When training a model on multiple tasks simultaneously hurts performance compared to training on individual tasks separately.