Computing

不要从AutoGen、LangGraph等框架入手学习Agentic AI——它们会锁定你的世界观。在领域快速演变期，应该从第一性原理出发自己构建。

Don't start Agentic AI with frameworks like AutoGen or LangGraph—they lock you into a worldview. Build from first principles while the field is still evolving rapidly.

对比GPT-4o和o1评测DeepSeek r1，发现其在指令遵循和思维深度上不及第一梯队，但在中文表达和创意写作方面展现出独特优势。

A critical review of DeepSeek r1 comparing it to GPT-4o and o1, finding it lacking in instruction following and reasoning depth, but exceptional in Chinese language mastery and creative writing.

用医院的分诊系统作比喻，通俗解释DeepSeek v2和v3的混合专家架构原理，包括专家分工、路由机制、负载均衡挑战和跨节点通信问题。

Explains DeepSeek v2 and v3's Mixture of Experts architecture using a hospital analogy, covering expert specialization, routing, load balancing challenges, and cross-node communication.

以一次debug经历说明三个AI管理技巧：克制抢键盘的冲动、提供可视化上下文而非模糊抱怨、授之以渔而非授之以鱼——即从IC到Manager的思维蜕变。

Uses a debugging session to illustrate three key AI management skills: resisting the urge to take over, providing visual context instead of vague complaints, and teaching methodology rather than giving answers—the shift from IC to manager mindset.

通过分离Planner和Executor、强制文档沟通、用o1当Planner三项改造，解决Cursor鬼打墙问题，实现质量显著提升的多智能体系统。

How separating Planner and Executor roles, enforcing document-based communication, and using o1 as Planner transformed Cursor from a simple assistant into a multi-agent system.