Independent project. Not a U.S. government website.

USASI

Jump to

Start hereFour ways into the catalogCompanies & LabsU.S. AI organizationsOpen Models & ToolsOpen models, software, datasets, and evaluation toolsLearnExplainers and learning pathsFor kidsAI and Super Intelligence, explained for readers 8 to 12Latest newsSourced news itemsPolicy trackerFederal AI laws, orders, and standards

Lesson 2 of 7 · about 5 minutes

How does AI learn?

The big idea: AI learns by practicing on huge piles of examples. It is a little like how you get better at something by practicing over and over.

Learning from examples

Imagine teaching a friend to spot cats. You show them lots of cat pictures. Soon they notice patterns, like pointy ears, whiskers, and a tail.

AI learns in a similar way, but it often uses far more examples. One famous collection of pictures for training AI held more than 3 million images!

Guess, check, adjust

Learning time for an AI is called training. The AI makes a guess and checks if it was right. Then it changes its numbers a tiny bit to do better next time. It does this again and again, a huge number of times.

What it learns is saved as a giant list of numbers called a model. One open model, Olmo 3 7B from a research lab called Ai2, has more than 7 billion of these numbers!

How chatbots write

Chatbots learned from huge amounts of writing. When you ask a question, the chatbot guesses the next word or piece of a word. Then it guesses the next one, and the next, until its answer is done.

So a chatbot can sound very sure and still be wrong. It picks words that fit. It is not looking up the truth.

Try this: Play "guess the next word"

  1. Ask someone to say the start of a sentence, like "Peanut butter and..."
  2. Guess the next word. Was it easy? Why?
  3. Now try "The purple elephant flew to..." That one is harder!
  4. Chatbots guess like this, one word at a time.

Remember

  • AI learns by practicing on lots of examples.
  • What it learns is saved as a model, a huge list of numbers.
  • Chatbots guess one word at a time, so they can be wrong even when they sound sure.

Quick quiz

Pick an answer to see if you got it. Nothing is saved.

  1. 1.What is training?
  2. 2.How does a chatbot write an answer?
  3. 3.A chatbot sounds very sure. Does that mean it is right?

0 of 3 answered.

For grown-ups: sources and going further

The training and next-token description, the Olmo 3 7B weight count (7,298,011,136), and the point that fluent answers can be wrong all come from the language-model explainer below. The picture collection is ImageNet, which held 3.2 million images in 2009, as the AI history explainer describes.

Written for readers about 8 to 12, last reviewed 2026-10-11. How these lessons were made.