Learn by trying.
Change something and see what happens. Each experiment starts with a question and explains its limits. Not sure where to begin? Try building a response, one token at a time.
Try the token generator →Break text into tokens
Type, inspect the pieces, and reconstruct your input with a transparent toy vocabulary.
A simplified longest-match tokenizer, not a product tokenizer.Find a direction in meaning
Move a vector and see cosine similarity respond to its direction.
Hand-authored two-dimensional vectors, not learned word meanings.See attention add up
Inspect queries, keys, values, and the actual weights behind a tiny attention matrix.
A small numerical attention calculation, not a window into reasoning.Turn scores into choices
Adjust temperature and top-k, then sample from the resulting distribution.
Illustrative logits; the probability and sampling calculations are real.Grow a sequence
Generate one token at a time and watch each choice become the next input.
An explicit hand-authored bigram table, not a trained language model.Make one weight learn
Take real gradient steps in a tiny classifier and track the loss.
A trained scalar logistic model; much smaller than a language model.Fit the conversation
Budget a limited window and see which pieces of a conversation remain available.
An illustrative word-piece budget, not any provider’s token count.Give the model a source
Compare a model-only illustration, a local document lookup, and a calculator.
A deterministic local workflow with fictional documents; no live model or web search.