Machine Learning and Metrology
There are many, many papers in machine learning these days. And this paper, taking a step back, and thinking about how researchers measure their results and how good a specific type of benchmarks even can be - crowdsourced golden sets. It brings a convincing example based on word similarity, using terminology and concepts from metrology, to show how many results that have been reported are actually not supported by the golden set, because the resolution of the golden set is actually insufficient. So there might be no improvement at all, and that new architecture might just be noise.
The story of the Swedish calendar
Frozen II in Korea