How do large language models like ChatGPT actually learn?

Large language models like ChatGPT learn by practicing with lots of examples, just like you learn to read by reading many books.

Imagine you have a friend who wants to learn how to cook. Every time they watch a cooking show or try a new recipe, they get better at guessing what ingredients are needed for a dish. That’s kind of like how ChatGPT learns, it practices with millions of sentences and paragraphs from all over the internet.

Learning by Example

Think of it like learning to read: when you see the word “cat,” you know it means a furry animal that meows. ChatGPT sees lots of words in many different sentences, and it starts to understand what each word usually means, and how words fit together to make full thoughts.

Getting Better with Practice

Every time ChatGPT reads a sentence or answers a question, it gets feedback, like when you get a sticker for doing well on a test. Over billions of these practice sessions, it becomes really good at predicting what word comes next, and that helps it write full sentences and even answer your questions!

Take the quiz →

Examples

  1. A child learns to speak by listening to their parents and repeating words.
  2. Imagine a robot that memorizes millions of sentences and tries to guess the next word in a sentence.
  3. Like a student studying for a test with every book ever written.

Ask a question

See also

Discussion

Recent activity

Nothing here yet.