I think the reason people are impressed is it's assumed the accuracy will improve quickly over time as architectures improve and the models increase in size, just as it did in other areas of machine learning like image classification.