The physical universe may be considered a hypercomputer — a system capable of computations beyond what our artificial computers can achieve. This perspective offers a compelling explanation for why embodied cognition is more efficient at knowledge acquisition than purely simulated entities. When an agent is physically embodied in our world, it can leverage the universe’s inherent computational capabilities rather than having to simulate them.
Physical embodiment enables agents to learn the “causal microstructure” of reality through direct interaction. This includes understanding fundamental concepts like gravity, object permanence, and the behavior of surfaces and materials. The physical world has a “smoothness property” — a fundamental continuity punctuated by meaningful discontinuities. This structure helps agents learn foundational concepts like containment, support, and spatial relationships naturally and efficiently.
Consider trying to simulate these physics-based interactions in a computer. To accurately model even simple physical interactions would require enormous computational resources. The simulation would need to account for countless variables and their complex interactions. In contrast, an embodied agent can simply interact with the real world, where these computations are effectively “free” — they’re handled by the universe itself.
Source: There Are Monsters in Your LLM.