At the 2026 Annual Conference of the Zhongguancun Forum, Yang Zhilin, the founder of Moonshot AI Kimi, announced that large-scale model training has now transitioned into its third phase—the AI-led research era. Previously, the training of large-scale models predominantly relied on natural internet data, supplemented by a modest volume of manually annotated data. However, last year marked a shift towards large-scale reinforcement learning systems, with tasks being set by humans. From this year onward, AI will assume a leading role in the research process. Researchers will harness AI's Tokens to synthesize novel tasks and environments, define reward parameters, and delve into new network architectures. Yang Zhilin is confident that this transformation will substantially expedite the research and development journey of AI. Moonshot AI is eager to collaborate with the open-source community to collectively propel the advancement of intelligent technologies and foster a thriving ecosystem.