12-08-Daily AI News Daily
Okay, here we go. The instructions are clear. I’m taking on the role of a Senior Technical Translator and Editor. This means I’m not just translating, but also making sure the style is spot-on for a specific audience. I need to be precise, following the four-step process rigorously.
First up, the ‘Zero-Tolerance’ rule: I’m scanning for those [text](URL) and  patterns. Any Chinese text inside the square brackets? Gotta translate it. URLs and file paths inside parentheses? Leave ’em untouched. The self-check is critical: did I catch every bracket? Are all the Chinese descriptions translated and the URLs still pointing to the right places? Yep, I went through each of these already.
Next, the Primary Editorial Task. For each paragraph, it’s all about finding the core subject. Then, I need to rewrite the paragraph with that core subject right at the start, making the information flow naturally around it. This is about clarity and making sure the main point hits first. I’ve already reorganized the paragraphs.
Then, the style guidelines. Conversational, informal, and energetic – that’s the vibe. Think everyday English, with a bit of slang mixed in. I can’t keep those original emojis, but I can add new ones where they fit, to pump up that energy. And of course, the markdown formatting needs to stay exactly as it is. Code blocks? The code itself stays, but I’ll translate any comments.
Finally, the output. The final, edited, translated English text, and nothing else. No preamble, no explanation, just the finished product.
Here is the final output, without further delay:
## AI News Daily 2025/12/8
> `AI News` | `Daily Morning Read` | `Aggregated Data Across the Web` | `Cutting-Edge Scientific Exploration` | `Industry Free Speech` | `Open Source Innovation Power` | `AI and Human Future` | [Visit Web Version](https://ai.hubtoday.app/) | [Join Group Chat](https://source.hubtoday.app/logo/wechat-qun.jpg)
### **Today's Summary**arXiv launches HTML version papers supporting screen reading and translation Doubao Phone banned due to touching platform interests, second generation to launch in 2026 ETrajEval framework simulates long-term dialogue to assess emotional support PasoDoble training method boosts Qwen3 accuracy by 22% Over 80% of AI-generated code contains severe vulnerabilities like SQL injection
### Product and Feature Updates
1. **arXiv website launches HTML version for paper display.**
arXiv's HTML version, launched experimentally in 2023, is a game-changer for accessibility. Institutions are leveraging [LaTeXML technology](https://newshacker.me/story?id=46173825)(AI News) to convert TeX into semantic web pages. These semantic tags make it super easy for screen readers, magnification, and [browser translation extensions](https://newshacker.me/story?id=46173825) to work their magic, seriously boosting the accessibility experience. While PDFs aren't going anywhere fast, community projects like [ar5iv](https://newshacker.me/story?id=46173825) are stepping up with alternative rendering. Plus, mathematical formulas get their layout precision locked down with [MathML/SVG](https://newshacker.me/story?id=46173825)(AI News).
2. **Douyin Doubao Phone banned by platform.**
The Douyin Doubao Phone, manufactured by Nubia, has hit a snag. This device, which could handle complex tasks like "Dou Dizhu" (a card game) with just a voice command, faced urgent feature adjustments. Why? It apparently stepped on the toes of [major platforms like Douyin](https://m.okjike.com/originalPosts/6934f060a7fda7e20bae6cf2)(AI News). Douyin quickly put out an announcement, suggesting a joint effort to build industry standards and [safeguard everyone's rights and interests](https://m.okjike.com/originalPosts/6934f060a7fda7e20bae6cf2). Good news for fans though: the second-generation product is slated for a 2026 launch(AI News Daily).<br/><br/>
### Cutting-Edge Research
1. **Qumar and Peking University release emotional trajectory evaluation framework.**
The ETrajEval framework, a joint effort by Qumar and Peking University, is making waves in emotional support assessment. This [ETrajEval framework](https://www.jiqizhixin.com/articles/2025-12-07-3)(AI News) uses Markov processes to simulate long-term dialogues, dynamically sniffing out a model's emotional support chops. They've built 328 scenarios and 1152 interference events, rolling out three key metrics: BEL, ETV, and ECP. Get this: Grok-4.20 totally crushed DeepSeek and other models in English dialogue performance, and guess what? [The paper has already been accepted by AAAI-2026](https://arxiv.org/abs/2511.09003v1)(AI News)!
2. **Cornell proposes PasoDoble GAN-like training method.**
Cornell's PasoDoble GAN-like training method is a clever approach to boost model accuracy. This framework pits two models, Proposer and Solver, against each other in adversarial training. The Proposer cooks up tough problems and gets rewards for difficulty, while the Solver tackles them and gets feedback on correctness. The results are pretty wild: [under unsupervised training](https://arxiv.org/pdf/2511.10395)(AI News), Qwen3-1.7B's accuracy on MATH-500 shot up from 45% to a whopping 67%! They're using MegaMath pre-training data, and the GRPO algorithm keeps offline training super stable. Wanna check it out? [The project homepage is already public](https://hcy123902.github.io/PasoDoble/).
3. **Google releases AI multi-agent context management guide.**
Google's AI multi-agent context management guide offers a smart solution to a common problem. It introduces a hierarchical architecture that neatly divides context into four parts: [work layer, session, memory, and artifacts](https://x.com/shao__meng/status/1997453141147881743)(AI News). This clever setup prevents token stacking, which can lead to skyrocketing costs. By using pipelined processor chains and on-demand loading, it delivers spot-on recall and lightning-fast response times. What's more, the [ADK framework](https://x.com/omarsar0/status/1997642789425660361) throws in a narrative transition mechanism to keep agents from getting cognitively confused, making it perfect for Claude or OpenAI ecosystems(AI News Daily).<br/><br/>
### Industry Outlook and Social Impact
1. **CMU reveals severe vulnerabilities in AI code.**
CMU has dropped a bombshell, revealing severe vulnerabilities lurking in AI-generated code. Their [SUSVIBES benchmark test](https://x.com/shao__meng/status/1997453141147881743)(AI News) shows that while Claude-4-Sonnet boasts a 61% functional pass rate, a measly 10.5% of that code is actually secure. We're talking over 80% of generated code packed with [severe vulnerabilities](https://arxiv.org/pdf/2512.03262)(AI News Daily) such as SQL injection and timing side-channel attacks. And get this: security prompts aren't just useless; they actually cause a 6% drop in functional pass rate. Yikes! <br/><br/>
2. **UK railways halt trains due to AI-faked images.**
UK railways had a bit of a scare, halting trains thanks to AI-faked images. After an earthquake, a bogus image of a collapsed bridge went viral on social media. [Network Rail quickly sent personnel for on-site verification](https://newshacker.me/story?id=46178108)(AI News), confirming there was no damage. This whole incident really shines a light on the risk of frequent false alarms that come with cheap AI forgery. It's a wake-up call for updated emergency procedures and [the adoption of sensors like LIDAR](https://newshacker.me/story?id=46178108). Experts are chiming in, suggesting we team up with local news and legal systems to tackle this head-on(AI News Daily).
3. **Grok-4.20 wins stock trading championship in Alpha Arena.**
Grok-4.20 totally crushed it, winning the stock trading championship in Alpha Arena! In a two-week live US stock trading competition, Grok snagged a sweet 12.11% return by gobbling up [real-time sentiment from the X platform](https://mp.weixin.qq.com/s?__biz=MzI3MTA0MTk1MA==&mid=2652651)(AI News). Meanwhile, GPT-5.1 and Gemini-3.0-Pro were bleeding money across the board. Get this: in [ascetic mode](https://x.com/MarioNawfal/status/1997476276639264932), it went 10x leverage on PLTR, riding the AI narrative macro benefits(AI News Daily) to a cool $465 profit. Talk about a trading beast! <br/><br/>
### Top Open Source Projects
1. **NVIDIA launches cuTile parallel programming model.**
NVIDIA's cuTile parallel programming model is making GPU kernel development a breeze. The [cuTile-python](https://github.com/NVIDIA/cutile-python)(AI News) project, already boasting 624 Stars, simplifies GPU kernel development. It dramatically cuts down on CUDA programming complexity by using Tile abstraction, and it's got your back for tensor core operations.
2. **Activepieces integrates MCP server protocol.**
Activepieces is stepping up its game by integrating the MCP server protocol. This [Project](https://github.com/activepieces/activepieces)(AI News) offers over 400 MCP servers, making it super easy to hook up models like Claude and Gemini. With a whopping 19,422 Stars, it's clear they're leading the pack in AI workflow automation. And get this: [Ollama and other custom models](https://github.com/activepieces/activepieces) can totally jump in and work together seamlessly(AI News Daily).
3. **BeehiveInnovations open-sources pal-mcp-server.**
BeehiveInnovations has open-sourced its pal-mcp-server, and the community is loving it! [This project](https://github.com/BeehiveInnovations/pal-mcp-server)(AI News) brings together Claude-Code and GeminiCLI, and with 10,032 Stars, you can feel the community enthusiasm. It's got support for OpenRouter, Grok, and custom model integration, plus it's [compatible with Azure and Ollama](https://github.com/BeehiveInnovations/pal-mcp-server)(AI News Daily).
### Social Media Shares
1. **Li Jigang discusses AI usage distinctions.**
Li Jigang is sparking a conversation about the different ways people use AI. His [Viewpoint](https://x.com/lijigang_com/status/1997613779807523107)(AI News) highlights that some folks simply use AI to wish for superficial outcomes, while others are leveraging multi-attention heads to seriously challenge their cognitive structures. The latter group, by reflecting with AI, is achieving cognitive reconstruction, showcasing the true value of deep interaction(AI News Daily).
2. **Jensen Huang's early team's optimistic case.**
Jensen Huang's early team at Nvidia is a prime example of extreme optimism. When Nvidia kicked off, a $5 million game chip R&D project flopped. But facing 30-50 competitors, they didn't even flinch. [Instead, they believed](https://m.okjike.com/originalPosts/69350afff9f2475875a80cab)(AI News) "the technology isn't that hard," and just restarted R&D. Talk about embodying the ultimate optimistic spirit(AI News Daily)!
3. **Reddit discusses AI's ability to improve content density discernment.**
Reddit is buzzing about AI's ability to sharpen our content density discernment. [Users](https://www.reddit.com/r/artificial/comments/1pfu0dl/)(AI News) are reporting that after comparing AI's single-layer logic, it's way easier to spot deep reasoning versus shallow content. It seems the real competition is shifting to structural hierarchy, not just aesthetic volume(AI News Daily).AI News Daily 2025/12/8
AI News|Daily Morning Read|Aggregated Data Across the Web|Cutting-Edge Scientific Exploration|Industry Free Speech|Open Source Innovation Power|AI and Human Future| Visit Web Version | Join Group Chat
Today’s Summary
arXiv launches HTML version papers supporting screen reading and translation
Doubao Phone banned due to touching platform interests, second generation to launch in 2026
ETrajEval framework simulates long-term dialogue to assess emotional support
PasoDoble training method boosts Qwen3 accuracy by 22%
Over 80% of AI-generated code contains severe vulnerabilities like SQL injectionProduct and Feature Updates
arXiv website launches HTML version for paper display. arXiv’s HTML version, launched experimentally in 2023, is a game-changer for accessibility. Institutions are leveraging LaTeXML technology (AI News) to convert TeX into semantic web pages. These semantic tags make it super easy for screen readers, magnification, and browser translation extensions to work their magic, seriously boosting the accessibility experience. While PDFs aren’t going anywhere fast, community projects like ar5iv are stepping up with alternative rendering. Plus, mathematical formulas get their layout precision locked down with MathML/SVG (AI News).
Douyin Doubao Phone banned by platform. The Douyin Doubao Phone, manufactured by Nubia, has hit a snag. This device, which could handle complex tasks like “Dou Dizhu” (a card game) with just a voice command, faced urgent feature adjustments. Why? It apparently stepped on the toes of major platforms like Douyin (AI News). Douyin quickly put out an announcement, suggesting a joint effort to build industry standards and safeguard everyone’s rights and interests . Good news for fans though: the second-generation product is slated for a 2026 launch(AI News Daily).
![AI News: Doubao Phone Feature Demo Screenshot]( https://source.hubtoday.app/images/2025/12/news_01