AI Field Notes · AI Field Notes
AI 研究档案AI research archive A quiet, bilingual reading room for technical essays on reasoning models, coding agents, and the systems around them.一个安静的双语阅读空间，整理关于推理模型、代码智能体及其系统环境的技术长文。
AI 研究档案AI research archive A quiet, bilingual reading room for technical essays on reasoning models, coding agents, and the systems around them.一个安静的双语阅读空间，整理关于推理模型、代码智能体及其系统环境的技术长文。
All supplied figures are kept as local assets. Choose a language from the menu; each page shows one language at a time.所有原始图表均作为本地资源保留。请使用右上角菜单切换语言；每个页面一次只显示一种语言。
文章目录Article index
Controlling Reasoning Effort in LLMs控制 LLM 的推理强度
July 18, 20262026 年 7 月 18 日
A field guide to training, inference scaling, thinking modes, and the engineering recipes behind adjustable reasoning effort.梳理推理模型的训练、推理扩展、思考模式，以及可调推理强度背后的工程实现路径。
From Vibe Coding to Agentic Engineering从 Vibe Coding 到 Agentic Engineering
July 23, 20262026 年 7 月 23 日
A bilingual research digest on how coding agents change the unit of engineering, while responsibility, verification, and product interfaces remain human concerns.双语研究摘要：代码智能体如何改变工程活动的基本单元，以及责任、验证与产品接口为何仍然不可外包。
Skill Issue: AutoResearch and the Loopy Era of AISkill Issue：AutoResearch 与 AI 的 Loopy Era
July 23, 20262026 年 7 月 23 日
A bilingual synthesis of persistent agents, constrained AutoResearch loops, program.md as meta-optimization, and the coordination costs of machine work.双语综合：持久智能体、受约束的 AutoResearch 循环、作为元优化层的 program.md，以及机器工作带来的协调成本。
GPT-5.6 Model Family and Reasoning-Effort ArchitectureGPT-5.6 模型家族与推理强度架构
July 23, 20262026 年 7 月 23 日
A public-evidence reconstruction of GPT-5.6 tiers, effort semantics, token accounting, tool loops, and possible training mechanisms.基于公开证据重建 GPT-5.6 的模型规格、推理强度语义、token 计费、工具循环与可能的训练机制。