It all comes down to generating tokens. The dominant paradigm is to compose the music by generating the next token based on the last few seen tokens, then take that and generate the next token, and so on. No one has been able to significantly improve upon that. All development in the last ten years or so has been about creating ever larger models with more data so that emitted sequences will stay coherent for longer. The problem with this paradigm is that it works amazingly well for short sequences, but often fails to stay coherent for longer sequences. So composing interesting ten seconds long instrumental music is a solved problem - composing full songs three-to-five minutes long is far beyond the reach of current state-of-the-art.

Text-generating models suffer from the same problem. E.g. GPT-3 will generate two to three-hundred tokens or so and then it will begin to spout nonsense. That it still works so well is because it's trained with a massive corpus and because its tokens are richer. It operates on the world level, while most music composition models operate on the note level which are more analogous to characters than words. If someone were to be able to figure out how to represent multi-instrumental compositions using "music vectors" (analogous to word vectors) that would probably lead to a major breakthrough.