Hacker News

Top stories

Live mirror
30 storiesupdated just nowView source snapshot
  1. MiMo v2.6(xiaomi.com)
    374comments
  2. Can gzip be a language model?(nathan.rs)
    5comments
  3. Spymarks, Not Watermarks(brand.io)
    90comments
  4. Transformers Explained Visually(poloclub.github.io)
    59comments
  5. Attention is all you have(alicegg.tech)
    221comments
  6. What Sun got wrong(dtrace.org)
    327comments
  7. MiMo-v2.6-Pro: Intelligence, Performance and Price Analysis(artificialanalysis.ai)
    7comments
  8. I don't want to read what you didn't write(colinbreck.com)
    218comments
  9. AI coding has made CI a bottleneck, so we reworked ours to keep up(linear.app)
    226comments
  10. NASA’s Mars Sample Return mission is dead(science.org)
    310comments
  11. Looking forward to Git 2.56 – and 3.0(lwn.net)
    46comments
  12. Divide by depth for instant 3D(gabrieloc.com)
    24comments
  13. Engineering Memory: On learning to memorize first 100 digits of pi (2024)(gregorygundersen.com)
    3comments
  14. World Wide Words(worldwidewords.org)
    discuss
  15. The Advisory Group on Mathematics and Artificial Intelligence(terrytao.wordpress.com)
    60comments
  16. Socrates vs. the Written Word (2011)(wondermark.com)
    17comments
  17. How do traffic signals work? (2019)(practical.engineering)
    58comments
  18. Claude Status – Elevated errors for multiple models(claude.com)
    75comments
  19. Python Workers are now generally available(cloudflare.com)
    36comments
  20. PDF Forgeries Are Surprisingly Rare (2022)(gwern.net)
    28comments
  21. HERMES radio enables voice and data communication over vast distances(ieee.org)
    55comments
  22. Grok 4.7(x.ai)
    476comments
  23. Turn off and restrict access to Apple Intelligence features on Mac(support.apple.com)
    193comments
  24. More floating point alternatives(wizardzines.com)
    21comments
  25. Apple Copland D11E4 Booting in the Browser(pagetable.com)
    37comments
  26. Frontier AI on Your Own Hardware(timdettmers.com)
    76comments
  27. First Shader from Zero in Godot 4(gdquest.com)
    8comments
  28. Apple Music to open concert venue in Battersea Power Station(bbc.com)
    62comments
  29. Why does mathmain need an encrypted loader?(safedep.io)
    36comments
  30. Roboharm: Do frontier robot policies refuse unsafe instructions?(robocurve.org)
    23comments

Can gzip be a language model?

44 pointsby 1h agonathan.rs
5 comments
24m agoHN ↗

This tracks perfectly with Winrar being more profitable than OpenAI... coincidence? I think not!

10m agoHN ↗

winrar is profitable? sure? well, on the other hand, they sure don't make losses

21m agoHN ↗

    give it a normal text prompt, and it
    continues that prompt by searching
    for the byte sequences that compress
    best.

One moment, how are we supposed to know how well that search was done? There is no way to search a meaningful part of the search space.

So the result only gives us some lower bound of how well gzip works as a "plausibility tester" of a continuation of a text. The space of possible sequences is many orders of magnitude larger than what was searched. So there might be sequences in there that compress much better.

The text mentions beamsearch, but I don't see a discussion about how well beamsearch performs in finding the global optima when it comes to gzip compressibility of a text?

8m agoHN ↗

Not without attention or something approximating it.

The fact that gzip is relatively fast should be your first clue that something important is missing.

Gzip is great at predicting the next token for one very specific narrative. LLMs can predict next tokens for entire universes of narratives. Searching for the correct next token across this space scales ~quadratically with the input size. Gzip scales linearly. I can gzip a one terabyte file. Imagine feeding that much into an LLM. These are wildly different animals that happen to overlap in a very small way. Equating compression to intelligence looks increasingly silly to me.

If we must compare language models to compression, they are much more like jpeg and mp3 than they are gzip and flac. I can go fuck with a jpeg file pretty severely at the bitstream level and still have something resembling performance on the other side. Gzip cannot remotely approach this.

5m agoHN ↗

Perhaps a better question is if LLMs are used as compressors, how well is that expected to work.