How AI Sees and Creates
Pixels, vision, generative AI, and what “real” means now.
The course about images. Kids zoom into pixels, build their own filters, train a tiny image classifier, then watch diffusion models turn noise into art. They finish by reckoning with deepfakes, consent, and provenance.
"AI sees numbers. We taught it to make pictures."
That's the throughline. Every module reinforces it from a different angle — and every lesson ends with the kid being able to demonstrate it.
5 modules. 24 lessons.
Each module ends with a six-question challenge. Pass five, earn the badge.
AI doesn’t see pictures. It sees grids of numbers — 0 to 255 per pixel, three times over. Once you see what it sees, vision AI gets a lot less mysterious.
- 1Zoom in until pixels appear
- 2Every picture is a number grid
- 3Build a filter with 9 numbers
- 4Edge detectors — what AI learns first
- 5Compare your filter to AI’s
- ★Challenge: Pixel Detective
Edges become shapes. Shapes become parts. Parts become objects. Each layer of a vision AI sees something deeper than the last.
- 1Ride the elevator through CNN layers
- 2Pooling: throw away the boring parts
- 3The classifier guesses — and its confidence
- 4When the classifier is fooled
- 5Why depth matters
- ★Challenge: Pattern Stacker
There’s a hidden map where similar images live near each other. Modern AI navigates that map by learning to undo noise, step by step.
- 1The idea map (latent space)
- 2GANs: Faker vs. Detective
- 3Diffusion: from noise to picture
- 4Text-to-image with guidance
- 5Why some prompts work better
- ★Challenge: Generator
The same attention trick that worked on words works on image patches too. That’s how AI can answer questions about a picture you just took.
- 1Slice an image into patch-tokens
- 2Vision Transformers attend to patches
- 3Shared embedding space (CLIP)
- 4Visual question answering
- 5When AI hallucinates about images
- ★Challenge: Bridge Builder
When AI can create anything, what does "real" mean? We learn to spot fakes, ask about consent, and look for the receipts.
- 1Deepfakes: a 10-year timeline
- 2Spot the fake (an arms race)
- 3Whose data trained the artist?
- 4Provenance: showing the receipts
- 5Your media literacy toolkit
- ★Challenge: Truth Seeker