March 21, 2026
Contents
Intro
Following my goal to do 24 startups in 12 months, I spent the past two weeks working on something that ultimately ended up nowhere. Here’s what happened.
The idea
At first, the thinking was to provide a web app where a user could generate either a song or a stream playing music which would be clear of any copyright or licensing concerns as it would be AI generated. Under the hood, this was to be done with LLMs producing strudel.cc programs so live coding but written by agents rather than humans.
Whether it’s song parodies or other videos I work on, the audio or instrumentals is an important part of that. I figured this could be of interest to streamers, content creators, or developers working with multimedia in some form.
What I did
I vibe coded a simple Hono app that would use a secure vm to process the generated strudel (since they’re technically untrusted programs). After a few manual iterations I got the stream and song generations to sound almost like real music; as though it was in this uncanny valley of sounding more like 8-bit than something produced by a human.
The realization
I spent bit more than a week drilling into a consumer product that was not yet reaching the threshold where I’d prefer it over something like Suno. Pausing to imagine a world where I do finish something viable, it occurred to me that I’d be frantically swinging it left and right rather than know who would be wanting the sort of product I was working towards.
Having alrady bought the domain timetomake.music, it was a bit disheartening that it was doomed to go no further than perhaps a presentational demo. However, I got audio generation to sound alright for various downtempo or EDM genres which gave me the idea of assembling long-form music compilations.
The attempted pivot
Having attempted to do more with live coding tools pre-LLMs, I was a bit tentative of the viability of programming and audio engineering (at least with my limited knowledge of music theory or music production). After searching through YouTube for 16 particular sub-genres and descriptions for each of them, I got to work on LLMs generating song programs…
Claude eventually had a setup using SuperCollider but it took more than ten times as long to produce audio as the audio itself was defined to be. Once it started suggesting restarting the computer and crossing my fingers, I knew I reached a point where I can either try another desperate pivot or bookmark this as not progressing further.
What went well
I was able to scope out the idea for the product and set up an infinite loop with Claude to run overnight despite tier limits. Plus, I managed to get Gemini to do live coding and make audio that sounded sorta like music.
Learnings
I need to be able to describe either the exact people I’m planning to build a service for or have a solid idea of the profile of customer I’m looking for. Even if I could do the napkin math in advance for figuring out pricing of songs or chats, economics don’t make a business, solving problems do.
The next two weeks will be a jab towards something with a clearly identifiable and winnable market.