How does an album develop an idea?
A theme can return with different words. A new voice can complicate what an earlier song appeared to say. An ending can resolve a tension, preserve it, or make us reconsider the beginning.
I study those relationships through close reading, NLP, embeddings, and LLM-based experiments. The question is whether a computational account preserves the distinctions that matter to an interpretation — and what evidence would let us disagree with it.
Music is a subject of this research in its own right. It also provides demanding examples for AI engineering: memory, source attribution, multilingual retrieval, and the cost of reasoning over evidence.
Start with the album-memory studies
What Should an Agent Remember? compares four ways of retaining evidence across four albums. It asks what is lost when an interpretation depends on songs far apart in the sequence.
Memory Is Not Context follows that question through retrieval, verification, and resource accounting. An answer may satisfy a model verifier while still making an unsupported claim.
These studies work with model-produced paraphrases of lyrics. They do not observe the full musical performance, and they do not establish definitive narratives for the albums.
Inspect the designs, outputs, and review materials →
Earlier explorations
When Lyrics Change Language: Aquamosh examines the association between language transitions and similarity thresholds. Its revised interpretation distinguishes an automated judge from human validation.
Attention Windows: Beatles and Pink Floyd now documents an unresolved numerical provenance issue and explains why a similarity statistic cannot simply be interpreted as listener attention.
Both articles have dated entries in the correction log.
Questions still open
Interpretation and agreement. Which relationships can readers support consistently, and where should disagreement remain part of the result?
Larger and more varied corpora. How much of an observed pattern belongs to selected songs, languages, or genres?
Lyrics and sound. What changes when harmony, timbre, recurrence, performance, and transitions enter the evidence?
Application. Which failures suggested by musical examples also occur in a particular retrieval or agent system? That transfer needs its own evaluation.
I am a vinyl collector and a serious listener. The listening keeps the computational representation in perspective: a model sees what we give it, while a musical work can hold more than the experiment observes.