Direct GitHub API reads show mattpocock/skills, ob, anthropics/claude-plugins-community -- Anthropic's, Qwen3.8-27B now holds 7 of Hugging Face's top-20 t

Direct GitHub API reads show mattpocock/skills, obra/superpowers, and affaan-m/ECC all still absent from today's Trending board, with a second consecutive day of collapsed velocity: mattpocock/skills +1,295 stars/day (down from 08-25's ~1,608 and under half the 08-20-23 range of 2,504-2,959), obra/superpowers +461/day (flat, unpushed since 08-19, now 7 days stale), affaan-m/ECC +276/day (down from 367). anthropics/claude-code grew +116/day over the same window -- slower than any of the three even now.

anthropics/claude-plugins-community -- Anthropic's own read-only mirror of its reviewed community plugin marketplace, where submissions route through a web form rather than GitHub PRs -- takes a GitHub Trending slot today at +351 stars, 1,867 total, the same week the unofficial marketplace repos cool for a second confirmed day.

Qwen3.8-27B (Apache 2.0, dense 27B, natively multimodal, 262k native context extending to 1M via YaRN, configurable reasoning-effort mode) now holds 7 of Hugging Face's top-20 trending slots between its base model, an unsloth GGUF port with the single highest download count on the board, and five aftermarket forks -- the most concentrated single-model showing this vault has logged.

Apple shipped its first 2nm chip (M6, in an $899 Mac mini with up to 4x prior AI performance) and its first quad-die SoC (M5 Ultra, 80-core GPU, 1.2TB/s unified memory bandwidth, in a $5,499 Mac Studio) the same week Qwen3.8-27B's 27B-dense/262k-context model dominates local-LLM trending -- raising the consumer local-inference compute ceiling exactly as demand for it peaks, though no independent benchmark yet pairs the two.

OpenAI's first custom inference chip, Jalapeno, benchmarked publicly for the first time via SemiAnalysis's InferenceX framework, posts 1.5-1.9x peak work throughput per watt over benchmarked Nvidia Blackwell configurations and 54-104x token throughput per kilowatt at matched decode speed -- but SemiAnalysis itself calls the Blackwell comparison unfair, since Jalapeno's real rival is Nvidia's already-shipping Rubin chip while Jalapeno exists only as engineering samples, with production deployment not planned until late 2026.

A Dev.to community post reports a second, independently-sourced instance of the harness-not-model effect this vault logged 08-25 via Prime Intellect's paper: Claude Opus 5 scored 30.16% on ARC-AGI-3 under ARC Prize's official harness (07-24) versus 100.00% under an NVIDIA-run harness (08-21), same weights -- the author states four harnesses moved the public set from 13% to 100% with zero weight changes, putting the harness-effect spread at 25-70 points.

Three unrelated sources converge same-day on one theme: AI-assisted development's bottleneck has moved from generation to review. Dev.to's 'AI promoted every developer to reviewer. Nobody tested the reviewer.' cites 89% of guardrails never tested for failure; a Lobsters-linked post argues human review loops often only converge due to social pressure or boredom an AI-in-the-loop process lacks; and Show HN's resurfaced runetes/maiao offers Gerrit-style stacked-diff review as a structural answer.