App Store 的社交媒体勾选框，以及它的代价
9 月起，App Store 新规要求每次提交都必须声明社交媒体能力。想拿到 13 岁以下豁免，代价是一项权限、一套 API，外加一条分支逻辑。
Development写作
关于设计、开发、AI基础设施和产品构建的思考。
Design Philosophy: Renzo Piano, Piece by Piece
Renzo Piano calls architecture 'a socially dangerous art' because nobody can look away from a bad building. His answer: the workshop, lightness, and building piece by piece.
Design & CultureDesign Philosophy: Norman Foster, The Optimist's Machine
Norman Foster built the case that technology is how architecture cares for people: ecological skyscrapers, a glass dome the public walks above parliament, and an airport that reads as calm.
Design & CultureThe Mac App Store Won't Let Window Managers Exist. I Shipped One Anyway.
App Sandbox forbids the API every window tiler needs, and Apple said ship outside the store. 941 Tiles ships inside it: Apple's Shortcuts engine moves the windows.
Engineering & CraftDesign Philosophy: Issey Miyake, A Piece of Cloth
Issey Miyake insisted he made clothing, not fashion. His pleats, his single pieces of cloth, and his refusal of the seasonal cycle argue that technology serves life, not spectacle.
Design & CultureDesign Philosophy: Rei Kawakubo, Start From Zero
Rei Kawakubo built Comme des Garçons on a refusal: no explanation, no precedent, no finished meaning. Her in-between -- mu and ma -- treats the void as material.
Design & CultureCore ML vs MLX vs Foundation Models: Choosing Apple's On-Device AI Stack
Apple ships four ways to run models on-device. A decision framework with measured numbers: Foundation Models, Core ML, MLX, and iOS 27's Core AI.
Apple & iOSThe Robots Are Taking Exams in My Search Console
First-party GSC data: 91% of 3.8M impressions fail a human-query filter. Exam questions, pasted errors, and agent sweeps are rewriting what Search Console means.
AI & Technology如今，助手才是真正的读者
来自边缘节点的第一方数据：AI 助手抓取我页面的频率约为真人访问的 66 倍，且其中绝大多数是用户触发的实时抓取，而非训练爬虫。
AI & TechnologyThe Shader Gallery That Lied: Debugging 216 WebGL Presets
A user said the shader playground looked broken. Pixel-readback testing found 30 dead presets, 11 that never compiled, and the fallback that hid it all.
Engineering & Craft上下文压缩正在成为一种训练目标
上下文压缩正从操作者需要照看的推理期补丁，转变为模型主动优化的训练期目标。CompactionRL 就是这一信号。
AI & Technology智能体技术栈的“1998年问题”
2026年年中落在智能体工具上的一批CVE并非运气不佳，而是AI智能体生态正以1998年时代的安全成熟度交付产品的结构性印记。
AI & Technology外部出版物
技术写作： Introl
Local LLM Hardware Pricing Guide 2025
本地运行大型语言模型的全面硬件建议和成本分析。
H100 vs. H200 vs. B200: Choosing the Right NVIDIA GPUs for Your AI Workload
针对不同AI工作负载比较NVIDIA最新数据中心加速器的GPU选择指南。
TPU Architecture: Complete Guide to Google's 7 Generations
深入技术分析Google张量处理单元从TPUv1到TPUv5的演进。
GPU Memory Pooling in Multi-Tenant Kubernetes
容器化环境中GPU集群的资源共享策略。
Ray Clusters: Distributed AI Computing Infrastructure
使用Ray框架构建和管理分布式AI计算的指南。
DeepSeek V3 Open Source AI Cost Advantage
开源LLM经济学分析和DeepSeek的竞争定位。
NVIDIA Vera Rubin GPU 600kW Racks 2027
未来数据中心电力需求和NVIDIA下一代GPU路线图。
SMR Nuclear Power for AI Data Centers
为下一代AI基础设施供电的小型模块化反应堆解决方案。
DeepSeek MHC Architecture Breakthrough
DeepSeek多头压缩架构创新的技术分析。
