AI Daily Dev — October 14, 2026

Torvalds Pushes Two LLM-Assisted Kernel Branches, Each Patch Tagged 'Assisted-by: LLM'

Oct 11-12: `noref-filemap-read-with-llm` branch in Torvalds' tree holds an 8-patch page-cache rework co-authored with Claude Code; eight-thread hot-cache read benchmark jumps from ~10M to 30-33M reads/sec on his laptop.
Same weekend: an `assoc-array-fixes` tag lands — Torvalds says the reported fix was 'too obscure for me to follow,' so he had an LLM refactor the library and build a userspace harness that caught three more bugs dating to the original 2013 commit.
Commit message credits 'Claude Code … under my direction' and each patch carries an `Assisted-by: LLM` trailer; targets the Linux 7.4 merge window, not 7.3.
Lands days after Torvalds publicly rejected calls to label or ban LLM-assisted kernel patches; HN and Phoronix threads active through the weekend.

community phoronix.com

Oct 12: Meta Model API opens to Muse Wonder at $4/$20 per M tokens (cached input $0.20/M), matching Opus 5.5 and GPT-6 Sol; 1M-token context, text/image/video/audio input.
Artificial Analysis Intelligence Index v4.3.2: Muse Wonder 52, Muse Wonder Full 56, Opus 5.5 58 — Full ties Sonnet 5.5 and trails Opus by two points on the independent index.
Meta's own numbers: Wonder Full 79.0 on DeepSWE 1.1 and 57.5 on SWE-Marathon (vs. ~74 and 46.3 for Opus 5.5), with a 59.0 self-reported AA figure Artificial Analysis itself publishes as 56.
Wonder Full is not sold through the API — only via $50 and $200/month Muse Code subscriptions; first frontier-tier release since Meta's Superintelligence Labs reorg.

models techcrunch.com

Oct 12: 320B-parameter MoE with 18B active per token and a 1M context window, fine-tuned on legal corpora and RL in legal environments.
Scores 19.2% on Harvey's LAB benchmark across 24 practice areas — +12.5 points over the GLM-5.3-Flash base; Z.ai also claims frontier-level LegalBench and PLawBench results without publishing numbers.
No public API yet: access gated behind a consultation form, and the model also ships inside Z.ai's AutoClaw desktop agent.
Second domain-specialist frontier model this quarter after Google's cyber-defense-first Gemini 4 Argon — the 'frontier for one job' pattern is a trend now, not an outlier.

models aiweekly.co

debpalash/VoiceStudio: 64.1k stars (+1,748 today) — AGPL-3.0 desktop app for voice cloning, dubbing, dictation, transcription, voice design, and audiobook generation, all local, no account, no API key.
Clones from ~3 seconds of audio; recent v0.5.x releases add a 'Voice match' consistency mode so each line of a speaker clones from one shared reference instead of per-line snippets.
Project claims coverage across 646 languages; renamed from OmniVoice Studio in v0.5.0 and now hitting GitHub Trending for the second time in six weeks.
Fits the broader 'run frontier-tier media models on your laptop' trend alongside AirLLM's Kimi K3 demo and ArtCraft's open-source creative suite.

open-source github.com

bilawalsidhu/gods-eye-view: 55.8k stars (+1,961 today) — Vite + CesiumJS web app that stacks live public feeds (aircraft, ships, satellites, earthquakes, traffic, cameras) on one interactive globe.
Keyless mode works offline with baseline imagery; optional provider keys unlock photorealistic 3D, Google Place search, live vessel feeds, NASA fire data, and AI voice control.
Formerly WorldView; author notes 'modeled view' labels for areas without live feeds — not a literal spy satellite, but close enough for a sci-fi UI that's lighting up HN and X.
Second-day #2 on GitHub Trending after morluto/rea; packaging warns against exposing it outside a trusted LAN since there's no auth.

open-source github.com