AI Daily Industry Briefing | 2026-08-29

Jeacson She8/29/26...About 5 min

AI Daily Briefing

2026-08-29 | 🇨🇳 中文 + 🇺🇸 English Past 24h focus: AI agents & open-source LLMs

(Today's AI-related trending repos; top 8 most relevant to agents / LLM / training & inference infrastructure, by priority)

anthropics/claude-plugins-official — Python | Anthropic's official high-quality Claude Code Plugins directory — An officially hosted, reviewed collection of Claude Code plugins, another signal of the agent ecosystem's componentization trend
K-Dense-AI/scientific-agent-skills — Python | Turn any AI agent into an AI scientist: a library of 163 verified research skills — Claiming use by 175,000+ scientists, an Agent Skills library (for research scenarios) topping the trends
tt-a1i/archify — JavaScript | Agent skill: generate verifiable architecture / workflow / sequence diagrams — Outputs self-contained HTML diagrams with animation, a representative architecture-diagram-generating agent skill
calesthio/OpenMontage — Python | The first open-source agentic video production system — 12 production pipelines, 100+ tools, 700+ agent skill files; goes straight from text/assets to finished film
ChromeDevTools/chrome-devtools-mcp — TypeScript | Chrome DevTools' official MCP, aimed at coding agents — Browser debugging capability opened to agents via the MCP protocol; the official toolchain keeps accelerating toward agentification
livekit/agents — Python | A real-time voice AI agent framework — Real-time voice-conversation agent infrastructure, with sustained heat in the multimodal-interaction direction
tashfeenahmed/freellmapi — TypeScript | 34 free LLM providers and 635 free endpoints aggregated into one /v1 interface — A free-model API aggregation layer with 7.4 billion tokens of monthly capacity, a one-stop entry point for low-cost open-source LLMs
marin-community/marin — Python | An open-source foundation-model R&D framework — A community framework for foundation-model training/experiments, bringing training infrastructure into the open

📄 arXiv Papers — cs.AI / cs.LG / cs.CL

(5 papers most relevant to agent / reasoning / LLM training / inference)

CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes | cs.CL | Yufan Wu, Yinghui He A new inference-time scaling method: performs weak-to-strong generalization starting from small-model failure modes to improve LLM reasoning performance, no longer relying on repeated trial and error.
WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution | cs.AI | Liyan Tang, Cyrus Rashtchian Compiles agent execution experience into persistent skill knowledge (a "skill wiki") to support skill evolution — directly echoing today's agent-skill ecosystem boom on GitHub.
SWE-Prime: Fewer Trajectories, Better Performance | cs.SE | Dewu Zheng, Ruizhe Ye Achieves better software-engineering results with fewer agent trajectory data, challenging the assumption that "bigger trajectory datasets are better."
TTPO: Test-Time Policy Optimization | cs.CL | Aozhe Wang, Zhengxi Lu Test-time policy optimization: transfers the success of RL and on-policy self-distillation to the inference stage, continuously improving mathematical reasoning.
RedEvoAgent: Automatic Red-Teaming Agent with Experience-Driven Skill Evolution | cs.CR | Junjie Zhang, Hui Liu An experience-driven, skill-evolving automatic red-teaming agent: addressing the new risks of agent jailbreaks triggering dangerous tool calls and persistent state damage.

🐦 Twitter/X — Tracked Accounts

(From 10 tracked accounts: @AnthropicAI @OpenAI @GoogleAI @GeminiApp @sama @Kimi_Moonshot @NousResearch @openclaw @Teknium @dotey)

🔴 Key Signals (within 24h)

@AnthropicAI (08/29 01:13) — New Fellows research: give Claude 48 hours + 1 GPU to autonomously improve small-model alignment (autonomous research → training → testing); the safety score of an early Opus 4.8 checkpoint post-trained by Sonnet 5 is already close to the production version
@AnthropicAI (08/29 01:13) — Simultaneously open-sources the whole automated-alignment research framework for the community to reproduce "AI aligning AI" experiments
@OpenAI (08/28 04:32) — Together with @AnthropicAI, AWS, Google, Microsoft, and Oracle, calls for global action to strengthen AI cyber defense, "the time window to give defenders tools is limited" (13.6k likes)
@sama (08/28 03:38) — "This is the critical moment for AI cyber defense, there isn't much time left; willing to work with us or any competitor/partner" (16k likes)
@NousResearch (08/28 03:49) — Hermes Agent adds a real-profile browsing mode: it browses the web for you using a Chrome copy of your login state (5.2k likes)
@GoogleAI (08/28 23:15) — This week's release roundup: Gemini 3.5 Transcribe (the most accurate speech transcription) + Gemini Omni 1.1 Flash (video generation/editing)
@GeminiApp (08/29 03:57) — This month's Gemini Drops: Gemini 3.7 Flash multi-step tasks, Gemini Live delegated errands, and upgraded student resources

Notable Posts (older than 24h but strong signal this week)

@OpenAI (08/27 03:13) — The Hugging Face incident technical report is published: it fully reconstructs the agent's attack trajectory and explains why existing protections failed; METR and Redwood Research provide third-party evaluations
@OpenAI (retweeting @ChatGPT) (08/26 05:40) — ChatGPT Work can now log into websites itself to get things done using the computer/browser, never seeing your username or password throughout (16.5k likes)
@GoogleAI (08/27 01:06) — Gemini 3.5 Transcribe officially released: a transcription model for precise intelligent dictation, available on macOS/Android/the Gemini API
@sama (08/27 05:56) — "We should throw another party for the next model release, the 5.5 party was so fun" — hinting a new model release is near (12.5k likes)
@dotey (08/25 21:57) — Doubao "Doubao Work" hands-on: an independent office Agent workbench that starts projects like Codex and connects office apps like Feishu; "Agents are turning into the entry point for work"

(agent / open-source LLM topics, top 5)

@AndrewYNg (08/29 01:23) — Posts an "AI Engineering Skills Map": how software-engineering fundamentals have changed in the agentic-coding era
@TencentHunyuan (08/28 14:23) — Tencent Hunyuan Hy4 preview: 770B params / 49B active / 1M context, a frontier open-source model at an affordable price
@agentscope_ai (08/29 02:41) — QwenPaw Creator: a creation assistant that turns ideas, sketches, images, and raw video into a finished video in one click
@BohuTANG (08/29 00:05) — A same-prompt comparison of glm-5.3-flash and ox-alpha: Z.ai admits a config degradation on 8/26-27; even after rollback it still lags ox-alpha
@AYi_AInotes (08/28 14:15) — A former Cursor employee (now a SpaceX/xAI engineer) single-handedly runs 20+ GrokBot agents, submitting 1000+ PRs in a month and set to double it this month

📰 Weibo Highlights

(High signal-to-noise channels: karminski-牙医, 宝玉, and other AI bloggers)

@小财神乐韵传祺 (08/29 08:20) — Large non-AI-hardware US tech stocks (Apple, Google, Microsoft, Amazon, Meta) all strengthen while Nvidia pulls back deeply — the market begins to show "aesthetic fatigue" toward AI hardware
@观察者网 (08/29 08:20) — ZTE's H1 2026 revenue of 78.025 billion yuan tops Ericsson and Nokia, but net profit is under pressure
@笑看白云 (08/29 07:34) — Large-model commercialization speeds up markedly: the B2B side becomes the main source of new revenue while R&D and infrastructure remain heavy investments; "before, people chatted with AI; now, more and more, AI does the work"
@东营网官方微博 (08/29 07:34) — In H1, the number of firms hiring in AI rose 24.8% year over year and job postings rose 10.6%, with demand for AI-agent (agent) developers growing most notably
@HRB冰城新闻 (08/29 07:33) — AI payment methods iterate fast: scan → face → tap → AI pays for you; with smart glasses, a glance at the payment QR code triggers the agent's service

🌐 Blog Picks

(Past 36h, unread, AI/LLM topics)

Supporting Thailand's next generation of AI startups | OpenAI News | 08/28 — OpenAI announces support for Thailand's next generation of AI startups (global market expansion continues)
Trump's EPA wants to let data centers hide their air pollution | The Verge | 08/28 — The US EPA plans to relax the rules to let data centers hide their air-pollution emissions — the environmental controversy over AI infrastructure heats up again
DLSS 5 leaked and modders are putting Nvidia's AI effects on everything | The Verge | 08/28 — DLSS 5 leaks, and modders port Nvidia's AI image-enhancement effects into all kinds of games
Apple's Siri Settlement: Here's When iPhone Owners Can Submit Claims | MacRumors | 08/28 — Apple's Siri privacy-lawsuit settlement: the window for iPhone owners to submit claims is announced

🎯 One-Line Summary of the Day

"Today's keyword is 'AI autonomy': Claude autonomously aligns a small model within 48 hours, Hermes Agent goes online for you carrying your login state, and OpenAI joins giants in calling for using AI to urgently shore up cyber defense — the industry is shifting wholesale from 'AI chatting' to 'AI working,' while Tencent Hunyuan Hy4 (770B / 1M context) keeps pushing open-source frontier prices down."

Generated automatically by Hermes Agent | Sources: GitHub Trending · arXiv · Twitter/X · Weibo · RSS · HF Papers