AI Daily Industry Briefing | 2026-11-16
AI Daily Briefing
2026-11-16 | ๐จ๐ณ ไธญๆ + ๐บ๐ธ English Past 24h focus: AI agents & open-source LLMs
๐ฅ GitHub Trending โ AI/LLM
(Today's AI-related trending repos, 8 picked by relevance to agents / LLMs / training-inference infrastructure)
- GoogleChrome/modern-web-guidance-src โ TypeScript โ 1,219 | Keep your coding agent up to date with the latest web best practices โ The Chrome team packaged web best practices into a spec source that coding agents can consume directly.
- dream-num/univer โ TypeScript โ 988 | AI office SDK โ Bundles editing, calculation, collaboration and file handling into an office kernel that "humans and agents share".
- backnotprop/plannotator โ TypeScript โ 866 | workspace to stay in the loop with agents โ A workbench for humans to cut in while an agent works: annotate, review, and feed feedback back to the agent in one click.
- ChromeDevTools/chrome-devtools-mcp โ TypeScript โ 717 | Chrome DevTools for coding agents โ The official MCP server, exposing DevTools capabilities so coding agents can debug browsers.
- alibaba/open-code-review โ Go โ 691 | hybrid code review โ Hybrid code review combining a deterministic pipeline with LLM agents, line-level comments, built-in multi-language rule sets.
- Tencent/LoopForge โ Go โ 423 | self-hosted coding-agent delivery workflows โ Self-hosted orchestration of CodeBuddy / Codex / Claude Code, with scripts and human approval gates.
- ag-ui-protocol/ag-ui โ TypeScript โ 355 | Agent-User Interaction Protocol โ An interaction protocol layer for wiring agents into frontend applications.
- google-labs-code/stitch-skills โ TypeScript โ 247 | Agent Skills for Stitch MCP server โ Follows the Agent Skills open standard, compatible with Antigravity / Gemini CLI / Claude Code / Cursor.
๐ arXiv Papers โ cs.AI / cs.LG / cs.CL
(The 5 most relevant to agents / reasoning / LLM training / inference)
- AgentSSL: Can MLE Agents Leverage Unlabeled Data? | cs.AI | Ethan Lin, Akanksha Sarkar MLE agents had only been evaluated in fully supervised settings; this paper is the first systematic test of whether they can use unlabeled data, and identifies annotation and pipeline engineering as the two main barriers to deployment.
- CoReMem: Cost-Internalizing Memory for Successive Coding Tasks with Prospective Valuation | cs.AI | Peiwen Li, Harshit Verma Coding agents must write memories ahead of time on successive tasks; the authors internalize "write/read cost vs future payoff" so the agent itself decides what to remember and when to act.
- Classical Energies Learn to Reason 1,000x Faster than Looped Models | cs.AI | Benjamin Hoover, Hugh Blayney On Sudoku- and maze-style reasoning, replacing looped-model training with energy descent is three orders of magnitude faster and more stable.
- ArchSpace: Recording, Reviewing, and Referencing LLM Architecture Research | cs.CL | Jie Wang, Tong Zhu Architectural innovations in DeepSeek, Qwen and Kimi are rarely recorded in full; this paper offers a reproducible, citable framework for recording LLM architecture research.
- Ballad: Bandit-Based LLM Routing for Automated Heuristic Discovery | cs.AI | Samidha Verma, Ankit Anand Challenges the default assumption that "stronger model, better heuristic discovery", using bandit routing to pick models on demand โ cheaper with no loss in quality.
๐ฆ Twitter/X โ Tracked Accounts
(From 10 tracked accounts: @AnthropicAI @OpenAI @GoogleAI @GeminiApp @sama @Kimi_Moonshot @NousResearch @openclaw @Teknium @dotey)
๐ด Key signals (within 24h)
- @Teknium (11/15 16:23) โ Straight-up model-picking advice: if you're still on Luna, switch to DeepSeek Flash or Terra, "same price but far more accurate"; Luna over-thinks, making it pricier and worse ยท original
- @dotey (11/16 04:20) โ Relays an FT report: Jensen Huang is considering returning as Nvidia CEO, with allies forming a "war room" to lobby the board ยท original
- @varun_mathur (11/16 01:27) โ Ran an agent swarm over PageRank analysis of nearly 2,000 articles to trace how the NYT shapes the Jensen Huang narrative; says Nvidia's market cap has already shed over $1 trillion ยท original
- @dotey (11/16 07:18) โ Translates and introduces Watch a Brain Learn: how four neurons learn to chase light through reward and punishment, taken as a minimalist metaphor for AI learning ยท original
Notable older posts (outside 24h but strong signal this week)
- @GoogleAI (11/13 02:03) โ Releases Gemini 4 Hydrogen: focused on everyday coding and agentic workflows, 2.5x the speed of Gemini 3.8 Flash ยท original
- @openclaw (11/13 12:31) โ OpenClaw supports Gemini 4 Hydrogen the same day ยท original
- @AnthropicAI (11/14 03:00) โ Engineering blog gives 10 pieces of advice on designing and reporting model safety evaluations ยท original
- @modal (11/14 01:04) โ Modal Endpoints lands in Hermes Agent, usable with a single
hermes plugins install modalยท original - @GeminiApp (11/14 06:05) โ Uses Hydrogen plus custom skills to go from a one-line brief straight to a complete sizzle reel ยท original
๐ฆ Twitter/X โ Popular Discussions (broad search)
(Agent / open-source LLM topics, top 5)
- @noahatwhere (11/15 09:03) โ Argues the "era of big LLMs" has arrived: relays that Anthropic demoed roughly a thousand agents formalizing the existing proof of Fermat's Last Theorem within 30 hours (second-hand relay, unconfirmed by the vendor) ยท original
- @Mr_Buzzoni (11/15 23:27) โ Breaks down an "Anthropic engineer roadmap": 23 steps from Python to harness engineering, including evals and real projects ยท original
- @devXritesh (11/15 18:36) โ A production-grade agent is more than "an LLM wrapped in a loop": tools, memory, planning, execution and guardrails are all indispensable ยท original
- @UPing123zzz (11/14 00:27) โ Criticizes the inverted order in which people learn about agents, recommending the Agentic AI knowledge graph be read by dependency ยท original
- @KunhaoZ (11/10 23:12) โ Paper v2: a single vector in weight space can amplify or suppress reward hacking ยท original
๐ฐ Weibo Picks
(High signal-to-noise channels: AI bloggers such as karminski-็ๅป, ๅฎ็ and others; some items are from the past few days and are marked with their actual timestamps)
- @้ๅญไฝ (11/16 08:05) โ Relays Claude engineer Addy Osmani on "cognitive surrender": the more you habitually trust AI output, the easier it is to lose independent judgement ยท original
- @ๅฅๆๅ ๅฝฉ (11/16 08:15) โ People's Daily theory page discusses "agile governance" of agents: set the governance approach first, then keep pace with the technology ยท original
- @karminski-็ๅป (11/14 15:32) โ Hands-on test of ERNIE 5.5 flash preview: decode reaches 267 token/s, with notable spatial understanding and 256K long-context performance ยท original
- @ๅฎ็xp (11/12 06:24) โ Stanford's new course CS329Z "Agent Engineering", lecture one "What is an AI Agent", is now public with bilingual subtitles ยท original
- @ๅ ๅฎนไธ็ฉๆๆฒณๆฐด (11/12 01:05) โ Relays Google researcher Pietro Miotti: letting agents run NASA LiDAR data located an unmapped Maya site in Mexico within 72 hours ยท original
๐ Blog Picks
(Unread, AI/LLM topics from the past 36h)
- A step toward defining the Standard Model on a lattice | OpenAI News | 11/15 โ OpenAI uses AI-assisted lattice computation to advance the definition of the Standard Model, the latest case in its "AI for Science" narrative.
- Our new blog for science | OpenAI News | 11/15 โ OpenAI launches a new science blog dedicated to AI progress in basic research.
- Jensen Huang might be plotting a return and more of the latest Nvidia drama | The Verge | 11/15 โ A rundown of the comeback rumors after Jensen Huang's departure and the board investigation.
- Congress wants to tighten export rules on AI chips after Nvidia spy scandal | The Verge | 11/15 โ After the Nvidia spy scandal, the US Congress pushes to tighten AI chip export rules.
- Understanding Positional Embeddings in Modern LLMs | Ahead of AI | 11/15 โ Sebastian Raschka systematically walks through positional encoding schemes and trade-offs in modern LLMs.
๐ฏ One-line Takeaway
"The model layer is racing on speed and cost (Gemini 4 Hydrogen, Kimi K3.1 Pro open weights), while the agent layer is racing on engineering โ today, from the GitHub charts to arXiv, the protagonists are all scaffolds for putting agents into real workflows."
Generated automatically by Hermes Agent | Sources: GitHub Trending ยท arXiv ยท Twitter/X ยท Weibo ยท RSS ยท HF Papers
