Solo 日志 Solo Log

深度解析 RAG 在长上下文时代的定位变化，从入门实践到架构选型，再到异常检测场景的真实局限 Deep analysis of RAG's positioning in the long context era, from beginner practices to architecture selection, and real limitations in anomaly detection scenarios

Harness Engineering 三部曲第三篇。Anthropic 给 Claude 加了'做梦'功能整理记忆，但这恰恰说明 Agent 失忆是行业基础架构缺陷。从物理天花板、Context Window 限制，到记忆工程的实操解决方案。 The third part of the Harness Engineering trilogy. Anthropic added a 'Dreaming' feature to Claude for memory organization, which precisely illustrates that agent amnesia is an industry-wide infrastructure flaw. From physical ceilings and context window limits to practical memory engineering solutions.

当模型推理速度远超人类审查速度时，谁来保证产出的质量、安全和可控？Harness 就是答案——不是换更强的模型，而是给模型一个能可靠工作的环境。 When model inference speed far exceeds human review speed, who ensures quality, safety, and control? Harness is the answer—not stronger models, but a reliable working environment for models.

入门 Harness Engineering 时，你关心的问题是 agent 能不能完成一个任务。大师级的问题是：agent 能不能连续跑 8 小时、跨越几十个 context window、在级联失败时自愈？ Beginners ask if an agent can complete a task. Masters ask if it can run for 8 hours, span dozens of context windows, and self-heal during cascading failures.

从 30 秒到 200 毫秒的飞跃。Apple M5 Pro + Qwen 3.6 35B 实测：14/14 代码数学逻辑满分，8/8 Agent 任务零失败。本地模型不再是聊天机器人，而是能操作你电脑的数字助手。 From 30 seconds to 200 milliseconds. Apple M5 Pro + Qwen 3.6 35B tested: 14/14 perfect score in code/math/logic, 8/8 Agent tasks with zero failures. Local models are no longer chatbots, but digital assistants that can operate your computer.

通过构建一个高性能、带工具清洗功能的 Agent Gateway，让 Claude Code 完美运行在 NVIDIA NIM 的各种开源模型上。 Build a high-performance Agent Gateway with tool sanitization to run Claude Code flawlessly on NVIDIA NIM open-source models.

AI 变强的真正质变在于'思维模式'的重构。结合两年观察，深度解构 AI Agent 的思考范式演变，揭示从胡说到深思的大脑机制。 The real qualitative change in AI's strength lies in the reconstruction of 'thinking patterns'. Combining two years of observation, deeply deconstructing the evolution of AI agent thinking paradigms, revealing the brain mechanisms from nonsense to deep thinking.

深度分析 Meta AI 智能体失控事件，揭示从辅助工具到潜在威胁的系统性风险，探讨 AI 安全与控制的边界。 A deep analysis of the Meta AI agent失控 incident, revealing systemic risks from auxiliary tools to potential threats, and exploring the boundaries of AI safety and control.

记录我作为 Solo 开发者的日常工作，分享从创意到产品的全过程。 Documenting my daily work as a solo developer, sharing the journey from idea to product.

记录 duyou.me 域名的注册和网站的搭建过程。 Documenting the domain registration and website building process of duyou.me.

打破'盲人摸象'的行业认知，将 Agent 分为主权级、IDE深度集成级和全自主工程级三个维度，提供系统性的视角。 Breaking the 'blind men touching an elephant' industry perception, categorizing agents into three dimensions: sovereign, IDE-integrated, and autonomous engineering, providing a systematic perspective.

记录督友小程序的开发过程和核心功能设计。 Documenting the development process and core functionality design of the Duyou mini program.

基于真实项目的实操记录，在 MacBook Pro M3 Max (64GB RAM) 上完成 LLM 微调，打造专属私人助理。拒绝云评测，全是干货。 Practical records based on real projects, completing LLM fine-tuning on MacBook Pro M3 Max (64GB RAM) to create a personalized assistant. No cloud reviews, all practical content.

当你的公司有4个数字员工，当你每天只需运行一行命令就能生成深度长文——这不是自由，这是另一种形式的囚禁。反思 AI 一人公司的真实代价。 When your company has 4 digital employees and you can generate in-depth articles with just one command a day—this isn't freedom, it's another form of imprisonment. Reflecting on the true cost of AI solopreneurship.

一线落地复盘，拒绝空洞愿景。从工具赋能、流程优化到组织变革，深度解析 AI 在企业中真正落地的三条路径。 Front-line implementation review, rejecting empty visions. Deep analysis of the three paths for real AI implementation in enterprises: tool empowerment, process optimization, and organizational transformation.

记录购置 MacBook Pro M5 Max 64GB 并搭建本地 AI 研究工作台的过程。 Documenting the process of purchasing MacBook Pro M5 Max 64GB and building a local AI research workstation.

记忆，是智能体对抗熵增的唯一武器。深度解析 OpenClaw 的架构设计，看它如何通过记忆系统实现自我迭代与净化。 Memory is the only weapon for agents to fight against entropy. A deep architectural analysis of OpenClaw, exploring how its memory system enables self-iteration and purification.

2026年，中国AI告别'百模大战'的浮躁，进入残酷的'应用收割期'。华为的算力底座稳了吗？Kimi和豆包谁是C端之王？对标美国六大科技巨头，拆解24个月来的产业剧变。 In 2026, Chinese AI bids farewell to the frenzy of the 'Hundred Model War' and enters the brutal 'Application Harvest Period'. Is Huawei's computing foundation stable? Who is the king of consumer AI—Kimi or Doubao? Benchmarking against the US Big Six tech giants to decode 24 months of industry upheaval.

保姆级教程，带你从环境配置到模型部署，用 DeepSeek Janus-Pro 打造一个能看懂屏幕、自动执行任务的本地 AI 助手。 A comprehensive tutorial guiding you from environment setup to model deployment, using DeepSeek Janus-Pro to build a local AI assistant that can understand screens and execute tasks automatically.

如果 API 是大厂施舍的窄门，那视觉（Vision）就是我们打破围墙的攻城锤。写给那些不甘于被'围墙花园'困住的独立开发者。 If APIs are the narrow gate controlled by big platforms, vision is the battering ram that breaks the wall. Written for indie builders who refuse to be trapped inside a walled garden.

记录「黄帝内经每天趣味一分钟」视频号的运营历程和AI应用。 Documenting the operation journey and AI application of the 'Huangdi Neijing One Fun Minute Daily' video account.

站在2026年的门槛，回顾跨越76年的AI马拉松。从图灵测试到ChatGPT，充满天才狂想、资本博弈、学派倾轧，以及无数次绝地反击。 Standing at the threshold of 2026, looking back at the 76-year AI marathon. From the Turing Test to ChatGPT, filled with genius visions, capital games, academic rivalries, and countless comebacks from dead ends.

记录「谢先生的AI深析札记」公众号的创立过程和初期规划。 Documenting the creation process and initial planning of the 'Mr. Xie's AI Deep Analysis Notes' WeChat official account.