Knowledge-based Capabilities of a Linguistic Neural Network

Abstract

Today, linguistic neural networks are penetrating all spheres of human activity, including science. This fact is generally considered positively, as it yields a clear economic payoff. According to popular predictions, the current generation of people will already face the emergence of advanced artificial intelligence (AI) developed on the basis of the GPT line of linguistic neural networks. They say that AI will surpass the human intelligence in all respects. However, these expectations seem to be inflated. The main reason lies in the fact that the domain of modern linguistic models of artificial intelligence is the language of human communication, but languages of intuitive thinking, without which the generation of new knowledge obviously does not occur, are not yet amenable to formalization by means of information technology. The purpose of the work was to evaluate the knowledge-based capabilities of the modern neural network ChatGPT-3.5 in the course of discussion of phenomena of different scales: the control means and the foundations of modern science in the West and in Russia and the problem of reversibility and irreversibility of time in physics reflected in the theories of crystal growth. Initially, a linguistic model is aimed at making an assertion that is as anticipated as possible. And this feature leads to the eclecticism of the whole set of responses related to the broad theme of a dialog. The results of communication with the neural network on a narrowly specialized topic have demonstrated its ignorance of a well-known physics problem and, more importantly, its inability to apply it to a theory of crystal growth where this problem is the key one. Preoccupations, unreasonable optimism or fears about AI that characterize the moods of contemporary society have so far had little to do with scientific practice, apart from the harm caused by the ever-increasing information noise in which neural networks are becoming involved.