Qwen's former lead argues AI is moving from reasoning to agents

MarkTechPost reports on former Qwen lead Junyang Lin's argument that AI progress is shifting from training better stand-alone reasoning models toward training agents. Lin says agentic reinforcement learning needs infrastructure, environments, and reward systems rather than just hybrid thinking modes. The piece also flags reward hacking as a central unresolved risk as teams move more capability into long-running tool-using agents.

Key details: Junyang Lin previously led work on Qwen, His thesis is that the field is moving from model training toward agent training, The argument centers on environments, reward design, and reward-hacking risk.

Why it matters: It captures a live strategy shift in frontier AI: from bigger reasoning modes toward infrastructure for autonomous agents.