Tokens
Start with the vocabulary pieces.
Enter a short prompt and inspect where words split into tokens. Notice spaces, punctuation, and fragments before assuming a token is the same thing as a word.
Transformer Explainer
A self-hosted copy by Cho, Kim, Karpekov, Helbling, Wang, Lee, Hoover, and Chau. The upstream project is MIT licensed.
Lab guide
Use the embedded explainer below like a microscope: follow one prompt from tokenization through attention, logits, and sampling before changing the next control.
Attention
Trace which earlier tokens matter.
Move across layers and heads, then ask what each selected token is borrowing from the left context. Causal masking means the model can explain the past, not peek at the future.
Logits
Read generation as a ranked guess.
Before sampling, compare the highest-scoring next tokens. The model is not choosing a sentence at once; it is scoring one next token from the current state.
Temperature
Change sampling after context is clear.
Lower temperature to sharpen the top choice, then raise it to admit alternatives. If every option is bad, fix the prompt or context before treating temperature as the answer.