This review article discusses the skepticism surrounding large language models (LLMs) as mere mimics of language. It evaluates evidence against the Blockhead analogy, indicating that LLMs may exceed predictions about non-classical systems. The authors argue for the need to investigate how LLMs understand language and their internal mechanisms, noting that current methods of analysis are insufficient. A new set of experimental methods is required to explore LLM behavior and cognition further, setting the stage for future research in this evolving field.
Our analysis revealed that the advanced capabilities of state-of-the-art LLMs challenge many of the traditional critiques aimed at artificial neural networks as potential models of human language and cognition.
Moving beyond the Blockhead analogy continues to depend upon careful scrutiny of the learning process and internal mechanisms of LLMs, which we are only beginning to understand.
We will explore these methods, their conceptual foundations, and new issues raised by the latest evolution of LLMs in Part II.
An understanding of what LLMs represent about the sentences they produce—and the world those sentences are about—calls for careful empirical investigation.
Collection
[
|
...
]