Qwen Spins Up 3,300 RL Worlds for Cents

🔥 What happened Qwen's team flipped the pipeline: solve a math model first, then render it into a full agentic environment. The result: 3,300 RL worlds at a few cents apiece. 💡 Why it matters Qwen3.6-35B-A3B jumps from 0.204 to 0.815 mean agentic score — and beats Qwen3.7-Max on E-Commerce Bench. The paper's kicker: most of the learnable gap sits in stateful interaction, not raw problem solving. ⚡ Our take Building environments before defining outcome rules is backwards. VHD-Play proves synthetic training worlds are now commodity — the moat moves to the mechanisms themselves.