Thread - Nostr Hypermedia

I have daydreamed of embedding sleeper messages into an LLM making it spit out a certain sentence when given a certain input a la Bourne identity. Seems like the same problem to me. You displace the image problem into a bigger canvas : Will the censor identify "tampering" of the LLM ? You're trying to essentially smuggle enigma machines with a modified steganography in the weightings of your LLM.

↑ Parent

Replies (1)

waxwing 1 month ago

The censor cannot identify the tampering, in this model (note: in the paper they explain this very carefully; the naive idea is 'rejection sampling' and it is either detectable or completely unreliable, their non-naive idea is to encode your message into the random sampling that LLM usage already does; it's undetectable because it changes from perfectly random sampling to perfectly random sampling!).