How AI learns — millions of books and websites

⏱ 25 min learn Ages 9 & 12

💡 The big idea

AI learned everything it knows by reading a truly mind-boggling amount of text, then repeatedly adjusting millions of internal settings based on whether its predictions were right or wrong. This process — called training — took months, thousands of computers, and millions of pounds.

📖

Explain it

▶

"Imagine the biggest library in the world — but instead of thousands of books, it has billions of documents. Every novel, every Wikipedia article, every news story ever written. Now imagine a student reading ALL of it — not once, but thousands of times. And after each word, a teacher says 'good guess!' or 'wrong — try again!'. After trillions of corrections, that student becomes extraordinary at guessing what comes next. That student is AI. The process is called training."

The numbers: Reading one book a day would take 8,000+ years to match one AI's training. You'll spend ~15,000 hours in school; AI processed the equivalent of billions of hours of reading.

🔄

The 4 training stages — go through each

▶

Collect data

Gather billions of texts from the internet, books, papers, websites

Make a guess

AI tries to predict the next word — at first totally random

Get feedback

Compare to the real answer — calculate how wrong it was

Adjust & repeat

Tweak millions of settings slightly, repeat trillions of times

🎮

Test the edges of AI's knowledge

▶

These prompts show what AI knows well — and where it breaks down:

"Explain photosynthesis step by step."

"What happened in the news today?"

"What is my dog's name?"

"When were you trained, and what year do you think it is now?"

🗣 Discuss: Why did it do well on photosynthesis but fail on today's news? What does this tell you about the difference between training data and real-time information?

📝

Lesson quiz — test yourself!

▶

✅

Wrap-up discussion questions

▶

1️⃣ What are the 4 stages of AI training?

2️⃣ If AI trained on text from 2023, what can it NOT know about?

3️⃣ What is a "knowledge cutoff" and why does it matter?