Large Language Models and Human Cognition: An Optimistic Perspective

Large Language Models (LLMs) are deep learning-based text generation tools that have shown remarkable improvement in producing coherent discourse and displaying unexpected abilities. This article explores how LLMs can contribute to our understanding of human language and cognition. The author argues that viewing LLMs merely as word predictors that mimic human language behavior oversimplifies their underlying mechanisms and representational capabilities. LLMs function as information acquisition and processing systems and, from a connectionist perspective, can serve as useful models of human cognition – or at least of certain aspects of it. The article begins by providing a brief account of criticisms concerning LLMs’ limitations, followed by an examination of the nature of representations within these models and a discussion on their architectural components. It further presents LLMs as general-purpose systems, highlighting their emerging non-linguistic capabilities. It is suggested that LLMs may have the potential to capture and effectively apply cultural constructs, which are primarily conveyed through language and encapsulate useful 'programs' for addressing a variety of tasks. Additionally, the article briefly examines the possibility of top-down processing and consciousness in these models. Ultimately, the author proposes that current and future generations of LLMs can contribute to a deeper understanding of human language activity and cognition, encouraging cognitive scientists to view them as more than mere engineering tools designed to mimic human language.