Better Gzip Language Model from Beam Search

The window for gzip is way too small to expect good results for this because it is representing the whole model as opposed to just the attention window in an attention model.

With the huge dictionary of LZMA

https://en.wikipedia.org/wiki/Lempel%E2%80%93Ziv%E2%80%93Mar...

you can build a better model but then your model will not fit in cache and performance will be worse.