12-26-Daily AI News Daily
## Aivora AI Daily 2025/12/26
> AI Daily
### **Today's Digest**
NVIDIA's pulled a slick move, poaching Groq's founding team – they want the talent, not the whole company, a page right out of Google's playbook! ByteDance's TRAE just smashed 1.6 million monthly active users, while Tsinghua open-sourced a video generation tool that's 200x faster. The AI toolkit is seriously speeding up. What's really buzzing today? Talent mobility and a full-blown efficiency revolution. Developers, you gotta check out these new open-source acceleration frameworks!
## 🚀 Quick Navigation
- [📰 Today's AI News](#今日ai资讯) - A quick glance at the latest happenings
> ✨ **Pro Tip**: Wanna be among the first to try out the latest AI models mentioned here (like Claude 4.5, GPT, Gemini 3 Pro)? No account? No problem! Head over to [**Aivora**](https://aivora.cn?utm_source=daily_news&utm_medium=mid_ad&utm_campaign=content) to grab an account. You'll be up and running in a minute, with hassle-free support.
## **Today's AI News**
### **👀 The One-Liner**
NVIDIA just snatched Groq's founding team, pulling a classic move: they want the brains, not the whole company. Talk about learning fast!
### **🔑 3 Keywords**
#TalentHarvest #AIProgrammingTakesOff #VideoGenAcceleration
---
## **🔥 Top 10 Heavy Hitters**
**1. NVIDIA "Acquires" Groq Founding Team, Wants People Not Company**
Remember Groq, the company boasting the "world's fastest inference chip"? Well, the news just dropped: **Groq founder Jonathan Ross and the core team are joining NVIDIA**. Plus, Groq's inference tech is being non-exclusively licensed to Jensen. This move feels super familiar – Google pulled the same stunt with Character.AI. GroqCloud will keep chugging along, but its heart and soul have a new owner. Jensen's play here is genius: he snagged the tech, harvested the talent, and didn't even have to shoulder the entire company's baggage. Talk about a win-win!
---
**2. TRAE Monthly Active Users Exceed 1.6 Million, AI Programming Ecosystem Data Revealed for the First Time**
ByteDance's AI programming tool, TRAE, just dropped its annual report, and the numbers are seriously impressive: **6 million total users, with 1.6 million monthly active users**. This might just make it the top AI programming tool in China. What's really interesting is the user behavior data – **40% use it for bug fixing, 30% for code generation**, while features like repository management and code optimization still haven't quite won everyone's trust. As for programming languages, Vue takes the lead, with Python and JS hot on its heels. Looks like front-end developers are the main force driving AI programming!
---
**3. Tsinghua University Teams Up with Shengshu Technology to Open Source TurboDiffusion, Accelerating Video Generation by 200x**
Remember how generating a 5-second, 720p video with a Diffusion model used to take a whopping 4767 seconds (over an hour)? Well, guess what? Now it's **done in 24 seconds, or even under 2 seconds at its fastest**! The Tsinghua TSAIL team and Shengshu Technology's open-sourced TurboDiffusion framework made this happen. Their secret sauce? A three-pronged approach: hybrid attention acceleration, efficient step distillation, and W8A8 quantization. And here's the kicker – **almost no loss in image quality**! This means even regular consumer-grade GPUs can now dabble in real-time video generation. AI video is officially moving from "offline waiting" to "real-time preview" era. Mind blown! 🤯
---
**4. OpenAI Plans to Raise $100 Billion, Valuation to Hit $830 Billion**
OpenAI is gearing up for another massive move. Reports say they're eyeing **up to $100 billion in funding**, which, if successful, would push their valuation to a staggering $830 billion. What does that even mean? It's more than a third of NVIDIA's market cap! While it's still early days and investor demand isn't set in stone, one thing's crystal clear: the AI arms race is burning through cash faster than most folks can even imagine. 💰
---
**5. ChatGPT to Launch "Skills" Feature in Early 2026**
OpenAI's new feature, internally codenamed "Hazelnut," has just been revealed! **Skills are modular capability units**, organized like folders, containing instructions, context, examples, and even executable code. The coolest part? Skills can be automatically recognized, dynamically loaded, and combined by AI. Plus, you can even convert existing GPTs into Skills with a single click. This means you'll soon be able to assemble AI capabilities like LEGO bricks, instead of training from scratch every time. Expect it to drop in January 2026.
---
**6. Liquid AI Releases 2.6B Parameter Model, Outperforming Billion-Parameter Competitors**
Small models are winning again! Liquid AI's LFM2-2.6B-Exp, with just 2.6 billion parameters, has **outperformed billion-parameter models** in multiple benchmarks, especially in instruction following. What's even more crucial? It was trained purely with reinforcement learning – no supervised fine-tuning, no large model distillation. It's fully open-source and supports edge device deployment. So, in the future, running a small model on your phone might just be even better than a massive cloud-based one.
---
**7. Kuaikan Manhua Teams Up with MiniMax, AI Interactive Comics Are Here**
Comic characters can finally "come alive"! Kuaikan Manhua and MiniMax have teamed up to launch AI companion interactive comics. They're using **AI tech to enable multi-modal interaction and emotional connection with comic characters**. You can chat and interact with your favorite characters, not just passively read their stories. This is being hailed as the "fourth track of AI experience" in the comic industry. For anime/manga fans, this might just be one of the most exciting product forms to drop this year! 🤩
---
**8. Huawei Noah's Ark Lab Releases SCOPE, Enabling Agents to Evolve from Errors**
What's the biggest headache with AI Agents? They **see error messages but keep making the same mistakes** next time. Huawei Noah's Ark Lab, in collaboration with CUHK, just dropped the SCOPE framework to fix this! It enables an Agent's Prompt to self-evolve, learning from execution trajectories. In tests on the HLE benchmark, task success rates shot up from 14.23% to 38.64% – more than doubling! This is a major breakthrough for building truly reliable AI Agents.
---
**9. Xiaohongshu Open Sources InstanceAssemble, Breaking Through Accuracy in Complex Multi-Instance Image Generation**
One of the biggest headaches in e-commerce? **Getting AI to precisely place multiple products in a single image**. Xiaohongshu's open-sourced InstanceAssemble framework tackles this head-on. It significantly boosts spatial alignment accuracy in complex scenarios using cascaded modeling and an Assemble-Attention mechanism. The best part? It only adds 0.84% to parameters and is compatible with mainstream base models. For anyone doing e-commerce design or poster generation, this is definitely a tool to keep an eye on!
---
**10. OpenAI Releases Holiday Custom Codex Model: GPT-5.2-Codex-XMas**
Okay, this one's purely an Easter egg! The OpenAI Codex team just dropped a **special Christmas edition model, GPT-5.2-Codex-XMas**. Its performance is identical to the regular version, but it comes with a "Santa Claus personality." When you call it via the command line, it'll chat with you about code in Santa's jolly tone. While it doesn't have much practical use, this little bit of humor is pretty adorable! 🎅
---
## **📌 Worth Noting**
**[Products]**
- [Lingguang AI Launched for a Month, Users Created 12 Million Flash Apps](https://www.aibase.com/zh/news/24060) - Quickly create mini-apps without coding; a fresh take on AI-native apps.
- [Lenovo to Unveil "Super Intelligent Agent" at CES, Integrating PC and Mobile Ecosystems](https://www.aibase.com/zh/news/24060) - Taking on Doubao, Lenovo aims to create a cross-device AI assistant.
**[Research]**
- [AI-Generated Faces Can Be Detected with 5 Minutes of Training](https://www.aibase.com/zh/news/24060) - Teeth, hairlines, and ear symmetry are the telltale signs.
- [MeViSv2 Multimodal Video Understanding Dataset Released](https://www.bestblogs.dev/en/article/3719b2da) - Now includes 150,000 seconds of audio data, pushing the limits of AI logical reasoning.
**[Open Source]**
- [LEANN: RAG for Everything, Saving 97% Storage Space](https://github.com/yichuan-w/LEANN) - Run fast, accurate, and private RAG on your personal device.
- [MediaCrawler: Xiaohongshu/Douyin/Bilibili Crawler Tool](https://github.com/NanmiCoder/MediaCrawler) - With 40k+ stars, it's a must-have for content analysis.
**[Misc]**
- [Epic is Giving Away "Disco Elysium: The Final Cut" for Free Today](https://x.com/op7418/status/2004398113826710001) - A masterpiece, seriously, go grab it!
---
## **❓ Related Questions**
### How to Experience ChatGPT's Latest Features (e.g., Upcoming Skills)?
ChatGPT's advanced features (like GPT-4o, upcoming Skills, etc.) typically require a ChatGPT Plus subscription. For users in mainland China, this often means facing payment hurdles or account registration restrictions.
**Solutions:**
- **[Aivora](https://aivora.cn)** offers ready-made ChatGPT Plus accounts.
- Super fast delivery: order, use immediately, no payment or registration hassles.
- Get a stable, exclusive account with worry-free after-sales support.
Visit [aivora.cn](https://aivora.cn) to check out the full list of AI account services.
```## Aivora AI Daily 2025/12/26
> AI Daily
### **Today's Digest**NVIDIA’s pulled a slick move, poaching Groq’s founding team – they want the talent, not the whole company, a page right out of Google’s playbook! ByteDance’s TRAE just smashed 1.6 million monthly active users, while Tsinghua open-sourced a video generation tool that’s 200x faster. The AI toolkit is seriously speeding up. What’s really buzzing today? Talent mobility and a full-blown efficiency revolution. Developers, you gotta check out these new open-source acceleration frameworks!
## 🚀 Quick Navigation
- [📰 Today's AI News](#今日ai资讯) - A quick glance at the latest happenings
> ✨ **Pro Tip**: Wanna be among the first to try out the latest AI models mentioned here (like Claude 4.5, GPT, Gemini 3 Pro)? No account? No problem! Head over to [**Aivora**](https://aivora.cn?utm_source=daily_news&utm_medium=mid_ad&utm_campaign=content) to grab an account. You'll be up and running in a minute, with hassle-free support.
## **Today's AI News**
### **👀 The One-Liner**
NVIDIA just snatched Groq's founding team, pulling a classic move: they want the brains, not the whole company. Talk about learning fast!
### **🔑 3 Keywords**
#TalentHarvest #AIProgrammingTakesOff #VideoGenAcceleration
---
## **🔥 Top 10 Heavy Hitters**
### 1. [NVIDIA "Acquires" Groq Founding Team, Wants People Not Company](https://x.com/op7418/status/2004377305138024889)
Remember Groq, the company boasting the "world's fastest inference chip"? Well, the news just dropped: **Groq founder Jonathan Ross and the core team are joining NVIDIA**. Plus, Groq's inference tech is being non-exclusively licensed to Jensen. This move feels super familiar – Google pulled the same stunt with Character.AI. GroqCloud will keep chugging along, but its heart and soul have a new owner. Jensen's play here is genius: he snagged the tech, harvested the talent, and didn't even have to shoulder the entire company's baggage. Talk about a win-win!
---
### 2. [TRAE Monthly Active Users Exceed 1.6 Million, AI Programming Ecosystem Data Revealed for the First Time](https://x.com/op7418/status/2004446870358339756)
ByteDance's AI programming tool, TRAE, just dropped its annual report, and the numbers are seriously impressive: **6 million total users, with 1.6 million monthly active users**. This might just make it the top AI programming tool in China. What's really interesting is the user behavior data – **40% use it for bug fixing, 30% for code generation**, while features like repository management and code optimization still haven't quite won everyone's trust. As for programming languages, Vue takes the lead, with Python and JS hot on its heels. Looks like front-end developers are the main force driving AI programming!
---
### 3. [Tsinghua University Teams Up with Shengshu Technology to Open Source TurboDiffusion, Accelerating Video Generation by 200x](https://www.bestblogs.dev/en/article/af2dec69)
Remember how generating a 5-second, 720p video with a Diffusion model used to take a whopping 4767 seconds (over an hour)? Well, guess what? Now it's **done in 24 seconds, or even under 2 seconds at its fastest**! The Tsinghua TSAIL team and Shengshu Technology's open-sourced TurboDiffusion framework made this happen. Their secret sauce? A three-pronged approach: hybrid attention acceleration, efficient step distillation, and W8A8 quantization. And here's the kicker – **almost no loss in image quality**! This means even regular consumer-grade GPUs can now dabble in real-time video generation. AI video is officially moving from "offline waiting" to "real-time preview" era. Mind blown! 🤯
---
### 4. [OpenAI Plans to Raise $100 Billion, Valuation to Hit $830 Billion](https://www.aibase.com/zh/news/24060)
OpenAI is gearing up for another massive move. Reports say they're eyeing **up to $100 billion in funding**, which, if successful, would push their valuation to a staggering $830 billion. What does that even mean? It's more than a third of NVIDIA's market cap! While it's still early days and investor demand isn't set in stone, one thing's crystal clear: the AI arms race is burning through cash faster than most folks can even imagine. 💰
---
### 5. [ChatGPT to Launch "Skills" Feature in Early 2026](https://www.aibase.com/zh/news/24060)
OpenAI's new feature, internally codenamed "Hazelnut," has just been revealed! **Skills are modular capability units**, organized like folders, containing instructions, context, examples, and even executable code. The coolest part? Skills can be automatically recognized, dynamically loaded, and combined by AI. Plus, you can even convert existing GPTs into Skills with a single click. This means you'll soon be able to assemble AI capabilities like LEGO bricks, instead of training from scratch every time. Expect it to drop in January 2026.
---
### 6. [Liquid AI Releases 2.6B Parameter Model, Outperforming Billion-Parameter Competitors](https://www.aibase.com/zh/news/24060)
Small models are winning again! Liquid AI's LFM2-2.6B-Exp, with just 2.6 billion parameters, has **outperformed billion-parameter models** in multiple benchmarks, especially in instruction following. What's even more crucial? It was trained purely with reinforcement learning – no supervised fine-tuning, no large model distillation. It's fully open-source and supports edge device deployment. So, in the future, running a small model on your phone might just be even better than a massive cloud-based one.
---
### 7. [Kuaikan Manhua Teams Up with MiniMax, AI Interactive Comics Are Here](https://www.aibase.com/zh/news/24060)
Comic characters can finally "come alive"! Kuaikan Manhua and MiniMax have teamed up to launch AI companion interactive comics. They're using **AI tech to enable multi-modal interaction and emotional connection with comic characters**. You can chat and interact with your favorite characters, not just passively read their stories. This is being hailed as the "fourth track of AI experience" in the comic industry. For anime/manga fans, this might just be one of the most exciting product forms to drop this year! 🤩
---
### 8. [Huawei Noah's Ark Lab Releases SCOPE, Enabling Agents to Evolve from Errors](https://www.bestblogs.dev/en/article/d02a4e21)
What's the biggest headache with AI Agents? They **see error messages but keep making the same mistakes** next time. Huawei Noah's Ark Lab, in collaboration with CUHK, just dropped the SCOPE framework to fix this! It enables an Agent's Prompt to self-evolve, learning from execution trajectories. In tests on the HLE benchmark, task success rates shot up from 14.23% to 38.64% – more than doubling! This is a major breakthrough for building truly reliable AI Agents.
---
### 9. [Xiaohongshu Open Sources InstanceAssemble, Breaking Through Accuracy in Complex Multi-Instance Image Generation](https://www.aibase.com/zh/news/24060)
One of the biggest headaches in e-commerce? **Getting AI to precisely place multiple products in a single image**. Xiaohongshu's open-sourced InstanceAssemble framework tackles this head-on. It significantly boosts spatial alignment accuracy in complex scenarios using cascaded modeling and an Assemble-Attention mechanism. The best part? It only adds 0.84% to parameters and is compatible with mainstream base models. For anyone doing e-commerce design or poster generation, this is definitely a tool to keep an eye on!
---
### 10. [OpenAI Releases Holiday Custom Codex Model: GPT-5.2-Codex-XMas](https://x.com/op7418/status/2004376604282400957)
Okay, this one's purely an Easter egg! The OpenAI Codex team just dropped a **special Christmas edition model, GPT-5.2-Codex-XMas**. Its performance is identical to the regular version, but it comes with a "Santa Claus personality." When you call it via the command line, it'll chat with you about code in Santa's jolly tone. While it doesn't have much practical use, this little bit of humor is pretty adorable! 🎅
---
## **📌 Worth Noting**
**[Products]**
- [Lingguang AI Launched for a Month, Users Created 12 Million Flash Apps](https://www.aibase.com/zh/news/24060) - Quickly create mini-apps without coding; a fresh take on AI-native apps.
- [Lenovo to Unveil "Super Intelligent Agent" at CES, Integrating PC and Mobile Ecosystems](https://www.aibase.com/zh/news/24060) - Taking on Doubao, Lenovo aims to create a cross-device AI assistant.
**[Research]**
- [AI-Generated Faces Can Be Detected with 5 Minutes of Training](https://www.aibase.com/zh/news/24060) - Teeth, hairlines, and ear symmetry are the telltale signs.
- [MeViSv2 Multimodal Video Understanding Dataset Released](https://www.bestblogs.dev/en/article/3719b2da) - Now includes 150,000 seconds of audio data, pushing the limits of AI logical reasoning.
**[Open Source]**
- [LEANN: RAG for Everything, Saving 97% Storage Space](https://github.com/yichuan-w/LEANN) - Run fast, accurate, and private RAG on your personal device.
- [MediaCrawler: Xiaohongshu/Douyin/Bilibili Crawler Tool](https://github.com/NanmiCoder/MediaCrawler) - With 40k+ stars, it's a must-have for content analysis.
**[Misc]**
- [Epic is Giving Away "Disco Elysium: The Final Cut" for Free Today](https://x.com/op7418/status/2004398113826710001) - A masterpiece, seriously, go grab it!
---
## **❓ Related Questions**
### How to Experience ChatGPT's Latest Features (e.g., Upcoming Skills)?
ChatGPT's advanced features (like GPT-4o, upcoming Skills, etc.) typically require a ChatGPT Plus subscription. For users in mainland China, this often means facing payment hurdles or account registration restrictions.
**Solutions:**
- **[Aivora](https://aivora.cn)** offers ready-made ChatGPT Plus accounts.
- Super fast delivery: order, use immediately, no payment or registration hassles.
- Get a stable, exclusive account with worry-free after-sales support.
Visit [aivora.cn](https://aivora.cn) to check out the full list of AI account services.Last updated on