AI Atlas

Learn AI from a single map

A 78-lesson route from Python to large language models, in the order to learn them. Lessons are being written in Vietnamese and English, free, no account needed.

Start learning Open the map

The 1838 map An Nam dai quoc hoa do, its title printed in both Chinese characters and Latin script. — An Nam dai quoc hoa do, Jean-Louis Taberd, 1838. Its title is printed in two scripts.

Seven parts, one fork

Walk through Python, Math, Machine Learning and Deep Learning in order. There the road splits three ways: language, vision, or large language models.

Python
The language every later lesson speaks.
7 lessons
Ada Lovelace’s Bernoulli table, 1843: the first published program.
Math for AI
Vectors, derivatives and probability: enough, no more.
3 lessons
$The Pythagorean theorem drawn as red, yellow and blue squares.$
The Pythagorean theorem in Oliver Byrne’s Euclid, 1847.
Machine Learning
The first models, and how to tell where they go wrong.
14 lessons
Francis Galton’s diagram, 1886, from the paper that named "regression".
Deep Learning
Stacking layers of neurons.
8 lessons
Purkinje cells, drawn by Santiago Ramón y Cajal, 1899.
After Deep Learning, pick a branch
Natural Language Processing
Machines that read: from tokenization to BERT and GPT.
13 lessons
The Rosetta Stone: one text, three scripts. Photo published in 1922.
Computer Vision
Machines that see: from CNNs to diffusion.
10 lessons
Sallie Gardner galloping frame by frame, Eadweard Muybridge, 1878.
Large Language Models
Using, building and training large language models.
23 lessons
A Vietnamese–Latin dictionary, compiled by Pigneau de Béhaine, printed by Taberd in 1838.

All 78 lessons

Every lesson knows which lessons lead to it. Pick a destination to see the route.

Destination

Route to Linear Regression. Lessons on the way: 8. Parts: 3.

Python 4 → Math for AI 2 → Machine Learning 2

Python

7 lessons

Getting Started: Google Colab and KaggleNot translated yet (Learned)
Basic SyntaxNot translated yet (Learned)
Data Structures: list, tuple, set, dictNot translated yet (Learned)
Functions, Strings, and File I/ONot translated yet (Learned)
Object-Oriented Programming (OOP)Not translated yet (Learned)
NumPy and PandasNot translated yet (Learned)
Data Visualization: Matplotlib and SeabornNot translated yet (Learned)

Math for AI

3 lessons

Linear AlgebraNot translated yet (Learned)
CalculusNot translated yet (Learned)
Probability and StatisticsNot translated yet (Learned)

Machine Learning

14 lessons

Introduction to Machine Learning (Learned)
Linear Regression (Learned)
Overfitting and Underfitting (Learned)
Ridge, Lasso, and Elastic Net (Learned)
Gradient Descent (Learned)
Advanced Optimizers (Learned)
K-means (Learned)
K-Nearest Neighbors (KNN) (Learned)
Logistic Regression (Learned)
Model Evaluation: Accuracy, Precision, Recall, F1 (Learned)
Softmax Regression (Learned)
Decision Trees and Random Forests (Learned)
Data Preprocessing (Learned)
An End-to-End ML Project (Learned)

Deep Learning

8 lessons

Neural NetworksNot translated yet (Learned)
Training Neural NetworksNot translated yet (Learned)
Convolutional Neural Networks (CNN)Not translated yet (Learned)
Modern CNN ArchitecturesNot translated yet (Learned)
Recurrent Neural Networks (RNN)Not translated yet (Learned)
GRUNot translated yet (Learned)
LSTM and BiLSTMNot translated yet (Learned)
TransformersNot translated yet (Learned)

After Deep Learning, pick a branch

Natural Language Processing

13 lessons

Regular Expressions (Regex)Not translated yet (Learned)
TokenizationNot translated yet (Learned)
Text NormalizationNot translated yet (Learned)
Vector Semantics and EmbeddingsNot translated yet (Learned)
Word2Vec (CBOW, Skip-gram), GloVe, FastTextNot translated yet (Learned)
BERTNot translated yet (Learned)
GPT and Distributed TrainingNot translated yet (Learned)
Advanced TransformersNot translated yet (Learned)
Information RetrievalNot translated yet (Learned)
Sentiment AnalysisNot translated yet (Learned)
Named Entity Recognition (NER)Not translated yet (Learned)
Text Generation and SummarizationNot translated yet (Learned)
Question AnsweringNot translated yet (Learned)

Computer Vision

10 lessons

Image Classification with CNNsNot translated yet (Learned)
Image Classification with Transformers: Attention, ViTNot translated yet (Learned)
Semantic Segmentation: FCN, UNet, FPN, DeepLab, HRNetNot translated yet (Learned)
Instance and Universal SegmentationNot translated yet (Learned)
Vision and Language (1): CLIP, BLIP, GLIPNot translated yet (Learned)
Vision and Language (2): LLaVA and Its VariantsNot translated yet (Learned)
Image Generation (1): VAE, GAN, CycleGAN, Pix2PixNot translated yet (Learned)
Image Generation (2): Diffusion, Stable Diffusion, LoRANot translated yet (Learned)
Data Augmentation in Computer VisionNot translated yet (Learned)
Self-Supervised Learning for ImagesNot translated yet (Learned)

Large Language Models

23 lessons

Building LLM Applications

Using Text Generation APIsNot translated yet (Learned)
Building UIs with Gradio and StreamlitNot translated yet (Learned)
Prompt EngineeringNot translated yet (Learned)
Docker and the CloudNot translated yet (Learned)
Deploying Your Own ModelNot translated yet (Learned)
RAG and Modern ChatbotsNot translated yet (Learned)
Web Crawling and MongoDBNot translated yet (Learned)
LangChain and LlamaIndexNot translated yet (Learned)
AgentsNot translated yet (Learned)
Graph RAGNot translated yet (Learned)
Image and Video APIsNot translated yet (Learned)

Building and Training LLMs

Building Your First LLMNot translated yet (Learned)
The GPT Family and Multi-Task Fine-TuningNot translated yet (Learned)
Preparing Pre-Training DataNot translated yet (Learned)
RLHFNot translated yet (Learned)
DPONot translated yet (Learned)
Model Evaluation and PEFTNot translated yet (Learned)
The LLaMA Family: RoPE and Parallel TrainingNot translated yet (Learned)
Sequence-to-Sequence Models (Seq2Seq)Not translated yet (Learned)
The DeepSeek FamilyNot translated yet (Learned)
Debugging Language ModelsNot translated yet (Learned)
Multimodal LLMsNot translated yet (Learned)
Deploying LLMs: Multi-GPU and Distributed ServingNot translated yet (Learned)

Bold: open nowFaded: coming soon

Inside a lesson

Every lesson follows the same order: intuition first, then the theory, then example code.

From Python to Machine Learning, the code runs right on the page. This is a sketch of Linear Regression, the first lesson to open.

Tabula Indiae Orientalis, Frederick de Wit, c. 1682.

Intuition
Bigger houses tend to cost more.
Draw one straight line close
to every (size, price) point.
That line guesses the price
of a house not yet sold.
Theory
$\hat{y}_i = w x_i + b$
$\mathcal{L} = \frac{1}{m} \sum_{i=1}^{m} (\hat{y}_i - y_i)^2$
Find $w$ and $b$
that make $\mathcal{L}$ smallest.

Code

import numpy as np

rng = np.random.default_rng(42)
m = 50
x = rng.uniform(0, 10, m)
# true parameters: w = 3, b = 2, plus noise
y = 3 * x + 2 + rng.normal(0, 1.5, m)

w, b = np.polyfit(x, y, 1)
loss = np.mean((w * x + b - y) ** 2)
print(f"w = {w:.3f}")
print(f"b = {b:.3f}")
print(f"L = {loss:.3f}")

Output
```
w = 3.018
b = 1.653
L = 1.290
```
The data came from w = 3, b = 2. Close, despite the noise.
Quick check
Every price doubles.
What is the new $w$ ?
Answer: twice the old $w$ .
The line gets twice as steep.
Summary
One line close to every point.
"Close" is measured by $\mathcal{L}$ .
Further reading: Galton (1886),
the paper that named "regression".

One idea, two names

In the Vietnamese lessons, each term carries its English name the first time it appears, so English material feels familiar.

A painting of a purple Iris versicolor flower. — Iris versicolor, Curtis’s Botanical Magazine, 1787: one flower, a Latin name and a common name.

linear regression: hồi quy tuyến tính
loss function: hàm mất mát
learning rate: tốc độ học
weight: trọng số
bias: hệ số tự do
hyperparameter: siêu tham số
training set: tập huấn luyện
supervised learning: học có giám sát
clustering: phân cụm
neural network: mạng nơ-ron
backpropagation: lan truyền ngược
attention mechanism: cơ chế chú ý
tokenization: tách từ
large language model: mô hình ngôn ngữ lớn

Why AI Atlas exists

Hello,

AI Atlas began as a course syllabus for learning AI. I rearranged its sessions into one map, so anyone can see what to learn first, what comes next, and why.

Each lesson is written in Vietnamese first, then translated into English. Every code example is run, every source is real.

The site is free and needs no account. Your progress is kept in your own browser and sent nowhere.

The first lesson is now open: Introduction to Machine Learning. See you there.

LuffyrealsMaker of AI Atlas

Content under CC BY-NC-SA 4.0. Code under the MIT license.

Read the license