That isnt relevant to whether NN models which record correlations in text are a model of language; they aren't. It isn't an open question.
This is an argument from ignorance: you dont know what caused the universe, so maybe god did.
Well, we do know much more than enough to regard any use of NNs today as having nothing to do with intelligence; and likewise, that any model of correlations in symptoms of human social activity (text, books, etc.) as not relevant to modelling/simulating intelligence.
A neural network model is just a set of averages over training data, which in this case, are simply recording correlations in word frequency. There's no neural "Neural network", it's just some mean()s.
> neural network model is just a set of averages over training data
We could claim the human brain is just 100 million neurons arranged in a mesh were each neuron has their own 'activation potential'.
Neither your or my statement are wrong. But again until we can pinpoint what (human) intelligence is, then claiming squid or neural networks are unable to be intelligent agents is just as emotional as the people you criticize.
Right, but it isnt. The neural network algorithm is just one in the whole set of non-parametric curve-fitting algorithms which approximate a dataset with a function whose structure is given by averages over the dataset its given.
It doesnt matter what algorithm in this family we choose, they all just produce correlative models of the data given. So if we give a NN a library of books, it produces a function whose structure is just correlations of words in those books.
To believe that the structure of this function has anything to do with language, the brain, or anything else is pseudoscience. It isnt close to science, it isnt an open or a philosophical question, it's homoeopathy.
Language use in humans is not a function of correlations of word frequencies. We know this, as much as we know where mars is.
No human is shown a trillion books and arrives at an internal model of language corresponding to word frequencies in those books. We know that it would be impossible to do this, since words refer to the world -- and correlations between word frequencies dont.
Your comment here defending this is equivalent to a homoepath saying "we dont know everything about chemistry". We dont know everything, but we do know that diluting with water doesnt make substances more efficaious.
And we do know that the correlational structure of text in books captured with any algorithm whatsoever, has nothing to do with modelling language. We know any such system cannot use language, as at least, it has no model of how words refer to the environment.
I fundamentally agree with your larger point that the map is not the territory when it comes to language.
However:
> Language use in humans is not a function of correlations of word frequencies.
I would argue that this is untrue. For example, people pick up on common expressions (and other behaviors) of those around them. If I hang out with people that say "catch my drift" a lot, I will probably also start using "catch my drift" more. I find it very plausible that "my internal model of language" is in many (but not all) aspects driven by the word correlations I observe around me.
I do wonder about the absoluteness of your statements though. Would a bed-ridden person that has no sense of touch/sight/smell etc. except hearing (i.e. only able to interact with the world via language) be "not human" to you? If not, how is "the environment" different for it and GPT-3 (other than context window)?
A model of gravity can be used to generate abitary night skys, but correlations of pixel patterns in sky-images isnt a model of gravity.
When I say "not a function of" i mean, explicitly, that the model does not have a term which depends on these correlations. Just as F=GMm/r^2 does not have a term for "image pixel weights".
Real Language is what wrote all those books, as much as gravity is what made the night sky. Pixels and word frequencies are symptoms, a few of an infinity of symptoms -- distant effects, whose correlations are not models of their causes.
There’s no such thing as real language. Language is loseless compression of perception to allow collaboration of agents. NNs also do loseless compression (from already compressed data - words). I think with more raw “senses” their compression will get better. Only missing thing is some memory trick. I would even bet that most humans use heuristic thinking and “logic” is just a way of validating legitimacy of these heuristic proposed patters. Once someone replicates this in AI it will get scary. Already things like “Let’s think step by step” are going in this direction. Hierarchical generation of actions might also help (and model understanding that these actions could be used to cause effects). I agree NNs aren’t enough for RL agents but it’s great bootstrap for world model component of agents that probably somewhere between cat-humanity level
>We could claim the human brain is just 100 million neurons arranged in a mesh were each neuron has their own 'activation potential'.
You actually can't because, unlike neural networks (we know exactly what they do), we don't have anything even close to an exhaustive view of the functioning of the human brain.
This is an argument from ignorance: you dont know what caused the universe, so maybe god did.
Well, we do know much more than enough to regard any use of NNs today as having nothing to do with intelligence; and likewise, that any model of correlations in symptoms of human social activity (text, books, etc.) as not relevant to modelling/simulating intelligence.
A neural network model is just a set of averages over training data, which in this case, are simply recording correlations in word frequency. There's no neural "Neural network", it's just some mean()s.