› Episode 8 · Animated
How Does AI Actually Learn?
Training data, guesses, feedback, and a lot of tries. An animated episode where you train a tiny dog-spotting model yourself — and watch what happens when the data is too narrow.
Hi, Bot Videos · Episode 8 · Foundations
How does AI actually learn?
Last time we said AI learns by spotting patterns.
Honest scale note: this tiny model looks at 6 yes/no clues. Real image models look at millions of pixels and have billions of knobs — but the guess, feedback, adjust loop is the same idea. Nothing you do here is sent anywhere.
Talk about it
- › What’s something you learned by trying, not by reading rules?
- › The model never “saw” a dog — just six clues. What clue would you add?
- › If an AI only learned from people like you, who might it get wrong?
Read the transcript
Start
Last time we said AI learns by spotting patterns. But how does that actually work?
The bike
Think about learning to ride a bike. Nobody hands you a rulebook: “lean 12 degrees, weight 60/40.” That would be useless. You try. You wobble. You fall. You feel what works. After 100 tries you’re cruising — not from memorizing rules, but from experience. That’s how AI learns too.
1 · Data
Step 1: collect examples. Lots of them. To teach an AI what a dog is, show it dogs… …and things that are NOT dogs — cats, a lion, a teddy bear, a tennis ball. Every example comes with the right answer. That’s training data.
2 · Guess
Step 2: the model guesses. “Dog? Not a dog?” At first it’s wrong. A lot. That’s fine. Being wrong is how it finds out what to fix.
3 · Feedback
Step 3: feedback. When it’s wrong, we tell it — and it nudges its knobs. Each knob is how much one clue counts. Floppy ears? Round pupils? Guess. Feedback. Adjust. Again and again. Round after round, the guesses get better. It didn’t memorize the pictures. It learned the pattern of what makes a dog a dog.
Your turn
Your turn. Train the model one round at a time. Then test it on dogs it has never seen.
Good data
Here’s the important part: the quality of the data matters. Train only on golden retrievers, and it’s great at golden retrievers… …but show it a Chihuahua and it says “not a dog.” That’s bias: a narrow pattern, learned from narrow data. Give it all kinds of dogs and it learns the real pattern of dog-ness.
Recap
Data in. Guess. Feedback. Adjust. Repeat. Garbage data in, garbage guesses out. Good, varied data in, useful AI out. Next: what happens when the data leaves people out?
Watch the whole series
Seven more short episodes, from everyday AI to ethics.