Large Language Models

As an evolved medium of human communication and later of knowledge transmission preservation, a human language is NOT merely sequential patterns of familiar and well-defined (or at least socially agreed upon) words, which presumably denote thing, processes and attributes, along with purely language-level constructs and common, standard idioms.

Understanding of a human language, by how and what for it has been evolved, implies extraction of the underlying meaning of a verbal communication, so called “deep-structure”, which, hopefully corresponds to an inner representation of some “fact” about out shared environment in which we happen to evolve.

So, it is a meaning-reconstruction, not pattern-recognition.

All the modern models operate on sequential patterns, not on underlying meanings, and by definition (by the underlying algorithms) are completely oblivious to existence of any underlying meaning behind a verbal communication.

There are obvious examples from real life.

When a person cannot construct an adequate verbal description of what is around him, we diagnose a neurological or a psychiatric disorder as an ultimate cause (which, of course, has to be scientifically confirmed). Society isolates such people not because they (only) hate them, but because the ability to adequately describe and understand verbal communication is absolutely essential to survive in a society.

This proves, however indirectly, what for a language has been evolved and therefore how it has to be used.

No one wants to listen to incomprehensible tirades a drunk or a dope, even to beloved children. It is crucial that every parent has an instinct urge to correct children’s language usage and freak out at any kind of speech disorders.

The exact problem with current language models is that they operate only on a “surface-structure” of a language, and even trying to capture the whole shape of it, instead of extracting the underlying “deep-structure” (meaning and intention) behind every single sentence.

This is just fundamentally wrong, but seems useful to an uneducated majority.

Notice that creation of “abstract” (disconnected from reality) theories and religions and philosophies (with abstract, non-existent entities) is considered as misuse of a language and an unnecesary burden on individuals and ensociety as a whole.

Disciplined use of a language, for writing or otherwise, has been a crucial concern of the earliest civilizations. An unambiguous and even precise use of a domain-specific language is at the core of mathematics, logic and of all sciences.

Nowadays programming languages (and layered “abstractions” and DSLs they produce) even more emphasize a necessary discipline.

ChatGPTs are just artificial biological parrots (cute birds which tends to mimic sounds). ChatGPTs mimic a “surface-structure” (and only it) of a human language.