Part of Temporal Constructs Appendix, which contains details behind the research: dates, quotations, descriptions of instruments, and sources.
General definition
AI models trained on large amounts of text to predict and generate language. Large language models underlie conversational assistants and AI agents.
Relation to the practice
Language models are the main kind of AI the professional doctorate uses. The doctorate uses no sound-generative AI in its works. They work in the symbolic domain, through natural language, programming languages and control data.
While I’m highly critical of the design principles behind today’s AI technologies, and how they endanger the livelihoods of many creatives, I do feel language models can make the past experiential in very unique and powerful ways.
artist statement
Because the command vocabulary is bounded, small models running locally can operate the system, and larger models can take on more decisions. Language models also allow reasoning, which here means breaking a task into smaller steps, and each step contributes to the final result. Reasoning allows a language model to follow the thought process that was once intended for human users.
Magenta RealTime 2 set beside HMSL, June 2026.
Audio models compared
The doctorate set an open music-generation model, Magenta RealTime 2, beside HMSL in June 2026. The comparison turned on transparency and on tuning. The model’s output is audio, and the decisions that produced it cannot be read the way a symbolic MIDI score can. Its control layer is a twelve-tone equal-tempered piano roll, so microtonality is outside it however much the model is fine-tuned. That exclusion matters here, because the lineage the doctorate works in, from Mills College, took experimental tuning as one of its subjects.
A camera view, the move to multimodal models.
Vision as input
Providing an agent with a camera view is the step from a symbolic-only domain to multimodal models. The camera view is what allows an agent to programme a physical instrument such as a robotic light from a visual reference.
The distinction drawn from Mahowald and others 2024.
Fluency and competence
The distinction the account makes is between a model’s fluency and its competence, from Mahowald and others 2024. Fluent output is not evidence of reasoning underneath it. That matters when the doctorate uses language as the way into a system. The system carries out a request phrased in speech only where it already has something to carry it to.
What comes back is a prediction of likely text. The command language and the protocol the model addresses do the work. To treat the reply as a computed result would claim more than the account can support.
Added as the research runs.
Unforeseen
Here the unforeseen situations, sometimes difficult, sometimes unexpectedly interesting, will be described as the project runs. Many of these unforeseen situations have already occurred and have been recorded. They will be added in the time to come.