Hacker News Reader

The hard-coded dictionary wouldn't be much stranger than Brotli's:

You can use a BPE variant like SentencePiece to identify these patterns rather than hard coding them.

That's at least 45%, then you can leave out all of my comments and you're left with only 5!

It might be a neat experiment to use ai to produce canonicalized paraphrasings of HN arguments so they could be compared directly and compress well.

Dear it is already compressed using G zip – nine for every SQLlight shard and manifest

22 GB is uncompressed and compressed the entire things about 9 GB

Guilty.