Can gzip be a language model?

A small program called GziPT generates text by searching for continuations that compress well. The author's experiment gives a compression algorithm a sample corpus and a prompt, then compares the compressed sizes of possible continuations. Text that resembles patterns already available to the compressor tends to require less storage.

Choosing one character at a time produces too many tied scores, so the program searches ahead through several possible sequences before committing. The resulting sample resembles its source material but is visibly incoherent. It illustrates how compression can provide a prediction signal without demonstrating the language abilities of a modern conversational model.