“Current AI training objectives may not extract the maximum value from each token of data; models could learn more if they were forced to do more computational work on certain tokens during training.”