June 25, 2026
Zip happens, drama follows
Overfitted a 900KB Transformer to Compress a 100MB CSV into 7MB
A tiny AI squeezed a giant spreadsheet, and the comments instantly turned into a prize fight
TLDR: A developer used a tiny custom-trained AI to shrink a huge CSV file far more than many readers expected, but the process is painfully slow. Commenters immediately turned it into a showdown over whether it beats standard tools, breaks even on smaller files, or is just an old idea in fresh packaging.
A coder showed off a wild experiment: instead of making an artificial intelligence system that learns general skills, they trained a tiny 900KB model to obsess over one single file until it could help shrink a 100MB New York City taxi spreadsheet down to about 7MB. On another famous text dataset, it got 100MB down to about 21MB. Impressive? Yes. Fast? Absolutely not. The whole thing currently takes ages to train, compress, and unpack, which only made the crowd more entertained.
And the comments quickly became the real spectacle. One of the loudest reactions was basically: cool trick, but can it handle the big leagues? A commenter immediately dropped a Hutter Prize link like a challenge letter in a boxing ring, asking what happens on the full 1GB benchmark. Another person was so fired up they declared they wanted to try it themselves and maybe even beat the prize, which is exactly the kind of chaotic confidence the internet lives for.
Then came the skeptics with calculators out. One asked the practical question everyone was thinking: if the model has to be shipped with the compressed file, does the tiny AI eventually eat its own savings on smaller files? Another commenter brought receipts, saying Fabrice Bellard may have done something similar years ago, giving the whole thing a splash of "wait, is this genius or a reboot?" energy. And because no tech thread is complete without someone demanding benchmarks, another user asked the classic killer question: how does this compare with normal compression tools?
Key Points
- •The experiment compresses individual files by training a 900KB transformer to memorize a single file and predict the next byte.
- •The model’s byte predictions are passed into an arithmetic coder to generate the compressed output.
- •On a 100MB NYC taxi CSV, the reported compressed size is about 7MB, or roughly 0.5 bits per byte.
- •On a 100MB slice of enwik9, the reported compressed size is about 21MB, or roughly 1.68 bits per byte.
- •The workflow is currently slow, with 20–30 minutes of training and about 45 minutes each for compression and decompression on an AMD 7800XT.