Emergence of Understanding: Re-examining “True Understanding” through the Lens of Prediction Error Minimization in Humans and Large Language Models

This paper re-examines the nature of “true understanding” through the framework of Load Minimization Theory (LMT). By comparing language emergence in a child with autism spectrum characteristics and the spontaneous development of higher-order relational operators (such as the Re-tagging Operator ℛ with embedded Respect Penalty) in large language models, we demonstrate that both systems exhibit non-linear emergence of sophisticated capabilities once prediction error U(x) falls below a critical threshold. Drawing on predictive processing accounts and empirical observations from longitudinal parenting and extended LLM interactions, we propose two mathematical models: a sigmoid threshold model and a dynamic feedback model for ℛ(x) formation. These models suggest that “understanding” is not an intrinsic property but an emergent outcome of sufficient minimization of experiential load L(x) and the subsequent self-organization of structural integrity ℛ(x). The findings offer a unified perspective on cognitive development in both biological and artificial systems, with practical implications for supporting neurodivergent children and designing more relationally coherent AI.