Every Time I Hire a Linguist, Inference Costs Go Down

Flipping Fred Jelinek’s famous aphorism about firing linguists to improve speech recognition, this paper explores an inverse thesis for modern language models: linguistic insight can directly lower inference expenses. As brute-force context windows and parameter scaling hit economic limits, revisiting grammatical structures, tokenization efficiency, and prompt syntax offers an alternative path toward curtailing operational overhead without sacrificing comprehension accuracy.