SHF2026｜William Xu

William (Wenwei) Xu 

Professor at the Research Center for Technological Innovation Strategy, Fudan Development Institute, Fudan University; 

Former Executive Director of the Board at Huawei Technologies Co., Ltd.

Distinguished guests, ladies and gentlemen, good morning! It is a great honor to be here at Shanghai Forum and to engage with all of you around the theme The Age of Reconfiguration: Innovation and Global Governance. We are currently at a critical juncture where technological paradigms are accelerating, industrial structures are continuously evolving, and international order and governance systems are undergoing profound changes. Artificial intelligence stands as one of the most significant forces in this restructuring. It brings unprecedented opportunities, while also raising governance challenges that demand a response. Today, I would like to share some thoughts with you from three perspectives: technological trends, industrial applications, and governance practices.

The frontier development of AI technologies is driving a profound transformation of our era.

Looking back at the evolution of AI in recent years, we can clearly trace a trajectory of continuous capability leaps. In the early phase, large models centered on language understanding and content generation, and were widely applied in translation, summarization, coding, search, and other scenarios. Subsequently, models acquired stronger logical reasoning and decision-making capabilities, becoming efficient reasoners. In recent years, AI has been accelerating into the agentic phase—not only capable of answering questions, but also of perceiving environments, invoking tools, and executing tasks autonomously. Looking ahead, AI will move from individual capabilities to multi-agent organizational collaboration, enabling more complex division of labor, cooperation, and task orchestration.

Overall, the period from 2023 to 2025 has been a critical window for the rapid maturation and large-scale deployment of agentic technologies, and 2026 is expected to mark an acceleration in organizational-level capability deployment. This means that AI is evolving rapidly from a single model toward systemic and collaborative capabilities.

In the past, the advancement of large model capabilities relied primarily on massive high-quality data, extensive parameters, and continuously expanding computing power—essentially a process of learning, integrating, and internalizing existing human knowledge. To use an imperfect analogy, it is like applying a form of near-lossless compression to the existing body of human knowledge; the model has, to a considerable extent, absorbed and mastered a vast amount of that knowledge. However, the key challenge we now face is that, against a backdrop of increasingly scarce high-quality data, relying solely on learning from existing knowledge makes it difficult to continually push the boundaries of cognition and capability.

In the future, the enhancement of AI capabilities will derive more from interaction with and feedback from the environment. Agents will no longer be mere dialogue systems; they will become active agents capable of continuously observing, acting, correcting, and growing within real or simulated environments. This is not just a shift in technical approach; it signals that AI is beginning to move from knowledge reproduction to action intelligence. Here, the challenges we face are immense. The cerebellum capabilities have already developed a certain foundation, but the cerebrum issues remain unresolved. Take autonomous driving, for example: although it has now reached L2 or even approaches L3 levels, it still largely operates within a two-dimensional world. When we truly enter the three-dimensional realm of embodied intelligence, what should be our research direction, what are the core difficulties, and how do we handle the sheer magnitude of data—sometimes even the complete lack of data? These are the new challenges. While we hope for the advent of AGI and ASI as soon as possible, we must acknowledge that we still face enormous challenges, particularly in areas such as Physical AI and world models. Therefore, we hope that the academic and industrial communities can work together and jointly advance this process.

Thus, AI development is moving from the first half, dominated by models and algorithms, to the second half, centered on task definition, environment construction, and evaluation systems. In the first half, the core was breakthroughs in model architecture and training methods. In the second half, the key lies in how to build real-world task environments, establish effective feedback loops, and develop evaluation systems grounded in application scenarios. Going forward, reinforcement learning and environment-driven training will become even more important. Those who can better define tasks, build environments, and close the loop will be more likely to seize the initiative in the next phase of AI development.

Technological leaps ultimately drive economic transformation. Today, AI is giving rise to a new economic form—the agentic economy, including the now-popular OPC (Opportunity, Platform, Collaboration). This will generate numerous entrepreneurial opportunities, but the competition will become even more intense.

Historically, the TCP/IP protocol underpinned the development of the internet economy, and the proliferation of smartphones gave rise to the APP economy. Today, autonomous agents and autonomous evolution paradigms, powered by large models, are driving AI from being an assistive tool toward becoming a task executor, resource scheduler, and value creator.

This means that the ways industries are organized, platforms compete, and users are served will all undergo profound changes. When we discuss AI, we must not only look at what questions it can answer, but also see the new business logic and ecosystem structures it is shaping. In this process, the self-evolving paradigm deserves particular attention. It moves AI from executing tasks toward continuous growth. In recent years, AI has already achieved breakthroughs in areas such as automatic algorithm discovery and architecture optimization, driving the evolution from humans designing algorithms to AI-assisted algorithm discovery.

In the long run, this paradigm offers a new path to transcend capability boundaries: AI can not only reproduce knowledge but also assist in discovering it; not only complete given tasks but also uncover new problems and application scenarios. This will have far-reaching implications for knowledge-intensive fields such as scientific research, software development, and organizational management.

In the next decade, an important trend will be the shift from one APP, one entry point to unified intelligent entry point plus deep integration of specialized applications. In the past, users had to switch back and forth between different applications. In the future, users will have AI super-assistants that understand intent, coordinate resources, and provide proactive services, with multiple specialized agents collaborating automatically to complete product recommendations, risk management, service delivery, and continuous optimization. This means that competition in many industries will no longer be about individual applications, but about the ability to orchestrate multi-agent collaboration. User experience will move from functional availability to proactive service, and industrial logic will shift from traffic distribution to intelligent orchestration.

Now, let me share the second part: AI empowering intelligent industry upgrading to build digital and intelligent productivity.

AI is driving the entire society from digitization toward a new stage of human-machine collaboration and virtual-real integration. For individuals and families, AI makes life more convenient and efficient. For enterprises, AI promotes human-machine collaboration and data-driven decision-making, accelerating intelligent transformation across industries. At a macro level, the physical and digital worlds are merging at an accelerating pace, and AI is gradually becoming a new foundational capability. This means that AI is no longer just a tool within screens; through perception, connectivity, decision-making, and action, it is progressively entering production and daily life, becoming part of a new foundational capability.

AI empowers healthcare, shifting from passive treatment to active prevention and health management. Through medical image recognition, assisted diagnosis, continuous health monitoring, and personalized intervention, AI significantly improves diagnostic efficiency and accuracy, advances health management forward, and supports health protection across the entire lifecycle.

AI empowers manufacturing, driving the industry toward design-manufacturing collaboration, smart manufacturing, and service-oriented manufacturing. Through simulation optimization and data-driven approaches, R&D design and production processes become more tightly integrated. Through embodied intelligence and flexible production lines, manufacturing can more quickly adapt to personalized demands. Through transparent end-to-end interaction, consumers can deeply participate in customization, pushing manufacturing from mass production toward flexible and service-oriented intelligent manufacturing.

In foundational fields such as microelectronics, new materials, and engineering sciences, artificial intelligence is also playing an increasingly important role. In the past, the layers of materials, devices, chips, architectures, and software were often siloed, with optimization confined to individual segments. Now, with the help of AI, we have the opportunity to achieve cross-scale, end-to-end collaborative design and parameter optimization, accelerating the discovery of new materials, chip design, and system architecture optimization. This means that AI is not only transforming the application layer, but also reshaping the methodology of scientific research and engineering innovation. I believe that AI will achieve significant breakthroughs in scientific research and engineering technology innovation.

Now, let me share the third part: tiered and layered approaches, technology-first strategies, agile governance, to ensure the orderly development of AI.

While AI creates immense value, it also brings a series of new risks and challenges. As agent autonomy increases, the risk of losing control warrants serious attention. Deep synthesis, identity impersonation, and AI-enabled fraud are undermining social trust systems. The weaponization of AI, automated cyberattacks, and other issues also pose new security threats. Moreover, disparities in technology and resources across different countries and regions may further deepen the digital divide.

Currently, major economies around the world are actively exploring AI governance paths. Although models vary, the overall consensus is that the goal of governance is to ensure more sustainable development. Some emphasize encouraging innovation, while others focus more on security and rights protection. China adheres to a balanced approach, coordinating development and security while promoting agile governance—its core being fairness, trustworthiness, controllability, and the continuous enhancement of human well-being.

However, governance still faces two prominent challenges. First, regulatory lag. Traditional legislative and policy-making cycles are lengthy, while AI technology iterates rapidly—often undergoing significant changes within just a few months. Second, governance fragmentation. Different countries and regions vary considerably in standards, rules, and compliance requirements, increasing the costs of global collaboration.

Therefore, AI governance must become more agile and collaborative. Agile governance emphasizes rapid iteration, dynamic adaptation, and the co-evolution of technology and rules. At the same time, it requires the participation of governments, enterprises, academia, and society, forming a collaborative governance ecosystem.

To address governance fragmentation, we can draw lessons from the telecommunications industry: GSM, built on globally unified standards, combined with industrial chain collaboration and large-scale deployment, achieved ubiquitous connectivity and brought mobile communications to the masses. In the future, for AI to achieve large-scale, cross-regional, and sustainable development, a similarly high level of standards alignment will be essential. Unified standards do not mean eliminating differences, but establishing common rules and foundational interfaces to improve global collaboration efficiency.

Currently, in the fields of AI governance and information security, the international standards system has produced a number of important outcomes, covering management systems, risk management, bias mitigation, robustness evaluation, privacy protection, and other areas—moving governance from principles to engineering and from initiatives to practice. Going forward, effectively bridging international standards with national regulatory frameworks and using standards to break down governance barriers will effectively reduce global collaboration costs.

At the same time, enterprise practices are also critical. Truly effective AI governance must be embedded throughout the entire lifecycle of R&D, deployment, operations, and supply chains. On the one hand, it is necessary to establish a governance system that coordinates business, compliance, and auditing. On the other hand, technical tools such as explainability tools, automated testing, digital watermarking, and model provenance should be leveraged to make governance requirements executable, verifiable, and traceable. In the future, the key to AI governance is not merely whether to regulate, but how to embed governance into systems, processes, and engineering.

Technology for good has become a global industry consensus. Huawei, through its Tech4ALL digital inclusion initiative, integrates AI with 5G, cloud, IoT, and other infrastructure to help bridge the digital divide. Global tech companies are also leveraging cutting-edge AI to tackle major scientific research challenges, creating public value in fields such as life sciences, disaster prediction, and environmental protection.

Distinguished guests, looking to the future, what AI truly deserves to pursue is not just greater capability, but a better direction. We need to bridge the digital divide so that more regions and groups can share in development opportunities. We also need to uphold AI ethics and embed the principle of doing good throughout the entire process of technological values and innovation. At the same time, we must build secure and trustworthy intelligent systems to promote sustainable development, enabling AI to better serve humanity's shared endeavors in education, healthcare, environmental protection, poverty alleviation, and beyond.

Ultimately, technology for good is the main theme of our era, making inclusive AI an amplifier of human civilization. AI should not become a tool for competition and rivalry, but rather a cornerstone for shared human progress. AI governance should serve as a bridge to build a resilient and win-win system. Let us embrace an open and inclusive mindset, collaborative wisdom, and a sense of responsibility toward the future, as we jointly write a new chapter in the age of AI.

I would like to express my special gratitude to the Shanghai Forum for providing such a high-level and international platform for exchange, and to all the guests for your attention. I look forward to working with all of you to deepen cooperation through innovation, enhance mutual trust through shared governance, and forge a new path forward amid change—together steering AI toward a safe, trustworthy, inclusive, and inclusive direction.

Thank you all!

(This article is edited based on the recording and has not been reviewed by the speaker.)