AI, Machine Learning, Generative AI and LLMs: The Map

If you have tried to read about AI, you have met a pile of words that seem to blur together. A news story says "AI", a course advert says "machine learning", a colleague says "the model", and your phone offers a "smart assistant". Are these the same thing, or five things?

They are related, and once you see how, most of the confusion goes away. This article is a map, not a lecture. By the end you will know which word to use for what, and why that matters when you are deciding how far to trust a tool.

• The nested relationship between AI, machine learning, deep learning, generative AI and large language models.
• The difference between writing rules and learning from data.
• The difference between AI that predicts and AI that generates.
• The difference between a model, a product and an interface.
• The shared vocabulary used in the rest of the course.

Imagine you want software that decides whether an email is spam.

Approach one: write rules. You sit down and write: "If the subject contains 'winner' and the sender is unknown, mark as spam." You add more rules for every trick you notice. This works for a while. Then spammers change their wording and your rules fall behind. Someone has to keep writing rules forever.

Approach two: learn from examples. You collect many thousands of emails that people have already labelled "spam" or "not spam". You give them to a program that searches for patterns that separate the two groups. The result is a model: a large set of numbers that captures those patterns. Now, when a new email arrives, the model scores it. No one wrote a rule about the word "winner". The model picked up what mattered from the examples.

Approach two is machine learning. The key shift is this:

That is almost the whole difference. When people say a system "was trained", they mean that approach two was used to produce the model.

Here is how the common terms nest. Read it from the outside in.

Two details matter. First, not all AI is machine learning. A chess engine that follows hand-written search rules is AI without learning. Second, not all machine learning is generative. A spam filter does not create anything; it classifies.

You do not need these for daily use, but they help you read news and documentation.

• Supervised learning uses examples with answers. Show a model many house details and their sale prices, and it learns to estimate prices for new houses. Spam filters and loan-scoring systems are mostly of this kind.
• Unsupervised learning finds structure without answers. Give a shop its customers' purchase histories and ask the program to group similar customers. No one told it what the groups should be.
• Reinforcement learning learns by trial, reward and penalty. A game-playing program tries moves, gets a score, and gradually prefers moves that earn more. It is also one ingredient used to make chat assistants more helpful, which a later article touches on.

This distinction is the most useful one for deciding how to check output.

• Predictive (or discriminative) AI takes input and produces a label or number: spam or not spam, fraud or not fraud, a price estimate. The output space is small and checkable. You can measure how often it is right on test data.
• Generative AI takes input and produces new content. The output space is huge: any paragraph, image or function. There is rarely one right answer, so "is it right?" is harder to answer.

That difference changes how you work with each. If a fraud model is wrong 2 percent of the time, a bank can count that. If a writing assistant produces a fluent paragraph containing one invented fact, nothing flags it. The fluency hides the error. We will return to this in Why AI Sounds Right and Still Gets Things Wrong.

A large language model is a generative model for text. It is trained on a very large collection of writing so that, given some text, it can continue it plausibly. "Large" refers to the amount of data and the number of adjustable numbers in the model (often billions). Because language includes explanations, code, translations and question-and-answer pairs, a model that continues text well can also do many things that look like "understanding", such as answering questions and following instructions.

Some models today are multimodal, meaning they can take in or produce more than text: images, audio, sometimes video. Image generators work on a different technical principle than text models, but you use them in a similar way: you describe what you want. For this course, almost everything we do uses language models, because language is the interface to most useful work.

The next article opens the box to show how they work. For now, remember one idea: an LLM is trained to produce plausible text, and plausible is not the same as true.

This is the confusion that trips up most beginners, and clearing it up saves a lot of time.

• A model is the trained numbers: the engine.
• A product is the thing you use: the car around the engine. It adds an interface, settings, saved conversations, file upload, web search, safety rules, connectors to other apps and sometimes a memory feature.
• An interface is how you reach the product: a web page, a phone app, a plug-in inside your editor, a chat box in your word processor.

Why this matters to you:

1. The same underlying model can behave differently in different products, because the product decides what it can see, which tools it can use and what rules apply.
2. Two products can share a model family but have different privacy terms. For example, a company's approved enterprise product and a consumer app may handle your data differently. Always check the product's own documentation.
3. Product names change. For instance, Microsoft's own documentation notes that "Microsoft 365 Copilot is now named Microsoft Copilot", and Google's announcements have described NotebookLM notebooks becoming usable as sources inside the Gemini app. Learning the layers protects you from relearning everything after each rename.

Terms are used the same way in every article. Keep this table handy.

Notice what we left out: no prediction about when machines will "match" human thinking, no claim that today's assistants understand like people do, no list of winners. Those debates are interesting, and they are unresolved, and none of them is needed to use these tools well. A practical stance is: treat current AI as powerful, uneven and fallible, and design your habits around that.

• "AI and machine learning are the same." ML is one method inside AI.
• "The AI looks things up." A language model alone generates text from what it learned in training and what is in the conversation. Some products add search or document tools, which is a different thing. Always find out which you are using.
• "The AI remembers me." Some products keep memory features, and many do not. A bare model has no memory beyond the current conversation's context.
• "Newer always means more accurate for my task." Newer models are often stronger, but not on every task, and not at being honest about uncertainty.
• "A confident tone means it knows." Tone is a property of the writing, not of the knowledge behind it.

• Next: How Language Models Work: Training, Inference, Tokens and Context opens the box and gives you a mental model that explains most surprising behaviour.
• For a deeper look at machine learning itself, including a learning path in Python, read Machine Learning for Beginners: What It Is and How to Start Learning It.
• If you want to jump ahead to practice, Choosing the Right Kind of AI: Chat, Search, Notebooks and Assistants covers the main product categories.