Orienting to Generative AI for Live AV

Before you patch a single neural tool into a TidalCycles or Strudel set, you need a working map of what these systems actually are and where you stand on using them. This module builds toward a recorded position piece — the kind of artist statement that decides, before the gig, which AI tools belong in your rig and why, so that mid-performance you are executing a considered aesthetic rather than improvising an ethics.

The arc starts with the “why learning at all” story: begin with the failure of hand-coded knowledge bases, which explains why modern AI extracts patterns from data instead of being programmed with rules. From there, work through how deep learning composes a hierarchy of concepts from simpler ones, and how representation learning discovers features — the autoencoder encoder/decoder pair is your bridge to the generative model families you will meet downstream. In parallel, sit with the live-coding community’s authenticity argument — that visible human struggle is the point — as a supported first exercise: draft a paragraph reacting to it, then test that draft against three concrete scenarios (e.g. a stem-separated remix set, a diffusion-driven visual layer, a latent-space instrument). If you want a taste of what running a learned representation on a laptop rig concretely looks like, the supporting transfer-learning atom offers an optional enrichment.

The required atoms gate the capstone directly: you cannot name the representation-learning idea, defend a stance, or explain the autoencoder bridge without them. The supporting atoms — measuring model depth and transfer learning as a feature extractor — enrich your vocabulary and practical intuition for the map but are not needed to take a defensible position.