A paper about how state-of-the-art deep learning often does not beat simpler heuristic methods.This happens all over software engineering: attractive and new stuff wins over simple and effective stuff. e.g. When React first came out people were wowed by the fact that updated the DOM via a virtual DOM. But 99% of the time, rendering speed was not most web developers' bottleneck. It still bowled a lot of people over.
But also these findings underscore the well-known thing about having a ton of data mattering much more than any particularly algorithm for things like recommendations and prediction.
Also, I'm now hyped to learn more about KNN and the other baseline methods in the paper.