quantbagel
quantbagel on Hacker News
we distilled GPT-6 Astra, into Nvidia's DreamZero world model. Astra is state-of-the-art on most robotics benchmarks, but its too slow. On the Dream Zero backbone its much faster! Of course, there's some degradation, but the speed improvement means it's actually usable on high-end robots to amateur arms.
Posted by
quantbagel on Thursday October 01, 2026 @10:46PM
from the abstract.inc dept.
Posted by
quantbagel on Wednesday July 29, 2026 @06:26PM
from the show-hn dept.
https://useglass.up.railway.app/#
A fully local tool I built so claude could search and see everything I do so I dont have to copy and paste screenshots all the time
Posted by
quantbagel on Saturday March 21, 2026 @08:21PM
from the show-hn dept.
https://tryreflex.ai
After distilling oss models, serving inference to robotics startups and getting to 200k in mrr im launching reflex. Upload policies and models and we do 1-click deploy https://x.com/quantbagel/status/2028924664190517364?s=20
Hi HN, I made a JIT compiler so that I can easily setup VLA experiments and iterate fast for my robot arms using a python interface but have cuda kernels and metal kernels in the background. I prefer the platform agnostic DX and the uv native simplicity whilst also gaining performance, Its a cool idea please give a few stars, im sure it will help the ecosystem
Well I wanted to implement light transport papers without having to deal with cpp. I think tinygrad, and more specifically tinyJIT are super useful abstractions. This is def not available in ts
Posted by
quantbagel on Wednesday January 21, 2026 @10:26PM
from the show-hn dept.
https://github.com/quantbagel/gtinygrad
I saw a tweet that tinygrad is so good that you could make a graphics library that wraps tg. So I’ve been hacking on a gtinygrad, and honestly it convinced me it could be used for legit research.The JIT + tensor model ends up being a really nice way to express light transport all in simple python, so I reimplemented some new research papers from SIGGRAPH like REstir PG and SZ and it just works....
When I made a swift package manager as Rust rewrite I realized that the language wasn't the issue, design is a lot more important. Rust just gave a boost to everything else. you can try Gust here
https://github.com/quantbagel/gust a lot better than using SwiftPM but there's room for improvement! Make issues with your ideas
near 300x speed improvement on warm build and resolved, parallelized installs, etc
Posted by
quantbagel on Thursday December 25, 2025 @08:32PM
from the github.com dept.
Also made an abstraction for tryhackme, hackthebox and pwncollege. So you can solve machines straight from the terminal with hints.
Posted by
quantbagel on Tuesday December 16, 2025 @12:58PM
from the trysonder.ai dept.
works iOS->linux, windows, Mac. Android->linux, windows, Mac, and vice versa.
open source, feel free to add issues, star or check out some if my other projects!
Posted by
quantbagel on Wednesday December 10, 2025 @05:09PM
from the github.com dept.
Sakana.ai improved on this by honing in on sample efficiency iirc with shinkaevolve (which is open source and not an ai slop project)