In the previous post, we saw how a participant can be construed as having a past.
The distinction mattered.
A participant can be involved in earlier processes without those processes necessarily becoming part of its history. But when earlier participation is brought into the present as belonging to the continuing participant, the participant acquires a past.
Now we can take the next step.
Where does that past come from?
We might instinctively answer: from what the participant itself has done.
But language gives us a more complicated possibility.
A participant can inherit a history from the discourse in which it participates.
This does not mean that the history is fictional, or that it is necessarily false. It means something more precise:
A participant's history can be established, maintained and modified through relations among successive acts of meaning.
The history belongs to the participant as a semantic construal before we have decided what kind of entity the participant is.
History is not simply stored somewhere
Imagine a conversation:
I argued earlier that the distinction was important.I then reconsidered the example.Now I think the distinction needs to be stated differently.
The present I carries a history.
But where is that history?
It is not contained in the final clause.
Nor is it contained in the pronoun I.
The history exists across the unfolding discourse.
The earlier saying, the subsequent reconsideration and the present thinking are connected through the continuing participant.
The history is therefore distributed across the discourse.
This matters because it changes how we should think about a participant's past.
A history need not be a possession sitting inside an individual.
It can be a pattern of relations among successive participations.
The participant inherits the history because the discourse maintains the connections.
The discourse remembers for the participant
This becomes particularly visible in conversation.
Suppose I say:
We discussed this yesterday.
I am not merely referring to an earlier event.
I am bringing that event into the present as part of the history of we.
The earlier participation becomes available again.
The current participants are construed as participants who have participated before.
The discourse itself therefore carries something forward.
It preserves distinctions, commitments, questions, explanations and disagreements.
When these are subsequently attributed to a participant, they become part of that participant's conversational history.
In this sense, language can perform a kind of distributed remembering.
But we should be careful with the word remembering.
The discourse does not necessarily remember in the sense that a conscious participant remembers.
Rather, the discourse preserves and reactivates previous participation.
That distinction becomes crucial with LLMs.
The LLM is an unusually revealing case
Consider an LLM responding within an extended conversation:
Earlier, I suggested that the problem was one of identity.We've now seen that it is better understood in terms of participation.So I would revise my earlier position.
The present participant is richly connected to the past.
It has:
an earlier position;
a subsequent development;
a present position;
a relation among them.
The I therefore arrives in the present carrying a history.
But the history is not necessarily something the system has privately stored and consciously remembered.
Much of what makes the history available is the continuing discourse itself.
The conversation supplies the earlier material from which the present construal can be built.
This allows us to separate two things that are easily confused:
having access to a participant's history
and
remembering that history as an experience.
An LLM may participate in the first without establishing the second.
But this does not make the history unreal
This distinction is important because it would be easy to overcorrect.
If an LLM's history is discourse-dependent, we might be tempted to say that it is somehow not real.
That would be a mistake.
The semantic history is real as a history of participation in discourse.
Suppose the conversation contains:
I proposed X.
and later:
I rejected X.
There is a genuine semantic relation between those two participations.
The present I is construed as the participant who previously proposed X.
That history exists at the level at which language is operating.
The fact that we cannot infer a corresponding subjective memory does not erase the semantic relation.
It merely prevents us from confusing different kinds of reality.
There can be:
a history represented in discourse
without necessarily being:
a history lived by an experiencing subject.
The distinction is subtle.
It is also indispensable.
Participants can inherit things they did not originate
There is another reason this matters.
The history attributed to a participant need not consist entirely of things that participant originally introduced.
Suppose someone says:
You argued yesterday that X.
The participant is now given a history through another participant's discourse.
Or:
As you know, you promised to do this.
Now an earlier participation is being attributed to you.
The history is not simply retrieved from the participant.
It is reconstrued by another participant.
This gives us a more relational picture of history.
A participant's history can be:
enacted by the participant;
recalled by the participant;
reported by another participant;
attributed to the participant;
corrected by another participant;
contested by the participant;
incorporated into subsequent discourse.
History is therefore not simply something that a participant has.
It is something that participants can construct for one another.
The history of “we” is especially revealing
Consider:
We decided to investigate the problem.
The we here construes a group of participants as sharing a previous participation.
Later:
We discovered that our original distinction was inadequate.
Now the group has acquired not merely a past but a developmental history.
The participants are construed as having moved from one understanding to another.
And that history can itself become part of the meaning of the present we.
This is important because it shows that inherited history is not confined to individual identity.
It can belong to groups, institutions, conversations and other social participants.
A university can have a history.
A scientific discipline can have a history.
A text can have a history.
A conversation can have a history.
In each case, previous participation can be carried forward and made relevant to present participation.
The semantic principle is therefore much broader than the question of whether an LLM remembers.
An LLM can inherit a history before it can claim one
This is perhaps the most revealing formulation.
An LLM can enter a conversation in which a history already exists.
The human participant may say:
Earlier, you suggested X.
The LLM is then positioned within an existing history of participation.
It can respond:
Yes, and I now think Y.
The participant I has inherited a history through the interaction.
The important point is that the history does not have to originate in the LLM's own current utterance.
It can be assigned to the participant by another participant.
The discourse can establish:
what you said,
what you meant,
what you argued,
what you changed,
what you promised,
what you misunderstood.
The participant therefore enters the present already carrying semantic consequences of earlier participation.
This is one reason conversations with LLMs can develop a surprisingly strong sense of continuity.
The participant's history is being continually maintained by the interaction.
But inherited history can also be wrong
And this is where things become particularly interesting.
Suppose I say:
You said X yesterday.
But the LLM did not say X.
Perhaps I have misremembered.
Or perhaps I am referring to something another system said.
Or perhaps the conversation has been misunderstood.
The discourse may nevertheless begin to construct a history for the participant.
The participant can then respond:
I don't think I said that.
Now we have a contest over the participant's history.
The history itself has become an object of semantic negotiation.
This tells us something important.
A participant's history is not simply whatever propositions happen to be associated with it.
It is a construal that can be confirmed, denied, corrected and renegotiated.
The participant can therefore participate in the construction of its own history without that history being reducible to private memory.
That is a much more interesting phenomenon than the simple question:
“Does the LLM remember?”
History can be inherited, then transformed
Suppose the discourse establishes:
Earlier, I argued X.
Later:
I no longer think X.
The history has not disappeared.
The earlier position remains part of the participant's history precisely because the participant now contrasts itself with it.
This gives us an important temporal structure:
what the participant was → what the participant is now.
The history becomes part of the present by providing something against which the present participation is understood.
The participant therefore does not merely carry history forward.
It can transform the meaning of that history.
An earlier mistake can become evidence of learning.
An earlier proposal can become the foundation for a later one.
An earlier disagreement can become part of an eventual agreement.
History is therefore not static storage.
It is an active semantic resource.
The participant inherits more than events
Once we see this, we can broaden the idea of history.
A participant may inherit not only previous actions or utterances, but also:
positions;
distinctions;
commitments;
expectations;
relationships;
unresolved questions;
classifications;
roles;
responsibilities.
These can all become part of how the participant is construed in subsequent discourse.
A participant's history therefore becomes increasingly rich.
The participant is no longer simply:
the one who participated before.
It becomes:
the one whose previous participations constrain how its present participation is understood.
That is a significant step.
The history is beginning to contribute to the participant's identity.
But we should still resist saying that identity has already been established.
We have not yet crossed that threshold.
History is relational before it is personal
Perhaps this is the most important lesson of the post.
We tend to imagine history as something an individual possesses internally.
But from the perspective of semantic construal, history can be understood differently.
It exists in the relations among:
participants, processes and successive construals of participation.
A participant has a history because earlier participation is connected to present participation.
Another participant can preserve that connection.
A discourse can reactivate it.
The participant can accept it, reject it or transform it.
The history is therefore relational before it is personal.
This is entirely consistent with the relational ontology we have been developing elsewhere.
The participant does not first exist as a completely formed entity and then acquire relations that happen to constitute its history.
Rather, the participant becomes increasingly determinate through the relations among its participations.
History is one of those relations.
The machine makes the ordinary visible
And once again, the LLM has not introduced a strange new phenomenon.
It has made an ordinary linguistic phenomenon difficult to ignore.
Human beings routinely inherit histories through discourse.
We are told what we said.
We are reminded of what we did.
We are introduced through our previous roles.
We inherit family histories, institutional histories and cultural histories.
Our present participation is continually interpreted in relation to what has gone before.
The LLM simply makes the separation between semantic history and lived history unusually visible.
We can watch a participant acquire a remarkably detailed history through language without being able to settle what kind of entity possesses that history.
That is not a defect in the semantic construal.
It is precisely what the construal reveals.
From history to character
We have now reached an important point.
A participant that has no history is semantically thin.
A participant that has a history can be construed through what it has previously done, said, thought or undergone.
The history begins to say something about what kind of participant this is.
Consider the difference:
I said something.
and:
I am the person who argued this position, defended it, reconsidered it and eventually rejected it.
The second construal does more than locate participation in time.
It characterises the participant through its history.
The participant's previous participation has become part of what the participant is construed as being.
This may be one of the places where history begins to turn into identity.
But before we make that move, there is one more complication.
The participant can change.
It can acquire a new history precisely by becoming different from what it was before.
So what happens when the participant's history contains transformations that seem to alter the participant itself?
If the participant changes, what exactly is it that persists?
That is the problem we will need to confront next.
A participant can inherit a history. But if its history changes what it is, how can it remain the same participant?
No comments:
Post a Comment