AI, machine learning, deep learning

The three terms people use interchangeably, what a model actually is, and where the boundary of "intelligence" sits.

Three nested ideas

AI is the goal: machines doing things we would call intelligent. Machine learning is the dominant method: instead of writing rules by hand, you show examples and let the program infer the rules. Deep learning is one family of ML methods — neural networks with many layers — which happens to work extremely well on images, audio and text.

TermWhat it meansExample
AIThe broad fieldChess engines, recommenders, chatbots
Machine learningLearning patterns from dataSpam filter trained on labelled email
Deep learningML with deep neural networksSpeech recognition, image segmentation
Generative AIModels that produce contentText, images, code, audio
LLMA very large text modelGPT-style, Claude, Llama, Mistral
💡
Classic rule-based programs are still AI in the broad sense. "It uses AI" says nothing about quality — plenty of products labelled AI are a decision tree or a regex.

What a model really is

A trained model is a large set of numbers (parameters) plus a fixed computation graph. Training adjusts those numbers to reduce a loss — a number measuring how wrong the predictions are. Nothing in a model "knows" anything; it encodes statistical regularities of its training data.

# the shape of every supervised learning problem
X, y = load_data()             # features and known answers
model = Model()

for epoch in range(EPOCHS):
    preds = model(X)
    loss  = criterion(preds, y)      # how wrong are we?
    loss.backward()                  # gradients
    optimizer.step()                 # nudge the parameters
    optimizer.zero_grad()

Two consequences follow directly from that definition. First, a model can only be as good as the data it saw — biases in, biases out. Second, it interpolates rather than reasons: give it an input unlike anything in training and the output is unreliable.

What current systems cannot do

  • No guaranteed correctness. Outputs are plausible, not verified. Anything consequential needs a check or a deterministic fallback.
  • No reliable self-knowledge. A model will state a wrong fact with the same confidence as a right one.
  • Bounded context. It only "sees" what is in its window plus what tools let it fetch.
  • Training cut-off. Facts change; the weights do not.
⚠️
Treat model output as a draft from a fast, confident, sometimes-wrong colleague. For code, medicine, law or money, a human or a test must own the decision.

FAQ

Do I need maths to work with AI?
To use APIs and build products, no. To train or fine-tune models, linear algebra, probability and calculus stop being optional — see the Math for AI course.
Is an LLM a database?
No. It has no rows to look up and no structured guarantees; it generates likely continuations. Anything factual should come from a database or search that you control.

The model lifecycle Machine learning in one page

Last refreshed 2026-09-18.