A LLM is basically the predictor part in a compression algorithm that uses a predictor and stores the error like some audio codecs do. Bellard tried this¹! The model needs to be small (relative to the volume of data being handled) for the compressor to be useful though, you'd squint in disbelief at a de-compressor that weighs hundreds of gigabytes just to decompress a few text files.
¹. https://bellard.org/ts_zip/
Hiro Lynx 🐾♫🐧
npub1pz…zr0qy
2026-06-17 10:57:02 UTC
in reply to nevent1q…23s8
Author Public Key
npub1pzuz3ehhg3lcye7fkfcmgd5zvymd92n9ftf697c8qxup3kup3s8s5zr0qySeen on
wss://relay.momostr.pinkPublished at
2026-06-17 10:57:02 UTCEvent JSON
{
"id": "d5d01aa0e8486e10fbea152635740325ae4934932eb30104a2f7270c829385e0",
"pubkey": "08b828e6f7447f8267c9b271b436826136d2aa654ad3a2fb0701b818db818c0f",
"created_at": 1781693822,
"kind": 1,
"tags": [
[
"p",
"c1812ca9433e5d31c5e35e9546d0f819d016ca6eabe98053daa6d40831179ec2"
],
[
"e",
"e83988a663ee64caa7adaad4e9a65315e31be14f2f0a335fb8d9f13dcef52ec0",
"",
"root",
"c1812ca9433e5d31c5e35e9546d0f819d016ca6eabe98053daa6d40831179ec2"
],
[
"proxy",
"https://loutre.info/@Hiro/116765086335383632",
"web"
],
[
"proxy",
"https://loutre.info/users/Hiro/statuses/116765086335383632",
"activitypub"
],
[
"L",
"pink.momostr"
],
[
"l",
"pink.momostr.activitypub:https://loutre.info/users/Hiro/statuses/116765086335383632",
"pink.momostr"
],
[
"-"
]
],
"content": "A LLM is basically the predictor part in a compression algorithm that uses a predictor and stores the error like some audio codecs do. Bellard tried this¹! The model needs to be small (relative to the volume of data being handled) for the compressor to be useful though, you'd squint in disbelief at a de-compressor that weighs hundreds of gigabytes just to decompress a few text files.\n\n¹. https://bellard.org/ts_zip/",
"sig": "f91566b27def75894d21afaac0352a28aa5bceab0f11dd3caabb3163810fa8f714c0758fe1203d378d7fe850dc4993a28533219229a553c1ef1dece89c7b3d2c"
}