Quinten Kock

Bloated patches: LM generated solutions of SWE-bench tasks are consistently longer than human-written gold solutions (and it's not just comments) 🧵

• web browser (noun) a fork of chromium • text editor (noun) a fork of vscode

When you're trying to do some AVX programming using ISPC but you're on a Zen4 machine.. Even doing this serialized is faster lmfao

Another common mistake I see with LLMs is sticking to prose inputs and outputs for the model Most users are actually pretty allergic to both typing out and reading large blocks of text. It's often better to structure inputs and outputs to be something other than sentences/paragraphs

I just started doing #adventofcode in #Haskell, but I've already done that for like the last three years, so maybe it's time for something else? Does anyone here have any fun suggestions?

QUICK REMINDER NEWSKIES: Pronouns are SUPER easy to add Just pop over here, sub, and then go to posts to select yours so they pop up as a profile label and you have more room in your bio!! bsky.app/profile/pron...

Hi, so I've spent the past almost-decade studying research uses of public social media data, like e.g. ML researchers using content from Twitter, Reddit, and Mastodon. Anyway, buckle up this is about to be a VERY long thread with lots of thoughts and links to papers. 🧵