Are Large Language Models obsolete? And are multimodal solutions inappropriate?

This brief commentary proposes a novel AI design paradigm inspired by an evolutionary theory of language and a dialogical cognitive model. Unlike traditional multimodal models, our approach integrates multimodality at the core of the design process, moving beyond the primary reliance on large language models (LLMs). Drawing from cognitive neuroscience, we propose an asymmetric, dual-network architecture where each network processes distinct modalities – LLMs and gestural language – translating inputs from one to the other. This design aims to enhance creativity and error tolerance in AI systems. We discuss the theoretical foundations, potential advantages, and future research directions of this approach.