12-18-Daily AI News Daily

AI News Daily 2025/12/18

AI Insights | Daily Read | Aggregated Global Data | Cutting-Edge Science Exploration | Industry Voices | Open-Source Innovation | AI & Our Future | Visit Web Version↗️ | Join Group Chat🤙

Today’s Rundown

Tencent Hunyuan World Model 1.5 launches, enabling text and image-based interactive world generation.
ByteDance Seedance achieves 100% audiovisual sync, now live on Jimeng and Doubao.
OpenAI releases FrontierScience benchmark; GPT-5.2 scores 77% in Olympiad track.
Yao Shunyu appointed Tencent's Chief AI Scientist, reporting to Martin Lau.
Nvidia acquires Slurm developer SchedMD, fortifying its computing power scheduling moat.

Product & Feature Drops

1. Tencent’s Hunyuan World Model 1.5 just launched, bringing the nation’s first 🎮 real-time interactive experience platform! Now you can conjure up interactive worlds from text or images and freely explore ’em with a keyboard, mouse, or even a gamepad (how cool is that?!). What’s more, they’ve open-sourced their entire training system for the first time, covering everything from data to inference and deployment. Talk about a full package! Experience Now (AI News)

2. Kling 2.6 Voice Control is officially here! Kuaiying AI 📢 has rolled out this awesome feature, letting you create even more captivating personalized content (AI News) using your very own voice. To celebrate, they’ve also kicked off a creative contest 🏆 with prizes up to $1000! Just submit your work for a chance to get featured on the homepage. Sweet deal, right?
   
   

3. ByteDance Seedance 1.5 Pro just dropped, and this next-gen audio-video model is bringing 🎬 100% audiovisual synchronization! We’re talking highly precise lip-sync, intonation, and performance rhythm for characters. It also supports natural expression in multiple languages and dialects, and can even pull off complex camera movements (AI News) like Hitchcock zooms. You can find it live on Jimeng AI and Doubao platforms now!
   
   

4. Meta just rolled out its SAM Audio model, extending the “🔊 segment anything” philosophy from images straight into the audio realm! Following their image segmentation success, this new model supports three prompting methods—text, visual, and temporal span—allowing you to precisely separate sounds, just like cutting out (AI News) objects from an image. Go give it a whirl on the Segment Anything Playground!
   
   

5. Xiaomi is opening up its 🤖MiMo large model series and CarIoT hardware ecosystem to developers! The AIoT platform has now connected over 1.04 billion devices, with the developer community growing to 1.2 million (AI News) strong. Plus, MiMo-V2-Flash has been open-sourced and has even ranked in the global TOP2 for open-source models in Agent evaluations. Pretty impressive, huh?
   
   

6. Meta has unveiled its new AI hearing-enhancing glasses! These new specs feature an open-speaker design that can amplify the voice of the person you’re chatting with 👓. They’re especially perfect for noisy environments (AI News) like cafes or busy streets, making everyday conversations a breeze. How cool is that for a little assist?

Cutting-Edge Research

1. OpenAI just unveiled its FrontierScience benchmark, specifically designed to evaluate expert-level scientific capabilities. It features hundreds of original problems across physics, chemistry, and biology! GPT-5.2 absolutely rocked it, scoring 77% in the Olympiad track and 🔬 25% in the research track, outperforming other cutting-edge models. Gemini 3 Pro even showed comparable performance (AI News) to GPT-5.2 in the Olympiad track. Impressive stuff!

2. The FreeKV framework is here to supercharge LLM inference efficiency! This bad boy tackles long-context KV cache issues with an algorithm-system co-optimization. By leveraging speculative retrieval and double-buffered streaming recall, it achieves 🚀 near-lossless accuracy and speeds things up by up to 13 times (AI News) compared to SOTA methods. Talk about a major boost!
   

3. Titans is giving AI a real memory! This paper, which even Google’s Jeff Dean gave a nod to, finally solves AI’s infamous “goldfish memory” problem (you know, the short attention span thing). It uses three distinct mechanisms—short-term, long-term, and persistent memory—each doing its own thing. The result? A whopping 96%+ accuracy in 2 million token ultra-long text comprehension tasks, absolutely crushing Mamba2’s 5.4% (AI News) . Mind blown! 🤯
   

Industry Outlook & Social Impact

1. It’s official! Yao Shunyu, a star scholar born in ‘95, has been appointed Tencent’s Chief AI Scientist within the “CEO/President’s Office,” reporting directly to Martin Lau. Tencent is leveling up its large model R&D architecture. Yao Shunyu will also head the AI Infra Department and the Large Language Model Department, set to 📈 fully bolster Tencent’s large model R&D system (AI News) . Big moves are happening!

2. Nvidia just quietly acquired SchedMD, the brains behind Slurm! This move is being hailed as “widening their moat” 💪. Slurm is the resource scheduling tool used by over half of the world’s TOP500 supercomputers, with giants like Meta, Mistral, and Thinking Machines relying on it. The kicker? Even if you’re rocking AMD chips, as long as you need compute power scheduling, you simply can’t bypass Nvidia (AI News) . Talk about strategic!

3. AI context management is stirring up some major privacy debates. Would you feel safe uploading all your life notes to a third-party server? While 🔥 feeding Obsidian notes to Claude can totally fetch personalized advice, community discussions show most folks are leaning towards controllable solutions (AI News) like local LLMs. Plus, some are warning that relying too much on AI summaries might actually erode our genuine grasp of knowledge. Food for thought, right?
   

4. GitHub Actions is hitting us with platform fees starting in 2026! Scheduling for private repositories and self-hosted runners will be billed at $0.002 per minute 💸. Yep, even if your compute power is on your own servers, you’ll still be paying a “tax.” Smaller teams are gonna feel this pinch harder, and the community is already scoping out Forgejo and other alternatives (AI News) like GitLab. Bummer, right?
   

5. Can AI really push formal verification into the mainstream? 🤔 The big debate centers on how tough it is to formalize specs and how often requirements shift. Optimists are pointing out that large models like Opus and GPT-5.2 🤖 have significantly sped up proof engineering. But the pessimists reckon that cultural and economic hurdles are the real obstacles to popularization (AI News) . What’s your take?
   

Top Open-Source Projects

1. Moore Threads just open-sourced its LiteGS fundamental library! This 3DGS reconstruction algorithm, which snagged a silver award 🥈 at SIGGRAPH Asia 2025, is now available. It finishes a 60-second task in just 34 seconds, achieving the same quality with only 10% of the original training time. We’re talking full-chain optimization, from GPU systems to algorithm design! The code is open on GitHub (AI News) , and it’s already caught the eye of the academic world. ⭐

2. Nvidia just dropped its Nemotron 3 open-source model! This MoE architecture supports a million-token context and comes in three sizes: Nano (30B), Super (100B), and Ultra (500B). The Nano version is already out, boasting a 🚀 4x throughput improvement over its predecessor, and it’s been hailed as the most open and efficient model (AI News) of its kind. Seriously fast!

3. Xiaomi just open-sourced its MiMo-V2-Flash model! This self-developed MoE large language model, packing 309B total parameters and 15B active ones, is specifically designed for ultimate inference efficiency. It boasts strong code and Agent capabilities 💡 with lightning-fast generation speeds. Plus, its API is free for a limited time and can connect with tools like Claude Code and Cursor (AI News) . Developers are absolutely loving it! ⭐
   
   

4. Chatterbox, an open-source TTS system, is making waves! Touted as the most advanced open-source text-to-speech system, it’s already racked up ⭐ 15,614 stars. Check out the project over at resemble-ai/chatterbox (AI News) .
   

5. Microsoft just open-sourced its TRELLIS.2 image-to-3D model! This 4B parameter model supports generating 3D models from images. An online demo is live, but community feedback is a bit mixed (shrug emoji). Some even think it’s not as good as previous versions. You can find the model released on Hugging Face (AI News) .
   
   

6. Meituan just open-sourced its LongCat virtual human model! Similar to ByteDance OmniHuman and Kuaishou Avatar, it lets you generate videos from audio-driven photos 🎤. It’s especially handy for streamers and music video scenes. You can find the project homepage and model released on Hugging Face (AI News) .
   
   

Social Media Buzz

1. A deep dive into Prompt Caching technology is making the rounds! What’s cached isn’t just text, it’s “thought states” 🧠. Essentially, it’s all about reusing the KV matrix, which can slash Token costs by roughly 90% and cut first-token latency in long texts by a whopping 85%. Real-world tests show Anthropic’s manual mode hits a 100% success rate (AI News) , while OpenAI’s automatic mode is only at 50%. Pretty wild, huh?
   
   

2. Gemini 3 Flash is now live and ready for use! Compared to the Pro version, it’s significantly faster, though the frontend visuals haven’t really changed (still looks great!). The aesthetics continue to outperform other models. ZenMux has it exclusively available for free right now, so access it here (AI News) . Go check it out!
   
   

3. Thoughts on building moats in the ‘Vibe Coding’ era are circulating. Tech isn’t the core competitive edge anymore 🤔. While it’s easy to grab a wave of traffic, building a real moat demands more thoughtful effort. Some folks are spotting flaws, while others are seeing opportunities (AI News) – and these opportunities aren’t for the nitpickers, ya know?
   

4. We’re testing out GPT Image 1.5’s image capabilities! It’s purely a drawing model, not a 🌍 world model like Banana Pro. The community’s buzzing that “Google is a generation ahead this time.” You can check out Baoyu’s test results (AI News) for city weather card generation – they’re pretty neat!
   
   

5. The AI hardware gadget Stickerbox is totally going viral! Voice input → AI automatically draws → instant sticker print 🖨️ – it helps kids turn their imagination into reality! It features a child-safe mode with no screen interaction, and this whole concept is bound to make its way into the 3D printing field (AI News) super soon. How cool is that for creative tech?
   
   

AI News Daily: Voice Edition

🎙️ Xiaoyuzhou	📹 Douyin
Reincarnation Tavern	Self-Media Account