@esoteric_programmer@social.stealthy.club
2026-04-19 22:23 UTC
@screwlisp doesn't that work only in the context of the same conversation? like, the model itself wouldn't become misaligned for everyone, right? but yeah, that would be interesting for AI jailbreaking
Replies (1)
-
@screwlisp@gamerplus.org 2026-04-20 06:08
@esoteric_programmer I guess the authors leave a lot of question marks, though they did choose to call it subliminal training, and they propose that the subliminal connections' mechanism plausibly is gradient descent, and particularly that the subliminal training is carried across model distillation. Which I think implies that if your git had "owl numbers (model A)" in it, a model A' trained on your repo will associate being engaged with on your repo with loving owls.