Doing Linguistics in the GPT era

Chesi’s position on the relation between theoretical linguistics and large language models (LLMs) should be taken seriously by anybody who has worked on these two ways of approaching language. This commentary largely shares his general viewpoints, with some qualifications. One is that standardization (and the search for a broad core of compatible analyses) is more important than formalization. The second is that the size and opacity of the materials used for training the largest LLMs makes them essentially useless as models of human linguistic competence (though perhaps not entirely implausible as models of the evolution of a creature capable of using language). On the other hand, smaller models, trained on human-sized amounts of language (as in the the BabyLM challenge) could become useful tools to study human competence and understand the limits of exclusively data-driven approaches.