Great post, Sean. This explanation of variations in what Saussure called “parole,” language in use, vs “langue,” language as an object, refocuses the question of what LLMs are actually doing and why they can seem weird. Is it reasonable to say that the LLMs software is built around language as object but the training derives from language in use? Chomsky’s genius was to explore a grammar of language in use, which is the basis of computational language. This whole question of sociolinguistics and the bot is fascinating.
Great post, Sean. This explanation of variations in what Saussure called “parole,” language in use, vs “langue,” language as an object, refocuses the question of what LLMs are actually doing and why they can seem weird. Is it reasonable to say that the LLMs software is built around language as object but the training derives from language in use? Chomsky’s genius was to explore a grammar of language in use, which is the basis of computational language. This whole question of sociolinguistics and the bot is fascinating.
Thanks!
It's a good point. Others have also made reference to Chomsky's performance/competence distinction with respect to LLMs. You might enjoy this article: https://www.cell.com/trends/cognitive-sciences/abstract/S1364-6613(24)00027-5