What Polyglot ultimately asymptotes to is a structured framework for humans to distill from AI.
Language is the realm for which it is easiest to build a curriculum that covers absolutely everything the model knows, weighted by how much each datum matters in real life, through dictionaries and corpus frequency statistics.
Learning by doing, using a really high fidelity (and relatively fun) simulation (LLM roleplay) of real life usage of that skill, is easiest with language.
Which is why learning languages is the ideal first step towards enabling humans to learn any skill they want through distillation from AI. (Coding's the optimal next step for our method, once we've perfected natural language.)
Such a structured framework for distillation need not necessarily be useful for humans alone ;)