“Applying techniques like dropout and training for many more epochs on existing text data could allow for the creation of more capable models, mitigating concerns about running out of new data.”