12-23-日报-AI资讯日报

## ⚡ 快速导航

- [📰 今日 AI 资讯](#今日ai资讯) - 最新动态速览



> 💡 **提示**：想第一时间体验文中提到的最新 AI 模型（Claude、GPT、Gemini）？没有账号？来 [**爱窝啦 Aivora**](https://aivora.cn?utm_source=daily_news&utm_medium=mid_ad&utm_campaign=content) 领个号，一分钟上手，售后无忧。

**Processing The Core Idea**

Okay, I've got a handle on the request's core. The "Senior AI Observer" role and tone are coming into focus. I'm focusing on weaving in natural language and emotional hooks to create a truly compelling piece. It's about bringing the user along with a natural conversational flow.


**Analyzing Inputs and Constraints**

I'm now diving deep into the raw input, applying the constraints as a filter. Late-night shares are in focus, so I'm thinking of creating some contrasts. I have some big news: OpenAI's year review by Brockman is a strong user-engagement item. I'm prioritizing the Yann LeCun vs. AGI debate, as it has high conflict and philosophical value. Anthropic's `skills` repo is a developer-focused, potentially valuable addition. I also found something about "Vibe Coding" I'm reviewing as well.


**Exploring Further Options**

I've just added several potentially newsworthy items to the mix. Ant Group's Anzhen'er model is now out, plus Google's A2UI which generates UIs looks like an interesting practical tool. LLMLingua-2 is pretty technical, but important for saving tokens. I've also noted that Waymo and Tesla both survived the recent San Francisco blackout! Finally, Claude Code/GLM 4.7 is out there, getting positive reviews.


**Incorporating Recent News**

I'm now integrating fresh details! Ant Group's new medical model is a significant domestic release. Plus, Google's A2UI, is a potential front-end disruptor I'm excited to explore. LLMLingua-2 is a practical token saver, which is important. Also, I'm noting the gap between GitHub stars and real-world agent frameworks, a reality check from DAIR.AI. GLM 4.7 is also making progress, getting some early reviews.


**Revising Key Perspectives**

I'm now revising my key perspectives. Coding, according to the Flask creator, is evolving into a more intuitive, AI-directed process focused on vibes rather than rigid syntax. San Francisco's blackout highlighted the contrasting approaches of Waymo and Tesla, with Tesla's vision system proving more resilient. Anthropic's new agent skills repo is a valuable resource. I'm also reviewing OpenAI's year in review, and a reality check from DAIR.AI shows 80% of current agent frameworks have major problems, and are not delivering.


**Prioritizing & Integrating Findings**

I'm now integrating the user feedback on agent frameworks and evaluating their utility, which is pretty surprising. I'm focusing on Ant Group's new medical model which seems strong, and Google's A2UI which offers a major user-facing tool. I'm also really interested in LLMLingua-2's potential cost savings. I'm keeping GLM-4.7 in mind, and its recent strong performance is showing it's a front-runner.


**Optimizing the Output**

I'm now generating the response and making key refinements. I've re-written some sections, particularly the medical model description, to emphasize "clinical thinking" rather than just retrieval. Also, I've simplified the LeCun discussion, making it more relatable through the "pen and paper" analogy. I've ensured all the provided images have been included. Finally, I've avoided corporate jargon and kept the markdown formatting clean for easy reading.


## **今日AI资讯**

### **👀 只有一句话**
旧金山大停电，Waymo 无人车集体瘫痪堵死路口，而不需要高精地图的 Tesla FSD 却淡定飘过，这波马斯克赢麻了。

### **🔑 3 个关键词**
#VibeCoding #LeCun开炮 #无人车大考

## **🔥 必看 TOP 10 (图文深度)**

### **1. [Flask 作者神预言：编程已死，Vibe Coding 当立](https://x.com/dotey/status/2003176742857691603)**
**这可能是今年最扎心的一篇文章。** 写了二十年代码的大神 Armin Ronacher（Flask 框架作者）坦言：2025年，他已经不再像以前那样写代码了。以前是亲自敲键盘，现在是做“虚拟实习生的技术领导”。
他提出了一个新词 **"Vibe Coding"**（氛围编程）：你不再纠结语法和最佳实践，而是凭直觉判断 AI 生成的代码“感觉”对不对。这种玄学编程法让传统工程经验失效，但也让人从繁琐实现中解脱。如果你还在死磕语法，可能真要被时代抛弃了。
![AI资讯图片](https://pbs.twimg.com/media/G8y2kXmXEAALk_o?format=jpg&name=orig)

### **2. [Yann LeCun 深夜开炮：你们说的“通用智能”根本不存在](https://x.com/ylecun/status/2003227393071439979)**
**就在大家都在吹 AGI（通用人工智能）快来了的时候，Meta 的首席科学家 LeCun 又出来泼冷水了。** 他发长文怒怼：人类的大脑是极度“专用”的，根本不是通用的。
他打了个比方：理论上只要给足够的时间和纸笔，人脑也能算图灵完备，但在实际任务（比如下棋或处理海量数据）中，人脑效率低得发指。他认为现在的 LLM 路线也是同理，总是幻想“通用”是死路一条。这老哥在离职前是真敢说啊。

### **3. [Waymo 遇停电瘫痪，纯视觉方案的胜利时刻？](https://ai.hubtoday.app//2025-12/2025-12-23/)**
**以前大家总觉得激光雷达+高精地图才安全，现实却狠狠打了一巴掌。** 旧金山突发大停电，红绿灯全灭。依赖高精地图和基础设施交互的 Waymo 无人车瞬间变成“智障”，不知道怎么走，直接把路堵死。
反观特斯拉 FSD，像人一样靠“看”来判断路况，在混乱中居然应对自如。这次事故直接暴露了依赖预设地图的脆弱性，视觉派的粉丝今晚可以开香槟了。

### **4. [OpenAI 官方发糖：看看你今年跟 ChatGPT 聊了啥](https://x.com/gdb/status/2003242320448352371)**
**属于 AI 时代的“网易云年度歌单”来了。** OpenAI 总裁 Greg Brockman 宣布推出 "Your Year with ChatGPT"。如果你开启了记忆功能，它会根据你一年的对话记录，分析你的兴趣、工作习惯，甚至是你自己都没发现的怪癖。
目前已在美国、英国等地推送，记得更新 App。这不仅仅是回顾，更像是一面镜子，让你看看这一年你是变强了，还是变懒了。
![AI资讯图片](https://pbs.twimg.com/media/G8zBqDIXYAAz41p?format=jpg&name=orig)

### **5. [Anthropic 悄悄开源了 Agents 的“必杀技”库](https://github.com/anthropics/skills)**
**Claude 的母公司 Anthropic 也是人狠话不多，直接在 GitHub 上开源了一个叫 `skills` 的仓库。**
这不是普通的工具包，而是官方定义的 Agent 技能标准。如果你正在开发智能体，还在头疼怎么让 AI 优雅地使用工具、操作电脑，这个库就是你的“武功秘籍”。拥有 25k Star 的关注度说明了一切：大厂下场教你怎么做 Agent 了。
![AI资讯图片](https://opengraph.githubassets.com/9f0ec7a83afaa277dff29ce0475e537a2fa27ffa3e93928c373e73e74e26a997/anthropics/skills)

### **6. [GitHub 星星多就是好框架？这份报告打脸了 80% 的开发者](https://x.com/omarsar0/status/2003180917528088758)**
**别再盲目跟风用 LangChain 或 AutoGen 了！** DAIR.AI 发布了一份硬核报告，分析了 1500 多个 Agent 项目。结论很扎心：GitHub Star 数根本代表不了好用程度。
比如 MetaGPT 有 5.9 万星，但在实际复杂项目中用得极少。反倒是 LangGraph 虽然星星少，但落地很稳。报告还指出，目前的框架在“逻辑失败”和“死循环”上简直是灾难。选工具前，先看看这篇避坑指南。
![AI资讯图片](https://pbs.twimg.com/media/G8y3_UeagAAcI0v?format=png&name=orig)

### **7. [智谱 GLM-4.7 偷跑？代码能力疑似大幅增强](https://linux.do/t/topic/1351367)**
**国产模型这几天有点东西。** 有开发者在使用智谱 API 时，惊讶地发现模型版本变成了 GLM-4.7。实测下来，思考时间变长了（这是好事，说明在推理），写代码时不再傻乎乎地乱加 Emoji，而是能听懂复杂的逻辑拆解。
这种“静默升级”通常意味着正式版快了。如果你嫌 Claude 太贵，这个也许是近期值得关注的平替。
![AI资讯图片](https://linux.do/uploads/default/optimized/4X/2/2/e/22e0f6a1497bf2734d1db4fa50e0f6e2a4410785_2_690x254.png)

### **8. [Google 开源 A2UI：以后写界面不用写前端代码了？](https://ai.hubtoday.app//2025-12/2025-12-23/)**
**前端工程师看完可能会失眠。** Google 开源了 A2UI 项目，核心逻辑粗暴有效：AI 不需要写 HTML/CSS 代码，直接生成 JSON 描述，然后前端 App 自动把这个 JSON 渲染成原生界面。
这意味着智能体可以直接“变”出一个操作面板给你，完全绕过了“AI 写代码会出错”的安全隐患。未来的 App 可能都是动态生成的，你需要啥界面，AI 现场给你画一个。
![AI资讯图片](https://source.hubtoday.app/images/2025/12/news_01kd3arb6vemevwb7wk30gghsg.avif)

### **9. [LLMLingua-2：微软教你如何把 Prompt 费用砍半](https://linux.do/t/topic/1351456)**
**这就是真金白银的省钱技巧。** 很多人的 Prompt 又臭又长，跑一次 API 心都在滴血。微软的 LLMLingua-2 就是专门干这个的：它能把你的长文本压缩（蒸馏），去掉废话保留精华。
测试数据显示，Token 用量能减少近 50%，但模型理解的准确率几乎不变。对于那些需要塞一堆文档给 AI 的应用来说，这简直是救命稻草。
![AI资讯图片](https://linux.do/uploads/default/original/4X/4/e/7/4e7901c36545accadcb660a4640ca68fbf64cfb6.png)

### **10. [蚂蚁发布“安诊儿”：千亿参数的 AI 医生](https://ai.hubtoday.app//2025-12/2025-12-23/)**
**这不是那种只会百度看病的 AI。** 蚂蚁集团发布了千亿参数的医疗模型，采用 MoE 架构。它的亮点在于具备“临床级诊疗思维”，也就是说它不是在背书，而是在像医生一样推理。
目前已经落地在心脏健康和心理咨询场景。在医疗这个容错率极低的领域，千亿参数的投入说明大厂是真的想解决“看病难”的问题，而不是单纯炫技。
![AI资讯图片](https://source.hubtoday.app/images/2025/12/news_01kd3ar1kyed6s0tcktyf28tyd.avif)

## 📂 分类速览

**✨ 产品与工具**
*   **[Suno AI 支持加密货币支付](https://linux.do/t/topic/1351364)**：以后买 AI 音乐会员不用找代购了，虽然不知道能坚持多久。
*   **[JustHTMLs - 开源 HTML 工具集](https://linux.do/t/topic/1351413)**：所有工具都是单文件 HTML，隐私党的福音，这才是 Web 该有的样子。
*   **[PentestGPT](https://github.com/GreyDGL/PentestGPT)**：AI 赋能的渗透测试工具，黑客（白帽子）的效率神器。
*   **[NotebookLM 变身记忆神器](https://ai.hubtoday.app//2025-12/2025-12-23/)**：谷歌这产品神了，现在能自动把笔记变成考研复习卡片。

**📊 商业与行业**
*   **[中国芯片的“DeepSeek 时刻”？](https://readhub.cn/topic/8pHnhxS22gm)**：彭博社预测，国产 AI 芯片将在 2026 年迎来爆发，英伟达要小心了。
*   **[罗永浩“科技春晚”回归](https://readhub.cn/topic/8pI2gz1Nz4A)**：老罗又要开讲了，门票两小时秒光。虽然不带货，但大家都想听单口相声。
*   **[山姆 App 美颜过头被喷](https://readhub.cn/topic/8pI2I3S5w1t)**：给烤鸡加滤镜？山姆评论区图片自动提高饱和度，网友：这鸡看着有毒。

**🧠 研究与观点**
*   **[Transformer 也能搞经济学？](https://x.com/emollick/status/2003172860064633321)**：Ethan Mollick 发现用 Transformer 预测经济模型效果出奇的好，看来经济学家也要用 GPU 了。
*   **[阿里千问 2025 十大高频词](https://ai.hubtoday.app//2025-12/2025-12-23/)**：排名第一的竟然是“股票预测”，还有“算命”和“离婚”，AI 承受了太多世俗的欲望。

**🔒 安全与隐私**
*   **[Flock 摄像头裸奔](https://www.reddit.com/r/artificial/comments/1pt4tnu/flock_exposed_its_aipowered_cameras_to_the/)**：这家的 AI 摄像头直接暴露在公网，谁都能看。大厂的安全神话就是个笑话。
*   **[DeepAudit 开源](https://github.com/lintsinghua/DeepAudit)**：国内首个开源的代码漏洞挖掘多智能体系统，让每个人都能拥有一支黑客战队。

## 🧾 全量索引

| # | 类型 | 标题 | 链接 |
|---|---|---|---|
| 1 | Social | OpenAI: Your Year with ChatGPT Rolling Out | [Link](https://x.com/gdb/status/2003242320448352371) |
| 2 | Social | Yann LeCun Rants on "General" Intelligence | [Link](https://x.com/ylecun/status/2003227393071439979) |
| 3 | Reddit | Lack of Open Weight Embeddings Providers | [Link](https://www.reddit.com/r/MachineLearning/comments/1pt9w0w/d_hosted_and_open_weight_embeddings/) |
| 4 | Social | Yann LeCun on Human Intelligence Specialization | [Link](https://x.com/ylecun/status/2003228003212628274) |
| 5 | Social | Automated Red Teaming for ChatGPT Security | [Link](https://x.com/gdb/status/2003190365982183918) |
| 6 | Social | Codex Experimental Background Terminals | [Link](https://x.com/gdb/status/2003189401472639199) |
| 7 | Reddit | Evaluation Metrics for Unsupervised Subsequence Matching | [Link](https://www.reddit.com/r/MachineLearning/comments/1pt8vg6/r_evaluation_metrics_for_unsupervised_subsequence/) |
| 8 | Social | Vibe Coding Tutorial Difficulties | [Link](https://x.com/dotey/status/2003180702293196987) |
| 9 | Reddit | Universal Reasoning Model Paper Discussion | [Link](https://www.reddit.com/r/MachineLearning/comments/1pt7vk8/r_universal_reasoning_model/) |
| 10 | Social | DAIR.AI Study on AI Agent Frameworks | [Link](https://x.com/omarsar0/status/2003178742844408299) |
| 11 | Social | Detailed Findings on AI Agent Frameworks | [Link](https://x.com/omarsar0/status/2003180917528088758) |
| 12 | Social | Armin Ronacher's "A Year Of Vibes" Recap | [Link](https://x.com/dotey/status/2003176742857691603) |
| 13 | Social | Transformers in Economic Modelling | [Link](https://x.com/emollick/status/2003172860064633321) |
| 14 | Reddit | Intel GenAI Examples on Old Xeon CPUs | [Link](https://www.reddit.com/r/artificial/comments/1pt619a/intel_releases_genai_examples_v15_while/) |
| 15 | Social | Wolfram on Srinivasa Ramanujan | [Link](https://x.com/EMostaque/status/2003170883670794711) |
| 16 | Reddit | Flock Exposed AI Cameras to Internet | [Link](https://www.reddit.com/r/artificial/comments/1pt4tnu/flock_exposed_its_aipowered_cameras_to_the/) |
| 17 | Social | 哥飞评大厂安全问题 | [Link](https://m.okjike.com/originalPosts/694973b18dab01fe53584bc3) |
| 18 | Reddit | How Google Gemini Leapfrogged ChatGPT | [Link](https://www.reddit.com/r/artificial/comments/1pt3vtc/how_google_gemini_leapfrogged_chatgpt_and_why_the/) |
| 19 | News | AI资讯日报 2025/12/23 (含多个新闻点) | [Link](https://ai.hubtoday.app//2025-12/2025-12-23/) |
| 20 | Forum | 关于省token的话题，LLMLingua-2 | [Link](https://linux.do/t/topic/1351456) |
| 21 | News | 比亚迪回应技术研发人员涨薪 | [Link](https://readhub.cn/topic/8pIGyX2Am85) |
| 22 | News | 《阿凡达 3》票房遇冷，博纳影业跌停 | [Link](https://readhub.cn/topic/8pHuAvYHkwm) |
| 23 | News | 彭博：中国产业级 AI 芯片加速崛起 | [Link](https://readhub.cn/topic/8pHnhxS22gm) |
| 24 | News | LeCun 离职前吐槽 Meta 封闭作风 | [Link](https://readhub.cn/topic/8pGc1tQKWX3) |
| 25 | News | 保时捷中国自建充电站将停止运营 | [Link](https://readhub.cn/topic/8pHseEbOSnq) |
| 26 | News | 山姆 App 评论照片自动「美颜」引争议 | [Link](https://readhub.cn/topic/8pI2I3S5w1t) |
| 27 | Forum | 智能驾驶车辆推荐询问 | [Link](https://linux.do/t/topic/1351438) |
| 28 | Forum | 免费esim流量体验 | [Link](https://linux.do/t/topic/1351436) |
| 29 | Social | Google AI Tips for 2025 | [Link](https://x.com/GoogleAI/status/2003232591248306599) |
| 30 | Forum | GPT plus 美国大兵申请教程求助 | [Link](https://linux.do/t/topic/1351431) |
| 31 | Forum | 深夜闲聊贴 | [Link](https://linux.do/t/topic/1351429) |
| 32 | Forum | IOS App Store 美区ID添加Fiat24失败问题 | [Link](https://linux.do/t/topic/1351422) |
| 33 | Forum | 建议关闭“网页加速器”油猴脚本 | [Link](https://linux.do/t/topic/1351418) |
| 34 | Forum | [开源自荐] JustHTMLs - HTML 工具集 | [Link](https://linux.do/t/topic/1351413) |
| 35 | Forum | 快手发生事件询问 | [Link](https://linux.do/t/topic/1351399) |
| 36 | News | 奈飞签署200亿美元贷款协议 | [Link](https://readhub.cn/topic/8pIVE1SYC8i) |
| 37 | News | 罗永浩“科技春晚”门票被抢光 | [Link](https://readhub.cn/topic/8pI2gz1Nz4A) |
| 38 | Forum | 用文本建构你的APP-OnSpace | [Link](https://linux.do/t/topic/1351396) |
| 39 | Forum | 开u卡当visa/mastercard验证卡询问 | [Link](https://linux.do/t/topic/1351373) |
| 40 | Forum | 智谱Coding Plan的GLM4.7体验 | [Link](https://linux.do/t/topic/1351367) |
| 41 | Forum | SUNO AI支持加密货币订阅 | [Link](https://linux.do/t/topic/1351364) |
| 42 | Social | Gemini 3 Flash for Small Business | [Link](https://x.com/GeminiApp/status/2003156416555782498) |
| 43 | Forum | 机械硬盘大华希捷检测询问 | [Link](https://linux.do/t/topic/1351355) |
| 44 | Forum | 免费生成UI设计图(可导出html) | [Link](https://linux.do/t/topic/1351354) |
| 45 | Forum | GPT Plus 军人认证讨论 | [Link](https://linux.do/t/topic/1351350) |
| 46 | Forum | GCP12台服务器项目推荐 | [Link](https://linux.do/t/topic/1351344) |
| 47 | Forum | Gemini 网页端显示地区和时间 | [Link](https://linux.do/t/topic/1351338) |
| 48 | Forum | 科学AI记录工作流推荐 | [Link](https://linux.do/t/topic/1351333) |
| 49 | Forum | GLM 4.7 分数讨论 | [Link](https://linux.do/t/topic/1351327) |
| 50 | Social | Telegram 图片分享 | [Link](https://t.me/woshadiao/191707) |
| 51 | GitHub | anthropics/skills Repo | [Link](https://github.com/anthropics/skills) |
| 52 | Forum | Claude Code 原生工具好评分享 | [Link](https://linux.do/t/topic/1351251) |
| 53 | GitHub | exo: Run AI cluster at home | [Link](https://github.com/exo-explore/exo) |
| 54 | GitHub | iptv: Public IPTV channels | [Link](https://github.com/iptv-org/iptv) |
| 55 | GitHub | PayloadsAllTheThings | [Link](https://github.com/swisskyrepo/PayloadsAllTheThings) |
| 56 | GitHub | PentestGPT: AI penetration testing | [Link](https://github.com/GreyDGL/PentestGPT) |
| 57 | GitHub | cocoindex: Data transformation framework | [Link](https://github.com/cocoindex-io/cocoindex) |
| 58 | GitHub | Fabric: AI augmenting humans | [Link](https://github.com/danielmiessler/Fabric) |
| 59 | GitHub | tensorflow | [Link](https://github.com/tensorflow/tensorflow) |
| 60 | GitHub | rendercv: Typst-based CV generator | [Link](https://github.com/rendercv/rendercv) |
| 61 | GitHub | home-assistant/core | [Link](https://github.com/home-assistant/core) |
| 62 | GitHub | EntraGoat: Vulnerable Entra ID environment | [Link](https://github.com/Semperis/EntraGoat) |
| 63 | GitHub | langextract: Info extraction library | [Link](https://github.com/google/langextract) |
| 64 | GitHub | expressjs | [Link](https://github.com/expressjs/express) |
| 65 | GitHub | DeepAudit: AI Code Auditing | [Link](https://github.com/lintsinghua/DeepAudit) |
| 66 | GitHub | Free-Certifications List | [Link](https://github.com/cloudcommunity/Free-Certifications) |
| 67 | GitHub | build-your-own-x | [Link](https://github.com/codecrafters-io/build-your-own-x) |
| 68 | GitHub | LocalAI: OpenAI Alternative | [Link](https://github.com/mudler/LocalAI) |
---

## **AI 账号极速发货: [爱窝啦 Aivora ⬆️](https://aivora.cn)**

还在为 ChatGPT Plus、Claude Pro、Midjourney 的支付问题烦恼？**爱窝啦 Aivora** 为您提供一站式 AI 账号解决方案！

✅ **极速发货**：下单即发，无需等待，即刻开启 AI 之旅。
✅ **稳定可靠**：精选优质独享账号，拒绝封号焦虑，售后无忧。
✅ **超全品类**：ChatGPT Plus、Claude Pro、Midjourney、Poe、Sunno 等热门 AI 工具账号应有尽有。
✅ **超高性价比**：比官方订阅更优惠的价格，享受同等尊贵服务。

🚀 **立即访问 [aivora.cn](https://aivora.cn) 选购您的 AI 助手，释放无限创造力！**