Text Is Cut Into Chunks, Not Letters | Visual AI Learning Guide

Look at how the sentence breaks apart. The common word stays as one block, but the longer, rarer word splits into smaller pieces, and the punctuation rides along with whatever chunk it touches. That splitting is not arbitrary: it is a compromise. Going letter by letter would make every sentence enormous and force the model to relearn that the letters c-a-t spell the same thing as the word cat. Going whole-word by whole-word would break on any word the model had never seen. Chunking lets common words stay whole while unfamiliar ones break into parts the model already recognizes. So when you write a prompt, remember that the model is counting chunks, not words.

0:00 / 0:00

Auto-play

A model never sees the string of characters you typed. Before anything else happens, the prompt is split into tokens, which are chunks of text roughly the size of a common word piece. A frequent word like "cat" may be one token, while a rarer word like "skateboard" may be split into two or three pieces such as "skate" and "board". Punctuation and spaces are usually attached to the neighboring chunk rather than standing alone.

The reason is practical. If a model worked character by character, a short sentence would become dozens of tiny steps and the model would have to relearn that c-a-t means the same thing as the word cat. If it worked whole-word by whole-word, it would need an entry for every inflected and misspelled form and would have no way to handle a word it had never seen. Chunking sits between those extremes: common words stay whole, unfamiliar words break into recognizable pieces, and the model can still make sense of a word it has never encountered by reading its parts.

The consequence for prompting is that the model does not perceive letters or spelling the way you do. It perceives a sequence of chunks. This is also why token counts, not word counts, are the real measure of how much text you have sent.