# Most AI Work Can Wait - Tomasz Tunguz

Router > 模型选择。三层架构：skill classifier → router → model selector。70-80%流量走本地模型，靠 queueing + 异步推理降费90%+。

**标签** ：ai-architecture、llm-ops、reading

---

**来源** ：[https://tomtunguz.com/ai-execution-routing](https://tomtunguz.com/ai-execution-routing)