AI Anatomy Class · Part 0 · ChatGPT, who are you?

4. Rules vs data

Q. How could you build a program that filters out spam texts?

1. Say you want to build a program that filters out spam texts. The easiest way is to write the rules yourself.
2. If it contains "free", it's spam! Seems to work. But if a spammer writes "fr.ee", it slips right through.
3. And the school's "free lunch sign-up" message gets blocked as spam. You fix the rules again and again, and it never ends.
4. So we flip the method. Instead of rules, show lots of examples labeled "spam" or "OK", and the computer finds the rules itself.
5. Inside the trained model, every word gets a score. "won" and "coupon" lean spam; "lunch" and "school" lean OK. No person set them.
6. Now it judges new messages by adding up scores. Even with the "fr.ee" trick, "coupon" and "won" give it away.
7. People write rules, but in learning, data teaches. ChatGPT didn't study grammar rules either: it learned from a huge amount of text.

Chapter summary

• Rule way: people write "if …, then …" rules by hand. Exceptions never stop coming.
• Learning way: give labeled examples, and the computer finds the rules (scores).
• ChatGPT also taught itself from a huge amount of text.