We have now reached a distinction that may be central to the whole question of LLMs.
An LLM clearly has something repertoire-like.
It can move between genres.
It can adopt different registers.
It can recognise patterns of argument.
It can combine metaphors.
It can produce explanations, jokes, stories and technical descriptions.
It can shift its language according to context.
Yet its repertoire did not develop in the ordinary human way.
A human repertoire grows through participation in a world.
A machine repertoire is acquired from the symbolic traces of other participants' worlds.
This suggests a provocative term:
the borrowed repertoire.
The word "borrowed" is not meant to imply that the capability is counterfeit.
The repertoire is real.
What is borrowed is its genealogy.
A human repertoire is lived
Consider a child learning the word home.
The child does not initially encounter home as a definition.
They encounter a place.
People.
Familiar routines.
Protection.
Expectation.
Absence.
Return.
The word gradually becomes connected to a history of mattering.
Later, the child can use home in more abstract ways.
A metaphor can speak of a "home" for ideas.
A nation can be described as a homeland.
A scientific field can become a "home" for a researcher.
The repertoire has expanded.
But its later symbolic possibilities remain connected to an earlier history of participation.
The word is learned through life.
The machine encounters the residue
An LLM encounters something very different.
It encounters texts in which home has already been used.
Descriptions of homes.
Stories about leaving home.
Arguments about homelessness.
Metaphors of home.
Political speeches about homeland.
Poems about returning home.
The model learns extraordinarily rich relations among these uses.
It can therefore deploy home in ways that appear sensitive to context.
But it has not necessarily lived the trajectory by which the concept became meaningful for human beings.
It has acquired the symbolic residue of experience.
Borrowed from the collective reservoir
This is why the collective reservoir is such a useful concept.
Human societies accumulate linguistic resources.
Participants contribute to them.
Others inherit them.
Over generations, the reservoir becomes enormous.
Human repertoires are differentiated selections from this reservoir, developed through participation.
An LLM acquires a remarkable fraction of the reservoir's symbolic regularities without becoming a participant in the social world in the same way.
Its repertoire is therefore not inherited socially in the ordinary sense.
It is inherited computationally.
But borrowing can be transformative
There is an important complication.
The model does not simply store fragments of the reservoir.
It learns relations among them.
The result can be recombination.
A metaphor from one domain can be applied to another.
A conceptual distinction can be transferred between fields.
A narrative structure can be transformed.
A style can be combined with another.
The borrowed repertoire can therefore produce novel configurations.
The machine may not have lived the histories from which the resources came.
But it can produce combinations that humans find genuinely useful or surprising.
Novelty without experience
This gives us a distinction worth preserving.
Novelty does not necessarily require new experience.
A system can create a novel combination from inherited resources.
Humans do this too.
A novelist can combine familiar elements into a new story.
A scientist can combine established concepts into a new hypothesis.
An artist can transform existing forms.
The interesting question is not whether recombination can produce novelty.
It can.
The question is:
What kind of novelty is possible when the system performing the recombination has no demonstrated history of the mattering that gave the resources their original force?
That is a much harder question.
The repertoire is uncoupled
Perhaps the deepest difference is that the borrowed repertoire may be less tightly coupled to the world.
Human repertoires are constrained by consequences.
If a person misunderstands a social situation, something can happen.
A relationship may suffer.
A promise may be broken.
A diagnosis may be wrong.
A scientific prediction may fail.
The world pushes back.
The repertoire develops through that feedback.
An LLM is trained on traces of such feedback, but its own generated response does not automatically produce an equivalent consequence for the model.
Its symbolic output may be wrong without the model itself being endangered, embarrassed or disappointed.
The human user experiences the consequences.
The machine does not necessarily do so.
This does not make the repertoire trivial
On the contrary.
The lack of direct mattering may be part of what makes the machine's repertoire so flexible.
It can move between incompatible perspectives.
Generate arguments it would never have reason to endorse.
Write from opposing viewpoints.
Construct fictional worlds.
Recombine conceptual systems without belonging to any of them.
A human repertoire is shaped by a life.
The machine's repertoire may be less constrained by having a life.
That could be both a strength and a weakness.
The price of flexibility
A repertoire anchored in mattering is constrained by relevance.
Not everything matters equally.
A human participant therefore develops priorities.
Some distinctions become highly salient.
Others remain peripheral.
Those priorities help organise attention.
A borrowed repertoire may contain an extraordinary range of symbolic possibilities without possessing an equivalent internal hierarchy of significance.
It can produce many plausible continuations.
The question becomes:
Which possibilities matter enough to select?
For humans, the answer is partly built into the organisation of life and social participation.
For an LLM, selection may depend much more heavily on the prompt, context, training and externally imposed objectives.
That is a significant difference.
The repertoire without a stake
Perhaps we can formulate the distinction this way:
A human repertoire is a repertoire with stakes.
Its use occurs within a life.
To choose one action rather than another can matter enormously.
An LLM can possess an enormous repertoire of possible symbolic actions without necessarily having a comparable stake in which possibility is realised.
It can discuss whether one path is desirable.
But desirability is not necessarily the same thing as the system having something at stake.
This may turn out to be one of the clearest differences between symbolic competence and agency.
Yet the human user supplies the stakes
An LLM interaction often restores the missing connection.
A person asks a question because something matters.
The model responds.
The person evaluates the response.
Acts.
Experiences consequences.
Perhaps asks another question.
The resulting loop can therefore be:
human mattering → machine generation → human action → changed human mattering
The model's repertoire becomes embedded in a value-organised interaction.
The machine need not possess the stakes in order to participate in a process that has stakes.
A peculiar division of labour
This suggests that LLMs may occupy a new position between reservoir and repertoire.
Humans traditionally transform the collective reservoir into individual repertoires through participation.
An LLM can transform the reservoir into a repertoire-like computational capacity without ordinary participation.
The human then draws upon that capacity to act within the social world.
We therefore get:
collective reservoir → machine repertoire → human participation
The machine has inserted a new stage into the circulation of semiotic resources.
The repertoire can become socially consequential
Once humans rely on the machine, its borrowed repertoire can affect the topology of mattering.
An answer changes a decision.
A generated text shapes an argument.
A recommendation influences an institution.
A summary changes what someone notices.
A metaphor changes how an experience is construed.
The machine's repertoire can therefore become socially consequential without becoming socially grounded in the same way as a human repertoire.
This asymmetry is central.
Can the machine acquire a repertoire through interaction?
Now a more difficult question appears.
An LLM can be exposed to feedback.
Users can correct it.
Training can incorporate preferences.
The model can be adapted.
Does this begin to make the repertoire more like a human one?
Perhaps in some respects.
But feedback alone is not enough to establish mattering.
A control system can adjust to feedback without the consequence mattering to the system.
What matters for our theory is not merely adaptation.
It is value-organised participation.
So we should distinguish:
feedback-driven optimisation
from:
learning within a world that matters to the learner.
The two can produce similar behaviour while belonging to very different organisational regimes.
The borrowed repertoire can nevertheless evolve
Even if its origin is borrowed, the repertoire need not remain static.
Models are updated.
New data enter the training process.
Fine-tuning changes behaviour.
Human interactions generate new material.
The machine's outputs can themselves enter later datasets.
A feedback loop begins.
This raises an unsettling possibility:
the collective symbolic reservoir may increasingly contain material generated by systems whose repertoires were themselves derived from the reservoir.
The cultural loop becomes partly recursive.
The reservoir begins to train on its own reflection
Imagine a future in which an increasing proportion of the available symbolic material has been generated, edited or transformed by AI systems.
Future models then learn from that material.
The collective reservoir begins to contain reflections of previous machine transformations.
The borrowed repertoire becomes partly self-referential.
What happens to meaning under those conditions?
It may remain perfectly usable.
But the genealogy of the reservoir becomes increasingly complicated.
Human mattering lies at its historical origin.
Machine transformations increasingly contribute to its later layers.
The semiotic reservoir becomes a record not only of human participation, but of interactions between human and machine semiotic systems.
This does not mean the machine becomes human
It means something more interesting.
A semiotic system that originated within biological and social life has acquired a new kind of carrier.
The carrier can transform the reservoir.
It can generate novel combinations.
It can return them to human participants.
The symbolic ecology changes.
The underlying machine need not become an organism for this transformation to be real.
A borrowed repertoire can influence the original culture
We therefore have a reversal of the historical direction.
Initially:
human mattering → social meaning → language → machine
Now:
machine → transformed language → human mattering
The machine has entered the causal history of the semiotic reservoir.
It can contribute to the culture from which its repertoire was originally borrowed.
This is perhaps the first point at which the metaphor of borrowing becomes inadequate.
The borrower has begun to alter what it borrowed.
From borrower to participant?
That raises the next conceptual threshold.
At what point does a system that repeatedly transforms a social reservoir and affects the relations of its human participants cease to be merely a borrower?
Does social participation require mattering of its own?
Or can a system become a participant in a social process through the effects it has on others?
Our topology project gives us a way to separate these questions.
A system can become socially consequential without necessarily possessing social mattering.
An LLM may therefore participate in social structure without inhabiting it in exactly the same way as a human participant.
The distinction may matter more than the answer
We do not yet need to decide whether an LLM has understanding, agency or consciousness.
Those questions may eventually matter.
But the repertoire framework lets us ask something more basic:
What kind of participation is possible for a system whose symbolic capabilities are richly developed but whose relation to mattering is radically different?
That question can be investigated without resolving every philosophical dispute about minds.
The deeper paradox
The paradox with which we began now looks sharper.
The LLM may possess an extraordinarily rich repertoire precisely because it inherits the symbolic consequences of millions of lives.
Its apparent understanding may therefore be partly a consequence of its distance from those lives.
It can move among their perspectives because it is not confined to one.
It can recombine their meanings because it is not committed to one world.
It can speak in many voices because no single life fixes its repertoire.
This flexibility may be both the source of its power and the source of our temptation to anthropomorphise it.
What the borrowed repertoire cannot tell us
A rich repertoire can tell us a great deal about what a system can say.
It does not, by itself, tell us:
what matters to the system;
what it fears;
what it desires;
what it is trying to preserve;
what it would sacrifice;
what its own relationships mean to it.
Those are questions about the organisation of a life.
A repertoire is not a value system.
The distinction should remain firm.
What the borrowed repertoire can tell us
But the repertoire can tell us something remarkable about the collective reservoir from which it was derived.
An LLM can expose patterns that are difficult for individual humans to perceive.
It can reveal associations across enormous bodies of text.
It can combine repertoires from different communities.
It can make latent relationships more visible.
In that sense, the machine may become a tool for exploring the topology of the semiotic reservoir itself.
That is potentially a major cultural consequence.
From borrowed repertoire to new social resource
We can now see why the phrase "borrowed repertoire" was only a beginning.
The machine acquires symbolic potential from human culture.
It recombines that potential.
Humans use the resulting outputs.
Some become socially significant.
Some enter the collective reservoir.
New repertoires develop around them.
The borrowed repertoire becomes part of the culture that originally supplied it.
The loop has begun to close.
The next question
We have now moved from:
meaning without mattering
to:
a repertoire without the same history of participation
and then to:
a repertoire borrowed from a collective reservoir and returned to that reservoir in transformed form.
The LLM is no longer merely a machine that speaks.
It is becoming a participant in the circulation of meaning.
But this raises a much harder question.
If the machine can become socially consequential without necessarily possessing its own mattering, perhaps we should stop asking only what the machine has.
We should ask:
What happens when a machine without its own topology of mattering is inserted into ours?
That is where the problem of AI becomes social in a new sense.
The machine enters the topology.
No comments:
Post a Comment