大模型 · Harness · Agent LLM · Harness · Agent
三个词，说清这两年 AI 为什么突然"能干活了"——以及如何用这些技术，加速你的射电天文数据处理与研究。 Three words that explain why AI suddenly "gets things done" these past two years — and how to use them to speed up your radio-astronomy data processing and research.
一句话读懂三概念 Three Concepts in One Line Each
① 大模型 LLM
会思考的"大脑"：在海量文本上训练，擅长理解与生成，但不会动手。
The thinking brain: trained on oceans of text, great at understanding and generating, but has no hands.
进入 → Enter 🛠️② Harness Harness
给大脑装上的"身体 + 工具箱"：文件、代码执行、沙箱、权限、循环与评测。
The "body + toolkit" for the brain: files, code execution, sandbox, permissions, loops, evaluation.
进入 → Enter 🤖③ Agent Agent
带着目标、自主循环、把事做完的系统——模型 + Harness + 工具 + 目标。
The system that pursues a goal autonomously — model + harness + tools + goal.
进入 → Enter站点导览 Site Map
1 · 大模型 What is an LLM
定义、工作原理（Token 预测演示）、能力与局限、代表模型。
Definition, how it works (interactive token demo), capabilities & limits, representative models.
2 · Harness What is a Harness
组件全景图、裸模型对比、MCP 协议、DeepSeek Harness 实例。
Component map, bare-model comparison, MCP, a DeepSeek Harness case study.
3 · Agent What is an Agent
ReAct 循环动画、Agent 类型、代表产品、风险与边界。
Animated ReAct loop, agent types, representative products, risks & boundaries.
4 · 主流 AI 编程工具 Popular AI Coding Tools
国内推荐（Qoder / Trae / Kimi Code / WorkBuddy）与国外简述（Claude Code / OpenCode 等）；从编程 Agent 到通用 Agent。
Domestic picks (Qoder / Trae / Kimi Code / WorkBuddy) & foreign briefs (Claude Code / OpenCode…); from coding to general agents.
4+ · 技能生态 Skills & Ecosystem
Skill / 插件 / MCP / 工作流与 SOP——未必要写代码，把流程写清楚同样有价值。
Skills, plugins, MCP, workflows & SOPs — documenting a clear process counts as programming.
5 · 射电天文实战 ⭐ Radio Astronomy in Practice
端到端案例：HI 谱线立方体 → 矩图 → CASA 交叉验证 → 科学图，每步配提示词模板。
End-to-end case: HI cube → moment maps → CASA cross-check → figure, with prompt templates.
6 · 学习资源 Resources
官方文档、公开数据（THOR / HI4PI / ALMA）、论文与学习路径。
Official docs, public data (THOR / HI4PI / ALMA), papers, learning path.
考一考：你学会了吗？ Quick Quiz: Did You Get It?
5 道题，选完即看答案。答对 4 题以上就大胆开始用 AI 编程工具吧！
Five questions, instant feedback. Score 4+ and you're ready to start using AI coding tools.