Reading 07 · World Models
Revisiting the Notion of “State”
Defining state correctly is the foundation of any reliable autonomous system. The physical world is deterministic; our models treat it statistically — and the gap is where safety lives or dies.
The thesis
State should be understood not as a single complete description but as a network of interconnected models — with unresolved questions of determinism, observability and the sufficiency of the data we train on.
The argument, in five moves
What we mean when we say “state”
Determinism vs. statistics
The physical world is deterministic, yet current approaches model it statistically. Safety-critical applications — autonomous surgery, for one — need deterministic guarantees, not probabilities.
A network, not a single state
State must span perspectives: the manipulator's own state, the environment's context such as water temperature, and the system's dynamics. Each subcontext calls for its own state semantics.
Observability is a modelling choice
Partial observability stems from incomplete data collection and modelling decisions, not inherent physical limits. Yet capturing every detail does not guarantee a better model either.
We can't be sure the data is enough
Building models solely from captured data leaves fundamental uncertainty about completeness, failure modes, and whether critical dynamics were missed entirely.
Coordinating many perspectives
Networked state models must integrate local agent viewpoints with shared and global ones — individual cars versus central traffic coordination — and this remains largely an open problem.
State: the underlying reality of the world — complete in principle, but never directly visible.
The Nyāya lens · अदृष्ट
Indian thought long reasoned about adṛṣṭa — the unseen determinants that shape an outcome without ever appearing in it. World models rediscover this as the frame problem: the state that matters is complete in principle yet never fully visible, and wisdom lies in knowing what we have chosen not to see.
Distilled from the original essay on engkraft.com.