“A potential continual learning process involves having AI model modules with both large and small representations, continually distilling the large version into the small one, then deleting the large one and adding new parameter capacity to the small one to learn more.”