There's a lot of good information in this article, but this one seems rather unwarranted.

From first principles it's certainly not intuitive that an assemblage of hundreds of millions of neurons--that is, a system built out of hundreds of millions of fundamental computational units--shares very many qualitative properties in common between its holistic output (e.g. speech) and the units themselves.

Indeed, you get novel properties even out of two neurons properly connected. They can function as a bistable switch when the individual neurons are not switchlike, or convert neurons that steadily fire at a moderate rate to oscillating bursters (when mutually inhibitory to each other, and depending on channel inactivation kinetics).

So the O = Ψ(I) functional relationship breaks down very quickly as you scale up. Indeed, part of the problem of the formalism is that neurons basically all have internal state, so they don't obey a functional relationship like O = Ψ(I); rather, at best, it's (O, S') = Ψ(I, S) where S' is the internal state following from S. (There are temporal dynamics to the state, but you can make those implicit in Ψ as long as you choose a sufficiently short update time between S and S'.)

Anyway, despite seeming very weird--as weird as if "Microsoft Word is nearly perfectly speaking in the language of transistors" were true--we could nonetheless ask whether speech mimics internal cognitive processing at some level.

And although natural language is extremely powerful and general-purpose--you can describe Turing computability with natural language, for instance--there are many instances of it not actually reflecting underlying computations very well at all.

Among a huge list of counterexamples we can note everything from the difficulty of writing computational neuroscience textbooks through to the observation that people "know what they mean" but can't find the words to express it, and that physical skills are almost always better taught by visual example than by verbal description.

What, one wonders, is the linguistic equivalent of the receptive field of a complex cell in V1?

It is true that language evolved because it enabled important survival behaviors (presumably involving coordination and sharing of information), but there is no reason that this has to be at the level of fundamental processing as opposed to good-enough activation of sufficiently common high-level patterns of activity. Rather, it seems more plausible that it is the universality of computation in general, not that language is particularly close to the biological workings of neurons, that gives language its incredible reach. (That plus the survival value of handling single events appropriately: you want to know that the pretty blue berry is poisonous the first time someone says it, and stop putting it in your mouth right away.)