This paper discusses world models, algorithmic representations of the real-world environment in which biological agents experience and interact. With the growing need to develop virtual agents with artificial (general) intelligence, world models have recently attracted significant attention, sparking debates about their definition, construction, utilization, and evaluation methods. Drawing on the imagination of the science fiction novel Dune and the concept of "hypothetical thinking" from the psychological literature, this paper critiques various world modeling approaches and argues that the primary goal of world models is to simulate all feasible real-world possibilities for purposeful reasoning and action. Building on this critique, we propose a novel architecture for general world models based on hierarchical, multilevel, mixed continuous/discrete representations and a generative self-supervised learning framework. We then present a vision for physical, functional, and nested (PAN) AGI systems enabled by these models.
