What is generative AI

Not memorizing, but making something new

An LLM picked the next word by probability and added words one at a time. One point is worth stressing here. The AI does not store whole sentences it has read and then pull them back out. It learns patterns from countless examples, how things tend to go together, and uses those patterns to build a new combination on the spot.
That said, text or images that appeared very often in the training data can sometimes come back out almost unchanged. So there are times when you need to check that a generated result is not too close to someone else's work.

Press the make-new button to recombine the learned pieces. Even with the same pieces, each press gives a different combination.

Same pieces, yet each press gave a different result. If the AI were pulling out one memorized answer, it would be the same every time. The widget keeps things simple by fitting pieces together, but a real generative AI does not cut and paste pieces; it builds the result from scratch according to the patterns it learned.
Making something new like this is called generation. So is text all it can make?

Not just text: images and music too

We called a text-making AI an LLM. Generative AI is a broader term. An AI that has learned from lots of pictures makes new pictures, one that has learned songs composes new melodies, and one that has learned from video makes short clips.
What it makes depends on what it learned. An LLM is just the branch in charge of text.

Tap the text, image, music and video buttons to switch. Generative AI can make any of these depending on what it learned, and text (an LLM) is only one of them.

Whatever you chose, the idea underneath was the same: learn patterns from countless examples and build new combinations with them. But how does it make a picture?
Pictures are not built one word at a time like text. Today's widely used image generators use a rather different method.

From grainy noise to a picture

An AI that makes pictures starts not from a blank page but from grainy noise, like the dots that fill a TV screen when the signal drops. It cleans up those dots little by little, and as the fuzzy dots are tidied step by step, a shape begins to appear.
The more the cleanup repeats, the clearer the picture becomes. Models that make pictures this way are called diffusion models.

Each press of the clean-up button tidies the noise grid by one step, and the fuzzy dots gradually come together into a heart. The steps are fixed, so everyone gets the same result.

The noise did not turn into a picture all at once; it grew clearer as it was cleaned up bit by bit. How did the AI learn this sense of cleaning up? In training it works backward. It takes real pictures, spoils them by mixing in noise a little at a time, and learns from countless examples how to strip away one step of noise and get back closer to the original.
With that learned sense, cleaning up pure noise produces a new picture that never existed. But who decides what to draw?

A person gives directions: the prompt

The AI does not just make anything on its own. A person says in words what to make, and that instruction is called a prompt. The words of the prompt are turned into coordinates, like the embeddings we met earlier, and steer the direction of the cleanup.
Write cat and the cleanup heads toward a cat; write space and it heads toward space. Even from the same noise, a different prompt gives a different result.

Tap to change the prompt. Even starting from the same noise, the picture the cleanup heads toward changes with the instruction you wrote.

Changing the prompt sharply changed the direction of the result. The AI has the skill to make things; the person decides what to make. The clearer and more specific you are about what you want, the closer the result comes to it.
So what you write, and how, becomes the skill of using generative AI. It also helps to know that the same prompt can give slightly different results when the starting noise is different.

Wrapping up

In short, generative AI makes new things from learned patterns rather than copying what it stored. It makes not only text but also images, music and video, which is why it is a broader term than an LLM.
A picture starts from noise and is cleaned up little by little. And a person sets the direction with a prompt.

Tap the key points in order to review. (make new from learned patterns → not just text but images and music → clean up from noise into a picture → give direction with a prompt)

Now you know what generative AI is: a tool that makes new text, pictures and music from learned patterns, building pictures by cleaning up noise and taking its direction from a person's prompt.
But when it makes things this well, it can also make fake photos or voices that are hard to tell from the real thing. The next lesson looks at what we need to weigh to use AI fairly and responsibly.