I work on making large language models cheaper to train and serve: aggressive quantization, quantization-aware training, and getting models to run on custom silicon without losing quality.
My interests sit where numerics meets systems. That means outlier suppression and rotations, 4-bit and ternary formats, integer datapaths, and the distributed training and MoE serving stacks that all of this has to fit into.
I’m a co-author of Selective Spectral Decay (S²D). I’m based in the San Francisco Bay Area.
This blog is where I write about what I’ve learned from pushing models to low precision, using public models and public research. Views are my own and do not represent my employer.