modern problems require modern solutions

I remember people cacelling their OpenAI subscription because they sold their AI to US Defense Department. Me: I’m not cancelling. They’re all same. It seems Anthropic is personally supporting NSA to spy on us. Like they have people at NSA onsite, helping them.

dang, where'dya think that green line's going? electrek.co/2026/05/20/i...

It's funny learning that the reason there's going to be a crippling helium supply shortage even though the United States once had a huge strategic reserve of it is literally because Republicans did the "jeff thinks the beans have to take turns lmao" thing to it

request for proposal: a theme park that is all just factory logistics systems where you get to ride in the cardboard boxes around roller belts

DeepSeek v4 PRO running via SSD streaming on my 128GB MacBook m5 max. 1.6 trillion parameters.

use @defined.net to ssh into ur agent boxes 💖

Gemma 4 12B Runs comfortably in 14GiB VRAM, but performs on par with Gemma 4 26B (which performed on par with gpt-oss-120b, which was up around o3) Also multimodal and awesome blog.google/innovation-a...

Google releases Gemma 4 12B, a new model that can run locally on 8GB RAM. Gemma 4 12B Unified model supports image, audio and 256K context. Run and train the model via Unsloth Studio. GGUF: huggingface.co/unsloth/gemm... Guide: unsloth.ai/docs/models/...

my "honest disclosures now" tshirt has people asking a lot of questions already answered by my shirt

12B Gemma 4!!! It's neat on its own, image and audio in, but architecturally it's super cool: instead of a visual or audio encoders, they directly train the model on vision and sound, with images using a "lightweight embedding module" and audio 'projected into the same space as text tokens'

DeepSWE is nice because it's a fresh benchmark. Less hill climbing. GPT has the Pareto frontier (perhaps ant should lower their prices). Kimi K2.6 got 24%, Opus 4.6 got 28% both with 4% error bars. Opus came out 2.5 months before Kimi, so consistent with a 4-5 month lag for open models.

Someone actually did a study a while back that showed the most effective way to get women out of TERF groups wasn’t people trying to change their minds, it was them getting literally any benign hobby, like learning other languages or taking up needlework

The past years have seen horrendous attacks towards trans people. We have a duty to protect the most vulnerable members of our communities. There is no Haskell without the trans people who have shaped it through the years. This June is a month of celebrations, but the struggle continues. ✊🏳️‍⚧️λ

Happy pride! The fascism only makes me more gay motheruckers

"Today, I'm calling on HPD to provide transparency, communication, and accountability regarding my daughter's death. I am not going away. My family's not going away. I will be a nightmare to the city of Houston."

Good evening, I'm at the Montrose Center here in Houston for a vigil for Persia Amarra Conway, the transgender woman who was found dead in SW Houston. Her family is going to speak the media soon and give us an update.

@sunshowers.io is there an up to date doc you have for this? github.com/sunshowers-c... messing around with making a cli in rust, not a fan of the deep drilling —help by default in clap

iddqd, or the hardest kind of unsafe Rust. My first post for the @oxide.computer blog! Goes deep into the pitfalls of unsafe Rust and how to reason about it. A lot of labor went into the post -- please have a read! oxide.computer/blog/iddqd-u...

some old GIS-related threejs projects

That's because it's correct. Agents are stupid. Borrow checker is smart. Borrow checker is love. Borrow checker is life. I love you, borrow checker.

they/them really cooked with this one! 🏳️‍🌈🏳️‍⚧️💛🤍💜🖤 I had a ton of fun making this and it was lovely to see how excited everyone in the @sentry.io slack was about it also psa: pride logos can be more interesting than slapping the flag directly in your logo mark

yep

happy pride month 🎉

????

Maudit.org: a Rust library to build static websites, by @bruits.org. I like this approach: SSGs tend to become bloated kitchen sinks that you have to re-learn every time you want to change something. A library in a strongly typed language allows building just what you need, using your mental model.

maybe google aint that bad

im back on my bullshit (using concepts i learned recently to improve the horse game)

I'm very excited. This has been my white whale for years. I'm gonna spend some more time exploring and then work towards building a delta-tracking path tracer for some offline PBR. Rendering is currently super rudimentary: just glPoints, additive blending, and bigger points for closer particles.

It's not a meme: atproto.com/blog/serving...

Ok, seems like Claude cooked with this one actually. Can't believe I ever doubted him. The future is amazing

Thinking Machines: Interaction Models A new class of LLM, this is an agentic model focused entirely on seamless interactions. It knows how to keep you engaged while it goes off and does other work thinkingmachines.ai/blog/interac...

yay my 2-bit quantized qwen3.6-27b works huggingfacely within 13gb

why the hell aren’t more people using late interaction embedding models in agents? this is crazy @lateinteraction.bsky.social are you here? it’s crazy. since you can pick out just a few words from individual matches, the agent doesn’t need to intentionally search, so it’s like sidecar RAG

what's the sitch on android eink? I want an analogue to a kindle, phone sized would definitely be cool, I think I saw one out there that attached via magsafe? I use readwise.io/read, which has e-ink styling support, so support for that would be cool af

DwarfStar distributed inference is now on GitHub: you can run 2 bit Flash using 2 64GB machines, or 4 bit Flash with two 128GB machines or 4 64GB, ad so forth. Prefill speed will increase thanks to pipelining. Generation will pay the latency.

I can imagine finetunes of this thing going insane. Mogs OSS-20B on agentic stuff. I assume it's really fucking fast because of how few active params it has. 38T tokens (more than deepseek-v4!!!) for an 8B model wew

the 4.6 -> 4.8 evolution is really not great here, just a slide towards simplicity, and a mild preference against (?) user competence with 4.8

v1.122 is rolling out today, featuring a better way to discover and share articles, blog posts and newsletters across the Atmosphere via new, dynamic Standard.site link cards. 🧵

Wh-what the fuck

www.patreon.com/DynamicLife love this stuff translate fluoddity to minecraft worldgen? 🫪

Oops All Oops All Paperclips Paperclips Posts

A new way to explore your node modules folder: nesbitt.io/heap

probably one of the best threads I've ever seen

telegraph.cee.wtf morse code only chatroom hop on lets have extended conversations about ai consciousness

Leaving everything else aside, the Solar Revolution is just an unbelievably positive development, we aren’t that far away from “people in the Global South can run A/C all summer with extremely cheap zero carbon energy” which would have been essentially unthinkable even like 10 years ago

I got six!