清华大学翟季冬：从“算得出”到“送得到”，“智能路由”打开 AI 基础设施新赛道

Core Insights - The development of AI infrastructure (AI Infra) is crucial for the evolution and industrial application of artificial intelligence, transitioning from theoretical research to practical implementation across various industries [1][4] - The focus of AI Infra is shifting from merely enhancing computational power to optimizing the "routing" of AI services to meet specific industry needs, ensuring efficient and effective deployment [3][12] Group 1: AI Infrastructure Development - AI Infra has supported the rapid iteration and deployment of large models through breakthroughs in high-performance operators, optimized training frameworks, and efficient inference engines [1] - The current phase of AI development emphasizes the integration of AI into various sectors, requiring AI Infra to facilitate the seamless flow of intelligent services to end business scenarios [1][3] Group 2: Intelligent Routing - The concept of "intelligent routing" is proposed to address the challenges of selecting the most suitable models and services for specific tasks, enhancing efficiency and cost-effectiveness [3][7] - Two main challenges in intelligent routing are identified: model routing, which involves selecting the optimal model for a given task, and service routing, which matches the best API service provider based on safety, efficiency, and cost [3][9] Group 3: Market Dynamics and Opportunities - The Chinese government aims to promote the deep application of general large models in manufacturing by 2027, creating specialized industry models and high-quality datasets [5] - The service routing approach offers a new commercial opportunity by allowing users to access AI services without needing to directly engage with domestic computing hardware, thus simplifying the integration process [10] Group 4: Performance and Optimization - The performance of domestic computing power is sufficient to meet most intelligent service demands, but further optimization of inference engines designed for domestic architectures is necessary to enhance service delivery [10][11] - The open-source inference engine "Chitu" demonstrates significant advantages on domestic platforms due to its native development and optimization, highlighting the importance of tailored solutions for domestic computing environments [11] Group 5: Future Directions - The exploration of intelligent routing is seen as a key area for expanding the technical boundaries of AI Infra, supporting the efficient and stable circulation of intelligent services [12] - As intelligent routing becomes a standard configuration in AI Infra, it is expected to facilitate the integration of AI services into various sectors, contributing to the construction of a digital economy in China [12]