On the Creation of Artificial Intelligence

The paper analyzes the fundamental gap between modern approaches to artificial intelligence, based primarily on machine learning methods, and the decision theory as a discipline focused on achieving the goals set by decision-makers. It is shown that the key problem in the development of strong AI is a mismatch between the target functions of machine learning models (prediction error minimization, perplexity, etc.) and the real values of end users of the technologies. Particular attention is paid to the limitations of the existing paradigms: the inability of large language models to solve combinatorial optimization problems, insufficient knowledge on the neurophysiological foundations of decision-making, and the difficulties of formalizing fuzzy human judgments within the framework of classical models. A synthetic approach is proposed to overcome these barriers through the integration of reinforcement learning, multicriteria decision analysis, and fuzzy logic. Three key components of the solution are detailed: the construction of an adaptive reward function based on fuzzy preference relations and the dynamic refinement of the decision-maker’s value system; the creation of a hierarchical model of the environment with learnable abstraction rules for scaling from the micro- to the macrolevel; and the architecture of a “strong” agent generating heuristics for “weak” agents. The creation of strong AI is impossible without the formation of a unified methodological basis that combines machine learning and decision theory, and it requires a targeted development of new mathematical tools for representing multicriteria problems and fuzzy preferences within the framework of the learning environment of intelligent agents.