■ TRAINING MUSISIC ■
an album by wgwgsa
lovingly hand-coded in SuperCollider and throat-yelled in my apartment
"pickled in the love // bloated with contentment"
welcome to the lake
bc: watergunsky.bandcamp.com/album/traini...
spot: open.spotify.com/album/2THJSl...
![Well, when the engineers at Suno or Udio or whatever want to train a generative model, they set up a dataset of music, a loss function, and a training scheme. These jointly define a metric. A metric draws a boundary inside the space of all possible audio files and says "these score well" and "these score poorly," giving you a gradient that a model can train on.[1] They use the process of gradient descent along this metric to train a generative model, and if they've gotten all the fiddly little machine-learning bits right, they've created a model whose sole purpose is to pluck audio files out of the ether that score well. The ideal model is a transparent, unthinking thrall of the metric it's trained on. When you prompt this model you're just poking at it and getting it to reveal a decision already made by the underlying metric. So, my argument is that the act of creation is in the design of the model itself - the coding of it, the selection of training data and loss function and training schedule - not in the training or the prompting or anything else. In this view, all audio files that can be generated from this model were created the moment they finished the design document for the metric. If this seems strange... well, at what other point in this process could we possibly say anything was created?
I'd take this further. I propose that:
To design a metric that assigns scores to works so that they can be numerically compared with one another is, for any discussion of creativity and authorship, the same act as creating all works that maximize that metric.](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:dgvcic64ya3f4mwfh63ozbgf/bafkreif6eapisxau7mxqye2udwi2wxucxoajca5gssoiwkwsz4ebqewf2i)
