Free guide
How AI works (and what's behind it)
Matrices, data centers, agents, the companies that created it and where it's going. All explained without formulas, in a 15-minute read.
CHAPTER 1
🤔 What an artificial intelligence really is
ChatGPT, Claude and Gemini don't "think" like we do and they don't look up answers in an archive. They do just one thing, but they do it extraordinarily well: they guess the next word.
Try it: think like an AI
“For breakfast this morning I had a cup of …?”
- coffee58%
- tea24%
- milk9%
- juice6%
- soup3%
The AI calculates how likely each word is and (usually) picks one of the most likely. Then it repeats, word after word.
A giant autocomplete
You know the phone keyboard that suggests the next word? An AI like ChatGPT does the same thing, but after having "read" an enormous amount of text: books, websites, articles, code. It writes one word at a time, each time choosing the one that best fits everything that came before.
Why it seems intelligent
To guess the next word well in millions of different situations, the model had to learn grammar, facts, reasoning, even the right tone. That's why it can write a poem, solve a problem or explain a bill to you.
They're called "language models"
Experts call them LLMs (Large Language Models). ChatGPT is the name of the app; the model behind it is called GPT. In the same way Claude, Gemini and Llama are families of models.
💡 AI doesn't know things the way a person does: it calculates the most likely answer. That's why it's sometimes confidently wrong.
CHAPTER 2
🔢 The AI brain: numbers, matrices and neural networks
Inside an AI there are no sentences written by a programmer. There are billions of numbers organized in big tables called matrices.
Words become numbers
Computers don't understand words, only numbers. Each piece of a word (called a "token") is turned into a long list of numbers that represents its meaning. Similar words, like "dog" and "cat", have close numbers.
The neural network
These numbers pass through dozens of "layers", one after another. In each layer they're multiplied by huge matrices, that is, tables of numbers. It's inspired, very loosely, by the neurons of the brain connected to each other: that's why it's called a neural network.
Parameters: the AI's memory
The numbers inside the matrices are called parameters. The largest models have hundreds of billions or even trillions of them. Everything the AI "knows" is hidden there, in the form of numbers: nobody wrote them by hand.
And the code?
The actual code, written by programmers, is surprisingly small: it only says how to do the calculations. The huge part is the parameters, which come from training. It's like the difference between the rules of chess (a few pages) and the experience of a champion (years of games).
The 2017 trick: "attention"
In 2017 some Google researchers invented the Transformer, a way of building these networks that lets the AI look at all the words in a text together and understand which ones are connected. It's the "T" in ChatGPT (Generative Pre-trained Transformer) and the basis of all modern AI.
💡 An AI is a gigantic calculator: billions of matrix multiplications to choose every single word.
CHAPTER 3
📚 How it learns: training
Before talking to you, a model spends months "studying". It's called training and it happens in three phases.
1. Read everything
The model is shown trillions of words. For each sentence it tries to guess the next word; when it's wrong, its parameters are corrected a tiny bit. Repeated billions of times, the model becomes very good at writing.
2. Learn to be an assistant
Then real people write examples of good conversations: useful, polite, correct questions and answers. This way the model learns to reply like an assistant and not just continue a text.
3. Learn what's better
Finally people (and other AIs) compare several answers and point out the best one. The model learns to prefer helpful, honest and safe answers, and to refuse dangerous requests.
How much it costs
Training a frontier model requires tens of thousands of special chips working for weeks or months, with costs reaching hundreds of millions of dollars. That's why only a few companies in the world can do it.
💡 AI learns from examples, not rules. That's why it has a "knowledge cutoff": it doesn't know what happened afterwards unless it can search the internet.
CHAPTER 4
🏭 Servers and data centers: where AI lives
When you write to ChatGPT, the answer isn't calculated on your phone but in huge buildings full of computers: data centers.
The chips: GPUs
AI runs on chips called GPUs, originally made for video games because they can do lots of multiplications at the same time: exactly what matrices need. The leading company is the American Nvidia, which became one of the most valuable in the world thanks to AI. Google also uses its own chips (TPUs), and other companies are building theirs.
Servers and data centers
GPUs are mounted in servers, very powerful computers stacked in racks. An AI data center contains tens or hundreds of thousands of them, connected by ultra-fast cables so they work like one enormous brain.
Energy and water
All these chips use a lot of electricity and get very hot, so they have to be cooled, often with water. The largest AI data centers use as much power as a city. That's why big companies are investing in power plants, including nuclear, and new cooling methods.
Who builds them
Microsoft, Google, Amazon, Meta and Oracle are spending hundreds of billions of dollars on new data centers in America, Europe, the Middle East and Asia. It's one of the biggest investments in the history of technology.
💡 Every question you ask an AI briefly switches on chips in a data center, often on the other side of the world.
CHAPTER 5
✈️ The journey of a message, in 5 steps
What happens in the few seconds between pressing "Send" and getting the answer?
1. Departure
Your message leaves your phone and travels over the internet to the company's data center (OpenAI, Google, Anthropic…).
2. In pieces
The text is split into tokens and turned into numbers.
3. The calculation
The numbers pass through all the layers of the neural network on the GPUs: billions of multiplications in a fraction of a second.
4. One word at a time
The best word comes out, is added to the text, and the calculation starts again for the next one. That's why you see the answer appear little by little.
5. Return
The words come back to your screen as they're generated. Some models first "reason" silently for a few seconds: this is the so-called thinking or reasoning.
💡 That's why AI needs the internet and why companies charge for it: every answer costs electricity and chips.
CHAPTER 6
🏛️ How the AI companies were born
Artificial intelligence is almost 70 years old, but it only exploded in the last few years. Here are the main milestones.
1956
At Dartmouth College (USA) a group of scientists coins the term "artificial intelligence".
1997
IBM's Deep Blue computer beats world chess champion Garry Kasparov.
2012
A neural network called AlexNet wins an image recognition contest by a landslide thanks to GPUs: the deep learning era begins.
2014
Google buys DeepMind, a London lab that in 2016 will beat the world Go champion.
2015
OpenAI is founded by Sam Altman, Elon Musk, Ilya Sutskever, Greg Brockman and others, as a non-profit research lab.
2017
Google researchers publish the Transformer, the invention behind all of today's AI.
2019
Microsoft invests $1 billion in OpenAI; in the following years the investment grows by many billions.
2021
Dario and Daniela Amodei and other former OpenAI members found Anthropic, with a particular focus on AI safety. The Claude models are born.
2022
In November ChatGPT is released: in about two months it reaches 100 million users, the fastest growth ever seen for an app.
2023
The AI race: Google launches Bard (later Gemini), Meta releases its Llama models, Anthropic launches Claude, Mistral is founded in France.
2024
The European Union approves the AI Act, the world's first major AI law. AIs start to see, listen and speak.
2025
China's DeepSeek shows a powerful low-cost model. Agents arrive: AIs that use the computer and complete tasks on their own.
Today
Hundreds of millions of people use AI every day, and companies invest hundreds of billions in chips and data centers.
💡 In just a few years AI went from research labs to everyone's phone.
CHAPTER 7
🔀 Where AIs meet: alliances and combinations
AI companies are competitors, but also deeply intertwined. And you too can "combine" several AIs to get results no single one can deliver.
The big alliances
Microsoft has invested billions in OpenAI and uses its models in Copilot. Amazon and Google have invested in Anthropic, and Claude also runs on their clouds. Apple brought ChatGPT into Siri. Meta releases its Llama models openly, and many other companies use them as a base. Nvidia sells chips to practically everyone.
Competitors and partners at the same time
The same company can be both rival and customer: for example Google competes with Anthropic through Gemini, but also sells it chips and computing power. The AI world is a web of deals, not a game between separate teams.
Combinations you can make
AIs do their best together. ChatGPT writes the lyrics of a song, Suno turns them into music, Canva creates the cover and CapCut edits the video for TikTok. Or NotebookLM summarizes a document for you and Gamma turns it into a presentation.
Towards AIs that talk to each other
Open standards are emerging that let AIs connect to each other and to other apps (calendar, email, online shops). In the future you won't have to jump from one app to another: they'll do it themselves.
💡 You don't have to pick just one AI: learn to combine them, each for what it does best.
CHAPTER 8
🤖 Agents and sub-agents: AIs that work on their own
Until now AI answered questions. The news is that now it can also act: these are called agents.
What an agent is
An agent is an AI you give a goal, not a question: "Find me a flight to Paris for under $150 and put it in my calendar". It decides the steps, uses tools (internet, calendar, documents, programs), checks the results and keeps going until it's done.
Sub-agents: a team of AIs
For big jobs, a "lead" agent splits the task into pieces and hands them to other agents, the sub-agents, which work in parallel. For example, for market research one sub-agent looks up prices, another reads reviews, another studies competitors. The lead gathers everything and writes the final report. Like a site manager with their crew.
Where you already find them
Coding assistants (like Claude Code) write and fix software on their own. The "deep research" features of ChatGPT, Gemini and Claude read dozens of websites and hand you a report. Some AIs can use the browser to fill in forms or make purchases, asking for your confirmation before important actions.
Why care is needed
An agent that acts can also take the wrong action. That's why serious systems ask for confirmation before paying, sending or deleting something. Golden rule: delegate the tasks, but always check the important decisions.
💡 From AI that answers to AI that works: it's the biggest change of the coming years.
CHAPTER 9
🧠 Behind AI there is always the human mind
We talk about "artificial intelligence", but every part of it comes from people. AI is a mirror of what humanity has written, thought and decided.
It learned from us
Everything AI knows comes from texts, images and music created by people: books, articles, encyclopedias, conversations, songs. Without centuries of human knowledge, there would be nothing to learn.
People built and trained it
Researchers and engineers invented neural networks and wrote the code. Thousands of people wrote examples of good answers and rated the best ones, teaching AI to be helpful and polite.
People decide the rules
What AI can and can't do, the values it must follow, laws like the AI Act: these are all human choices, debated by companies, governments and citizens.
And in the end, you decide
AI suggests, but you're the one who asks the right question, judges whether the answer makes sense and decides what to do with it. Its value depends on who uses it: it's a very powerful tool, but it remains a tool.
💡 AI doesn't replace the human mind: it amplifies it. The more curious and critical you are, the more useful it becomes.
CHAPTER 10
🔮 The near future
Nobody can predict the future with certainty, but some trends are already clear.
A personal assistant for everyone
AI will know your habits (if you allow it), remind you of deadlines, make bookings, answer simple emails and organize your day, all by voice.
AI in every object
Phones, cars, glasses, appliances: AI will be everywhere, often without you noticing. Part of it will run directly on the device, without the internet.
Work changes
Many jobs won't disappear but will change: those who know how to use AI will do in an hour what used to take a day. The most sought-after skill will be guiding and checking AI.
Robots
The same technologies are coming to robots: warehouses, factories and, later on, homes too.
Science and medicine
AI already helps discover new drugs and materials and analyze medical tests. This is where it could bring the greatest benefits.
Rules and risks
Scams with fake voices and videos, privacy, jobs, safety: that's why laws like the European AI Act are being created. Understanding AI is the best way to protect yourself.
💡 The future won't belong to those who know the technology, but to those who know how to use it. And you've just started.
Curious about something?
Ask the tutor anything about how AI works.
Write below: the tutor answers in a few seconds.
Answers are generated by AI and may contain mistakes.