腾讯研究院AI速递 20260107

Group 1: Generative AI Developments - Nvidia officially launched the Vera Rubin supercomputing architecture, achieving a 5x increase in inference performance and a 3.5x increase in training performance while reducing costs by 90%, set to be mass-produced and available in the second half of 2026 [1] - AMD introduced the Helios all-liquid-cooled rack platform featuring the MI455X GPU, which has 320 billion transistors and 432GB of HBM4 memory, offering a 10x performance improvement over the MI355X, with a planned release of the 2nm MI500 in 2027 [2] - Intel released the third-generation Core Ultra processor, the first based on Intel's 18A process (1.8nm), achieving 180 TOPS of edge AI computing power, with a 60% increase in multi-threaded performance and a 77% increase in gaming performance [3] Group 2: Key Personnel Changes in AI Companies - OpenAI's VP of Research, Jerry Tworek, announced his departure after seven years, citing a desire to pursue research that cannot be conducted at OpenAI, marking a significant loss of talent following the exits of other key figures [4] Group 3: AI Innovations and Experiments - MiroMind launched the MiroThinker 1.5 model, which, despite having only 30B and 235B parameters, set a new record in the BrowseComp test with a single call cost of just $0.07, innovating through an internalized training mechanism [6] - A professor at Hong Kong University of Science and Technology conducted an experiment using AI glasses powered by GPT-5.2, achieving a score of 92.5 in a computer networking exam, outperforming 95% of students [7] - Boston Dynamics unveiled the new Atlas robot, which stands 1.9 meters tall and weighs 90 kg, with a production goal of 30,000 units annually by 2028, supported by a partnership with Google DeepMind [8] Group 4: AI Training and Performance Enhancements - The ZhiYuan Institute proposed the SOP (Scalable Online Post-training) framework, integrating online, distributed, and multi-task mechanisms for real-world training, achieving a 92.5% success rate in parallel learning experiments [9] - Anthropic's community lead shared 31 practical tips for using Claude Code, emphasizing the importance of understanding when to use specific modes and how to construct prompts effectively [10][11]