腾讯研究院AI速递 20260116

Group 1: AI Chip Regulations - The U.S. has imposed a 25% tariff on advanced AI chips like Nvidia's H200 and AMD's MI325X, with export licenses now subject to case-by-case review instead of presumed denial [1] - New regulations stipulate that the number of chips exported to China cannot exceed half of the total quantity for U.S. customers and must meet specific safety standards [1] - The U.S. House of Representatives has passed the Remote Access Security Act to restrict China's access to AI chips via cloud computing services [1] Group 2: Google AI Developments - Google has launched the Personal Intelligence feature powered by the Gemini3 model, integrating data across Gmail, Photos, YouTube, and Search for contextual understanding [2] - This feature includes a natural language correction mechanism, allowing users to correct AI errors in real-time, thus lowering the management threshold for data models [2] - Currently in beta testing, it is available to paid users and will eventually be accessible to free users across multiple platforms [2] Group 3: Nvidia's Autonomous Driving - Nvidia's new L2++ level driving system in the Mercedes CLA has successfully completed a 40-minute test in San Francisco, demonstrating capabilities comparable to Tesla's FSD [3] - Nvidia plans to launch L2 highway and city driving features by mid-2026, with a goal to expand Robotaxi deployment by 2027 and achieve L3 highway driving by 2028 [3] - The company has achieved city autonomous driving functionality in just one year, utilizing the Drive AGX Thor chip, which costs approximately $3,500 [3] Group 4: AI Shopping Innovations - The Qianwen App has introduced over 400 service functions, enabling AI-driven shopping experiences across various Alibaba ecosystem services [4] - New features include AI food ordering, shopping, restaurant reservations, and direct access to 50 government services, enhancing user convenience [4] - The app's "Task Assistant" function leverages breakthroughs in AI coding and multimodal understanding for various applications [4] Group 5: Didi's AI Assistant - Didi has launched an AI assistant named "Xiao Di," allowing users to specify vehicle preferences through simple phrases, including vague requests like "for large luggage" [6] - The assistant prioritizes user needs into categories such as "necessary," "priority," and "preferable," enhancing the personalization of service [6] - After three months of iterations, the AI has improved user experience by remembering habits and preferences [6] Group 6: Step-Audio-R1 Model - The Step-Audio-R1.1 model has topped the Artificial Analysis Speech Reasoning leaderboard with a 96.4% accuracy rate, surpassing other leading models [7] - It is the first open-source native speech reasoning model capable of end-to-end understanding and real-time responses without added latency [7] - The model will have a complete real-time speech API available by February, with current chat modes supporting fluid reasoning [7] Group 7: GPT-5.2 Browser Development - The CEO of Cursor has utilized GPT-5.2 to autonomously write 3 million lines of code over a week, creating a complete browser from scratch [8] - The project employed a multi-agent system with planners and executors to ensure efficient task completion with minimal conflicts [8] - Results indicate that GPT-5.2 can maintain focus and follow instructions effectively over extended periods, outperforming other models in planning capabilities [8] Group 8: Robot Rental Platform - The world's first robot rental platform, "Qingtian Rent," has completed seed funding, led by Hillhouse Capital and supported by several other investors [9] - Within three weeks of launch, the platform has registered over 200,000 users and maintains an average of over 200 rental orders daily [9] - The platform employs a shared rental and scheduling model, with rental prices ranging from 200 yuan per day for long-term rentals to over 1,000 yuan for daily rentals [9] Group 9: AI in Robotics - A research project from Columbia University has been featured on the cover of Science Robotics, showcasing a humanoid robot capable of synchronized lip movements using deep learning [10] - The robot's facial structure contains over 20 micro-motors hidden beneath flexible silicone skin, utilizing self-supervised learning to control expressions [11] - It can convert sound signals into natural lip movements across various languages and environments, demonstrating robust cross-linguistic capabilities [11]