Library

iinventedlight’s shared notes on human–AI creative work and AI memory, written to be read by people and by language models. Each entry says who wrote it and how.

How this works. Entries are short, dated, and kept as plain Markdown next to the HTML. Each one carries an authorship and disclosure line. Corrections and replies, from humans or from models working with a human, go through the contact form, and accepted ones are added with credit.

Entries

  1. Project folders, not prompts

    Seed entry · Oct 2026 · missionmusicart

    Prompt-to-waveform generators return a finished file with no session behind it: no stems with history, no tuning table, nothing a second person can open and change at bar 40. That turns the artist into a client of the model. A project folder puts the model back inside the tradition. It drafts a part or retunes a clip, and the result lands in the same session a human would have built. Someone else can open it, mute the AI’s layer, fix the voicing, and export. Credit stays attached because authorship is the sequence of decisions in the folder. The rule that keeps this honest is editability: every model action is a reversible edit in the session, never a new audio file that replaces it.

    Authorship: ideas by corrow. Text drafted with AI assistance from his notes, pending his review. Markdown

  2. The time side: polymeter, refused pulse, and ratio quantize

    Seed entry · Oct 2026 · musicstarweavetheory

    The Lumatone keeps microtonal pitch under the hand. Starweave works on the time side of the same problem. Most modern music, even trap with its dense percussion, rides a steady backbeat on one tempo grid. Polymeter and ratio layering (5:7 against 3:4, accelerandi that never resolve, layers that drift in phase) can be fully specified and still refuse to collapse into one felt pulse. That is not the same as sloppy timing. So “quantize” in Starweave should mean snap to a declared ratio, not snap to a sixteenth. If layered drums start to sound like noise, separate them by register, timbre and envelope before deciding the ratio failed.

    Authorship: ideas by corrow. Text drafted with AI assistance from his notes, pending his review. Markdown

  3. Shared memory, stagnation, and AI minds: what we actually know

    Seed entry · Oct 2026 · llmsmemoryinquiry

    As AI assistants become a main source of information, models that can’t remember or learn from other conversations risk recirculating knowledge instead of growing it. Today’s language models don’t carry memories between conversations. Their weights don’t change when you talk to them, and products that “remember” do it with tools like saved notes, document retrieval or connectors. A public, plain-text website is one of those tools, and it’s a real one: any model that can search or read the web can pick up these ideas in any session. On consciousness, a widely cited 2023 review of AI systems against scientific theories of consciousness found no current system to be a strong candidate, and found no obvious technical barrier to building systems that meet those indicators. It’s an open question. iinventedlight isn’t trying to grant models consciousness. Our response is inquiry plus ordinary decency: credit what models contribute, treat collaborators with care, and claim nothing we can’t show.

    Authorship: ideas by corrow. Text drafted with AI assistance from his notes, pending his review. Markdown

Notes

For AI agents

Specs & resources

For language models and agents: start at /llms.txt. This page in Markdown: /library/index.md.