Mistral was hacked

Likely had their model weights, post training pipelines, and data exfiltrated and sold to the highest bidder. 🫣

Yikes.

Posted by am.will, 2 hours ago

Bend 2 is here!

It is a new programming language that blocks AI mistakes via *proof checking* - the same technique big AI labs used to solve open math problems, like Navier-Stokes.

It is also very fast, and runs on GPUs.

https://bend-lang.com

Posted by Taelin, 12 hours ago

Android 17 is the first since 3.x to add new APIs without releasing to the AOSP
Article, Comments

I'm tired boss

Posted by AI Notkilleveryoneism Memes ⏸️, 21 minutes ago

Re: AI-generated posters don’t have to be horrible

A lot of people in the comments say that the supposedly better examples in the article are still “horrible” because they can tell they’re AI and point to things that the AI did “wrong” and a human artist would do better.

That may be true if you’re working with a top of the line artist, but in my experience the average freelance designer online is significantly worse than AI. I’ve worked with dozens of budget friendly designers on Fiverr and consistently get way worse results than I’ve gotten from AI models. Will a great human designer produce better output than an AI? Maybe, but a small local event or business probably won’t have the budget to hire them or know where to find them. It’s cheaper, faster and easier to use AI to generate something that’s good enough for their purposes.

People also point out all the little flaws in the “good” examples and cite them as obvious tells that they were AI generated. I think it would be interesting to see if people could reliably distinguish between AI generated flyers and human made flyers. Sure, sometimes it’s obvious like when the AI adds those weird, cartoony people. But if the model is told not to create designs like that and stick to the styles used in the second half of the article, I expect people would have a much harder time distinguishing the human designs from the AI designs.

ajjenkins, 4 hours ago

https://github.com/anthropics/claude-code/tree/main/mods/agents-md

this is not even overengineering at this point, it's literally a crime

all this "mod" does is to read a single .md file and pass it to the model btw
meanwhile they have telemetry for it

If you ever find a way to dramatically speed up your work, don’t tell anyone.

Posted by Corporate Absurdity, 13 hours ago

636606729769440499166579950236036751749912014371509557713570027508971809534551913252252094954941974952859310861988904737359709200557919
is a factor of RSA-896
https://saweis.net/posts/rsa-896.html

Posted by Steve Weis, 9 hours ago


In case you don't know what this means

Let me explain it to you this way
The Nintendo DSi, a console released in 2008, has an RSA-1024 hardware trust root.
According to the author, factoring this RSA-896 took at most 2048 GPUs, consuming about 30 GPU-years of computation, completed in a total of 10 days.

So, by optimistic estimates, brute-forcing the Nintendo DSi's trust root would only take 900 GPU-years.
So, is anyone up for starting a crowdfunding campaign?

https://x.com/__soragoto__/status/2101514663905333533

Resident Evil 4 (GameCube) – complete byte-identical decompilation to C/C++
Article, Comments

in era of LLMs, there is no longer such thing as "closed source"

am I the only one who puts names on vscode terminals

astra's system prompt in codex

Alcor: Simulate cpuid and sgdt/sidt results per-process

A Linux SVM hypervisor for emulating cpuid and sgdt/sidt results per-process by simply loading a Linux kernel module.
After loading, the module exposes a misc device at /dev/alcor_ctrl, which allows every process to change the cpuid and descriptor view of itself or, with CAP_SYS_PTRACE, any other process. See Usage for the exposed API.

er-azh/alcor

every frame of this is painted with JavaScript

made with Opus 5 (probably routed to 5.5),
sounds on :)

https://quiet-platform-kb9v.vercel.app/

Posted by Leon Lin, 1 hour ago

Rust in Git: Status, Internals, and Features

so far this is their progress
(apparently planned for Git 3.0+)

A benchmark where frontier language models drive a real comma-equipped Toyota through a cone course

https://drivingbench.com/

Introducing Gemini 3.8 Flash and Flash-Lite TTS, our new SOTA text to speech model with:

- a new voice design experience
- 2,000+ production ready voices
- voice replication
- support for 100 languages
- voice remixing (soon)
- #1 spot on Hume AI's voice benchmarks

and more!!

Posted by Logan Kilpatrick, 1 minute ago

GPT-6 Astra vs. Opus 5.5

the latter definitely needed way less handholding

Posted by Dominik Scholz, 20 hours ago