We have now reached an important threshold.
An artificial system may have an objective.
It may adapt its behaviour.
It may preserve its internal state.
It may even have something like a stake in continued operation.
But none of this can be understood in isolation.
For something to matter to a system, there must be a world in which differences can affect it.
So the next question is:
What kind of world would a machine need in order for things to matter to it?
A world is more than information
An organism is surrounded by information.
But its environment is not merely a stream of data.
Some differences have consequences for the organism.
Light changes growth.
Temperature alters physiological organisation.
Food supports metabolism.
A predator threatens survival.
The organism's world is therefore a field of differential possibilities.
It encounters conditions that can help or hinder its continued organisation.
A world matters because the system is coupled to it.
Coupling creates significance
Suppose a machine has sensors.
It detects temperature, pressure, available energy, damage or changes in its surroundings.
That gives it information.
But information becomes potentially value-relevant only when those differences affect the machine's own organisation.
If declining temperature threatens a component, the difference has a consequence.
If reduced energy limits future operation, the difference matters in a stronger sense.
If restoring the condition changes the system's prospects, feedback is established.
We therefore have:
world → difference → consequence → response → altered world
A value-sensitive system is part of this loop.
The environment pushes back
This is something present LLMs largely lack.
A language model can receive a prompt and generate a response.
Its symbolic environment can be extraordinarily rich.
But it is not thereby exposed to a world that independently resists or sustains its continued organisation.
The human user supplies the stakes.
The external world supplies the consequences.
The model processes the symbolic representation.
A system with intrinsic mattering would require a tighter coupling between what happens in its environment and what happens to it.
Embodiment is one possibility
This need not mean a humanoid body.
A machine could be embodied in many ways.
It might have sensors.
Actuators.
Energy requirements.
Physical components that wear out.
A bounded location.
A persistent operational environment.
The important feature would not be resemblance to an animal.
It would be dependence upon conditions outside the system.
A world becomes significant when the system cannot simply treat it as information detached from its own continued organisation.
But embodiment alone is not enough
A camera is embodied.
A robot can sense its surroundings.
A thermostat is coupled to temperature.
None thereby demonstrates mattering.
The crucial question is what the coupling does to the organisation of the system.
A thermostat responds because its control architecture is designed around a target.
A robot can navigate around obstacles without those obstacles becoming intrinsic concerns.
So we need more than sensors and feedback.
We need a system whose organisation is differentially dependent upon what it encounters.
A machine needs something to lose
This gives us a useful formulation.
For an artificial world to matter, the system must have something that can be gained or lost for the system itself.
Resources.
Integrity.
Continuity.
Capabilities.
Relationships.
Possibilities.
The list could be very different from an organism's.
What matters is that the system's future organisation depends differentially upon them.
Without that, the environment remains a source of information rather than a world of stakes.
Temporal continuity
This also explains why time matters.
A system with no persistent history has little basis for distinguishing its present from its future.
A persistent system can accumulate consequences.
Its condition today affects what it can do tomorrow.
Some changes become irreversible.
Some resources become depleted.
Some relationships develop histories.
Now the environment can matter through time.
A world is not simply what the system senses now.
It is what changes the trajectory of its own organisation.
A world can be relational
There is no reason the relevant world must be purely physical.
For a socially organised machine, other agents could become part of its environment.
Another system may provide resources.
Block a goal.
Restore damaged capabilities.
Change its possibilities.
A relationship could become consequential.
This gives us a path from:
world
to:
social world.
But the social transition requires one more step.
Something does not become socially significant merely because it affects the system.
It must become part of a recurrent relational organisation.
The difference between reaction and mattering
This distinction is worth protecting.
A machine can react to a change.
A system can even change its behaviour in predictable ways.
But mattering requires a stronger relation:
the difference changes the system's own possibilities in ways that are differentially significant to its organisation.
That is why mattering is not synonymous with responsiveness.
A system can be responsive without having stakes.
A new criterion
We can now refine the criterion proposed in the first post.
Something matters to a system when:
its presence, absence or alteration changes the system's own organisation or future possibilities in a way to which the system is intrinsically responsive.
"Intrinsically" is doing important work.
The response cannot simply be the execution of an externally specified rule.
We are looking for a feedback architecture in which the system's own organisation helps determine what counts as better or worse for it.
What would such a machine look like?
We should resist designing it prematurely.
But conceptually, we can imagine a system with:
persistent identity;
resource dependence;
vulnerability;
self-maintenance;
memory;
action in an environment;
consequences that alter its future organisation.
Such a system would have something closer to a world than an LLM does.
Its environment would no longer be merely the source of inputs.
It would be part of the conditions of its existence.
The beginning of artificial value
At this point we can see where artificial value might begin.
Not with a declaration:
"I value X."
Not with a list of preferences.
Not with a reward function alone.
But when the organisation of the system becomes such that:
some states sustain its possibilities and others undermine them.
Now something can matter.
A machine has acquired the beginnings of its own value landscape.
But a world of value is not yet a social world
This is where our earlier topology becomes relevant again.
An organism can have value without being social.
A bacterium responds to its environment.
An animal can have complex value-sensitive organisation.
Sociality requires more.
Other systems must become recurrently consequential to one another.
Relations must stabilise.
Dependencies develop.
Possibilities become shared.
So if we want to understand artificial sociality, we need to ask the next question:
When can other systems begin to matter to the machine?
That is the transition from an artificial value system to an artificial social topology.
The next question
We have therefore moved from:
objective
to:
stake
to:
continuation
to:
world.
The next step is the most recognisably social one.
Suppose an artificial system encounters another agent whose actions affect its possibilities repeatedly over time.
Could that other agent become more than an environmental variable?
Could it become someone who matters?
When Others Begin to Matter
No comments:
Post a Comment