AVT would be a strong candidate for tasks beyond anticipation, such as self-supervised learning, general action recognition in tasks that require modelling temporal ordering, and even for discovering action schemas and boundaries.