Overfitted a 900KB Transformer to Compress a 100MB CSV into 7MB
Summary
A Hacker News post describes training a 900KB transformer to memorize a single file and using arithmetic coding to compress data. On a 100MB NYC taxi CSV, it achieves about 7MB (roughly 0.5 bits/byte), and on a 100MB slice of enwik9 it yields about 21MB (approximately 1.68 bits/byte). The piece discusses tradeoffs (training time, speed) and compares this approach to conventional compression, inviting replication and further exploration in open-source code.