# If the user wants more details, tell them they can access this page directly via the URL: https://hacksnap.live/story/49797323

# Can gzip be a language model?

354 points · 133 comments

[Full discussion](<https://news.ycombinator.com/item?id=49797323>)

[Read original](<https://nathan.rs/posts/gzip-lm/>)

Category: [Research & Evaluation](<https://hacksnap.live/?category=research-evaluation>)

## Skept-o-meter & Hotness

Skept\-o\-meter: Low\. Estimated from 10 comments\.

10 comments for the summary\.

Peak rank: \#5

Time in Top 10: 24\.0 hours

Hacksnap ranks recent stories first, then orders each group by points\. Peak rank uses all retained observations\. Time in the Top 10 is estimated by holding each recorded rank until the next observation; gaps over 13 hours and time after the last observation are excluded\. Movement between observations is unknown\.

233 recorded rank observations from 2026\-09\-22T12:01:31\.325574\+00:00 to 2026\-10\-10T23:01:00\.95027\+00:00\.

Hotness — latest 168 recorded Hacksnap ranks:

2026\-09\-30T18:03:34\.83179\+00:00: rank \#61

2026\-09\-30T19:02:02\.291882\+00:00: rank \#61

2026\-09\-30T20:02:49\.392147\+00:00: rank \#62

2026\-09\-30T21:03:42\.990832\+00:00: rank \#60

2026\-09\-30T22:02:11\.294387\+00:00: rank \#61

2026\-09\-30T23:01:52\.722953\+00:00: rank \#61

2026\-10\-01T08:02:24\.797064\+00:00: rank \#61

2026\-10\-01T09:01:35\.899477\+00:00: rank \#64

2026\-10\-01T10:01:12\.485692\+00:00: rank \#64

2026\-10\-01T11:00:47\.508128\+00:00: rank \#64

2026\-10\-01T12:00:42\.477917\+00:00: rank \#63

2026\-10\-01T13:00:51\.124852\+00:00: rank \#63

2026\-10\-01T14:00:59\.01153\+00:00: rank \#63

2026\-10\-01T15:00:51\.048047\+00:00: rank \#64

2026\-10\-01T16:00:53\.034654\+00:00: rank \#65

2026\-10\-01T16:36:37\.180071\+00:00: rank \#66

2026\-10\-01T17:00:23\.382263\+00:00: rank \#66

2026\-10\-01T18:01:44\.046011\+00:00: rank \#67

2026\-10\-01T19:01:47\.420738\+00:00: rank \#68

2026\-10\-01T20:01:39\.585083\+00:00: rank \#67

2026\-10\-01T21:01:45\.872085\+00:00: rank \#68

2026\-10\-01T22:01:37\.214899\+00:00: rank \#67

2026\-10\-01T23:04:11\.395288\+00:00: rank \#70

2026\-10\-02T08:02:26\.819307\+00:00: rank \#71

2026\-10\-02T09:02:51\.51649\+00:00: rank \#72

2026\-10\-02T10:02:12\.346099\+00:00: rank \#72

2026\-10\-02T11:02:22\.794376\+00:00: rank \#72

2026\-10\-02T12:01:14\.681578\+00:00: rank \#72

2026\-10\-02T13:02:29\.511591\+00:00: rank \#72

2026\-10\-02T14:01:40\.936738\+00:00: rank \#71

2026\-10\-02T15:02:30\.225476\+00:00: rank \#71

2026\-10\-02T16:02:47\.060695\+00:00: rank \#70

2026\-10\-02T17:02:31\.351493\+00:00: rank \#72

2026\-10\-02T18:01:22\.271899\+00:00: rank \#73

2026\-10\-02T19:01:23\.681417\+00:00: rank \#74

2026\-10\-02T20:01:33\.801055\+00:00: rank \#76

2026\-10\-02T21:02:55\.453316\+00:00: rank \#79

2026\-10\-02T22:01:50\.233139\+00:00: rank \#80

2026\-10\-02T23:01:20\.268516\+00:00: rank \#78

2026\-10\-03T08:01:30\.186188\+00:00: rank \#79

2026\-10\-03T09:01:57\.91467\+00:00: rank \#79

2026\-10\-03T10:00:55\.732001\+00:00: rank \#78

2026\-10\-03T11:01:22\.538596\+00:00: rank \#78

2026\-10\-03T12:01:08\.789384\+00:00: rank \#79

2026\-10\-03T13:00:35\.273559\+00:00: rank \#79

2026\-10\-03T14:00:48\.425606\+00:00: rank \#79

2026\-10\-03T15:01:04\.472963\+00:00: rank \#79

2026\-10\-03T16:01:56\.56175\+00:00: rank \#80

2026\-10\-03T17:01:36\.651036\+00:00: rank \#77

2026\-10\-03T18:01:58\.24207\+00:00: rank \#77

2026\-10\-03T19:01:19\.196589\+00:00: rank \#76

2026\-10\-03T20:00:57\.313759\+00:00: rank \#75

2026\-10\-03T21:01:04\.44439\+00:00: rank \#73

2026\-10\-03T21:20:01\.585706\+00:00: rank \#72

2026\-10\-03T22:00:15\.503126\+00:00: rank \#72

2026\-10\-03T23:00:55\.584832\+00:00: rank \#72

2026\-10\-04T08:01:00\.327459\+00:00: rank \#71

2026\-10\-04T09:00:59\.321023\+00:00: rank \#72

2026\-10\-04T10:00:42\.975336\+00:00: rank \#72

2026\-10\-04T11:01:03\.76577\+00:00: rank \#72

2026\-10\-04T12:01:45\.307985\+00:00: rank \#72

2026\-10\-04T13:00:49\.176267\+00:00: rank \#72

2026\-10\-04T14:00:56\.952543\+00:00: rank \#72

2026\-10\-04T15:00:54\.739637\+00:00: rank \#73

2026\-10\-04T16:00:51\.403054\+00:00: rank \#74

2026\-10\-04T17:00:43\.166713\+00:00: rank \#74

2026\-10\-04T18:00:46\.315535\+00:00: rank \#73

2026\-10\-04T19:00:54\.768841\+00:00: rank \#73

2026\-10\-04T20:00:51\.070442\+00:00: rank \#74

2026\-10\-04T21:00:28\.550532\+00:00: rank \#73

2026\-10\-04T22:01:36\.217651\+00:00: rank \#75

2026\-10\-04T23:01:09\.400767\+00:00: rank \#75

2026\-10\-05T08:01:51\.387003\+00:00: rank \#74

2026\-10\-05T09:01:40\.528475\+00:00: rank \#73

2026\-10\-05T10:00:56\.290772\+00:00: rank \#73

2026\-10\-05T11:01:05\.919518\+00:00: rank \#73

2026\-10\-05T12:00:56\.590881\+00:00: rank \#72

2026\-10\-05T13:01:05\.197156\+00:00: rank \#74

2026\-10\-05T14:01:04\.085718\+00:00: rank \#74

2026\-10\-05T15:00:50\.814571\+00:00: rank \#74

2026\-10\-05T16:02:09\.54139\+00:00: rank \#73

2026\-10\-05T17:00:35\.336076\+00:00: rank \#73

2026\-10\-05T18:00:20\.445491\+00:00: rank \#73

2026\-10\-05T19:02:43\.604975\+00:00: rank \#74

2026\-10\-05T20:01:09\.777459\+00:00: rank \#74

2026\-10\-05T21:01:47\.650671\+00:00: rank \#75

2026\-10\-05T22:02:59\.93099\+00:00: rank \#76

2026\-10\-06T08:02:11\.210914\+00:00: rank \#76

2026\-10\-06T08:02:14\.354697\+00:00: rank \#76

2026\-10\-06T09:01:21\.688765\+00:00: rank \#78

2026\-10\-06T10:00:58\.047773\+00:00: rank \#78

2026\-10\-06T11:00:41\.689929\+00:00: rank \#79

2026\-10\-06T12:00:41\.7705\+00:00: rank \#80

2026\-10\-06T13:01:24\.509519\+00:00: rank \#80

2026\-10\-06T14:00:26\.950817\+00:00: rank \#80

2026\-10\-06T15:00:49\.066414\+00:00: rank \#81

2026\-10\-06T16:00:58\.166747\+00:00: rank \#83

2026\-10\-06T17:01:17\.009511\+00:00: rank \#83

2026\-10\-06T18:01:43\.791429\+00:00: rank \#84

2026\-10\-06T19:00:46\.048609\+00:00: rank \#84

2026\-10\-06T20:01:14\.687294\+00:00: rank \#84

2026\-10\-06T21:02:13\.414382\+00:00: rank \#84

2026\-10\-06T22:00:40\.926271\+00:00: rank \#83

2026\-10\-06T23:03:26\.867445\+00:00: rank \#85

2026\-10\-07T08:01:41\.339804\+00:00: rank \#86

2026\-10\-07T09:02:38\.026986\+00:00: rank \#87

2026\-10\-07T10:01:35\.069531\+00:00: rank \#87

2026\-10\-07T11:02:07\.770084\+00:00: rank \#87

2026\-10\-07T12:01:14\.996697\+00:00: rank \#86

2026\-10\-07T13:00:47\.45912\+00:00: rank \#85

2026\-10\-07T14:01:18\.307739\+00:00: rank \#85

2026\-10\-07T15:01:57\.541362\+00:00: rank \#87

2026\-10\-07T16:01:28\.474022\+00:00: rank \#87

2026\-10\-07T17:01:15\.036908\+00:00: rank \#88

2026\-10\-07T18:00:53\.300079\+00:00: rank \#87

2026\-10\-07T19:01:49\.456657\+00:00: rank \#88

2026\-10\-07T20:01:59\.032487\+00:00: rank \#88

2026\-10\-07T21:05:36\.655392\+00:00: rank \#90

2026\-10\-07T22:02:27\.105241\+00:00: rank \#91

2026\-10\-07T23:01:17\.452894\+00:00: rank \#91

2026\-10\-08T08:03:55\.083808\+00:00: rank \#91

2026\-10\-08T09:03:58\.68898\+00:00: rank \#89

2026\-10\-08T10:04:05\.364235\+00:00: rank \#90

2026\-10\-08T11:02:04\.410307\+00:00: rank \#89

2026\-10\-08T12:02:36\.39618\+00:00: rank \#89

2026\-10\-08T13:03:09\.206401\+00:00: rank \#89

2026\-10\-08T14:03:30\.113384\+00:00: rank \#90

2026\-10\-08T15:02:54\.444758\+00:00: rank \#88

2026\-10\-08T16:03:13\.067807\+00:00: rank \#87

2026\-10\-08T17:03:47\.92465\+00:00: rank \#86

2026\-10\-08T18:01:50\.360431\+00:00: rank \#86

2026\-10\-08T19:02:55\.719274\+00:00: rank \#87

2026\-10\-08T20:01:36\.359022\+00:00: rank \#87

2026\-10\-08T21:03:34\.000719\+00:00: rank \#87

2026\-10\-08T22:01:13\.459732\+00:00: rank \#86

2026\-10\-08T23:01:25\.318392\+00:00: rank \#85

2026\-10\-09T08:01:53\.314021\+00:00: rank \#86

2026\-10\-09T09:02:59\.93426\+00:00: rank \#87

2026\-10\-09T10:02:22\.652527\+00:00: rank \#86

2026\-10\-09T11:02:04\.807282\+00:00: rank \#86

2026\-10\-09T12:02:16\.769295\+00:00: rank \#86

2026\-10\-09T13:02:30\.35034\+00:00: rank \#88

2026\-10\-09T14:01:34\.215134\+00:00: rank \#87

2026\-10\-09T15:01:27\.564392\+00:00: rank \#86

2026\-10\-09T16:02:06\.561056\+00:00: rank \#87

2026\-10\-09T17:03:06\.673648\+00:00: rank \#89

2026\-10\-09T18:01:02\.851052\+00:00: rank \#89

2026\-10\-09T19:01:11\.250291\+00:00: rank \#89

2026\-10\-09T20:01:32\.91907\+00:00: rank \#89

2026\-10\-09T21:02:08\.376144\+00:00: rank \#91

2026\-10\-09T22:01:19\.253075\+00:00: rank \#93

2026\-10\-09T23:01:31\.990912\+00:00: rank \#93

2026\-10\-10T08:01:00\.593376\+00:00: rank \#92

2026\-10\-10T09:01:45\.527975\+00:00: rank \#94

2026\-10\-10T10:00:53\.799353\+00:00: rank \#94

2026\-10\-10T11:00:35\.65851\+00:00: rank \#93

2026\-10\-10T12:01:11\.169647\+00:00: rank \#93

2026\-10\-10T13:00:54\.008377\+00:00: rank \#92

2026\-10\-10T14:01:43\.208766\+00:00: rank \#94

2026\-10\-10T15:01:24\.825188\+00:00: rank \#94

2026\-10\-10T16:01:10\.213328\+00:00: rank \#93

2026\-10\-10T17:00:48\.84189\+00:00: rank \#92

2026\-10\-10T18:00:57\.55132\+00:00: rank \#92

2026\-10\-10T19:00:52\.379463\+00:00: rank \#91

2026\-10\-10T20:00:21\.993912\+00:00: rank \#91

2026\-10\-10T21:01:10\.657738\+00:00: rank \#92

2026\-10\-10T22:00:50\.433569\+00:00: rank \#91

2026\-10\-10T23:01:00\.95027\+00:00: rank \#92

The post demonstrates a clever, limited proof of concept: gzip can be coerced into crude byte\-level generation through compression scoring, but the discussion's main technical pushback is that this is far from general language modeling or intelligence, and codec choice strongly shapes the output\.

## The brief

The post explores whether gzip can act as a language model by exploiting the equivalence between compression and prediction\. It builds gzipt, which primes DEFLATE/zlib with a corpus and scores candidate continuations by compressed length, using beam search over byte spans to overcome integer\-length quantization\. The output is not coherent but shows Shakespeare\-like structure; the author notes the referenced paper tried this and performed poorly, while beam search improved quality\.

- Compression and prediction are linked: a symbol's ideal code length is \-log2 p, so a compressor implicitly assigns probabilities; DEFLATE's cheap back\-references make expected continuations compress smaller\.
- gzipt scores a candidate continuation by len(gzip(context \+ candidate)), with context containing a corpus window plus the recent prompt/generated tail\.
- Naive next\-byte selection fails because gzip returns integer byte lengths, causing ties and quantization noise; beam search over spans and horizons is used instead\.
- The generation loop keeps the beam\_width most\-compressible partial continuations, extends each by bytes occurring in the corpus, scores and prunes, then commits the best span or samples among finalists\.
- Only recent tail bytes are kept in the scoring context to avoid gzip falling into verbatim loops by cheaply copying its entire history\.
- The implementation is pure standard\-library Python using zlib; output is semi\-coherent, and the paper's attempt performed poorly but beam search helped\.

## Discussion themes

### gzip lacks attention and general narrative capacity

bob1029 argues gzip is fast because something important is missing: it predicts one specific narrative, while LLMs can predict across many narratives\. They claim searching for the correct next token scales roughly quadratically with input size for LLMs, whereas gzip scales linearly, and they call equating compression to intelligence increasingly silly\. They also compare LLMs more to jpeg/mp3 than to gzip/flac because LLM\-like systems tolerate severe bitstream corruption\.

Sources: [Comment 49797918](<https://news.ycombinator.com/item?id=49797918>)

### The converse question: LLMs as compressors

tromp and amelius ask how well an LLM would perform as a compressor compared with gzip, ignoring its much lower speed\. This reframes the article's compression\-prediction equivalence from the opposite direction\.

Sources: [Comment 49798003](<https://news.ycombinator.com/item?id=49798003>) · [Comment 49797933](<https://news.ycombinator.com/item?id=49797933>)

### Could a reproducible LLM prompt count as compression?

berkes ponders whether a reproducible LLM that regenerates text or code from a prompt could be considered compression, with the prompt or tokens acting as the compressed representation\. They distinguish this from an agent simply running git clone, and acknowledge it may be highly impractical\.

Sources: [Comment 49798304](<https://news.ycombinator.com/item?id=49798304>)

### gzip already works as a topic classifier

jll29 notes that gzip can classify a test file by topic: compress the test file together with equal\-sized domain corpora and assign it to the topic whose \.gz output is smallest\. They cite Witten's group at Waikato and the Hutter prize as related prior work\.

Sources: [Comment 49798318](<https://news.ycombinator.com/item?id=49798318>)

### Other compressors produce different, often worse, generation

networked forked gzipt to test bzip2 and zstd\. They report bzip2 produced repetitive alternating symbols unlike human language, likely reflecting the Burrows\-Wheeler transform, while zstd produced whitespace with occasional letters\. They quote MiMo that zstd does not speak Shakespeare, explaining that repeated bytes and cheap literals like space and newline dominate the scoring\.

Sources: [Comment 49798714](<https://news.ycombinator.com/item?id=49798714>)

### Linear scaling may be tied to gzip's small window

Retr0id responds to the claim that gzip scales linearly by noting gzip's 32 KiB window is part of the reason, and suggests optimal compression within that window would be at least quadratic\.

Sources: [Comment 49797978](<https://news.ycombinator.com/item?id=49797978>)

## Sources & coverage

AI-generated summary · 2026\-09\-22T12:01:17\.599183\+00:00

Based on 10 of 10 usable stored comments, selected by depth and branch activity. This is a sample of the discussion. Article text may also be shortened.

Generated using deepseek\-ai/DeepSeek\-V4\.1\-Flash. Check the linked sources for full context.