Projects

A llama.cpp fork that runs a 27B model at 33 tokens a second on an M4 Max. Along the way I found a precision bug that hit every Q4_K model on Metal.

I taught a vision model with RL to zoom in only when it’s worth the time. Same accuracy, 2.18x faster on an M4 Max.

Sharper 3D scenes with diffusion models. My first-author paper at WACV 2025.

PaperCode

Two agents learn soccer from nothing by playing each other. They get good after an hour of training on a Mac.

Stable Diffusion in raw CUDA and C++, written to learn how GPUs actually work.