オープンソースAI

主要ニュース

1. 1

   [Hands-on] Build Your Own AI Avatar With Human-like Memory

   (100% open-source, works in real-time)

2. 2

   Banning Open Source AI Would Be A Mistake

   This post was originally an op-ed co-authored with Kevin Xu of Interconnected for a general, non-technical audience.

3. 3

   My suitcase robot gets high off a real gas sensor wired into the LLM sampler

   Article URL: https://old.reddit.com/r/LocalLLaMA/comments/1u9a17y/my_suitcase_robot_gets_high_now_off_a_real_gas/ Comments URL: https://news.ycombinator.com/item?id=48595528 Points: 1 # Comments: 0

4. 4

   Applied AI Engineer/ Product Builder/ Data Science

   Location: Santa Cruz, CA Remote: Yes Willing to relocate: Remote Preferred. SF/Socal Hybrid ok Applied AI / product builder (Full Stack). I’m strongest at turning messy ideas, workflows, datasets, and APIs into useful AI-assisted products and prototypes; especially where product judgment, user understanding, automation, and modern LLM/agent tooling matter. In a world where AI is become exponentially more intelligent in general, a recurring challenge has become harnessing that intelligence for specific, personalized real world applications, not just generic generalized intelligence. This is what I do in addition to and in conjunction with Product Development, Data Science, and Bot Building. I scope problems by optimizing for specific organizational or enterprise goals, and choose the tech stack that helps support the achievement of those goals in both the short and long term. I work with a lot of open source, personalized, persistent AI agent harnesses and platforms; avoiding the lock i

5. 5

   Fearless Concurrency on the GPU: Safe GPU inference in Rust, competitive with vLLM/SGLang [R]

   I maintain cuTile Rust and just posted the paper "Fearless Concurrency on the GPU." As more GPU code gets AI-generated, the bottleneck moves from writing it to trusting it. cuTile Rust lets you write or generate GPU kernels whose memory safety and data-race freedom are verified by the compiler, through Rust's ownership and borrow checking. You get those guarantees by construction. It's a tile-based programming model that lowers to CUDA Tile IR, carrying Rust's ownership model across the launch boundary. You partition a mutable output into disjoint mutable sub-tensors, pass inputs as shared references, and write tile kernels with single-threaded semantics that the compiler maps to thread blocks. End to end, we built Grout, a Qwen3 inference engine, on cuTile Rust with Hugging Face. At batch-1 decode it reaches 171 tok/s for Qwen3-4B on an RTX 5090 and 82 tok/s for Qwen3-32B on a B200, competitive with vLLM and SGLang. Batch-1 decode is memory-bandwidth-bound, and Grout's throughput