Learning Paths: Three Routes

The one-line definition: the 51 pages of this handbook are not a linear textbook to be read cover to cover, but a roadmap you can combine on demand and jump around in at any time — all you need to know is where you're headed, then pick up the pages along the matching route below. Every page is ordered and organized to serve "building context fast," not "memorizing in strict sequence."

Why a roadmap rather than a textbook? Because the world of hot AI concepts has extremely high information density and loosely coupled pages: you can read Transformers and Attention without first finishing A Brief History of the AI Boom, and you can build a RAG application without clearing every concept page first. A textbook assumes readers have long blocks of time and a uniform set of prerequisites; a roadmap assumes readers have different goals, different schedules, and different stocks of existing knowledge. This page is the overview of that roadmap — and the unified index of every page on the site.

Below are three routes for "three kinds of readers": the job-hunting sprint (2 weeks), systematic deepening (8 weeks), and the desk reference (forever). You can enter through any of them and switch lanes along the way — they share the same starting point and converge on the same finish line.

1. The Three Routes at a Glance

Route	Duration	Who it's for	Core goal	Main sections
Route 1 · Job-Hunting Sprint	About 2 weeks	Anyone interviewing within 1–3 months	Cover the high-frequency interview topics in minimal reading, and train until you can explain, write, and field questions	Concepts → Case studies → Career
Route 2 · Systematic Deepening	About 8 weeks	Beginners and career changers who want a complete knowledge system	Concepts, case studies, hands-on practice, and papers as one integrated whole	Primer → Concepts → Practice → Papers
Route 3 · Desk Reference	On demand (permanent)	People already doing the work who look things up as problems come up	The fastest path to the page that solves the problem	Resources + Concepts + Practice

The relationship among the three routes fits in one diagram:

            ┌─────────────────────────────────────────┐
            │     Shared start: /guide/what-is-ai     │
            └───────────────────┬─────────────────────┘
                                │
          ┌─────────────────────┼─────────────────────┐
          ▼                     ▼                     ▼
Route 1 · Job sprint  Route 2 · Systematic   Route 3 · Desk ref
 (2 weeks, density)   (8 weeks, structure)   (on demand, speed)
          │                     │                     │
          └─────────────────────┼─────────────────────┘
                                ▼
       Shared finish: /practice/pitfalls + /resources/glossary

The one-line decision

Short on time? Take Route 1. Thin foundations? Route 2. Already doing the work? Route 3. None of the three routes asks you to read all 51 pages — each only asks for "the part your current goal needs," and the rest of the pages will always be there to pull on demand.

2. Route 1: The Job-Hunting Sprint (About 2 Weeks)

1. Who this route is for

• Anyone with interviews coming up within 1–3 months for roles such as LLM algorithms, AI applications, or AI products;
• Anyone with some engineering/algorithm foundation who, when the interviewer probes LLM concepts, "can run the code but can't explain the principles";
• Anyone who needs to quickly assemble three weapons: a conceptual skeleton, case-study ammunition, and answer frameworks.

The logic of this route is high-frequency topics first: establish the concept boundaries (so you don't answer the wrong question), then take down the 5 concept pages that show up most often in interviews, then use 2 classic case studies to anchor your arguments, and finally use the question bank to convert knowledge into answering ability.

2. Week 1: The conceptual skeleton

Spend the first 2 days on What Are AI Hot Concepts? and AI vs ML vs DL vs GenAI vs Agents to hammer the site-wide map and the five concept boundaries into your head — they are the coordinate system for every answer that follows; over the remaining 5 days, take down 5 core high-frequency topics in order:

Day	Topic	Required reading	Same-day check
D1	Landscape and boundaries	What Are AI Hot Concepts?, Concept Boundaries	Can state in one sentence how AI/ML/DL/GenAI/Agents nest inside one another; can say what each of the site's 7 sections covers
D2	Large language models	Large Language Models (LLMs)	Can explain the pretraining → fine-tuning → inference three-stage pipeline; can name 2 mainstream models and their parameter scales
D3	Transformer	Transformers and Attention	Can sketch the Self-Attention structure from memory and spell out the roles of Q/K/V
D4	Prompt engineering	Prompt Engineering	Can write a prompt for a real task and justify every design decision in it
D5	RAG	Retrieval-Augmented Generation (RAG)	Can draw the retrieve → augment → generate pipeline; can name two sources of hallucination and two mitigation techniques
D6	Agent	AI Agents	Can explain how an agent differs from "an LLM that chats"; can name the three components: tool calling, planning, memory
D7	Review and gap-checking	Re-read your weak pages + Glossary	Can re-explain each of the 5 topics on a whiteboard in 3 minutes; scores 8 out of 10 on a spot quiz

3. Week 2: Case studies and the question bank

With the conceptual skeleton in place, this week turns "abstract concepts" into "case studies with names," then organizes the cases into reusable interview answers:

Day	Topic	Required reading	Same-day check
D8	Case study: ChatGPT	ChatGPT and Conversational AI	Can explain, in a technology → product → business structure, why ChatGPT ignited this wave
D9	Case study: DeepSeek-R1	DeepSeek-R1 and Reasoning Models	Can explain how a reasoning model differs from an ordinary LLM, and what role reinforcement learning plays
D10–11	Mapping against the JD	JD Knowledge-Point Breakdown	Build your own topic-coverage matrix against the target JD, flag 3 weak spots, and close them one by one
D12–13	Question drilling	Interview Question Bank	Complete 8–10 questions per day; for the hard ones, write full prose answers instead of bullet points
D14	Mock interview and final check	Site-wide self-check + Resume Analysis	Run a full 45-minute mock interview; finish revising your resume line by line against the JD

4. The universal daily check

Whichever page you read, run this self-check before you stop for the day — "I've read it" is not the bar; "I can explain it" is:

Check	Passing standard	How to self-test
Whiteboard re-explanation	Explain the day's topic to a "non-technical friend" in 3 minutes	Record yourself and play it back: you get through it without getting stuck
Structural sketch	Draw the day's architecture/pipeline diagram from memory	Compare your drawing against the original page, detail by detail
Terminology	Define every term on the day's page in one sentence	Pick 5 at random and quiz yourself
Question quota	Complete the day's interview/practice questions	Log them on a tracker; make up any missed day right away

30 minutes of daily "Feynman re-explanation"

In Week 1, spend 30 minutes a day on one thing: explain the day's topic to an imaginary interviewer, then go back to the page and check "what did I leave out?" During this skeleton-building phase, the cardinal sin is "the eyes learned it but the mouth can't say it" — interviews test the mouth.

3. Route 2: Systematic Deepening (About 8 Weeks)

1. Who this route is for

• Career changers, students, and early-career engineers who want to build a complete knowledge system from scratch;
• People with relatively ample time (about 10–15 hours a week) who are willing to attack concepts, case studies, practice, and papers on four fronts at once;
• People whose goal is a complete system, not the fastest possible job readiness.

This route moves through the rhythm "primer builds the map → concepts lay the foundation → generation and optimization add depth → case studies → hands-on practice → close reading of papers," across 8 weeks, with a clear deliverable and a check for every week.

2. The 8-week schedule

Week	Phase	Required reading	Weekly check
W1	Primer · Landscape	What Are AI Hot Concepts?, Concept Boundaries	Can hand-draw a map of the whole site: 7 sections, all 51 pages in their places; can articulate the boundaries of the five concept pairs
W2	Primer · History and architecture	A Brief History of the AI Boom, Anatomy of a Modern AI System	Can list the 5 milestones of the brief history; can draw the system architecture diagram of "data → training → inference → applications"
W3	Concept foundations (I)	Large Language Models (LLMs), Transformers and Attention, Multimodal Models	Can sketch the Transformer structure; can explain what LLMs and multimodal models share and where they differ
W4	Concept foundations (II)	Prompt Engineering, RAG, AI Agents	Can draw the RAG pipeline and the agent component diagram; can describe how the three divide the work and fit together
W5	Generation and optimization (I)	Diffusion Models and Generative AI, Fine-Tuning and PEFT	Can explain how diffusion models denoise; can explain why LoRA saves compute
W6	Generation and optimization (II)	Alignment, Inference Optimization	Can compare RLHF and DPO — similarities and differences; can name several typical quantization approaches and their costs
W7	Case studies	Pick 4–6 pages from ChatGPT, DeepSeek-R1, Perplexity and AI Search, Manus and Agent Applications, GitHub Copilot, and Midjourney and Image Generation	Write a technology → product → business three-part note for each case, building your own case library
W8	Hands-on practice	Build a RAG Application from Scratch, Build an Agent from Scratch, Fine-Tune Your Own LLM	Complete at least one runnable project and write a reproducible README
W9	Paper reading and gap-checking	Reading Paths, Classic Papers in Depth, The Paper Map	Read 2 classic papers closely and write a "problem–method–result" card for each

Note: the table has 9 rows because "case studies" flexibly spans two weeks by topic. Grouped by phase: W1–W2 are the primer and concept foundations, W3–W4 bridge concepts with generation and optimization, W5–W6 are case studies, W7 is hands-on practice, and W8 is paper reading and gap-checking. Better to slow down by a week or two than to skip around — the whole value of systematic deepening is "connecting the full chain end to end."

3. Fixed weekly rituals

Ritual	Frequency	Notes
Concept re-explanation	3 times a week	Explain the week's topic to someone, or record yourself; 10 minutes or more
Term drilling	Daily	Drill 10 terms from the Glossary, by letter or by theme
Practice log	Weekly	Record the week's deliverables; after W8 they can be polished into a portfolio
Blind-spot list	Weekly	Note the questions you "couldn't explain clearly"; close them all in W9

The most common mistake in systematic deepening

Bookmarking instead of reading, reading instead of re-explaining, re-explaining instead of building. In the 8-week route, concept pages account for more than half, but they are only "raw material"; what actually internalizes the knowledge is the W7 case notes and the W8 runnable project. If you catch yourself three weeks in having read nothing but concepts and written not a line of code, switch immediately to hands-on practice for a catch-up round.

4. Route 3: The Desk Reference

1. Who this route is for

• People already working on AI applications, productivity gains, technology selection, or interview prep who want the fastest possible answer to a concrete problem;
• People who don't need a weekly schedule and just want to know "which page should I read for this question?"

The desk reference has no order, only an index. When a problem comes up, scan the table below and jump straight to the hit; if that page uses a term you don't know, follow the link to the next page — that is exactly what this handbook's "dense cross-linking" is designed for.

2. Problem → page quick index

Your question	Go-to page	Why read it
The model keeps making things up (hallucinations)	Retrieval-Augmented Generation (RAG)	RAG suppresses hallucinations by wiring in external sources of fact
How to make the model do what I want	Prompt Engineering, Alignment	Prompt design plus alignment training, working both ends
How to hook up my own private data	RAG, Vector Databases and Semantic Search	The standard pairing of retrieval + vectorization
How to build an agent	AI Agents, Build an Agent from Scratch	Conceptual framework + hands-on implementation
How to choose a model	Models and Leaderboards Quick Reference	Mainstream models, parameter counts, and leaderboard comparisons
What interviews ask	Interview Question Bank	High-frequency questions and answer frameworks
Want to truly understand the Transformer	Transformers and Attention	The foundation beneath every large model
The model runs too slowly or costs too much	Inference Optimization and Quantization, Deployment in Practice	Optimization techniques and production engineering
How to evaluate whether a model is any good	LLM Evaluation and Benchmarks, Build an LLM Eval Suite	Metrics + evaluation as engineering
Want to fine-tune my own model	Fine-Tuning and PEFT (LoRA), Fine-Tune Your Own LLM	Concepts + the full workflow hands-on
Want to add image/video generation to a product	Diffusion Models and Generative AI, The Midjourney Case, The Sora Case	Principles + business cases
Want to understand AI safety and compliance	AI Safety and Governance	Risks, alignment, and governance frameworks
How to make my resume stand out	Resume Analysis	Translating knowledge points into JD language
Where to find datasets and tools	Datasets and Tools Directory	An index of data and tools
Which paper to start with	Reading Paths, The Paper Map	Layered reading order and the paper genealogy
What roles the market is hiring for right now	The JD List	Open AI roles at big tech firms at home and abroad
Stuck on a term	Glossary	The site-wide dictionary, always at hand

Even quick lookups deserve a proper close

The trap of desk-reference use is "look it up, leave, forget it." After finishing any page, spend 3 minutes following one or two links in that page's Further Reading section — those are "next stops" the author planted in advance, and they're usually the way into your deeper problem.

5. The Shared Start and Finish of the Three Routes

1. The shared starting point

Whichever route you take, the first page is What Are AI Hot Concepts?. This is not a formality — it's about efficiency:

• It defines the coordinates of all 51 pages on the site: what each of the 7 sections holds and how the concepts are organized;
• It sets the "compass" for everything you read afterward — only by knowing where you are on the whole map can you avoid getting lost in the cross-links.

After that page, the three routes fork: the job-hunting sprint jumps straight to Concept Boundaries to build its coordinate system; systematic deepening continues with A Brief History of the AI Boom and Anatomy of a Modern AI System; the desk reference goes straight to the quick index above.

2. The shared finish line

The closing moves of all three routes converge on the same set of "pitfall-proofing + permanent companion" pages:

Closing item	Page	Purpose
Final pitfall check	Common Pitfalls and Antipatterns	Once a route's concepts are read and its projects written, use this page to self-check for the most common mistakes
Permanent dictionary	Glossary	A lifelong companion: come back whenever you hit an unfamiliar term
Capstone reading	Classic Papers in Depth	The job-hunting sprint may skip it, but systematic deepening and serious practitioners close with it
Final job check	Module Guide and Role Landscape	Only job seekers need it, as the final loop-closer for resume and interviews

3. Switching lanes at any time

The three routes are not isolated classrooms; they are three well-traveled paths on the same map. A few typical lane-change scenarios:

• While using the desk reference, you realize you can't even explain the concepts → pause the lookups and switch to W3–W4 of systematic deepening to shore up the concept foundations;
• During the job-hunting sprint, an interviewer pushes on a "why" question → temporarily switch to systematic deepening's paper week to add depth;
• Midway through systematic deepening, an interview opportunity lands → compress the schedule into a 2-week job-hunting sprint, then come back and continue where you left off.

The one-line rule: routes are starting points, not shackles; being able to combine them on demand is what using this roadmap well really means.

6. A One-Line Index of Every Page on the Site

Below is a one-line index of all 51 pages on the site, grouped by the 7 major sections. One line per page — it doubles as "one sentence per page" and as an extension of Route 3: when you don't know which page a given kind of question belongs to, scan here first.

Primer (/guide/)

Page	One-line summary
Learning Paths: Three Routes	This page: the site-wide roadmap and unified index you are reading now
What Are AI Hot Concepts?	The panoramic overview: the map of hot concepts and site-wide navigation; the starting point of every route
AI vs ML vs DL vs GenAI vs Agents	Boundary analysis of five high-frequency concept pairs; the coordinate system for interview answers
A Brief History of the AI Boom	A 70-year timeline from symbolism to large models
Anatomy of a Modern AI System	What modules make up a modern AI system and how they fit together

Core Knowledge (/concepts/)

Page	One-line summary
Large Language Models (LLMs)	What large models are, what they can do, and what limits them
Transformers and Attention	The attention mechanism and the shared foundation of GPT/BERT
Multimodal Models	Unified understanding and generation across text + images + audio + video
Prompt Engineering	Steering models without writing code: the methodology of prompt design
Retrieval-Augmented Generation (RAG)	The mainstream way to attach an external knowledge base to a model and suppress hallucinations
AI Agents	Autonomous systems that can plan and call tools
Vector Databases and Semantic Search	The ground RAG stands on: vectorization and similarity search
Knowledge Graphs and Knowledge Injection	Using structured knowledge to fill the model's blind spots
Diffusion Models and Generative AI	The core technology behind mainstream image/video generation
Fine-Tuning and PEFT (LoRA)	How to adapt open-source models on a budget
Alignment: RLHF and DPO	Training methods that make models obedient, helpful, and harmless
Inference Optimization and Quantization	Making large models run faster and cheaper
LLM Evaluation and Benchmarks	How to measure model quality scientifically
AI Safety and Governance	Risks, alignment, compliance, and responsible AI

Case Studies (/case-studies/)

Page	One-line summary
ChatGPT and Conversational AI	The product that ignited this wave and the engineering behind it
DeepSeek-R1 and Reasoning Models	Training models that "think" with reinforcement learning
Midjourney and Image Generation	The benchmark for commercializing text-to-image
Sora and Video Generation	World models and the frontier of video generation
Speech AI: Whisper and TTS	Classic engineering in speech recognition and speech synthesis
GitHub Copilot and Code Intelligence	How large models are changing the craft of programming
Manus and Agent Applications	The product form that moves from "chatting" to "doing the work for you"
Perplexity and AI Search	The template for shipping RAG in a real product
AlphaFold and AI for Science	A scientific milestone in predicting protein structures
Recommendation Systems in the Era of Large Models	How generative AI is reshaping recommendation

Papers (/papers/)

Page	One-line summary
Paper Reading: Start Here	The primer for the papers section and how to enter it
Reading Paths	Paper reading sequences tiered by level
The Paper Map	The genealogy of the key papers and how they cite one another
Classic Papers in Depth	Close readings of must-read papers such as Attention Is All You Need
Frontier Advances	New directions from the past two years worth watching
Reading Discipline and FAQ	How to read papers so the effort actually pays off

Practice Guides (/practice/)

Page	One-line summary
Build a RAG Application from Scratch	The full end-to-end build of a RAG application
Build an Agent from Scratch	Implementing tool calling, planning, and memory
Fine-Tune Your Own LLM	Complete hands-on LoRA fine-tuning
Build an LLM Eval Suite	Turning evaluation into sustainable engineering
The Prompt Playbook	A field checklist of prompting techniques
Deployment and Inference Optimization in Practice	Shipping models, load testing, and performance tuning
Common Pitfalls and Antipatterns	The potholes others fell into first — reading it saves you six months of tuition

Careers and Roles (/career/)

Page	One-line summary
Module Guide and Role Landscape	The entrance to and overall map of the careers section
The JD List	AI-related openings at big tech firms at home and abroad
JD Knowledge-Point Breakdown	The topic map hidden behind the JD
Resume Analysis	What your resume should highlight and how to benchmark it against the role
Interview Question Bank	High-frequency interview questions and answer frameworks

Resources (/resources/)

Page	One-line summary
Glossary	The quick-reference dictionary for every term on the site
Curated Resource List	A hand-picked collection of tutorials, blogs, and tools
Datasets and Tools Directory	First-hand records of datasets and tools
Models and Leaderboards Quick Reference	Side-by-side comparisons of mainstream models and leaderboards

How to get the most out of this index

The index tables are "table-of-contents entries": scan quickly for your target, click through, and read the body. The in-body site links are "exploratory entries": follow the natural connections between concepts. The two working together is how this roadmap is meant to be used.

Further Reading

• What Are AI Hot Concepts? — the unified starting point for all 51 pages; every route enters here
• AI vs ML vs DL vs GenAI vs Agents — the definitive boundary analysis; required reading on D1 of the job-hunting sprint
• A Brief History of the AI Boom — required in W2 of systematic deepening; builds the 70-year timeline
• Anatomy of a Modern AI System — nails down "what an AI system is made of"
• Glossary — the shared finish line of all three routes; your permanent pocket dictionary
• Common Pitfalls and Antipatterns — the pitfall checklist to run before wrapping up any route
• Curated Resource List — resources beyond this handbook; once the system is built, exit through here

References

• Peter Norvig. Teach Yourself Programming in Ten Years — a computer scientist's classic response to "quick fixes," and a reminder of why Route 2 needs 8 weeks rather than 8 days
• S. Keshav. How to Read a Paper — the three-pass reading method; the underlying approach for the paper weeks of systematic deepening
• Coursera: Learning How to Learn — a learning-methods course from a cognitive-science perspective; the basis for checks like "re-explaining / sketching / spaced repetition"
• Wikipedia: Spaced repetition — the mechanism of spaced repetition; the principle behind daily term drilling
• Anki — open-source spaced-repetition software; you can turn this site's glossary into a deck
• Wikipedia: STAR method — the STAR framework for interview answers; use it alongside the Interview Question Bank
• OpenAI API Documentation — first-hand API reference for the hands-on work (RAG, agents)
• Hugging Face Docs — documentation for the mainstream open-source toolchain for fine-tuning and deployment