开始之前
先找到自己当前的学习位置,弄清怎么学才真正过脑子,再进入正式课程。
- 我们在哪里
- 怎样学才有效
- 为什么要打基础
From first principles to agents, in one coherent course.
157 core topics, 398 interactive lessons, 3 languages: build AI intuition, then understand models, products, coding and agents.
Study the core course in order, then choose independent chapters and bonus material as needed.
先找到自己当前的学习位置,弄清怎么学才真正过脑子,再进入正式课程。
从零建立 AI 直觉:它能做什么、为什么会错、怎样沟通、哪些任务可以交给它。
串起训练数据、Token、GPT 和幻觉,建立判断模型边界的底层框架。
学习上下文、Prompt、安全、Agent、工具调用和成本,让大模型从聊天变成可用系统。
沿真实 Agent 应用拆解从 Demo 到产品的关键环节:循环、记忆、协作、权限与 MCP。
拆生产级 Agent 的设计模式,重点看上下文、工具、评测、长运行和安全边界。
理解 Harness 如何让系统持续改进,从固定流程走向能优化自身协作方式的 Agent。
建立一套可复用的 AI 协作流程:目标、上下文、验收、安全闸门和文档沉淀。
训练判断好不好看的基本能力,把层级、留白、克制和一致性翻译成 AI 能执行的要求。
从状态、防错、流程和控件选择入手,让 AI 做出的界面不只是能跑,还能顺手使用。
理解用户为什么觉得 AI 产品好用或难用,把等待、信任、错误和付费感受设计进体验。
看懂 Token 成本和模型计费,再从语法、语义、架构、输出四层压缩浪费。
用 AI 场景理解数据结构:数组、缓存、检索、图和队列分别解决什么问题。
用大模型里的真实机制理解复杂度、排序、分治、图搜索和采样,补足工程直觉。
可选源码深潜:读 Grok Build 的工程结构,理解一个生产级 Coding Agent 怎么组织。
可选源码深潜:读 DeepSeek Harness 的插件内核、会话循环、沙箱和可重建日志。
从权重、许可证、蒸馏到本地部署,看懂开源模型到底开放了什么,以及自己怎么跑起来。
用章节自测检查概念是否真的会了,发现薄弱点后回到对应课程补齐。
AI 产品进入真实世界前,先补确权与合规基础:商标、软著、专利、域名和股权。
产品做出来后,学习如何让搜索引擎和回答型 AI 看得懂、找得到、愿意引用。
正课后的创业方法补充:产品、口碑、融资、估值、股权和现金流。
Five routes scope the catalog by goal. Choose one, the course contents trim down, and you can switch anytime without losing progress.
This path skips theory and engineering, and focuses on using AI well: what it is doing, why it invents, how to ask, and what you can safely hand off. It still keeps the practical Harness themes and a few collaboration lessons for everyday work.
This path adds the full LLM fundamentals part and the full collaboration-methods part, plus two programming-basics lessons on vocabulary and vectors. You get judgment about model limits without entering engineering implementation.
This path adds the full Harness set, design patterns, evaluation, cost engineering, and the self-test center. It keeps product-side slices of the practicum and leaves code walkthroughs, long-running agents, sandboxing, and source-code electives to the build path.
This path takes every core lesson, including full code walkthroughs and self-improvement. Follow the hands-on track end to end and you leave with a working Agent plus your own collaboration playbook.
This adds Grok Build source anatomy, DeepSeek Harness internals, and open-source model distillation plus local deployment beyond the build-it-yourself path. These are the hardest sections and the biggest differentiator.
Unsure which track fits? That is fine. All five start from the same lesson; two lessons in, the right depth usually becomes clear.