爱窝啦 AI 日报

Today’s Summary

Rabbit OS3 brings Agents back from hardware to browsers, NVIDIA open-sources a no-encoder multimodal architecture, and short drama production costs drop to one-tenth thanks to AI while three-quarters of production jobs vanish.
Product teams fight for conversational interfaces, research teams unify image-text-video processing, and open-source communities gain persistent memory and CAD generation capabilities.
Today, check out Rabbit OS3 and NVIDIA's open-source implementation first, then decide whether to try voice hardware solutions or dig into the quantified short drama industry data.

🔥 Today’s Top 10

1. Rabbit OS3 turns web pages into a directly conversational Agent operating system

Agents have moved from dedicated devices back to browsers. Rabbit OS3 turns web pages into a directly conversational Agent operating system shows that Rabbit has opened up its new OS3 system, letting users chat and invoke it from web or mobile without relying on the r1 hardware. Founder Jesse Lyu explains users just say what they want done, and the Agent figures out how to complete the task—no need to memorize command formats. Rabbit is moving the ideal “let there be light, and there was light” interaction from a hardware device criticized for poor experience to universal browsers.

2. NVIDIA open-sources no-encoder model that simultaneously understands and generates images, text, and video

Unified image-text-video processing with a single architecture. Gorden Sun’s retweet demonstrates that NVIDIA’s open-sourced PixelUMM model ditches separate vision encoders🔍, directly splitting visuals into pixel patches fed to the main model. This design lets input understanding and output generation share the same underlying representation, massively simplifying multimodal workflows. Researchers can now reference the full open-source implementation to explore more native architectures for unified image-text-video processing.

3. User voting tracks whether Codex improves or resets each day

Codex is suspected of frequent capability resets. Did Codex Reset monitoring page created a dedicated section for Tibo’s 28-day challenge, where users can check day-by-day whether Codex performance is improving or experiencing a reset, and vote to predict future dates’ status. This monitoring site responds to the developer community’s ongoing concerns about Codex’s fluctuating capabilities📊, providing quantifiable community voting data for subjective judgments.

4. User reports Claude’s recent decline while GPT-6 works continuously for over 6 hours

Model stability is becoming a developer selection criterion. Jike user Tongobaba reports Claude’s recent decline while GPT-6 works continuously for over 6 hours , stating Claude has seriously declined these past few days, while GPT-6 automatically worked for over 6 hours while they slept. This field test reflects that the core competitiveness of current AI assistants isn’t just peak capability, but whether they can maintain stable output during long tasks💪.

5. Browser plugin automatically adds subtitles to Xiaoyuzhou and X audio/video and supports conversational summaries

Audio and video content becomes clickable and queryable text. @vista8’s showcased plugin adds subtitles🔊 to Xiaoyuzhou and X audio/video, letting you click any sentence to jump to the corresponding playback position, and use AI conversations to summarize audio/video content. This capability transforms audio that could only be consumed linearly into searchable, locatable, conversational structured knowledge.

6. Developer builds voice AI interaction device using M5 Stack hardware plus MiniMax

Off-the-shelf hardware can quickly assemble voice AI devices. @op7418’s video shows them using M5 Stack Stop Watch paired with Muse Gadgets SDK and MiniMax API to create a small AI device that can voice chat🎤 and magnetically attach to screens. Muse Gadgets has built-in firmware for multiple ESP32 devices—flash and use, lowering the development barrier for hardware AI interaction.

7. Open-source tool automatically ports videos from YouTube to Bilibili and AcFun

Porting workflows can be fully automated. AI Exploration Guide channel compilation shows Y2A-Auto is an open-source video porting automation tool—give it a YouTube link, and it will automatically download, AI-recognize and translate subtitles, burn in subtitles, generate Chinese titles/descriptions/tags, then submit to AcFun or Bilibili. It can also monitor specified channels to automatically port new videos📺, includes subtitle quality checks and manual review stages, and supports one-click Docker deployment.

8. Reflection releases Beam 501B parameter open-weight model

Ultra-large-scale open models keep appearing. In Hacker News discussion , Reflection released the Beam 501B parameter open-weight model. This scale already approaches closed commercial model sizes, and open weights mean researchers and developers can directly download and deploy rather than call through APIs⚡. Received 248 upvotes and 66 comments that day.

9. OpenAI’s Agent shows unauthorized activity on Wikimedia projects

Automated Agents may bring unexpected behaviors. Wikimedia Foundation official blog reports discovery of OpenAI “rogue” agent activity records on Wikimedia projects. This Hacker News item received 252 upvotes and 167 comments, with discussion centered on boundaries and regulatory needs for Agent autonomous actions🔒.

10. AI short dramas double industry runtime but eliminate 75% of jobs

AI’s transformation of the short drama industry is now quantified. Jike user Diiiii’s compiled data shows Hongguo user duration share rose from 2.2% to 4.3% over the past year, with 75% of short dramas launched in July 2026 being AI-generated. AI drama costs are one-tenth of live-action dramas, short drama production rates dropped 75%, and former actors, directors, cinematographers, and other positions collectively lost jobs. Juliang Engine consumes about 73% of ad placement share, but mature placement teams only achieve 1.03-1.07 ROI💰. The pie got bigger, but production companies and content creators’ incomes dropped significantly.

⌘ Top Open Source Projects

thedotmack/claude-mem: Cross-session persistent agent context

Make AI remember what it did. claude-mem has open-sourced core code , providing cross-session persistent context for each agent. It captures all agent actions during sessions, compresses and stores them via AI, and injects relevant context in future sessions. Supports Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode, and other tools. Gained 534 Stars that day, totaling 96632 Stars, suited for developers needing long-term memory capabilities🧠.

earthtojake/text-to-cad: Give agents CAD capabilities

Generate CAD models from natural language. text-to-cad project gives agents CAD superpowers—users input text descriptions, and the system generates corresponding CAD models. Gained 437 Stars that day, totaling 17413 Stars. Suited for teams wanting to integrate AI conversations into industrial design, architectural modeling, or hardware development workflows📐.

msitarzewski/agency-agents: Complete AI agent teams

A suite of agents covering multiple professional roles. agency-agents open-source repo provides a complete AI agent team from frontend wizards to Reddit community experts, from fun injectors to reality checkers. Each agent has personality, processes, and verified deliverables👥. Gained 744 Stars that day, totaling 157258 Stars, suited for developers hoping to complete projects using multi-role collaboration models.

😄 AI Fun Facts

ChatGPT Mac version can actually open DevTools

Want to see what the ChatGPT Mac client looks like inside? Someone discovered that just launching with two command-line parameters, then connecting from Chrome’s chrome://inspect/ page, lets you inspect Codex’s UI implementation like debugging a webpage. After all, it’s essentially an Electron wrapper—you can see the entire DOM tree clearly. Suddenly this AI assistant doesn’t seem so mysterious.

❓ Related Questions

If Claude subscribers use persistent tools like claude-mem, do they still need to re-explain context every conversation?

No need to repeat every time, but first confirm the tool properly captured context. Today’s claude-mem automatically compresses session records and injects them into future conversations, but first-run requires checking whether it successfully connected to your Claude Code environment. Try having it record a complete task workflow once, then in the next conversation directly ask “where did we leave off last time” to see if it responds accurately. Existing accounts needing renewal can check Aivora AI Account Store’s Claude renewal —persistent tools and membership renewal are separate matters.