This paper analyzes the problem of the low practical efficiency of modern large language models (LLMs). We argue that the limiting factor is not a limit of the technology itself, but the organization of reasoning and the architectural role assigned to the language model. We claim no novelty for the "executor — overseer" scheme: analogous configurations have been described in the literature on multi-agent and self-verifying systems, and the field observation presented here is treated as an empirical illustration of a particular case of a more general hypothesis. The main contribution consists of four parts. (1) Diagnosis: a modern LLM is interpreted as the implementation of predominantly a single functional zone — the speech zone. Mind as a process is amodal and not bound to natural language; language is merely one of the serializers operated by the speech zone. (2) An architectural framework: a zonal-modular architecture with an explicit control circuit and with memory moved out into separate addressable zones — together with an ontogenetic program for its formation: zones and connections are grown in stages, in a sequence reproducing the stages of human cognitive development, with verification gates between stages, an offline consolidation phase ("sleep"), and an upbringing (alignment) stage built into the ontogeny with a graduation gate. (3) Experimental verification of two mechanisms of this program. Experiment 1 (training small transformers from scratch, two synthetic domains, four arms): staged formation with worked solutions and gates outperforms training on a shuffled corpus at an equal budget (+0.11 and +0.17 final-exam accuracy), and the accuracy gap between the training set and the exam falls from ~0.2 to ~0 — the model stores the rule rather than memorized answers. Experiment 2 (an agentic task stream): the cost of linearly accumulated context grows quadratically in the agent's lifetime horizon, while the cost of periodic consolidation grows linearly, at equal answer quality; the break-even point is reached already at short horizons. (4) A motivating field observation on a real engineering task, from which the hypothesis is derived. In this optic, modern LLMs are merely the first, speech zone of the future architecture; a full-fledged artificial mind requires building out the remaining zones operating outside language, a correct memory architecture, a coordinating circuit operating in an amodal medium, and a staged ontogeny with verification gates.
Dmitriy Borzov· Zenodo (CERN European Organi...· 0 citations
This paper analyzes the problem of the low practical efficiency of modern large language models (LLMs). We argue that the limiting factor is not a limit of the technology itself, but the organization of reasoning and the architectural role assigned to the language model. We claim no novelty for the "executor — overseer" scheme: analogous configurations have been described in the literature on multi-agent and self-verifying systems, and the field observation presented here is treated as an empirical illustration of a particular case of a more general hypothesis. The main contribution consists of four parts. (1) Diagnosis: a modern LLM is interpreted as the implementation of predominantly a single functional zone — the speech zone. Mind as a process is amodal and not bound to natural language; language is merely one of the serializers operated by the speech zone. (2) An architectural framework: a zonal-modular architecture with an explicit control circuit and with memory moved out into separate addressable zones — together with an ontogenetic program for its formation: zones and connections are grown in stages, in a sequence reproducing the stages of human cognitive development, with verification gates between stages, an offline consolidation phase ("sleep"), and an upbringing (alignment) stage built into the ontogeny with a graduation gate. (3) Experimental verification of two mechanisms of this program. Experiment 1 (training small transformers from scratch, two synthetic domains, four arms): staged formation with worked solutions and gates outperforms training on a shuffled corpus at an equal budget (+0.11 and +0.17 final-exam accuracy), and the accuracy gap between the training set and the exam falls from ~0.2 to ~0 — the model stores the rule rather than memorized answers. Experiment 2 (an agentic task stream): the cost of linearly accumulated context grows quadratically in the agent's lifetime horizon, while the cost of periodic consolidation grows linearly, at equal answer quality; the break-even point is reached already at short horizons. (4) A motivating field observation on a real engineering task, from which the hypothesis is derived. In this optic, modern LLMs are merely the first, speech zone of the future architecture; a full-fledged artificial mind requires building out the remaining zones operating outside language, a correct memory architecture, a coordinating circuit operating in an amodal medium, and a staged ontogeny with verification gates.
Dmitriy Borzov· Zenodo (CERN European Organi...· 0 citations