“Effective distillation techniques could enable the creation of very small, specialized models that mimic large model performance on narrow tasks like reformatting tables from Markdown to LaTeX.”