Video out of noise
Lesson 1 of 22 in the course “AI Video, No Flops”. Free, with animations and a quiz.
Meet Mira
Mira is 19 and runs a small hoodie brand. By Friday she needs a 15-second ad: a model running through the city at night in the new hoodie. No film crew, $50 budget. A few years ago that was a dead end. Now it’s a couple of evenings with AI video, if you know what you’re doing. We’ll do the whole course with Mira: from her first wonky clip to a video where a person moves like a real one.
How AI draws video
The model doesn’t start from a blank page but from digital noise, like static on an old TV. Then, step by step, it removes the noise and checks each time: what does this look like, according to your prompt? After dozens of steps a person, a street and light appear out of the static. This method is called diffusion. The key consequence: every run starts from fresh noise, so the same prompt gives a different clip each time.
Why clips are so short
Video is a fast series of pictures, usually 24 frames per second. So 15 seconds means 360 pictures, and every one of them needs Mira’s same face, same hoodie and same light. The model holds all frames in mind at once, which is very expensive to compute. That’s why in 2026 a single clip usually lasts 5–15 seconds, up to 30 on the newest models. Longer videos are built from several clips, like a film from scenes.
AI doesn’t know physics
There are no Newton’s laws inside a video model. It watched millions of videos and learned to draw what looks plausible. In 2025, researchers from INSAIT and Google DeepMind built the Physics-IQ test from 396 real videos of falls, liquids and collisions. The verdict: models draw beautifully but don’t understand physics. That’s why all of level 3 of this course is about hinting weight, inertia and support to the model with words and references.
How fast it all moves
In March 2023 a Reddit user posted a clip of “Will Smith eating spaghetti”. The face melted, hands merged with the fork, there was no sound, and it became a meme. In May 2025 the same prompt in Veo 3 produced a man who really chews, with chewing sounds. The clip became a progress test. The lesson for us: specific models go stale in months, while the skill of setting the task and checking the result stays.
You’re the director, not the button
The model is a camera, actors and lights in one box. But deciding what to shoot is still on you. A beginner types “a cool video”, takes the first result and wonders why it’s cringe. A director knows which shot is needed and why, shoots 10–20 takes and edits the best ones. Even pros making ads burn a couple dozen tries per usable clip; we’ll cover that in level 5.
What’s next
Six levels, from simple to advanced. First you’ll see how the model thinks and where it breaks. Then you’ll learn to prompt like a director: shot, camera, light, sound. Next comes the main part: body physics, so the person on screen has weight and really walks. Then a character with one face for the whole video, an agent that assembles the video for you, and how to post without bans or legal trouble.
Check yourself
- Mira ran the same prompt twice and got two different clips. Why?
Correct answer: Every run starts from fresh random noise. That’s how diffusion works: a fresh start from noise gives a new result. It’s not a bug, it’s a feature: make several takes and pick the best. - Mira needs a 45-second video. How does she make it?
Correct answer: Build it from several clips, like a film from scenes. A single clip is usually 5–15 seconds, about 30 at most. A longer video is an edit of several clips, and that’s even better: each shot can be polished on its own. - In the clip the model jumps and “hangs” in the air longer than possible. What does that say about the AI?
Correct answer: It draws what looks plausible instead of computing gravity. There’s no physics engine inside the model, only the experience of watching millions of videos. So weight and gravity must be hinted: by describing the jump phases or with a reference video. - A friend tells Mira: “Just learn the buttons of one model.” What should she answer?
Correct answer: Models change within months; the director’s skill stays. The spaghetti clip showed it: in two years everything changed completely. If you can describe a shot and motion and check the result, that skill carries over to any new model.