The building bricks of language

Language makes infinite use of finite means, as Chomsky would have it. There is an interesting functional similitude between Chomsky’s generative grammars and the latest generative language models such as ChatGPT. They both generate language; they are able, broadly speaking, to convert an idea into various wordings, within the bounds of what is grammatical. They differ as well: Chomsky’s economizing principles stand in stark contrast with the careless ‘maximalism’ of Large Language Models, fueled by the motto “there’s no data like more data”. How do these models relate to each other, and also with respect to an important third model, the human language system, implemented in countless independent brains? Drawing on linguistic, computational and neurobiological research I discuss how these different views offer explanations for individual language variation, by focusing on the building bricks, the ‘finite means’, they appear to be using.