Caution: The Myths Surrounding the “Generality” of AI Reasoning Abilities

Last week, Apple researchers published a provocative paper titled “The Illusion of Thinking,” suggesting that current language models (LLMs) face significant limitations in reasoning. While the findings sparked interest, notable critiques, including from Gary Marcus, indicate the paper may overstate its conclusions, reflecting general sloppiness and lack of depth. The paper presents LLMs struggling with four reasoning tasks, claiming performance drops to zero past a certain complexity, implying fundamental limitations in reasoning capabilities. However, counterarguments emphasize that many observances may relate to inherent task complexities rather than intrinsic flaws in LLMs. Notably, LLMs can execute tasks via programming, suggesting a different form of reasoning than the authors assessed. Critics argue effective critique should include empirical evidence, and they caution against dichotomizing reasoning ability based solely on performance on toy examples. Overall, skepticism toward LLM capabilities is warranted, but comprehensive analyses grounded in real-world applications are essential.

Source link