OsamaJaber

5 ms·

31.
AutoMegaKernel: Compile an LLM into one provably-correct CUDA megakernel(github.com)
4 points by OsamaJaber 4mo ago|0 comments
32.
StreamIndex: Memory-bounded compressed sparse attention via streaming top-k(arxiv.org)
4 points by OsamaJaber 5mo ago|0 comments
33.
Show HN: AutoKernel, Auto GPU Kernel Optimization(arxiv.org)
2 points by OsamaJaber 5mo ago|0 comments
34.
DeepSeek V4's indexer dies at 65K. We got it to 1M on 6GB(arxiv.org)
5 points by OsamaJaber 5mo ago|0 comments
35.
AutoKernel: Autonomous GPU Kernel Optimization via Iterative Agent-Driven Search(arxiv.org)
4 points by OsamaJaber 5mo ago|0 comments
36.
by OsamaJaber 5mo ago
the bottleneck isn't generation, it's verification
37.
DeepSeek V4's indexer OOMs at 65K context. We got it to 1M in 6G(arxiv.org)
8 points by OsamaJaber 6mo ago|0 comments
38.
Ouroboros: Dynamic Weight Generation for Recursive Transformers(arxiv.org)
2 points by OsamaJaber 6mo ago|0 comments
39.
by OsamaJaber 6mo ago
Good to see this exist. Inference providers quietly swap quant levels. Most users never check. A standard verifier from the model maker is the right move, would love to see other labs ship the same
40.
by OsamaJaber 6mo ago
The modified MIT clause is sneakier than people think. Hit 100M users or $20M a month and you have to slap "Kimi K2.6" on your UI. That covers any consumer app worth building. Not really open, more like free until you matter. Llam
41.
by OsamaJaber 6mo ago
That's why infra needs stricter internal walls than normal SaaS
42.
by OsamaJaber 6mo ago
Small models in the browser are a different optimization problem than small models on a server. On server you chase throughput so you batch. In browser you're stuck at batch size 1, which means kernel launch overhead and memory bandwid
43.
Tide: Token-Informed Depth Execution for Per-Token Early Exit in LLM Inference(arxiv.org)
3 points by OsamaJaber 6mo ago|1 comments
44.
Own your AI. Optimized down to the kernel(runinfra.ai)
1 points by OsamaJaber 7mo ago|0 comments
45.
by OsamaJaber 8mo ago
Appreciate it :)
46.
Agents with "Hands"(openfang.sh)
7 points by OsamaJaber 8mo ago|18 comments
47.
Open-Source Agent Operating System(github.com)
11 points by OsamaJaber 8mo ago|3 comments
48.
PicoLM: Run a 1B parameter LLM on a $10 board(github.com)
4 points by OsamaJaber 8mo ago|1 comments
49.
The Floating Dock for Developers(snap-dock.co)
2 points by OsamaJaber 8mo ago|0 comments
50.
by OsamaJaber 9mo ago
Keep talking like that and people will colonize space
51.
by OsamaJaber 9mo ago
MCP support is the real story here Means you're not locked into Claude or Codex Can plug in whatever agent you want
52.
by OsamaJaber 9mo ago
Nice. The config fragmentation across tools is a real annoyance Does it handle conflicts if tools expect different formats for the same setting?
53.
by OsamaJaber 9mo ago
The requirements gap point is underrated. AI guesses where a human would ask By the time you catch it in review, you've already wasted the time you saved -_-
54.
by OsamaJaber 9mo ago
Surprised it took this long for someone to write it up properly
55.
by OsamaJaber 9mo ago
250 C files were deleted. 2032 to go. Watching Zig slowly eat libc from the inside is one of the more satisfying long term projects to follow
56.
by OsamaJaber 9mo ago
30+ years maintaining one of the most critical pieces of infrastructure on nearly every Linux and Unix system, and he's currently looking for a sponsor to fund continued development. Every company running sudo in production owes this m
57.
by OsamaJaber 9mo ago
Great job! This is the kind of project that should exist for every complex system Systems like vLLM's codebase are massive and hard to follow Would love to see the same approach for other infra (a nano-Kubernetes, nano Postgres.....
58.
by OsamaJaber 9mo ago
Sooooo underrated Incredibly handy when you're away from your machine
59.
by OsamaJaber 9mo ago
The real issue was never AI in Windows It was AI with no clear user benefit. A Copilot button in Notepad doesn't solve a problem anyone has Good to see them pulling back, but the test will be whether the features they keep actually ear
60.
by OsamaJaber 9mo ago
We did something similar with TypeScript strict mode Turned it on per file with a ratchet count, and over a few months, the whole codebase was strict without ever blocking anyone

More ›