My interpretation is that it's the other way around. The language model trainer's job is to find the network weights that make the model best at compressing the data in the training set. So what this means is that, say, professional work-speak text samples and hacker l33t-speak text samples are different enough that they end up being predicted by different sparse sub-networks; it was apparently too hard to find a smaller solution in which the same sub-network weights predict both outputs.

All LLM behavior is mediated through language by construction. That doesn't mean the same applies to humans.

I think specifically, certain psychological modes require different levels of articulation, and language is one way to get there in a bandwidth-limited system.

See also: https://en.wikipedia.org/wiki/Newspeak

Language constrains your perception of reality to only the set of concepts conceivable within that language.

Agents who only speak Rust have no conception of what runtime errors are, for instance. Fascists won't understand concepts like "universal human rights" as in their worldview there is nothing universal about humanity as a whole.