| Sumario: | Since the release of ChatGPT in November 2022, large language models (LLMs), once the purview of machine learning experts, computational linguists, and cultural analytics scholars, have entered the public discourse. Many different metaphors have been used to describe these models, and all of them reach past thinking of large language models as simply instruments for using language (or, as the companies that sell chatbots would have it, as artificial intelligence itself)—instead, many scholars ask us to consider the model as something that both uses language and is made of language. This essay will contextualize the history of LLMs within a larger history of compilation in text technologies and the digital humanities. In this broader historical view, it is possible to understand LLMs as programs that use compilations as training inputs to produce compilations as readable output but are not themselves compilations within their model architecture. The metaphor of compilation is useful, alongside the many other metaphors used to understand these often-opaque models, as a way of capturing the historical continuities in how language models are trained and how their outputs are read.
|
|---|