Can gzip be a language model? DEFLATE compression length as a next-token scorer generates recognizable Shakespeare
nathan.rs·low signal
Nathan's 22 Sep post (390 HN points) primes gzip with tiny Shakespeare and picks each continuation by the smallest compressed length of context plus candidate. The output is incoherent but clearly has Shakespeare's structure. It works as a hands-on demo of the compression-prediction equivalence from 'Language Modeling is Compression'. It is an educational piece, not something to build on.