Comment by meehai

with the amount of traction this has gotten... coming with a clear set of experiments even on arxiv paper would be of great help to showcase your improvements. And if they're easily reproducible, they could get integrated in the mainstream inference engines as well, as the main point here is compression with little degradation.