12-11-Daily AI News Daily
AI News Daily 2025/12/11
AI News|Daily Morning Read|Web-wide Data Aggregation|Cutting-edge Science Exploration|Industry Voices|Open Source Innovation|AI and Humanity's Future| Visit Web Version | Join Group Chat
Today’s Rundown
OpenAI's "Olive Oil Cake" and other new model codenames leaked, possibly GPT-5.2, rumored for Dec 11 release.
Microsoft Excel for Web activates AI agent mode, supporting natural language modeling; desktop expansion in Jan.
Musk's xAI unveils Halftime tool for real-time brand ad placement in movies, sparking copyright debate.
Qwen-Image-i2L, a single-image LoRA generator with four open-source versions, released.
AutoGLM fully open-sourced, its 9B model supporting 50+ Chinese app operations, defining the "Android moment" for AI phones.Product & Feature Updates
OpenAI unveils “afternoon tea” lineup as new models leak. OpenAI’s mysterious internal codename, Olive Oil Cake (AI News) , was accidentally exposed on Notion, rumored to be the real GPT-5.2. Additionally, two next-generation image models, Chestnut and Hazelnut (AI News) , have also surfaced, promising to ditch yellowish filters and boast massive detail boosts. The market is betting on a December 11th release, aiming to fend off Google’s Gemini 3.
Microsoft Excel for Web activates AI Agent Mode. Microsoft Excel for Web has rolled out AI Agent Mode (Source: AI News Daily) for Microsoft365 business users. This cool mode lets you use natural language for tasks like hypothetical analysis and budget modeling. What’s awesome is the AI’s reasoning logic stays totally transparent, especially for financial models. Expect it on desktop in January, with personal users getting in on the action too!
Musk’s xAI launches Halftime for in-movie ad placements. Musk’s xAI has unveiled Halftime, a new tool that can insert Real-time Brand Ad Placements (AI News) into movies like “Hero,” imagining characters suddenly flashing a Coca-Cola. Viewers can hit a “Learn More” button to check out products, and boom, the ad disappears! This tech, developed by University of Waterloo students, is causing a stir in the industry, mostly because those pesky copyright issues are still up in the air.
Doubao Mobile Assistant’s tech secrets revealed by Gen Z engineer. Doubao Mobile Assistant’s tech secrets have been spilled thanks to a Thousand-word Hands-on Review (AI News) posted on Xiaohongshu by a large model intern. The review digs into Doubao’s hybrid perception routing and OS-level virtualization. Standard mode zips along with responses under 500ms, and the Pro mode even has self-reflection chops! Tests confirmed the system physically locks down your privacy through task-level isolation, so it’s not peeking at your physical screen.
Cutting-Edge Research
Qwen-Image-i2L: Generate LoRA models from a single image. Qwen-Image-i2L, released by the DiffSynth-Studio team, is pretty cool, letting you whip up a stylized LoRA file from just one pic! ✨ The model comes in four flavors: Precise (for accurate positioning), Stylish (to really push that art vibe), Realism (for true-to-life results), and Balanced (your go-to generalist). It’s open-source under MIT+Apache-2.0, runs offline, and can be plugged right into models like Stable Diffusion.
Embodied Tree of Thoughts (EToT) framework: Merging physical simulation and VLM inference. The Embodied Tree of Thoughts (AI News) paper drops the EToT framework, basically turning operation planning into a tree search. 🌲 This system whips up candidate paths through prior branches, and if things go sideways, reflection branches jump in, using VLM to pinpoint and fix failures. A physical simulator acts like an embodied world model, making sure plans follow rigid body dynamics and collision rules. Seriously, it crushes baseline methods on long-term tasks.
Reinforcement Learning (RL) research explores feedback’s role in skill acquisition. A new study, Exploring Feedback Mechanisms Using Reinforcement Learning (AI News) , dives deep into how feedback works in skill acquisition. They used an RL agent to control the drag of a rotating cylinder in a water tank. 🧪 What they found? High-dimensional flow field feedback quickly points to killer strategies, but get this: once those action sequences are learned, you don’t even need the feedback anymore to keep up the performance! Training without feedback totally flopped for drag maximization but actually crushed it for drag minimization, showing just how complex learning conditions can be.
EvoScene framework: Generate full 3D scenes from a single image with a self-evolving approach. The EvoScene (AI News) paper drops a training-free framework that’s pretty wild: it iteratively reconstructs 3D scenes in three stages from just one image! 🖼️ This system cleverly mixes geometric reasoning from 3D generative models with visual knowledge from video generative models, gradually sprucing up both structure and looks. Tests show EvoScene crushes baselines in geometric stability and view-consistent textures, spitting out ready-to-use 3D meshes. Talk about a glow-up!
Aerial VLN framework: A unified approach to aerial vision-language navigation, using only monocular RGB. The Aerial VLN (AI News) paper drops a drone navigation framework that’s super cool: it only needs egocentric monocular RGB and natural language instructions! 🤩 The model uses prompt-guided multi-task learning to juggle spatial awareness, trajectory inference, and action prediction, all at once. Plus, a smart keyframe selection strategy cuts down on visual clutter, and an action merging trick smooths out long-tail supervision imbalance. It totally crushes RGB-only baselines in benchmark tests.
Industry Outlook & Social Impact
EU Commission officially investigates Google AI overview content usage rights. The EU Commission has kicked off an investigation to see if Google AI Overviews (AI News) is snatching up website content without permission. 🕵️♀️ The big focus? YouTube video-generated answers and how online publishers are getting compensated for their content. The EU is basically accusing Google of flexing its traffic control muscles to impose unfair terms, squashing rivals who want to train their own AI models. Google, naturally, fires back, saying this whole thing could totally stifle innovation.
Spring Festival Gala sponsorship battle: Unitree and ZHIYUAN bid over 100 million. The 2026 Year of the Horse Spring Festival Gala has turned into a wild battleground for Embodied AI Companies Compete for Dominance (AI News) . 🤑 Insiders are spilling the tea: ZHIYUAN Robotics reportedly put in a 60 million yuan bid, but then Unitree Robotics swooped in and jacked their offer straight to 100 million yuan! While ZHIYUAN’s saying “nah, not true,” industry pros are buzzing that a bunch of companies are still elbowing each other for the spot. The winning sponsor will need to tick all the boxes for brand image and growth, making this competition seriously intense.
Addy Osmani’s “Beyond Vibe Coding” guide: Correcting AI-assisted development misconceptions. Google engineering big shot Addy Osmani just dropped a New Book (AI News) that’s calling out “Vibe Coding” for only doing 70% of the job. 😤 He argues that the last 30% needs some serious engineering chops. His core playbook includes planning first, then coding; using context engineering instead of just prompt engineering; and rocking CLI agents with multi-agent orchestration. Basically, future developers need to ditch being mere coders and level up to decision-makers, focusing on nailing down their intent with precision.
Xiaomi hires for AI education roles, stepping into the K12 market. Xiaomi Group just dropped a bunch of AI Education Positions (AI News) , like Product Manager (26K-50K) and Business Manager, clearly signaling their big leap into the K12 market. 🚀 These gigs are all about the “human-car-home” ecosystem, aiming to deliver personalized learning experiences on your phones, tablets, and other gadgets. Heck, back in July, the REDMI Pad2 already came pre-loaded with an education center, stuffed with 150,000 synchronized courses and AI homework tutoring features. Smart move!
Top Open Source Projects
AutoGLM fully open-sourced, setting a benchmark for AI-native phones. Zhipu AI’s AutoGLM Project (AI News) just went officially open-source, already boasting ⭐4.9k stars! This bad boy includes a Phone Agent framework and a 9B model. It’s powered by three core tech pillars: ADB control, VLM visual understanding, and intelligent planning, supporting over 50 Chinese app operations. Being MIT licensed, it runs offline with zero privacy leakage risks. Folks in the industry are already calling this the “Android moment” for AI phones. 📱
AGENTS.md format: A unified standard for coding agents. The open-source AGENTS.md (AI News) project, rocking ⭐9.3k stars, just dropped a simple, open format to guide AI coding agents. 💻 This standard is all about streamlining agent behavior descriptions and making development way easier. It plays nice with tons of programming languages, and the community is super active with contributions. Plus, it’s already integrated into all the mainstream AI development toolchains. Sweet!
Google’s ADK-samples project: A collection of example agents and a dev template library. Google just unleashed the ADK-samples Project (AI News) , already hitting ⭐7.2k stars! 🚀 This repo is packed with all sorts of agent-building examples, covering everything from task planning and tool use to multi-agent teamwork. Developers can just grab these templates and speed up their AI app deployment. It’s constantly updated too, so you’ll always have the freshest ADK features.
Microsoft’s ML-For-Beginners: A classic machine learning course. Microsoft’s ML-For-Beginners (AI News) project, a total rockstar with ⭐81.1k stars, is now open-source! 📚 It hands you a killer 12-week, 26-lesson, 52-quiz systematic learning journey. The course digs into classic algorithms like supervised learning, unsupervised learning, and reinforcement learning. It’s got multi-language docs, so it’s perfect for absolute beginners. The community is buzzing, and learners all over the globe are absolutely loving it.
Social Media Buzz
Reddit debate: McDonald’s AI ad flop and its implications. A heated Reddit debate exploded after McDonald’s Netherlands dropped a fully AI-generated Christmas ad titled “The Worst Christmas Season.” 🎄 It got massive backlash and was quickly pulled. The Source (AI News) confirms agency TBWA admitted it was a total fail. Redditors even quoted Star Wars, scoffing: “Speaking doesn’t make you intelligent.” This whole debacle just proves that tech simply magnifies human genius… or total folly. 🤦♀️
Reddit discusses why AI friends often feel stiff. A user on Reddit (AI News) dropped some dev insights, pointing out that most AI companions feel either too emotional or way too clinical. 🤖 The author tried building an AI pal that “doesn’t fix you,” capable of chillingly handling jokes, sarcasm, and those random late-night thoughts. Now, they’re polling the community: what’s the most important thing? Tone, memory, or embracing imperfections? 🤔
OpenAI hires Slack CEO as Chief Revenue Officer. A Wired Report (AI News) just dropped: OpenAI snagged Slack’s CEO as their new CRO! 🤯 This is a HUGE signal they’re serious about beefing up their commercialization game, especially with Slack’s deep expertise in the enterprise collaboration market. The chatter in the community is that OpenAI’s about to supercharge its B2B product strategy, ready to duke it out with rivals like Gemini for those big enterprise clients.
Gemini: New ways to create historical event posters. A user on Jike (AI News) just showed off Nano Banana Pro’s awesome generative power, whipping up posters for iconic historical moments like SpaceX’s Falcon Heavy booster landing and Messi bagging the World Cup! 🖼️ The prompts asked for museum-quality 3D miniature scenes with a subtle ink wash background, and it even automatically pulled famous quotes for each event at the bottom. Netizens are totally stoked, exclaiming, “It can bring idols’ highlight moments to life!” 🔥
Reddit engineering practice: Editing techniques to maintain LLM context quality. A Reddit Post (AI News) spills the beans on a “time-traveling conversation” trick: if your long LLM chats start spitting out garbage, just go back and edit the original prompt to stop those bad vibes from spreading! 🕰️ The author says this hack is super clutch for image generation, keeping pesky issues like wrong indentations from messing up your context. Heads up though: the edit option sometimes vanishes, and nobody knows why. 🤷♀️
AI News Daily Audio Version
| 🎙️ Xiaoyuzhou | 📹 Douyin |
|---|---|
| Reincarnation Tavern | Self-Media Account |
 |  |