Open your phone and type the first half of a text message. The little suggestions above the keyboard — "running late", "on my way" — are a tiny language model guessing what comes next. Now imagine that same trick, scaled up a billion times and trained on a large slice of everything humans have written. Ask it to "write a polite email declining a meeting" and it doesn't pick from a menu of canned replies — it composes a brand-new one, word by word, that fits your exact situation. That leap, from suggesting to composing, is the whole story of Generative AI.
Generative AI in one sentence
Here's the definition we'll use for the rest of the course, stripped of jargon:
The key word is creates. For most of computing history, software was a machine for following instructions: if the user clicks this, show that; if the number is above 100, flag it. Every behaviour had to be spelled out by a programmer in advance. Generative AI breaks that mould. Nobody wrote a rule for how to answer your specific question — the model generates a fresh answer on the spot, having learned the underlying patterns of language (or images, or sound) from enormous amounts of examples.
That's why people say AI is "eating software." Tasks that used to need a custom-built program — summarising a document, drafting marketing copy, translating a paragraph, answering a support ticket — can now be handled by a single general-purpose model with the right instructions. You're not writing the logic anymore; you're directing a model that already has it.
Where it fits: AI, ML, deep learning, GenAI
"AI" is a crowded word. A 1980s chess program, a spam filter, and ChatGPT are all called "AI," yet they work in completely different ways. The cleanest way to keep them straight is a set of nested circles — each one a more specific case of the one around it.
- Artificial Intelligence (AI) is the broadest goal: get machines to do things that look intelligent. A hand-written rulebook for playing tic-tac-toe counts.
- Machine Learning (ML) is one way to get there — instead of coding the rules, you let the system learn patterns from data. A spam filter that learns from examples of spam lives here.
- Deep Learning (DL) is ML done with multi-layered neural networks. It's what made speech recognition and image labelling suddenly work well around 2012.
- Generative AI is deep learning pointed at creating new content rather than just labelling existing content. The large language models (LLMs) at the heart of this course are its flagship example.
When someone says "we use AI," ask one question: does it learn from data, and does it create something new? The answers instantly tell you whether you're looking at an old rulebook, a classic ML model, or true Generative AI — and that changes how you'd build, test, and trust it.
Two kinds of "smart": deciding vs creating
The single most useful distinction in this whole field is between models that decide and models that create. Get this and a lot of confusion melts away.
A discriminative model answers "which bucket does this belong in?" Is this email spam or not? Is this review positive or negative? Is this transaction fraud? It draws boundaries between categories. These models power a huge amount of valuable software, and they're not going anywhere.
A generative model answers "what's a plausible new example?" Write a review. Draft the email. Suggest a melody. Continue this paragraph. Instead of sorting what exists, it produces what doesn't exist yet. The simple test: does it pick from options, or make something new?
Assuming "Generative AI" means "smarter AI." It doesn't — it means a different job. For "flag risky transactions," a focused discriminative model is often cheaper, faster, and more reliable than an LLM. Knowing when not to reach for generation is a senior skill we'll keep sharpening.
Why it's happening now
Generative models aren't a brand-new idea — researchers have tinkered with them for decades. Three things converged to make them suddenly, startlingly good:
- Data. The internet produced an ocean of text, images, and code to learn from.
- Compute. Graphics chips (GPUs) made it practical to train enormous models on that ocean.
- A new architecture. The transformer (2017) let models weigh the relationships between every word in a passage at once — and scale beautifully. We'll meet it properly in Section 3.
Scale up a transformer on enough data with enough compute and something remarkable happens: the model picks up abilities nobody explicitly trained it for — translation, summarisation, basic reasoning, writing code. That surprise, called emergence, is why a single "foundation model" can power thousands of different products through nothing more than a well-written prompt.
Think about one task in your week that involves producing text, code, or images from scratch. That task is almost certainly something a generative model could draft for you. Hold onto it — you'll meet tools in this course that automate exactly that kind of work.
Your destination
It's easy to get lost in a long course, so let's name the summit on day one. By the end, you won't just understand Generative AI — you'll have built a real, production-shaped system with it. You'll go from "I can call a model's API" to "I can design, evaluate, secure, and deploy an AI application I'd trust in front of users."
The vehicle for that is a single product you'll grow section by section — an AI assistant called Sage, built for a fictional online-learning startup. You'll meet Sage properly in the very next lesson. For now, the only thing to remember is this: you learn by building, one layer at a time, not by collecting disconnected demos.
No code yet — just sharpen the core distinction. For each item below, decide whether it's a job for a discriminative ("decide") or generative ("create") model:
- Sorting incoming support tickets into "billing", "bug", or "other".
- Writing a friendly first-draft reply to a support ticket.
- Detecting whether a photo contains a cat.
- Producing alt-text describing what's in that photo.
Show the answers
1 — discriminative (sort into buckets). 2 — generative (compose new text). 3 — discriminative (cat / not-cat). 4 — generative (create a description). Notice how 1&2 and 3&4 are the same domain but opposite jobs. That decide-vs-create instinct will guide dozens of design choices ahead.
- Generative AI = software that creates new content, not just labels existing content.
- Nested circles: AI ⊃ ML ⊃ Deep Learning ⊃ Generative AI.
- Discriminative models decide; generative models create — same domains, opposite jobs.
- Data + compute + the transformer made it work now; emergence is why one model does many tasks.
- You'll learn by building one real product — Sage — layer by layer.