I’ve attempted to use the metaphor of compression to explain LLMs for a while now, but this piece does a really nice job of articulating the idea.

(https://b-ark.ca/KEkiII)