“Imprecise” language models are smaller, speedier, and nearly as accurate

Discussion