If you learned about AI before 2022 — or from a college textbook — you learned about systems that judge things: spam or not spam, fraud or genuine, cat or dog, will this customer churn. Useful, invisible, everywhere. Then generative AI arrived and did something the textbooks did not predict would work this well: it started making things. Whole essays. Working Python. Photorealistic images. This page explains what changed, how the technology works without the jargon, and what it honestly can and cannot do.
Generative vs traditional AI — the table worth memorising
| Traditional (discriminative) AI | Generative AI | |
|---|---|---|
| Core job | Classify / predict from data | Create new content |
| Example task | "Is this transaction fraud?" | "Draft the fraud-alert email in Hindi and English" |
| Typical output | A label, score, or number | Text, code, images, audio, video |
| How it is built | One custom model trained per task | One huge pre-trained model, steered by prompts |
| Skill to use it | Train models (ML engineering) | Instruct and integrate models (prompting + engineering) |
| Famous examples | Netflix recommendations, UPI fraud checks | ChatGPT, Claude, Gemini, Midjourney |
Both kinds are “real AI”, and companies run both — the fraud model does not disappear because a chatbot arrived. If the wider AI/ML/deep-learning family tree is fuzzy, our AI vs ML vs deep learning explainer sorts it out; for the ground-floor question, start at what is artificial intelligence.
How it works — the one-idea explanation
Strip away the mystique and a large language model does exactly one thing: given some text, predict the next word. That is the entire trick. “The capital of France is ___” → “Paris”. Trained on a substantial fraction of everything humans have written, with billions of adjustable parameters, that single skill turns out to contain multitudes — because predicting the next word well in a legal document, a Python file, or a Hindi poem requires absorbing law, programming, and poetry.
Two more steps complete the picture. The raw predictor is then tuned with human feedback until it behaves like a helpful assistant rather than an autocomplete. And at answer time it generates one token after another, each conditioned on everything so far — which is why responses stream word by word, and why the same question can produce different phrasings on different runs. Image models (Midjourney, DALL·E) run a cousin of this idea: they learn to sculpt random noise into a picture that matches your description.
What it genuinely cannot do (the part hype skips)
- •It can state falsehoods fluently. LLMs predict plausible text, not verified truth. When the model lacks the fact, it may produce something confident and wrong — called a hallucination. This is why serious business systems attach models to real documents using RAG.
- •Its knowledge has a cutoff. A model knows what its training data knew. Yesterday's RBI circular or your company's internal policy? Not in there — unless engineered in via retrieval or search.
- •It does not "understand" like a person. No goals, no experience of the world, no idea whether it just helped or harmed. It is an extraordinarily capable text machine, and treating it as an infallible oracle is the #1 beginner mistake.
- •It cannot act on its own. A chat model only talks. Making models actually do things — search, book, execute, verify — is a separate engineering layer called AI agents.
Those four limitations are not footnotes — they are the job market. Almost every GenAI engineering role in India exists to work around exactly these gaps: grounding models in company data (RAG), instructing them reliably (prompt engineering), and giving them safe hands (AI agents).
Where Indian companies actually use it (beyond chatbots)
Customer support is the obvious one, but the quieter deployments are more instructive: banks summarising loan documents and drafting relationship-manager notes; e-commerce teams generating and localising product descriptions across languages; IT services firms using coding assistants to cut delivery time; hospitals drafting discharge summaries for doctor review; edtechs generating practice questions. The pattern across all of them — a human stays in the loop, and the model is grounded in the company’s own documents. Pure “let the chatbot answer from memory” deployments are where the embarrassing screenshots come from.
Why this changed AI careers
Before 2023, working in AI mostly meant training models — months of data work, real mathematics, one task per model. Generative AI added a second career lane: building applications on top of pre-trained models. The skills are Python, APIs, retrieval, evaluation, and judgement — closer to software engineering than to research, and needing notably less maths. It is currently the fastest-growing slice of Indian AI hiring, with GenAI/LLM roles at ₹18–30 LPA mid-level, and it is why our AI course dedicates its later modules to the GenAI stack — LLMs, RAG, LangChain, and agents — rather than stopping at classical ML.
Frequently asked questions
ChatGPT is one product built on generative AI — the relationship of WhatsApp to "messaging". Claude, Gemini, Llama, Midjourney and many others are siblings in the same family.
Large Language Model — the text engine underneath tools like ChatGPT and Claude. "Large" refers to billions of learned parameters; "language model" means its core skill is predicting text.
It replaces tasks faster than jobs. Fluent first drafts — of code or prose — are now nearly free, which moves human value to specifying, reviewing, and owning the result. The people most at risk are those who refuse to use it; the people gaining are those who direct it well.
To build applications — no; a working mental model like the one on this page is enough to start. To train or fine-tune models at a research level — yes, and that is a much smaller job market.