I can ask a LLM to write a haiku about the loss function of Stable Diffusion. Or I can have it do zero shot translation, between a pair of languages not covered in the training set. Can your "language JPEG" do that?

I think "it's just compression" and "it's just parroting" are flawed metaphors. Especially when the model was trained with RLHF and RL/reasoning. Maybe a better metaphor is "LLM is like a piano, I play the keyboard and it makes 'music'". Or maybe it's a bycicle, I push the pedals and it takes me where I point it.