Today's Nvidia-Hugging Face coverage draws 1,867 H, Claude in Chrome reached general availability on 2, A new arXiv paper, WikiSkill, formalizes a framewo

Today's Nvidia-Hugging Face coverage draws 1,867 Hacker News points / 866 comments -- the largest single-story engagement this vault has logged this month -- but also surfaces a Business Insider detail that talks 'had not yet produced a signed agreement and could still atomize,' plus reporting that HF CEO Clem Delangue had spent much of 2026 publicly aligning with Nvidia's open-source stance before the deal. This pulls confidence in the deal's finality back from yesterday's 0.8 to 0.65.

Claude in Chrome reached general availability on 2026-08-26 across every paid Claude plan, with Anthropic reporting 0% attack success against Sonnet 5 and Opus 5 (0.3% against Fable 5) after a content-screening probe and action-verification classifier, down from 17.6%/3.8% pre-mitigation. This is the first vendor-published near-zero figure this vault has logged against its own standing 0.85-confidence thesis that no shipping AI browser has adequate prompt-injection isolation -- unverified by independent testing using Black Hat's own named techniques.

A new arXiv paper, WikiSkill, formalizes a framework that compiles agent execution experience into a persistent shared wiki that subsequent skill updates build from -- ablations show removing the persistent-wiki layer measurably degrades skill quality versus keeping only scattered optimization-history logs. This is independent, formally-measured validation of the wiki-compounding mechanism this vault's own daily research pipeline (scrape to OKF bundle to wiki ingest) already runs on.

Google will require Play Store apps to meet new Android memory thresholds starting February 2027 -- foreground apps on 8GB-RAM devices capped at 2.25GB, background/user-perceived services at 1.5GB, bitmap memory at 200-400MB -- driven by AI-datacenter demand tightening the global memory-chip supply. Non-compliant apps face reduced visibility and publishing capability rather than an outright ban, enforced through Play's app-review process.

"Small Models Have Arrived" (Hacker News #2, 566 points/253 comments) argues small, specialized models paired with RAG and task-specific tuning have reached a 'good enough' deployment stage -- a community-sentiment complement to this vault's ongoing tracking of Qwen3.8-27B, GLM-5.3-Flash, and Ornith-1.5's unreproduced benchmark-parity claims, arguing the deployment decision no longer waits on those claims resolving.

SourceHut now prohibits training or fine-tuning LLMs on its hosted code without written permission and bars automated scraping for training datasets, applying to public and private repositories alike -- announced the same day a two-month-old post, 'Please stop flooding our projects with AI slop to furnish your CV,' resurfaces simultaneously on Hacker News and Lobsters, arguing AI-generated PRs and vulnerability reports optimize for contributor-profile metrics over genuine project improvement.

Google Play's top-30 shows com.openai.chatgpt at #3 (4.8 stars) -- the first visible AI-chat incumbent since this vault's AI-assistant-free streak logged 08-19 through 08-26. No Claude or Gemini appears today; held to this page's own standard (a second consecutive day) before reading as a resumption rather than a one-day flicker.