Modern neural networks have shown impressive success in approximating complex multidimensional transformations and generating realistic signals such as images, videos, texts, and speech. They can imagine new situations, indicating the potential for the artificial general intelligence (AGI) development, which differs from narrow artificial intelligence in that it can in-dependently find new knowledge without direct human involvement, based on the analysis of input data and existing knowledge. The AGI creation requires the neural structures development capable of independent task understanding, breaking it down into components and formulating intermediate goals. This paper presents ideas for designing AGI neural network structures capable of acquiring new knowledge in complex environments.
