At the 2026 Zhongguancun Forum annual meeting held on March 25, Yang Zhilin, founder of Moonshot AI and creator of Kimi, shared insights indicating that large model training is advancing into its third stage—one characterized by artificial intelligence taking the lead in research.

Yang Zhilin pointed out that three years ago, large model training relied predominantly on natural data collected from the internet, supplemented by limited human-annotated data aligned with specific values or preferences. By last year, the focus had shifted toward large-scale reinforcement learning systems, where humans curated high-quality tasks—though the tasks themselves were still human-defined—and improved model performance through reinforcement learning on those tasks.

"However, from this year through the next several years, the methodology of AI research and development will undergo significant transformation. AI will increasingly steer the research process," Yang stated. He explained that each researcher will be equipped with a substantial allocation of AI tokens, which will assist in synthesizing new tasks and environments, defining optimal reward parameters, and even exploring novel neural architectures.

Under this new paradigm, Yang believes the pace of AI development will accelerate considerably. Moonshot AI aims to collaborate with the open-source community to advance intelligent technologies and foster a more robust ecosystem.