On March 25, during the opening ceremony of the 2026 Zhongguancun Forum, Yang Zhilin, the founder and CEO of Moonshot AI, delivered a keynote speech focusing on open-source AI. He highlighted that open-source models, exemplified by Kimi K2.5, are steadily emerging as the new industry norm, necessitating hardware manufacturers to verify performance enhancements through open-source model evaluation benchmarks.

Yang Zhilin also noted a paradigm shift in the research methodology within the AI domain. Traditionally, the industry heavily depended on vast amounts of internet data supplemented by a limited volume of human annotations, with human judgment determining alignment with human values and preferences. However, since the previous year, there has been a growing emphasis on reinforcement learning within the industry, although the selection of high-quality tasks still necessitates human intervention. The advancements observed in programming and mathematics are a direct result of this approach.

Looking ahead, AI is poised to assume an even more significant role in research, with each researcher being equipped with a substantial number of tokens to synthesize novel tasks and environments. AI will be responsible for defining the optimal reward functions within these environments and may even venture into exploring new network architectures, thereby expediting the entire AI research and development process.