AGI AND LINGUOFETISHISM
This chapter compares the current attention paid to natural language in artificial intelligence with how natural language is used by natural intelligence.
A century ago, the first movies appeared - including those with the legendary Charlie Chaplin; the situations in these films were quite understandable, despite the fact that there was no sound due to the lack of the required technologies. Today, assembly instructions for furniture bought at IKEA do not have text - all the required information is conveyed by pictures. Maps and drawings appeared with our prehistoric ancestors long before the alphabet appeared, and the preservation of technical information to this day is based primarily not on the use of documents in natural language, but on drawings, diagrams, tables, and photos. The advent of computers initiated the emergence of artificial programming languages, which people use without problems. When repairing devices today, everyone knows that it is useful to take a photo to make it easier to explain the situation, order parts, and correctly assemble the disassembled - no one thinks of describing this in text or voice in natural language. When a quick response to events is required, the notification of events does not use natural language - visual signals or acoustic signals are used. Messages are exchanged with pilots using specially developed formalized jargon, as when writing computer programs. Information delivered to the public by the media increasingly uses graphical representation of information as technical capabilities for this appear, since it is faster and easier to perceive. Learning new actions - physical exercises, cooking, car repair, etc. - is radically more practical and convenient when implemented in the form of videos.
De facto, the only area of ​​application of natural language, where it is not limited to a specific, restricted, and easily formalized subset, is dialogue between people (directly or using text as an intermediary). However, even in this case, as a rule, other methods of transmitting information are used - images, gestures, and facial expressions.
It should be noted that from a theoretical point of view, formalized languages, exhaustively described by a set of rules (not only programming languages but also those used, for example, in technical documentation), are capable of conveying any information that an "unlimited" natural language is capable of conveying; any arbitrarily complex phrase of a natural language can be reduced to an equivalent set of phrases in a formalized language - which is actually used in teaching natural language by gradually expanding proficiency in natural language and is the basis of explanatory dictionaries.
In practical applications, the slowness and high level of resource consumption prevent LLM (Large Language Models) systems, which use natural language to store information, from being used as the "brain" of robots, autonomous cars, drones, etc.
However, despite the apparent negative aspects of natural language and the inefficiency of its use in technical systems, LLM systems that use natural language outside its natural sphere of application as a means of human communication are being intensively developed and offered to the public.
However, despite the apparent negative aspects of natural language and the inefficiency of its use in technical systems, LLM systems that use natural language outside its natural sphere of application as a means of human communication are being intensively developed and offered to the public.
This could be explained by the presence of a large number of accumulated texts in natural language, requiring use in computerized systems, but modern systems, essentially based on the representation of knowledge by texts in natural language, are proposed as a foundation for the creation of universal-purpose intelligent systems comparable in capabilities to human intelligence. These intentions (and promises) take place despite the fact that the speed of natural language processing by these computer systems (ChatGPT, Grok, Claude, etc.) is hundreds and thousands of times lower than the speed of processing texts in formalized languages ​​(programming languages, professional jargons, etc.) and requires correspondingly as many more resources. This takes place despite the fact that in a natural intelligent system - in the human brain - information is certainly not presented in the form of texts in natural language.
The system of information representation in the brain developed long before the emergence of natural language as a means of information exchange and uses a unified representation of various types of information - visual, acoustic, tactile, emotional, etc., which is not verbal. The most obvious confirmation of this is the situation familiar to everyone when difficulties arise in verbally describing something absolutely familiar/known - this is caused by the complexity of converting non-verbal information representation in the brain into verbal representation and the ability to do this in many different ways. Mastering natural language as the ability to convert internal representation into verbal uses the existing internal non-verbal representation of knowledge to establish a connection with the verbal representation.
The current rejection of the option implemented by nature - non-verbal internal representation of information with verbalization of information in those cases when it is required (in communication with people). The real reason, as it seems to us, is the reliance on the false thesis that an artificial neural network is a natural way to create a strong artificial intelligence, which was discussed earlier in AGI AND NEUROFETISHISM in combination with the spectacular demonstration of the verbal capabilities of LLM systems based on artificial neural networks, masking the lack of ability to reason.
Probably the second factor that determined the direction of AI development, leading to the creation of LLM systems, is the implicit assumption of the initiators of the developments that the system being created should be used as an Internet mass service system; this allows ignoring the bulkiness and slowness of the system, which is unacceptable for local control systems of robots, drones, cars, etc.
The obvious and unavoidable errors of LLM, which are a consequence of both the innate specificity of artificial neural networks and the informality of natural language, confirm the impracticability of the approach to creating strong artificial intelligence based on the verbal representation of knowledge in artificial neural networks. Success should be expected in systems using a non-verbal representation of knowledge - a semantic graph (permanently modifiable ontology) describing the connections between concepts and a traditional database for the rest of the information. Natural language is represented by a subsystem for interpreting natural language into an internal representation and a subsystem for generating texts in natural language based on the internal representation of knowledge/data.
Of course, LLM has a broad scope of applicability, where full-fledged ability to reason and other capabilities of strong (real) artificial intelligence are not required.