Two friends look at the same food photo. One says "that's paneer tikka, maybe 4 stars." The other says "give me two minutes" and writes a full Zomato review with a joke in it. Both are intelligent — but they're doing completely different jobs. One is judging what exists; the other is creating something new. Hold that picture — it's the entire first question of most AI interviews, and by the end of this lesson you'll answer it better than the interviewer expects.
The judge and the artist
Almost every AI system you've ever met falls into one of two families. The judge looks at something and gives a verdict: spam or not spam, fraud or genuine, cat or dog, approve or reject. The artist is given a request and produces something that did not exist a second ago: an email, a poem, working code, an image of "a cow doing yoga on Marine Drive."
The judge's output is always small and fixed — one label from a known list, or a score. The artist's output is open-ended — there are infinite valid product descriptions, and it composes one of them. This single difference decides how the models are built, trained, and used.
Now the proper names: generative vs discriminative
Time to earn the vocabulary. The judge is a discriminative model — it discriminates (separates) between classes. It learns the boundary: "emails with these patterns fall on the spam side of the line." The artist is a generative model — it learned the distribution of the data itself, deeply enough that it can sample new examples from it. It doesn't just know where the line is; it knows what the territory looks like well enough to draw new places on the map.
Discriminative models learn the boundary between classes; generative models learn the data well enough to create new samples of it. That sentence, word for word, is a complete interview answer.
| Discriminative (judge) | Generative (artist) | |
|---|---|---|
| Question it answers | "Which class is this?" | "Create one like this." |
| Output | a label / score from a fixed set | new open-ended content |
| Examples | spam filter, fraud detection, face unlock | ChatGPT, image generators, Copilot |
| Classic interview line | learns P(label | input) — the boundary | learns the data distribution — can sample from it |
Selection-round radar: "Difference between generative and discriminative models" opens GenAI rounds at TCS, Infosys and Accenture constantly. Give the boundary-vs-distribution line, then ONE example of each. Bonus point: mention that translation and summarization count as generative — the output is newly created text, even though input text exists.
Where LLMs fit in the GenAI family
"Generative AI" is the umbrella. Under it: text generators, image generators (text-to-image), audio/music generators, video generators, code generators. The stars of the show — and of this course — are LLMs, Large Language Models: generative models for text (ChatGPT, Claude, Gemini). "Large" is literal: billions of learned parameters, trained on a giant slice of the internet.
One more term you'll hear in every job description: foundation model. It means a big model pre-trained once on broad data, which everyone then adapts for their own tasks — the way one strong engine powers many different vehicles. GPT-4-class models, Claude, Gemini and Llama are all foundation models. You don't train them; you build on them — with prompting, RAG and fine-tuning, which are literally lessons 6, 9 and 10 of this course.
GenAI vs traditional ML — what changed for developers
In the traditional ML world, every task meant a project: collect labelled data, train a model for that one task, deploy it, repeat for the next task. A spam model can't summarize; a summarizer can't classify. In the GenAI world, one pre-trained model does thousands of tasks — you steer it with instructions written in plain language. Same model: summarize this contract, extract these fields, write this test, translate this paragraph.
That shift changed the job itself. The scarce skill moved from "can you train a model?" to "can you get reliable, safe, affordable results out of a model someone else trained?" — which is exactly what GenAI interviews probe, and exactly what the rest of this course teaches.
Wait — if it's so smart, why did it just make up a court case?
Here's the "wait, what?" moment of this lesson. Ask an LLM for a famous poem — perfect. Ask it for the exact citation of a specific 2019 High Court judgment — it may invent one, fluently, confidently, with a realistic-looking case number. Why? A generative model's whole talent is producing plausible content. Usually plausible and true overlap. When they don't, the model picks plausible — because plausible is all it ever learned to be. That failure has a name, hallucination, and it gets a full lesson (lesson 7). For now, lock in the mindset: GenAI output is a draft to verify, not a fact to trust.
Where you'll actually meet GenAI at work
In Indian IT projects right now, the common builds are: customer support assistants answering from company documents, document processing (contracts, invoices, claims → structured data), code assistants and test generators inside dev teams, content and marketing generation, and meeting/report summarizers. Notice a pattern: most of them are language in, language out, grounded in the company's own data — which is why RAG (lesson 9) is the most-asked applied topic in interviews.
💡 Tip: when an interviewer asks "have you used GenAI?", don't say "yes, ChatGPT." Say what you built or automated with it, even if small — "I built a script that summarizes error logs with an LLM API" beats "I chat with it daily."
Common mistakes (interviewers hear these weekly)
- Calling every AI "generative" — face unlock and fraud detection are discriminative. If the output is a label, it's not generation.
- Missing that translation/summarization ARE generative — new text is created, even though input text exists.
- "GenAI is always better than traditional ML" — for numeric prediction on structured data (sales forecasting, credit scoring), classic ML is still the right tool.
- Treating model output as verified fact — it's a plausible draft; important outputs need grounding or review.
- Saying "the model searches the internet" — a plain LLM has no live search; it generates from learned patterns (unless a tool explicitly gives it search — lesson 11).
Quick recap
| Concept | One-liner |
|---|---|
| Discriminative model | judges — learns the boundary, outputs a label |
| Generative model | creates — learned the data's patterns, samples new content |
| LLM | a generative model for text, with billions of parameters |
| Foundation model | pre-trained once, adapted by everyone for many tasks |
| The developer's job | steer existing models: prompts, data, guardrails — not training |
| The catch | plausible ≠ true — GenAI output is a draft to verify |
Practice Zone — PYQs from real selection rounds
Six click-to-answer questions and three hands-on tasks — under each one you'll see the companies where that pattern has been asked. Attempt first, reveal second.
Which of these is a generative AI task, rather than a discriminative one?
Asked in
The cleanest one-line difference between discriminative and generative models is:
Asked in
Which of the following is NOT a typical generative AI use case?
Asked in
Compared with traditional ML models, the biggest practical difference of modern GenAI (LLMs) for a developer is:
Asked in
A foundation model is best described as:
Asked in
Your manager says: "Use GenAI to predict next month's sales number from our past sales data." The honest technical answer is:
Asked in
Now the hands-on tasks. Think your answer out loud (seriously — say it), then reveal:
Label each task G (generative) or D (discriminative): 1) Face unlock on your phone, 2) GitHub Copilot suggesting the next line of code, 3) Netflix deciding which thumbnail to show you, 4) Translating a paragraph from Hindi to English, 5) A bank flagging a UPI transaction as fraud, 6) Composing background music for a reel.
Asked in
Interviewers often ask: "Explain generative AI to a non-technical manager in 2–3 sentences." Write your version, then compare.
Asked in
A bank wants three things: (a) a chatbot that answers customer questions about its loan products, (b) a system that approves/rejects loan applications, (c) monthly summaries of customer complaints for management. For each, say whether you'd reach for GenAI, classic ML, or both — and why.
Asked in
FAQ
Is generative AI the same thing as AI?
No — it's one family inside AI. AI also includes discriminative ML (classifiers, recommenders), optimization, robotics and more. GenAI is the family that creates content, and it's the one currently reshaping hiring.
Is ChatGPT a model or a product?
A product — a chat application built around LLMs (GPT-4-class models and successors), plus safety layers, tools and memory features. In interviews, use the precise word: the model generates; the product wraps it.
Do I need to know GANs and diffusion models for placements?
Only at one-line level for most roles: GANs pit a generator against a discriminator; diffusion models generate images by learning to reverse noise. LLM-side questions dominate placement rounds, which is why this course focuses there.
What should I learn after this lesson?
Go straight to lesson 2 — How LLMs Work. Tokens and next-token prediction are the foundation every other lesson (and most interview questions) stand on.
Next lesson: the machine behind the magic — Lesson 2: How LLMs Work →