catid

@MrCatid

Training the models. Prior: Saronic, Anduril, Oculus VR, Game Closure, MSEE@GATech

Austin, TX

Born 1984

Pinned
catid
@MrCatid
Oct 9
Real-time image processing: Introducing the SoftMenon algorithm for image demosaicking. It runs as fast as bilinear on CUDA, less than 2x slower than bilinear on CPU. It has higher quality reconstruction than the Menon 2007 paper result previous best.
GitHub - catid/SoftMenon: Real-time debayering/demosaicking at higher quality in 2026
From github.com
catid
@MrCatid
16h
That would be kind of funny actually - hacking all the machines on the Internet to force-patch them to latest PHP. I'm not gonna do it but someone else's abliterated agent swarm might.
catid
@MrCatid
16h
DS4.1Flash is a sol-6 (released last month) class model you can run in swarms of 2000 with one 8xB300 server. How fast is it compared to the 64xB200 baseline DeepSeek shared, one just one server? Here's what it looks like. Very competitive, and faster for prefill.
3
catid
@MrCatid
Oct 20
Getting insane token/second prefill on 8xB300 on DS4.1 Flash due to YOCO architecture and DeepInfer. 90k token/s per GPU!
9
catid
@MrCatid
Oct 20
After reading "Already Clean" mechint blog post that shares a lot of evidence that models are very interpretable and that behaviors crystallize around early signals in localized areas, this result makes a lot of sense: alignment.anthropic.com/2026/graft/
1