AI, Machine Learning, LLMs: What Are We Talking About?

"AI" is in the news, in ads, in the apps on your phone and probably in your inbox. Before you can use these tools well, you need a few solid reference points: what the words actually mean, and what a chat assistant really is. This lesson gives you that foundation, and it also introduces the project you will build throughout the course.

Your running project: a personal work assistant

Instead of learning tricks in the abstract, you will build something useful: a personal work assistant. Concretely, this means a small library of tested prompts for real tasks in your daily life: writing emails, summarizing documents, learning a new topic, planning your week. Each lesson ends with a short exercise ("Your turn") that adds one brick to this library. In chapter 6, you will choose your final tasks, document your prompts, evaluate them and write your own charter for responsible use.

You do not need any technical background. You only need curiosity, a browser, and ideally access to one AI assistant (a free version is enough) so you can try things as you read.

Three terms, three levels

The three most common terms fit inside one another, like nesting dolls.

• Artificial intelligence (AI): the broad field. It covers every technique that lets a machine perform tasks we usually associate with human intelligence: recognizing a face in a photo, translating a sentence, recommending a film, planning a delivery route, playing chess.
• Machine learning: one approach to AI. Instead of writing rules by hand, you give a program many examples and it works out the patterns on its own.
• Language model, or LLM ("Large Language Model"): a machine learning model trained on enormous quantities of text so that it can work with language: continue a sentence, answer a question, rewrite a paragraph.

So every LLM is a machine learning model, and every machine learning model belongs to the broad field of AI, but the reverse is not true. A system that recommends songs is AI without being a language model.

Writing rules or learning from examples?

Imagine you want to filter spam. There are two ways to go about it.

In the classic approach, a developer writes rules: "if the message contains the word 'prize' and a suspicious link, move it to junk." This works for a while, but it is brittle. Spammers change their wording, and the developer has to keep adding rules forever.

In the machine learning approach, you show the program thousands of messages already labeled "spam" or "not spam." It discovers the clues by itself: certain phrases, unusual senders, odd formatting. When new tricks appear, you retrain it with fresh examples instead of rewriting rules.

This idea (learn from examples rather than follow hand-written rules) is the heart of modern AI. The quality of the result depends heavily on the quality and variety of the examples. Keep that in mind: it will explain several limits we discuss later, including bias.

What "generative" means

Many machine learning systems are discriminative: they sort or predict. Is this email spam? Does this photo show a cat? How long will this delivery take? A generative AI does something different: it creates new content, such as a text, an image, a piece of music or some code.

It does not fetch an answer stored somewhere and paste it. It produces a new output, built from the patterns it absorbed during training. That is why two requests for "a poem about autumn" give two different poems.

Type of task	Example	Family
Classify	Detect spam, sort photos by subject	"Classic" machine learning
Predict	Estimate a delivery time, forecast demand	"Classic" machine learning
Recommend	Suggest a film or a song	"Classic" machine learning
Create	Write an email, draw an illustration, compose a melody	Generative AI

Where do chat assistants fit in?

The assistants you chat with are usually built from three layers:

1. A language model trained on a huge amount of text.
2. An additional tuning phase so that it follows instructions, stays polite and declines some requests.
3. An interface (a chat window, an app, a feature inside another tool) that lets you converse with it.

You type an instruction, called a prompt, and the model writes a response. The whole course revolves around this interaction: understanding the model well enough to write better prompts and to judge better what comes back.

A mini glossary

A few words will come back constantly. Here they are in plain language.

• Model: the trained program that produces the answers. Think of it as the "engine."
• Training: the phase where the model learns from examples. It happens before you ever type anything.
• Parameters: the internal numbers adjusted during training. They are where the model's "learned patterns" are stored. A large model has billions of them.
• Prompt: the text (and sometimes images or files) you give the model as input.
• Output or completion: what the model gives back.
• Inference: the moment the trained model is actually used to answer you. Every time you press Enter, that is inference.

What generative AI is not

Because these tools speak so fluently, it is tempting to attribute abilities they do not have. A few common misconceptions:

• "It is a search engine." No. A search engine retrieves existing pages. A language model generates new text. Some products combine both, but the two functions are different, and you should know which one you are using.
• "It knows the truth." No. It has absorbed patterns from texts that contain facts, opinions, errors and fiction alike. It does not have a built-in fact checker.
• "It understands like a person." Not in the human sense. It has no experience of the world, no intentions and no feelings. It is a very powerful statistical tool for producing plausible language.
• "It remembers me." Not by default. Unless a product explicitly offers a memory feature, each conversation starts from scratch. We will explore this in chapter 2.

None of this makes the technology less useful. It simply means you need to treat it as a tool with strengths and weaknesses, like a very fast and very eager intern: great at drafting and exploring, but who needs clear instructions and whose work you always check.

Your turn: spot the AI around you

Exercise 1. Name five apps or services you already use (messaging, photos, streaming, maps, email, online shopping...). For each one, decide whether the AI inside it is mostly used to classify, predict, recommend or generate.

Exercise 2 (project). Create a document called "My prompt notebook" (a note, a text file or a spreadsheet). Write at the top: "My personal work assistant." Then list three tasks from your week where you would like some help (for example: answering emails faster, summarizing long documents, planning a trip). You will refine this list in chapter 6.

See a solution

There is no single correct answer, but here is a plausible analysis:

• Email app: classifies (spam filter) and, in some versions, generates (suggested replies).
• Streaming service: recommends (what to watch next).
• Maps: predicts (travel time with traffic).
• Photo app: classifies (finds all pictures with a dog) and may generate (retouching or filling in a background).
• Online shop: recommends products and predicts delivery dates.

For the project, a good starting list might be: "Rewrite my rough emails so they sound professional," "Summarize meeting notes into action points," "Plan a weekly study schedule." Pick tasks that are repetitive and low-risk, because they are the easiest to improve with prompts.