爱窝啦 AI 日报
每日 AI 行业动态、产品更新、前沿研究、开源项目与社媒信号
Today’s Roundup
Pulley shuts down at year-end after raising $50M, Muse can log into any site with your credentials and save it as a connector, and RedFruit mini-dramas hit 168M DAU—surpassing long-form video platforms combined.
Equity management tools are being redefined by AI, production costs for mini-dramas and barriers to agent connectors are dropping in parallel, and the common thread in both products and open source is reducing complexity to the bare minimum.
Today we'll test Muse's login capability and Raycast's infographic generation, then examine the window of SaaS replacement behind Pulley's shutdown.🔥 Today’s Top 10
1. Pulley Announces Shutdown After Raising $50M
Voluntarily closing after 7 years. According to 36Kr, equity management platform Pulley will cease all operations by year-end , with founder Yin Wu confirming the decision on social media. The company was valued at $250M💰 and raised $50M total. The timing collides with the window where AI tools are aggressively replacing traditional SaaS, and equity management tools are being redefined.
2. Muse Can Log Into Any Site with Your Credentials
Connector barrier drops to zero. In a tweet, Xiang Yang Qiao Mu demonstrates that Muse can receive a website URL, username, and password, then log in and operate on your behalf , saving the operation as a connector🔌. Already tested logging into Z-Library to download ebooks. Steam uses tokens, DeepSeek uses API keys, Bilibili uses QR codes—the model never sees the raw credentials. This means you can have an agent operate any service with login state.

3. RedFruit Mini-Dramas Hit 168M DAU, Exceeding Long-Form Platforms Combined
Mini-drama users break 100M. According to 36Kr’s summary, Nomura’s monthly report shows RedFruit mini-dramas reached 168M DAU , up 107% YoY📈. This scale already exceeds the combined DAU of iQiyi, Tencent Video, and Youku. China’s broadcasting regulator reports total micro-drama users have surpassed 100M. AI-generated mini-drama costs and production cycles are being redefined.
4. ChatGPT Moves Plugins and Library to Left Sidebar
Major navigation overhaul. Gui Cang notes in an Okjike post that ChatGPT moved plugins and library to the left sidebar 📂. New library and image pages added, with the library consolidating all images, documents, and AI-generated content. Previously everything was in the top bar—you’d have to scroll back with lots of projects. Some users find it unfamiliar, but long-term it’s better for managing and reusing assets.

5. Xiang Yang Qiao Mu Uses Raycast AI to Generate Infographics from Selected Text
One-click share images. Xiang Yang Qiao Mu demonstrates in a tweet that Raycast AI can generate Chinese infographics based on selected text 🖼️. The prompt demands design standards of an excellent Chinese magazine art director, ensuring accurate content, clear expression, and appropriate style. The product topped Product Hunt today. Can generate two different styles at once, easily readable on mobile without zooming.

6. Oscar-Winning Editor Says Emotion is 51%, Raw-to-Final Ratio 95:1
Emotion matters more than story. Xiang Yang Qiao Mu mentions in a tweet that Walter Murch’s book on editing Apocalypse Now emphasizes emotion at 51% 🎬. Murch won three Oscars with a raw-to-final ratio of 95:1. The book notes that people blink when clarifying a thought—blinking is the brain’s “cut point.” This ratio aligns closely with recent data from Han Qing Yuri’s concert.

7. Magpie New Version Supports Semantic Routing and Manual Grouping
Smart routing capability launched. Baoyu forwards yetone’s intro that Magpie’s new version supports semantic routing and manual grouping 🔀. Can automatically identify intent, convenient for smart routing. For scenarios requiring automatic model selection based on question type, this feature reduces manual configuration. Official docs updated with usage instructions.
8. Records of the Grand Historian Knowledge Base Extracts 14K Entities with AI
Records becomes a clickable graph. Telegram channel recommends the open-source project shiji-kb, which uses AI to restructure 577K characters from Records of the Grand Historian 📜. Currently organized 14K+ entities, 3,198 historical events, 7,600+ event relationships, and 130 timelines. Click a person and drill down through person→event→location→time→original text. Browse Records like Wikipedia.
9. Tech Stocks Now 39.4% of S&P 500 Market Cap, Exceeding 2000 Bubble
Bet reaches historical extremes. Telegram channel cites Reuters stats showing information technology stocks now represent 39.4% of S&P 500 market cap 📊, exceeding the 35% peak of the 2000 internet bubble. Including Amazon, Alphabet, and Meta in a broad tech definition, related companies exceed 50% of the entire S&P 500. If compute demand falls short in coming years or AI revenue can’t cover massive CapEx, the entire U.S. market could be dragged into a bear market.
10. Muse Connector Design Reduces Complexity to Minimum
Works with adapted and unadapted connectors. Gui Cang notes in an Okjike post that Muse’s connector and context acquisition system design is very complete . Adapted connectors work, unadapted ones work too. Steam gives an option and button to fill token, DeepSeek has an input box for keys, Bilibili directly provides QR code🔐. Dissolves complexity beautifully—can already read Bilibili viewing history and Steam games.

⚡ Product & Feature Updates
ABB Stock Up 50% in a Year, Powering AI Data Centers
AI infrastructure needs power too. Telegram channel mentions that Swedish company ABB’s stock rose over 50% in the AI infrastructure wave ⚡. About one-quarter of global data centers use ABB technology, and the company is collaborating with NVIDIA to develop next-gen power architecture for gigawatt-scale AI data centers. AI’s endpoint isn’t just compute—it’s power.
⌘ Top Open Source Projects
zhaoxuya520/reverse-skill: Reverse Engineering Skill Routing Package
AI-driven routing + on-demand toolchain. reverse-skill gained 361 stars today, total 38,008 . The project provides AI routing🔒, on-demand toolchain bootstrapping, and self-evolving knowledge bases for reverse engineering, authorized penetration testing, and security research. Supports Claude Code, Kiro, Cursor, Cline, and other AI coding clients. Suitable for security researchers and reverse engineers.
mobile-next/mobile-mcp: Mobile Device Automation Protocol Server
Covers iOS, Android, and simulators. mobile-mcp gained 168 stars today, total 7,348 📱. The project implements a Model Context Protocol server for mobile automation and scraping, supporting iOS, Android, simulators, emulators, and real devices. Suitable for developers building mobile automation or testing tools.
◉ Social Media Highlights
Xiang Yang Qiao Mu Develops Features Based on Obsidian and Open Sources Them
Consumed 20x Codex resets. Xiang Yang Qiao Mu tweets that he redeveloped all the features he needed based on Obsidian and open sourced them . Consumed 20x Codex with 3 resets, received 100+ yuan in donations💰. Though not breaking even, very happy. This experience reflects the actual cost of developing tools with AI and the motivation for open-source sharing.

Simon Lee Uses Codex to Create Van Gogh-Style Scene Generator
Moving Blender nodes into the browser. Baoyu forwards Simon Lee’s preview, where he worked intermittently for a month with Codex 🎨, moving previously built Blender geometry nodes and oil painting materials into the browser. Primary model is GPT-Astra High. He discovered the right approach isn’t directly translating old workflows but having Astra research existing Three.js open-source projects first, then extrapolate. Project goal is to parametrically generate every element, building oil painting scenes like LEGO blocks.
Baoyu Has Codex Create “What is DINOv3” Educational Video
One prompt generates educational content. Baoyu shares in a tweet that he had Codex create a “What is DINOv3” video using JS 🎬, requiring depth and accessibility for high school students. Prompt allows internet search and tool installation. This is his second educational video after “What is Transformer,” both completed with Claude Code + Opus 5.5.
😄 AI Fun Facts
Master Cang’s CodePilot Library Launched Months Before Codex
Codex’s major overhaul moved the library to the left sidebar, and Gui Cang deadpan notes that CodePilot added a library months ago 📂. Can consolidate AI-generated images, videos, web pages, audio, and jump back to the chat from results to see specific operations. Many still find it unfamiliar and less convenient. Major overhauls always have this effect, but long-term it’s definitely easier to find things. Master Cang got ahead of Codex by one version—turned in the exam early.