传统大模型有何局限性？欧洲科学院院士金耀初：能耗比较大且没有自主学习能力

Core Insights - The "2025 Technology Wind List" annual event was held on January 15, 2026, in Beijing, focusing on the theme "Inspiring New Intelligence, Embarking on a New Journey" [1][4] - Jin Yaochu, an academician of the European Academy of Sciences and founder of the Academy of Trustworthy and General Artificial Intelligence at Westlake University, emphasized the importance of safety in embodied intelligence, noting that large models have inherent hallucinations and safety risks [1][4][5] Group 1 - Jin Yaochu discussed the need to integrate pulse-based information processing methods into embodied intelligent systems, which is currently being explored [3][7] - He highlighted the limitations of traditional large models, including high energy consumption and lack of autonomous learning capabilities, contrasting them with pulse neural networks that are event-driven and energy-efficient [3][7] - The integration of brain mechanisms such as plasticity and neural regulation into data-driven and large model-driven approaches is essential for achieving autonomous perception and decision-making [3][7] Group 2 - Jin Yaochu raised concerns about aligning the values of embodied intelligent systems with human values to ensure harmonious coexistence [3][7] - He pointed out that safety becomes increasingly critical when integrating large models into embodied systems, particularly regarding perception, reasoning, and interaction with robots and humans [1][5]