张亚勤｜未来，每个人、每个设备都将拥有智能体

2025年11月18-20日，以“青年之智，解码未来”为主题的2025全球青年领袖年度对话会在北京成功举办。本届对话会由全球化智库（CCG）主办、“国际青年领袖对话（GYLD）”项目秘书处协办，并得到北京市海淀区人才工作局和海淀区人民政府外事办公室支持。清华大学智能科学讲席教授、智能产业研究院院长张亚勤出席11月19日的开幕式，并发表题为《拥抱人工智能新浪潮：人工智能向善》 的主旨演讲。

Themed "Decoding the Future with Young Minds," the Global Young Leaders Dialogue Annual Forum 2025 was successfully held in Beijing from November 18 to 20. The event was hosted by the Center for China and Globalization (CCG), co-organized by the Secretariat of the Global Young Leaders Dialogue (GYLD) program, and supported by the Haidian District Human Resources Bureau and the Haidian District Foreign Affairs Office.At the Opening Ceremony on November 19, Dr. Ya-Qin Zhang, Chair Professor of AI Science and Dean of the Institute for AI Industry Research (AIR) at Tsinghua University, delivered a keynote speech titled “Embracing the New Wave of AI: AI for Better.”

以下附张亚勤博士演讲重点摘要。英文转录和中文译文均仅供参考，具体内容以视频演讲为准。

Below are selected highlights from Dr. Ya-Qin Zhang’s speech. The English transcription and Chinese translation are for reference only; please refer to the video speech for the accurate content.

Many thanks, Huiyao and Miao Lu for the very kind invitation. It's always exciting to interact with young leaders.

非常感谢辉耀和苗绿的热情邀请，与青年领袖交流总令人振奋。

I want to talk about some new exciting developments on AI, a subject I've been working on for over 30 years, from Microsoft to Baidu, and now at Tsinghua University. The theme is AI for better. I want to talk about some megatrends in technology, some state of the art in AI, comparisons between China, the US, and the rest of the world, and finally, future trends in the AI industry.

我想谈谈人工智能领域一些激动人心的新进展，这已是我深耕三十余年的课题——从微软到百度，再到如今的清华大学。我的主题是“人工智能向善”。我将谈及技术领域的宏观趋势、人工智能的前沿案例、中美及全球AI发展比较，并最后展望人工智能产业的未来走向。

Indeed, AI is the technology engine behind the Fourth Industrial Revolution. This new wave of AI is a convergence of digital AI, physical AI, and biological AI. This is the backdrop for why I started AIR, the Institute for AI Industry Research at Tsinghua University, about five years ago. It's very lucky that we are able to recruit some of the best minds in the world. Now we have over 20 professors and 400 top-notch PhD students, postdocs, and research scientists.

的确，人工智能是第四次工业革命背后的技术引擎。当前这一轮人工智能浪潮，本质上是数字AI、物理AI与生物AI的深度融合。正是在这样的时代背景下，我在大约五年前创立了清华大学智能产业研究院（AIR）。我们非常幸运能够招募全球最优秀的研究者。目前，研究院拥有20多位教授，以及400多名顶尖博士生、博士后和研究科学家。

I will use a few examples to illustrate some of the state of the art in digital AI, physical AI and biological AI.

接下来，我将通过几个具体案例，介绍数字AI、物理AI和生物AI的一些前沿案例。

First, our work on large language models. We work closely with leading companies like ByteDance, Alibaba, and DeepSeek. A major advance in large language models recently is reinforcement learning and reasoning models. Our professors and researchers developed DAPO with ByteDance, which improved reinforcement learning performance and efficiency by several times. With DeepSeek, we work on defining a better reward model for GRM. With Alibaba's Qwen, we work on agentic AI. One thing about our work is that we publish everything—our research, algorithms, code, and data.

首先是我们在大语言模型方面的工作。我们与字节跳动、阿里巴巴、深度求索等领先企业保持着紧密合作。近来，大语言模型的一个重要突破在于强化学习和推理模型方面。我们的教授和研究人员与字节跳动联合研发了DAPO算法，使强化学习的性能和效率提升了数倍。我们与深度求索合作，优化GRM通用奖励模型；并与阿里巴巴的千问团队共同推进AI智能体研究。值得一提的是，我们的研究成果、算法、代码和数据都全面公开发布。

The second example on physical AI is autonomous driving—driverless cars. When I was President of Baidu, I started a program called the Apollo. An example of this work is Apollo Go. Right now, we have the largest self-driving fleet in the world, particularly in Wuhan, with over 1,500 cars covering 17 million people across 3,000 square kilometers. This is the most challenging part of physical AI or embodied AI. A self-driving car is a special robot that accumulates all the critical pieces of AI. A critical advancement is the large language models in the past few years. Now we're able to generate data, cover some of the long-tail cases, do a lot of simulation, and conduct road tests. Our data, after 200 million kilometers [of driving], shows [autonomous driving] is 17 times safer than people. We haven't had a single casualty or major accident. This will be the foundation for embodied AI, and we will use the same foundation model for industrial, social, or home robots.

第二个例子来自物理AI领域的自动驾驶，也就是无人驾驶汽车。我在担任百度总裁期间，启动了“阿波罗”计划，其代表性应用便是“萝卜快跑”。目前，我们已经拥有全球规模最大的自动驾驶车队，尤其是在武汉，超过1500辆无人车覆盖约3000平方公里、服务1700万人口。自动驾驶是物理AI或具身智能中最具挑战性的方向之一。一辆无人车，本质上是一种集成了几乎所有关键AI能力的特殊机器人。近年来，大语言模型的发展成为关键突破，使我们能够生成更多数据、覆盖长尾场景、开展大规模仿真和道路测试。基于累计2亿公里的行驶数据分析，无人驾驶的安全性已达到人类驾驶的17倍。我们目前为止没有发生一起伤亡或一起重大事故。这将成为具身智能的基础，未来我们也会将这一基础模型用于工业机器人、社会服务机器人以及家庭机器人。

Third, biological AI. Professor Liu Yang developed the world's first AI Agent Hospital. It simulates a real hospital in Beijing, and everything is virtual: doctors, patients, nurses, and pharmacists are all agents. We emulate 21 different departments. The agents are evolving in this virtual world at a much faster speed than the real world. The result is that in two days, we're able to cover the work of a large 3A hospital in two or three years, with much higher accuracy—achieving an accuracy of 92% on the USMLE (US Medical Licensing Examination) benchmark. For an average doctor who has a license in the US, the average is 65%. [AI doctors] are not trying to replace real doctors, but to be agents and assistants to doctors for faster and more accurate diagnosis—also useful for rural areas without good medical access.

第三个例子是生物AI。刘洋教授团队开发了全球首个“AI智能体医院”。它完整模拟了一家北京的真实医院，医生、患者、护士和药剂师全部由智能体构成，共覆盖21个科室。在这个虚拟世界中，智能体的学习和进化速度远快于现实世界。结果是，仅用两天时间，就能完成一家大型三甲医院两到三年的工作量，并以美国执业医师考试（USMLE）为基准取得了92%的准确率，而美国持证医生的平均水平约为65%。需要强调的是，这些AI医生并非试图取代真实的医生，而是作为智能体助手，帮助医生更快、更准确地作出诊断，尤其对医疗资源不足的农村地区来说帮助性很大。

We also have about 1/3 of our professors working on AI for new drug development. I agree with Nobel laureate Demis Hassabis that AI is going to accelerate drug development dramatically. We are changing how drug is developed and how trials are conducted. Obviously, AI is becoming super powerful but also risky. I personally work with a few leading AI scientists and a few Nobel laureates to form the International Dialogue on AI Safety. The risk for AI is real, especially when expanding from digital to physical and biological domains, and we need specific and very prompt measures to take care of this.

此外，我们约有三分之一的教授正从事AI驱动的新药研发工作。我认同诺贝尔奖得主德米斯·哈萨比斯的判断：人工智能将极大加速药物研发进程。我们正在改变药物研发和临床试验的方式。当然，AI在变得越来越强大的同时，也伴随着风险。我本人也与多位顶尖AI科学家和诺贝尔奖得主共同召开人工智能安全国际对话。随着AI从数字领域拓展到物理和生物领域，风险将更加复杂，我们必须采取具体、及时的治理措施。

Let me also talk about AI in China versus the US. I get asked all the time: what’s the difference? Which one is ahead? In the US, OpenAI, Microsoft, Google Deep Mind, Anthropic, Grok, and Meta are hyperscalers. In China, models like DeepSeek, Qwen from Alibaba, and Doubao from ByteDance stand out. In terms of chips and infrastructure, the US is ahead, with the exception of electricity. China is actually way ahead in terms of electricity grid. In terms of models and software, China is more open and has smarter models with more efficient architectures. The US has more frontier, bigger models, and they are more closed. In terms of applications, China is probably ahead. My view is that it's not about who is going to win. I think both China and the US are going to win, and Europe and Africa are going to win. Everybody can be a winner. We are at the very beginning of a new revolution and everybody will benefit from this revolution. Healthy competition is wonderful. This is not a zero-sum game.

关于中美人工智能的比较，我经常被问到：区别是什么？谁更领先？在美国，有OpenAI、微软、谷歌DeepMind、Anthropic、Grok和Meta等超大规模机构；在中国，则有深度求索、阿里巴巴的通义千问、字节跳动的豆包等优秀模型。在芯片和基础设施方面，美国整体领先，但电力供应是一个例外。事实上，中国在电力网络方面明显领先。在模型和软件层面，中国更为开源，模型架构也更高效；美国则拥有更多前沿、规模更大的模型，但整体更为闭源。在应用软件方面，中国可能走在前列。但在我看来，这并不是一场你输我赢的竞争。中美会赢，欧洲和非洲同样会赢，所有人都将是赢家。我们正站在一场全新革命的起点，所有人都将从中受益。健康的竞争是好事，这不是零和博弈。

Finally, what is the future of AI? Just a disclaimer: I don’t represent anybody. My personal view is that, first, we are moving from generative AI to agentic AI, and we are actually moving from agentic AI to agentic network. We're moving from internet of PCs, internet of mobiles, internet of things, to the internet of agents. In the future, every person or every device will have agents. They will all interact, work together, evolve and come back with answer to serve the people. Secondly, both open and closed models will move forward. I see 80% of models will be open and 20% will be closed. We are also going to see a lot of foundation models, but the biggest opportunities will reside in vertical models—for robots, biology, and prediction. The opportunity there is at least 100 times bigger than that of foundation models. You can also have a lot of edge models, meaning models on your phones, your PCs, and your glasses, and that’s going to be a huge market as well. Third, a critical piece of generative AI is the scaling law. Scaling in pre-training is already slowing down, because data is almost depleted. You can still see the growth, but it has flattened out. The key intelligence lies in post-training, in agents, inference, reinforcement learning, and post-training scaling.

最后谈谈我个人对AI未来的看法。需要说明的是，我的观点不代表其他任何人。第一，我们正在从生成式AI走向AI智能体，并进一步走向智能体网络。从PC互联网、移动互联网、物联网，逐步迈向智能体互联网。未来，每个人、每个设备都将拥有智能体，它们相互协作、持续进化，最终带着答案更好地服务人类。第二，开源与闭源模型将并行发展。我预计大约80%的模型会是开源的，20%是闭源的。基础模型会很多，但最大的机会将来自垂直领域模型——例如机器人、生物和预测等领域。这些垂类模型所蕴含的机会，至少是基础模型的百倍。此外，还会出现大量端侧模型，也就是部署在手机、个人电脑、智能眼镜等设备上的模型，这同样将形成一个巨大的市场。第三，生成式人工智能的一个关键机制是“规模定律（或尺度定律）”。目前，预训练阶段的规模扩展已经明显放缓，主要原因在于可用数据正在接近枯竭。虽然能力仍在增长，但增速已趋于平缓。未来真正的智能突破，将更多来自后训练阶段，包括智能体、推理、强化学习以及后训练的扩展。

People always ask me when we are going to get AGI, Artificial General Intelligence. But there are hurdles to get there. We need new architectures, a better understanding of the physical world, a world model, and we need to do better job with memory, which is a key part of human intelligence. My personal take is that we can achieve AGI in digital or informational AI in probably less than five years. In physical AI, driverless cars will be the first to pass a Turing test, probably in three to five years. For humanoids, probably ten years. For biological AI—like a neural link of different sensors connecting your brain with AI—that will probably take another 15 to 20 years.

很多人问我，通用人工智能（AGI）什么时候会到来？实现AGI仍面临诸多挑战：我们需要新的模型架构，对物理世界更深入的理解，需要真正的世界模型，还需要在“记忆”这一人类智力的关键要素上取得突破。我的个人判断是，在数字或信息AI领域，AGI可能在五年内实现；在物理AI中，无人驾驶可能率先通过图灵测试，大约需要三到五年；人形机器人可能需要十年；而生物AI，例如通过多种传感器将人脑与人工智能连接的脑机接口技术，或许还需要十五到二十年。

Overall, I'm very optimistic about what AI can do for the world. It can do tremendously good, and it can also hurt people if we don't take the right direction. The future relies on young talents, on all of you.

总体而言，我对人工智能为世界带来的可能性保持高度乐观。若能把握正确方向，它将带来巨大的福祉；反之也可能给人类带来伤害。未来最终掌握在年轻一代手中，也正是在座各位的手中。

Thank you.

谢谢大家。

CCG 图书

● 出版 | 中国科学技术出版社

● 作者 | 苗绿

图书介绍

《在全球化的世界中行走》讲述了苗绿博士作为全球化智库联合创始人，在个人成长、海内外求学、创办智库、国际交流、民间外交、为国家建言献策等过程中的诸多故事与心路历程。作为慕尼黑安全会议青年领袖代表，苗绿博士曾对话联合国秘书长古特雷斯，开启 2021 慕安会第一问；她是比利时国王会见的七位全球青年领袖之一；她发起的“国际青年领袖对话”项目，推动了国际间青年的交流互鉴，得到中国国家领导人的回信；她经常受邀参加国际高端论坛，在巴黎和平论坛、多哈论坛等重要国际场合，参与设置议程，打造国际交流新叙事，以全球视野讲述时代中国，展现了新时代中国智库人的风采。

● 出版 | 中信出版集团

● 作者 | 王辉耀，苗绿

图书介绍

《21世纪的中国与全球化》首先梳理了全球化的变迁与理论发展，从技术与人本等新的视角观察全球化，并做出全球化的界定，总结了后疫情时代新型全球化具备的特征，然后对中国融入全球化的历史与现实进行了全面总结，用数据与事实说明，中国正在从全球化的受益者发展为反哺者，正在通过自身发展推动全球化进程，并尝试承担起更多国际责任，为全球治理创新贡献方案。作者对全球化发展的理论和文献做了梳理，回顾了全球化在世界和中国的发展历程，指出全球化走到了一个十字路口。本书从第四章开始，两位作者对中国推动全球化实现包容性和公平性发展的路径进行了探索，通过发挥中国的优势和特点，让中国为全球化发展注入新动力。作者基于长期的研究以及与国内国际、官产学各界有影响力重要人士的对话交流等，对中国的全球化发展路径及全球治理创新等形成了新的思考，提出中国推动全球化发展的三大支柱与七大路径。

● 出版 | 中国科学技术出版社

● 作者 | 王辉耀

图书介绍

本书深度剖析了中国在全球化浪潮中的角色演变与抉择，及其对全球未来的影响。全书分为三部分，第一部分回顾了中国融入全球化的历程，展示了中国从一个封闭的农业国家逐步转型为全球第二大经济体的过程。书中详细探讨了中国在贸易、投资、跨国企业崛起等方面的角色变迁，以及教育、人才和文化纽带在这一进程中的重要作用。第二部分探讨了中国在国际舞台上的崛起及其对全球治理的影响。作者分析了中国在多极化世界中的地位变化，风云激荡中的中国外交，中美关系的复杂性，以及中国在崛起的、更加一体化的亚洲中的角色。同时，还讨论了中欧关系的发展与挑战。第三部分审视了多边主义面临的挑战和改革。书中探讨了如何共同应对全球性挑战，寻找自由贸易的发展方向，以及“一带一路”倡议的发展。通过这些讨论，展示出中国在全球治理中的积极参与和贡献。

● 出版 | 中国科学技术出版社

● 编著 | 王辉耀、苗绿

图书介绍

《对话世界：理解新时代的全球化》全书分为三部分：第一部分“全球化发展史”回顾了全球化的历程，从古代贸易到现代经济转型，探讨了全球化的起源与演变。通过与耶鲁大学教授瓦莱丽·韩森、《金融时报》首席经济评论员马丁·沃尔夫和《世界是平的》作者托马斯·弗里德曼的对话，揭示了全球化的多层次发展。第二部分“弥合全球不平等与赤字”探讨了全球化带来的不平等和治理赤字问题。诺贝尔经济学奖得主安格斯·迪顿、巴黎和平论坛主席帕斯卡尔·拉米、亚洲协会副所长温迪·卡特勒等嘉宾，分享了他们对全球经济不平等、贸易体系和制度改革的看法。第三部分《权力转移与大国关系》分析了21世纪的权力转移和大国关系，特别是中美关系的复杂性。通过与哈佛大学教授格雷厄姆·艾利森、“软实力之父”约瑟夫·奈、布鲁金斯学会主席约翰·桑顿等专家的对话，讨论了大国竞争、合作以及全球治理的未来。

● 出版 | 中信出版集团

● 编著 | 王辉耀

图书介绍

作为海内外决策层和广大公众理解中美关系时广泛引用的框架，“修昔底德陷阱”将成为未来几十年对全球秩序有决定性影响的问题。在与全球化智库(CCG)理事长王辉耀的对话中，格雷厄姆·艾利森就中美关系和中美地缘政治竞争、中国崛起、美国外交政策、美苏关系、全球地缘政治、核武器、朝鲜问题、新冠疫情及影响等议题进行了深入阐述；全面、系统性地展示了艾利森对“修昔底德陷阱”和中美经济、金融、科技、军事、外交等多个方面存在的结构性矛盾和竞争的看法；深入而透彻地分析了中美双方实力的变化，以及发生战争的风险；坦诚而直率地提出了跨越“修昔底德陷阱”的方法和建议。

● 出版 | 中信出版集团

● 作者 | 王辉耀，苗绿

图书介绍

《我向世界说中国》是由全球化智库（CCG）主任王辉耀和秘书长苗绿基于“世界新格局下的中国对外叙事及话语权重塑”问题研究的重要成果，由中信出版集团出版。据悉，该书讲述了全球化智库近年来立足芒克辩论会、慕尼黑安全会议、巴黎和平论坛、达沃斯论坛等知名国际舞台，与各国政商学界知名人士畅谈国际时局与未来趋势，回应各方对于中国的关切和质疑，诠释中国的发展模式，降低外界对中国的误解，通过多层次、多主体、多元化、多渠道国际交流及传播，以全球视野讲述时代中国，积极塑造可信可爱可敬的中国形象的生动故事。同时，本书立足国际形势变化和全球传播新格局，针对中国应当如何开展对外交流和传播工作、如何创新外宣方式讲好中国故事等问题进行了深入浅出的剖析。