I saw bellard neural net compression benchmarks. If you have time and performant system, based on the semantic of the data to compress, neural net seems tough to beat.

It is a matter of good compromises: speed and efficacy. I even wonder if LZMA2 is worth replacing bzip2.