Runner retweetledi
Little side project I’m building...
I’m indexing the lyrics of every top Billboard hit into a giant vector database to finally create GOOD lyrics for AI-generated music.
The biggest problem with AI music right now is the lyrics. They sound like AI. They default to predictable, cheesy vocabulary. It has no soul.
The solution? I'm not letting the AI write a single word from scratch.
Instead, I'm forcing the LLM to strictly use the "Cut-up technique." It uses semantic search to extract the exact words, vowels, and disjointed phrases from proven hit songs in the database, and collages them into entirely new phonetic combinations.
It inherits the raw "singability" and gritty vocabulary of a hit, without stealing the song itself.
This is actually a legendary songwriting trick. David Bowie famously built a custom Mac program in the 90s called the 'Verbasizer' to do exactly this. Kurt Cobain and Thom Yorke used physical paper cut-ups to escape predictable rhyming patterns and find weird, brilliant juxtapositions.
We’re essentially just rebuilding Bowie’s Verbasizer using high-dimensional vector embeddings and modern LLM orchestrators. The AI acts as the collager, the database acts as the paint.
Let's see if we can get a machine-generated song onto the Billboard Hot 100 in the next couple of years. 🎧
English
