爱窝啦 AI 日报

Today’s Summary

OpenAI open-sources 722 math papers and Haiku 5.5 slashes prices by 75%, lowering barriers on both model costs and data sources.
Multiple products are pushing outputs beyond text—GPT-6 dynamic interfaces, Kling's independent listing, Grokbot's multi-model orchestration all point toward more complete deliverables.
Check today's price shifts and Octop's local solution first, then decide whether to swap existing models or deploy offline assistants.

🔥 Today’s Top 10 Focus

1. OpenAI Open-Sources 722 Math Papers in One Go

Math research drops in bulk. Machine Heart reports that OpenAI published 722 math-related materials , covering cutting-edge topics like quasi-Riemann conjectures and 4D Kakutani. This batch release saw rapid GitHub Star growth. Math researchers and AI trainers now have fresh data sources to tap into.

2. Grokbot Switches to Multi-Model Orchestration

Subscriptions now tap more models. Guizang relays Musk’s announcement that Grokbot routes tasks to the best-fit model , not just Grok family members. It’ll also call Opus 5.5, Midjourney, Suno APIs to build content or run tasks🎨. Some features are live—certain tasks already run on Opus 5.5—but users can’t manually pick models yet. Image generation and music creation are pending deployment. Subscribers get one entry point to multi-vendor outputs.

3. Kuaishou Preps Kling for Independent IPO

Standalone listing enters prep stage. 36Kr reports that Kuaishou has picked investment banks to prepare Kling for a Hong Kong listing , targeting earliest 2027, raising at least $1 billion💰. Kling’s Q2 revenue exceeded 850 million yuan, up over 200% YoY, far outpacing the parent’s 1.4%. A spinoff lets capital invest directly in video generation. But it also means Kuaishou’s parent needs fresh reasons to hold investor attention.

4. Tencent Open-Sources Local Multi-Agent Platform

Team assistants can run fully offline. Gorden Sun introduces Octop—single-process Web console, desktop client, scheduled tasks , integrating WeChat, Feishu, DingTalk. Supports multi-expert collaboration with permission-gated CLI and browser automation🔒. All conversations, knowledge bases, and memory stay on local devices. Privacy-conscious teams can pilot automation assistants with this setup.

5. Backend Copy Rewritten Into Product-Speak Snippets

AI rewrites menus so users grasp them instantly. Xiangyang Qiaomu tested a prompt that turns backend jargon into product-style short phrases . Requirements: 2–4 characters for names, under 20 for descriptions📝, swap jargon for everyday language. Test output shows before-and-after comparison tables. Developers can use this to make settings screens friendlier.

6. Records of the Grand Historian Turned Into Searchable Knowledge Graph

577,000 characters of ancient text become a relationship network. AI Exploration Guide introduces the open-source project shiji-kb, transforming Records of the Grand Historian into a searchable knowledge graph . Click a figure and drill down through events, places, dates, original passages🗺️. The project also includes a Grand Historian subway map, drawing each chapter as a line and events as stations. Readers can trace timelines without mental gymnastics.

7. Chrome 155 to Support JPEG XL Decoding

Image sizes shrink another 30–50%. The Chrome team announced that starting Chrome 155, JPEG XL decoding support arrives . The format handles lossless compression, HDR, lossless JPEG transcoding🖼️, and fine-grained progressive decoding. Implementation uses the pure Rust decoder jxl-rs; fuzz testing and AI code reviews found no memory-safety bugs. The team suggests trying both AVIF and JPEG XL—not every image beats AVIF. This is browser decoding capability, not universal app or OS support yet.

8. Google Releases Cross-Modal Embedding Model

Unified space maps text, code, images, audio, and video. Google’s blog shows that EmbeddingGemma 2 has 740 million total parameters , loads modules on demand, targets on-device semantic search🔍, cross-modal retrieval, and RAG. The model handles 8K token contexts—up to roughly 5.5 minutes of audio, 29 images, or 58 video frames. Through Matryoshka representation learning, output vectors can truncate to 512, 256, or 128 dimensions, saving up to 6x in vector storage. Google reports that on Pixel 11 Pro quantized configs, text-only and full-modal active memory can run as low as roughly 191 MB and 567 MB respectively.

9. Community Roasts Opus 5.5 vs. Competitors

Sarcastic post hints at model-choice shifts. A LINUX DO user posted “Opus 5.5 sucks, don’t migrate to Claude,” then tagged it as a Reddit joke . The post drags OpenAI founder Sam and reset-record holder tibo into the comparison⚖️. This kind of ironic banter reflects some developers reassessing model-switching costs and usability.

10. Ted Chiang Explains Why AI Text Feels Off

Number of choices made defines a text’s soul. My Brother Tiema relays that Ted Chiang’s 2024 New Yorker piece answers this puzzle . When you write a 10,000-word story yourself, you’re essentially making 10,000 choices✍️. But when you prompt AI to generate 10,000 words, you might only make 100 to 1,000 word-level choices—AI handles the other 9,000. It’s this effort of making loads of choices that gives text its soul. This also explains why the more precise and effortful your prompt, the closer the output lands to your expectations.

⌘ Top Open-Source Projects

manaflow-ai/cmux: macOS Terminal Supporting AI Coding Agent Notifications

manaflow-ai/cmux open-sourced its core code , built on Ghostty with vertical tabs and AI coding agent notifications🔔. Gained 44 Stars today, reaching 27,837 total. Built for multitasking, organization, and programmability. Suits developers who want to receive AI agent status updates in the terminal.

◉ Social Media Picks

GPT-6 Answers Turn Into Interactive Smart Interfaces

Conversation outputs are no longer static text. Baoyu relays that OpenAI is rolling out GPT-6 in ChatGPT , headlining a feature called Intelligent UI. Plus, Pro, Business, and Enterprise users get access starting today. With this feature, GPT-6 decides how to shape its answer based on your question: comparison questions get side-by-side layouts🖼️, principle explanations come with drag-and-drop parameter diagrams. Official example: you’re hosting friends for Sunday roast lamb—GPT-6 delivers a shopping list with a guest-count slider. Bump the number from 5 to 8, and the lamb and potato quantities adjust on the fly.

Claude Haiku 5.5 Cuts API Costs by 75%

Small-model pricing keeps dropping. Anthropic’s official account announces that Claude Haiku 5.5 is the cheapest, fastest, most capable small model yet . Average run cost drops roughly 75% compared to Claude Haiku 4.5💰. Developers can swap this version into existing small-model calls to slash inference costs.

SynthID Detector Opens to Everyone

Detection tool spots content from multiple vendors. Google DeepMind announces that SynthID Detector is now available to all 🌐. The tool checks whether online content was generated by Google AI or partners like OpenAI, NVIDIA, Kakao—Apple joining soon. Users can try it at synthid.com.

😄 AI Fun Facts

ChatGPT Can Now Draw Interfaces for You

OpenAI just launched GPT-6 and Intelligent UI, letting ChatGPT generate interactive custom interfaces to answer questions. Ask about the weather and it won’t just tell you the temp—it’ll draw a live-updating weather card; ask for a recipe and you get a timer plus step-by-step checklist. In the demo video Sam Altman posted, the chat window suddenly sprouted sliders, buttons, and charts—the chat box turned into an app store.