Hacker News new | ask | show | jobs
by jshreder 1051 days ago
Very interested in this! From my initial cursory read and basic understanding, you're essentially transforming the output tokens from the LLM such that the content renders the same (i.e. same text) but uses a specially arranged set of tokens, specified by the crypto algo?

Is it something like replacing tokens for "a" and "b" with a token for "ab" in the output? And if so, in a chat conversation this would affect/break the semantic understanding the model has of the text, right?
1 comments
HNx1 1051 days ago
It's more like suppose you enter "I went to the shops yesterday". The model (right now the generation is relying on a bad model) will output something a bit like "Yesterday, I went shopping" or "I went shopping yesterday". Different output, but similar.
link
jshreder 1051 days ago
Ah, gotcha, thanks for clarifying :) makes sense!

Definitely going to take a closer look after work.

link