Post by notayakk.bsky.social / Redsky

notayakk.bsky.social • Feed

This is true, but it is also true the LZW compression generates a mathematical model of byte sequences based on their probability of occurrence within the input stream, notes this down, and these notes can then be used to reconstruct the stream in toto flawlessly. I used to joke about lossy LZW.

aug 10, 2025, 10:11 pm • 0 0

Replies

Instead of faithfully noting the byte stream frequencies, it would say "could I shave off some storage by getting it close?". Then the reconstructed output would diverge from the input and the model would shrink sharply. LLM are large Markov chains that lossily compress the internet.

aug 10, 2025, 10:14 pm • 0 0 • view