12-20-Daily AI News Daily

AI Daily News 2025/12/20

AI News | Daily Briefing | Web Data Aggregation | Cutting-Edge Science Exploration | Industry Voice | Open Source Innovation | AI & Human Future | Visit Web Version↗️ | Join Group Chat🤙

Today’s Headlines

Google releases 270M parameter FunctionGemma with 85% accuracy.
GPT-5.2-Codex becomes the strongest coding model, achieving 56.4% on SWE-Bench.
Renmin University & Tencent confirm noise accumulation in long reasoning chains, propose Adaptive Think.
Manus hits $100M ARR in eight months, setting a global growth record.
Pieter Abbeel takes over as Amazon AGI Head, focusing on frontier research.

Product & Feature Updates

1. Google launches FunctionGemma. Google FunctionGemma, a 🔥tiny 270M parameter model, can directly handle converting natural language into device commands (AI News) . Its test accuracy has soared from 58% to an impressive 85%. Say “Set a reminder to feed the cat at 8 PM,” and it instantly understands and calls system APIs ✨, upgrading from a chatbot to a truly capable 🚀smart agent.
   
   

2. Google Gemini can detect AI videos. Google Gemini now allows users to upload videos💡to directly check if they were generated by Google AI, leveraging SynthID watermarking technology to inspect visual and audio tracks separately. This feature supports videos up to 100MB and 90 seconds, and it’s available globally for free (AI News) – no subscription needed! 🎉

3. OpenAI releases GPT-5.2-Codex. OpenAI’s GPT-5.2-Codex is currently the 🚀strongest agent programming model, boasting an impressive 56.4% accuracy on SWE-Bench Pro. It can focus on complex tasks for extended periods without losing track of progress. Its defensive cybersecurity 💡capabilities are also top-notch, even helping researchers discover a critical vulnerability in the React framework (AI News) .
   
   

4. Kling 2.6 motion control feature goes live. Kling 2.6’s motion control feature lets users define how image characters move 🔥! Participate in the creation contest and win up to $1000 cash. Five first-place winners will also receive 16,000 points. The deadline is December 31st 🎉, and submissions even have a chance to be featured on the official homepage (AI News) .
   
   

5. Mistral releases OCR 3. Mistral OCR 3 boasts a 74% win rate over its predecessor when processing scanned forms and handwritten content. It costs just 💡$2 per thousand pages, with bulk discounts bringing it down to $1. It can preserve complex table structures and supports direct Markdown output (AI News) .
   
   

Frontier Research

1. Large models’ “overthinking errors” confirmed. The phenomenon of large models ’thinking themselves into errors’ has been confirmed. The Renmin University and Tencent team used information theory to discover 🔥that overly long reasoning chains accumulate noise, proposing the Adaptive Think strategy to make models “stop when confident enough.” On GSM8K, Token consumption was halved, and accuracy even improved (AI News) ! The paper was selected for NeurIPS 2025 Spotlight 🎉.

2. JARVIS framework enhances visual reasoning. The JARVIS framework, an 💡I-JEPA-inspired self-supervised learning framework (AI News) , enhances visual reasoning. It enables multimodal large models to learn visuals without solely relying on text descriptions. Experiments show it 🚀consistently improves performance on vision-centric tasks without impacting multimodal reasoning capabilities. The code is open-sourced on GitHub.

3. AIMM detects social media stock market manipulation. AIMM, an AI framework, detects social media manipulation in the stock market. It integrates Reddit activity and OHLCV data to generate 💡daily manipulation risk scores, issuing a warning 🤯 22 days before the GME event. A truth dataset with 33 labeled samples is open-sourced (AI News) .

4. Pull-based protocol solves AI collaboration challenges. A pull-based protocol addresses AI collaboration challenges. A paper found that knowledgeable 💡Leaders often fail to guide Followers correctly due to a lack of theory of mind, with success rates plummeting from 35% to 17%. Experiments show that an active questioning Pull protocol is more stable than Push instructions (AI News) , with clarification request frequency 🚀doubling.
   

Industry Outlook & Social Impact

1. Manus hits $100M ARR in 8 months. Manus, a Singaporean AI agency, set a 🔥global record for the fastest growth, with a monthly compound growth rate exceeding 20% and processing 147 trillion tokens. It can autonomously handle complex tasks (AI News) from resume screening to full-stack development, all with a lean team of only 105 people 🤯.
   
   

2. Amazon AGI Head steps down. Amazon’s AGI Head has stepped down. Rohit Prasad 🔥left after his two-year term, and reinforcement learning guru Pieter Abbeel took over the frontier research team. This Berkeley professor’s former students include OpenAI co-founders (AI News) , and he boasts 231,000 academic citations.

3. ByteDance’s AI phone solution revealed. ByteDance’s AI phone solution has been revealed. It waives token sharing and custom development fees 💡in exchange for entry points, negotiating with Vivo, Lenovo, and Transsion to pre-install Doubao Assistant (AI News) . Phone manufacturers can share in traffic and membership revenue, precisely addressing the pain point of previously high token costs 🚀.

4. AWS CEO opposes laying off junior developers. AWS CEO Matt Garman opposes laying off junior developers. He believes replacing newcomers with AI is 🔥“the stupidest idea” because junior employees are better at using AI tools. He emphasized that talent pipelines are like sports teams, and not nurturing new talent will create a talent gap (AI News) . He predicts AI will create more jobs in the long run.

Top Open Source Projects

1. PentestGPT: A penetration testing powerhouse. PentestGPT, a GPT-driven security tool ⭐9495, is a penetration testing powerhouse. It can automate penetration testing processes 🔥, helping security researchers discover system vulnerabilities. It supports various attack vector analyses and is open-sourced and free to use (AI News) .
   

2. Stanford CS229 Cheat Sheet. The Stanford CS229 Cheat Sheet is a must-have. This 💡VIP cheat sheet ⭐18921 for the classic machine learning course covers core concepts like supervised learning and deep learning. It’s the condensed essence (AI News) for review and exam preparation.
   

3. Metabase: An open-source BI tool. Metabase, an open-source BI tool, is a ⭐45061 business intelligence powerhouse that makes it easy for 🚀everyone to handle data, supporting embedded analytics and visualization. Its enterprise-grade features are fully open-sourced (AI News) , a godsend for small and medium-sized teams! 🙌
   

Social Media Shares

1. Context engineering emerges as a new moat. Context engineering emerges as a new moat. The Box CEO analyzed the evolution of AI agents from 💡“model capability” to “system architecture,” noting that the root cause of failure is no longer logical flaws but information asymmetry. Context engineering is essentially reverse-engineering 🚀what information input (AI News) an expert needs.
   
   

2. ByteDance’s 35% salary increase is shocking! ByteDance’s 35% salary increase is shocking! When everyone else stopped growing, the average increase was surprisingly high 🔥. Netizens are universally expressing fuming with envy (AI News) 😱.
   
   

3. Xiaohongshu AI video goes viral with 100K likes. Xiaohongshu AI video goes viral with 100K likes. Uncle Yingfeng’s work 💡cleverly avoided AI breathing pauses, and the sound transitions and rhythm control were 🚀both precise and impactful. Gaining 100K likes in 10 days, the long-tail recommendation power is terrifying (AI News) !
   
   

4. Claude Code is surprisingly powerful! Claude Code is surprisingly powerful! Li Mo demonstrated how Feishu applications can act as databases 💡for one-click collection and publishing to Xiaohongshu, and can also be packaged as an API using the Claude Agent SDK to run on a schedule. When running a dozen tasks in parallel, if an error occurs, it will self-corrects code errors (AI News) and rerun 🤯.
   

5. Plan Mode architecture: A deep dive into its barriers. Plan Mode architecture: A deep dive into its barriers. The Flask author pointed out 🔥that the native plan mode is deeply integrated with the IDE toolchain, allowing real-time perception of file status. Users can intercept approvals at atomic-level steps, transforming from coders to reviewers (AI News) .
   
   

6. 16-year-old hacker breaches four major tech companies. A 16-year-old hacker breached four major tech companies. Through the Mintlify SVG/XSS vulnerability💡, he took down Discord, Vercel, Cursor, and X, but the bounty of only a few thousand dollars sparked controversy 🤔. Discussion suggests that placing third-party content on the main domain is a root cause of risk (AI News) .
   

7. Google Conductor promotes context-driven development. Google Conductor promotes context-driven development. This Gemini CLI extension can 🚀automatically scan project structures to extract relevant code, packaging it into rich context requests for models. Say goodbye to manual copy-pasting; AI is no longer flying blind (AI News) ! ✨
   
   

AI Daily News Audio Version

🎙️ Xiaoyuzhou	📹 Douyin
Laisheng Xiaojiuguan Podcast	Self-Media Account