Too theoretical to be pivotal. Years ago we had results about how all sorts of learning was reducible to kernel SVM and that drove a big fad for kernel SVM rese

Too theoretical to be pivotal. Years ago we had results about how all sorts of learning was reducible to kernel SVM and that drove a big fad for kernel SVM research at NeurIPS and other hot conferences. Nobody cares about SVM anymore. So we don't have a good enough grasp on the theory of AI to predict what is going to be a breakthrough. Only rigorous empirical results convince. Second, the idea of meta-learning gradient descent, or that some learning algorithms can do it, isn't new or specific to Transformer architectures.