🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Over the 10 days, I wrote a new Racket library for "expect testing", a style that Yaron (Ron) Minsky has advocated for in OCaml. github.com/samth/recspecs/ As an experiment, I built it entirely with Codex (the OpenAI async AI programming tool). I have some thoughts.

LLM game changer for fixing build errors in OSS PRs on our internal build system

A lot of folks on Twitter swearing by voice transcription, but personally, I find it easier to think clearly typing things out; like I would literally prefer to type and speak simultaneously than to only speak

Hi Twitter, what are the current SOTA OSS models that have reproducible training recipes? Pre/post train OK. EleutherAI/pythia is an old example in this space. infly/OpenCoder a newer post-train one.

Edward Z. Yang Tulu 3 comes to mind allenai.org/tulu

OK, so, uh, has anyone gotten claude code's new hooks feature to work. Because it is definitely not working for me

Oh wow, did you guys know that torch.compile can compile numpy code? And even run it on GPU? This is pretty neat for all kinds of "surrounding" code besides the model (like evals and fancy metrics) that I used to do with numba/numexpr (cuz CPU-XLA was pretty meh). Poll below

I rarely say this, but it would actually be pretty useful to have a PL-style formulation in Greek of DTensor propagation rules

Been spending some time with the GSPMD paper recently. It's funny seeing all the work on making convolution work; 2021 truly was a different era