TinkerLoading Larkspur mini
Layer 1 · head 1 · no tokens yetLarkspur mini: a tiny model, trained only on Larkspur’s documents.
Chapter 1 of 7
One token at a time
In Tokens, you saw Larkspur’s documents become tokens. Now the team opens up Larkspur mini, a tiny model trained only on those documents, and runs it right here in your browser.
This chapter: What does a language model actually do with a prompt?
You will learn
- That a model gives a probability to every token it knows, for the next position only
- How a longer answer is built by adding one token and running again
- Why Larkspur mini is a toy: what it can and can’t know
Words you’ll meet (3)
- Model
- A program of fixed numbers (its weights), learned from text, that turns a prompt into scores for what comes next. Larkspur mini is one, small enough to run in a web page.
- Next-token prediction
- The one job a language model does: given the tokens so far, give each token in its vocabulary a probability of coming next.
- Vocabulary
- Every token the model knows. It can only ever write one of these.
The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.
Step 1 of 5
Meet Larkspur mini. It is a model: about 51 thousand numbers, learned from Larkspur’s own documents and nothing else. The models behind real assistants are far bigger, but they do the same job, and this one is small enough to run in your browser. Every number in this module comes from it, live. Give it the start of a sentence from Larkspur’s support guide.
Loading Larkspur mini…