B[Beijing: Promote global competition in basic artificial intelligence models] On September 9, the Beijing Municipal People’s Government issued the “Beijing Digital Economy Development Plan during the 15th Five-Year Plan Period.” It pointed out that the level of scientific research on basic theories of artificial intelligence should be improved. Enhance the long-range reasoning, code generation, and intent understanding capabilities of large language models, tackle multimodal model understanding, generation, and alignment technologies, and achieve world-leading breakthroughs in basic models. Explore new model architecture, swarm intelligence, brain-like intelligence and other directions to seize the commanding heights of the next generation of artificial intelligence technology. Improve engineering levels such as model compression, efficient inference, and lightweight deployment. Construct a classification and grading model evaluation system with international influence. B[Seven departments: Strengthen the research and development of key supporting technologies such as storage and computing integration, high-performance storage, data compression, distributed training framework, DB4AI (database for AI services)] On September 4, seven departments including the Central Cyberspace Administration of China jointly issued the "Implementation Plan for Promoting Digital Green Collaborative Transformation Development (2026-2030)". It pointed out that the research and development of key supporting technologies such as storage and computing integration, high-performance storage, data compression, distributed training framework, and DB4AI (database for AI services) should be strengthened, and core technologies such as high-speed interconnection, high-bandwidth memory, and low-power computing should be accelerated to improve the efficiency of data transmission and model training. Accelerate the research on more efficient model training and inference methods, actively promote the optimization and iteration of model architectures such as hybrid experts and dynamic sparse, strengthen research on model compression technologies such as quantification and pruning, explore lightweight deployment methods, and achieve a balance between ultra-large-scale parameter model performance and resource overhead. Accelerate the construction of full-stack collaboration capabilities of software and hardware, strengthen R&D adaptation of low-power chips, operator acceleration libraries, communications, compilation, frameworks, algorithms, and multi-agent collaboration to support rapid iteration of artificial intelligence technology and universal ease of use. Build a resource consumption assessment system for the entire life cycle of generative artificial intelligence, guide generative artificial intelligence service providers to proactively disclose large model resource utilization, and guide the green and efficient development of artificial intelligence technology.
— All loaded —
