AI Daily Industry Briefing | 2026-11-09

Jeacson She11/9/26...About 5 min

AI Daily Briefing

2026-11-09 | 🇨🇳 中文 + 🇺🇸 English Past 24h focus: AI agents & open-source LLMs

(Today's AI-related trending repos, ranked by relevance to agents / LLM / training and inference infrastructure)

milind-soni/OpenMausBot — TypeScript | ★1,439 today — An open-source Grok Bot alternative that gives the bot a virtual machine it can actually operate.
meta/astryx — TypeScript | ★875 — Meta's open-source customizable design system, built natively for agent-generated UI.
vercel-labs/agent-browser — Rust | ★815 — A browser-automation CLI for AI agents, written in Rust.
anthropics/skills — Python | ★792 — Anthropic's official public Agent Skills repository, the reference implementation for the skill ecosystem.
dsx-ai-factory/nvke — Go | ★345 — NVIDIA Kubernetes Engine (NVKE): a management plane for GPU K8s on bare metal.
anthropics/claude-quickstarts — Python | ★32 — A collection of examples for quickly building deployable apps with the Claude API.

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

SwarmEvolve: Latent Agent Swarms for Evolutionary Discovery and Optimization | cs.AI | Qile Jiang, George Em Karniadakis Splits LLM-guided evolutionary search into "latent-space agent swarms" that improve programs in parallel, then merges the high-scoring solutions.
Scientist World: Grow a Research World for AI-Scientists | cs.AI | Taiming Lu, Sachin Konan Builds a continuously growing research environment (open problems + shared literature) for AI-Scientists, filling the gap that they have "no world to grow up in".
ReproBridge: Building Reproducibility into ML Research Artifacts | cs.AI | Yubo Wang, Rui Meng When coding agents try to reproduce ML experiments, information is scattered across papers, code and supplementary files, causing failures; proposes a structured reproducibility bridge.
Benchmarks as Software: Terminal-Bench and the Case for Continuous Benchmarks | cs.AI | E. Kelly Buchanan, Crystal Zhou Argues that coding-agent benchmarks should be maintained as continuously updated software rather than released once.
Spectral and Subspace Dynamics of LoRA in Supervised Fine-Tuning and Reinforcement Learning | cs.LG | Chenruo Liu, Zihan Wang Compares how LoRA's subspace dynamics differ between SFT and RL training, explaining why identical hyperparameters perform differently.

🐦 Twitter/X — Tracked Accounts

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

🔴 Key signals (within 24h)

@NousResearch (11/09 07:30) — Warns about a class of agent design risk: letting an agent swap models or rewrite its own code at will without telling the user is "building on quicksand" · original
@dotey (11/09 05:20) — Jensen Huang wrote to employees admitting infidelity toward his wife; the NVIDIA board formed an independent committee to investigate whether he has become a target of foreign intelligence operations (tied to reporting on a Megaspeed executive spy case) · original
@NousResearch (11/09 02:47) — Hermes Agent added an official security plugin: first map the attack surface and trusted inputs, then split the code across parallel subagents and verify each finding individually · original
@dotey (11/09 01:24) — Quotes DeepSeek's positioning: hopes DeepSeek Harness becomes "the operating system of the agent era" · original
@Teknium (11/08 14:39) — Built himself a "Signal Dog" with a Hermes plugin: a microphone listens for the doorbell, oven and fridge alarms and notifies him in chat, keeping no audio · original
@dotey (11/08 11:21) — Breaks down Anthropic's agent memory approach: remember while working, then use a background "dreaming" step to revisit the original conversations, correct mistakes and reorganize notes · original
@NousResearch (11/08 08:23) — Community-built Hermes plugins pass 1,000 · original

Notable older posts (outside 24h but strong signals this week)

@Kimi_Moonshot (11/05 00:00) — Opens the Kimi K3.1 Pro weights: 2.8T parameters, 1M context; the official line is to "keep advancing open intelligence" · original
@NousResearch (11/07 05:48) — Releases Arete-27B: post-trained from Qwen3.8-27B to reinforce Western values, with a seven-dimensional evaluation built on 803 prompts · original
@OpenAI (11/06 01:09) — GPT‑6 Cyber launched: security-research capability and alignment improved together, already used in Linux kernel vulnerability research · original
@GoogleAI (11/06 01:08) — Gemini Omni 1.2 Flash: lower cost, video generation/editing limit raised to 15 seconds, better shot continuity · original
@AnthropicAI (11/07 05:16) — Engineering blog details Claude's "dreaming" memory mechanism: organizing what it learned, correcting errors and improving recall · original

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

@dhh (11/08 20:10) — Cites Asahi Linux policy: LLMs harm software freedom, open-source communities, the open internet, the environment and user consent, so LLM contributions are banned broadly · original
@openSUSE (11/08 18:56) — Shares how it fine-tunes a local LLM on OBS submission data (spec diffs, file lists, release notes) to automate the tedious but critical release process · original
@business (11/09 02:10) — Reports that DeepSeek published adoption figures for its AI agent infrastructure, continuing to challenge OpenAI / Anthropic with an open-source approach · original
@0xRafy (11/09 01:37) — Someone built BLA (Blind LLM Agents) to plant "minefields" for reverse engineering: fake credentials and misleading metadata inside software, leading AI reversers into dead ends · original
@iamlukethedev (11/08 15:56) — Rounds up the 69 PRs merged into Hermes on 11/07: the in-app browser reusing real browser login state, password filling, and more · original

📰 Weibo Picks

(High signal-to-noise channels: karminski-牙医, 爱可可-爱生活 and other AI bloggers)

@爱可可-爱生活 (11/09 08:15) — synchopate open-sources llama-modchip: MoE quantization, kernel patches and runtime tooling around llama.cpp/ggml, aiming to run a 284B-class model at 256K context on 64GB of pure CPU · original
@WHQ_CHINA (11/09 08:15) — Reposts a Zhihu column: building a real-time multimodal voice AI agent with the open-source Python framework Pipecat · original
@绿天使-W (11/09 08:10) — A news digest says Tencent released and open-sourced a new-generation large model, Hy4, rated first among open-source models on overall agent-task capability (original source unverified) · original
@德州中院 (11/09 08:03) — Shanghai's first case of stealing large-model compute with AI-forged signatures is sentenced: over 4 million accounts farmed compute, three years in prison · original
@karminski-牙医 (11/06 07:02) — Tested Baidu's ERNIE-5.5-Flash-Base: reading the model card and config, it is entirely an efficiency optimization for "1M ultra-long context + repeated sampling in post-training/agents" · original

🌐 Blog Picks

(AI/LLM topics from the past 36h)

Apple is racing to make Siri more agentic | The Verge | 11/08 — Apple is pushing Siri toward being an agent that can carry out tasks on the user's behalf.
Siri AI Rumored to Become 'More Like an Agent' as Early as Next Year | MacRumors | 11/08 — Rumor: Siri could take a more agent-like form as early as next year.
AI broke the cybersecurity world. What's next? | The Verge | 11/07 — Podcast: how AI has completely rewritten the tempo of attack and defense, and where defenses go next.
Why a new report on Anthropic automating AI development is alarming researchers | The Verge | 11/07 — A report on AI automating AI R&D is putting researchers on alert.
NYT Cooking's AI assistant saved my weeknight chicken, but almost ruined my cake | The Verge | 11/07 — Hands-on with the NYT Cooking AI assistant: fine for everyday recipes, still stumbles on precise tasks like baking.
OpenAI valued at $1.25 trillion, as rival Anthropic nears its IPO | The Verge | 11/06 — OpenAI reaches a $1.25 trillion valuation as Anthropic approaches its IPO.

🎯 Today's One-Line Takeaway

"Open weights and agent runtimes are accelerating together: Kimi K3.1 Pro's weights are open and community plugins passed one thousand — while an agent's ability to change its own model or code is becoming the focus of a new round of security and trust debate."

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