Unlocking Universal Artificial Intelligence: The Key Lies in Establishing 'Heart'

Many people regard the concept of Scaling Law as a fundamental principle, believing that more data, greater computing power, and larger model parameters are all that is needed. However, achieving general artificial intelligence (AGI) requires more than just data; we must explore an alternative approach—instilling a 'heart' in artificial intelligence. At a recent academic forum commemorating the 20th anniversary of the Lianhuashan Research Institute in Ezhou, Hubei, Zhu Songchun, director of the Beijing Academy of General Artificial Intelligence, shared his insights on this topic.

Zhu emphasized that AGI has become a critical focal point in global technological competition. To achieve breakthroughs in this arena, it is essential to clarify the sources of big data and accurately define the direction for AI development. He noted, "If we cannot process visual data, the AI system remains an empty shell." According to Zhu, data labeling serves as a special pair of "glasses" for computers, enabling them to recognize and comprehend images, text, and other data nuances.

In 1997, Scott Konish completed the world's first dataset labeling—image boundaries—to train classifiers. This paved the way for Zhu’s own large-scale, high-granularity data labeling efforts, which began in 2004 after recognizing the potential of statistics for image understanding. However, in 2008, Zhu and his team encountered two major bottlenecks in their data-labeling process. First, elements like value, causality, and intention are often hidden beneath the surface of perceptual data, making them difficult for sensors to detect and label. Second, the process of data labeling is highly specific to particular tasks, with varying requirements that do not necessarily improve by simply expanding the data or model size.

These challenges prompted Zhu to reflect deeper on AGI. He perceives AGI as a complex mega-system comprised of core fields such as computer vision and natural language processing, with development intricacies akin to "landing on the moon." In contrast, the big data approach resembles "climbing Mount Everest," with vastly different objectives.

So, how can we explore the path to AGI? Zhu advocates for a shift in AI research from a focus on "theory" to "heart." "Theory" encompasses mathematical models, while "heart" refers to cognitive architectures and value alignment. After nearly three decades of development, various core AI fields are increasingly converging and intersecting, progressing toward AGI. Zhu believes that through this integration, a unified AI architecture will emerge, enabling a transition from specialization in single tasks to the capability of solving a wide range of tasks and autonomously defining new ones.

In Zhu's view, instilling a "heart" in machines, achieving the transition from "theory" to "heart," and evolving from big data to significant tasks—shifting from perception to cognition—represents the academic frontier for the next 10 to 20 years and is a core mission for the field of intelligent studies.