You could imagine some GitHub Copilot like model that is trained on available music then uses a language like Melrose to predict music. That way the model sends

You could imagine some GitHub Copilot like model that is trained on available music then uses a language like Melrose to predict music. That way the model sends notation to a synthesizer instead of directly creating a waveform. It would prevent the odd artifacts currently present in GAN music.