Aller au contenu

Le contenu principal de cette page est en anglais.Traduire cette page

IA et sujets tendance

AI Basics & the AI Family Tree roadmap

New to AI? Start here. This roadmap explains the family tree of AI, from classic rules-based systems to machine learning, deep learning, generative AI, LLMs and agents, so every other topic makes sense.

Temps
⏱ 2–3 weeks
Level
Beginner
Steps
6 steps
Cheat sheet ↓5 interview Q&As ↓
Partager

The roadmap

Click a step to open the full guide.

AI Basics & the AI Family Tree cheat sheet

The essentials on one screen. Bookmark it for revision.

The AI family in one line each

IAMachines doing intelligent tasks
Machine learningLearning patterns from data
Deep learningML with big neural networks
IA générativeDeep learning that creates content
LLMGenerative AI for language
RAGLLM + your documents
MCPStandard for connecting outils
AgentLLM + tools + loop + goal

Learning types

SupervisedLabelled examples → predict labels
UnsupervisedNo labels → find groups
ReinforcementRewards → learn actions
Self-supervisedPredict hidden parts of data (how LLMs pre-train)

Improve performance

  • Pick the smallest model that meets your quality bar; it's faster and cheaper.
  • Give models only the context they need; long prompts are slower.
  • Use specialised models (vision, speech) for those tasks instead of forcing a text model.

Security checklist

  • Don't paste confidential or personal data into AI tools that aren't approved for it.
  • Verify important facts; AI can be confidently wrong.
  • Be careful with AI that can act (agents): limit what it can do and review actions.

Most important interview questions

Click a question to see a strong answer. Each step also has its own questions.

1Qu'est-ce que the difference between AI, ML, deep learning and generative AI ?

Each is a subset of the previous: AI is the field, ML learns from data, deep learning uses large neural networks, and generative AI uses deep learning to create new content.

2Qu'est-ce que a large language model ?

A deep learning model, usually a transformer, trained on huge amounts of text to predict the next token, then tuned to follow instructions.

3Qu'est-ce que the difference between generative AI and agentic AI ?

Generative AI creates content on request. Agentic AI uses a generative model in a loop with tools to plan and take actions toward a goal.

4Qu'est-ce que a transformer ?

A neural network architecture built on attention, which lets each token weigh the importance of every other token. It powers modern LLMs.

5Name three limitations of LLMs.

Hallucinations, a knowledge cut-off date, and no built-in access to private or real-time data (fixed with RAG and tools).

Related feuilles de route

6

AI Engineer

6 steps · 4–6 months · Intermediate

Build products on top of large language models.

View roadmap →
6

RAG: Retrieval-Augmented Generation

6 steps · 3–4 weeks · Intermediate

Build AI that answers from your own documents.

View roadmap →
6

AI Agents

6 steps · 3–4 weeks · Intermediate

Build AI that plans, uses tools and gets work done.

View roadmap →
5

MCP: Model Context Protocol

5 steps · 2–3 weeks · Intermediate

Connect AI apps to tools and data with one open standard.

View roadmap →
1

AI Basics & the AI Family Tree · Step 1 of 6 · 2 days

The AI family tree

See how all the AI terms relate.

Think of AI as nested circles. Artificial intelligence is the broad goal of machines doing tasks that need intelligence. Machine learning is the part of AI where systems learn patterns from data instead of following hand-written rules. Deep learning is machine learning with large neural networks. IA générative uses deep learning to create new text, images, audio or code. Large language models (LLMs) are generative models for language. AI agents wrap LLMs in a loop with tools so they can act, not just answer.

Alongside these sit other branches: computer vision, speech, robotics, recommendation systems and reinforcement learning.

What you'll learn

  • Artificial intelligence (the whole field)
  • Machine learning
  • Deep learning
  • IA générative
  • Large language models
  • AI agents

AI as nested layers (outer to inner)

  1. Artificial intelligence
  2. Machine learning
  3. Deep learning
  4. IA générative
  5. Large language models
  6. Agents (LLMs + tools + loop)

Key concepts

Artificial intelligence
Machines doing tasks that normally need human intelligence.
Subset
Each term is a smaller part of the one above: every LLM is generative AI, but not all AI is an LLM.

Step by step

  1. 1.1

    Learn the nested circles

    AI ⊃ machine learning ⊃ deep learning ⊃ generative AI ⊃ LLMs.

  2. 1.2

    Add the 'doing' layer

    RAG, tools/MCP and agents are ways of using LLMs.

  3. 1.3

    Place the other branches

    Vision, speech, recommendations, robotics and reinforcement learning.

Artificial Intelligence
├── Rule-based / expert systems
├── Search & planning
├── Machine Learning
│   ├── Supervised (classification, regression)
│   ├── Unsupervised (clustering)
│   ├── Reinforcement learning
│   └── Deep Learning (neural networks)
│       ├── Computer vision (CNNs, vision transformers)
│       ├── Speech & audio
│       └── Generative AI
│           ├── Image, audio & video generation (diffusion)
│           └── Large Language Models (transformers)
│               ├── RAG (answer from your data)
│               ├── Tool use & MCP
│               └── AI Agents (plan → act → observe)
└── Robotics (perception + planning + control)

Notes

  • Every LLM is generative AI, but not all AI is generative.
  • Agents are a way of using models, not a new kind of model.

Tips & tricks

  • Draw the tree yourself from memory once a day for three days.

Interview questions

Qu'est-ce que the difference between AI, ML and deep learning ?

AI is the broad field. Machine learning is AI that learns from data. Deep learning is machine learning with large neural networks. Each is a subset of the previous one.

The quick way to study this

Memorise the six-level chain: AI → ML → deep learning → generative AI → LLMs → agents.

The AI family tree explained →Generative AI vs agentic AI →
2

AI Basics & the AI Family Tree · Step 2 of 6 · 3 days

Machine learning in plain words

How machines learn from examples.

Machine learning finds patterns in examples. Show a model thousands of emails labelled spam or not spam, and it learns to predict the label for new emails.

The key risk is overfitting: memorising the training examples instead of learning general patterns. That is why models are tested on data they have never seen.

What you'll learn

  • Training data and labels
  • Models and predictions
  • Supervised vs unsupervised learning
  • Overfitting
  • Evaluation

How machine learning works

  1. Collect examples
  2. Train a model
  3. Test on new data
  4. Use for predictions
  5. Retrain with new data

Key concepts

Training data
Examples the model learns from.
Label
The correct answer for an example, like 'spam'.
Overfitting
Memorising training examples instead of learning general patterns.

Step by step

  1. 2.1

    Supervised learning

    Learn from labelled examples (spam / not spam).

  2. 2.2

    Unsupervised learning

    Find groups or patterns without labels.

  3. 2.3

    Reinforcement learning

    Learn by trial, error and rewards.

  4. 2.4

    Testing

    Always check on data the model has never seen.

Notes

  • Better data usually beats a fancier model.

Tips & tricks

  • Try a no-code ML demo to see training in action.

Interview questions

Qu'est-ce que overfitting and how do you detect it ?

When a model performs well on training data but poorly on new data. Detect it by comparing training and validation scores; a big gap means overfitting.

Supervised vs unsupervised learning?

Supervised uses labelled examples to predict a known target. Unsupervised finds structure (like clusters) in unlabelled data.

The quick way to study this

Understand: data in, patterns learned, predictions out, tested on new data.

3

AI Basics & the AI Family Tree · Step 3 of 6 · 3 days

Neural networks and deep learning

The engine behind modern AI.

A neural network is layers of simple units with adjustable weights. Training nudges the weights to reduce errors. With enough data, layers and compute, these networks learn very complex patterns.

The transformer architecture, built on attention, powers today's language models.

What you'll learn

  • Neurons, layers and weights
  • Training by gradient descent
  • Why depth helps
  • Transformers and attention
  • GPUs

A neural network learning

  1. Input
  2. Hidden layers
  3. Prediction
  4. Compare with answer
  5. Adjust weights

Key concepts

Neural network
Layers of simple units with adjustable weights.
Gradient descent
Repeatedly adjusting weights in the direction that reduces error.
Attention
A mechanism that decides which other words matter most for each word.

Step by step

  1. 3.1

    Neurons and weights

    Each unit multiplies inputs by weights and passes the result on.

  2. 3.2

    Training

    Compare predictions to answers, then nudge weights to reduce the error (gradient descent).

  3. 3.3

    Depth

    More layers learn more abstract features: edges → shapes → objects.

  4. 3.4

    Transformers

    Attention lets every token look at every other token for context.

Notes

  • GPUs matter because training is huge amounts of parallel arithmetic.

Tips & tricks

  • Watch a visual explainer of a neural network learning.

Interview questions

Why are GPUs used for deep learning?

Training is mostly huge matrix multiplications that can run in parallel, which GPUs do far faster than CPUs.

The quick way to study this

Learn the idea of weights + training + layers; skip the maths for now.

4

AI Basics & the AI Family Tree · Step 4 of 6 · 3 days

Generative AI and LLMs

Models that create.

LLMs learn to predict the next token from huge amounts of text, then are tuned to follow instructions. Image generators usually use diffusion: starting from noise and refining it step by step.

Because they generate plausible output, they can be confidently wrong (hallucinate). RAG and tools reduce this.

What you'll learn

  • Tokens and next-token prediction
  • Pre-training and instruction tuning
  • Diffusion models for images
  • Hallucinations
  • Reasoning models
  • Multimodal models

How an LLM is made

  1. Huge text data
  2. Pre-training (next token)
  3. Instruction tuning
  4. Safety tuning
  5. Assistant model

Key concepts

Hallucination
A confident but false answer.
Knowledge cut-off
The date after which the model's training data stops.
Multimodal
Works with more than text: images, audio, video.

Step by step

  1. 4.1

    Next-token prediction

    LLMs learn by predicting the next token in huge amounts of text.

  2. 4.2

    Instruction tuning

    Extra training on examples of following instructions helpfully and safely.

  3. 4.3

    Diffusion for images

    Start from noise and denoise step by step into an image.

  4. 4.4

    Know the limits

    Hallucinations, knowledge cut-off, no built-in access to your data.

Notes

  • LLMs do not look things up unless you give them a search or retrieval tool.

Tips & tricks

  • Compare answers from two different AI assistants on the same question.

Interview questions

Why do LLMs hallucinate?

They generate plausible text based on patterns, not by looking facts up. Without retrieval or tools, they can produce confident errors.

The quick way to study this

Use three AI tools for a week and note where each succeeds and fails.

How generative AI works →
5

AI Basics & the AI Family Tree · Step 5 of 6 · 3 days

RAG, tools, MCP and agents

How LLMs connect to data and take action.

On their own, LLMs only know their training data. RAG retrieves relevant passages from your documents and gives them to the model. Tool calling lets models run functions such as search or a calculator. MCP is an open standard for plugging tools and data into AI apps. Agents combine these in a loop to complete multi-step tasks.

What you'll learn

  • RAG: answering from your documents
  • Tool calling
  • MCP: a standard for connecting outils
  • Agents: models in a loop
  • Human approval

What each adds to an LLM

RAG

  • Knowledge
  • Your documents
  • Answers with sources

Tools & MCP

  • Abilities
  • APIs and apps
  • Search, calculate, update

Agents

  • Autonomy
  • Plan → act → observe
  • Multi-step tasks

Key concepts

RAG
Gives LLMs knowledge they weren't trained on.
Tools / MCP
Give LLMs abilities to act.
Agents
Give LLMs autonomy to complete goals.

Step by step

  1. 5.1

    RAG for knowledge

    Retrieve your documents and answer from them.

  2. 5.2

    Tools for abilities

    Let the model call functions: search, calculate, look up orders.

  3. 5.3

    MCP for connections

    Plug tools into many AI apps with one standard.

  4. 5.4

    Agents for autonomy

    Let the model plan and act in a loop with limits.

Notes

  • Most business AI products are a mix of RAG, tools and a little agent logic.

Tips & tricks

  • Read the RAG, MCP and AI agents roadmaps next.

Interview questions

Explain the difference between RAG and an AI agent.

RAG retrieves information to answer a question in one pass. An agent decides and takes multiple actions (which may include retrieval) to complete a goal.

The quick way to study this

Remember: RAG = knowledge, tools/MCP = abilities, agents = autonomy.

RAG roadmap →MCP roadmap →AI agents roadmap →
6

AI Basics & the AI Family Tree · Step 6 of 6 · 2 days

Responsible AI and next steps

Use AI safely and choose your path.

AI systems can reflect bias in their data, leak private information or be tricked by malicious instructions. Responsible use means checking outputs, protecting data and keeping humans in control of important decisions.

Next, choose a path: AI engineer, ML engineer, data scientist or generative AI engineer.

What you'll learn

  • Bias and fairness
  • Privacy
  • Copyright and content rules
  • Security risks such as prompt injection
  • Choosing a career path

Where to go next

  • You (after AI basics)
  • Build AI products → AI Engineer
  • Go deeper into models → Generative AI Engineer
  • Train and deploy models → ML Engineer
  • Analyse data → Data Scientist / Data Analyst
  • Physical AI → Robotics / IoT

Key concepts

Bias
Unfair patterns a model learns from skewed data.
Human oversight
People review and can override AI decisions that matter.

Step by step

  1. 6.1

    Know the risks

    Bias, privacy, misinformation, copyright and security.

  2. 6.2

    Use AI responsibly

    Check outputs, protect data, keep humans in charge of important decisions.

  3. 6.3

    Choose a path

    Pick a career roadmap and follow it for 90 days.

Notes

  • Never paste confidential data into tools that are not approved for it.

Tips & tricks

  • Pick one career roadmap and follow it for 90 days.

Interview questions

How would you make an AI system more responsible?

Test for bias on different groups, protect personal data, explain limitations to users, keep a human in the loop for high-stakes decisions, and monitor for harmful outputs.

The quick way to study this

Read one AI safety checklist and pick your next roadmap today.

AI Engineer roadmap →

Partager cette page