Agents as Webs of Beliefs (www.lesswrong.com)

A new conceptual model titled "belief webs" has been introduced to understand intelligent agents by integrating beliefs, goals, and actions into a unified framework. This model draws from active inference, agent foundations, and machine learning, and addresses the common issue of inconsistencies within an agent's beliefs. Unlike traditional frameworks that rely on a single probability distribution, belief webs account for local consistency among beliefs while acknowledging global inconsistencies. The model utilizes probabilistic dependency graphs (PDGs) and Garrabrant induction to formalize how local beliefs are interconnected, forming a multi-layered structure that aids in hierarchical concept formation. The significance of this model lies in its potential to enhance our understanding of agent behavior by merging cognitive processes with decision-making. It introduces the self-predictive model, which posits that actions arise from a subset of beliefs, linking thought and action in real-world scenarios. This model emphasizes the need to manage beliefs as essential components of agency, providing insights into phenomena like procrastination and internal conflict. Moreover, the idea of "drives" suggests a new approach to navigating belief systems without fixing goals artificially, allowing for a more coherent integration of desires and empirical beliefs. This framework could reshape our approach to AI systems, moving toward more adaptable and context-sensitive agents.