Was thinking of work like https://arxiv.org/pdf/1811.12817.pdf although the compression improvements vs classical methods are much more impressive for lossy compression
Are you referring to the work of Fabrice Bellard? I thought he only showed the proof of concept for text data, using language models, only?
I am referring to the work by Fabian Mentzer and related approaches, which covers both lossless and lossy compression.