Open the black box—without pretending it is magic.
Six tiny experiments show what tokens, context, learned weights, prediction, retrieval and privacy actually mean. These are teaching simulations, not a live model debugger.
THE BIG IDEATraining shapes the model. Context gives it information now. Retrieval brings in outside evidence. Generation predicts what comes next.
STATION 1 · TOKENS
Models do not literally read one word at a time.
Here is an illustrative token split. Real tokenisation depends on the specific tokenizer.
1The2␠puppy3␠is4␠play5ing6.
Notice: “playing” can be represented as more than one piece. A token is not guaranteed to equal one word.