人工智能发展趋势和策略探讨  

Research on the development trends and strategies in artificial intelligence

作　　者：肖子玉 张鹏飞 XIAO Zi-yu;ZHANG Peng-fei(China Mobile Group Design Institute Co.,Ltd.,Beijing 100080,China;China Mobile Group Co.,Ltd.,Beijing 100033,China)

机构地区：[1]中国移动通信集团设计院有限公司,北京100080 [2]中国移动通信集团有限公司,北京100033

出　　处：《电信工程技术与标准化》2024年第S02期1-12,共12页Telecom Engineering Technics and Standardization

摘　　要：当前算力需求指数增长,算法效率显著提升,大模型训练领域的竞争加剧了技术的快速迭代与应用进程。AI基础设施成为稀缺资源,能耗和成本问题已日益凸显。本文研判未来的研发趋势,包括高质量数据、高效率训练方法、模型架构优化、AI for Science加速科学发现以及结合具身智能开发适应性更强的机器人,提出加强多领域研发布局、针对边缘和端侧应用场景全面融入自有生态、实施影子模式增强模型自适应能力、推进全栈国产化、边缘和端侧部署深度定制推理硬件、构建梯次化的人工智能计算布局等发展策略。This paper analyzes the current state of artificial intelligence development,where the demand for computational power has grown exponentially while algorithmic efficiency has significantly improved.The intensified competition in large model training has accelerated the rapid iteration and application process of technologies.AI infrastructure has become a scarce resource,with energy consumption and cost issues increasingly prominent.Anticipated future research trends encompass high-quality data utilization,efficient training methodologies,optimized model architectures,the acceleration of scientific discovery through AI for Science,and the development of more adaptive robots that integrate embodied intelligence.The paper proposes strategies for advancement,including strengthening multi-disciplinary R&D layouts,flly integrating edge and endpoint scenarios into proprietary ecosystems,implementing shadow mode to enhance model adaptability,promoting full-stack domestication,deploying deeply customized inference hardware at the edge and endpoints,and constructing a tiered AI computing infrastructure framework.

关 键 词：大语言模型 大视觉模型 多模态模型 算力 能耗 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]