The idea behind this experiment: lyrics may make ideas about love, success, power, and identity feel familiar before we consciously examine them. Build a mind from the words you listen to, then question the worldview that emerges. What do you recognize? What would you challenge?
Train the mind, then give it a question.See what a life made only of lyrics can say.
Only your description reaches the text model. It has no vision training.
Responses are raw model continuations, not factual answers. Questions never become training data.
The idea: music may carry value systems in through emotion, repetition, and familiarity. This project asks whether interacting with a mind built from the lyrics you hear can help you notice and question those values. That is an exploratory premise, not a claim that music bypasses a specific brain region or that this model reproduces your actual subconscious.
This is a small, real neural model trained from random weights in your browser. Its alphabet and next-character patterns come only from the active lyric corpus. There is no language model API, pretrained embedding, persona prompt, or outside knowledge in the model. Lyric acquisition can call LRCLIB before training.
The distinction: learning lyric patterns is not the same as learning to answer questions. Expect fragments, invented words, repetition, and memorization. A stable personality is a hypothesis to evaluate, not a feature we can promise.
Every request starts with fresh recurrent state. It uses your current question as a text prefix; earlier conversation, metadata, and image pixels do not enter training. Unknown characters are skipped and reported.
Your playlist: paste an Apple Music text export, Spotify playlist JSON, a CSV, or a list of titles and artists. Then reuse cached lyrics or look up missing songs. The lab shows every missing song and requires lyrics for the complete selected playlist before training. It never fills the gaps with demo songs. Playlist titles, artists, and descriptions do not enter training. Repeated song entries are trained once; this pilot does not model listening frequency.
Billboard plan: use ranks 1–40 from each genre chart, collect 2025–2026 weeks, and deduplicate recurring songs. Pop uses Pop Airplay, country uses Hot Country Songs, hip-hop uses Hot Rap Songs, and R&B uses Hot R&B Songs. Some charts have fewer than 40 entries. Imports are user-supplied records; this lab validates their format but does not independently verify chart claims.
Use lyrics you can use for this experiment and record their source and permitted use. A lyrics display license may not include model training. The bundled examples are original demo verses, not real Top 40 songs. Lyric lookups and library imports are cached in this browser. Playlists and trained weights last only while this tab remains open; export your model to keep a record.
Images: a text-only training corpus teaches no visual recognition. Attach a picture and type your own description to condition a response. Adding pretrained vision would change the experiment.
Paste song titles and artists. Your playlist can cross genres and years. These exports identify songs. The lab reuses stored lyrics and looks up missing tracks.
Apple Music on Mac: select your playlist, then File → Library → Export Playlist. Choose Text and paste the file's contents here.
Spotify: its account-data download includes playlist JSON. Paste the playlist file here and choose one playlist if the export contains several. CSV exports with song and artist columns also work.
Simple lists use Song title — Artist, one song per line. Playlist links and direct account connections aren't supported yet.
Song title — Artist
Choose the correct artist and song version. The lab will not guess between different lyric texts.