Introduction: The Dilemma of the AI Learner
Imagine walking into a library to learn artificial intelligence. One shelf holds dense textbooks that assume you already know functional analysis and probabilistic graphical models. Another holds pop-science pamphlets like "Understand AI in Three Minutes" that leave you with vocabulary but no real understanding. Too deep or too shallow — the middle ground seems missing.
Enter ConardLi, a front-end engineer who has written technical articles for 8 years (JavaScript closures, React concurrent rendering, V8 garbage collection, web performance). In 2023 he began systematically learning AI — and instead of writing academically or as joke compilations, he did something unusual: he rebuilt how AI knowledge is presented, using a front-end engineer's mindset. The project is called easy-learn-ai.
Why a Front-End Engineer Explains AI More Clearly
The best explainers are not experts (who suffer from the curse of knowledge) nor complete outsiders, but people who just crossed the threshold — they remember every pitfall and every metaphor that triggered the "aha" moment.
Example — explaining LoRA (Low-Rank Adaptation):
> A typical explanation: LoRA adds low-rank matrices BA beside 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))...
Correct, but confusing. ConardLi's version: imagine a completed building (the pretrained model) you want to convert into a hospital (fine-tuning). Full fine-tuning demolishes and rebuilds — prohibitively expensive. LoRA instead keeps the main structure untouched and adds a few thin, adjustable supports on the outside — cleverly placed, low-rank, tuned so the building serves its new purpose. Core intuition: don't change the model itself; learn an "adaptation layer."
"Easy Understanding" Is Not Simplification — It's Restructuring
easy-learn-ai contains 35+ tutorials forming a deliberate learning sequence:
1. Token — the smallest unit the model reads, like letters for English. Explains why models sometimes break words mid-generation, why effective context differs between Chinese and English, and why prompts stress comma separation. 2. Transformer — self-attention as the "looking back" your brain does when resolving a pronoun, made computable and learnable. Each word attends to all previous words with different weights, forming a learned semantic-association map. 3. Fine-tuning trio — LoRA (adjustable supports), RLHF (a teacher grading homework so the model learns what humans prefer), quantization (compressing weights like PNG to JPEG — smaller, visually indistinguishable). 4. From model to system — RAG (retrieve first, then generate, fixing hallucinations about private data), Agents (Function Calling as "hands and feet" so the model can act, not just chat), and MCP (Anthropic's open standard unifying AI-tool interactions, like USB for peripherals).
From easy-learn-ai to Easy Dataset
The author also built Easy Dataset, a 5k+ star GitHub tool solving a real pain point: constructing high-quality LLM fine-tuning datasets from scratch. It visualizes the workflow:
- Upload documents and auto-slice them
- AI-assisted question generation
- Tag trees for knowledge organization
- Batch answer generation (with chain-of-thought support)
- Export to Alpaca or ShareGPT formats
Both projects share one methodology: break complex engineering workflows into steps ordinary people can complete one by one. easy-learn-ai teaches the "why" and "what"; Easy Dataset enables the "how."
The Underlying Logic of "Easy Understanding"
Three principles:
1. Cognitive scaffolding — give a metaphor or picture first, then map its parts to technical concepts. Humans learn top-down: big picture, then details. 2. A minimal dependency graph — no concept arrives out of nowhere; the 35-article order follows prerequisite structure without detours. 3. Engineering perspective — articles like "Understanding Model Deployment: Ollama vs vLLM" and "Understanding Function Calling" are guides to real decisions: what to choose, how to configure, how to debug.
Suggested Learning Path
Stage 1 — Fundamentals: Token, Transformer, LLM
Stage 2 — Model principles: GPT (Decoder-Only), BERT (Encoder-Only), LLaMA, DeepSeek R1
Stage 3 — Optimization: Quantization, distillation, LoRA
Stage 4 — Engineering practice: Deployment (Ollama vs vLLM), RAG, Agent, MCP
Epilogue
easy-learn-ai is not a breakthrough research result, but a footnote to knowledge democratization: high-quality AI education is mostly in English and aimed at people with CS backgrounds. This project builds a bridge for everyone else — in Chinese, in plain language, with an engineer's pragmatism. As Feynman said: "If you think you understand something, try to explain it in simple language."
Project: https://github.com/ConardLi/easy-learn-ai