AI Atlas
LEARNING PATH

Learn AI from zero — in the right order.

The terms are laid out in a reading order — each section builds on the previous. Start anywhere; reading time is shown next to each section.

118 terms ~271 min total
Step 1

Foundations

5 terms · ~9 min total

The big picture first: what is AI, how does ML fit, why were neural networks a breakthrough, and how does a model actually 'run'?

  1. 1 Artificial Intelligence AI · ~2 min
  2. 2 Machine Learning ML · ~2 min
  3. 3 Neural Network Layered learning model · ~2 min
  4. 4 Inference The model's runtime phase · ~2 min
  5. 5 Computer Vision Machine seeing · ~1 min
Step 2

Machine Learning · In Depth

28 terms · ~55 min total

The three ML styles, classical algorithms, honest evaluation, and the basics of training neural networks — in order.

Core

  1. 1 Supervised Learning Learning from labeled examples · ~2 min
  2. 2 Unsupervised Learning Finding structure without labels · ~2 min
  3. 3 Reinforcement Learning RL · Learning by reward · ~2 min
  4. 4 Classification Assigning a label from a fixed set · ~2 min
  5. 5 Regression Predicting a continuous number · ~2 min
  6. 6 Clustering Grouping by similarity, no labels · ~2 min

Algorithms

  1. 7 Linear Regression Fitting a straight line · ~2 min
  2. 8 Logistic Regression Linear classifier with calibrated probabilities · ~2 min
  3. 9 Decision Tree Recursive yes/no splits · ~2 min
  4. 10 Random Forest An ensemble of decision trees · ~2 min
  5. 11 Gradient Boosting Sequentially correcting trees · ~2 min
  6. 12 Support Vector Machine SVM · Maximum-margin classifier · ~2 min
  7. 13 k-Nearest Neighbors k-NN · Predict by your neighbors · ~2 min
  8. 14 K-means Partition data into K clusters · ~2 min
  9. 15 Naive Bayes Probabilistic classifier with the independence trick · ~2 min

Evaluation

  1. 16 Overfitting Memorizing instead of learning · ~1 min
  2. 17 Bias-Variance Trade-off The decomposition of error · ~2 min
  3. 18 Cross-Validation Honest performance estimation · ~2 min
  4. 19 Loss Function What the model optimizes · ~2 min
  5. 20 Precision and Recall Two complementary metrics · ~2 min
  6. 21 Confusion Matrix Where the model gets it right and wrong · ~2 min
  7. 22 ROC and AUC Threshold-independent classifier ranking · ~2 min

Deep Learning

  1. 23 Backpropagation Backprop · Computing gradients across layers · ~2 min
  2. 24 Gradient Descent The optimizer underneath everything · ~2 min
  3. 25 Optimizer Adam, AdamW, SGD — what makes training work · ~2 min
  4. 26 Activation Function Where neurons get their nonlinearity · ~2 min
  5. 27 Dropout Random neuron deactivation · ~2 min
  6. 28 Batch Normalization Stabilizing layer activations · ~2 min
Step 3

Language Models

5 terms · ~10 min total

The core of modern AI: LLMs, their architecture (Transformer), the atomic unit (token), and the context they can see.

  1. 1 LLM Large Language Model · ~2 min
  2. 2 Transformer The architecture behind LLMs · ~2 min
  3. 3 Token The atomic unit of text for an LLM · ~2 min
  4. 4 Context Window How much an LLM can see at once · ~2 min
  5. 5 Multimodal Many modes, one model · ~2 min
Step 4

Talking to LLMs

9 terms · ~18 min total

The craft of steering a model: prompts, system messages, sampling parameters, and learning by example.

  1. 1 Prompt The instruction you give an LLM · ~2 min
  2. 2 System Prompt The persistent instruction · ~2 min
  3. 3 Temperature The randomness dial · ~2 min
  4. 4 Top-p Nucleus Sampling · ~2 min
  5. 5 Top-k Top-k Sampling · ~2 min
  6. 6 max_tokens Output length cap · ~2 min
  7. 7 Few-shot Learning Teaching by example · ~2 min
  8. 8 Chain-of-Thought CoT — Step-by-Step Reasoning · ~2 min
  9. 9 Streaming Token-by-token output · ~2 min
Step 5

Limits & Risks

4 terms · ~8 min total

Failure modes, security gaps, and the safety layers you must add in production.

  1. 1 Hallucination Confident wrong answers · ~2 min
  2. 2 Jailbreak Bypassing safety guardrails · ~2 min
  3. 3 Prompt Injection Hidden commands in user data · ~2 min
  4. 4 Guardrails Production safety filters · ~2 min
Step 6

Vectors & Meaning

3 terms · ~4 min total

The infrastructure of RAG: meaning as numbers, vector databases.

  1. 1 Vector Numeric representation · ~1 min
  2. 2 Embedding Meaning represented as numbers · ~2 min
  3. 3 Vector Database Similarity search store · ~1 min
Step 7

Custom Knowledge with RAG

7 terms · ~13 min total

Giving a model external knowledge without retraining — RAG's full stack and advanced variants.

  1. 1 RAG Retrieval-Augmented Generation · ~2 min
  2. 2 Chunking Splitting documents for retrieval · ~2 min
  3. 3 Hybrid Search Keyword + semantic, combined · ~1 min
  4. 4 BM25 Best Matching 25 — classic relevance score · ~2 min
  5. 5 Reranker The second-pass ranker · ~2 min
  6. 6 Knowledge Graph Structured knowledge as nodes + edges · ~2 min
  7. 7 RAG-Fusion Multi-query + fused ranking · ~2 min
Step 8

Agents & Tools

4 terms · ~9 min total

Going from chat to action: tool use, agents, the MCP standard.

  1. 1 Function Calling Tool Use · ~2 min
  2. 2 AI Agent An LLM that takes actions · ~2 min
  3. 3 MCP Model Context Protocol · ~2 min
  4. 4 llms.txt A Markdown overview of a site for LLMs and agents · ~3 min
Step 9

The Harness: The Machine Around the Agent

11 terms · ~33 min total

The layer that turns a model into a useful agent: the loop, context management, subagents, hooks, sandboxing and memory.

  1. 1 Agent Harness The software layer around the model · ~3 min
  2. 2 Agent Loop The tool-use loop · ~3 min
  3. 3 Context Engineering Designing what the model sees · ~3 min
  4. 4 Context Compaction Summarize, then keep going · ~3 min
  5. 5 Subagent A helper agent with its own context · ~3 min
  6. 6 Agent Hooks Lifecycle hooks for agents · ~3 min
  7. 7 Agent Skills Capability packs loaded on demand · ~3 min
  8. 8 Agent Sandbox Isolation and permission modes · ~3 min
  9. 9 Agent Memory Short-term and long-term memory for agents · ~3 min
  10. 10 Human-in-the-Loop Approval checkpoints for agents · ~3 min
  11. 11 Coding Agent AI that takes a task and changes the code itself · ~3 min
Step 10

Training & Optimization

7 terms · ~14 min total

How models are trained, aligned, compressed, and evaluated.

  1. 1 Fine-tuning Specializing a pretrained model · ~2 min
  2. 2 RLHF Reinforcement Learning from Human Feedback · ~2 min
  3. 3 Alignment Tuning models to human values · ~2 min
  4. 4 LoRA Low-Rank Adaptation · ~2 min
  5. 5 Quantization Model compression · ~2 min
  6. 6 Knowledge Distillation Teacher → Student transfer · ~2 min
  7. 7 Benchmark Standardized evaluation · ~2 min
Step 11

Running Models

10 terms · ~26 min total

Running models efficiently in production / locally — runtime parameters and memory management.

  1. 1 Small Language Model (SLM) Compact models that run where you are · ~3 min
  2. 2 num_ctx Runtime context length parameter · ~3 min
  3. 3 KV Cache Key-Value Cache · ~2 min
  4. 4 n_gpu_layers GPU layer offload count · ~3 min
  5. 5 num_thread CPU thread count (llama.cpp -t / --threads) · ~2 min
  6. 6 Batch Size Sequences processed in parallel · ~2 min
  7. 7 keep_alive How long Ollama keeps a model in memory · ~2 min
  8. 8 Flash Attention Memory-efficient attention and -fa / OLLAMA_FLASH_ATTENTION · ~3 min
  9. 9 min_p Min-p sampling · ~3 min
  10. 10 repeat_penalty Repetition penalty and repeat_last_n · ~3 min
Step 12

Advanced & Generative

10 terms · ~21 min total

Frontier topics: reasoning, MoE, and modern generative architectures (image, audio, video).

  1. 1 Reasoning Model Models that think first · ~2 min
  2. 2 System One Model Decisions in a single pass · ~3 min
  3. 3 Mixture of Experts MoE · ~2 min
  4. 4 Diffusion Model Generation by gradual denoising · ~2 min
  5. 5 GAN Generative Adversarial Network · ~2 min
  6. 6 Image Generation Text-to-image · ~2 min
  7. 7 TTS Text-to-Speech · ~2 min
  8. 8 ASR Automatic Speech Recognition · ~2 min
  9. 9 World Model An internal simulator inside the model · ~2 min
  10. 10 AGI Artificial General Intelligence · ~2 min
Step 13

Applied AI

6 terms · ~13 min total

The AI/ML applications most products actually ship: forecasting, anomaly detection, recommendations, sentiment.

  1. 1 Forecasting Predicting future values · ~2 min
  2. 2 Time Series Ordered observations over time · ~2 min
  3. 3 Anomaly Detection Spotting the unusual · ~2 min
  4. 4 Recommendation System Personalization at scale · ~2 min
  5. 5 Sentiment Analysis Inferring tone from text · ~2 min
  6. 6 Generative Engine Optimization (GEO) Making content citable by AI answer engines · ~3 min
Step 14

AI Governance & Regulation

9 terms · ~38 min total

Using AI responsibly and lawfully: the EU AI Act, NIST AI RMF, ISO/IEC 42001, risk classification, audits and data privacy.

  1. 1 AI Governance Who decides, by which rules, with what proof · ~4 min
  2. 2 EU AI Act Regulation (EU) 2024/1689 · ~5 min
  3. 3 NIST AI RMF NIST AI Risk Management Framework · ~4 min
  4. 4 ISO/IEC 42001 The AI management system (AIMS) standard · ~4 min
  5. 5 AI Risk Classification Putting each use case in the right box · ~4 min
  6. 6 Model Card A model's ID and instructions for use · ~3 min
  7. 7 AI Audit Algorithmic audits and the evidence trail · ~4 min
  8. 8 AI Data Privacy Applying GDPR and KVKK to AI · ~4 min
  9. 9 AI Watermarking & Content Labeling Watermarks, C2PA Content Credentials and visible labels · ~6 min