Kimi Yang Zhilin: The Training of Large Models Has Entered the Third Stage of AI-Driven Research

Moonshot founder Yang Zhilin clearly stated at the Zhongguancun Forum annual meeting on March 25, 2026, that large model training is entering a third key stage led by AI. This paradigm shift marks that the development of large models is evolving from relying on natural data and manual annotation to highly automated self-evolution.

Looking back at the technical path, Yang Zhilin divided the evolution of large models into three periods: the first stage three years ago mainly relied on natural internet data and a small amount of manually annotated value alignment; the second stage last year focused on large-scale reinforcement learning, with researchers selecting high-quality tasks to improve model performance. Entering 2026, there has been a fundamental change in AI research methods, and the role of researchers is shifting towards "AI compute scheduler." In this new stage, the research process will be driven by AI using a large number of Tokens to autonomously synthesize new tasks and environments, define the most suitable reward parameters, and even deeply participate in exploring new network architectures.

This trend indicates that AI research and development efficiency will enter an exponential acceleration period. Moonshot stated that its core product Kimi will focus on pushing the boundaries of intelligent technology and building a collaborative evolutionary technology ecosystem with the open-source community. The transition from "human teaching AI" to "AI guiding research" is not only an upgrade in training methods but also an important milestone in the path to achieving general artificial intelligence (AGI), marking a shift from passive learning to autonomous exploration.