>We’ll need to break out of the Chomsky hierarchies and develop some new theories of language.

We might need to develop some "new theories of language", covering other areas, but why exactly do you think we'd need to "break out of the Chomsky hierarchies"?

Nothing about LLMs challenges Chomsky hierarchies.

They do challenge Chomsky's ideas about language being an innate human quality, but that's about it. Besides that's not related to the Chomsky hierarchy of language, e.g. regular, context-free and so on. Those will continue to remain a hierarchy describing languages based on capability levels, whether there's AI, and AGI, or not.

The Chomsky hierarchy will remain unchanged and very useful for constructed languages. I didn't mean break out, as in create a higher level. LLMs are not going to magically become more than a turing machine or anything. I meant we need to stop thinking about spoken language in terms of the hierarchy and try to find another approach.

You can't tell me with a straight face that actual spoken language fits well in that hierarchy. It seems we can shoehorn it in with a 100B parameter turing machine definition, but that just makes my point that the constructed language hierarchy is not well suited for describing spoken languages if it takes such a large definition.

We should be able to find something like a new probabilistic theory of language that will do a much better job of describing spoken languages. Such a theory could help explain some LLM behaviors we see, and apply generally to other forms of noisy communication.

> 100B parameter ... not well suited for describing spoken languages if it takes such a large definition.

Humans by contrast use about 700T synapses to support our own understanding of language.

> To me it shows how fundamental language is to intelligence.

I believe most human intelligence is based on language. We are smart because language gives us the keys to solving all sorts of situations, based on the experience of prior generations. We are wondering if GPT architecture can generate real understanding, but forgetting the training corpus. The language corpus is the real magic here. The source and accumulation point of intelligence.

Take a baby and give it access to language, you get a modern human. Take a random init neural net, use language corpus for training, you get GPT4. It doesn't matter what the model is - brain or transformer, it doesn't matter if it's a transformer or just a RNN (RWKV). What matters is the training signal.

We should not be wondering how come computers got so smart now, we should be wondering how smart is Language. Language is a separate evolutionary system of ideas and concepts. It is a self replicator, like DNA.

You're just describing information theory. Every message is contextualized within the codebooks used to encode/decode it. Sometimes we call that codebook "language", sometimes it uses another representation.

Yes, we should develop a new theory of language based on an information theory approach rather than a discrete symbol approach, or maybe something involving both.

But you can’t say information theory as it stands gives us a completely satisfactory model of human language. Something will need to be built on top of it.