“Language modeling benefits from a virtually infinite amount of training data available by scraping the web.”