Envoyer par SMS: Solving inefficiency of self-supervised representation learning