12-28-Daily AI News Daily

## AI Daily 2025/12/28

>  `AI Daily`

### **Today's Summary**

OpenAI is urgently recruiting a Head of Preparedness, with Sam Altman admitting models can now find critical vulnerabilities, highlighting rapidly escalating risks. VSCode has officially transformed into an "AI code editor," and MiniMax has open-sourced M2.1, indicating a full AI integration across the developer toolchain. The message today is clear: as AI capabilities grow, safety concerns become more pressing; don't just focus on new tools.

## ⚡ Quick Navigation

- [📰 Today's AI News](#today-ai-news) - A quick look at the latest updates

> 💡 **Pro Tip**: Wanna be among the first to try out the latest AI models mentioned here (Claude 4.5, GPT, Gemini 3 Pro)? No account? No worries! Grab one from [**Aivora**](https://aivora.cn?utm_source=daily_news&utm_medium=mid_ad&utm_campaign=content) – get started in a minute, with worry-free after-sales support. 🚀

## **Today's AI News**

### **👀 One-Liner**
OpenAI is on the hunt for an "AI Safety Head," with Sam Altman himself posting that models are already capable of finding vulnerabilities.

### **🔑 3 Keywords**
#SafetyAlert #OpenSourceFiesta #VSCodeTransformation

---

## **🔥 Top 10 Heavy Hitters**

### 1. [OpenAI Urgently Hiring "Head of Preparedness," Sam Altman: Models Can Now Find Critical Vulnerabilities](https://openai.com/careers/head-of-preparedness-san-francisco/)

Sam Altman, in a late-night post, revealed that **AI models are already powerful enough to uncover critical vulnerabilities in cybersecurity**. Thought AI safety was a distant concern? Think again! By 2025, they're already grappling with AI's potential impact on mental health, and now they're staring down the barrel of biological capability release and self-improving system safety issues. This "Head of Preparedness" role demands you "dive straight into the deep end," promising immense pressure but profound impact. OpenAI is finally fessing up: **as model capabilities skyrocket, so do the risks.** 😬

### 2. [VSCode's Official Website Quietly Changed Its Slogan: Now Called "Open Source AI Code Editor"](https://code.visualstudio.com/)

VSCode's official website now proudly displays its new slogan: **"The open source AI code editor."** This isn't just marketing fluff—VSCode now natively supports AI auto-completion, code modification confirmation, and more. Microsoft is essentially declaring that **the AI era for code editors has officially arrived**, letting you enjoy AI superpowers without needing extra plugins. How cool is that? ✨

![AI News Image](https://pbs.twimg.com/media/G9MXyerW4AArkPO?format=jpg&name=orig)

### 3. [MiniMax Open-Sources M2.1 Model, Significantly Enhancing Code and Tool Usage Capabilities](https://huggingface.co/MiniMaxAI/MiniMax-M2.1)

MiniMax has made another splash in the domestic large model scene, open-sourcing its M2.1 model. This new iteration boasts significant improvements over its predecessor in **code generation, tool invocation, and task planning**. For developers keen on running open-source models locally, this is definitely a new, exciting option worth checking out. 🤩

![AI News Image](https://pbs.twimg.com/media/G9LSYSZaQAEIXYD?format=jpg&name=orig)

### 4. [Zhipu AI Releases Z Code: One API Key to Switch Between Multiple AI Programming Agents](https://zcode-ai.com/)

Zhipu AI has launched **Z Code**, a lightweight visual desktop application that integrates the capabilities of various AI programming agents like Claude Code, Codex, and Gemini CLI. Tired of those command-line tools feeling too high-barrier? With Z Code, **a single API Key lets you seamlessly switch between agents**, no more fiddling with endless configurations. For anyone eager to try multiple AI programming tools without the hassle, this one's a sweet deal! 🍬

![AI News Image](https://pbs.twimg.com/media/G9LUU_6bIAAI9Gt?format=jpg&name=orig)

### 5. [StoryMem: Enabling AI Video Models to "Cut Shots" and Generate Multi-Shot Long Videos](https://github.com/Kevin-thu/StoryMem)

StoryMem is here to spice up your AI video game! Tired of single-shot AI videos? This tool enhances single-shot models like Wan2.2 into a **multi-shot video model**, capable of generating long videos with dynamic scene cuts, much like Sora2. The open-source code and models are already out there, so if you're into AI video creation, give it a whirl! 🎬

### 6. [High School Student Uses AI to Discover 1M+ Astronomical Objects, NASA Director Personally Tweets Recruitment Offer Including Fighter Jet Experience](https://x.com/kevinweil/status/2004953446353699224)

This is one seriously inspiring story! A high school student, leveraging AI, **discovered over a million astronomical objects that humans had previously missed** in cosmic data. NASA's Jared Isaacman was so impressed he publicly tweeted a recruitment offer, even promising a fighter jet flight as a signing bonus! OpenAI's CPO Kevin Weil called it the perfect collision of all his worlds—truly, **AI is empowering ordinary people to make extraordinary discoveries.** 🤯

### 7. [ezff: Process Videos with a Single Sentence, No Need to Memorize FFmpeg Commands](https://x.com/vista8/status/2004948184565776386)

ezff is here to save your day if you're constantly looking up FFmpeg commands! After installing with `npm install -g ezff`, you can process videos using natural language: **"ff convert video.mp4 to gif" or "ff compress video.mp4 to 10mb."** This tool handles format conversion, compression, cropping, audio extraction, and speed adjustments—all with plain English commands. Talk about a game-changer! 🚀

![AI News Image](https://pbs.twimg.com/media/G9MApOrb0AA-o6l?format=jpg&name=orig)

### 8. [ES-CoT: Enabling Reasoning Models to "Quit While Ahead," Saving 41% of Tokens](https://arxiv.org/abs/2509.14004)

ES-CoT is a new research paper addressing the cost of Chain-of-Thought reasoning models, which are powerful but pricey. This new approach, **ES-CoT**, enables models to detect if an answer has converged during the reasoning process and stop early if it has. The best part? **No fine-tuning or architectural changes needed**, and it saves an average of 41% of inference tokens without sacrificing accuracy. This is a super practical trick for production environments looking to keep costs down. 💰

![AI News Image](https://pbs.twimg.com/media/G9LjwEjbgAcYX4B?format=png&name=orig)

### 9. [Claude Code Transcripts: Turning Claude Code Conversations into Beautiful HTML Documents](https://github.com/simonw/claude-code-transcripts)

Claude Code Transcripts, a new tool by Simon Willison, transforms your Claude Code JSON sessions into **paginated, mobile-friendly HTML pages**. When you're coding with Claude Code, those conversation logs are actually valuable project documentation. This tool makes it super easy to convert them and even share them to GitHub Gist with a single click. For anyone looking to solidify their AI programming experience, this is incredibly useful! 📝

![AI News Image](https://pbs.twimg.com/media/G9LgLxcWwAAIBGG?format=jpg&name=orig)

### 10. [System 3 Arrives: Enabling AI Agents to "Self-Evolve"](https://x.com/omarsar0/status/2004909360431108405)

System 3 is here, and it's a game-changer for AI Agents! While System 1 handles rapid perception and System 2 deep reasoning, what exactly is **System 3**? This paper introduces **Sophia**, a metacognitive framework that empowers agents with continuous adaptation, autonomous goal generation, and identity recognition. During a 36-hour deployment test, it demonstrated significant self-improvement capabilities. **This could very well be the pivotal step in transforming agents from mere "tools" into true "partners."** 🤖

![AI News Image](https://pbs.twimg.com/media/G9LcoofbgAIEMiQ?format=png&name=orig)

---

## **📌 Worth Noting**

**[Product]** [One-Click GIF Sticker Generator Tool](https://x.com/Gorden_Sun/status/2004962899266322813) - Upload an image, and it automatically generates 16 animated stickers in 6 different styles. So fun! 😂

**[Open Source]** [awesome-llm-apps](https://github.com/Shubhamsaboo/awesome-llm-apps) - A collection of LLM applications with 84k stars, including RAG and Agent examples. A treasure trove! 💎

**[Open Source]** [Yuxi-Know](https://github.com/xerrors/Yuxi-Know) - An Agent platform combining LightRAG and knowledge graphs, supporting MCP. Pretty advanced stuff! 🧠

**[Tool]** [rendercv](https://github.com/rendercv/rendercv) - A YAML to PDF resume generator, specifically for academics and engineers. Super handy for job hunting! 💼

**[Research]** [Review of AI Agent Memory Systems](https://www.bestblogs.dev/en/article/f4144ca1) - A review of 200+ papers, proposing three forms of memory: Token, Parameter, and Latent. Deep dive material! 📚

**[Discussion]** [Over 20% of Videos Recommended to New YouTube Users Are AI-Generated Content](https://www.reddit.com/r/artificial/comments/1px5wjv/more_than_20_of_videos_shown_to_new_youtube_users/) - Another piece of evidence for the proliferation of AI slop. Yikes! 😬

**[Industry]** [China Launches National Distributed AI Computing Network](https://www.reddit.com/r/artificial/comments/1pwzlpl/china_activates_a_nationwide_distributed_ai/) - Connecting data centers over 2000 kilometers apart. Big moves! 🇨🇳

---

## **❓ FAQs**

### How to Experience Claude Code?

Claude Code, an AI programming tool from Anthropic, currently requires a Claude Pro or Max subscription to use. For users in mainland China, this can often mean payment hurdles or account registration restrictions. Bummer, right?

**Solution:**
- **[Aivora](https://aivora.cn)** offers Claude-related account services.
- Super fast delivery: order and use immediately, no payment or registration headaches.
- Get a stable, dedicated account with worry-free after-sales support.

Check out [aivora.cn](https://aivora.cn) for the full list of AI account services.

OpenAI is urgently recruiting a Head of Preparedness, with Sam Altman admitting models can now find critical vulnerabilities, highlighting rapidly escalating risks. VSCode has officially transformed into an "AI code editor," and MiniMax has open-sourced M2.1, indicating a full AI integration across the developer toolchain. The message today is clear: as AI capabilities grow, safety concerns become more pressing; don't just focus on new tools.