About me
MoreI build AI-native applications, developer tools, and human-centered systems that connect research, interaction, and real-world implementation.
My current question: how can AI become part of human memory, work, research, and everyday thinking?
Currently working on
Research
I use multimodal data to estimate and analyze human states, exploring how to detect and understand changes over time.
Engineering
Full-stack prototypes, local AI workflows, and tools that move quickly from idea to real use.
Personal projects
Starting with what I want to use every day, I build and refine tools that support everyday life and learning.
Currently using
Point at one, or throw it around
Latest Activities
2026
Nov
Benjamini–Hochberg 方法(BH 校正)错误发现率(FDR)p 值
Knowledge base
Sep
Knowledge base从有限增长模型走到概率接口与神经网络激活函数的 S 形函数。
Aug
Activation Functions: Why a Network Needs a Bend
Essays深度学习:Basics · Chapter 2 of 12A linear model can rank, but it can never enclose. Starting from a rule as ordinary as "comfortable between 18 and 24 degrees", this post shows what one straight line cannot do, and how a single bend fixes it.
Essays从公开 API 的输入、推理、结构化输出、工具、状态、实时与异步接口出发,重新理解现代模型的能力,以及模型与平台之间的边界。
Knowledge base衡量二分类概率预测与真实标签有多不一致的损失函数,会显著惩罚自信但错误的预测。
Essays深度学习:Basics · Chapter 1 of 12从规则、数据与参数的边界出发,沿人工神经元的历史走到一个可训练的 Sigmoid 神经元。
Essays当系统开始解释 · Chapter 2 of 18从批处理、分时系统、命令行、图形界面、Web 到移动设备,追踪人与计算机如何从提交任务走向连续操作。
Essays当系统开始解释 · Chapter 3 of 18当自然语言成为计算入口,人与系统的关系如何从调用预设功能,转向对表达、意图与共同任务的持续解释。
Essays当系统开始解释 · Chapter 1 of 18从“交互”、interaction 与日语中的「交互」「相互作用」「インタラクション」出发,追溯交互如何从彼此作用生长为一门设计学科。
管理 Agent Skills:为什么我关掉了大部分全局 Skills
Essays从 Agent Harness 与 Codex、Claude Code、Kimi Code 的加载机制出发,讨论 Vibe Coding 中用户级与项目级 Skills 的边界,以及模型变强之后,Skills 还留下什么价值。
Agent Harness智能体 Agent智能体技能 Agent Skill
Knowledge base
