How would you modify a standard transformer architecture to better handle long-form temporal dependencies in video at Mirage?
