Making Sense of the Technical Strengths and Limitations of LLMs

Abstract This article considers a question left open by Samuel R. Bowman's Critical AI article “Eight Things to Know about Large Language Models,” namely, how humanlike these systems’ representations might be. Reviewing the evidence Bowman provides and contrasting it with results from cognitive science, the article argues that these systems at best approximate the real-world information that people rely on, while lacking the compositional and productive mechanisms needed to put this information to use in humanlike ways. More generally, talk of representation is an inherently anthropomorphizing strategy that demands more rigorous practice from computer scientists but opens generative possibilities for humanistic interventions.