It might not be surprising when they implemented it by supervised learning, but when they also solved it by reconstructing the input it was like a rediscovery.