以美光 HBM2E 为例，从 DRAM 存储单元、阵列和 bank 讲到 GPU 旁的内存堆叠。
Explore Micron HBM2E from DRAM cells, arrays, and banks to the memory stacks beside a GPU.
Daniel Wang
做些自己觉得有用，也希望你会觉得有用的东西。
Making things I find useful, and hopefully you will too.
完整走一遍 decoder-only Transformer 推理的 prefill 和 decode 阶段，重点理解 Attention 与 KV Cache。
Walk through prefill and decode in decoder-only Transformer inference, with a focus on Attention and KV Cache.
通过简单的 Explore → Build → Reuse 工作流，让现有 Web App 对 Agent 更友好。
A simple Explore → Build → Reuse workflow for making existing web apps agent-friendly.
完全在浏览器中制作透明背景 MOV 视频特效，无需上传或安装，生成后可直接放入 CapCut 或剪映。
Create transparent-background MOV video effects entirely in your browser—no uploads or installs—ready to drop into CapCut or JianYing.
把你在 Chrome 中写下的文本变成学习材料；暂停输入后，文本框下方会自动出现翻译卡片。
Use the text you write in Chrome as your learning material, with a passive translation card appearing below text inputs after you pause typing.
从 AI 回复的任意位置创建分支，每个分支绑定自己的 Agent 或 Chatbot session，把线性聊天变成可探索的对话画布。
Fork from any part of an AI reply, and each branch is bound to its own Agent or Chatbot session, turning a linear chat into an explorable conversation canvas.
把你的 Opencode Agent 嵌入 Obsidian vault，再加上一点我的风格。
Embed your Opencode agent in Obsidian vault, with a bit of my own flavor.
把线性 AI 对话转成树状图，获得自上而下、目标驱动的学习体验。
Transform linear AI conversations into tree-shaped graphs for a top-down, goal-oriented learning experience.
通过简单的 Explore → Build → Reuse 工作流，让现有 Web App 对 Agent 更友好。
A simple Explore → Build → Reuse workflow for making existing web apps agent-friendly.
用一个 Chrome 扩展工具栏图标，集中追踪进行中和未读的 ChatGPT、DeepSeek 与 Claude 对话。
Track working and unread ChatGPT, DeepSeek, and Claude conversations from one Chrome extension toolbar icon.
在设定的空闲时间后自动模糊 Chrome 标签页，并在你回来操作时立即恢复。
Automatically blur inactive Chrome tabs after a timer, then uncover them the moment you return.
完全在浏览器中制作透明背景 MOV 视频特效，无需上传或安装，生成后可直接放入 CapCut 或剪映。
Create transparent-background MOV video effects entirely in your browser—no uploads or installs—ready to drop into CapCut or JianYing.
重新对齐 SRT 与字级 JSON 的时间轴，编辑字幕断句并实时检查改动，之后直接导出 MP4，扩展原有的 ai-jian-koubo 工作流。
Re-time SRT and word-level JSON, edit caption splitting, and review changes live—then export MP4 directly, extending the original ai-jian-koubo workflow.
为口播视频添加语义缩放特效与字幕强调，在本地 review 页面逐项微调后，再由 FFmpeg 渲染。
Add semantic zoom effects and caption emphasis to talking-head videos, then fine-tune every choice in a local review page before FFmpeg rendering.
以美光 HBM2E 为例，从 DRAM 存储单元、阵列和 bank 讲到 GPU 旁的内存堆叠。
Explore Micron HBM2E from DRAM cells, arrays, and banks to the memory stacks beside a GPU.
完整走一遍 decoder-only Transformer 推理的 prefill 和 decode 阶段，重点理解 Attention 与 KV Cache。
Walk through prefill and decode in decoder-only Transformer inference, with a focus on Attention and KV Cache.