The Action-Oriented Future of AI

Lin Junyang, who spearheaded Alibaba's Qwen large language model development before his recent departure, broke his silence with a thought-provoking analysis of AI's next evolutionary leap. His revelations challenge how we think about artificial intelligence's capabilities and limitations.

From Contemplation to Action

"We've been obsessed with making models think longer," Lin observed during his March 26 address, "but that's like praising a chef for how long they can chop vegetables rather than the quality of their dishes." His analogy captures the industry's pivotal shift from "reasoning-based thinking" to what he calls "agent-based thinking" - where AI doesn't just ponder problems but actively works to solve them.

The Qwen Experiment: When Thinking and Doing Collide

Lin pulled back the curtain on Qwen's early growing pains in 2025. The team ambitiously tried creating a unified system that could toggle between deep reasoning and quick execution based on task complexity. "We thought we could build the AI equivalent of an Olympic decathlete," he admitted. "Turns out, even AIs need specialization."

The hybrid approach produced mediocre results - models that hesitated too much when thinking and acted too hastily when executing. This painful lesson led to Qwen's strategic pivot: separate "Instruct" and "Thinking" versions that each excelled at their specific functions.

Rethinking Intelligence Metrics

Lin challenges conventional wisdom about what makes an AI model "smart":

• Longer reasoning ≠ smarter AI: Complex thought chains often waste computational resources without improving outcomes
• Actionability is key: A model's value lies in its ability to translate thoughts into effective real-world actions
• Environment matters: Future breakthroughs will come from training complete agent systems (model + environment) rather than isolated models

"Imagine judging a basketball player solely on their vertical leap," Lin suggested. "That measurement means nothing without considering how it translates to actual gameplay."

Key Points:

• AI is evolving from passive reasoning engines to active problem-solving agents
• Early attempts to combine thinking and action functions proved ineffective
• Future development will focus on complete agent systems rather than standalone models
• Performance metrics must prioritize actionable results over theoretical capabilities