Large language models are not about natural language (arxiv.org)

A new paper argues that large language models (LLMs) fundamentally miss the mark when it comes to understanding human language. Authored by Johan Bolhuis, the study critiques LLMs as probabilistic systems that merely analyze vast datasets of text, lacking the innate, hierarchical thought structures inherent to human communication. Instead of learning language through minimal external input like humans do, these models rely on extensive data without grasping the underlying principles of language formation and interpretation. This insight is significant for the AI/ML community as it challenges the current trajectory of LLM development. Bolhuis emphasizes that linguistic models need to incorporate a deeper understanding of the cognitive processes involved in language acquisition and use, rather than solely focusing on data-driven approaches. His findings may spur a rethink in how researchers and developers approach artificial intelligence, highlighting the need for models that mimic human-like linguistic capabilities beyond surface-level data patterns. As the debate over the future direction of AI in language processing continues, such critiques are crucial for refining methodologies and enhancing the performance of language technologies.