Contra Chomsky: LLMs Cannot Learn (at Least Some) Languages That Are Impossible for Humans to Learn.

You may remember Noam Chomsky’s NYT article on ChatGPT from last year. https://t.co/HyL1U2JXZr
He and others have claimed that LLMs are equally capable of learning possible and impossible languages. We set out to empirically test this claim.

— Julie Kallini ✨ (@JulieKallini) January 15, 2024

We find that GPT-2 struggles to learn impossible languages. In Experiment 1, we find that models trained on exceedingly complex languages learn the least efficiently, while possible languages are learned the most efficiently, measured through perplexities over training steps. pic.twitter.com/JPtGGe5ZtI

— Julie Kallini ✨ (@JulieKallini) January 15, 2024

A big thank you to my wonderful co-authors: @isabelpapad, @rljfutrell, @kmahowald, @ChrisGPotts.
This tweet will self-destruct in 5 seconds. đŸ”Ľ

— Julie Kallini ✨ (@JulieKallini) January 15, 2024

That Chomsky should make such a claim is not terribly surprising. The actual (physical) process by language strings are produced and comprehended is of no interest to him. For all practical purposes Chomsky treats language as existing in a timeless realm of Platonic ideals. Thus it wouldn't occur to him to think that the nature of the process by which LLMs produce language strings places restrictions on the nature of the languages they can deal with.