There is another way of reducing what a language model does:
“It’s just imitating us.”
The claim seems almost self-evident.
A language model is trained on human-produced language. It learns patterns from what people have written and then generates new language that resembles those patterns.
So whatever appears to be intelligence is really only imitation.
Again, there is something true here.
But once again, there is a small word that changes the whole sentence.
Just.
Because imitation is not necessarily a simple thing.
A parrot can imitate a sound.
A child can imitate an adult.
An actor can imitate a person.
A musician can imitate a style.
A scientist can imitate an experimental result.
A species can evolve to imitate the appearance of another species.
These are all forms of imitation, but they are not the same activity.
So what exactly does a language model imitate?
It does not copy a particular person.
It does not reproduce a particular sentence every time it generates language.
It does not even need to have encountered the precise sequence it produces.
What it learns is something more abstract: regularities in the ways people use language.
That distinction matters.
If I imitate your sentence by repeating it word for word, the relationship between your utterance and mine is fairly straightforward.
But if I learn your way of arguing, your vocabulary, your habits of explanation, the kinds of distinctions you tend to make, and the ways those things combine in unfamiliar circumstances, the relationship is different.
I am no longer merely copying an utterance.
I have acquired some capacity to reproduce a pattern.
That is still imitation.
But it is imitation at a different level of organisation.
And there is something else worth noticing.
Human beings themselves learn an extraordinary amount through imitation.
Children imitate sounds before they understand the grammatical system those sounds participate in. They imitate gestures, expressions, behaviours and social conventions. Apprentices imitate skilled practitioners. Writers imitate genres. Musicians imitate styles before developing their own.
Imitation is not the opposite of learning.
It is one of the ways learning occurs.
Indeed, social life would be difficult to imagine without it.
This creates an awkward question for the claim that an LLM is “just imitating us.”
If imitation is enough to dismiss machine behaviour as unintelligent, why isn't imitation enough to explain a considerable amount of human learning?
The obvious answer is that humans do more.
We have bodies.
We have needs.
We inhabit physical environments.
We enter relationships.
We act on the world and are acted upon by it.
We have histories that are not merely textual.
All of this is important.
But notice what follows.
These differences give us reasons to investigate what the machine lacks and what those absences make impossible.
They do not establish that imitation itself is trivial.
There is another complication.
A language model is trained not simply on individual examples but on a vast accumulation of linguistic activity. Human beings have already used language to describe objects, events, relationships, arguments, emotions, institutions, scientific theories, fictional worlds and imagined possibilities.
The model's training data therefore contains traces of a world of social meaning.
The model does not need to have lived that world in order to have learned something about the regularities of its linguistic representation.
That does not mean it has acquired the world itself.
But neither does “imitation” tell us what it has acquired.
Perhaps the most revealing comparison is with an apprentice.
An apprentice learns a craft partly by imitating those who already practise it.
At first the imitation may be crude.
Gradually it becomes more flexible.
Eventually the apprentice may perform an action that nobody has demonstrated before, while still acting within the organisation of the learned practice.
At that point, would we say that the apprentice is only imitating?
We might.
But the word would no longer explain very much.
The interesting question would have become:
What has been learned that makes novel performance possible?
That question applies to language models too.
A model can produce combinations of words that no human ever wrote.
It can respond to novel prompts.
It can adapt its response to preceding context.
It can combine patterns acquired from very different domains.
Whether these capacities amount to understanding, thinking or anything else remains open.
But they cannot be described adequately simply by saying that the system repeats what it has seen.
The stronger claim — that the system merely imitates — would require us to specify what kind of imitation we mean and what the learned organisation consists of.
And perhaps there is a deeper irony.
Language itself is profoundly imitative.
Every speaker inherits a language they did not invent.
We reproduce words, grammatical patterns, metaphors, genres, conventions and ways of making distinctions that existed before us.
We do not begin with language and create it from nothing.
We enter an already existing social semiotic system and learn how to participate in it.
Human linguistic creativity does not consist in escaping that inheritance.
It consists partly in doing new things with what we have inherited.
That makes the contrast between human language and machine language less simple than “we create; it imitates.”
Both operate within histories of prior usage.
Both can produce novel combinations from inherited resources.
The differences between them remain profound.
But “imitation” does not tell us what those differences are.
It tells us where the material came from.
It does not tell us what the system has become capable of doing with it.
And so the familiar sentence turns out to have the same structure as the others.
“It’s just imitating us.”
Perhaps.
But just has again transformed a description into a dismissal.
The more interesting question is:
When does imitation become learning, and when does learned imitation become a capacity for something new?
No comments:
Post a Comment