关系的核心紧密相连。找到与治理机构、非营利组织和其他组织合作开发技术以赢得信任的方法非常重要。

Yeah, this is one of the most important questions. Perhaps even more so than the technology itself, it may govern how successfully the technology diffuses. We think deeply about this because our vision for the technology is to deploy personal agents that know you better than most people know you—about your health, goals, relationships, family, and friends—and can help you live your life more effectively, be healthier, accomplish projects you never had time for, build the business you always wanted, and explore your interests. To properly build that requires immense trust from consumers, governments, and effectively everyone. This is an interesting moment because the form of the technology we want to take is so intimate and deeply tied to the future of consumer life and the core of the relationship with technology. Figuring out ways to develop the technology in partnership with governing bodies, nonprofits, and other organizations to garner that trust is very important.

这其实是我们之前在 Meta 做 WhatsApp 时看到的情况。我们在后台也聊过，WhatsApp 之所以变得无处不在，成为全球许多人的实用工具，很大程度上是因为它对隐私的承诺以及我们围绕 WhatsApp 建立起来的信任。很多时候，这并不一定是为了优化我们的商业目标，而是为了优化 WhatsApp 作为平台的信任度。我认为现代 AI 产品也需要经历类似的历程。我们已经看到现有 AI 产品和实验室之间各种争论，关于他们的行为是否真正建立了这种信任。而你在布莱切利园以全球性方式提出的安全问题，比以往任何时候都更重要。我认为我们已经看到模型的能力达到了某种水平，让我们对监控到的风险——比如各种国家安全风险或网络安全风险——感到需要谨慎。因此，正确测试并构建合适的系统来安全地监控和部署这些模型，其重要性怎么强调都不为过。

This is something we've seen at Meta actually with WhatsApp. We were talking about this backstage, but WhatsApp has become incredibly ubiquitous and a real utility for much of the world, in large part due to its commitment to privacy and the trust we've built around WhatsApp as an organization. Oftentimes, it's not necessarily for the optimization of our business objectives, but for the optimization of trust in WhatsApp as a platform. I think a similar journey will be necessary for the AI products of the modern age. We've already seen various kerfuffles or jabs back and forth between existing AI products and labs about whether their actions are truly building these levels of trust. And then the topic you really kicked off at Bletchley in a global way around safety is more important than ever. I think we're already seeing that the models are achieving levels of capability that cause some pause around the risks we monitor for these models, such as various national security risks or cybersecurity risks. So the importance of properly testing and building the right systems to monitor and deploy these models safely is of paramount importance.

你和我多年来一直在讨论同一个话题。这正是我们在布莱切利园的关键目标：在英国建立 AI 安全研究所来专门做这项工作。当时的想法是，政府之间存在技术能力缺口，无法评估其中一些风险。那么我们英国如何能够既发挥领导作用，又在某种意义上提供全球公共产品，与像贵公司这样的企业以及其他机构合作呢？

You and I have been talking about the same topic for some years. And that was key to us at Bletchley: establishing the AI Security Institute in the UK to do exactly that work. The thought was there was a gap where the technical capability didn't exist amongst governments to evaluate some of these risks. So how could we in the UK hopefully take a lead but in a sense provide a global public good, working in partnership with companies like yours and others?