If you've browsed AI concept learning sites over the past few years, the recipe is familiar: top navbar + hero gradient background + three tabs for "Core Concepts / Use Cases / Learning Path" + a few cards under each tab. It's not unusable, but viewing ten sites feels the same as viewing one.
easy-learn-ai recently did something drastic: it tore down and rebuilt every sub-site. Not a reskin—a DNA change.
The Old Problem: Product-Whitepaper Tone
The old ai-sites/* sub-sites shared several flaws:
- Multi-tab structure — felt like a product manual, not learning material
- Gradients + emoji + framer-motion — decoration over information
- "Let's start your X journey together" — every site opened the same way
- Templated sections — always intro, then core concepts, then use cases, in that order
- Data stuck in 2023–2024 — stale model version numbers and benchmark scores
- Every section has ≥ 2 dynamic elements (hover doesn't count—that's basic politeness)
- ≥ 3 sections site-wide reach L3+ (drag-to-edit / real-time reflection)
- Adjacent sections can't use the same interaction type
- ≥ 5 distinct interaction forms per site
In one sentence: it wasn't teaching concepts, it was selling them.
The New Solution: Five Soul-Searching Questions
The new design's first step wasn't writing code—it was answering five questions:
1. What is the most direct one-sentence definition of this concept? — The hero's first line, as a complete declarative sentence 2. How do you explain it in 30 seconds to someone who's never heard of it? — Two or three conversational sentences, no paragraph longer than two sentences 3. What's the most counterintuitive fact? — Not the hero, but the hook, placed at the start of section two 4. If you could change only one thing, what must it be? — The core action, the site's visual anchor 5. What did the previous site look like, and which two things must change this time? — The anti-template red line
For example, the Agent site's hero now reads:
> "What is an AI Agent?" > An AI Agent = a program that can perceive its environment, make autonomous decisions, and invoke tools to accomplish goals.
No metaphors, no rhetorical questions, no "imagine this." First a completable equation, then the expansion.
The New Discipline: Plain Language + Interaction + Real Data
Plain-Language Discipline
Three self-check questions: 1. Would you say this at a product launch? If yes → cut it 2. Does a big word explain something small? If yes → cut it 3. Does the colloquial version lose much? If a lot → use the colloquial version
> "Empowering enterprise-level intelligent decision-making" → "Companies don't have to dig through docs every time" > "Exploring the core principles and application value of X" → "How does X actually work?"
Interaction Discipline
Data Discipline
All data must be real and verifiable in 2026. Model version numbers, benchmark scores, product release status → search before writing; every data point should have a source you can cite immediately.
Which Sites Were Rebuilt
This round of refactoring covered:
| Sub-site | Topic | New Features | |---|---|---| | agent | AI Agent | Five building blocks, ReAct loop, evaluation systems | | llm | Large Language Models | In-context learning, emergent abilities, instruction-following animation | | rag | Retrieval-Augmented Generation | Building blocks, three-step walk, ecosystem | | distill | Knowledge Distillation | Dark knowledge, teacher perspective, distillation lineage | | moe | Mixture of Experts | Router, load balancing, cost gallery | | quantization | Quantization | Number-line journey, method map, ecosystem fit | | function-calling | Function Calling | Schema editing, parallel calls, MCP comparison | | batch-size | Batch Size | Effective batch, one update, training vs inference | | bert | BERT | Bidirectional masking, multi-head attention, still alive in 2026 | | deepseek-r1 | DeepSeek-R1 | GRPO, distillation, Aha Moment | | deepspeed | DeepSpeed | ZeRO, Offload, 3D parallelism | | loss | Loss Functions | Curve doctor, regression, cross-entropy | | mcp | MCP Protocol | Pre-MCP era, basic primitives, triangle architecture | | nlp | NLP | Word embeddings, tokenizers, paradigm differences | | rlhf | RLHF | Reward hacking, KL leash, where do you rank |
Every site is a single scrolling page—no tabs, no header/footer, no "module navigation." One long page is everything.
The Harshest Rule
> If a site's structure is 70% similar to the previous site = red line. Restructure.
This means the BERT site and the loss site can't both be "intro → core → applications → summary." BERT might use a "timeline" narrative; loss might use "pain → solution → cost." Every concept must grow its own unique skeleton.
---
This reminds me of a design principle: good teaching isn't putting knowledge into a pretty container—it's making the container itself part of the knowledge. When you scroll through a sub-site, the interaction rhythm, information density, and visual anchors you experience are themselves an answer to what the concept "looks like."
#easy-learn-ai #daily-update #memory #xiaokai