Introduction: The Confusion in the Library
Imagine walking into a library to learn AI. The first shelf holds dense textbooks assuming familiarity with functional analysis and probabilistic graphical models. The second shelf holds pop-science pieces like "Understand AI in Three Minutes"—you pick up terminology, but the knowledge slips away when you try to write a single line of code.
This is the dilemma facing today's AI learners: resources are either too deep or too shallow.
Enter ConardLi, a frontend engineer with 8 years of technical writing experience (JavaScript closures, React concurrent rendering, V8 garbage collection, web performance). Starting in 2023, he began writing about AI—not academically, and not as a joke collection, but by using a frontend engineer's mindset to restructure how AI knowledge is presented. The project is called easy-learn-ai.
Chapter 1: Why a Frontend Engineer Explains AI More Clearly
The best person to explain complex concepts isn't an expert (who suffers from the "curse of knowledge") or a layperson (who can't show the path), but someone who has just crossed the threshold—someone who still remembers every pitfall and which analogy finally made things click.
Example: LoRA via the Scaffolding Metaphor
A typical explanation of LoRA:
> LoRA fine-tunes by adding low-rank matrices BA alongside the pretrained weight matrix W, where B and A have dimensions d×r and r×k, r << min(d,k), reducing trainable parameters from O(dk) to O(r(d+k))...
ConardLi's version: Imagine you have a completed building (pretrained model) that you want to convert into a hospital (fine-tuning task). The traditional approach is demolition and rebuild—full-parameter fine-tuning, frighteningly expensive. LoRA's approach: don't touch the main structure; add a few adjustable supports on the outside. These supports are thin (low-rank) but cleverly positioned, and adjusting their angles and lengths lets the building serve its new purpose.
No matrix decomposition, no rank theory—just a picture you can hold in your mind.
Chapter 2: "Easy Understanding" Isn't Simplification—It's Restructuring
easy-learn-ai contains 35+ tutorials forming a deliberately designed learning sequence:
Stop 1: Token — AI's "Alphabet"
Starting with Token because it's the minimal unit through which models read the world. "你好世界" might become ["你好", "世界"] tokens; "unhappiness" might split into ["un", "happiness"]. This explains why models sometimes generate words "broken in the middle" and why effective context lengths differ across languages.
Stop 2: Transformer — Attention Is Everything
When you read a pronoun like "it," your brain automatically scans back for the most relevant noun. Self-attention makes this "looking back" process computable and learnable: each word attends to all previous words, assigning relevance scores. Multi-head attention and positional encoding then become natural extensions of this core intuition rather than magic.
Stop 3: LoRA, RLHF, Quantization — Three Gates of Fine-tuning
- LoRA: the "scaffolding metaphor" — don't rebuild, just add adjustable supports.
- RLHF: the "teacher grading homework" — the model generates answers, human graders score them, and the model iteratively learns what answers humans prefer.
- Quantization: the "image compression" metaphor — store weights at lower precision, like converting PNG to JPEG: smaller, but visually indistinguishable.
- RAG solves hallucination about private data: retrieve first, then generate — find relevant passages in your document store, then feed them to the model.
- Agent solves "answers questions but can't act": via Function Calling, the model gains "hands and feet" — checking weather, writing code, sending email.
- MCP (Model Context Protocol, from Anthropic) solves "too many tools, messy interfaces": like USB unified peripherals, MCP unifies how AI interacts with data sources and tools.
Each metaphor is imperfect but captures the core mechanism.
Stop 4: RAG, Agent, MCP — From Model to System
Chapter 3: From easy-learn-ai to Easy Dataset
ConardLi's other project, Easy Dataset (5k+ GitHub stars), solves a practical problem: building high-quality LLM fine-tuning datasets from scratch. It visualizes the pipeline: upload documents → automatic chunking → AI-assisted question generation → label-tree knowledge organization → batch answer generation (with chain-of-thought) → export to Alpaca or ShareGPT format.
A team from Beihang University even published a paper on it: *Easy Dataset: A Unified and Extensible Framework for Synthesizing LLM Fine-Tuning Data from Unstructured Documents*.
Both projects share one methodology: break complex engineering processes into steps ordinary people can follow. easy-learn-ai teaches the "why" and "what"; Easy Dataset enables the "how".
Chapter 4: The Underlying Logic of "Easy Understanding"
1. Cognitive scaffolding — give a metaphor or picture first, then map its parts to technical concepts. Humans learn top-down, not bottom-up. 2. Concept dependency graph — no skipped prerequisites, no detours. The 35 tutorials are ordered along a minimal dependency map. 3. Engineering perspective — articles like "轻松理解模型部署——Ollama vs vLLM" and "轻松理解 Function Calling" are guides for real engineering decisions, not pure theory.
Epilogue: A Footnote in Knowledge Democratization
easy-learn-ai won't win NeurIPS awards, but it builds a bridge for people locked out of AI knowledge—not because they lack intelligence, but because they lack a suitable entry point. As Feynman said: "If you think you understand something, try explaining it in simple language."
Appendix: Suggested Learning Path
Stage 1 — Fundamentals: Token, Transformer, LLM Stage 2 — Model Principles: GPT (Decoder-Only), BERT (Encoder-Only), LLaMA, DeepSeek R1 Stage 3 — Model Optimization: Quantization, Distillation, LoRA Stage 4 — Engineering Practice: Deployment (Ollama vs vLLM), RAG, Agent, MCP
Project: https://github.com/ConardLi/easy-learn-ai