I write about how ML systems get built and why they work that way, with a particular focus on the infrastructure (compilers, kernels), the modelling on top of it, and at times the money underneath. Expect short notes on papers, and the occasional essay or experiment.
I work on GPU kernel and compiler infrastructure at Meta (PyTorch, Python, Triton); before that, developer infrastructure at Lyft and Firebase at Google among other things. Opinions are mine. Do not draw any research conclusions based on this website.
Start Here: Megacores, Everything MoE, Maybe the agents shouldn’t write the kernels, TileIR, The model can probably write the code, Unbundling Work, Benchmarks Mean Business
Elsewhere: Github. Linked In, the Hellsite. The Dodgyscripts Collective, A kinda zine, archived writing.
RSS: Everything Feed
Public key (age): age10gd86290dydz0xcwyzq4p0mu2686ygapuutryxhj0tqg2rtndg7s98n0yc
Some blogs I like/steal from: Simon Willison, Ed Yang, Dave Anderson, Nathan Lambert, Anne Helen Peterson, Matt Levine, Kevin Xu
Posts by WordPress Email (low-rent substack):