Wrote a bit this morning on the internals of ML compilers, and how I'm going about porting JAX to pure JS

i am at home for a couple days to rest, did some jax-js stuff. added float16 and going to run Apple's MobileCLIP2 next, just to see how it goes! here it is loading the cocktail of safetensors from CDN, cached in OPFS :D. also claude-coded a nice CLI for converting weights ✨

I’ve been reading things I wrote when I was 18-21 and it’s honestly good. had not figured out a lot of stuff yet, very turbulent. but there’s a raw simplicity and rightness about my convictions at the time, perhaps I could only write like so when first developing a core worldview.

saw a very strange movie today, and it made me remember that life can be a bit surprising at times, but at least I have the people I’ve grown to appreciate very much

RT @ekzhang1: "no, you're an infrastructure company, you can't have cute UI and information visualization! everything has to be tables and….

over the past 4 days, I read the really great LLM scaling book from google DeepMind/JAX and took some (very abridged) notes. it's a long book, but I think this provides an okay tl;dr of the plot at some level of technical background.

omg they're sharing notebooks now 😄 sign of what's to come soon.

[digression] i've always been a little bit confused by the idea of 'career progression' in tech because i've always just written software and worked out problems w people. seems if your work is your goal, the tools are all there for everyone? the high school student has the same.

if web development is mostly centering divs, and theoretical CS is just Chernoff bounds, then it seems like ML systems is mostly just AllReduce + lots of coffee.

RT @bernhardsson: What does @modal do? How does it work? What's different about AI infra? Why did we throw out Kubernetes and Docker built….

it’s so interesting, how life changes and the world blurs, you do something once and then remember it again years later, maybe lost a piece along the way. i really don’t remember. but in quiet moments there’s more to it, deeper truths to believe in than your day-to-day job

RT @modal: It’s been exciting to help support @cartesia_ai on Line, their new code-first voice agent platform — solving how they instantly….

TIL that kernel panics are just a fact of life for people who write GPU compilers as their job. y’all are so brilliant. I guess with jax-js, since the GPU compiler is in the browser, difference is that I can gift the kernel panic to all my users on the web, with just a link :D.

so this is awkward… I can get a consistent macOS kernel panic after ~8 epochs of my convnet training code in jax-js's WebGPU backend — well that's one hell of a bug to figure out. never thought a web page would trigger a kernel panic in 2025, let alone a library I made myself 😅

I got general convolution (n-dim, pads, dilated, strides) working in jax-js! Kind of a pain but it is all working now, reasonably fast & with backprop. Couple more features, some minor bugfixes, optimizations left, + a Wasm backend, and then I think I'll release the library :).

Feeling a bit of pride but also mostly irked, today I found out that my is now a tool of choice for crypto miners trying to abuse free cloud credits. Maybe it’s the e2e encryption, UX or ease of installing, idk. But yeah, I guess that’s just OSS in use 😕.

one unexpected skill I've learned recently is how to mentally rotate / do dimensional analysis on tensor indices in my head. this is from building jax-js, but also randomly from reading diff geo book and trying to wrap my head around funky tensors like riemann curvature. it's a.

RT @BennyKokMusic: @ekzhang1 man, didn't know you created sshx, one time one hacker used it to get into our modal instance to abuse our res….

reflecting more, that I appreciate my colleagues for hiring me back then in 2021. back then I had no past job, just ~20k github stars from 10 years of OSS and ideas for Rust code + good web dev. they offered me way more than anyone else given my 0 YoE, and it must've been a bet.

I think when I signed on a 2-person tech startup in 2021, I didn't really think abt that if things went well, I would be working at a a 60-person startup. or maybe I did but it felt so fast. now my coworkers are people who sign on to join 60-person startups.

for NYSRG this august, I didn't really know what we should read about, so we'll read some of the SSA Book