Not very impressive/interesting, they aren't releasing their code or models, and it seems that they had to use lots of rather obscure hacks to make the whole mu

Not very impressive/interesting, they aren't releasing their code or models, and it seems that they had to use lots of rather obscure hacks to make the whole multi-instrument thing work. (Though I suppose it definitely is impressive in other ways, especially hacking their existing GPT architecture to do something worthwhile with non-text tokens. So, I'm not saying that any sort of work should be dismissed outright!) I do still think that the good old https://github.com/hexahedria/biaxial-rnn-music-composition https://github.com/hexahedria/biaxial-rnn-music-composition (Hexahedria's Biaxial RNN/Tied Parallel Networks, also published a paper achieving SOTA at least for that time, on a variety of midi datasets) has more interesting/compelling output musically, despite starting from a comparatively tiny dataset and using rather elegant and easily-undestandable convolution- and RNN-based techniques. Too bad that the implementation relies on Theano which is quite endangered at this point (doesn't seem to support up-to-date python 3?), but I do think it's a compelling starting point if anyone really wants to work on this domain.