“Current large language models have ingested the entire corpus of the internet and now require data from domain experts to achieve further improvement, as they are now better than human generalists.”