The development of a single deep learning model that can be trained once and used for various tasks, whether in language or vision, presents an exciting opportunity to make models more powerful. However, there are still challenges to be overcome in terms of growing the model's capacity and determining how its weights should change as new tasks are added.