X Weekly Digest — Week 22, 2026
Fetched: 2026-05-31 | Accounts: 3 | Posts: 581
Cross-Account Highlights
- NVIDIA Qwen3.6 MoE (quantized) — NVFP4-quantized 35B/3B-active model released on HuggingFace; memory reduced ~3x with near-zero accuracy loss (source)
- Microsoft SkillOpt — Train agent skills in text space without touching model weights; best or tied-best on 52/52 settings across 6 benchmarks and 7 models (source)
- 500-hour AI Infrastructure Curriculum — Viral open-source curriculum for AI engineering (1,294♥), covering infra, MLOps, and deployment (source)
- book-to-skill — Tool that converts technical books into Claude Code skill files for on-demand reference; highly discussed in Chinese dev community (source)
- LocateAnything (CVPR 2026, trending #1 on HuggingFace) — NVIDIA vision-language detection model using Parallel Box Decoding for 2.5x faster inference; replaces slow token-by-token coordinate generation (source)
@tom_doerr — Tom Dörr
Posts this week: 440
Summary (EN): Tom shared a high volume of open-source tool discoveries this week, spanning AI infrastructure, security, developer tooling, and creative applications. Standout posts included a 500-hour AI infrastructure curriculum (1,294 likes), an OSINT tool hunting accounts across 400+ networks, a document-to-JSON converter for LLMs, and a Claude MCP server with 27 security intelligence tools. His feed leans toward curated GitHub repos with concise one-line descriptions.
本週摘要: Tom 本週分享了大量開源工具,涵蓋 AI 基礎設施、資安、開發工具與創意應用。最受關注的是一個 500 小時 AI 基礎設施課程(1,294 讚)、跨 400+ 平台的 OSINT 帳號搜尋工具、文件轉結構化 JSON 工具,以及一個整合 27 項資安情報工具的 Claude MCP Server。整體風格是簡潔一行描述 + GitHub 連結。
Notable posts:
- 500-hour AI infrastructure engineering curriculum —
2026-05-26↻182 ♥1294 - Hunts down social media accounts across 400+ networks —
2026-05-25↻157 ♥1236 - Converts documents and media into structured JSON for LLMs —
2026-05-25↻138 ♥1158 - Removes AI writing patterns from prose —
2026-05-24↻94 ♥1063 - 70 tiered cybersecurity projects plus certification roadmaps —
2026-05-27↻146 ♥945 - Runs Windows apps integrated into Linux desktop —
2026-05-24↻97 ♥936 - Turns Telegram accounts into unlimited cloud storage —
2026-05-30↻115 ♥860 - Python DDoS script with 57 attack methods —
2026-05-28↻137 ♥853 - Interactive step-by-step visualizations for 24 algorithms —
2026-05-25↻101 ♥830 - Gives Claude access to 27 security intelligence tools —
2026-05-27↻131 ♥789 - Maps digital infrastructure with interactive graphs —
2026-05-30↻86 ♥745 - Generates real-time talking avatars from photos and voice —
2026-05-28↻103 ♥686 - Real-time video-driven motion capture for 3D avatars —
2026-05-27↻75 ♥571 - 18 AI personas deliberate complex decisions —
2026-05-26↻88 ♥563
@GitHub_Daily
Posts this week: 35
Summary (EN): GitHub_Daily focuses on Chinese-language coverage of GitHub trending projects. This week highlighted AI tooling for the Claude Code ecosystem (book-to-skill, AgentHub, Huashu Design skill), academic workflow tools (Paper2Any for generating figures/slides from papers), Stanford REAP’s 23,000+ agent skill library for social science research, and Obsidian-related tools. The account skews toward practical AI productivity tools and learning resources.
本週摘要: GitHub_Daily 以中文報導 GitHub 熱門專案。本週重點包含 Claude Code 生態工具(book-to-skill、AgentHub 虛擬開發團隊、Huashu Design 技能)、學術工作流工具(Paper2Any 一鍵從論文生成架構圖與 PPT)、斯坦福 REAP 的 23,000+ Agent 技能庫,以及 Obsidian Digital Garden 發佈插件。整體偏向實用 AI 生產力工具與學習資源。
Notable posts:
- Animal-Island-UI — 動物森友會風格前端組件庫,開箱即用 —
2026-05-24↻74 ♥560 - book-to-skill — 把技術書轉成 Claude Code 技能文件 —
2026-05-30↻73 ♥338 - Taste Skill (19,200+ Stars) — 提升 AI 生成前端界面質感 —
2026-05-26↻36 ♥207 - Paper2Any — 一鍵從論文生成架構圖、PPT、圖表 —
2026-05-26↻30 ♥134 - pyVideoTrans — 開源視頻語音識別+翻譯+配音全流程 —
2026-05-27↻34 ♥121 - Stanford REAP — 23,000+ Agent技能庫覆蓋8大社科學科 —
2026-05-30↻38 ♥113 - Prompt-Engineering-Jumpstart — 14章提示詞教程 —
2026-05-29↻25 ♥100 - ai-learning-roadmaps — 入門到進階AI學習路線圖 —
2026-05-25↻34 ♥95 - Obsidian Digital Garden — 將Obsidian筆記發布成個人網站 —
2026-05-28↻11 ♥95 - AgentHub — 為Claude Code組建46角色虛擬開發團隊 —
2026-05-28↻16 ♥87
@HuggingPapers
Posts this week: 106
Summary (EN): HuggingPapers covered a busy week of model releases and research papers. NVIDIA dominated with several releases: quantized Qwen3.6 MoE, Kokoro TTS (82M param, commercial-use), LocateAnything (CVPR 2026, trending #1), and OneFormer for unified image segmentation. Research highlights include Microsoft SkillOpt (text-space skill training), AutoResearchClaw (23-stage autonomous research pipeline), SciAtlas (43M papers knowledge graph), and NEO-ov (VLM without image encoders). On-policy distillation was flagged as a hot post-training technique.
本週摘要: HuggingPapers 報導了密集的模型發布與研究論文。NVIDIA 主導本週發布:量化版 Qwen3.6 MoE、Kokoro TTS(82M 參數,商業可用)、CVPR 2026 熱榜第一 LocateAnything,以及統一分割模型 OneFormer。研究亮點包含 Microsoft SkillOpt(文本空間技能訓練)、AutoResearchClaw(23 階段自主研究流水線)、SciAtlas(4300 萬篇論文知識圖譜),以及 NEO-ov(無圖像編碼器的視覺語言模型)。On-policy distillation 被標記為當前熱門後訓練技術。
Notable posts:
- NVIDIA Qwen3.6 MoE NVFP4 — 35B/3B-active, memory ~3x reduction —
2026-05-29↻74 ♥903 - NVIDIA Kokoro TTS — 82M param lightweight speech synthesizer, commercial use —
2026-05-29↻48 ♥451 - Microsoft SkillOpt — text-space agent skill training, SOTA on 52/52 settings —
2026-05-25↻26 ♥160 - Self-Improving LLMs with Bidirectional Evolutionary Search —
2026-05-28↻24 ♥147 - AutoResearchClaw — 23-stage autonomous research pipeline with multi-agent debate —
2026-05-24↻20 ♥105 - Full Attention Strikes Back (RTPurbo) — 9.36x prefill speedup via retrieval head isolation —
2026-05-24↻14 ♥110 - SciAtlas — 43M papers, 157M entities, 3B triplets knowledge graph —
2026-05-25↻23 ♥98 - NVIDIA OneFormer — unified semantic/instance/panoptic segmentation —
2026-05-29↻16 ♥87 - NVIDIA LocateAnything — VL detection, 2.5x faster via Parallel Box Decoding (CVPR 2026) —
2026-05-27↻17 ♥82 - NEO-ov — VLM without image encoders, end-to-end pixel-to-word learning —
2026-05-28↻15 ♥82 - BeliefTrack — Contextual Belief Management for long-horizon LLM reasoning —
2026-05-29↻14 ♥71 - Diffusion-Adaptive Routing (DAR) — Alibaba drop-in for Diffusion Transformers —
2026-05-25↻11 ♥64 - Survey of Large Audio Language Models — trustworthiness across 6 pillars —
2026-05-24↻17 ♥68