Hangzhou once again embraced the wave of intelligence as the 5th Innovation Forum on Intelligent Computing, co-organized by Zhejiang Lab (ZJ Lab) and Science/American Association for the Advancement of Science (Science/AAAS), opened on November 20, 2025. Now in its fifth consecutive year, this annual event brings together global leading academics in intelligent computing to exchange insights on cutting-edge research trends in computing and intelligence and foster East-West intellectual dialogue.

18 top experts and scholars from China, the UK, the US, Canada, Switzerland, and other countries delivered intensive keynote presentations over two days, envisioning a vivid blueprint for the future intelligent world.

Intelligence Evolution of Foundation Models

"Science describes dimensional spaces far beyond the reach of language. To better understand our physical world, we must transcend linguistic boundaries and delve into spectra, molecules, and genes," stated XUE Guirong, the chief engineer of ZJ Lab's comprehensive planning regime for scientific models, in his opening remarks. He highlighted that science foundation models will initiate a new paradigm for addressing scientific challenges, thereby better serving research. He introduced ZJ Lab's efforts in developing the 021 Science Foundation Model from scratch, emphasizing that a key aspect lies in tokenizing diverse scientific data and embedding it into a unified space for training and reasoning, thus tackling complex interdisciplinary problems. He expressed his anticipation for deep collaboration with scientists across various fields to drive a paradigm shift in scientific research.

JIANG Daxin, Co-Founder and CEO of StepFun AI, China, presented recent observations and future trends of foundation models. "Research indicates that the token consumption required for enhancing foundation model capabilities has grown exponentially. In the age of electricity, electricity consumption served as a key barometer of a country's economic activity; in the age of AI, this role may be taken over by token consumption." He further noted that, amid surging demand for computing power, model reasoning efficiency has emerged as a pivotal factor in determining the large-scale application of AI. Enhancing this efficiency necessitates collaborative optimization across upstream and downstream industries, as well as model-chip co-design. As reasoning capabilities advance, large language models (LLMs) are transitioning from merely chatting to performing tasks. "In the future, digital agents will increasingly extend their effective working hours, while simultaneously expanding their presence from the digital realm into the physical world and moving from empirical learning to self-evolution."

AI+Science Accelerates Scientific Discovery

As intelligent technologies evolve, AI is profoundly reshaping our understanding of Earth's interior, the vast universe, and even the code of life.

Prof. John Ludden, Member of Academia Europaea (MAE) and Co-Chair of the GeoGPT Governance Committee, highlighted in his keynote talk that AI has demonstrated immense potential in combating climate change, providing early warnings for natural disasters, and advancing complex geological exploration. Richard Chuchla, Co-Chair of the GeoGPT Governance Committee, further elaborated on integrating AI into earth science. He emphasized the critical role of multidisciplinary integration, systems thinking, and four-dimensional spatiotemporal analysis in tackling the challenges posed by the Earth system's complexity. SONG Xiaodong, Chair Professor at Peking University, presented AI applications in deep Earth exploration through the GlobalTomo dataset—the world's first AI Earth imaging database, seismic wave simulator and inversion network. This dataset, comprising 160 million seismograms and 2.6TB of data, serves as a powerful tool for investigating Earth's internal dynamics.

The integration of astronomy and AI is advancing. Heloise Stevance, a research fellow at the University of Oxford, opened her speech by introducing how to discover transients. She emphasized that current transient research is confronted with a vast amount of observational data and limited spectroscopic resources. Stevance introduced the application of virtual research assistants in processing massive survey data, while cautioning against blind adherence to AI trends. Instead, she advocated for deep integration of domain-specific expertise into foundation models to ensure the interpretability of research outcomes. Prof. Jean-Paul Kneib from École Polytechnique Fédérale de Lausanne further highlighted the "data tsunami" in cosmic observations, driven by next-generation astronomical facilities such as JADES, MUST, SKAO, LISA, and Einstein Telescope. He underscored the pivotal role of AI and high-performance computing in astrophysical data processing, noting their potential to advance our understanding of frontier topics including dark matter, dark energy, gravitational waves, and the early universe. LUO Ali, a research fellow at the National Astronomical Observatories of the Chinese Academy of Sciences (NAOC), presented the OneAstronomy model jointly developed by ZJ Lab and NAOC, and the Global Open Transient Telescope Array (GOTTA) project included in the first programs endorsed by UNESCO's International Decade of Sciences for Sustainable Development.

At the forum, JIN Yaochu, a professor at Westlake University and Member of Academia Europaea, systematically expounded a research pathway advancing from "brain-like computing" to "brain-like embodied intelligence". WANG Jinlan, Distinguished Professor at Southeast University, presented her team's work on machine learning material design and synthesis for small datasets.

AI Applications Empower a Myriad of Industries

AI application is another hot topic at the forum. ZHANG Jie, a Professor at the University of Science and Technology of China and Member of the US National Academy of Engineering, discussed "intelligent listening" technology. Unlike vision, sound and vibration data exhibit lower volumes and energy consumption, enabling all-weather monitoring. "Intelligent listening" allows machines to perceive, interpret, and convert sounds and vibrations into actionable information. This technology holds broad prospects for applications like anomaly detection, road monitoring, and early warning of natural disasters. It is already deployed in Hefei's CitySeis project, which uses over 50,000 sensors to cover various scenarios including urban safety, traffic, and meteorology.

Enabling agents to move from the digital to the physical world requires technological breakthroughs that allow machines to autonomously adapt to dynamic environments. Prof. Osmar Zaïane from the University of Alberta discussed robot robustness, flexibility, and human-AI alignment in scenarios involving faults, complex environments, and human collaboration. He highlighted that developing next-generation adaptive machine systems—capable of fault adaptation, dynamic environment adaptation, and seamless human collaboration—requires integrating cutting-edge technologies such as reinforcement learning, large language models, multimodal reasoning, and multi-agent communication.

Prof. JIANG Jun from the University of Science and Technology of China introduced a robotic AI-Chemist developed by his team. Traditional chemical research often relies on trial and error, which suffers from low efficiency and sparse performance data. In contrast, the AI-Chemist performs tasks tirelessly with precise automation. Simultaneously, its “chemical brain” conducts quantum chemical simulations and integrates theoretical big data with experimental small data to generate a predictive AI model. This model ultimately yields an ‌algorithmically predicted and validated global optimal formula. The AI-Chemist's application has significantly enhanced the efficiency of material and molecular design, with successful validation across diverse systems including catalysts, optical polymers, and antibacterial peptides.

Prof. YU Duli from Beijing University of Chemical Technology presented the application of spatial AI in high-speed railway subgrade monitoring. He highlighted that integrating spatial AI, distributed sensor network and edge computing technologies enables the development of a monitoring system with real-time perception, intelligent analysis and quick response capabilities. This integration significantly enhances the operational safety and reliability of high-speed railways. In the long run, the spatially-driven monitoring system is expected to evolve towards autonomy, where sensor networks, AI algorithms, and cloud infrastructure will synergistically advance, achieving fully automated monitoring, diagnosis and decision-making without human intervention.

PANG Zhibo, a tenured professor at Peking University, then delivered a keynote talk on AI-driven next-generation industrial infrastructure. He systematically elaborated on AI applications across diverse industrial automation scenarios, including machine perception, automated mining, operational assistance, and autonomous robotics. Additionally, he introduced "Cloud Fog Automation", a new design paradigm for industrial control systems, as well as the embodied AI framework developed on this basis.

The forum also featured a dedicated panel discussion session, where John Ludden and Richard Chuchla (co-chairs of the GeoGPT Governance Committee), ZENG Yi (a research fellow at Institute of Automation, CAS), Heloise Stevance (a research fellow at the University of Oxford), YU Duli (a professor at Beijing University of Chemical Technology), and other guests engaged in an in-depth exploration of "Open, Ethical, and Responsible AI Development and Governance". XIAO Yitian, Chief Scientist at Zhejiang Deep-Time Digital Earth Research Center, moderated the panel discussion.

Despite the conclusion of the forum, ripple effects of scientific enlightenment have continued to spread. As emphasized by TONG Guili (Secretary of the Party Committee of ZJ Lab) in her opening remarks, the multi-dimensional insights and profound analyses shared during the event will positively influence the evolution of knowledge systems and methodological innovation in intelligent computing. "ZJ Lab eagerly collaborates with global peers to forge a tighter-knit network and push the frontiers of intelligent computing," she said. Similarly, Bill Moran, Publisher of the Science Family of Journals at AAAS, underscored the importance of shared vision and collaboration in his address. "Alone we can do so little; together we can do so much," he remarked.