Artificial Intelligence is any technique that lets a machine do something we'd normally call "intelligent" — recognising a face, understanding a sentence, planning a route, beating you at chess.

That's a huge umbrella. A chess program from 1997 and ChatGPT are both AI — but they work in completely different ways. The first big idea to get straight is that "AI" is a family, not a single thing.

Each box sits inside the one before it. All generative AI is deep learning; all deep learning is machine learning; all of it is AI — but plenty of older AI (like rule-based systems) isn't machine learning at all.

The simplest way to feel the difference: traditional AI chooses from what exists; generative AI makes something that didn't. A spam filter picks one of two boxes. An LLM writes a sentence no one has written before. This course starts with the traditional ideas (Days 1–3) because generative AI is built directly on top of them.

A useful distinction you'll hear: supervised learning (you give labelled examples — classification & regression) vs. unsupervised learning (the machine finds structure itself — clustering). Day 1 is supervised.

Nobody hands a child a rulebook for ripe mangoes.

Machine learning works exactly like the child at the fruit stall: examples go in, a pattern comes out. The machine writes its own rulebook — one it usually can't even put into words.

There are no rules inside that box. It shrinks your drawing to a 16×16 sketch and compares it to every example you taught — picking the label of the ones it looks most like. More examples → better guesses.

A doctor doesn't memorise every patient.

Two things to notice: she chose which features matter (not shoe size), and she stores a boundary, not the patients. A machine-learned model is exactly the same.

Now hit tricky one: no straight line can separate it. Hold that thought — fixing it is exactly why Phase 2 builds neural networks.

Pretend the AI has just one knob it can turn.

The "valley" in the demo below is not a real place — it's a picture of that one idea. Before you touch it, here's exactly how to read the picture:

The orange ball is the AI's current guess for the knob. Like a hiker in thick fog (above), it can't see the whole curve — it only feels the slope right under it, and steps in the downhill direction (toward fewer mistakes). The size of each step is the learning rate. Roll downhill enough times and the ball settles in the dip — the AI has found its best setting.

In short: downhill = less wrong. Watch the ball, the dashed slope line, and the live equation move together as you press Step.

Three experiments: a tiny rate (slow), a big rate past ~1.6 (it flies off — divergence is real!), and dropping the ball left vs. right to land in a deep valley vs. a shallow trap.

Picture a panel of judges scoring a dish.

A neuron takes its inputs, multiplies each by a weight (how much it trusts that input), adds them up with a bias, and passes the total through a switch (the activation) that decides how strongly it fires.

A relay team passing a baton — and passing back the blame.

Forward pass = make a prediction. Backpropagation = trace the error back through every weight so each one knows which way to nudge. It's the chain rule from calculus, applied layer by layer — then it's just gradient descent (Day 3) on all of them at once.

An assembly line, or building with Lego.

"Deep" just means many layers. Each layer composes the previous one's output into something more abstract. Nobody tells it what an edge or an eye is — it invents these useful features itself. That's representation learning.

A map — but of meaning instead of places.

Each word becomes a vector — a point in space. Direction and distance encode meaning. The astonishing part: this lets you do arithmetic on meaning. king − man + woman lands right on queen.

A library where books aren't shelved by title — they're shelved by what they're about.

The word "bank" alone tells you nothing.

A spotlight in a dark theatre.

Each word sends out a question, every word offers a key, and the match scores decide how much attention to pay. High match → look here. It's dot products → softmax → a weighted blend.

Keeping it honest: a real transformer learns its own meaning-vectors. Here each word has a small hand-built vector so the mechanism is visible — but the engine (match → soften → blend) is the genuine article.

A very well-read improv actor.

A transformer stacks attention layers to read all the context, then predicts the next token. Temperature controls the choice: low plays it safe and predictable; high gets adventurous and creative.

Real LLMs split text into "tokens" — often word-pieces. Our toy model uses whole words. The order matters: each token gets a position.

Each token is looked up in a giant table — the embedding matrix. This is the vector DB: row = token, columns = the 8 numbers that define its meaning. (Real models use 768 or more.) Each strip below is one row of the DB.

Now the model asks every token: "which other tokens are you most similar to?" Cosine similarity between every pair. The hotter the cell, the more aligned the meaning.

Cosines become attention weights by passing through a softmax: the top scores get amplified, the rest get squashed. The query token then gathers a weighted mix of every other token's vector.

A transformer doesn't run attention once — it stacks layers. Each layer takes the previous layer's vectors and re-mixes them through attention. The same word's vector changes at every layer, getting richer with context. That's why "bank" near "river" ends up very different from "bank" near "money".

To produce the next word, the model takes the last token's vector from the top layer and compares it against every word in the vocabulary — cosine again. Softmax over those scores gives a probability distribution. Sample from it: that's your next word.