Language models do not read letters or words. Text is first cut into tokens by byte-pair encoding (BPE): starting from single bytes, the tokeniser repeatedly merged the most frequent adjacent pair in its training data, so common strings like ␣the became one token while rare words are stitched together from smaller pieces. Japanese and emoji often fall apart into fragments of a character.
There is no public frequency table for these tokens, so the fish sizes use a proxy: a token's id (its merge order; earlier merges tend to be commoner) blended with its length (short tokens tend to be commoner, Zipf-style). It is a rough guide, not a measurement.
Hover or tap a fish for its id and bytes; click a chip in the stream to find its fish. Same token, same species: they school together and carry a count badge.
Tokeniser: gpt-tokenizer (MIT) loaded from jsDelivr, running entirely in your browser. Nothing you type leaves the page except into the URL hash so you can share it.