Data- and interaction-driven approaches for sustained musical practices with machine learning

While contemporary discourse on AI often fixates on large ‘foundation’ models, these types of models often offer only limited ability for musicians to express their musical intentions and agency. In this article, following a brief reflection on and critique of the musicality of large music models informed by our own experimentation, we describe a richer set of possibilities for integrating machine learning into musical practices. We outline several alternative approaches within machine learning which can better support sustained musical practices. These include approaches underpinned by an understanding of training data as a vehicle for communicating intention (as opposed to a representation of ‘ground truth’), approaches that leverage small datasets and models, and approaches that employ other interactive mechanisms to capture human intention and agency. We present examples of our own musical projects that illustrate these approaches, and we discuss the implications of our alternative perspectives on data and intention for system development, musical practice, future research, and the future of musicking.