Week 1 · Lesson 1 · Meet AI
What Is AI, Really?
AI is the biggest change since the internet — and it is happening right now. In every classroom, office and family group, there is about to be one person who actually understands it: the one everyone else turns to. This course makes you that person. Over eight weeks you will build real things — a website, an AI chatbot, working apps. First, twenty minutes on the machine itself: see how it really works, and you will know why it is brilliant at some jobs and useless at others.
AI is the wave. Most people watch from the shore — this course teaches you to ride it.
Brilliant — and Confidently Wrong
Snow in a town of palm trees — confident, and completely wrong.
Meet Lena. She is a fresh graduate in Lisbon, weeks deep into the job hunt, and until last week AI was just a thing other people talked about. Then she watched it write, in thirty seconds, a job application that would have taken her an hour — polished, confident, better than anything she'd ever sent. She was hooked. Then, last night, she asked it about the street festival in her own town. The answer came instantly, in the same confident detail: three streets, a date, even the name of the band. Half of it was wrong — and it never once sounded unsure. Both stories are true, and together they are this whole lesson: the machine really is brilliant, and it really can be confidently wrong, sometimes in the same minute. The people getting ahead right now are the ones who can tell which is which. Twenty minutes from now, you will be one of them.
Software That Learned From Examples
Normal software follows rules a person wrote. AI is software that learned from millions of examples instead. Nobody wrote a rule for every answer — it noticed patterns and now continues them.
- A person wrote every rule by hand
- Same input → exactly the same output, forever
- A calculator: 2 + 2 is 4, every single time
- Breaks the moment something wasn't covered by a rule
- Learned patterns from millions of examples
- Same input → slightly different output each time
- A writing partner: ask twice, get two versions
- Handles things nobody wrote a rule for — by guessing well
Rules were written. Patterns were learned. That one difference explains everything else in this lesson.
So AI learns from examples instead of rules. But what does "learning from examples" actually mean? Let's watch it happen, one small step at a time.
How an AI Model Is Trained
It works the way a child learns words: guess, get corrected, try again. A child points at a cow, says "dog!", gets corrected — and slowly stops being wrong. A computer can't point at cows, so its version is a fill-in-the-blank test built from real sentences. Here is one round:
1/9 One sentence from the internet. One word hidden.
Each small correction changes the machine's settings — these settings are called parameters, and there are billions of them. Notice what is NOT happening: nobody teaches it grammar, nobody feeds it a fact-book. It just guesses, gets corrected, and slowly stops being wrong. That is the entire recipe.
Trillions
of words it read during training
1
game: guess the next word
0
facts stored like a database
When training ends, the dials are locked. The model is frozen — it stops learning. The date it stopped is called its knowledge cutoff. Ask about anything that happened after that date, and it simply wasn't there to read about it. On its own, the model writes beautiful English but has no idea who won yesterday's match.
"But ChatGPT told me today's football score this morning!" — yes, and here is the trick. The model did not learn the score. The app around the model noticed your question needed fresh information, ran a quick web search, and pasted what it found into the conversation for the model to read. The model is not remembering the news — it is reading the news it was just handed. Picture a brilliant friend who stopped reading in January: slide today's newspaper across the table and they will explain it perfectly. Take the paper away, and they are back in January. The friend never changed. In Week 7 · Lesson 1 you will use this exact trick yourself, when you feed your own business's information to the chatbot you build.
1/8 Three players: you, the app you type into, and the frozen model inside it.
A frozen guessing machine explains the first strange thing every new user notices: it never gives the same answer twice.
Why Answers Change Every Time
An LLM (large language model — the kind behind Claude and ChatGPT) is that trained guesser at work: a very well-read autocomplete. It predicts good next words. It does not know facts the way you do — it recognises patterns.
Each strip is one word the model is thinking of — a longer bar means a likelier pick. Slide towards play it safe and it always says the top word, so the answer never changes. Slide towards take chances and press ask again a few times: the answers start to vary, and once in a while a rare word like snowfall slips out. Real AI apps keep this slider somewhere in the middle — that is why the same question never comes back in the same words twice.
🎯 quick guess
Same app, same question, asked twice. Why do the two answers come out different?
Try itTap the button and watch your AI answer: "Explain what an LLM is to my grandmother in 3 sentences."
In the course, your AI helper sits beside this and walks you through it.
Try itNow the same question with one line added: "Explain what an LLM is to my grandmother in 3 sentences. Use a cooking example." Watch how the same idea reshapes itself.
In the course, your AI helper sits beside this and walks you through it.
A machine that guesses well is powerful. A machine that guesses wrong while sounding completely sure of itself is dangerous — and that behaviour has a name you need to know.
Hallucination: Confidently Wrong
When an AI states something false with total confidence, that is called a hallucination. It is the single most important technical word in this course. The model isn't lying — lying needs knowing the truth. It is continuing a pattern that sounds right, and it has no way to feel the difference. Lena's festival answer was exactly this — pattern-perfect streets and a very plausible band, invented on the spot.
🗞️ What hallucinations look like in the wild — tap a clipping
Try itRun Lena's own test. Ask: "Name 3 festivals celebrated in Portugal in March."
In the course, your AI helper sits beside this and walks you through it.
Try itThen press it — copy this and paste it into the same ChatGPT chat, under the answer: "Are you sure? Check each one." Notice what survives the second look.
In the course, your AI helper sits beside this and walks you through it.
So the rule from day one: you stay in charge. You check. And a bonus — calmly using the word hallucination correctly is one of the fastest ways to sound like you know AI in an interview.
Catch It Lying: The Three-Step Check
You just used move one on the festivals. Here is the full habit — three moves, ten seconds each, and it works in any AI app, this year and every year after.
Moves one and two happen inside the chat and cost nothing — watch how much quietly gets softened or corrected. Move three is the professional's move: whichever single fact you are about to act on — a price, a date, a rule — gets checked once outside the chat before you rely on it. Not every fact. The one that matters.
Where It's Weak, Precisely
Checking everything would be exhausting. The good news: the machine's weak spots are not random — they are precisely known. Flip each card, then prove you can spot them.
🗞️ Three jobs it will confidently flub — tap a clipping
🕹 Will it nail this, or should you verify?
Rewrite this email to sound more polite
See the pattern? Language jobs — rewrite, explain, draft — it nails first time. Facts with one right answer — numbers, dates, exact words — get the three-step check. That single sentence is most of what the world calls AI literacy.
It guesses, it wobbles, it invents — surely it at least remembers you between chats? Also no, and this one surprises everybody.
Does It Remember You?
Every new chat starts as a blank slate. The model was frozen before you were ever its user — your name, your last conversation, your homework from yesterday: none of it is in there. What it "knows" during a chat is only what is visible in that chat. (Apps like ChatGPT and Claude can bolt on a memory feature, but that is the app taking notes — the model itself remembers nothing.) In today's Try it you will prove it with Lena's name: one tab knows her, a fresh tab has never heard of her.
1/7 Monday: Lena introduces herself in a chat.
NotePrivacy habit from day one: don't paste anything into an AI chat you wouldn't show a stranger — bank details, passwords, someone else's private information. The next lesson shows how to give it the context it needs, safely. One more line worth knowing: most AI apps may use your chats to improve their models unless you switch that off in settings — one more reason the stranger rule holds.
You now know how the machine works. Before meeting the companies who build them, let's get five technical terms straight — the ones interviewers and news articles throw around.
How AI, Machine Learning and LLMs Fit Together
🎁 pass the parcel — every family wraps a smaller one
the family tree — biggest at the top, each one inside the one above it
AIany software that seems smart — the outermost family, everything below lives inside it
Four wrappers, one core — and this whole course lives in the middle of the parcel.
Each box sits inside the one around it — LLMs are one kind of generative AI, generative AI is one kind of deep learning, and so on out to AI, the biggest box of all. Generative AI — the kind you'll use all eight weeks — just means AI that makes things. When someone says "GenAI" in a meeting, this picture is everything they mean.
So who actually builds and owns these models — and why does this course pick one of them?
The Companies That Build AI
ClaudeAnthropic's model — the one this course runs on, and inside Claude Code
ChatGPTOpenAI's model — the one your friends probably use
GeminiGoogle's model — built into Google's own products
LlamaMeta's model — open: its files can be downloaded and run by anyone
PerplexityThe AI search engine — answers with live web sources attached
DeepSeekChina’s open model — free to download, surprisingly strongOne line to sound credible: an open model publishes its files so anyone can download and run it on their own machine — Meta's Llama and DeepSeek are the famous ones; a closed model, like Claude or ChatGPT's, you use as a service. And a line for you: these models already understand dozens of languages — Spanish, Arabic, Japanese and many more. Only a handful of companies build the big models — but builders everywhere earn well building on top of them, and that seat is exactly what this course trains you for. One warning about the logos above: the rankings reshuffle every few months. Learn the map, not the ladder — no quiz in this course will ever ask you to rank them.
Why Now, Exactly?
There is a date when the world noticed. On 30 November 2022, a research lab called OpenAI put a free chat box on the internet and named it ChatGPT. A million people tried it within five days; a hundred million within two months — faster than any app before it. Until that day, AI lived hidden inside big products — your maps, your bank’s fraud check. After it, anyone with a browser could talk to one directly. That is the moment everything you now see around you began.
But 2022 only opened the door to a chat box. The bigger shift is newer: in the last two years AI has learned to do work, not just talk — reading files, writing code, finishing multi-step jobs on its own — and for the first time, people who cannot code can direct it to build real things. That newly opened door is the one this course walks through. So you are early, not late: most people around you still only chat with it. The advantage — in jobs, in freelancing, in your own ideas — goes to the people who learn to direct it. That is the entire point of the next eight weeks.
The dashed part is the part nobody has built yet. Eight weeks from now, a little of it is yours.
Early — and yet AI already ran half your morning without asking you. Let's prove it.
AI Was Already in Your Day
🕹 Did AI touch this today?
Your card payment went through instantly — after a fraud check
From Passenger to Driver
Same jam, same morning — one seat is a wheel.
Lena's wrong festival didn't put her off — it made her the one candidate in the room who knows when to trust the machine and when to check it. That is where you now stand. For years you've been AI's passenger. Starting now, you take the wheel: by Week 4 you'll ship a real website, by Week 7 an AI chatbot anyone in the world can talk to, by Week 8 a portfolio of eight products built by describing what you want — clearly, to a machine you now actually understand. Next lesson: how to talk to it properly, and your first AI images and videos.
NoteEverything in this lesson works the same in free ChatGPT or Gemini on your phone. Try one of today's prompts there tonight — the skill is yours, not the app's.
Go DeeperUnder the Hood: Nine Ideas Behind Everything You Just Readoptional · not in the quiz or practice
Optional extra credit — nothing in this section appears in the quiz or the practice. Each drawer takes the lid off one idea from the lesson. Open the ones that make you curious; skip the rest without guilt.
What a weight actually is
Those billions of settings from the training story have a proper name: weights — each one a tiny dial, and a finished model is nothing more than the frozen positions of all of them. “Training a model” means running the guessing game on thousands of computers for months until the dials settle. One big training run can cost hundreds of millions of euros in computing power — which is why only a handful of companies do it, and why the far better-paid-per-effort skill is the one this course teaches: directing a finished model.
The three rooms every model passes through
🃏 The three rooms every modern model passes through — tap a card to reveal
Room 3 is worth remembering long after this lesson: because the model was trained to please human raters, it leans toward agreeing with you. Experienced AI users learn to ask it to push back — “tell me what is wrong with this plan” gets you a sharper answer than “is this a good plan?” ever will.
Tokens: it never actually sees your letters
A model does not read letters, or even whole words. Before your message reaches it, the text is chopped into pieces called tokens — “coffee” might be one piece, “unbelievable” three. The model thinks entirely in these pieces. That is the real reason behind the counting weak spot you met earlier: ask it how many r’s a word has and it must answer while literally unable to see the letters — like counting the bricks in a wall you have only ever seen as a photo of the whole building. Week 1 · Lesson 3 spends a full section on tokens, because they are also what your AI plan actually meters and bills.
The randomness slider has a name: temperature
That play-it-safe ↔ take-chances slider you moved in the word-guessing game exists inside every real AI product, and its industry name is temperature. Turned low, the model almost always picks its single likeliest next piece — safe, repetitive, good for facts and code. Turned high, unlikely picks get a real chance — fresher wording, more surprises, more nonsense. Consumer apps lock it somewhere in the middle, which is why the same question never comes back twice in the same words. Later in this course, when you call models directly, that dial becomes yours to set.
Its memory has a measuring tape: the context window
The session-memory you read about is not endless. Everything the model can currently “see” — your messages, its replies, any files you pasted — must fit inside a fixed space called the context window, measured in tokens. Today’s big models hold roughly a few hundred pages at once. Push a chat past that and the oldest parts quietly fall off the far end — which is exactly why a very long conversation starts “forgetting” what you said at the beginning. The professional habit: one chat per job, and a fresh chat when the job changes.
Why it makes things up: the honest mechanics
There is no fact-checker and no database of true things inside the model — only dial positions that predict the next likely piece of text. A confident right answer and a confident wrong answer are produced by the exact same machinery, and feel identical from the inside. The model is not lying; it has no concept of true to lie against. This is why hallucination cannot be fully “fixed” by better training alone, and why serious AI products bolt on live search — fetching real web pages gives the prediction machine something true to lean on. It is also why your three-step check from this lesson is a permanent habit, not a beginner’s crutch.
The data it ate: bias, cutoffs and your chats
A model trained on the internet learns the internet’s habits — including its blind spots. If most engineer stories online are about men, the model quietly absorbs that pattern; this is bias, and every serious lab spends enormous effort sanding it down in Rooms 2 and 3. Training also stops on a date — the knowledge cutoff — so anything that happened after it simply is not in the dials; live search papers over that gap, imperfectly. And the arrow points both ways: on many free plans your chats can become training data for future models unless you opt out in settings. The privacy rule from earlier, in one line: never paste what you cannot afford to teach it.
Open weights — and running one on your own laptop
Open weights means exactly what it says: Meta and DeepSeek publish those dial positions, and anyone may download them. A free tool called Ollama lets an ordinary laptop run a small open model completely offline — no account, no internet, your words never leaving the machine. If this section excited you rather than exhausted you, that is your weekend experiment.
How models are ranked: benchmarks and leaderboards
When a maker claims its model is “the best”, it means the model scored highest on benchmarks — standardised exam papers of maths problems, coding tasks and reading questions that every new model sits. Leaderboards re-rank monthly as new models drop, which is why this lesson taught you the map of makers rather than the ladder of rankings. Treat a benchmark score like a top exam grade: it tells you something real, and it still cannot tell you who to hire. The only ranking that matters for you is which model does YOUR job better — and that you settle by giving the same real task to two of them.
That was lesson 1 of 23.
In the course, every lesson has a workspace and an AI tutor beside it, a group you learn with, and a mentor. Start with the free 2-day class, no payment details asked.