Modern AI models are typically considered specialised, functioning well only in specific situations for which they have been trained. In contrast, the human brain is seen as general, capable of handling a wide range of situations. This is because our mammalian ancestors, dating back a hundred million years, evolved to use imagination as a tool for survival. They employed strategies such as vicarious trial-and-error, counterfactual learning (using ‘what-if’ scenarios to learn from future possibilities rather than just historical data), and episodic memory to outplan predators.
For instance, an ancient relative of today’s cats sitting in a tree could survey the branches and plan its next moves for a more effective jump. These ancient animals learned quickly and exhibited better motor skills than their vertebrate predecessors. While most vertebrates (such as modern fish and lizards) can move rapidly, recall patterns, and track time, they don’t plan their movements or imagine using their limbs before acting.
In essence, modern mammals like cats and humans can simulate internally. The neocortex enables this predictive power, reconstructing the three-dimensional world around an animal to anticipate exactly what will happen next as objects and other animals move within it.
Various subregions of the neocortex represent different aspects of the external environment based on the information they gather. Our understanding of how the neocortex manages a multitude of simulations to create a vivid, three-dimensional world filled with perceivable objects remains limited. The neocortex may be inherently designed to make certain assumptions.
The focus of AI research has been to develop more general AI systems, akin to the neocortex of a sophisticated mammal. The challenge lies in how we can mimic the biological innovation of the imaginarium in creating AI technologies. To reverse-engineer the imaginarium, we need to comprehend the mechanisms of predictive modelling in the neocortex. These mechanisms may be significantly more specific: the neocortex could make narrow, explicit assumptions about the world, and it is likely these very assumptions that grant it its general capabilities. The specificity of these assumptions may benefit an animal by providing a foundational world model.Â
For example, the neocortex may be naturally programmed to assume that incoming sensory information—whether visual, auditory, or tactile—represents three-dimensional objects that exist independently of ourselves and can move on their own. Consequently, it does not need to learn about concepts such as space, time, or the distinction between self and others. Instead, it aims to interpret all incoming sensory information by presuming it arises from a three-dimensional world unfolding over time.Â
It seems that the neocortex learns to address the search problem by smartly selecting which paths to simulate and deciding when to simulate them. These simulations allow mammals to experiment through trial and error and to choose which path to follow based on imagined outcomes. Learning through “what-if” scenarios, known as counterfactual learning, is the most sophisticated solution to the credit assignment problem, enabling mammals to attribute credit based on causal connections. The simulations also facilitate episodic memory, allowing the recollection of past events and actions and thereby enabling the adjustment of future behaviours.
Current AIs lack a world model and cannot use model-based reinforcement learning to advance. Many leading AI researchers and practitioners have highlighted this reality, even as the broader industry continues to funnel substantial capital into scaling fragile models built on deep learning and transformer frameworks. There have been only a limited number of efforts to tackle the absence of a world model in AI, with one approach focusing on integrating causality. I will discuss this issue in several podcast episodes scheduled for 2026, as well as in this blog.