The process of training a system to predict videos involves learning a representation of those video clips that is maximally informative but abstract, allowing for easy prediction of the subsequent video clip. This ensures that important information is retained while non-essential details are dropped, making prediction possible but not overly complicated.