There is a peculiar confidence in the air.
Large Language Models produce text that is fluent, contextually appropriate, and often uncannily persuasive. From this success, a familiar inference is drawn: if statistical models over symbol sequences can do this, then meaning itself must be statistical in nature. Probability, it seems, has finally explained semantics.
This inference is understandable. It is also wrong.
What probabilistic models have captured with extraordinary sophistication is not the source of meaning, but the grain of instantiation: the patterned unevenness with which semiotic potential is actualised in use. Fluency lives there. Meaning does not.
This post draws a clean cut. It does not deny the relevance of information theory, probability, corpora, or large-scale modelling. On the contrary, it situates them precisely—showing why they matter, what they explain, and, crucially, what they cannot.
1. The Seduction of Fluency
Fluency is deceptive.
When a system produces text that is locally coherent, genre-appropriate, and pragmatically attuned, it invites a category error: the slide from functional adequacy to ontological explanation. The temptation is to say: whatever produces this must therefore explain meaning itself.
But fluency is not meaning. Fluency is a property of deployment—of how semiotic resources are marshalled once meaning is already in play. It is entirely possible to generate fluent text without participating in the construal of any phenomenon whatsoever.
This is not a new confusion. It recurs whenever performance overwhelms theory.
What has changed is scale.
Large Language Models operate over vast corpora of sedimented human discourse. They model, with extraordinary resolution, which semiotic choices tend to follow which others under which conditions. The result is text that feels meaningful because it mirrors the distributional trace of prior meaning-making.
But mirroring the trace of construal is not the same as construing.
2. Information Without Meaning
Claude Shannon’s information theory is often invoked at this point, usually as a quiet ontological upgrade: information reduces uncertainty; meaning reduces uncertainty; therefore information is meaning.
This syllogism collapses under inspection.
Shannon information is rigorously defined over symbol distributions. It measures the expected reduction of uncertainty given a probability space. It is indifferent to interpretation, reference, intention, experience, or construal. Two messages with identical probability distributions have identical information content, regardless of what they are taken to mean.
This indifference is not a flaw; it is the condition of the theory’s power. Information theory abstracts away from meaning in order to model transmission efficiency, redundancy, and noise.
Meaning, by contrast, is irreducibly semiotic. It arises only through construal—through a perspectival cut that brings a phenomenon into experience as something. There is no meaning independent of such cuts, and no construal that can be derived from probability alone.
Information theory, therefore, is not a theory of meaning. It is a theory of patterned selection under uncertainty. Its relevance to linguistics lies not in explaining what meaning is, but in modelling how semiotic systems are used.
3. System, Instance, and the Weighting of Potential
This is where a Hallidayan system–instance relation becomes indispensable.
A linguistic system is not a catalogue of forms but a structured space of potential—an organised theory of what can be meant. An instance is not a temporal process but a perspectival actualisation: a cut from potential to event.
Probability enters here—not as an ontological ground, but as a property of systems in use.
Over time, repeated actualisations weight the system unevenly. Some options become more probable than others in particular contexts. These weightings can be modelled statistically. Corpora make them visible. Information theory provides tools for describing their distribution.
But none of this explains construal itself.
Probability does not generate meaning; it reflects the history of meaning-making. It tells us about the grain of instantiation: the texture left behind by countless prior cuts from potential to event.
To mistake this grain for the source is to confuse sediment with spring.
4. Large Language Models as Second-Order Phenomena
Large Language Models operate entirely within this grain.
They do not encounter phenomena. They do not construe situations. They do not make perspectival cuts from semiotic potential to lived event. Instead, they model relations among already-produced semiotic artefacts.
In this sense, LLMs are second-order through and through. They operate over distributions of construals, not over the world those construals bring into being.
Their success is therefore unsurprising. If fluency is a matter of aligning with the probabilistic texture of prior discourse, then systems optimised to learn that texture will excel.
What would be surprising is if such systems failed to sound meaningful.
But sounding meaningful is not the same as meaning.
5. The Category Error
The contemporary error is not to take probability seriously. It is to take it too seriously, in the wrong way.
When probability is treated as explanatory of meaning itself, a category mistake has already occurred. The grain of instantiation has been mistaken for the ground of meaning.
This mistake is encouraged by the apparent autonomy of fluent text. Detached from its conditions of production, language appears to float free, as though meaning were self-generating. Statistical models seem to confirm this illusion by reproducing the surface behaviour of discourse without participating in its semiotic work.
A relational ontology dissolves the illusion.
Meaning is not in the symbols, nor in their probabilities, nor in the model that predicts them. Meaning arises only in the act of construal—in the perspectival actualisation of semiotic potential within lived situations.
Probability explains why some actualisations are smoother, more expected, more fluent than others. It does not explain why anything is meaningful at all.
6. The Cut
We can now state the cut cleanly:
Information theory models the grain of instantiation, not the source of meaning.
Once this is seen, much of the contemporary confusion evaporates. Corpora matter. Probability matters. Large-scale modelling matters. But none of these replace semiotic theory; they presuppose it.
In the next post, we will sharpen this cut further by distinguishing first-order phenomena from second-order patterning—and by showing why the success of probabilistic models depends entirely on the prior existence of meaning they do not, and cannot, explain.
Fluency, after all, is a surface achievement.
Meaning is a cut.
No comments:
Post a Comment