A small submarine moves through a dark ocean.
The pilot inside cannot directly touch the ocean, see all its dangers, or know with certainty what lies beyond the hull.
Instead, the pilot receives signals: sonar echoes, pressure changes, vibrations, turbulence, instrument readings.
The question is no longer only: What caused my sensations? It is also: What should I do next?
The hull is a boundary β but not a wall of isolation.
The submarine hull separates the inside of the vessel from the outside ocean. But it also allows contact between them.
This is the intuitive idea behind a Markov blanket: a boundary that separates internal states from external states while mediating their interaction through sensory and active states.
rocks, whales, currents
sonar, pressure, vibration
beliefs, expectations, model
turning, pinging, ascending
Perception and action form a loop.
The submarine is not a passive camera. It actively samples the world. It can send a sonar ping, slow down, turn away, or ascend.
External states
Things outside: rock walls, whales, open water, wrecks, currents, pressure zones.
Sensory states
Signals arriving at the blanket: sonar echoes, pressure, vibration, turbulence.
Internal states
The model inside: beliefs about what is out there and what will happen next.
Active states
Actions through the blanket: steering, ascending, slowing, sending sonar.
Prediction error
A mismatch between expected signals and incoming signals.
Active inference
Update beliefs and act to reduce uncertainty and remain viable.
Interactive: what should the submarine do?
Current posterior belief
active step: choose an action to change future sensory evidence
Choose an action
Each action changes what the submarine samples next. In active inference, action is part of inference.
Why this is active inference
In the Windowless Room demo, the being updated its beliefs about hidden causes.
Here, the submarine does something more. It acts. By turning, pinging, slowing, or ascending, it changes the sensory stream it will receive next.
The system therefore reduces uncertainty in two linked ways:
- by changing internal beliefs to better explain sensory evidence;
- by changing the world/body relation so that future evidence becomes safer, clearer, or more predictable.
The core loop
This is why active inference is not just a theory of perception. It is a theory of embodied, self-maintaining systems.
Everyday active inference
The submarine is only a teaching image. The same logic appears throughout ordinary life.
You slow down, turn your head, listen harder, and move to create better sensory evidence.
You do not just think harder. You lift cushions, retrace steps, call it, and sample the world.
You look, listen, pause, and adjust your movement to reduce uncertainty about cars.
You may freeze, listen again, switch on a light, or walk toward the sound. Action tests hypotheses.
Your hand moves around the object to infer its shape. Perception is guided exploration.
You lean forward, ask βsorry?β, watch the face, and use action to clarify uncertain speech.
Final teaching point
Living systems do not only infer what is out there. They act to sample the world, reduce uncertainty, and keep themselves viable.
The submarine helps us see the logic of active inference. It is enclosed, but not isolated. It is separated from the ocean by a boundary, but that boundary is exactly what allows perception and action.
The Markov blanket is therefore not simply a barrier. It is the structured interface through which a system maintains itself as a system.
From room to submarine
The Windowless Room shows Bayesian perception: hidden causes must be inferred from sensory effects.
The Submarine adds embodiment and action: the system acts through its boundary to generate better evidence and remain within safe states.