Abstract

LLMs are deep neural networks trained on vast amounts of text to predict the next word based on a sequence of previous words. In doing so, they develop a compact representation of the linguistic world. The most advanced LLMs display unexpected abilities in a wide variety of cognitive tasks such as analogical and counterfactual reasoning, problem solving, and understanding metaphors. When tested with classic paradigms of cognitive psychology, LLMs replicate patterns of responses and errors very similar to those of humans, thus making them promising models of cognitive processes. LLMs overcome many criticisms previously directed at associators, thanks to the attention mechanism that allows them to create associations even at long distances. There are still limitations in planning and autonomous goal generation, but progress in fine-tuning autonomous agents is very promising. LLMs have the potential to significantly impact cognitive psychology, reviving associationism as a unified theory of the mind and paving the way for new collaborative scenarios between AI and cognitive psychology