05. AI vs ML vs Deep Learning

People use these terms interchangeably. They are not the same thing. Each is a subset of the one before it.

graph TD
    AI["AI\nAny machine simulating human intelligence\nRule-based chatbots, search, robotics"]
    ML["Machine Learning\nAI that learns from data\nDecision trees, SVMs, clustering"]
    DL["Deep Learning\nML with multi-layer neural networks\nCNNs, RNNs, autoencoders"]
    TR["Transformers\nAttention-based deep learning\nBERT, GPT, T5, ViT"]
    LLM["LLMs\nLarge transformers trained on text\nGPT-4, Claude, Gemini, Llama"]

    AI --> ML
    ML --> DL
    DL --> TR
    TR --> LLM

    style AI fill:#1e1b4b,stroke:#7c3aed,color:#e2e8f0
    style ML fill:#172554,stroke:#3b82f6,color:#e2e8f0
    style DL fill:#14532d,stroke:#059669,color:#e2e8f0
    style TR fill:#451a03,stroke:#f59e0b,color:#e2e8f0
    style LLM fill:#450a0a,stroke:#ef4444,color:#e2e8f0

Definition: Any technique that enables machines to mimic human behavior.

Scope: Broadest term. Includes rule-based expert systems, search algorithms, ML, robotics — anything that makes machines act “smart.”

Example: A chess program using hardcoded rules is AI, even if it doesn’t learn.

Definition: A subset of AI where systems learn from data without being explicitly programmed.

Key idea: Give the machine data + expected outputs → it finds patterns → applies them to new inputs.

Input: [house size, location, bedrooms]
Output: predicted price

ML finds the relationship without being told "multiply size by X"

Examples: Linear regression, decision trees, random forests, SVMs, k-means clustering.

Definition: A subset of ML that uses multi-layer neural networks to learn representations.

Why “deep”? Many layers (depth) of processing. Each layer learns increasingly abstract features.

Image → edges → shapes → faces → "that's a cat"
(layer 1) (layer 2) (layer 3) (layer 4)

Examples: CNNs (vision), RNNs (sequences), autoencoders.

What deep learning unlocked: Tasks that classical ML couldn’t crack — images, audio, video, natural language.

Definition: A specific deep learning architecture using attention mechanisms to process sequences in parallel.

Published: 2017 (“Attention Is All You Need” — Google).

Why transformers matter: Previous architectures (RNNs) processed sequences word-by-word — slow and forgetful. Transformers process the whole sequence at once and learn which parts to “pay attention to.”

Examples: BERT, GPT, T5, ViT (vision transformers).

Definition: Transformers trained on massive text datasets with billions of parameters.

“Large” = billions–trillions of parameters trained on internet-scale data.

Examples: GPT-4, Claude, Gemini, Llama, Mistral.

What LLMs can do: Write, summarize, translate, reason, code, answer questions — in any domain they’ve been trained on.

Q: What is the relationship between AI, ML, Deep Learning, and LLMs?

A: They are nested subsets. AI is the broadest concept — any machine simulating intelligence. ML is a subset of AI that learns from data. Deep Learning is ML using multi-layer neural networks. Transformers are a deep learning architecture. LLMs are large transformer models trained on text. Every LLM is a transformer, every transformer is deep learning, every deep learning model is ML, and every ML model is AI.

Q: Why did transformers replace RNNs for NLP?

A: RNNs processed sequences sequentially — slow to train and prone to forgetting distant context (vanishing gradient). Transformers use self-attention to process all tokens in parallel and directly model relationships between any two positions in a sequence, regardless of distance. This made training faster and results dramatically better.