An engineering notebook on AI, systems, and everyday experiments.

5分钟用AI给全天相机加上云量识别功能：利用GPT-4V理解云量，让AI自动编写调用API的代码，展示AI时代开发效率的革命性提升。

构想一个AI知识引擎：自动收集会议和日常对话、提取有价值的知识并索引、在输出时及时提醒。三大挑战是数据采集、知识提取和实时参与。

Envisioning an AI knowledge engine that records conversations, extracts insights, and surfaces relevant knowledge when needed. Three key challenges: data collection, knowledge extraction, and real-time participation.

分享开发视频语音识别工具的技术心得：YouTube下载OAuth认证、异步任务API设计、Python并发挑战、GPT偷懒问题和iOS Shortcut集成。

Technical lessons from building an AI video transcription tool: YouTube OAuth2 authentication, async task-based APIs, Python concurrency patterns, GPT laziness issues, and iOS Shortcuts integration.

反思为何正确性不是思考的核心目标，提出真正重要的是纲举目张（Prioritization）、举一反三（Connection）和知行合一（Verification）三种能力。

视频信息密度低、不可检索、难以跳转，但up主赚大钱而写文章不赚钱。核心在于视频让"作者"站到前台，观众更容易产生情感连接和付费意愿。

Videos are terrible for learning—can't skim, search, or navigate easily. Yet creators make millions because video makes the "author" tangible, lowering the barrier to payment. Articles hide the author behind ideas.

提出AI辅助编程的核心从数据结构和算法转向注释和DocString的质量。类比管理AI与管理人类团队，强调了解能力边界、任务分解、委托和质量检查等开发经理式的思维方式。

Argues that effective AI-assisted programming shifts focus from data structures and algorithms to writing detailed comments and DocStrings. Draws parallels between managing AI and managing human teams, emphasizing delegation, context provision, and quality checks.