There have been many attempts to model and emulate human syntactic acquisition and processing, but the general consensus is that it cannot be done without presupposing some mechanism that enables hierarchical structure. The number of tokens a child needs to learn syntax is the tiniest fraction of the amount of tokens an LLM is trained on.

Humans can also lose parts of their language processing capabilities, without losing others (start at e.g. https://en.wikipedia.org/wiki/Language_disorder), which is highly suggestive of modular language development. The only question on which there isn't much consensus concerns the origin of that modularity. And humans can lose knowledge while still being able to speak and understand, or lose language while retaining knowledge.

LLMs don't have that at all: they predict the next token.

LLMs does have that, or at least it’s very likely that we will eventually be able to manipulate LLMs in a modular way (see https://news.ycombinator.com/item?id=40429540). One point remains: humans learn language with much fewer tokens than LLMs need, which suggests presence of a priori knowledge about the world. The LLM metaphor is finetuning, so babies are born with a base model and then finetuned with environment data, but it’s still within LLM scope.

> presence of a priori knowledge about the world

1. A certain architecture (e.g. a module that enables syntactic processing) is not knowledge about the world.

2. We model the world according to our capabilities.

3. Modular language models have been tried, but did not meet with success.

4. The link you include is about the conceptual space, which is not (directly) related to human syntactic processing.

5. The question is not about metaphors, but about reality.

6. Babies aren't born with a base model and fine-tuned. They learn. This is the metaphor NNs are actually based on.

I don't think LLMs have all that much to do with "innate grammar".

"Innate grammar" are essentially the meta-rules that govern why the rules are what they are. For instance, an English phrase can be recognized as valid or invalid by other native speakers according to the rules of the language. But why are the rules what they are?

This is especially puzzling due to the dazzling variety of human languages. And the fact that, after a period of immersion, humans seem to have the natural capacity to learn all of them.

How do LLMs fit into this? Well, I think it would be interesting if we left a group of LLM to talk to each other for 1000 years. Then see if 1) they developed a new language branch 2) that could be relearned by humans through immersion alone.

It's true that LLMs have learned (have they? I suppose that's a loaded word) human languages like English. But it's unclear if they are governed by the same meta-rules that both constrains and drives the evolution of humanities thousands of distinct languages.

No. Innate grammar has always been about how humans aquire language, not how any possible system which understands human language must posses that innate grammar.

Trying to put his in an uncontroversial way: the human brain (or a brain plus paper and a pencil) can be turning complete/equivalent. Therefor a human sitting down with a pen and pencil could, in a painstakingly long time, compute the backwards and forward passes of a transformer network.

Therefor a human with no understanding of grammar/language, and using no innate biological circuits, could process grammar and respond with language.

The flaw in this argument would be how to teach a human to do this without grammar ...

But that has never been proven that this is how indeed human acquire language; it is essentially a hypothesis. We may as well do it the way LLMs do - some undifferentiated networks acquires the grammar by unknown means.

LLMs are universal approximators and can pick up patterns in sequences that are very different from Human languages. Sure, they don't have many inductive biases and can understand language, but as a consequence require a tremendous amount of data to work. Humans don't, which implies a certain bias towards Human language built into our heads. A bias is also implied by the similarities across Human languages, though what structure(s) in the brain are responsible is not exactly clear.

It still does not proof anything, as claiming that "there is certain bias for Human Language built into our heads" is quite different thing that saying there is some universal grammar in the brain structures, as much we do not have innate abilities to comprehend calculus or play chess, yet we still able to learn it, with a lot less training information than LLMs. In fact 2 books will suffice for the both.

My comment was more of a response to

> We may as well do it the way LLMs do

We almost certainly don't learn the way LLMs do, it's just too data inefficient.

And I don't see what current LLMs can say about a universal grammar in the Human brain, unless there is proof that a LLM-style attention mechanism exists in the brain, and that it is somehow related to language understanding.

We don’t learn language from textbooks though.

Chomsky has explicitly answered this: Moro has shown in experiments that humans do not appear to be able to learn arbitrary grammatical structures in the same way as human-like (hierarchical) languages. Non-human like languages take longer to interpret and use non-language parts of the brain.

LLMs on the other hand can easily learn these non-human grammatical structures which means that they are not the way humans do it.

Compared to an LLM, how many hundreds of gigabytes of text do humans need to acquire a language? And isn’t that disparity already proof that some sort of innate structure must be going on?

Or that llm learning algos should be further improved, which will happen at some point. I remember Kasparov's tirades to the tune of I have an eternal soul therefore computers can never beat me in chess.

LLMs are an approximation of all the human media they consume. An LLM cannot exist with out human circuitry. It's at best an ersatz language user.

Unrelated to what I said, with all due respect.

It isn't tho, if you look at the bulk of tokens needed to train gen1 over LLMs and what is possible with better data and smaller models.

The fact that LLMs trained on dumptrucks full of data cannot achieve what a middle schooler begrudgingly achieves using existence and snide remarks.

https://www.discovermagazine.com/the-sciences/fruit-fly-brai... He mentioned a structure and scientists hacked a fruit fly Kenyon organ to process language which it does pretty well, also at MIT.

The approach is relatively straightforward. The team began by using a computer program to recreate the network that mushroom bodies rely on — a number of projection neurons feeding data to about 2,000 Kenyon cells. The team then trained the network to recognize the correlations between words in the text.

The task is based on the idea that a word can be characterized by it its context, or the other words that usually appear near it. The idea is to start with a corpus of text and then, for each word, to analyze those words that appear before and after it.

Somehow the transformer architecture does pretty well at this task, and other architectures do not. You could say a transformer has "innate grammar", while other architectures do not.

That an LLM does well at grammar doesn't prove or disprove this possibility. A more poignant criticism of "innate grammar" would be that it's not a hypothesis that can be disproven, and as such not really a scientific statement.

I think the popular perception is that his theory is extremely important, as far as I know the academic consensus is that while hugely influential it is long obsolete.