---
linkTitle: AI Daily
title: AI Daily-AI资讯日报
breadcrumbs: false
next: /en/2025-12/2025-12-19
description: Your daily source for curated AI news, practical tools, and actionable
tutorials to master Artificial Intelligence;
cascade:
type: docs
---
## AI Daily News 2025/12/20
> AI News | Daily Briefing | Web Data Aggregation | Frontier Science Exploration | Industry Insights | Open-Source Innovation | AI & Human Future | [Visit Web Version](https://ai.hubtoday.app/) | [Join Group Chat](https://source.hubtoday.app/logo/wechat-qun.jpg)
### **Today's Digest**
```
Google releases 270M parameter FunctionGemma with 85% accuracy
GPT-5.2-Codex becomes the strongest programming model, reaching 56.4% on SWE-Bench
RUC-Tencent confirms long reasoning chains accumulate noise, proposes Adaptive Think
Manus hits $100M ARR in eight months, fastest growth record globally
Pieter Abbeel takes over as Amazon AGI head, leading frontier research
```
### Product & Feature Updates
1. **Google unveils FunctionGemma.**
**FunctionGemma**, a small 270M parameter model, can now directly [convert natural language into device commands (AI News)](https://www.xiaohu.ai/c/a066c4/google-functiongemma)! Its accuracy skyrocketed from 58% to an impressive **85%** in tests. Imagine saying "set a reminder to feed the cat at 8 PM" and it instantly gets it, calling the system API. This isn't just a chatbot anymore; it's a powerful 🚀 smart agent ready to get things done.
2. **Google Gemini can detect AI-generated videos.**
**Google Gemini** now lets users upload videos ⬆️ to directly check if they were generated by Google AI. It leverages **SynthID watermark technology** to inspect both visual and audio tracks. This cool feature supports videos up to 100MB and 90 seconds, and it's [free to use globally (AI News)](https://www.aibase.com/zh/news/23831)—no subscription needed! ✨
3. **OpenAI releases GPT-5.2-Codex.**
**GPT-5.2-Codex** is officially here, and it's currently the most powerful agent programming model out there! 🤯 It boasts a **56.4% accuracy on SWE-Bench Pro** and can stay focused on complex tasks for extended periods without losing its place. Its defensive cybersecurity capabilities are also top-tier, even helping researchers uncover a [critical React framework vulnerability (AI News)](https://x.com/OpenAI/status/2001766212494332013).
4. **Kling 2.6's motion control feature is live.**
**Kling 2.6** just dropped a new motion control feature, letting users define how characters in their images move! 🤩 You can even join a creation contest for a chance to win up to **$1000 cash**! Five first-prize winners will also snag 16,000 points, and if you submit by December 31st, your work might even get [featured on the official homepage (AI News)](https://x.com/Kling_ai/status/2001891240359632965). Don't miss out!
5. **Mistral releases OCR 3.**
**Mistral OCR 3** is here, crushing its predecessor with a **74% win rate** when handling scanned forms and handwritten content! 📈 It costs just $2 per thousand pages, with bulk discounts bringing it down to a sweet $1. Plus, it can preserve complex table structures and even [supports direct Markdown output (AI News)](https://mistral.ai/news/mistral-ocr-3). Talk about efficiency!
### Frontier Research
1. **Large models' "thinking too much leads to errors" confirmed.**
The **RUC-Tencent team** has officially confirmed the "thinking too much leads to errors" phenomenon in large models! 🤯 Using information theory, they discovered that excessively long reasoning chains accumulate noise. Their solution? A new **Adaptive Think** strategy that tells the model to "stop when confident." This approach slashed Token consumption on GSM8K by half, and even [improved accuracy (AI News)](https://arxiv.org/abs/2505.18237). No wonder their paper was selected as a NeurIPS 2025 Spotlight! ✨
2. **JARVIS framework enhances visual reasoning.**
The **JARVIS framework**, a [self-supervised learning framework (AI News)](https://arxiv.org/abs/2512.15885) inspired by I-JEPA, is boosting visual reasoning for multimodal large models! 🧠 It helps them learn visually without relying solely on text descriptions. Experiments consistently show significant improvements on vision-centric tasks without compromising other multimodal reasoning abilities. The code is already open-sourced on GitHub – go check it out! 🚀
3. **AIMM detects social media stock market manipulation.**
**AIMM**, a new AI framework, is here to sniff out stock market manipulation on social media! 🧐 It combines Reddit activity and OHLCV data to generate a daily manipulation risk score. Amazingly, it issued a warning **22 days before the GME event**! The [truth dataset (AI News)](https://arxiv.org/abs/2512.16103), containing 33 labeled samples, has already been open-sourced. Take that, market manipulators! 📉
4. **Pull-based protocols solve AI collaboration challenges.**
A recent paper dives into AI collaboration, finding that knowledgeable Leaders often struggle to guide Followers effectively due to a lack of "theory of mind," causing success rates to plummet from 35% to a mere 17%. 📉 But here's the kicker: experiments proved that active, question-driven **[Pull protocols are more stable than Push commands (AI News)](https://arxiv.org/abs/2512.15776)**, doubling the frequency of clarification requests. It seems asking is better than telling! 🤔
### Industry Outlook & Social Impact
1. **Manus hits $100M ARR in 8 months.**
**Manus**, a Singaporean AI agent company, has just set a new global record, smashing past $100 million in ARR in just eight months! 🚀 With a monthly compound growth rate exceeding **20%**, it has processed a staggering 147 trillion tokens. This powerhouse can autonomously handle [complex tasks (AI News)](https://www.aibase.com/zh/news/23862), from resume screening to full-stack development, all with a lean team of just 105 people. Mind-blowing! 🤯
2. **Amazon AGI head steps down.**
**Pieter Abbeel**, the reinforcement learning guru, is taking the reins of Amazon's frontier research team, replacing Rohit Prasad after his two-year tenure. This UC Berkeley professor's former students include [OpenAI co-founders (AI News)](https://www.jiqizhixin.com/articles/2025-12-19-2), and his academic citations total a whopping 231,000! Talk about a big-name hire! 🌟
3. **ByteDance AI phone solution unveiled.**
**ByteDance's AI phone solution** is shaking things up! 📱 They're waiving token sharing and custom development fees, asking only for a prominent entry point. They're already in talks with Vivo, Lenovo, and Transsion to [pre-install Doubao Assistant (AI News)](https://www.aibase.com/zh/news/23851). This means phone manufacturers can rake in a share of traffic and membership revenue, directly hitting the previous pain point of sky-high token costs. Smart move! 💸
4. **AWS CEO opposes laying off junior developers.**
**AWS CEO Matt Garman** is calling out the "dumbest idea ever": replacing junior developers with AI. 🙅♂️ He argues that junior employees are actually better at using AI tools. Garman emphasizes that the talent pipeline is like a sports team; [not nurturing new talent will lead to a gap (AI News)](https://www.jiqizhixin.com/articles/2025-12-19-2) down the line. He believes AI will create even more jobs in the long run. Good point! 💡
### Top Open-Source Projects
1. **PentestGPT: A penetration testing powerhouse.**
**PentestGPT**, a GPT-driven security tool, is automating penetration testing workflows, helping security researchers uncover system vulnerabilities faster! 🛡️ It supports analysis across various attack vectors and is [open-source and free to use (AI News)](https://github.com/GreyDGL/PentestGPT). Sweet! 👍
2. **Stanford CS229 Cheatsheet.**
This **VIP cheatsheet** for Stanford's classic CS229 Machine Learning course is a goldmine! 📚 It covers core concepts like supervised learning and deep learning. An absolute must-have for review and exam prep, it's truly a [condensed essence (AI News)](https://github.com/afshinea/stanford-cs-229-machine-learning) of knowledge. Get studying! 🧑💻
3. **Metabase: Open-source BI tool.**
**Metabase**, a business intelligence powerhouse, makes data handling a breeze for everyone! 📊 It supports embedded analytics and visualization, and its enterprise-grade features are [fully open-source (AI News)](https://github.com/metabase/metabase). This is truly great news for small and medium-sized teams! 🎉
### Social Media Share
1. **Context engineering becomes a new moat.**
The **Box CEO** made a killer point: AI agents are evolving from "model capabilities" to "system architecture," and the root cause of failure isn't logical flaws, but **information asymmetry**. 🤯 He argues that context engineering is essentially reverse-engineering what [information input (AI News)](https://x.com/shao__meng/status/2001980022773645663) an expert needs. This is the new moat! 🏰
2. **ByteDance's 35% salary increase is insane.**
While everyone else is hitting the brakes on growth, **ByteDance** just announced an insane average salary increase of **35%**! 💰 Netizens are collectively expressing [envy, jealousy, and hatred (AI News)](https://x.com/op7418/status/2001979689846587723) – and who can blame them?! Wild! 🤑
3. **Xiaohongshu AI video goes viral with 100K likes.**
**Uncle Yingfeng's viral AI video on Xiaohongshu** just racked up 100,000 likes! 📈 His work ingeniously avoided the dreaded AI breathing pauses, and the sound transitions and rhythm were both precise and impactful. Gaining 100K likes in just 10 days proves the [terrifying power of long-tail recommendations (AI News)](https://x.com/huangyun_122/status/2001962295501766768). Check it out! 👇
4. **Claude Code is surprisingly powerful.**
**Li Mo** just showed off how surprisingly powerful **Claude Agent SDK** is! 🤯 He demonstrated using a Feishu app as a database for one-click collection and publishing to Xiaohongshu, and even wrapping it as an API to run periodically. The coolest part? When running a dozen tasks in parallel, if there's an error, it will [self-correct its code (AI News)](https://m.okjike.com/originalPosts/6944d57475a476d43923630d) and rerun! That's next-level automation. ✨
5. **Dissecting Plan Mode's Architectural Moat.**
The **Flask author** is shedding light on **Plan Mode's** architectural moat, pointing out that its native implementation is deeply integrated with IDE toolchains, allowing it to perceive file states in real-time. This means users can intercept approvals at **atomic-level steps**, essentially [transforming from a coder to a reviewer (AI News)](https://lucumr.pocoo.org/2025/12/17/what-is-plan-mode/)! Talk about control! 🧐
6. **16-year-old hacker breaks into four tech giants.**
A **16-year-old hacker** managed to breach Discord, Vercel, Cursor, and X through a Mintlify SVG/XSS vulnerability! 🤯 However, the bounty payments amounted to only a few thousand dollars, sparking controversy. The discussion highlighted that putting third-party content on the main domain is the [root cause of creating risk (AI News)](https://newshacker.me/story?id=46317098) in the first place. Food for thought! 🤔
7. **Google Conductor introduces context-driven development.**
**Google Conductor**, a new Gemini CLI extension, is here to revolutionize development with context-driven AI! ✨ It automatically scans your project structure, extracts relevant code, and packages it into rich context requests for your model. Say goodbye to tedious manual copy-pasting, and ensure [AI is no longer "feeling the elephant in the dark" (AI News)](https://developers.googleblog.com/conductor-introducing-context-driven-development-for-gemini-cli/). Genius! 💡
---
## **AI Daily News Audio Edition**
| **Xiaoyuzhou** | **Douyin** |
| --- | --- |
| [Afterlife Tavern](https://www.xiaoyuzhoufm.com/podcast/683c62b7c1ca9cf575a5030e) | [Self-media Account](https://www.douyin.com/user/MS4wLjABAAAAwpwqPQlu38sO38VyWgw9ZjDEnN4bMR5j8x111UxpseHR9DpB6-CveI5KRXOWuFwG)|
|  |  |