How I Design Claude Code Skills
Published:
A skill is crystallized consensus from repeated experience, not something designed in the abstract. Eight skills and what I learned about knowing what’s worth putting in them.
Posts by Tags
AI
How to Build AI Agents: Lessons from Five Projects
Published:
How much should I specify upfront, and how much should I let the agent figure out? Lessons from five agent projects, grounded in Anthropic’s engineering philosophy.
From Agent Mania to Research Automation: AI Breaking Through the Frontiers of Knowledge
Published:
Models are so powerful now that anyone can bring ideas to life. But what comes after agent mania? A reflection on data flywheels, distillation, and building AI that does research.
Ascend
From Attention to KV Cache Compression: Notes from a KernelGen Optimization
Published:
Starting from Attention and KV Cache, this note walks through KernelGen Challenge 8: a DeepSeek-style KV compression operator, its Triton implementation, and the separate optimization paths for GPGPU and Ascend.
Asterinas
我们如何在 RISC-V 实板上启动 Asterinas
Published:
从 U-Boot 的一次 booti，到早期页表、OSTD、四核启动、rootfs 和 PID 1：复盘我们怎样定义、推进并验证一次真实 RISC-V 开发板上的操作系统启动。
DeepSeek
From Attention to KV Cache Compression: Notes from a KernelGen Optimization
Published:
Starting from Attention and KV Cache, this note walks through KernelGen Challenge 8: a DeepSeek-style KV compression operator, its Triton implementation, and the separate optimization paths for GPGPU and Ascend.
GPU
From SwiGLU Backward to INT8 Quantization: Notes from a KernelGen Challenge 9 Win
Published:
A reconstruction of our KernelGen Challenge 9 optimization: starting from the SwiGLU backward formula, then walking through 128x128 tiling, fused INT8 quantization, memory bandwidth limits, and per-backend Triton engineering.
From Attention to KV Cache Compression: Notes from a KernelGen Optimization
Published:
Starting from Attention and KV Cache, this note walks through KernelGen Challenge 8: a DeepSeek-style KV compression operator, its Triton implementation, and the separate optimization paths for GPGPU and Ascend.
Huffman
形式化证明 Huffman 最优性：我们的 splitLeaf / 交换论证方法
Published:
整理我们在 Lean 4 中形式化 Huffman 编码最优性的方法：把一次贪心合并反过来写成 splitLeaf，用交换论证证明这个逆操作保持最优性，再用 splitLeaf 与 huffman 的交换律把归纳假设传回原始森林。
KV Cache
From Attention to KV Cache Compression: Notes from a KernelGen Optimization
Published:
Starting from Attention and KV Cache, this note walks through KernelGen Challenge 8: a DeepSeek-style KV compression operator, its Triton implementation, and the separate optimization paths for GPGPU and Ascend.
KernelGen
From SwiGLU Backward to INT8 Quantization: Notes from a KernelGen Challenge 9 Win
Published:
A reconstruction of our KernelGen Challenge 9 optimization: starting from the SwiGLU backward formula, then walking through 128x128 tiling, fused INT8 quantization, memory bandwidth limits, and per-backend Triton engineering.
From Attention to KV Cache Compression: Notes from a KernelGen Optimization
Published:
Starting from Attention and KV Cache, this note walks through KernelGen Challenge 8: a DeepSeek-style KV compression operator, its Triton implementation, and the separate optimization paths for GPGPU and Ascend.
Lean4
形式化证明 Huffman 最优性：我们的 splitLeaf / 交换论证方法
Published:
整理我们在 Lean 4 中形式化 Huffman 编码最优性的方法：把一次贪心合并反过来写成 splitLeaf，用交换论证证明这个逆操作保持最优性，再用 splitLeaf 与 huffman 的交换律把归纳假设传回原始森林。
Memory
OSTEP Memory Virtualization Notes
Published:
Notes on memory virtualization from OSTEP—address spaces, page tables, TLBs, and the mechanisms that make virtual memory work.
Milk-V Megrez
我们如何在 RISC-V 实板上启动 Asterinas
Published:
从 U-Boot 的一次 booti，到早期页表、OSTD、四核启动、rootfs 和 PID 1：复盘我们怎样定义、推进并验证一次真实 RISC-V 开发板上的操作系统启动。
MoE
From SwiGLU Backward to INT8 Quantization: Notes from a KernelGen Challenge 9 Win
Published:
A reconstruction of our KernelGen Challenge 9 optimization: starting from the SwiGLU backward formula, then walking through 128x128 tiling, fused INT8 quantization, memory bandwidth limits, and per-backend Triton engineering.
Operating System
OSTEP Memory Virtualization Notes
Published:
Notes on memory virtualization from OSTEP—address spaces, page tables, TLBs, and the mechanisms that make virtual memory work.
Operating Systems
我们如何在 RISC-V 实板上启动 Asterinas
Published:
从 U-Boot 的一次 booti，到早期页表、OSTD、四核启动、rootfs 和 PID 1：复盘我们怎样定义、推进并验证一次真实 RISC-V 开发板上的操作系统启动。
Quantization
From SwiGLU Backward to INT8 Quantization: Notes from a KernelGen Challenge 9 Win
Published:
A reconstruction of our KernelGen Challenge 9 optimization: starting from the SwiGLU backward formula, then walking through 128x128 tiling, fused INT8 quantization, memory bandwidth limits, and per-backend Triton engineering.
RISC-V
我们如何在 RISC-V 实板上启动 Asterinas
Published:
从 U-Boot 的一次 booti，到早期页表、OSTD、四核启动、rootfs 和 PID 1：复盘我们怎样定义、推进并验证一次真实 RISC-V 开发板上的操作系统启动。
SwiGLU
From SwiGLU Backward to INT8 Quantization: Notes from a KernelGen Challenge 9 Win
Published:
A reconstruction of our KernelGen Challenge 9 optimization: starting from the SwiGLU backward formula, then walking through 128x128 tiling, fused INT8 quantization, memory bandwidth limits, and per-backend Triton engineering.
Triton
From SwiGLU Backward to INT8 Quantization: Notes from a KernelGen Challenge 9 Win
Published:
A reconstruction of our KernelGen Challenge 9 optimization: starting from the SwiGLU backward formula, then walking through 128x128 tiling, fused INT8 quantization, memory bandwidth limits, and per-backend Triton engineering.
From Attention to KV Cache Compression: Notes from a KernelGen Optimization
Published:
Starting from Attention and KV Cache, this note walks through KernelGen Challenge 8: a DeepSeek-style KV compression operator, its Triton implementation, and the separate optimization paths for GPGPU and Ascend.
Virtualization
OSTEP Memory Virtualization Notes
Published:
Notes on memory virtualization from OSTEP—address spaces, page tables, TLBs, and the mechanisms that make virtual memory work.
agent
How to Build AI Agents: Lessons from Five Projects
Published:
How much should I specify upfront, and how much should I let the agent figure out? Lessons from five agent projects, grounded in Anthropic’s engineering philosophy.
From Agent Mania to Research Automation: AI Breaking Through the Frontiers of Knowledge
Published:
Models are so powerful now that anyone can bring ideas to life. But what comes after agent mania? A reflection on data flywheels, distillation, and building AI that does research.
ai
After Vibe Code, Vibe Clean: Turning ‘It Runs’ Back Into ‘I Understand’
Published:
After vibe coding comes vibe cleaning—turning a messy but working AI-generated codebase back into something a human can understand, maintain, and extend.
asterinas
当板子沉默时：OS Bringup 的调试方法与工程原则
Published:
OS bringup 的早期阶段几乎没有任何可观测性——一行串口就是全部。这篇文章讨论在板子沉默时如何定位故障、如何设计受控实验、以及从 10+ 轮真机迭代中提炼出的工程原则和反面教训。
从源码到 riscv_boot：为什么「编译成功」离「内核启动」还有一万个坑
Published:
把 Asterinas 移植到 RISC-V 真机时，编译器说一切正常、QEMU 里也跑到了用户态——但真机只输出一行 Starting kernel … 就沉默了。这篇文章从第一性原理出发，逐层拆解构建机、固件、内核三个世界之间的隐秘契约。
biophysics
A Single-Molecule Bound on Cryptochrome Radical Pair Compass Sensitivity and Its Implications for Avian Magnetoreception
Published:
A single-molecule bound on cryptochrome radical pair compass sensitivity—what quantum spin dynamics tells us about the physical limits of avian magnetoreception.
boot
从源码到 riscv_boot：为什么「编译成功」离「内核启动」还有一万个坑
Published:
把 Asterinas 移植到 RISC-V 真机时，编译器说一切正常、QEMU 里也跑到了用户态——但真机只输出一行 Starting kernel … 就沉默了。这篇文章从第一性原理出发，逐层拆解构建机、固件、内核三个世界之间的隐秘契约。
bringup
当板子沉默时：OS Bringup 的调试方法与工程原则
Published:
OS bringup 的早期阶段几乎没有任何可观测性——一行串口就是全部。这篇文章讨论在板子沉默时如何定位故障、如何设计受控实验、以及从 10+ 轮真机迭代中提炼出的工程原则和反面教训。
从源码到 riscv_boot：为什么「编译成功」离「内核启动」还有一万个坑
Published:
把 Asterinas 移植到 RISC-V 真机时，编译器说一切正常、QEMU 里也跑到了用户态——但真机只输出一行 Starting kernel … 就沉默了。这篇文章从第一性原理出发，逐层拆解构建机、固件、内核三个世界之间的隐秘契约。
camera-geometry
相机如何把光变成数据: 从投影、曝光到神经场的统一数学视角
Published:
从一条光线和一次有限测量出发, 把投影几何、镜头曝光、焦点堆栈、光场与光度成像统一成同一个反问题, 再用零空间、Bayes 风险和可辨识性解释深度学习论文为什么这样设计。
claude-code
How I Design Claude Code Skills
Published:
A skill is crystallized consensus from repeated experience, not something designed in the abstract. Eight skills and what I learned about knowing what’s worth putting in them.
computational-optics
相机如何把光变成数据: 从投影、曝光到神经场的统一数学视角
Published:
从一条光线和一次有限测量出发, 把投影几何、镜头曝光、焦点堆栈、光场与光度成像统一成同一个反问题, 再用零空间、Bayes 风险和可辨识性解释深度学习论文为什么这样设计。
computational-photography
从 RAW 到图像: 相机前端的统一数学视角
Published:
这篇文章不把 learned ISP 当成“用神经网络修图”, 而是从 RAW 观测、Bayes 推断、损失函数、结构先验和任务目标出发, 把相机前端统一看成一个带物理模型和渲染偏好的数学决策问题。
相机如何把光变成数据: 从投影、曝光到神经场的统一数学视角
Published:
从一条光线和一次有限测量出发, 把投影几何、镜头曝光、焦点堆栈、光场与光度成像统一成同一个反问题, 再用零空间、Bayes 风险和可辨识性解释深度学习论文为什么这样设计。
concurrency
Rust Concurrency Notes (3): Elegant Fixes for Five Real Bugs
Published:
Rust concurrency notes part 3: five real-world concurrency bugs and their elegant solutions—deadlocks, data races, and performance pathologies caught at compile time.
Rust Concurrency Notes (2): Threads, Locks, Channels, and Concurrency Control
Published:
Rust concurrency notes part 2: threads, locks, channels, and concurrency control primitives—Arc, Mutex, RwLock, Condvar, Barrier, and channel selection patterns.
Rust Concurrency Notes (1): Ownership, Types, and Errors
Published:
Rust concurrency notes part 1: how ownership, type system, and error handling create a foundation for fearless concurrency—Send, Sync, and what the compiler actually enforces.
cryptochrome
A Single-Molecule Bound on Cryptochrome Radical Pair Compass Sensitivity and Its Implications for Avian Magnetoreception
Published:
A single-molecule bound on cryptochrome radical pair compass sensitivity—what quantum spin dynamics tells us about the physical limits of avian magnetoreception.
debugging
当板子沉默时：OS Bringup 的调试方法与工程原则
Published:
OS bringup 的早期阶段几乎没有任何可观测性——一行串口就是全部。这篇文章讨论在板子沉默时如何定位故障、如何设计受控实验、以及从 10+ 轮真机迭代中提炼出的工程原则和反面教训。
Docker Socket Exists but Connection Is Refused: Debugging a snap + apt Double Installation
Published:
docker ps returned ECONNREFUSED but the socket file existed and the daemon was running. A debugging walkthrough through pgrep, strace, and journalctl to find the root cause: zombie container state from a snap + apt dual installation.
deep-learning
从 RAW 到图像: 相机前端的统一数学视角
Published:
这篇文章不把 learned ISP 当成“用神经网络修图”, 而是从 RAW 观测、Bayes 推断、损失函数、结构先验和任务目标出发, 把相机前端统一看成一个带物理模型和渲染偏好的数学决策问题。
相机如何把光变成数据: 从投影、曝光到神经场的统一数学视角
Published:
从一条光线和一次有限测量出发, 把投影几何、镜头曝光、焦点堆栈、光场与光度成像统一成同一个反问题, 再用零空间、Bayes 风险和可辨识性解释深度学习论文为什么这样设计。
deepseek
From Agent Mania to Research Automation: AI Breaking Through the Frontiers of Knowledge
Published:
Models are so powerful now that anyone can bring ideas to life. But what comes after agent mania? A reflection on data flywheels, distillation, and building AI that does research.
docker
Docker Socket Exists but Connection Is Refused: Debugging a snap + apt Double Installation
Published:
docker ps returned ECONNREFUSED but the socket file existed and the daemon was running. A debugging walkthrough through pgrep, strace, and journalctl to find the root cause: zombie container state from a snap + apt dual installation.
greedy
形式化证明 Huffman 最优性：我们的 splitLeaf / 交换论证方法
Published:
整理我们在 Lean 4 中形式化 Huffman 编码最优性的方法：把一次贪心合并反过来写成 splitLeaf，用交换论证证明这个逆操作保持最优性，再用 splitLeaf 与 huffman 的交换律把归纳假设传回原始森林。
infrastructure
Design Dimensions for Research Infrastructure
Published:
Four research infrastructure projects, and what they taught me about failure recovery, data integrity, observability, system abstraction, and the mistakes I’d avoid starting over.
inverse-rendering
相机如何把光变成数据: 从投影、曝光到神经场的统一数学视角
Published:
从一条光线和一次有限测量出发, 把投影几何、镜头曝光、焦点堆栈、光场与光度成像统一成同一个反问题, 再用零空间、Bayes 风险和可辨识性解释深度学习论文为什么这样设计。
isp
从 RAW 到图像: 相机前端的统一数学视角
Published:
这篇文章不把 learned ISP 当成“用神经网络修图”, 而是从 RAW 观测、Bayes 推断、损失函数、结构先验和任务目标出发, 把相机前端统一看成一个带物理模型和渲染偏好的数学决策问题。
kernel
当板子沉默时：OS Bringup 的调试方法与工程原则
Published:
OS bringup 的早期阶段几乎没有任何可观测性——一行串口就是全部。这篇文章讨论在板子沉默时如何定位故障、如何设计受控实验、以及从 10+ 轮真机迭代中提炼出的工程原则和反面教训。
从源码到 riscv_boot：为什么「编译成功」离「内核启动」还有一万个坑
Published:
把 Asterinas 移植到 RISC-V 真机时，编译器说一切正常、QEMU 里也跑到了用户态——但真机只输出一行 Starting kernel … 就沉默了。这篇文章从第一性原理出发，逐层拆解构建机、固件、内核三个世界之间的隐秘契约。
light-field
相机如何把光变成数据: 从投影、曝光到神经场的统一数学视角
Published:
从一条光线和一次有限测量出发, 把投影几何、镜头曝光、焦点堆栈、光场与光度成像统一成同一个反问题, 再用零空间、Bayes 风险和可辨识性解释深度学习论文为什么这样设计。
linux
Docker Socket Exists but Connection Is Refused: Debugging a snap + apt Double Installation
Published:
docker ps returned ECONNREFUSED but the socket file existed and the daemon was running. A debugging walkthrough through pgrep, strace, and journalctl to find the root cause: zombie container state from a snap + apt dual installation.
magnetoreception
A Single-Molecule Bound on Cryptochrome Radical Pair Compass Sensitivity and Its Implications for Avian Magnetoreception
Published:
A single-molecule bound on cryptochrome radical pair compass sensitivity—what quantum spin dynamics tells us about the physical limits of avian magnetoreception.
methodology
当板子沉默时：OS Bringup 的调试方法与工程原则
Published:
OS bringup 的早期阶段几乎没有任何可观测性——一行串口就是全部。这篇文章讨论在板子沉默时如何定位故障、如何设计受控实验、以及从 10+ 轮真机迭代中提炼出的工程原则和反面教训。
How I Design Claude Code Skills
Published:
A skill is crystallized consensus from repeated experience, not something designed in the abstract. Eight skills and what I learned about knowing what’s worth putting in them.
How to Build AI Agents: Lessons from Five Projects
Published:
How much should I specify upfront, and how much should I let the agent figure out? Lessons from five agent projects, grounded in Anthropic’s engineering philosophy.
opensource
Design Dimensions for Research Infrastructure
Published:
Four research infrastructure projects, and what they taught me about failure recovery, data integrity, observability, system abstraction, and the mistakes I’d avoid starting over.
operating-system
从源码到 riscv_boot：为什么「编译成功」离「内核启动」还有一万个坑
Published:
把 Asterinas 移植到 RISC-V 真机时，编译器说一切正常、QEMU 里也跑到了用户态——但真机只输出一行 Starting kernel … 就沉默了。这篇文章从第一性原理出发，逐层拆解构建机、固件、内核三个世界之间的隐秘契约。
optimality
形式化证明 Huffman 最优性：我们的 splitLeaf / 交换论证方法
Published:
整理我们在 Lean 4 中形式化 Huffman 编码最优性的方法：把一次贪心合并反过来写成 splitLeaf，用交换论证证明这个逆操作保持最优性，再用 splitLeaf 与 huffman 的交换律把归纳假设传回原始森林。
production
Rust Concurrency Notes (3): Elegant Fixes for Five Real Bugs
Published:
Rust concurrency notes part 3: five real-world concurrency bugs and their elegant solutions—deadlocks, data races, and performance pathologies caught at compile time.
proof-methodology
形式化证明 Huffman 最优性：我们的 splitLeaf / 交换论证方法
Published:
整理我们在 Lean 4 中形式化 Huffman 编码最优性的方法：把一次贪心合并反过来写成 splitLeaf，用交换论证证明这个逆操作保持最优性，再用 splitLeaf 与 huffman 的交换律把归纳假设传回原始森林。
python
After Vibe Code, Vibe Clean: Turning ‘It Runs’ Back Into ‘I Understand’
Published:
After vibe coding comes vibe cleaning—turning a messy but working AI-generated codebase back into something a human can understand, maintain, and extend.
quantum-biology
A Single-Molecule Bound on Cryptochrome Radical Pair Compass Sensitivity and Its Implications for Avian Magnetoreception
Published:
A single-molecule bound on cryptochrome radical pair compass sensitivity—what quantum spin dynamics tells us about the physical limits of avian magnetoreception.
radical-pair
A Single-Molecule Bound on Cryptochrome Radical Pair Compass Sensitivity and Its Implications for Avian Magnetoreception
Published:
A single-molecule bound on cryptochrome radical pair compass sensitivity—what quantum spin dynamics tells us about the physical limits of avian magnetoreception.
raw
从 RAW 到图像: 相机前端的统一数学视角
Published:
这篇文章不把 learned ISP 当成“用神经网络修图”, 而是从 RAW 观测、Bayes 推断、损失函数、结构先验和任务目标出发, 把相机前端统一看成一个带物理模型和渲染偏好的数学决策问题。
research
From Agent Mania to Research Automation: AI Breaking Through the Frontiers of Knowledge
Published:
Models are so powerful now that anyone can bring ideas to life. But what comes after agent mania? A reflection on data flywheels, distillation, and building AI that does research.
research-methodology
Design Dimensions for Research Infrastructure
Published:
Four research infrastructure projects, and what they taught me about failure recovery, data integrity, observability, system abstraction, and the mistakes I’d avoid starting over.
riscv
当板子沉默时：OS Bringup 的调试方法与工程原则
Published:
OS bringup 的早期阶段几乎没有任何可观测性——一行串口就是全部。这篇文章讨论在板子沉默时如何定位故障、如何设计受控实验、以及从 10+ 轮真机迭代中提炼出的工程原则和反面教训。
从源码到 riscv_boot：为什么「编译成功」离「内核启动」还有一万个坑
Published:
把 Asterinas 移植到 RISC-V 真机时，编译器说一切正常、QEMU 里也跑到了用户态——但真机只输出一行 Starting kernel … 就沉默了。这篇文章从第一性原理出发，逐层拆解构建机、固件、内核三个世界之间的隐秘契约。
rust
Rust Concurrency Notes (3): Elegant Fixes for Five Real Bugs
Published:
Rust concurrency notes part 3: five real-world concurrency bugs and their elegant solutions—deadlocks, data races, and performance pathologies caught at compile time.
Rust Concurrency Notes (2): Threads, Locks, Channels, and Concurrency Control
Published:
Rust concurrency notes part 2: threads, locks, channels, and concurrency control primitives—Arc, Mutex, RwLock, Condvar, Barrier, and channel selection patterns.
Rust Concurrency Notes (1): Ownership, Types, and Errors
Published:
Rust concurrency notes part 1: how ownership, type system, and error handling create a foundation for fearless concurrency—Send, Sync, and what the compiler actually enforces.
skill
How I Design Claude Code Skills
Published:
A skill is crystallized consensus from repeated experience, not something designed in the abstract. Eight skills and what I learned about knowing what’s worth putting in them.
snap
Docker Socket Exists but Connection Is Refused: Debugging a snap + apt Double Installation
Published:
docker ps returned ECONNREFUSED but the socket file existed and the daemon was running. A debugging walkthrough through pgrep, strace, and journalctl to find the root cause: zombie container state from a snap + apt dual installation.
software-engineering
How I Design Claude Code Skills
Published:
A skill is crystallized consensus from repeated experience, not something designed in the abstract. Eight skills and what I learned about knowing what’s worth putting in them.
How to Build AI Agents: Lessons from Five Projects
Published:
How much should I specify upfront, and how much should I let the agent figure out? Lessons from five agent projects, grounded in Anthropic’s engineering philosophy.
From Agent Mania to Research Automation: AI Breaking Through the Frontiers of Knowledge
Published:
Models are so powerful now that anyone can bring ideas to life. But what comes after agent mania? A reflection on data flywheels, distillation, and building AI that does research.
After Vibe Code, Vibe Clean: Turning ‘It Runs’ Back Into ‘I Understand’
Published:
After vibe coding comes vibe cleaning—turning a messy but working AI-generated codebase back into something a human can understand, maintain, and extend.
spin-dynamics
A Single-Molecule Bound on Cryptochrome Radical Pair Compass Sensitivity and Its Implications for Avian Magnetoreception
Published:
A single-molecule bound on cryptochrome radical pair compass sensitivity—what quantum spin dynamics tells us about the physical limits of avian magnetoreception.
systems
How to Build AI Agents: Lessons from Five Projects
Published:
How much should I specify upfront, and how much should I let the agent figure out? Lessons from five agent projects, grounded in Anthropic’s engineering philosophy.
Design Dimensions for Research Infrastructure
Published:
Four research infrastructure projects, and what they taught me about failure recovery, data integrity, observability, system abstraction, and the mistakes I’d avoid starting over.
systems-programming
Rust Concurrency Notes (3): Elegant Fixes for Five Real Bugs
Published:
Rust concurrency notes part 3: five real-world concurrency bugs and their elegant solutions—deadlocks, data races, and performance pathologies caught at compile time.
Rust Concurrency Notes (2): Threads, Locks, Channels, and Concurrency Control
Published:
Rust concurrency notes part 2: threads, locks, channels, and concurrency control primitives—Arc, Mutex, RwLock, Condvar, Barrier, and channel selection patterns.
Rust Concurrency Notes (1): Ownership, Types, and Errors
Published:
Rust concurrency notes part 1: how ownership, type system, and error handling create a foundation for fearless concurrency—Send, Sync, and what the compiler actually enforces.