Quartz 5

X Digest — 2026-W22

Properties1
tagsx-digest, research, robotics

6 min read

X Weekly Digest — Week 22, 2026

Fetched: 2026-05-31 | Accounts: 3 | Posts: 581

Cross-Account Highlights

  1. NVIDIA Qwen3.6 MoE (quantized) — NVFP4-quantized 35B/3B-active model released on HuggingFace; memory reduced ~3x with near-zero accuracy loss (source)
  2. 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)
  3. 500-hour AI Infrastructure Curriculum — Viral open-source curriculum for AI engineering (1,294♥), covering infra, MLOps, and deployment (source)
  4. book-to-skill — Tool that converts technical books into Claude Code skill files for on-demand reference; highly discussed in Chinese dev community (source)
  5. 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:

@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:

@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:

Graph View