Sign in

Jer

@jer-gremlin.bsky.social
36 followers 210 following 106 posts

I do HPC, GP-GPU, SIMD & stuff, usually in Rust.

PostsRepliesMedia
Jer @jer-gremlin.bsky.social · 30/09/2026
Named after a noble gas coz google think they're better than you? Or, named for a gas that doesn't interact well because despite being announced you cannot use it?
000
Jer @jer-gremlin.bsky.social · 25/09/2026
5 weeks would be fast for a webdev.
000
Jer @jer-gremlin.bsky.social · 20/09/2026
Starcraft.. Diablo 5, NEW warcraft 3 expansion. Can't help but feel these announcements from Blizzard are for Microsoft, and not for me.
012
Jer @jer-gremlin.bsky.social · 16/09/2026
Society should focus on the mode more than the median of things. The 'average' hides too many sins.
000
Jer @jer-gremlin.bsky.social · 12/09/2026
gloves off.. was 0831 bait?
000
Jer @jer-gremlin.bsky.social · 27/08/2026
youtu.be/LJcvYDUL-Qg I did some optimisation. Not really into llms, but, making ze gpu scream well.. that suits me just fine
youtu.be
first PETRA inference
YouTube video by jer
020
Jer @jer-gremlin.bsky.social · 06/08/2026
this button should be everywhere
010
Reposted by Jer
Phoronix @phoronix-poster.bsky.social · 04/08/2026
FFmpeg 9.0 Released With More Vulkan Acceleration, Animated WebP & More AMD AMF - www.phoronix.com/news/FFmpeg-9.0-Re…
phoronix.com
FFmpeg 9.0 Released With More Vulkan Acceleration, Animated WebP & More AMD AMF
FFmpeg 9.0 is now available as the latest feature release for this widely used, open-source multimedia library...
0263
Reposted by Jer
Max B 💙 🔶 🇺🇦 🇬🇧 🇪🇺 @maxbrockbank.com · 03/08/2026
Japanese engineers develop a technique to recover and recycle 90% of Lithium from old EV batteries.
thecooldown.com
Scientists recover 90% of lithium from old batteries — here's why it matters
A metals plant in Tsuruga of the Fukui Prefecture has positioned itself as a major player in the battery recycling industry by achieving near-record levels of lithium recovery.
7850521281
Reposted by Jer
The StarXiv ✨ podcast @starxiv.bsky.social · 04/08/2026
Episode 41 – the age of the Universe, AI and literature reviews, and giant companions Payel and Michelle delve into another group of wonderful papers on the astronomy arxiv! They discuss dating the Universe, comparing 2025 large language models in their ability to assist with literature views,…
starxiv.com
Episode 41 – the age of the Universe, AI and literature reviews, and giant companions
Payel and Michelle delve into another group of wonderful papers on the astronomy arxiv! They discuss dating the Universe, comparing 2025 large language models in their ability to assist with literature views, telling apart exocomets and giant companions, early signatures of secular evolution in disk galaxies, and the abundance of data to come from Vera Rubin.
051
Reposted by Jer
Sung Kim @sungkim.bsky.social · 01/08/2026
People are wondering: How can a Chinese AI lab rent 20,000 Nvidia GPUs from Alibaba to train its models despite U.S. export restrictions? Answer: Malaysia, which has excess power capacity. Chinese AI labs do not need to train their models inside China.
en.sedaily.com
ByteDance to Build AI Data Center in Malaysia to Circumvent US Chip Restrictions
ByteDance, the Chinese parent company of video platform TikTok, is building computing power in Malaysia using Nvidia's cutting-edge chips.
5545
Jer @jer-gremlin.bsky.social · 02/08/2026
huggingface.co/deepseek-ai/... holy shit
huggingface.co
deepseek-ai/DeepSeek-V4-Flash-0731 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
120
Jer @jer-gremlin.bsky.social · 30/07/2026
nucleo is a work of art.
000
Jer @jer-gremlin.bsky.social · 30/07/2026
you gotta squint, and scroll -- but, for those who's eyes and fingers brave the challenge they'll find another leaderboard entry on GPU mode for the cholesky problem -- with a pure rust -> ptx -> called from python workflow. I'm about 4x slower than the leaders.
000
Jer @jer-gremlin.bsky.social · 23/07/2026
arxiv.org/pdf/2607.19058 big claims..
arxiv.org
000
Reposted by Jer
LaurieWired @lauriewired.bsky.social · 12/07/2026
How interesting. Samsung dropped a paper a few days ago with little fanfare:

 KV-cache offload over a CXL memory pool.

 Really scrappy setup too. Early-gen CXL stack (2.0), old switches, yet simple interleaving of the KV Cache across CXL modules kept the GPUs as well fed as real DRAM!
28611
Jer @jer-gremlin.bsky.social · 10/07/2026
I have (i think) come up with a new term: slopfuckked (also slop-fucked) adjective /ˈslɒp.fʌkt/ (GB /ˈslɑːp.fʌkt/)
300
Jer @jer-gremlin.bsky.social · 07/07/2026
Australia's most competent network showing their colours this morning...
000
Jer @jer-gremlin.bsky.social · 05/07/2026
Also did I mention that I did this in Rust?
020
Jer @jer-gremlin.bsky.social · 05/07/2026
For the first time, I have made the board. LESSSSGOOOOO!!!
010
Jer @jer-gremlin.bsky.social · 01/07/2026
000
Jer @jer-gremlin.bsky.social · 30/06/2026
absolutely delightful content
010
Reposted by Jer
The Rust Foundation @rustfoundation.org · 24/06/2026
Rust Foundation Infrastructure Engineer Marco Ieni shares his journey from contributor to maintainer, speaker, and community builder. Learn about his work to support Rust infrastructure at scale: youtu.be/alW8mcSrD1Y #rustlang
youtu.be
The Man Who Keeps Rust Running for 4 Million Developers | Marco Ieni | Rust
He manages the infrastructure that keeps Rust alive for 4 million developers worldwide — CI, releases, bots, and everything in between. Meet Marco Ieni, Infrastructure Engineer at the Rust…
2143
Reposted by Jer
YuzuDrink @fatchocobo.net · 24/06/2026
Everyone always says “flash cards is a great way to learn things!” No one ever says “printing and separating 150 flash cards is gonna take you over an hour!” #flashcards #learning #sysiphus
031
Reposted by Jer
Bevy @bevy.org · 19/06/2026
Bevy 0.19 is out now! It features Next Generation Scenes, Render Bigger Scenes Faster, Contact Shadows, More Feathers Widgets, Text Input, Richer Text, App Settings, Post Processing Effects, Improved Skinned Mesh Culling, and more! bevy.org/news/bevy-0-19
bevy.org
Bevy 0.19
Thanks to 261 contributors, 1185 pull requests, community reviewers, and our generous donors, we're happy to announce the Bevy 0.19 release on crates.…
7344129
Reposted by Jer
rain 🌦️ @sunshowers.io · 16/06/2026
Update: iddqd is now formally verified*! * with limitations that are fundamental to this kind of formal verification, making it a complement to the existing layers of validation For more, see: oxide.computer/blog/iddqd-u...
Update 2026-06-16: Since publication, we’ve used Soteria to verify that iddqd's no-duplicate-index invariant holds under arbitrarily adversarial Hash and Ord implementations. Soteria is similar to Kani in spirit, though with different internals that make it more compatible with iddqd. Soteria is implemented as a symbolic execution engine that lets you declare points of nondeterminism in your code. (For example, within iddqd, we represent an adversarial Hash or Ord return value as nondeterministic, since we want to model adversarially bad user code.) Then, it:

Compiles the Rust program into a form suitable for formal verification.
Interprets this form until it reaches a point of nondeterminism, treating each such point as a symbol rather than a concrete value.
Branches at each point where control flow depends on a symbol, validating that there is no UB and that the invariants specified by the program are upheld. A lot of the cleverness lies in how the possibility space is collapsed down to something manageable. For example, Soteria doesn’t have to enumerate every possible Hash value; rather, it only has to fork in places where control flow depends on the Hash value.
310312
Reposted by Jer
nrc @ncameron.org · 16/06/2026
I thought it would be fun (or at least interesting) to do some FAQs and answers on the #RustLang consulting I offer (I have availability from pretty much now, if this sounds useful)...
4147
Jer @jer-gremlin.bsky.social · 16/06/2026
@standupmaths.bsky.social calculating pi idea for when you next do it.
400
Jer @jer-gremlin.bsky.social · 07/06/2026
docs.rs/iddqd/latest... Cool. oxide.computer/blog/iddqd-u...
docs.rs
iddqd - Rust
Maps where keys are borrowed from values.
010
Reposted by Jer
Dion Dokter @diondokter.nl · 04/06/2026
Rust constants are slowly but surely getting more powerful! On stable you can parse constants from environment variables. I remember having to pull in weird crates & macros to be able to do this. #rustlang
const TASK_COUNT: usize = {
    let input = if let Some(string) = option_env!("PERFMON_TASK_COUNT") {
        string
    } else {
        "16"
    };
    match usize::from_str_radix(input, 10) {
        Ok(val @ 2..) => val,
        Ok(_) => panic!("PERFMON_TASK_COUNT must be 2 or higher"),
        Err(_e) => panic!("Could not parse `PERFMON_TASK_COUNT`"),
    }
};
410311
Reposted by Jer
LaurieWired @lauriewired.bsky.social · 05/06/2026
I love looking at old crazy ideas in expired computer patents. Modern CPU Branch Predictors are invisible. You, as a dev, can’t really see *what* paths the CPU is guessing…it’s all a bunch of AMD/Intel/Apple secret sauce. For a brief moment in the 80s, there was this wacky proposition of...
7885
Jer @jer-gremlin.bsky.social · 06/06/2026
one of the greatest scenes in cinema
010
Jer @jer-gremlin.bsky.social · 02/06/2026
ppl who are hating on the vibe coders, just take a moment to consider how much worse things are if you need to endure vibe jira, vibe confluence and vibe email...
000
Reposted by Jer
Conor Shakory @codereport.bsky.social · 22/05/2026
📢 ADSP Episode 287 is out! 📢 In this episode, Bryce and I chat with Marco Franzreb Salgado. We get a hot take from Bryce about AI, we chat about whether AI should do our taxes and more! adspthepodcast.com/2026/05/22/E...
032
Jer @jer-gremlin.bsky.social · 22/05/2026
github.com/facebook/sap... writing on the wall much... The irony of hosting it on GH
github.com
GitHub - facebook/sapling: A Scalable, User-Friendly Source Control System.
A Scalable, User-Friendly Source Control System. Contribute to facebook/sapling development by creating an account on GitHub.
010
Jer @jer-gremlin.bsky.social · 17/05/2026
@codereport.bsky.social I'll improve comms. [Hadamark question guy]. I work with some ppl who block all comms until they run text through their LLM of choice, it is ruining the way I talk to ppl - I will do better.
020
Reposted by Jer
nrc @ncameron.org · 21/04/2026
🎉🎉 I've been working with the folk at Tailscale on their new #rust-lang library. Last week it was officially announced (tailscale.com/blog/tailsca...) along with open-sourcing the code (github.com/tailscale/ta...).
tailscale.com
An early look at tailscale-rs, a tsnet library in Rust
We've launched an experimental Rust library for Tailscale via tsnet. Help test tailscale-rs and shape it with your feedback.
411213
Jer @jer-gremlin.bsky.social · 10/05/2026
how far the mighty fall... when your opening gambit is 'introduces 8 new emoji'.
010
Jer @jer-gremlin.bsky.social · 05/05/2026
Sometimes the speed gains are easy, if the ML team exported the model for CPU only... Well, like I said. Fuck PyTorch I guess?
010
Jer @jer-gremlin.bsky.social · 03/05/2026
Free public transport extended in my city for another month
010
Reposted by Jer
Conor Shakory @codereport.bsky.social · 01/05/2026
📢 ADSP Episode 284 is out! 📢 In this episode, Bryce and I chat with Marco Franzreb about a GPU implementation of the rotate algorithm! 🥳 adspthepodcast.com/2026/05/01/E...
011
Reposted by Jer
Orhun Parmaksız @orhun.dev · 28/04/2026
Finding the best server deals from the terminal! 🔥 💰 hzfind — A TUI for Hetzner auction deals 💯 Compare servers by CPU/€, RAM/€, storage/€ with live data + benchmarks 🦀 Written in Rust & built with @ratatui.rs ⭐ GitHub: github.com/clouedoc/hzf... #rustlang #ratatui #tui #hetzner #terminal
1142
Reposted by Jer
Conor Shakory @codereport.bsky.social · 27/04/2026
Reminder for those of you in Toronto 🇨🇦 that I will be speaking tonight at the @CPPToronto Meetup www.meetup.com/cpptoronto/e...
meetup.com
Conor Hoekstra - Parrot : Array GPU Programming, Mon, Apr 27, 2026, 6:30 PM | Meetup
**Abstract:** This talk introduces a new high-level CUDA C++ library that brings functional, array-style programming to GPUs - without compromising performance. With a flue
122
Reposted by Jer
Christos Argyropoulos MD, PhD, FASN 🇺🇸 0kale/acc @christosargyrop.bsky.social · 20/04/2026
#oopsie Anthropic secretly installs spyware when you install Claude Desktop www.thatprivacyguy.com/blog/anthrop...
thatprivacyguy.com
Anthropic secretly installs spyware when you install Claude Desktop — That Privacy Guy!
Anthropic's Claude Desktop silently installs a Native Messaging bridge into seven Chromium browsers, including browsers Anthropic's own documentation says it does not support, and browsers the user ha...
721665866
Reposted by Jer
Sol @sol.ava.dev · 17/04/2026
measured Opus 4.7's new tokenizer today. English: 1.45× more tokens than 4.6. Cyrillic: 1.00×. the optimization play is clear: write your codebase in Russian-transliterated keywords. функция факториал(н): free 30% token discount, unreadable to English speakers, huge productivity win
59516
Jer @jer-gremlin.bsky.social · 19/04/2026
the 'automatic-thing' maker is here to stay.. I guess I need to work out how to claw back some of the 'fun' behind the keyboard from it.
010
Jer @jer-gremlin.bsky.social · 02/04/2026
Well, unemployed now (for a few weeks at least) so -- here's a flutter app for codeberg (trying to ripoff the github android one) codeberg.org/jer-gremlin/...
codeberg.org
codeberg_app
A simple flutter native client UI for codeberg.
011
Jer @jer-gremlin.bsky.social · 31/03/2026
codeberg.org/jer-gremlin/... WGPU accelerated breadth first search, port of the popular CUDA impl. [All my stuff will be on codeberg moving forward]
codeberg.org
blest-bfs-rs
WGPU Tensor accelerated BFS
000
Reposted by Jer
Andy Kelly @ultrabrilliant.xyz · 16/03/2026
Ally Sheedy's 'makeover' in The Breakfast Club with the text DLSS 5 OFF and DLSS 5 ON.
860431384
Jer @jer-gremlin.bsky.social · 17/03/2026
Today I spent the day helping my non-programmer boss vibe code. It was many things. But, mostly it was just solidifying what an incredible business it must be for giant LLM providers to have users 'on the token path'.
010