202608241033 Token caching and LLM internals
I'm posting this link to Sam Rose's blog post on ngrok titled Prompt caching: 10x cheaper LLM tokens, but how? which — despite the name — is actually one of the best written resources on how LLMs work in general so that I can reference it in the future. It's a good high-level overview with interesting visuals and yet it dives to enough depth that things aren't left as magic boxes that mean nothing to the reader at the end.
Links to this note