From Thought to Action: The Next Frontier for AI

Lin Junyang knows what it takes to build cutting-edge AI. As the former lead engineer behind Alibaba's Qwen large language model, he spent years pushing the boundaries of artificial intelligence. Now, weeks after leaving the company, he's sharing revelations that could reshape how we develop AI systems.

The Agent Revolution

"We've been obsessed with making models think longer," Lin explains in his first public statement since departing Alibaba. "But the real breakthrough comes when they learn to think in order to act."

This shift from passive reasoning to active decision-making represents what Lin calls "Agentic Thinking" - where AI doesn't just process information but continuously refines its plans through real-world interaction. Imagine an assistant that doesn't just answer questions but actually completes tasks while learning from each attempt.

Lessons from the Qwen Project

The path to this future wasn't smooth. Lin openly discusses early struggles during Qwen's development in 2025, when his team tried forcing "thinking" and "instruction" capabilities into a single system.

"It was like trying to merge two different languages," he recalls. The resulting model performed poorly at both tasks - overthinking simple commands while making rushed decisions on complex problems. These painful lessons led Qwen to separate its "Instruct" and "Thinking" versions, a move that later became an industry reference point.

Rethinking Intelligence Metrics

Lin challenges conventional wisdom about AI intelligence: "Longer reasoning chains don't necessarily mean smarter models. Sometimes they just mean wasted computing power."

He predicts research will shift focus from training standalone models to developing complete "model + environment" agent systems. The new benchmark? Not how much a model can process, but how effectively it can translate thought into action.

Key Points:

• Active over passive: Future AI needs to do more than reason - it must act and adapt
• Quality beats quantity: Longer reasoning doesn't always mean better performance
• System thinking: The next breakthrough requires designing complete agent environments
• Practical intelligence: True smarts come from effective real-world interaction