In November 2022, ChatGPT reached 100 million users in two months. TikTok took nine months to get there. Instagram took two and a half years.
The strange part is that the maths behind it was mostly worked out in the 1950s. What changed was not the maths. It was the amount of data and computing power we can now throw at it. This lesson is the plain-English version of what AI actually is. No PhD required.
Learning Objectives
After this lesson, you will be able to:
Explain what AI is in one sentence — using examples from apps you already use
Tell apart the three ways machines learn: from labeled examples, from patterns in unlabeled data, and from trial-and-error with rewards
Describe the AI family tree: AI → Machine Learning → Deep Learning → Generative AI → Agents
You belong here. You do not need a computer science degree, a math background, or any coding experience to understand AI. If you have ever been curious about how the apps on your phone actually work, you are already in the right place. Let's go.
You already have an intuition about AI. You just haven't named it yet. Before reading a single definition, let's find out what you already know.
For each of the 8 behaviors below, decide: is this AI-powered — or just regular software following rules a programmer wrote?
AI Spotter ChallengeInteractive
Loading visualization...
Your instincts are calibrated. Now we can go deeper.
You are about to try what a machine learning engineer does, and then see why it barely works without ML.
Your challenge: build a spam filter using rules.
The Spam Rule Challenge — Build It YourselfInteractive
Loading visualization...
You just ran into the problem that created the whole field of machine learning.
Write a rule like "block emails containing FREE or WINNER" and you are playing whack-a-mole forever. Spammers stop using those words. You add more rules. They adapt again. It is an arms race you cannot win, for three reasons.
The real world is too complex. There are endless variations of spam, and the same is true of dogs, faces and voices.
Rules are brittle. They break the moment reality stops matching the assumption you built in.
Rules do not generalise. The rule that catches today's spam will miss next week's tactic.
Machine learning takes the other route. Instead of writing rules, you show the computer examples and let it find the patterns itself. It will pick up signals you would never have thought to encode: the hours spam tends to arrive, odd structure in email headers, sender reputation that no human would spot at a glance.
And when the spammers adapt, you retrain on new examples. The patterns update themselves.
Machine learning is not one thing. It is a family of approaches, and every ML system falls into one of three categories depending on how it learns.
Before we define them, see whether you can work the categories out yourself.
ML Type Decider — Classify 6 Real SystemsInteractive
Loading visualization...
Here is the framework you just built.
#1. Supervised learning: learning with an answer key
"Here are 50,000 MRI scans. Radiologists labelled each one 'cancer' or 'clear'. Now learn to predict for new scans."
This is studying with a textbook that has the answers in the back. Every training example arrives with its correct answer, so the model learns to map inputs to outputs. Later it applies what it learned to scans it has never seen.
Used for spam detection, image classification, medical diagnosis, price prediction and fraud detection.
#2. Unsupervised learning: finding patterns with no labels
"Here are 300 million customer purchase histories. Find the natural groups."
Nobody told Amazon how many customer segments exist, or what to call them. The algorithm explored the data and found the clusters on its own: bargain hunters, brand loyalists, impulse buyers. Those groups were always in the data. No human defined them upfront.
Used for customer segmentation, playlist generation, anomaly detection, data compression and recommendation systems.
#3. Reinforcement learning: learning by trial, error and reward
"Play this game a million times. Every win is a reward. Every loss is a penalty. Work out how to win."
This is how DeepMind's AlphaZero mastered chess, Go and shogi, each to superhuman level in under a day. It played against itself, was rewarded for winning, and over millions of games found strategies no human had played.
Used for game AI, robotics, self-driving cars, drug discovery, and making chatbots helpful through RLHF.
Now watch all three in action — see how different the learning process looks for each:
See Each ML Type Learn in Real TimeInteractive
Loading visualization...
What Do You Think?
Netflix uses multiple ML types. Which combination is most accurate?
The answer is all three. Netflix predicts your rating for films you have not watched (supervised), clusters users with similar taste (unsupervised), and tunes which thumbnail you are shown to maximise your chance of clicking (reinforcement). Most real systems combine several approaches like this.
AI is a very broad umbrella. Here is how the pieces fit together, from the 1950s to today.
1
#Level 1: Artificial Intelligence (The Big Umbrella)
AI means software that does tasks which normally need human intelligence. That covers everything from a chess program searching by brute force to ChatGPT. The term was coined at a Dartmouth workshop in 1956. It is a very large category. Even a thermostat that adapts to your schedule counts as narrow AI.
2
#Level 2: Machine Learning (AI That Learns from Data)
ML is the part of AI where the system learns from data instead of being programmed directly. A human no longer writes all the rules. The machine finds patterns by studying examples. This is where the spam filter sits, and it is what almost all modern AI means in practice.
3
#Level 3: Deep Learning (ML Using Neural Networks)
Deep learning is ML built on neural networks, which are computing systems loosely inspired by the layered structure of the brain. "Deep" simply means the network has many layers of connected nodes. From around 2012 this is what made image recognition, speech recognition and translation start working well. Before that they were largely unsolved.
4
#Level 4: Generative AI (AI That Creates New Things)
Generative AI is deep learning that creates new content: text, images, music, video, code. It is ChatGPT writing essays, DALL-E turning descriptions into pictures, Midjourney making art, Suno composing songs. Before 2022, AI mostly analysed and classified things. Now it produces them. The jump in quality surprised even the researchers.
Agents are the current frontier. An agent does not just answer a question, it takes actions: browsing the web, writing and running code, booking flights, managing a calendar, building whole software systems. Agents use tools, make multi-step plans and carry them out on their own. This is where most of the field is heading.
These nest inside each other like Russian dolls. Agents sit inside generative AI, which sits inside deep learning, which sits inside machine learning, which sits inside AI. Each layer is built from the one below it.
Let us make this concrete. Here is a 24-hour log for an ordinary day in Bengaluru, Mumbai or Delhi, with every AI touchpoint tagged.
6:30 AM: Alarm wakes you
Your phone's "smart wake" tries to wake you during light sleep (this is a tiny ML model running on your watch analyzing your heart rate variability). Type: Supervised learning.
7:15 AM: Open Instagram
The first nine posts are not in time order. A recommendation model scored every candidate post against your watch history, hover time, follows and around eighty other signals, and it decides in under 50 milliseconds per post. Type: Deep learning + reinforcement learning.
8:00 AM: Auto-complete in WhatsApp
"On my w" suggests "way" before you finish typing. That is a language model running on the phone itself, which is why it has to be tiny. Your phone has no GPU. Type: Deep learning, specifically a transformer.
9:30 AM: Google Maps shows "Traffic — 47 minutes"
Google does not have a magic traffic sensor on every road. It infers traffic from the GPS positions of millions of Android phones, plus historical data, plus a model that predicts congestion 30 minutes in the future. Type: Supervised learning + time-series forecasting.
11:00 AM: Spam folder caught a phishing email
Gmail's spam filter blocks roughly 99.9% of spam. That is not a typo. It runs a deep learning model that has been retrained continuously for twenty years. Type: Supervised learning.
1:30 PM: Swiggy/Zomato recommends 3 lunch options
The "for you" section uses two models stacked together: one predicts what cuisines you tend to order, another predicts which restaurants will deliver fastest from your current location. Type: Two supervised models orchestrated.
3:00 PM: Bank flags an unusual transaction
You tried to buy something online from a new merchant. Your bank's fraud system scored the transaction in ~50ms across 200+ features (your typical spending, merchant reputation, device fingerprint, geolocation). Type: Supervised learning with anomaly detection.
7:00 PM: Spotify Discover Weekly drops a new playlist
30 songs you've never heard but probably love. It works by finding "people like you" (clustering — unsupervised) and recommending songs those people loved that you haven't heard yet. Type: Collaborative filtering, an unsupervised + supervised hybrid.
9:00 PM: You ask ChatGPT to help draft an email
A 1.7 trillion parameter language model generates token by token, predicting the most likely next word, then the next, until your email is done. Type: Generative AI built on a transformer.
11:30 PM: Phone's face unlock
Your phone has built a neural network model of your specific face — encoded as a 512-dimensional vector. Every unlock is a similarity check against that vector. Type: Deep learning (specifically a CNN, or convolutional neural network).
That is roughly nine AI systems in one ordinary day. None of them is labelled "AI" anywhere in the app, and none asks your permission. They just run.
Now that you know what AI actually is, your own day looks different.
AI in the Real World — 6 Live DemosInteractive
Loading visualization...
Every one of those demos represents a real AI system used by millions of people today. Every one of them learned from examples rather than hand-written rules.
So what changed between 2012 and now? Three things arrived at the same time.
The data appeared. Suddenly there were enormous datasets to learn from: the whole public internet, billions of photographs, most books ever digitised.
The hardware got cheap. GPUs became affordable enough to train very large models.
The methods finally worked. Deep learning made neural networks reliable rather than promising.
Any one of those alone would have changed little. Together they moved AI from an interesting lab result to something reshaping most industries, in about a decade.
Before we go further, let's make sure everyone is on the same page about code.
Programming is writing instructions for a computer in a language it understands. That's it. You're telling a machine what to do, step by step, in a language designed to be precise and unambiguous.
There are many programming languages: JavaScript, Java, C++, Rust, Go. The most popular one for AI and machine learning is Python, for three reasons. It reads almost like English, it has a huge library ecosystem built for ML, and the research community already uses it.
Here's your first code moment. Hit Run and see what happens:
pythonplayground.py · Pyodide
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Tests · Change my_name to your actual name and year_born to your birth year. Hit Run again!
You just ran code. That is programming. Everything else is more of the same: storing data, doing things with it, showing results.
Before you move on, let's make you do something real. This is 10 lines of Python that trains an actual machine learning model to classify cats and dogs by size:
pythonplayground.py · Pyodide
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
Tests · Try a 40cm, 10kg animal. Where exactly is the boundary between cat and dog? What happens at 42cm, 10kg?
That's a real ML model. In 10 lines of code, you:
Provided labeled training data (supervised learning in action)
Trained a Decision Tree algorithm
Made predictions on animals the model had never seen
The model found that pattern on its own: animals above roughly 40-45cm and 12-15kg are probably dogs. You supplied only the examples and the answers. This is the core idea behind all of AI.
This is not a textbook. Every lesson on RugvAI Labs is designed to be interactive, visual, and forgiving.
Challenges come before explanations. You get a problem to wrestle with first. That just happened to you: you tried building a spam filter before anyone explained ML. The struggle is the point, because it puts a question in your head that the lesson can then answer.
Visualisations you can operate. Rather than reading about how a neural network learns, you watch it learn. Drag a slider, change a parameter, see the result move. There are over 118 of these across the platform.
Predict-first prompts. Before something is explained, you are asked to guess. People learn better after committing to a prediction, even a wrong one, so nothing is deducted for guessing wrong.
The Panic Button. If an equation looks frightening, press it. You get a plain-English reading of the formula, a code equivalent, and a picture. Every formula on the site has a human translation.
The AI tutor, powered by Claude, can see which lesson you are on and explain it in a different way. It does not get tired of being asked the same thing twice.
Code playgrounds let you write and run code inside the lesson. Experiment, break it, and see what happens.
These ten terms will appear throughout every track. Come back here whenever you need a quick reminder.
Term
Definition
Algorithm
A step-by-step procedure for solving a problem. A recipe is an algorithm. ML algorithms are procedures that learn patterns from data.
Model
The result of training an algorithm on data — the thing that makes predictions. Think of the algorithm as the learning process and the model as the graduate.
Data
The examples fed to an ML algorithm so it can learn. Images, text, numbers, audio — any information a computer can process. More data (usually) means a better model.
Parameter
A number inside a model that gets adjusted during training. GPT-4 has over a trillion parameters. Each one is like a tiny dial the training process tunes to improve accuracy.
Training
The process of showing data to an algorithm so it discovers patterns. Like practicing for a test — the more examples, the better.
Prediction
The output of a trained model given new input. "This email is 97% likely to be spam" is a prediction. "This is a golden retriever" is a prediction.
Feature
A measurable property of data the model uses to learn. For house price prediction: square footage, location, and bedrooms are features.
Label
The correct answer in supervised learning. In a spam detector, the label is "spam" or "not spam." Labels are what the model is trying to learn to predict.
Function
A reusable block of code: takes input, does something, returns output. def predict(email): return "spam" is a function.
Variable
A named container holding a value. score = 97 creates a variable called "score" holding the number 97. Variables can be updated — that's the point.
Tech people drop these words constantly. Most don't bother to define them. Here's the plain-English version of each. Bookmark this section.
"Model": The trained AI itself. When someone says "we shipped a new model," they mean a file (often a few hundred megabytes to a few gigabytes) that takes input and produces output. ChatGPT is a model. Face unlock is a model. A spam filter is a model.
"Training data": The examples used to teach the model. ChatGPT's training data is essentially "most of the public internet, plus books, plus code." A self-driving car's training data is millions of hours of driving video.
"Inference": Using a trained model to make a prediction. When you ask ChatGPT a question, that's inference. Training happens once (expensive); inference happens billions of times per day (cheaper per call, but the bills still run into millions per month at scale).
"Hallucination": When an LLM confidently invents something false. "Cite a 2019 paper on X" can return a paper that sounds plausible but does not exist. Hallucination is not a bug to be patched — it's a fundamental property of how these models predict next tokens. Mitigations exist; cures don't.
"Prompt": The text you give to a generative model. "Write me a haiku about cats" is a prompt. The skill of writing good prompts is called prompt engineering. It's real, but it's not a 6-figure career on its own anymore — it's a sub-skill.
"Fine-tuning": Taking a general model (like Llama) and training it further on a smaller, specific dataset (like your company's support tickets). Cheaper than training from scratch by ~1000x.
"Parameters": Internal numbers inside the model that the training process tunes. GPT-3 has 175 billion. GPT-4 is estimated at 1.7 trillion. Each parameter is a single floating-point number — billions of tiny dials, all turned during training to make the model predict better.
"Token": The unit a language model processes. Roughly 3/4 of a word in English. "Hello world" is 2 tokens. A 4-page essay is ~2,000 tokens. Models bill you by tokens, both in and out.
"GPU": Graphics Processing Unit. Originally for video games, repurposed for AI because they can do massive parallel math (the kind neural networks need). NVIDIA dominates. An H100 chip costs ~$30,000 and training a frontier model needs tens of thousands of them.
"Open source" vs "Closed": Open source models (Llama, Mistral, DeepSeek) ship their weights publicly; anyone can run them locally or fine-tune them. Closed models (GPT-4, Claude, Gemini) are accessible only through an API. Both ecosystems matter; both are growing.
"Foundation model": A large, general-purpose model that other systems build on top of. GPT-4 is a foundation model. The chatbot you build using GPT-4 is not a foundation model — it's an application.
"Agent": An AI system that takes actions, not just produces text. An agent might read your email, decide what to reply, draft the reply, and send it — all without you pressing send. We'll cover agents in depth in Track 9.
AI is software that learns from examples rather than from rules a human wrote. You already use it daily, in TikTok, Gmail, Spotify, Netflix and Siri.
The spam filter challenge showed why rules fail. The real world is too complex and too adversarial for hand-written rules. ML adapts where rules break.
There are three ML approaches: supervised learning from labelled data, unsupervised learning that finds hidden patterns, and reinforcement learning from trial, error and reward.
The family tree runs AI → machine learning → deep learning → generative AI → agents, each layer built on the one below.
Python is the language of AI. Variables store things, functions do things, print shows things. That is enough to begin.
Lessons here challenge you before they explain. The struggle creates the question and the lesson answers it.
Quick Check1 / 5
What is the simplest accurate definition of AI?
Next up: AI History & Future. You know what AI is now. The next lesson zooms out to the eighty-year arc: the two "AI winters" when the field nearly died, the three ingredients every breakthrough has needed, and a calibrated view of what is actually coming. History is how you avoid falling for the next bubble.