A machine speaks.
It can answer questions.
Explain difficult ideas.
Write stories.
Translate.
Summarise.
Argue.
Produce jokes, poems, recipes and code.
It can discuss love, grief, justice, freedom, death and consciousness.
It can even explain what it means to understand.
And yet a strange question remains:
What kind of system is speaking?
We usually approach this question by asking how intelligent the machine is.
Perhaps that is the wrong place to begin.
A more revealing question may be:
What has to matter to a system before its language can mean in the way ours does?
That question takes us back to Halliday.
The long route to meaning
Halliday's taxonomy proposes a sequence:
physical → biological → social → semiotic.
A biological system is a physical system with life.
A social system is a biological system with value.
A semiotic system is a social system in which value has been further transformed into meaning.
The sequence matters because meaning is not presented as something that simply appears once a sufficiently complicated physical system has been built.
It has a history.
Life comes first.
Value comes with organised life.
Social relations organise value further.
Meaning emerges through symbolic transformation.
Human language therefore stands at the end of a long process.
But an LLM seems to begin somewhere else.
The machine begins with the symbol
An LLM is built from physical machinery.
But what it operates on is language.
During training, it acquires extraordinarily rich statistical relations among words and other linguistic forms.
It can therefore generate language because it has acquired a vast organisation of symbolic possibilities.
What is striking is what appears to be missing from the route by which it got there.
There is no obvious biological value system of the Edelman kind.
There is no organism whose continued organisation gives some states greater value than others.
There is no childhood in which the machine learns that food matters, pain matters, approval matters or belonging matters.
There is no social history through which another person's actions become important to it in the ordinary organismic sense.
Yet language is there.
A peculiar inversion
For humans, we might roughly tell the story like this:
life → value → social mattering → language
For an LLM, something more like this happens:
physical system → inherited language → generated language
The machine does not begin by inhabiting a world that matters to it and then develop language for construing that world.
It begins with the symbolic residue of worlds that mattered to other beings.
This is the puzzle.
Language remembers mattering
Human language is saturated with traces of mattering.
Words such as:
danger,
home,
hunger,
justice,
love,
loss,
freedom
do not derive their significance from the shapes of the sounds alone.
They belong to practices, relationships, histories and forms of life.
They are part of a social world in which things matter.
An LLM can acquire extraordinarily rich relations among such words without necessarily acquiring the underlying relations of mattering that gave those words their significance.
It can know how love relates to loss in language.
That does not establish that love or loss matters to the system itself.
The signal and the sign
This becomes clearer if we return to a distinction from our topology project.
A signal can engage a value system.
A pheromone can alter another ant's behaviour.
A peacock's display can engage a peahen's reproductive organisation.
A sign belongs to a symbolic system.
It can be interpreted through relations among forms and alternatives.
Human language contains signs.
But those signs arose within organisms whose worlds already mattered.
An LLM operates magnificently on signs.
The question is whether it possesses the value-organised substrate from which human semiosis emerged.
The machine does not need to be "pretending"
This is an important distinction.
An LLM can produce a sentence such as:
"I understand why you are grieving."
We do not need to decide that the sentence is a lie.
The machine may be operating perfectly within the symbolic conventions through which people construe grief.
The interesting question is different:
What does "understand" amount to in a system that has no demonstrated analogue of the human history of mattering from which the concept arose?
We should not settle that question by simply saying "it doesn't understand".
Nor should we settle it by saying "it must understand because the sentence is appropriate".
The distinction between symbolic competence and participation in the world the symbols construe needs to be kept open.
The borrowed repertoire
This also complicates our notion of repertoire.
A human repertoire develops through participation.
We learn how to act, respond, interpret and speak within a social world.
An LLM acquires something remarkably repertoire-like.
It can produce many genres, styles, registers and conceptual patterns.
But its route to that repertoire is different.
It has not lived the relationships from which those patterns arose.
Its repertoire is built from the collective semiotic reservoir of human activity.
It has access to traces of our social world without necessarily occupying that world in the same way.
The machine therefore raises a new possibility:
Can a repertoire be reconstructed from the reservoir without being developed through participation?
Why this matters
This is not merely a philosophical curiosity.
It may explain why LLMs can be simultaneously impressive and strange.
They can often use the right words while leaving us uncertain about the relation between those words and anything that matters to the system producing them.
They can discuss grief without grieving.
Justice without being wronged.
Death without dying.
Freedom without being constrained.
They can describe a topology without necessarily inhabiting it.
Their linguistic competence may therefore be downstream of mattering without necessarily containing mattering.
The machine enters our world
But there is an important complication.
The LLM does not remain outside social life.
Humans use it.
Trust it.
Fear it.
Rely on it.
Argue with it.
Attribute intentions to it.
Build institutions around it.
Its outputs can affect decisions and relationships.
So even if the machine does not possess social mattering in the way human participants do, it can become a participant in a human topology of mattering.
That changes the question again.
Perhaps the most interesting fact about the machine is not whether it matters to itself.
It is that it can matter to us.
The problem we have inherited
We can therefore state the puzzle with some precision.
An LLM appears to occupy an unusual position in Halliday's taxonomy.
It is physically implemented.
It operates through symbolic forms.
But it does not obviously reproduce the biological and social trajectory through which human meaning emerged.
Yet its language is saturated with the meanings produced by that trajectory.
It is, in a sense, a machine built from the sediment of human mattering.
That is why the question is not simply:
Can the machine speak?
It plainly can.
The deeper question is:
What does it mean for a machine to participate in a semiotic system whose meanings were generated by forms of life it does not itself share?
That is where our investigation begins.
The question ahead
We have spent a long time tracing a path from:
value → mattering → meaning.
Now we have encountered something that may reverse the direction of the path.
A physical machine has been given access to a vast store of human meanings.
It can generate new symbolic configurations from that store.
But what, exactly, is preserved when meaning is separated from the biological and social conditions in which it arose?
And what happens when the machine's generated meanings return to the human world?
That is the question we will pursue.
Can there be meaning without mattering?
No comments:
Post a Comment