Qwen Architect Reveals: AI Models Are Shifting from Thinking to Doing
From Thought to Action: The Next AI Revolution
Lin Junyang, the engineering mastermind behind Alibaba's Qwen large language model, broke his silence weeks after departure with a startling prediction: "We're witnessing AI's transformation from calculators to collaborators."
The Agentic Thinking Breakthrough
Speaking exclusively about his post-Alibaba reflections, Lin describes an industry at a crossroads. "For years, we obsessed over making models think longer and deeper," he admits. "But last year revealed the missing piece - thought means nothing without action."
His concept of "Agentic Thinking" represents a seismic shift. Imagine an AI that doesn't just answer questions but:
- Dynamically adjusts its approach based on real-world feedback
- Develops and refines action plans through interaction
- Balances deliberation with decisive execution
Lessons from the Qwen Trenches
The road to this revelation wasn't smooth. Lin openly shares their 2025 struggle to create a unified thinking-doing system. "We tried building one model to rule them all," he recalls with a wry smile. "Turns out, asking an AI to be both philosopher and plumber creates mediocre results at both."
The team discovered fundamental tensions:
- Reasoning vs Execution: Deep thought requires patience; actions demand speed
- Cost vs Reliability: Complex deliberation drained resources without guaranteeing better outcomes
- Flexibility vs Focus: General-purpose systems often underperform specialized ones
These hard-won insights led Qwen to separate its "Instruct" and "Thinking" versions - a move now considered industry best practice.
Rethinking Intelligence Itself
Lin challenges conventional wisdom about AI capability metrics: "Longer reasoning chains don't equal smarter models," he asserts. "Sometimes they're just expensive wheel-spinning."
His proposed new benchmark? Actionable intelligence - how effectively thoughts translate into real-world results. This shifts the R&D focus from training isolated models to developing complete agent systems that:
- Continuously learn from environmental feedback
- Optimize for practical outcomes rather than theoretical perfection
- Balance computational costs against tangible benefits
The implications could reshape everything from customer service bots to scientific research assistants.
Key Points:
- Paradigm Shift: AI moving from passive reasoning to active problem-solving
- Hard Lessons: Early attempts at unified systems revealed fundamental design challenges
- New Metrics: Effectiveness replacing complexity as the gold standard
- System Thinking: Future development focuses on complete agent-environment ecosystems
