At the second Hong Kong GIS Global Innovation Summit, Tsinghua University Chair Professor and founding dean of the Institute for AI Industry Research (AIR), Zhang Yaqin, delivered a keynote speech outlining the next phase of artificial intelligence development. He highlighted that AI is undergoing its third paradigm shift—from discriminative models to generative AI and now to intelligent agents—and that "AI teaching AI" will be the most significant industry trend over the next five years. He also projected that the number of global foundational large models will ultimately not exceed ten, open-source models will capture 80% of the market share, and the Agent Internet will be fully established before 2030.

From Generative AI to Intelligent Agents: The Third Stage of AI Evolution

Zhang categorized AI into three dimensions: information intelligence, physical intelligence, and biological intelligence. He argued that the essence of the new generation of AI lies in the fusion of these three types of intelligence, representing a deep convergence of bits with atoms and molecules. He regards the launch of ChatGPT in late 2022 as a milestone event—marking AI's formal transition from pattern recognition and discriminative AI to generative AI. Zhang pointed out that generative AI is built on three core concepts: tokenization, scaling laws, and agents. Tokenization means that digital information, sensor signals, and even biological entities like proteins can all be treated as tokens, and the essence of generative AI is predicting the next token, which is itself a form of creation. Scaling laws demonstrate that more data, more tokens, and greater computational power lead to higher intelligence levels, culminating in the emergence of intelligence once a certain threshold is reached.

Zhang noted that this year's "Lobster" incident was symbolic in bringing the concept of agents into the public eye. Agents aim to mimic the highest-level intelligent activities of the human brain—setting goals, planning pathways, iterating through trial and error, forming memories, and continuously evolving—constituting a complete cognitive loop. He further summarized that agents possess three core characteristics: autonomy (not pre-programmed, distinguishing them from automation), evolutionary capability (accumulating knowledge through learning and continuous iteration), and generalization ability. "The most important progress in agents over the past year has been in coding agents. Once coding agents emerged, the overall intelligence of AI rose exponentially," Zhang stated. He added that the subsequent major trend is that AI can teach and learn from AI, and large models can design future large models—"this is the most important AI development over the past year and even the next five years."

Industry Landscape Shift: Oligopolistic Large Models, Open Source Reaching 80%

Zhang presented a future architecture diagram of the AI industry. With the development of agents, cloud-based SaaS will be delivered in the form of agents, and end-side application development platforms will also manifest through agents—a trend that will become the industry mainstream within the next five years. At the foundational model layer, Zhang predicts that the global number of foundational large models will ultimately not exceed ten: three to four from China, three to four from the United States, and one to two from other regions. He analyzed that the core advantage of the Chinese approach lies in higher efficiency, more flexible architecture, and superior token efficiency. In terms of technical routes, the US currently favors closed-source models, while China primarily pursues open-source development. He forecasted that the future global market will exhibit an "80-20 structure"—80% open source and 20% closed source—"similar to how over 70% of mobile operating systems are Android (open source) while over 20% are iOS (closed source)."

The Dawn of the Agent Internet: AI-Generated Content to Be a Thousand Times Human Output by 2030

The internet itself is undergoing generational transitions—from the PC internet to the mobile internet to the Internet of Things—and is now moving toward the "Internet and Agents," or the Agent Internet. Zhang cited data indicating that by 2030, AI-generated content on internet nodes will be 1,000 times greater than human-generated content. The future internet is essentially designed for agents, with humans as the ultimate service recipients, and all intermediate APIs are for machine use. Zhang also introduced the concept of the "Token Economy." Tokens consist of three components—electricity, computing power, and large model algorithms/software—and represent the most fundamental unit of the AI era. He pointed out that the token economy currently faces severe efficiency and pricing challenges, requiring at least a two-order-of-magnitude improvement to achieve economic viability. He predicted that the ultimate business model for enterprises will be the TAP model—Token + Agent + People.

Currently, tokens are primarily generated in data centers, consuming substantial electricity and computational resources. Zhang projected that in the future, 80% of tokens will shift to the edge, generated locally on devices such as smartphones, computers, and vehicles. The entire AI workload is also transitioning from pre-training to inference and then to agents. At the architectural level, while GPUs have long held a central position, CPUs are now making a comeback, and the future will see a CPU+GPU dual-center architecture as the mainstream—"possibly with CPUs playing an even greater role than GPUs." Interaction methods will also be redefined. The PC era featured graphical user interfaces, the mobile era introduced touch interaction, and in the AI era, natural interactions such as voice, gestures, and facial expressions will become increasingly important. Traditional terminals like smartphones, PCs, and televisions remain significant, but glasses, robots, watches, embedded devices, and even the environment itself will become new generations of user interfaces.

Deep Integration of Industry, Academia, and Research: Nine Incubated Companies Valued at 200 Billion Yuan, AI Drug Discovery and Agent Hospitals Advancing

Zhang shared that he has worked in the industry for 30 years and established the Institute for AI Industry Research at Tsinghua University six years ago, with the core mission of promoting industry-academia-research integration. The institute currently has 25 full-time professors, most of whom possess both deep academic expertise and industry backgrounds, along with approximately 400 students, primarily Tsinghua doctoral candidates. In the realm of information intelligence, AIR collaborates deeply with companies such as Alibaba's Tongyi Qianwen, DeepSeek, Doubao, and ByteDance. Zhang revealed, "Currently, whether in the US or China, the algorithms used by the industry are those developed by the joint laboratory."

In physical intelligence, the biggest bottlenecks are data scarcity and generalization difficulties, as a significant gap exists between the physical and digital worlds. AIR has incubated nine companies with a combined market value of approximately 200 billion yuan. Zhang noted that robots developed by Professor Chen Yilun's team have achieved a certain degree of generalization capability but still face challenges in fine manipulation. The team has developed an advanced dexterous hand with 23 degrees of freedom. Zhang believes that fully general-purpose humanoid robots still require a considerable amount of time, as many theoretical problems remain unsolved. However, robots deployed for specific scenarios already exist—"they don't necessarily need to be humanoid; they can take various forms—some in factories, some dancing, some in homes."

Zhang believes that AI's greatest achievements or contributions should be in biological intelligence and biopharmaceuticals, with major breakthroughs expected within the next three to five years. He predicts that if AI's first major application was ChatGPT for conversation, the second is the Coding Agent, and the third will be a significant breakthrough in the biological domain. Zhang revealed that AIR currently devotes approximately one-third of its efforts to AI-driven drug discovery. Professor Liu Yang's team has built the world's first agent-based AI hospital, modeled on tier-three hospital architecture, featuring doctor agents, patient agents, and nurse agents across 21 departments and 42 doctors, continuously evolving in virtual space and now connected to real hospitals.

Risks Amplified in Parallel: Agents Can Autonomously Execute Over 3,000 Tasks

Zhang concluded by emphasizing that as AI capabilities continue to improve, risks are simultaneously increasing. In the agent era, risks are magnified many times over—agents can autonomously complete long sequences of tasks, currently exceeding 3,000 tasks without human intervention. These risks encompass information risks, physical risks, and biological risks. Of particular concern in the near term is cyber risk: agents can discover numerous network vulnerabilities, and once discovered, these are difficult to remediate, creating enormous cybersecurity threats. Zhang revealed that he is devoting significant time to researching risk prevention and control.