Module 3 · Generative AI & LLMs — Foundations
What Makes AI “Generative”
60 min
Learning objectives
- Distinguish generative models from discriminative (classification/prediction) models
- Explain what fundamentally changed to make today's generative AI possible
- Identify whether a given task is generative or discriminative
Two jobs: judging vs. creating
Most AI you have already met does one job: it judges an input. Is this email spam or not? Is this transaction fraud? What rating will this customer give? These are discriminative models — they draw boundaries between categories or predict a value. Generative AI does a different job: it creates new content that resembles its training data — a paragraph, an image, a melody, a block of code.
Discriminative model — A model that maps an input to a label or score — it learns the boundary between categories (e.g., spam vs. not-spam).
Generative model — A model that learns the patterns of the data well enough to produce new, plausible examples of it (e.g., a sentence, a picture, a song).
Analogy
A discriminative model is an art critic: shown a painting, it tells you the style and whether it's any good. A generative model is the artist: handed a theme, it paints something new in that style. The critic recognizes; the artist produces.
What actually changed
Generative techniques are not new — but until recently their output was crude. Three things changed around the late 2010s. Internet-scale data gave models enough examples to capture the deep structure of language and images. Cheaper, parallel compute (GPUs) made training enormous models feasible. And the transformer architecture (2017) let models learn long-range patterns far better than earlier designs. Together these turned 'generative' from a research curiosity into fluent, useful output.
| Question you ask | Type of model | Example output |
|---|---|---|
| Is this review positive? | Discriminative | Label: positive |
| Will this customer churn? | Discriminative | Probability: 0.82 |
| Write a reply to this review | Generative | A new paragraph of text |
| Make an image of a red barn | Generative | A new image |
Discriminative models choose among existing options; generative models produce new content. Both are 'AI', but they answer fundamentally different kinds of questions.
Example — Same data, two uses
Given thousands of product reviews: a discriminative use is classifying each as positive or negative. A generative use is writing a brand-new summary review in the same voice. The first sorts what exists; the second creates what didn't.
Watch out
“Generative” does not mean “correct” or “original in the human sense.” The model produces output that is statistically plausible given its training — which can be wrong, derivative, or a blend of memorized material. Plausible is not the same as true.
Knowledge check
Quick practice — not part of your exam score.
Which task is best handled by a generative model rather than a discriminative one?
Which combination of factors most directly enabled the recent leap in generative AI quality?
A model that outputs a probability that a transaction is fraudulent is best described as:
Sign in to track your progress and mark lessons complete.
Sign in