Open Source Radar — October 7: agents, CUDA kernels, robot brains
Today's daily trending board splits three ways: a skill that fixes how your coding agent talks to you, DeepSeek's inference math made readable, and Stanford open-sourcing how robot world models get built. All three verified at the source.
i-have-adhd (Python, MIT, about 54,600 stars) is a ten-rule skill that stops your coding agent burying the answer: lead with the next action, number multi-step tasks, end with one concrete next step, no preamble, no recap, no "Hope this helps!". The repo's before/after is the whole pitch — a rambling paragraph about an auth flow becomes three numbered commands and a test to run — and it's climbing daily trending because it's one small file you can drop into Claude Code or any agent that reads skills, with effects you feel on the very next prompt. If you've ever scrolled past throat-clearing to find the fix, this is the cheapest correction on the board.
DeepGEMM (CUDA, MIT, about 8,800 stars) is DeepSeek's tensor-core kernel library: one codebase for the matrix math its models run on — FP8, FP4 and BF16 GEMMs, fused MoE with overlapped communication, MQA scoring for its lightning indexer — all compiled on the fly, so nothing needs CUDA compilation just to install. It's back on daily trending after a September update packed in sparse-indexer and MoE optimizations, and DeepSeek claims it matches or beats expert-tuned libraries across common matrix shapes while staying small enough that a human can actually read the kernels. This is where inference cost comes from; if you want to see what serving a MoE model really spends cycles on, this is the cleanest open example there is.
OpenWAM (Python, brand new, Stanford) is an open framework for world-action models — the robot systems that predict what they'll see next and decide how to act, in one model or as separate parts you bolt together. It ships a shared video-and-action architecture pretrained on roughly 3.34 million robot and human-manipulation trajectories (14,640 hours of video), plus interchangeable "interaction programs" that let you predict video first, actions first, or both together under otherwise identical conditions — the comparison the field currently can't run because every lab uses its own backbone and training recipe. Code, quickstart and pretrained weights are all public on day one; the repo has barely over a hundred stars, so you're early. It's a research framework, not a product, but it's the most credible open crack at "what does a robot brain look like" in a while.
Worth watching this week.
Which of these would you actually install first? Tell us in the comments.
