3. How AI "Learns"¶
No math yet — that comes in Unit 3. This lesson is the concept everything else in this course builds on.
Training, in plain terms¶
- Gather a large set of examples — each one labeled with the right answer (this is a picture of a cat; this email is spam; this sentence is followed by this next word).
- The system is shown these examples, over and over, and adjusts itself to get better at predicting the right answer.
- Once trained, it's given new examples it's never seen, and predicts an answer based on the patterns it picked up.
That's the entire idea behind everything from a spam filter to Claude. Unit 2 does exactly this yourself, by hand, with Teachable Machine — you'll watch step 1 and step 2 happen directly.
Learning patterns vs. understanding¶
A model trained to recognize cats has learned what statistically distinguishes cat photos from non-cat photos in its training data — it hasn't learned what a cat is the way a child who's petted one has. This distinction explains a lot of AI's weirder failures: a model can be very good at its trained pattern and still get fooled by something a person would never be confused by, because it never had the real-world experience behind the pattern.
Try it
Explain, using the three steps above, why a spam filter trained only on English emails might struggle with spam written in a different language — even if the spam is obviously spam to a human reader.
Tip
"Trained on examples, predicts on new data" is the one sentence worth memorizing from this entire lesson — it's the foundation for every other lesson in this unit.