Editors’ Summary: In this post, Salvaggio considers how a lot of current scholarship on LLMs is fixated on proving that they are conducting something “akin to thought.” He discusses a recent paper that tested the ability of LLMs to make sense of ‘Jabberwocky’ language in which most or all content words have been randomly replaced by nonsense strings. He argues that the ability (or not) of LLMs to “think” is not what makes them important and calls for a reframing to the debate. This post claims that models are capable of transforming text, not writing it, and the prompt (written by a human) is what informs the quality of the response.
Editors’ Choice: Unreasonable Effectiveness
