Chongqing - Can AI apply what it learns to unfamiliar tasks and situations? That is a central goal of artificial general intelligence (AGI), explored by researchers and industry representatives at the Conference on General Artificial Intelligence in Chongqing on October 21–23, with a focus on real-world applications.

China is a major player in the global AI race. According to the China Academy of Information and Communications Technology, China had more than 6,600 AI companies by June 2026, accounting for around 15% of the global total.

The Conference on General Artificial Intelligence was held in Chongqing from October 21 to 23. (Photo/Huxin Luo)

Beyond task performance: what does AGI require

Today's AI can perform increasingly complex tasks, but success in specific settings does not necessarily mean they can generalize knowledge or act autonomously in unfamiliar situations.

Fields Medalist David Mumford, an honorary professor at Harvard and Brown universities, said agency, or the ability to act independently, remains a major challenge for AI. He also emphasized that AI needs appropriate values to act autonomously and interact responsibly with human society.

At the conference, director of the Beijing Institute for General Artificial Intelligence, Zhu Songchun, presented his team's CUV framework for studying general intelligence.

CUV combines cognitive architecture (C), which organizes cognitive modules and their communication; a capability system (U), encompassing vision, language, movement, learning and cognition; and a value system (V), which provides the underlying motivations guiding decisions and actions. 

The framework explores how these elements can work together in the pursuit of more general and adaptable intelligence.

Robotics provides a practical setting for exploring these challenges. At the conference, experts discussed whether world models could help robots overcome deployment challenges, including limited data, changing environments and difficulties generalizing across tasks.

Feng Liang, deputy dean of Chongqing University's School of Computer Science, noted that data generated by world models needs to reflect physical laws and match real tasks if it is to help robots move beyond laboratory demonstrations.

From computing power to applications: progress and limits

Developing more adaptable AI requires not only advances in models, but also computing systems capable of supporting complex tasks and applications that reveal where current capabilities fall short.

Li Zicheng, a representative of Hygon, a Chinese semiconductor company specializing in high-end processors and AI accelerators, said that cross-domain reasoning and generalization require more than simply adding computing power. 

As computing systems support model training and inference alongside AI agents' task planning, tool use and workflow orchestration, coordination between CPUs and accelerators, along with high-speed interconnects, becomes increasingly important.

At Alibaba Group's Alimama Marketing and Experience Center, AI has advanced from generating individual marketing assets and creative content to supporting merchants' end-to-end marketing workflows, according to the center's head Li Mu.

Yet brand strategy, business decisions and deeper consumer insights still require substantial human involvement. AI can offer useful insights, but the effectiveness of its recommendations needs further validation.

These examples show progress in both computing infrastructure and multi-step task execution, without suggesting that AGI has already been achieved.

Visitors film an expression-recognition robot as it scans a person's face to learn their current facial expression. (Photo/Conference on General Artificial Intelligence)

Modeling social intelligence, applying AI in Chongqing

The need for AI to understand complex environments extends beyond the physical world to human society, where outcomes depend on interactions among people, information and changing circumstances.

At the conference, Wang Hui, vice president of Communication University of China presented research on simulating the evolution of public opinion through interactions among social, information and cognitive factors. By modeling how AI agents interact and respond to changing information, the research explores how social events may evolve under different conditions.

Turning such research into practical applications requires mechanisms to connect technological advances with real-world needs. In Chongqing, the Mingjing platform launched at the conference aims to bridge this gap by linking original research with practical applications.

It supports the development, training, testing and iterative improvement of AI agents, with applications spanning areas such as urban governance, public services and industry. By connecting research capabilities with practical demands, the platform provides a way to test and refine AI systems in real-world settings.

These initiatives do not demonstrate that AGI has been achieved. Rather, they illustrate how real-world applications can help researchers evaluate system performance, identify limitations and use feedback to guide further development.

AGI remains a long-term research goal, with no universally agreed timeline for its realization. China's AI industry and Chongqing's research initiatives, manufacturing capabilities and urban applications provide settings for exploring how these elements can work together.