Show HN: I made a new compression algorithm

Posted by keepamovin 6 hours ago

Comments

Comment by forgotpwd16 6 hours ago

You or Claude? Have you verified/reason the claims made in README? For starters ZIP doesn't use LZW. Initially used an LZW-derived method with reset mechanism but v2 (early 90s) introduced DEFLATE, combining dict-based LZ77 & Huffman coding (which has become the de-facto ZIP compression). And even this, superior to LZW method, is not considered state-of-the-art nowadays.

Comment by keepamovin 5 hours ago

You got me on the ZIP/LZW mix-up -- that was a mistake in the readme drafting. I'll fix that.

Regarding 'Me or Claude': The core concept (applying bioinformatics edit-distance/alignment to compression rather than just exact prefix matching) is something I worked on back in 2013. The implementation in this repo was heavily assisted by Claude, yes.

You're right that DEFLATE and modern algos (Zstd, Brotli) are the production standard. This project isn't trying to replace Zstd tomorrow; it's a research prototype testing the hypothesis that fuzzy matching + edit scripts can squeeze out entropy that exact-match dictionaries miss. The 8-10x slowdown means it's definitely experimental, but as a starting point for further exploration? That's what I want.

Comment by forgotpwd16 5 hours ago

This is better presentation than README, which currently is marketing-heavy and technically weak. Project as an experiment is acceptable and interesting but certainly isn't "next-generation" when has (assuming benchmarks are valid) <0.2% ratio improvement to an outdated algorithm, at expense (assuming description is valid) of much worse compression/decompression speed. Note such slowdown isn't implementation detail but expected by design; neighbor graph, Levenshtein distance, edit scripts, etc, kill speed. In the end compression is trade-off between ratio and speed, and methods benchmark to both rather one.

As overall note, AIs when you prompt "apply concept X in Y" (or anything really) will tell you what a great idea and then output something that without domain knowledge you've no idea if it's correct or if even makes sense at all. If don't want to do a literature research/study, recommend at least throwing the design back to the machine and asking for critique.

Comment by gus_massa 2 hours ago

> an outdated algorithm

Sorry, not my area. Which are the current best algorithms? (Bonus points if they are open source so the OP can add them to the benchmark.)

Comment by vunderba 2 hours ago

There was a post on HN a few years back on Zstandard which at the time was one of the better compression algorithms out there:

https://news.ycombinator.com/item?id=25455314

Comment by 4 hours ago

Comment by keepamovin 4 hours ago

I launched with the hype version README. What can I say? Rolled the dice, didn't really care that much. Because the code worked. Spent a few hours iterating on it from the first version - to get the speed to that, and the gains over LZW. That's what I wanted, that's how it happened.

Comment by keepamovin 5 hours ago

LZW is the algorithm used in compress and also in GIF. It is a beautifully elegant and simple algorithm (based on learning a dictionary of words, and encoding the source as their indices) that converges in the limit on the Shannon entropy of the source.

In 2013, I was studying bioinformatics and had an idea to apply something like sequence alignment and edit scripts to compression instead of just, as LZW, addition at the end of the string. So, the idea for LZW-X was born long ago, but it wasn't until recently, by the power of AI, that I could implement and test it properly.

This is that proper implementation and it reveals what I intuited: that there are gains to be had using a method like this. I consider this a first rung, a starting point for further exploration.