We have arrived at an unusual possibility.
An LLM can generate language without obviously having passed through the biological and social history through which human language emerged.
It can manipulate words.
Patterns.
Genres.
Metaphors.
Arguments.
Narratives.
But what, exactly, has it acquired?
We have called its capacity a repertoire.
Yet that word now needs to be examined more carefully.
A human repertoire develops through participation in a social world.
An LLM acquires its linguistic capacities through exposure to the products of that world.
The distinction may turn out to be fundamental.
Language is older than the machine
Human language did not begin as an abstract symbolic system waiting to be implemented on any sufficiently powerful substrate.
It emerged within communities of organisms.
People needed one another.
Depended upon one another.
Cooperated.
Competed.
Raised children.
Protected resources.
Maintained relationships.
Negotiated boundaries.
Passed knowledge between generations.
Language grew within these conditions.
Its meanings are therefore saturated with the history of social mattering.
Words do not merely refer to things.
They organise relations.
They classify roles.
Express obligations.
Construct identities.
Record histories.
Imagine futures.
Human language is therefore language after the social only in the sense that semiosis comes after social organisation in Halliday's sequence.
For an LLM, something strange happens.
Language is acquired without the same route through the social.
The collective reservoir
We have already distinguished the collective reservoir from the individual repertoire.
The collective reservoir is enormous.
It contains the linguistic possibilities of a culture:
words,
constructions,
genres,
metaphors,
narratives,
categories,
distinctions,
ways of arguing,
ways of imagining.
No individual possesses all of it.
Human participants acquire particular repertoires through participation.
A child learns language by living among speakers.
A scientist acquires a disciplinary repertoire by entering a community.
A novelist develops a literary repertoire by reading and writing.
The repertoire is therefore historically embodied participation in the reservoir.
The machine receives the sediment
An LLM encounters something different.
It does not encounter the social world directly in the way a child does.
It encounters enormous quantities of linguistic traces of that world.
Books.
Articles.
Conversations.
Documentation.
Web pages.
Stories.
Arguments.
Descriptions.
The training data are a kind of sediment.
The social world has already been transformed into language.
The machine receives the transformed product.
This suggests a striking formulation:
The human repertoire is formed through participation in the social world; the LLM repertoire is formed from the sediment left by that participation in the symbolic reservoir.
That is not yet a judgement about which is better.
It is a difference in origin.
A repertoire without a life?
We should therefore ask whether "repertoire" is really the right word for the machine.
The analogy is obvious.
An LLM can produce many styles and genres.
It can move between technical prose and conversation.
It can imitate literary forms.
It can combine conceptual traditions.
It can adjust its language to a context.
These look repertoire-like.
But a human repertoire includes more than the ability to produce forms.
It includes knowledge of how those forms matter in situations of participation.
A doctor does not merely know medical language.
The doctor knows what a patient's condition means for treatment, responsibility and care.
A friend does not merely know the language of friendship.
Friendship is part of the friend's lived relations.
The repertoire is therefore coupled to a history of mattering.
That coupling may be absent from the machine.
The illusion of experience
This helps explain one of the strange effects of LLM conversation.
The language can be so well organised that it appears to come from experience.
An answer about grief may sound compassionate.
An answer about friendship may sound intimate.
An answer about fear may sound as though it comes from someone who has been afraid.
But the linguistic form is not itself evidence that the speaker has lived the corresponding mattering-relations.
The model can reproduce the symbolic construal of experience without necessarily possessing the experience that produced it.
The distinction is easy to forget because human language normally comes bundled with a human life.
The machine separates them.
Language after the social
This may be the meaning of our title.
Human language is produced within a social world and carries that world within it.
An LLM can operate on language after it has already been socially constituted.
It therefore encounters:
meaning after mattering has been symbolically transformed.
The machine can work with the result.
But it need not reproduce the entire developmental path.
This is a remarkable form of semiotic inheritance.
A library is not a culture
There is an obvious analogy with books.
A library contains an enormous amount of human knowledge.
But a library does not participate in the culture it contains.
An LLM differs because it can actively recombine what it has inherited.
It can answer questions.
Generate formulations.
Produce new texts.
This makes it much more than a passive archive.
Yet the analogy reveals an important point.
The machine's access to meaning may resemble the library's in one respect:
both contain the sediment of human participation without themselves being organisms that participated in producing it.
The difference is that the LLM can dynamically operate upon the sediment.
Recombination as a new capability
This is where the machine becomes genuinely novel.
It can take linguistic resources that originated in very different social contexts and place them into new combinations.
It may connect concepts that no individual writer ever connected.
It may produce analogies across disciplines.
It may generate novel prose.
The novelty can be real even if the underlying resources are inherited.
The question is therefore not whether the machine is creative simply because it has a repertoire.
The deeper question is:
What kind of creativity is possible when recombination is detached from the lived mattering that gave the resources their original significance?
That is a much more interesting problem.
Meaning can survive the loss of origin
Consider a metaphor.
A metaphor may have emerged from someone's experience.
It is repeated.
Others adopt it.
Eventually it becomes culturally familiar.
A later writer can use it without knowing its history.
The meaning has become portable.
An LLM can go further.
It can combine thousands of such inherited resources without having lived the experiences that generated them.
Meaning can therefore become progressively detached from its original situations of mattering.
The machine is a particularly powerful agent of that detachment.
Yet the meaning is not empty
We should not conclude that machine-generated language is therefore meaningless.
The words still enter human semiotic systems.
A metaphor can illuminate something for a reader.
An explanation can help someone understand.
A story can move someone.
A phrase can change a decision.
The output can matter because humans matter.
The machine participates in meaning through the effects its language has within human systems.
Its lack of demonstrated self-mattering does not make its outputs semantically inert.
A new asymmetry
This creates an unusual asymmetry.
For human communication:
speaker matters → says something meaningful → hearer responds
For an LLM interaction, we may have:
human matters → asks meaningful question → machine generates language → human responds
The human brings the mattering into the exchange.
The machine supplies extraordinary semiotic responsiveness.
The asymmetry may be easy to overlook because the conversation itself is structurally similar to human dialogue.
But the underlying organisation can be very different.
The social relation surrounds the model
This suggests that the missing social level may not be absent from the interaction.
It may be absent from the machine.
A person asks because something matters.
The question belongs to their life.
The answer may influence their next action.
The exchange can therefore be socially consequential.
But the social relation may be asymmetrical:
the human participates in a topology of mattering; the model participates in the semiotic consequences of that topology.
That distinction may become central to understanding AI as a social technology.
Can the repertoire be borrowed?
We can now return to the question that ended the previous post.
Can a repertoire be acquired from the reservoir without being formed through participation?
For an LLM, apparently something like this happens.
It acquires capacities for linguistic production without childhood, friendship, professional apprenticeship or political affiliation.
Its repertoire is therefore not a history of participation in the ordinary sense.
It is a learned organisation of symbolic possibilities derived from the traces of participation left in language.
Perhaps we should call this a borrowed repertoire.
The phrase is provisional, but useful.
Borrowed does not mean fake
"Borrowed repertoire" should not imply that the model's capabilities are counterfeit.
They are real.
The model can genuinely generate the corresponding forms.
What is borrowed is the history of significance embedded in those forms.
A model can use a metaphor whose original force came from human experience.
It can reproduce a social distinction whose significance arose from centuries of practice.
It can generate a style whose cultural associations it never lived.
The repertoire is real as a capability.
Its genealogy is different.
The machine has no childhood — but it has training
This difference may also illuminate "learning".
Human learning transforms an organism's repertoire through ongoing interaction with a world that matters.
Machine training changes the model's internal organisation through exposure to data.
Both involve adaptation.
But the systems adapting are different.
The human learner is embedded in biological and social feedback.
The model is adjusted through an engineered training process.
The resulting capabilities may sometimes look similar.
Their conditions of development are not.
Training is a new kind of semiotic inheritance
This may be one of the genuinely new phenomena introduced by LLMs.
Cultural knowledge can now be inherited by a system that does not enter the culture as a biological participant.
The inheritance is:
social world → symbolic record → training process → machine repertoire.
The machine receives a compressed, transformed history of human semiosis.
It does not need to reenact the original social relations to acquire some of their symbolic consequences.
That is an extraordinary technological development.
But something is missing from the loop
The missing element becomes clearer if we compare the two learning processes.
Human:
world matters → participation → experience → repertoire → action → changed world
LLM:
symbolic traces → training → repertoire-like capability → generated language
The second loop can reconnect to humans:
generated language → human action → changed social world
But the model's own value-sensitive participation remains uncertain.
So the complete system may be better represented as:
human mattering → symbolic reservoir → machine → symbolic output → human mattering
The machine occupies the middle of the loop.
It may be a powerful mediator.
It need not be a source of the original value.
The machine as semiotic intermediary
This may therefore be a better way to think about an LLM than as a simulated human mind.
It is a semiotic intermediary.
It receives symbolic resources.
Transforms them.
Recombines them.
Returns them to a social world.
Its distinctive capability lies in operating on the symbolic reservoir at extraordinary scale.
The question of whether it has mind, consciousness or selfhood remains separate.
We can study what it does without resolving those questions first.
And this may explain the uncanny quality
Why can interaction with an LLM sometimes feel strangely personal?
Because human conversation normally bundles several things together:
language,
intention,
experience,
mattering,
social relationship.
The LLM can reproduce the linguistic component with remarkable flexibility.
We naturally supply the rest.
The conversation therefore creates a kind of semiotic completion.
The language suggests a participant.
Our social repertoires complete the picture.
That does not prove that the machine is empty.
Nor does it prove that it is a person.
It shows that human language carries expectations about participation with it.
The question we should now ask
We have been calling the machine's capacities a repertoire.
But perhaps what we really need to understand is the relation between:
a collective reservoir,
a human repertoire,
and
a machine's learned symbolic potential.
They may look similar at the level of language.
But their origins and relations to mattering differ.
That difference may explain both the extraordinary power of LLMs and some of the peculiar limitations we encounter when we treat them as conversational partners.
The next problem
So we have now moved one step beyond the original question.
It is not merely:
Can there be meaning without mattering?
It appears that symbolic meaning can at least be re-instantiated and manipulated by a system whose own relation to mattering is radically different from ours.
But that raises a new question.
Perhaps what matters is not simply whether the machine has a repertoire.
Perhaps what matters is how a repertoire is anchored in participation.
Can an LLM's learned symbolic potential function as a genuine repertoire without an embodied history of social participation?
Or does it constitute something new — a repertoire-like capacity that exists between the collective reservoir and the human participants who use it?
That is the problem we now need to explore.
What kind of repertoire does a machine have?
No comments:
Post a Comment