Artificial intelligence (AI) is software that does things we usually associate with human thinking: recognising images and speech, understanding and writing language, making predictions and planning actions. Most modern AI works by learning patterns from large amounts of data (machine learning), and today's best-known systems, large language models (LLMs), are deep neural networks trained on vast amounts of text.

What does artificial intelligence actually mean?

The term was coined in 1956 for a research workshop at Dartmouth College, with the ambition of making machines that "think". Seventy years later, the working definition is more practical: an AI system takes some input (a photo, a sentence, a spreadsheet, a sensor reading), and produces an output that would otherwise need human judgement (a label, a reply, a forecast, a decision).

There are two broad ways to build such a system:

  • Rules written by people. "If the transaction is above this amount and from a new device, flag it." Easy to understand, but brittle: people cannot write rules for every situation.
  • Patterns learned from data. Show the system thousands of past transactions labelled "fraud" or "not fraud" and let it learn the patterns itself. This is machine learning, and it powers almost every AI product you use.

How does AI learn from data?

At its core, a machine learning model is a mathematical function with adjustable numbers called parameters. Training works like this:

  1. The model makes a prediction for an example.
  2. A loss function measures how wrong the prediction was.
  3. An optimisation method (usually gradient descent) nudges the parameters to reduce that error.
  4. Repeat over millions of examples until predictions are good on data the model has never seen.

A small model might have a few dozen parameters; a large language model has billions. The principle is the same, which is why learning the fundamentals properly pays off all the way up to LLMs.

What are the main branches of AI?

BranchWhat it doesEveryday example
Machine learningLearns patterns from data to predict or classifyLoan approval scoring, demand forecasting
Deep learningUses many-layered neural networks for complex dataFace unlock, voice typing
Natural language processingUnderstands and generates human languageTranslation, chatbots, search
Computer visionUnderstands images and videoDocument scanning, crop-disease detection
Generative AI and LLMsCreates new text, code, images or audioWriting assistants, coding assistants
Reinforcement learningLearns by trial, error and rewardGame-playing agents, robotics, model alignment

These overlap. An LLM is a deep learning model for NLP; an AI agent combines an LLM with tools and planning. Our explainer on ML vs deep learning vs AI untangles the terms further.

Where is AI already used in India?

You use AI many times a day, often without noticing:

  • Payments: fraud-detection models screen digital payments, including the UPI transactions that NPCI reports in the billions every month.
  • Language: India has 22 scheduled languages, and translation and speech systems such as the government's Bhashini platform aim to make digital services usable in all of them.
  • Agriculture: image models identify crop diseases from phone photos; forecasting models help with weather and prices.
  • Healthcare: models assist in reading scans and screening, always alongside doctors.
  • Commerce and media: recommendations, search ranking and customer-support chatbots.

The government's IndiaAI Mission, approved in March 2024 with an outlay of ₹10,371.92 crore, funds compute, datasets, startups and skills, a sign of how central AI has become to India's plans. Our piece on how AI is used across Indian industries goes deeper.

What is the difference between narrow AI and general AI?

Every AI system in use today is narrow AI: very good at specific tasks, with no general understanding of the world. Artificial general intelligence (AGI), a system as broadly capable as a person, does not exist yet, and experts disagree about when or whether it will. Read the types of artificial intelligence for the full classification.

What skills does it take to build AI?

Using AI tools takes an afternoon. Building AI systems takes a stack of skills, each resting on the one below:

Weeks Program Zero spends on each layer of the AI skill stack
Foundations & programming14
DSA & databases15
Full-stack & system design12
Machine learning6
Deep learning6
NLP, transformers & LLMs16
MLOps, security & capstone9

Source: the published Program Zero curriculum (78 weeks in total).

Notice that the engineering foundations take more than half the time. That is deliberate: models are only useful inside working software, and engineers who can build the whole system are the ones teams rely on.

What are the risks and limits of AI?

RiskWhat it looks likeWhat engineers do about it
BiasModels repeat unfair patterns in their training dataAudit data, test across groups, monitor outcomes
HallucinationLLMs state false things confidentlyGround answers in sources (RAG), evaluate, show citations
PrivacyPersonal data leaks through training or promptsMinimise data, follow India's DPDP Act, control access
SecurityPrompt injection, model misuseLimit permissions, validate inputs and outputs
Over-reliancePeople trust outputs without checkingHuman review for important decisions

See why AI hallucinates and prompt injection explained for two of the most practical risks.

How is AI different from ordinary software?

Ordinary software is written as explicit instructions. A programmer decides every step: if the user clicks here, show that screen; if the balance is below zero, block the payment. The behaviour is fully specified in advance, so it is predictable and easy to test, but it can only handle situations the programmer foresaw.

AI software is different in one crucial way: part of its behaviour is learned rather than written. Nobody writes a rule that says "these pixel patterns mean a cat". Instead, engineers choose a model architecture, collect examples, define what "good" looks like, and let training discover the rules. The result can handle messy, varied input that no hand-written rulebook could cover, such as handwriting, accents, or the many ways people phrase the same question.

That flexibility comes with new engineering problems. A learned model can be wrong in surprising ways, can drift as the world changes, and cannot be fully understood by reading its code. This is why serious AI work includes evaluation sets, monitoring, versioning and human review, not just model training. Engineers who understand both sides, traditional software and learned models, are the ones who ship AI that works in production.

What does an AI project look like from start to finish?

Whether it is a fraud detector or a customer-support assistant, most AI projects follow the same life cycle:

StageWhat happensSkills used
1. Frame the problemDecide what to predict or generate, and how success is measuredDomain understanding, metrics
2. Collect and clean dataGather examples, fix errors, remove duplicates, protect privacySQL, Python, Pandas
3. Build a baselineStart with the simplest model that could workStatistics, scikit-learn
4. Improve the modelTry better features, algorithms or neural networksML, deep learning
5. Evaluate honestlyTest on data the model has never seen, across groups and edge casesValidation, error analysis
6. DeployWrap the model in an API, connect it to the productFastAPI, Docker, cloud
7. Monitor and retrainWatch quality, cost and drift; update as data changesMLOps

Notice how little of this is "the model" itself. Most of the work is data, engineering and evaluation, which is why a strong programming foundation matters so much.

How should a beginner start learning AI?

  1. Learn Python and basic problem solving.
  2. Learn to handle data: SQL, NumPy, Pandas and statistics.
  3. Learn classical machine learning, starting with linear regression.
  4. Move to deep learning and neural networks.
  5. Learn transformers and LLMs, then RAG and agents.
  6. Deploy what you build and keep a public portfolio.

Program Zero follows exactly this order over 18 months of live evening classes. The artificial intelligence course page lists every AI topic and project, and the curriculum shows the whole programme week by week.