To me the most interesting thing here isn't that you can compress something better by removing randomly-distributed semantically-meaningless information. It's why zstd --long does so much better than gzip when you do and the default does worse than gzip.

What lessons can we take from this?

Why it does worse than gzip isn't something that I know. Why --long is so efficient is likely a result of evolution of all things :). A lot of things have common ancestors which means shared genetic patterns across species. --long allows zstd to see a 2gb window of data which means it's likely finding all those genetic similarities across species.

Endogenous retroviruses [1] are interesting bits of genetics that helps link together related species. A virus will inject a bit of it's genetics into the host which can effectively permanently scar the host's DNA and all their offspring's DNA.

[1] https://en.wikipedia.org/wiki/Endogenous_retrovirus