量子位；2025年度AI十大趋势报告

Group 1: Core Trends in AI Development - The report outlines ten key trends in AI development for 2025, categorized into four dimensions: infrastructure, model evolution, application landscape, and China's development [1][7]. - AI infrastructure is characterized by the rapid growth of computing power and chip innovation, with a strategic focus on building large-scale data centers and transitioning chips towards AI-native designs [1][9]. Group 2: Infrastructure Trends - The demand for computing power is skyrocketing, leading to a surge in the construction of large-scale data centers, which are evolving into integrated powerhouses for computation, storage, and networking [17][19]. - Countries are elevating computing power infrastructure to a strategic level, with China planning a national integrated computing network to optimize resources and support digital transformation across industries [19][20]. Group 3: Model Evolution - Innovations in model architecture, such as the Mixture of Experts (MoE) and linear/sparse attention mechanisms, are addressing efficiency bottlenecks in large models, allowing for better performance without significantly increasing computational costs [29][33]. - The trend towards model distillation is gaining traction, enabling smaller models to replicate the capabilities of larger models, which is crucial for deploying AI on edge devices with limited computational resources [38]. Group 4: Application Landscape - The emergence of the Agentic Internet is reshaping user interaction, with AI agents capable of closing the loop from perception to execution, thus lowering development barriers and enabling multi-scenario applications [2][9]. - Multi-modal applications are becoming essential, with advancements in video, 3D, and code generation technologies driving productivity across various industries [2][9]. Group 5: China's Unique Opportunities - China is seizing unique opportunities in the global AI landscape, with a focus on developing core technologies and enhancing the influence of domestic open-source models [3][9]. - The country is accelerating the domestic replacement of AI chips and software, achieving significant milestones in training large models with domestic technologies, thus ensuring a self-sufficient AI ecosystem [25][27].