3 ms·

ChatGPT does feel a bit like a refutation of his life's work. Because whether or not ChatGPT can think, it certainly has a good command of language without sta

by mcv 3y ago

ChatGPT does feel a bit like a refutation of his life's work. Because whether or not ChatGPT can think, it certainly has a good command of language without starting out with an inborn sense of grammar. It learns grammar by copying. Just like kids do, I suspect.

theodorethomas 3y ago
I suspect not. Children are exposed to far less text than LLMs. LLMs are parlour tricks that teach us nothing about how humans do it.
coldtea 3y ago
>without starting out with an inborn sense of grammar I'm not so sure though, because it starts by being fed sentences assembled by people. Sort of how we wouldn't say that a book itself has a "good command of language", but that the writer has one, and put it in the book. So you might need a seed with "an inborn sense of grammar" to create language first and a training corpus second, but once that's available in huge volumes you can mechanically train an LLM to appear to have a "good command of language" just by brute force. Would an LLM gain a "good command of language" if it was just fed all the taling a human hears and the words it reads in its first 18 years of life?
- ajuc 3y ago
  > Would an LLM gain a "good command of language" if it was just fed all the taling a human hears and the words it reads in its first 18 years of life? Depends on what its starting weights are ;)
  - BobbyJo 3y ago
    Aaaaaaand we're back to inborn sense.
benpacker 3y ago
Kids learn grammar by copying, but a group of kids with no language will invent one (see Nicaraguan sign language), and empty transformer models will not do this. Kids raised hearing a mix of French and African languages will also invent a new language (Haitian creole) and a LLM will not do this, it would alternate speaking one or the other. ChatGPT proves that language is not so complex that it can’t be modeled, but the core Chomsky arguments are unaffected. I haven’t looked into the alternative perspectives (emergent grammar, etc) in the discipline since undergrad so I’m not saying Chomsky is right, just saying LLMs don’t disprove it.
- mcv 3y ago
  > empty transformer models will not do this. Do we have evidence for that? Of course those kids did have a real world to communicate about, and that creates the need for a language. I think if you had AI models that had a world to interact with and other models to communicate with and learn from, they might be able to develop their own language too. But this does get to another of Chomsky's points: the need to have a world model. To current LLMs, the only world model they have is text they read on the internet. And they only communicate with us, and I don't think they learn from their communication with us. So that's unavoidably limiting them. But I think it would be a very interesting experiment to give them a world to interact with and other similar models to communicate with, to see if they develop some sort of language (which would be completely incomprehensible to us of course; that's another hurdle). I agree that LLMs don't disprove his inborn grammar argument, but they don't support it either. I think more research is needed.
dllthomas 3y ago
I suppose I can grant "feels a bit like", but unless I'm missing something (and to be fair I have not rtfa...) it hasn't actually reached refutation. The observation that there is enough information to determine grammar rules in the training set for an LLM is only very weak evidence that there is also enough information in the much smaller training set for a typical human.
- mcv 3y ago
  I think it's pretty obvious that there is more than enough information in the training set of a typical human to learn grammar. Because we do. And people in different countries learn different grammars. Also keep in mind that people get more feedback over their control of language ad grammar while they're learning. We don't just listen, we communicate.
  - dllthomas 3y ago
    That's the point, though. As you say there is obviously enough information to learn grammar given our starting point, but Chomsky's point is that there's not enough information to distinguish between all possible languages, so (paraphrasing from a decades-old recollection and probably very much putting my spin on it) learning a (first?) language as a human is probably a matter of figuring out how to apply existing "machinery" that imposes some of its own structure rather than operating entirely from the information provided to the senses. I'm not saying it's necessarily correct, but it does seem well formed, and LLMs only undermine it in a very weak sense (so far?). To defend the conclusion itself I would have to know a whole lot more about how he estimated information content (needed and available).

by mcv 3y ago

theodorethomas 3y ago

I suspect not. Children are exposed to far less text than LLMs. LLMs are parlour tricks that teach us nothing about how humans do it.

coldtea 3y ago

>without starting out with an inborn sense of grammar I'm not so sure though, because it starts by being fed sentences assembled by people. Sort of how we wouldn't say that a book itself has a "good command of language", but that the writer has one, and put it in the book. So you might need a seed with "an inborn sense of grammar" to create language first and a training corpus second, but once that's available in huge volumes you can mechanically train an LLM to appear to have a "good command of language" just by brute force. Would an LLM gain a "good command of language" if it was just fed all the taling a human hears and the words it reads in its first 18 years of life?

ajuc 3y ago
> Would an LLM gain a "good command of language" if it was just fed all the taling a human hears and the words it reads in its first 18 years of life? Depends on what its starting weights are ;)
- BobbyJo 3y ago
  Aaaaaaand we're back to inborn sense.

benpacker 3y ago

Kids learn grammar by copying, but a group of kids with no language will invent one (see Nicaraguan sign language), and empty transformer models will not do this. Kids raised hearing a mix of French and African languages will also invent a new language (Haitian creole) and a LLM will not do this, it would alternate speaking one or the other. ChatGPT proves that language is not so complex that it can’t be modeled, but the core Chomsky arguments are unaffected. I haven’t looked into the alternative perspectives (emergent grammar, etc) in the discipline since undergrad so I’m not saying Chomsky is right, just saying LLMs don’t disprove it.

mcv 3y ago
> empty transformer models will not do this. Do we have evidence for that? Of course those kids did have a real world to communicate about, and that creates the need for a language. I think if you had AI models that had a world to interact with and other models to communicate with and learn from, they might be able to develop their own language too. But this does get to another of Chomsky's points: the need to have a world model. To current LLMs, the only world model they have is text they read on the internet. And they only communicate with us, and I don't think they learn from their communication with us. So that's unavoidably limiting them. But I think it would be a very interesting experiment to give them a world to interact with and other similar models to communicate with, to see if they develop some sort of language (which would be completely incomprehensible to us of course; that's another hurdle). I agree that LLMs don't disprove his inborn grammar argument, but they don't support it either. I think more research is needed.

dllthomas 3y ago

I suppose I can grant "feels a bit like", but unless I'm missing something (and to be fair I have not rtfa...) it hasn't actually reached refutation. The observation that there is enough information to determine grammar rules in the training set for an LLM is only very weak evidence that there is also enough information in the much smaller training set for a typical human.

mcv 3y ago
I think it's pretty obvious that there is more than enough information in the training set of a typical human to learn grammar. Because we do. And people in different countries learn different grammars. Also keep in mind that people get more feedback over their control of language ad grammar while they're learning. We don't just listen, we communicate.
- dllthomas 3y ago
  That's the point, though. As you say there is obviously enough information to learn grammar given our starting point, but Chomsky's point is that there's not enough information to distinguish between all possible languages, so (paraphrasing from a decades-old recollection and probably very much putting my spin on it) learning a (first?) language as a human is probably a matter of figuring out how to apply existing "machinery" that imposes some of its own structure rather than operating entirely from the information provided to the senses. I'm not saying it's necessarily correct, but it does seem well formed, and LLMs only undermine it in a very weak sense (so far?). To defend the conclusion itself I would have to know a whole lot more about how he estimated information content (needed and available).