The authors of fine article did not tested against cache-friendly binary search, for one. This approach can easily give speed up of 2x and more. Another thing

The authors of fine article did not tested against cache-friendly binary search, for one. This approach can easily give speed up of 2x and more. Another thing they didn’t test against is data-dependent middle index choice. Which would play nice with RLE encoding of first bytes of keys, especially if one choose reasonably big number of keys per page, at least 1024, not small number of 128 (or was it 256) in the paper. Especially for log-structured merge trees where learned indices can be used safely. Anyway, original paper was very sloppy at establishing baseline and very disappointing because of that.