> he would probably say the inputs into an NLP model would need to capture a much wider range of human culture and behavior
I get what you're saying - that something deserving the name "language" way outstrips the simple word embedding strategies in TFA, right? and I wholeheartedly agree - but my point is that the position of PI, at least as I read it, is fundamentally opposed to the idea that modeling - what he calls "a mental mechanism" - is how language proceeds. This is one of the major points of divergence between PI and the Tractatus (which posits an explicit correspondence or modeling relation between language and the world.)
The most explicit formulation of this that comes to mind is in PI #689:
> "There must surely be a further, different connexion between my talk and N, for otherwise I should still not have meant HIM."
> Certainly such a connexion exists. Only not as you imagine it: namely by means of a mental mechanism.
The point is very much that PI's emergent view of meaning is quite in tune with the connectionist sentiment that a bunch of simple elements (i.e. neurons, either in the brain or in some AI project) can exhibit the rule learning and rule following capacities at the centre of PI's treatment of language, without any model of the elements of language needing to be explicitly encoded.
The degree to which such a model may or may not exist, embedded within the parameterisation of a language-capable set of neural components, is an interesting question. But even granting its "existence" I think the point would be that the basic properties of the set of components, coupled with the training and samples provide the capability and the "model," if we want to grant its existence, is effectively a by-product - epiphenomenal if you like. :)