This is very interesting but for people who are not from the field note 1. The compressors he is comparing to *are* neural networks themselves (PAQ) 2.

This is very interesting but for people who are not from the field note 1. The compressors he is comparing to *are* neural networks themselves (PAQ) 2. He is using a rather small corpus enwik8 (100,000,000 bytes), instead of the common enwik9 3. He is not publishing the amount of RAM or hardware he used http://mattmahoney.net/dc/dce.html http://mattmahoney.net/dc/dce.html (seems down) https://web.archive.org/web/20210216231934/http://mattmahoney.net/dc/dce.html#Section_436 https://web.archive.org/web/20210216231934/http://mattmahone... It seems Bellard is starting to play with these. I bet he will find interesting things and I would love to see him interact in the Russian compression webforum (please don't link). Big names in compression like Yann Collet, Matt Mahoney, and Jarek Duda used to hang out there.