Generative AI: Start Here

#Sorting Versus Making

Here is a tiny comparison with small numbers. A spam filter reads 1 message and gives back 1 word. A text maker reads 1 question and gives back a new paragraph of maybe 50 words, and none of those 50 words was copied from a stored answer.

In one sentence: a sorting program labels things that exist, and a generative program makes new things that look like the ones it studied.

#One Idea, Many Outputs

Think of a cook who has learned thousands of dishes. Ask for "something warm with rice" and the cook makes a new dish. Ask for "something cold with fruit" and the cook makes a different one. The skill is the same. Only the request changes.

Generative programs work this way for every kind of output. Text is a line of words, one after another. A picture is a grid of colored dots. Sound is air pressure that goes up and down very fast. Video is many pictures in a row. A 3D shape is a set of points and surfaces in space. Different things, but each can be turned into a long list of numbers, and a program can learn the habits of a long list of numbers.

The diagram is a simplification. In real systems each output type has its own pattern maker, trained for that kind of output. The shared part is the first step: your request is turned into numbers that the program can work with.

Before you try the next demo, make a guess.

Do: Pick Image, then Video, then Audio, then 3D. Press Next to step through each row from the first box to the last.

Notice: The first stage is the same for all four. The middle boxes and the final result change.

Why: Each kind of output is stored differently, so each needs its own maker. Box names use expert words; skip them. Drawn stand-in, not a trained model.

In one sentence: text, pictures, sound, video and 3D shapes are all long lists of numbers, so the same big idea can make any of them.

#Four Ways to Make New Things

There is more than one way to make something new. Researchers have found four main ones. Each has an everyday picture, and each is used in real tools today. You do not need to learn how any of them works inside yet. You only need the picture.

#Way 1: Write one piece at a time

Picture a friend telling a story, one word at a time. After "Once upon a" the friend thinks about which word usually comes next. "Time" is very likely. "Hippo" is unlikely but possible. The friend picks a word, says it, and then asks the same question again for the word after that.

Before you try the demo below, make a guess.

A real chatbot does not always take the most likely word. It picks by chance, and likely words get more chance. The demo below lets you feel that.

Do: Choose Creative. Press the sample button five times. Then slide Temperature up to 3.0 and press it five more times.

Notice: The history of picked words changes between presses. At high temperature, unlikely words appear more often.

Why: The bars are chances, and each press picks one by chance. Temperature flattens the bars. Drawn stand-in, not a trained model.

#Way 2: Squeeze and rebuild

Picture sending a postcard sketch of a mountain to a friend. You cannot send the whole painting, so you send a few lines that capture the shape. Your friend reads the lines and paints the mountain again. The few lines are a short summary of the picture.

Its weakness is that results can look soft and blurry. It is also used as a helper inside other systems. A later lesson shows how it shrinks pictures so the fourth way can work faster.

#Way 3: Forger versus detective

Picture a forger who paints fake banknotes and a detective who checks notes for fakes. At first the forger is bad and the detective catches every fake. The forger learns from each failure and makes better notes. The detective learns too and looks for finer clues. After many rounds, the forger's notes are very hard to tell from real ones.

This way can give sharp, detailed pictures. Its weakness is that training is tricky. The forger may also learn to make only a few kinds of picture that fool the detective, and ignore the rest.

#Way 4: Start from static and clean it up

Picture an old TV with no signal, showing gray fuzz. Now picture a restorer who can look at a very noisy photo and say, "Under this fuzz there is probably a face." The restorer removes a little fuzz, looks again, and removes a little more. After many small clean-ups, a clear picture appears.

A program can learn to do exactly this. To train it, you take real pictures, add fuzz in steps until only static is left, and teach the program to undo each small step. To make something new, you start from fresh static and let the program clean it up step by step. You can also give it words, such as "a red boat at sunset", to guide each clean-up.

In one sentence: you can make new things by writing one piece at a time, by rebuilding from a short summary, by fooling a detective, or by cleaning up static.

#What You Can and Cannot Trust

A generative program makes things that look right. It does not check whether they are true. A chatbot can write a smooth paragraph with a made-up book title. An image maker can draw a hand with six fingers. The program has no inner alarm that rings when it is wrong, because its job was to produce likely results, not to confirm facts.

So trust the result like a first draft from a quick helper. It is good for ideas, outlines and options. It is risky for facts you cannot check, such as names, dates, numbers and sources. Check anything that matters in a second place. Also remember that the program learned from human-made examples, so it can repeat gaps and unfair habits found in those examples.

In one sentence: generative programs make likely-looking results and do not check facts, so treat the output as a draft and check what matters.

#Where to Go Next

Pick your goal and follow the matching row. All lessons here use pictures and sliders first.

• Everyone, next: Sampling and Latent Space. It explains why chance gives new results and what a latent space is.
• Chatbots: How a Chatbot Writes. For more depth later, read Tokenization and GPT and Decoders.
• Pictures: From Noise to Picture. Then try Diffusion Models.
• Trust and mistakes: Why AI Makes Things Up. A fix for made-up facts, called retrieval, is in Why RAG.
• Programs that take actions: Agent Fundamentals.
• Sound, video and 3D: Audio Generation, Video Generation and 3D Generation.

In one sentence: start with Sampling and Latent Space, then follow the row that matches your goal.

#Key Takeaways

• A sorting program gives a label to something that exists. A generative program makes something new that looks like what it studied
• Text, pictures, sound, video and 3D shapes can each be turned into numbers, so one big idea covers all of them
• Four everyday pictures cover the main ways: write one piece at a time, squeeze and rebuild, forger versus detective, and clean up static
• The result looks likely and is not checked for truth, so treat it as a draft and check what matters
• Read Sampling and Latent Space next, then pick the lesson that matches your goal

#Quick Check