If true, this seems like a good thing. Is there any general way proposed in the paper to go from deep --> shallow while preserving the learning function?

This list is pretty good but I suspect infrastructure work is far from done. Also Marketplaces aren't quite native by this definition-- cdixon is more so pointing out that their reaches can be expanded to the unbanked. I think the real magic of Bitcoin lies in the Blockchain itself and its application to the real world via smart contracts. The fact that bitcoin is a currency might turn out to have been a big distraction or just a side effect of the real applications. If this is true it further points to the need for ubiquitous infrastructure.