A conventional engineering instinct begins with preparation.
Before the machine acts, give it a map. Before it encounters an object, give it an object model. Before it enters a room, describe the room. Before it meets uncertainty, reduce the uncertainty.
The underlying assumption is reasonable: competence improves as the internal model becomes more complete.
From this follows an architecture familiar across robotics and artificial intelligence: represent → model → plan → act.
The world is first translated into an internal form. The internal form is processed. A decision is produced. The decision is returned to the world as action.
This works.
But nature appears to use another possibility as well.
A squirrel does not begin life carrying a high-resolution simulation of every forest it might enter. It does not need the Cartesian coordinates of every branch before jumping. It does not continuously reconstruct every object within view.
It begins with inherited structure, sensing, movement, reflexes, thresholds, and a capacity to change.
Then it enters the world.
The world supplies the rest.
The alternative sequence is therefore closer to: act → couple → receive residual → change only what the residual requires → act again.
This is not merely a smaller version of the first architecture.
It reverses the order.
The World Before the Model
Consider a branch.
Its stiffness is already present in the branch. Its geometry is already present in its shape. Its surface already possesses friction. Gravity is already acting.
The squirrel does not need to reproduce these properties internally before they can influence behaviour.
They participate directly in the encounter.
When paw meets branch, the branch responds. When weight shifts, the body responds. When the branch bends, vision, vestibular sensing, muscle tension, grip, and posture change together.
Much of the information required for successful behaviour is therefore not waiting to be reconstructed. It is already physically active.
The world is not merely an external database containing facts about branches. The branch itself is part of the constraint system producing the next state.
Do not compute what physics can supply directly.
A machine that calculates every consequence of a contact while already physically undergoing that contact may be computing something the interaction itself is providing.
This does not mean that computation disappears. It means that computation should be reserved for what the coupling does not already resolve.
The Library Is Outside
The same economy applies to memory.
A complete world is too large to carry. Fortunately, it remains available.
A room can be looked at again. An object can be approached again. A surface can be touched again. A sound can be sampled again. An external archive can be queried again.
The world therefore functions as the highest-resolution store available to an embodied agent.
Internal memory need not duplicate it.
The useful internal residue may instead be compact: where to look, what matters, what usually follows, what changed last time, what relation was learned, what action succeeded, and what should trigger another observation.
The organism or robot does not need the library. It needs an index.
Do not store what the world can be queried for again.
This is not an argument for having no memory. Memory becomes more selective.
Some structure must remain available internally because reacquisition would be too slow, too expensive, impossible, or dangerous.
Recent events may remain in cache. Frequently used relations may remain close. Critical responses may become deeply installed. Rare detail can remain outside.
The architecture is therefore neither amnesia nor encyclopaedia. It is calibrated access.
Attention Before Identification
Something moves.
The animal does not yet know what it is. That is enough.
Attention changes. The head turns. The eyes orient. The sensory system samples again.
Before there is identification, there is a pointer: there.
This is an important ordering.
The system does not need to identify every object in the scene continuously. A residual appears. Attention points toward its source. Then discrimination can become progressively more detailed.
Something in the air. Bird-like. Large. Approaching. Predator-like. Hawk.
But the sequence need not always reach the last step.
If “predator-like” is already sufficient to determine action, further identification may be unnecessary.
The squirrel does not receive additional survival value from performing ornithology while a hawk is approaching.
Recognition can stop when action becomes clear.
Do not identify more than action requires.
The difference is substantial.
A conventional recognition pipeline often treats correct object identity as the goal. A natural action system may treat identity only as an intermediate variable.
The question is not necessarily: What is this?
It may instead be: Does this require a different action?
A pigeon may resolve quickly into: continue chewing.
A hawk may resolve quickly into: seek cover.
A large balloon may initially trigger attention, survive several rounds of discrimination, and then collapse into: irrelevant.
The system has succeeded without producing a complete description of the balloon.
Progressive Recognition
This suggests that perception may operate through progressive refinement rather than complete reconstruction.
A coarse pattern appears. It activates a candidate. Further sensing tests the candidate. If uncertainty remains relevant, more detail is acquired. If the action is already determined, the process stops.
The loop becomes: notice → point → sample → match → act.
If the match is insufficient: sample again.
The system therefore spends resolution where resolution becomes useful.
Much of the visual world can remain computationally dark. Only a local disturbance needs to become bright.
This may be one reason attention is so fundamental.
Attention is not merely a spotlight placed over an already complete internal scene. It may be a request for additional evidence: that part of the world — more information from there.
The world remains available throughout.
Learning Configures Structure
Now a deeper distinction appears.
Light responds. A squirrel responds and learns.
Both are physical systems whose current structure constrains what happens next. But the squirrel's structure can be changed by experience.
A useful abstraction is: current structure + current perturbation → response.
Learning adds another process: residual history → structural change → changed future response.
The important possibility is that successful learning gradually moves work from active correction into structure.
At first, a situation may require attention, hesitation, exploration, and repeated adjustment.
Later, the same situation may produce a fast, quiet response.
The organism does not necessarily solve the same problem more rapidly. Part of the problem may no longer need to be solved. It has become configuration.
Learning turns history into structure.
Inference may often be configured structure responding.
The expert does not always calculate faster than the novice.
Sometimes the expert calculates less.
The structure has changed.
Residual First
This reverses another common intuition.
Intelligence is often associated with continuous internal activity. More processing appears to imply more intelligence.
But in a well-adapted system, successful coupling may require very little activity.
Most interactions can remain quiet. Only a consequential mismatch needs escalation.
A useful loop is: act → couple → residual.
If the residual is negligible, continue.
If the residual crosses an important boundary: attention → additional sensing → discrimination → adjustment.
If the mismatch persists: learning → structural change.
The architecture is therefore residual-driven.
Expensive activity is not the normal state. It is recruited when the cheap path fails.
Only spend internal work when a residual proves that the cheap path is insufficient.
Successful learning should therefore increase the fraction of behaviour that can remain dark.
Movement as Inquiry
Movement is not merely the output of perception. It can also be part of perception.
Turn the head. Move closer. Change angle. Touch the object. Wait. Walk around it.
These actions alter the incoming information.
The agent does not have to infer everything from a single passive snapshot. It can change the relationship between itself and the unknown.
This turns action into experiment.
Instead of asking the internal model to settle every ambiguity, the machine can interrogate the world.
The environment becomes part of the discrimination process.
This is especially important for robots.
A robot need not always solve uncertainty before acting.
Sometimes the cheapest way to know is to move.
A Different Robot
Imagine two robots entering an unfamiliar room.
The first tries to establish a complete scene representation.
Objects are detected. Depth is estimated. Surfaces are segmented. A map is updated. Relations are calculated. The internal model attempts to remain current.
The second robot begins with less.
It detects boundaries, motion, contact, hazards, opportunities, and unexplained residuals.
Most of the room remains unresolved.
When something matters, attention points. The robot samples again. Recognition deepens only as far as necessary. The resulting relation is stored compactly. Detail remains outside.
The first robot attempts to reduce uncertainty globally.
The second permits uncertainty to persist wherever it has no consequence.
That is the central architectural difference.
One asks: How complete can my internal world become?
The other asks: What is the minimum internal structure required for the next competent action?
The Economics of Intelligence
This is partly an energy problem. It is partly a memory problem. It is partly a latency problem. And it is partly a cost problem.
Every additional internal representation has a price.
It must be created. Updated. Stored. Moved. Compared. Kept coherent with a changing world.
A representation that is never used has still consumed resources.
A robot designed for open environments therefore has another option: keep only what earns its cost.
Frequently required structure can remain local. Recent structure can remain cached. Rare detail can remain external. Physical properties can remain physical. Object identity can remain unresolved until identity changes action.
This is not minimalism for aesthetic reasons. It is resource allocation.
Nature operates under energy constraints.
A machine should be allowed to do the same.
The Three Refusals
The architecture can be reduced to three refusals:
Do not compute what physics can supply directly.
Do not store what the world can be queried for again.
Do not identify more than action requires.
Together they reject unnecessary internal work.
They do not reject computation. They do not reject memory. They do not reject representation.
They impose a burden of proof on each.
Why compute this? Why store this? Why identify this further?
If the world already supplies the answer, leave it there.
If the information can be reacquired cheaply, leave it outside.
If action is already determined, stop identifying.
The remaining internal structure may become surprisingly small.
Uncertainty Is Not Failure
Perhaps the deepest difference concerns uncertainty.
Conventional design often tries to remove uncertainty before action.
But an organism in an open world never possesses complete certainty.
It acts anyway.
The alternative architecture therefore treats structured uncertainty as normal.
Unknown detail can remain unknown.
Only uncertainty that threatens successful action needs resolution.
This turns uncertainty from a defect into a resource boundary.
The system asks: Which uncertainty matters now?
Everything else can remain unresolved.
That may be closer to natural intelligence than the attempt to know everything before moving.
The Natural Machine
A natural machine would therefore not attempt to become a miniature universe carrying a complete copy of its surroundings.
It would enter the world lightly.
It would retain compact learned structure.
It would use physics.
It would use movement.
It would use attention.
It would use the world as external detail.
It would cache what matters now.
It would learn from residuals.
And whenever possible, it would remain quiet.
Its intelligence would not be measured by how much of the world it reproduces internally.
It would be measured by how little internal work is required for competent behaviour.
The world would remain outside.
The calibration would remain inside.
And the machine would carry exactly enough of the world to use the world.