a new era for kobimusic
a 4-year journey in my youth creating the SOTA midi generation model, and the future of kobimusic.
Kobimusic started years before I made a website, and years before I made the itch.io app. It actually started on Scratch.
Some years ago, roughly 2020, after the announcement of gpt-3, I became interested in AI, and the hobby of making music around then too. I had been using Scratch for years, but never built anything substantial until that point. So I decided to finally take on my first large project, ever.
The goal: build a model to continue MIDI music. The plan? None. With my later AI scratch projects, I honestly don't even know remotely how they even slightly functioned at all. My strategy was literally, multiplying numbers in a list. Then randomizing the numbers a bit each iteration, if it got a lower loss, keep evolving, else revert. So it's kind-of like a psuedo evolutionary model? Probably the worst part, was the 'tokenizer'. With the MIDI model, I did have a tokenizer, but the chosen index from the list, what the result of 2 (float) multiplications, which, obviously didn't work.
Did the project work? .. No. But of course, over the years I've gotten more-and-more experience, and now I'd like to imagine I kind-of know what I'm doing.
The first real success I had, was ~late 2022. I took a stock gpt-2 notebook and trained it on direct csv-ified MIDI events. I tried it on the midis I had in my downloads folder (still building up to this day, 10,000 and counting). It actually kind-of worked (followed music theory a bit, short phrases that were plausable). There was a specific generation that I realised, this might actually be possible. That encouraged me to keep going.
June 2023 I created the first version of the itch.io app and make my first 'foundation' model. My Junior year of highschool would go on to be the most transformative year of KobiMusic yet. By the end of the school year, I was beginning to superseed Musenet in capabilities. At the same time, I was blowing up on the internet on my promotion channel. In March 2024, I remember sitting in class, my phone constantly being buzzed about either a new itch.io purchase, or a new youtube comment. It wasunlike anything I had experienced before. Naturally I worked harder and harder, to create, not just a competitive model to musenet, not just a crappy app. But an indistinguishable model, and a UI I could be proud of.
By the time I finished though, the hype had dried up. Mostly, because my youtube was pivoting on a fad and not being sustainable. But users wouldn't stick because, it's simply wasn't a good app. All of my effort was basically for nothing, as I thought.
For the next couple years, I would stay dormant from the AI scene. I was honestly, burnt out. All of my projects from that period were everything but that. Including a project that's an ai music generator .. without AI.
April 2026, during my first year of college, I mostly was helping people with their startups to make cash on the side. No startups on my own, a limited amount of personal projects. But, I'm not sure what hit me, but I was reminded of the field, looked it up again. I realised since I had left the scene, basically nothing had changed. To this day, there is still so much to do. So I got to work. Over these past few months (Mostly took a pause due to my AI internship), I've finally started again. Instead of focusing on one small piece of research, I aim to contribute to AI research as a whole.
Throughout the years, there had been 2 phases of KobiMusic. The early years, had passion, stubborness but naive-ity. Later, it grew more into growing experince, wisdom, but at the cost of burnout. Now, with the next phase, I plan to usher in a new era of both passion, and expertise built from my years. I'd like to try my best to be an open-source beacon for musicians around the world, building what they actually want.
Anyways, I'm starved for compute right now (I bought a computer with a 5060, so AI experiments are very slow). If you're a cloud provider wanting to give me compute, please, please do (contact me).