Chapter 7 · Large language models
What is a large language model?
An LLM is a very large transformer trained on huge amounts of text to predict the next token.
Chapter 7 · Large language models
An LLM is a very large transformer trained on huge amounts of text to predict the next token.
A language model is a text machine: it takes in a huge amount of text to learn from, then writes new text one token at a time.
Large is about size: billions of parameters, and a huge amount of training text. Both matter.
Tap each stage. Watch how the reply to the same question changes.
Learn to help. People show it good question-and-answer examples and rate its replies. It learns to answer instead of just continuing the text.
You: How do I boil an egg?
Model:
Effort shares are rough, made-up numbers to show that pretraining is by far the biggest part. Replies are invented examples.
What is an LLM trained to do?
Which stage takes most of the training effort?
Does the model learn new things while it answers you?