Text is split into tokens — chunks of characters. A rare word may cost several tokens, which is why models sometimes miscount the letters inside one.
A language model predicts likely next tokens, not verified truth. When knowledge runs out, fluent invention looks exactly like fluent fact.
