Short writing on research systems, AI agents, and technical work.
回顾从 2023 年开始使用 LLM 时坚持的四个原则：不微调、不用 RAG、不做 constrained decoding、不把 temperature 设为 0，以及它们如何逐渐汇成我的 agent 方法论。
A look back at four principles I followed when I began working with LLMs in 2023—no fine-tuning, no RAG, no constrained decoding, and no temperature 0—and how they grew into my approach to agents.
（内容为戏仿作品）
围绕 P 与 NP 问题，本文提出一种基于 benchmark saturation 的“经验性证明”叙事：用大规模 3-SAT 基准、模型准确率和 scaling law 外推来宣称数学结论。
(Parody work)
A paper-style narrative around P versus NP that proposes an empirical proof via benchmark saturation, using a large 3-SAT benchmark, model accuracy, and scaling-law extrapolation to announce a mathematical conclusion.
从工作中的“对齐”经验出发，讨论人在 AI agent 时代的主体性：人和 agent 的区别不只是能力，而是信念、历史、责任，以及是否需要为自己的判断辩护。
A reflection on subjectivity in the age of AI agents: humans differ from agents not only by capability, but by belief, history, responsibility, and the need to stand behind their own judgments.