Kimi Yang Zhilin: Large Model Training Enters Third Stage of AI-Driven Research
Moonshot founder Yang Zhilin stated at the Zhongguancun Forum annual meeting on March 25, 2026, that large model training is entering a third critical stage driven by AI. This paradigm shift signals a move from relying on natural data and manual annotation toward highly automated self-evolution.

Reflecting on the technical roadmap, Yang Zhilin outlined three phases in large model evolution: the first phase three years ago primarily depended on natural internet data and limited manually annotated value alignment; the second phase last year centered on large-scale reinforcement learning, with researchers curating high-quality tasks to enhance model performance. In 2026, a fundamental shift has occurred in AI research methods, and the role of researchers is evolving into an "AI compute scheduler." In this new stage, the research process is driven by AI, which uses vast numbers of tokens to autonomously synthesize new tasks and environments, define optimal reward parameters, and even actively engage in exploring new network architectures.
This trend suggests that AI research and development efficiency is poised for exponential acceleration. Moonshot announced that its core product Kimi will focus on advancing the frontiers of intelligent technology and fostering a collaborative, evolving 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 a significant milestone on the path to achieving general artificial intelligence (AGI), signaling a shift from passive learning to autonomous exploration.
Related article
SpaceX IPO Filing Highlights Satellite Internet and AI Expansion Ambitions
In its S-1 registration statement filed ahead of a planned IPO, SpaceX recently unveiled a number of impressive business metrics that highlight its strong footprint in aerospace communications and artificial intelligence:Starlink subscribers surpass
Alibaba Tuhao M890 Debuts with Triple Performance, Ushering in Full-Stack Agent Era for Chip-Cloud-Model-Inference
On May 20, 2026, at the Alibaba Cloud Summit, Alibaba Cloud announced the completion of a full-stack technology system upgrade designed for the Agentic era. The transformation reshaped the entire pipeline—from underlying chips and cloud platform to m
Pentium 4 Revival: 20-Year-Old CPU Runs Meta Llama 3 Large Model
Recently, the YouTube tech channel Fully Buffered carried out an impressive and hardcore experiment: successfully running Meta's latest Llama 3.2 3B large model on the Pentium 4 641 processor, a chip released in 2006.This test forced modern artificia
Related Special Topic Recommendations
Comments (0)
0/500

Reflecting on the technical roadmap, Yang Zhilin outlined three phases in large model evolution: the first phase three years ago primarily depended on natural internet data and limited manually annotated value alignment; the second phase last year centered on large-scale reinforcement learning, with researchers curating high-quality tasks to enhance model performance. In 2026, a fundamental shift has occurred in AI research methods, and the role of researchers is evolving into an "AI compute scheduler." In this new stage, the research process is driven by AI, which uses vast numbers of tokens to autonomously synthesize new tasks and environments, define optimal reward parameters, and even actively engage in exploring new network architectures.
This trend suggests that AI research and development efficiency is poised for exponential acceleration. Moonshot announced that its core product
SpaceX IPO Filing Highlights Satellite Internet and AI Expansion Ambitions
In its S-1 registration statement filed ahead of a planned IPO, SpaceX recently unveiled a number of impressive business metrics that highlight its strong footprint in aerospace communications and artificial intelligence:Starlink subscribers surpass
Alibaba Tuhao M890 Debuts with Triple Performance, Ushering in Full-Stack Agent Era for Chip-Cloud-Model-Inference
On May 20, 2026, at the Alibaba Cloud Summit, Alibaba Cloud announced the completion of a full-stack technology system upgrade designed for the Agentic era. The transformation reshaped the entire pipeline—from underlying chips and cloud platform to m
Pentium 4 Revival: 20-Year-Old CPU Runs Meta Llama 3 Large Model
Recently, the YouTube tech channel Fully Buffered carried out an impressive and hardcore experiment: successfully running Meta's latest Llama 3.2 3B large model on the Pentium 4 641 processor, a chip released in 2006.This test forced modern artificia





Home






