blog

Write-ups of the things that don't fit in a README: the bugs that took a week, the ideas that didn't pan out, and the occasional one that did.

• A month on Trainium, mostly being wrong (2026-10-10)
  Training a language model in 30 minutes on a chip I'd never used, with AI agents writing the code and me trying to catch their mistakes.
• Hello, world, and why this site exists (2026-06-30)
  A new home for notes on the ML I reimplement from scratch and the systems I build on top.
• Eight months, five rewrites: building a diffusion model by hand (2026-06-22)
  A Qwen-Image-style MMDiT rectified-flow image generator, built from scratch over eight months, and the "minor details" that turned out to be everything.