Torvalds shipped 7.3 on Oct 18 pushing the tree to ~40.98M lines: AMD Zen 6 prep, faster Btrfs (logging, fsync, direct I/O), stable Intel Nova Lake iGPU support, Valve-driven VRAM management improvements for 8GB-class GPUs, and kernel drivers for the 2026 Steam Controller. Landing in Arch kernels within weeks. The meta-story: Torvalds openly credits AI coding agents for the kernel's growth — the "agents writing infra code" thesis your whole stack is built on, blessed from the top. Same caveat as last time: your omarchy kernel is its own package, so omarchy-update won't pull 7.3 until the linux-omarchy build lands.
LinuxCompatible — Oct 18
Attackers sat inside a developer's GitHub account for 6+ months, then trojanized OVHcloud CLI versions 0.11.0–0.15.0 with an info-stealer aimed at secrets on disk — including Linux x86_64/ARM64 binaries. Root cause: GitHub's release immutability setting was disabled, so existing releases were rewritten without CI/CD ever running. This is a supply-chain hit on exactly the kind of tool you'd curl | install without thinking. Two action items on the record: pin checksums on every new tool install (Strata included), and enable release immutability on any repo you publish.
BleepingComputer — Oct 20 (disclosed Oct 19)
Runs the same task through different coding agents (Copilot, Claude Code, …) and scores outputs against a dependency-graph of criteria with an LLM Judge — logs, workspace snapshots, versioned agent profiles, Docker Compose or K8s. This is exactly the missing measurement layer for your Glimmer-vs-Qwen shootout and the context-mode/harness adjudication work: a purpose-built rig for "which agent setup is actually better" instead of eyeballing trajectories. MIT, self-hosted, and it's the kind of thing Hermes should be running before any more shootout opinions get aired.
It's FOSS — Oct 20
Google (with Meta) open-sourced its native PyTorch stack for TPUs — training, serving, deploy, with integrations into vLLM, SGLang, llm-d 1.0, TorchTitan, VeRL, Ray and HF Transformers v5 (device_map="tpu"). LiteRT (ex-TF Lite) also joined the PyTorch ecosystem for on-device inference. The through-line: PyTorch/vLLM/SGLang is being cemented as the portable serving layer, so anything built locally on llama.cpp/vLLM idioms stays portable if cloud TPUs ever enter the picture. Awareness-only, but it rhymes with the llm-d story above.
GixTools (via Google Cloud Blog) — Oct 20
