AI Takes the Wheel: Yang Zhilin on the Third Stage of Large Model Training

The way we do AI research is getting a major overhaul. At the 2026 Zhongguancun Forum in Beijing, Yang Zhilin, the founder of Moonshot AI—a big name in large language models—dropped a bombshell: we're now entering the third phase of AI development, what he calls "AI-led research."

From Human-Led to AI-Led

Yang explained that starting this year, the research process in AI is undergoing a qualitative leap. In the past, training AI models was a hands-on affair: human researchers meticulously designed rules, tweaked parameters, and fine-tuned every step. But that model is crumbling. With recent advances, AI systems are starting to call the shots. They're evolving and optimizing themselves, speeding up the entire R&D cycle dramatically. Think of it as moving from a driver's ed car with a human instructor to a self-driving vehicle that learns the road on its own.

Open Source Takes the Lead

A key part of this shift, Yang noted, is that open-source models have become the de facto standard. They're no longer just alternatives; they're the core engine driving technology forward and spreading innovation across the industry. This isn't just about code sharing—it's about setting the pace for the whole field.

The Forum: Where Tech Meets Industry

The Zhongguancun Forum, themed "Deep Integration of Scientific and Technological Innovation with Industrial Innovation," wasn't just about tech breakthroughs. It was a five-day event packed with over 100 activities across five major sections. Thousands of experts and business leaders gathered, and the agenda included more than 20 high-density tech matchmaking sessions and over 500 technology project roadshows. The goal? To build a bridge between cutting-edge research and real-world applications.

What This Means for Developers and Enterprises

Yang's message is clear: as AI starts to lead its own research, the efficiency and depth of large model training will hit new highs. The barriers to entry are shifting, and the old rules of AI development are being rewritten. For developers and companies, this means adapting to a landscape where AI isn't just a tool but a partner—or even a boss—in the lab.

Key Points

AI-led research: Starting in 2026, AI systems take the lead in research, self-evolving and self-optimizing.
Open-source dominance: Open-source models are now industry standards, driving innovation.
Forum scale: Over 100 events, 500 project roadshows, and thousands of participants at Zhongguancun Forum.
Paradigm shift: The way AI is developed is being restructured, with new barriers and opportunities.

AI Takes the Wheel: Yang Zhilin on the Third Stage of Large Model Training