Erica Windisch
@ewindisch.bsky.social
240 followers 240 following 360 posts
building hyprstream. machine intelligence with zero consensus - peer to peer AI training and inference. https://github.com/hyprstream/hyprstream
Posts Media Videos Starter Packs
Pinned
ewindisch.bsky.social
We're building an AI that's free, public, and openly trained by the community, with open weights shared via torrent git chunks.

Streaming weekdays 11-3pm on Twitch.

Come hang out Monday, or give our repos some stars! ⭐️
ewindisch.bsky.social
Literally said this to my school board president and head of curriculum, in part as a response to AI, and recognition of linear's growing impact across fields
ewindisch.bsky.social
Ten years ago I made a commit with a message to temporarily disable enforcement to collect feedback. Loool
ewindisch.bsky.social
Reviewing the status of my old work to add an apparmor policy for the Docker engine (moby). The good news is that it still exists and is maintained. The bad news is that its non-enforcing and I don't think its even loading by default 🤔
ewindisch.bsky.social
In the ye ol days of the internet web hosting before Know Your Customer and a web full of frontpage and PHP installations where novice developers cargo culted "chmod 777" permissions, I was a sysadmin of 1000s of machines connected to the internet with no firewall. Uh... yeah.
ewindisch.bsky.social
The trick is having the AI pay for its own compute
ewindisch.bsky.social
It feels mainstream, sadly. Fear is a fantastic tool of propaganda; people have reasonable fears of AI upon a backdrop of capitalism.
ewindisch.bsky.social
Umm. Kids with disabilities are supposed to have the least restrictive environment. That means they should be in the classroom, not separated. That was true under democrats too.
ewindisch.bsky.social
Kids are learning scratch first these days. Mine has been reluctant to learn python... we might skip to Rust
ewindisch.bsky.social
If these learnings end up mapping to human intelligence...
ewindisch.bsky.social
I'm bullish on AMD. I've been accidentally-exclusively buying their hardware. The Strix Halo and mi210 are too well priced to pass up $/GB, and the memory bandwidth is fantastic. The BC-250 is also a great way to build cheap dev labs for tiny models.
ewindisch.bsky.social
Open source hardware and robotics is behind software but has a thriving community. Thankfully China has invested heavily there, otherwise America would be even further behind.
ewindisch.bsky.social
I built a battery powered wearable computer with a SBC in the year 2000.

It looked like this.
ewindisch.bsky.social
Loved my Octane with the R10000. Many regrets on giving the machine away
ewindisch.bsky.social
Git is the best operating system written by Linus Torvalds, and I love Linux.
ewindisch.bsky.social
I thought this was a new quantization.
ewindisch.bsky.social
👍 knowing which quants you ran would also be helpful.

I may get one just to help support it better. My llm engine is targeting ROCm as a primary target, but ROCm hasn't been well supported on the consumer APUs.

The earlier BC-250 APU also does well with Vulkan, crashing with ROCm.
ewindisch.bsky.social
John and I were the same age. I went into datacenters and cloud, hanging out with yall on IRC from \#loki to \#icculus. His death was tough.

The before times were such a blast. Back when Linux gaming was a small community of believers and wishful hackers.

He never saw a Steamdeck 😞
ewindisch.bsky.social
That's so disappointing. Those are all MoE models so your test is 1.5B-8B active parameters in each. The memory bandwidth really is constrained.

How does it do with a large dense model like a DeepSeek-R1-Distill-Llama-70B?
ewindisch.bsky.social
Small models can be as powerful as the biggest models. Local models and human ingenuity can figure out anything tbh.

Regulation is more likely to empower monopolies and destroy independent media than any good it may bring.

Regulate away the "Wikipedia of AI" and we'll only have OpenAI/Grok/Claude.
ewindisch.bsky.social
sharky6000.bsky.social
Hello all! 👋 🚨 New Preprint Alert! 🚨

Code World Models for General Game-Playing. ♟️🎲 ♣️♥️♠️♦️

I am pleased to announce our new paper, which provides an extremely sample-efficient way to create an agent that can perform well in multi-agent, partially-observed, symbolic environments!

🧵 1/N
ewindisch.bsky.social
have you seem google's new preprint?