rian butala
CS major at UCLA. I reimplement modern ML from the papers up in PyTorch: diffusion, world models, LLMs, because I personally only trust things I can rebuild myself.
currently
- Research assistant at UCLA, working with neural recordings and decision transformers. I joined the lab during my senior year of high school and stayed on.
- Writing the software at WakeMates, where it's just me and the founder getting the product out the door.
- Iterating on my latent diffusion model. The fifth rewrite in eight months finally trains stably; it turned out the problem was the data pipeline, not the architecture. You can sample from it here.
some numbers
- 1.5126 bits/byte, from my under-16MB language model for OpenAI's parameter golf, trained in 600 seconds on a single A5000.
- 144x, how much NeuroZip compresses EEG while keeping it searchable by free text. It took 3rd place at the QBI Hackathon at UCSF.
- 5, the number of attempts it took to get latent diffusion training stably. The first four taught me most of what I know about it.
- 4 years as founding software lead and captain of FRC 10252 / FTC 21781, where I wrote our rookie-year codebase and then taught the newer members how it all worked.
- 0 CSS files on this page.
before ucla
I went to Stratford Preparatory, where I founded the aerospace club and we built autonomous long-range RC aircraft on Pixhawk, ArduPilot, and ExpressLRS. I also built a theft-detection system that the school actually ran, played a lot of tennis, and came out a National Merit Finalist.
contact
Email is the fastest way to reach me: rianbutala@gmail.com. My code is on github and my trained models are on huggingface.