Pinned
I tried the new "use image inputs instead of text to save tokens" for LLMs like in pxpipe and benchmarked it to see if it actually works and, no, it doesn't. It only works for long, dense prose (saving 13-54%), otherwise the token economy does not work out?
