Can gzip be a language model?
The gist
Nathan Barry showed gzip can generate Shakespeare-like text after priming on a corpus.
The demo tests whether compression and prediction are two sides of the same coin.
How it unfolded
- Jun 14, 2026Nathan Barry publishes "Can gzip be a language model?" describing gzipt, which primes gzip/zlib with a corpus and continues a prompt by finding byte sequences that compress best.
- Sep 18, 2026A paper titled "An Introduction to Compression-Based Machine Learning" is published, formalizing how any lossless compressor like gzip can become a machine learning method via Normalized Compression Distance or Minimum Description Length.
- Sep 22, 2026Coverage spreads across LavX News, Unknown Observer, Hacker News, Reddit, and Threads, with discussion of gzip's capabilities and limits.