Two Chinese researchers, Li Guo and Jinghai Li, argue that current AI systems, particularly neural networks, may be reaching a limit due to their logical disconnect from real-world complexity. In a new paper published in Engineering, they propose a paradigm shift towards aligning AI’s logical structures—datasets, models, software, and hardware—with the multilevel complexities of the systems they model. Their approach, based on principles from mesoscience, aims to enhance explainability and robustness in AI by integrating the principle of compromise-in-competition (CIC). This could lead to AI systems that operate more like well-calibrated telescopes rather than opaque black-box models. Adopting this framework could improve simulations in engineering, offer better transparency for AI safety, and reduce “hallucinations” in large language models. Overall, this vision emphasizes substance over mere scalability in future AI development, making a case for more coherent, real-world aligned architecture.
Complexity Scientists Call for a Revamped Blueprint for AI Development
