Why I care
Why the concept belongs in my research, engineering, product thinking, or worldview.
Digital garden
Evolving notes around the concepts that shape my research, engineering, product thinking, and worldview. These are working nodes, not final definitions.
A concept card is not an encyclopedia entry. It is a working snapshot of an idea at its current stage. Sources make it traceable, not automatically authoritative; perspectives, tensions, and a separate sources section remain optional.
Shapes and grounds
Why the concept belongs in my research, engineering, product thinking, or worldview.
Makes facts, definitions, and disagreements traceable; cite close to the relevant claim when possible.
Different roles, scales, or disciplinary lenses through which to view the concept.
The definition and boundaries I am willing to use now, without presenting them as permanent conclusions.
Extends and tests
Places the concept in a knowledge network by relating it to other ideas and practices.
Values or requirements that pull against each other within the concept, not a list of its opposites.
What the current account cannot yet explain and where the note might grow next.
When and why the understanding changed, keeping the revision process visible.
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。
根据损失对参数的梯度反复更新参数,让模型一步步走向损失更小的位置。
先对特征做线性组合,再用 Sigmoid 将分数转换成概率的二分类模型。
把任意实数平滑地压缩到 0 和 1 之间的 S 形函数,常用于二分类模型的输出层。
一个关于 Agent Harness 如何围绕模型组织上下文、工具、状态、权限与反馈循环,使模型成为可行动系统的工作笔记。
一个关于 Agent 如何在环境中围绕目标,根据观察选择行动并利用反馈继续调整的工作笔记。
一个关于 Agent Skills 如何把特定任务的知识、指令与资源封装成可移植工作方法的工作笔记。
一个关于光栅作为用像素网格表达连续视觉世界的表征方式的工作笔记。
A working note on SSE as a simple HTTP-based streaming protocol for one-way server-to-client updates.
A working note on baseline as the reference point that makes an experiment, model, or design decision interpretable.
A working note on daemon as a long-running background process that keeps a capability available without direct foreground interaction.
A working note on fitting as the process where a model becomes informed by data, not one fixed algorithmic procedure.
A working note on protocol as a shared set of rules that makes interaction, comparison, and collaboration predictable.
A working note on turning a vague research interest into a bounded, answerable, and evidence-driven question.
A working note on the capacity to observe, model, regulate, and redesign one's own thinking processes.
A thought experiment about beings that behave like conscious humans while lacking subjective experience.
A working note on the view that mental, cognitive, and social phenomena remain continuous with physical reality rather than floating above it.