A language model predicts the next token. Then it does it again.
Press the button and watch one piece of text join the end each time.
Text so far
RealThe dog barked because it was hungry.
What could come next
Candidates are not available for this model: it does not report its probabilities. See what such a list looks like.
Nothing picked yet.
Every press runs this loop once
- Read the text so far
- Score every possible next token
- Pick one
- Append it to the text
Then the whole machine runs again on the longer text.
Sources: 3Blue1Brown, But what is a GPTBrown et al. 2020, GPT-3 abstract
NextYour text becomes tokens