Comment by js8
15 days ago
I don't see how the two things are related. Whether acquisition of human language is nature or nurture - it is still learning of some sort.
Yes, maybe we can reproduce that learning process in LLMs, but that doesn't mean the LLMs imitate only the nurture part (might as well be just finetuning), and not the nature part.
An airplane is not an explanation for a bird's flight.
The great breakthrough in AI turned out to be LLMs.
Nature, for an LLM, is its design: graph, starting weights, etc.
Environment, for an LLM, is what happens during training.
LLMs are capable of learning grammar entirely from their environment, which suggests that infants are too, which is bad for Chomsky's position that the basics of grammar are baked into human DNA.
LLMs require vastly more data than humans and still struggle with some more esoteric grammatical rules like parasitic gaps. The fact grammar can be approximated given trillions of words doesn't explain how babies learn language from a much more modest dataset.
I think it does. I think LLM showed us possibility that maybe there's no language but just pile of memes and supplemental compression scheme that is grammar.
LLM had really destroyed Chomsky's positions in multiple different ways: nothing perform even close to LLM in language generation, yet it didn't grow a UG for natural languages, while it did develop a shared logic for non-natural languages and abstract concepts, while dataset needing to be heavily English biased to be English fluent, and parameter count needing to be truly massive as multiple hundred billion parameters large, so on and on.
Those are all circumstantial evidences at best, a random paraphernalia of statements that aren't even appropriate to bring into discussions, all meaningless - in the sense that an open hand of a person observing another individual aligned to a line between standing position of the person to the center of nearest opening of a wall would be meaningless.
2 replies →
It's not that the invention of LLMs conclusively disproves Chomksy's position.
However, we now have a proof-of-concept that a computer can learn grammar in a sophisticated way, from the ground up.
We have yet to code something procedural that approaches the same calibre via a hard-coded universal grammar.
That may not obliterate Chomksy's position, but it looks bad.
6 replies →