Writing parsers in C++ just sucks, I'm spending days on what I could've done in hours in Python. This was a dumb idea, turns out MLIR even has good Python bindings.

New post: How function inlining can change floating point results through triggering FMA Fusion. siboehm.com/articles/23/In…

Matei Zaharia James Zou June GPT-4 started surrounding code with ```python markdown, which you didn't strip. I forked your code, removed that markdown, and re-submitted the output to Leetcode for judging. Now the June version does significantly better than the march version of GPT-4.

Really surprising just how much untrusted code execution software devs do all day (installing pip packages, pipe-to-sh, go get <github repo>, ...) and how it almost never goes wrong.

We're happy to announce the world's first compute auction site is now in beta. Bid on time on our 23,000 Nvidia H100s. Per-hour access for any AI project! flopauction.com

Looks like Triton conference talks are out: youtube.com/@Triton-openai…

I got Triton working on (local) Godbolt instances, including Python <-> SASS line correspondence. Pretty nifty. Still got some infra issues to figure out, meanwhile PR is here if anyone wants to try it: github.com/compiler-explo…

My biggest contribution to scientific progress has been telling every researcher I've ever worked with about mamba (to replace conda), and gifting them a copy of Barrett's Linux Pocket Guide.

I have to decide between training for a Factorio speedrun, or for an Ironman. Can only do one, pls help.

Here's Claude 3 Haiku running at >200 tokens/s (>2x as fast as prod)! We've been working on capacity optimizations but we can have fun testing those as speed optimizations via overly-costly low batch size. Come work with me at Anthropic on things like this, more info in thread 🧵

Every hour I've spent learning about floating point numerics and PCIe internals has payed for itself many times.

I'll write a kernel for AMD today! Wohoo.

Reading the EUV history I feel like if they'd known it would take 20y to ship EUV instead of the projected 5y it would've never gotten funding. Instrumentally useful self-deception.