Kernel review, automated
Sashiko is a tool Google engineers have been quietly building for the past several months — an agentic system that reviews Linux kernel patches. It’s now open source, donated to the Linux Foundation under Apache 2.0, and as of this week it’s running on every submission to the Linux kernel mailing list.
The Linux kernel maintainer problem is real and not new: thousands of patches per month, a small group of deeply specialized people who understand any given subsystem, and a well-documented pattern of burnout. Sashiko isn’t trying to replace that human judgment — it runs a multi-stage protocol that checks everything from architectural correctness down to concurrency issues and memory lifecycle, then flags problems before a maintainer ever sees the patch.
The benchmark they’re using is a bit unusual but honest: given the last 1,000 commits with “Fixes:” tags (meaning patches that were accepted and later found to be buggy), Sashiko catches 53.6% of them using Gemini 3.1 Pro. The framing in their announcement is deliberately dry about this: “In some sense, it’s already above the human level given that 100% of these bugs made it through human-driven code reviews and were accepted to the main tree.”
That’s not a boast, it’s a description of how hard kernel review is. False positive rate is harder to measure — they estimate under 20%, which for a first-pass triage tool is workable. The community reaction is predictably mixed; there’s always tension in open source around automated reviewers. But the tool is written in Rust, builds on per-subsystem prompts developed by Chris Mason, and has a GitHub repo that’s open to inspect. That’s the right way to introduce something like this.
Tinybox gets an upgrade
Less philosophically charged but practically interesting: tinygrad’s Tinybox v2 lineup is now shipping with current-generation hardware. The red v2 puts four AMD Radeon RX 9070 XT cards (RDNA 4) in a single 15A-plug enclosure — 64GB of VRAM, 778 TFLOPS FP16, for around $12K. The green v2 Blackwell goes the other direction: four RTX Pro 6000 cards, 384GB of VRAM, $65K, made to order.
The red v2 is the more interesting product. The 9070 XT is a gaming GPU being repurposed for training and inference, and the $12K price point puts a real training machine within reach of a small team or a serious independent researcher. It’s the kind of thing that would have required a cloud budget two years ago.
George Hotz has been working toward a fully sovereign AMD software stack — tinygrad, custom driver, runtime, libraries — and was reportedly one assembler away from completing it as of early 2025. Whether that matters to most buyers is debatable, but for people who want to run models entirely outside the CUDA/NVIDIA dependency chain, it’s a genuine option.
The two stories don’t obviously connect, but there’s something common in the direction: AI is showing up in places that used to be purely about human craft — kernel maintenance, low-level tooling, offline training infrastructure. Not replacing those things, but changing the economics and the workflow around them.