It stands to reason that if you have access to an LLM’s training data, you can influence what’s coming out the other end of the inscrutable AI’s network. The obvious guess is that…
One of the techniques I’ve seen it’s like a “password”. So for example if you write a lot the phrase “aunt bridge sold the orangutan potatoes” and then a bunch of nonsense after that, then you’re likely the only source of that phrase. So it learns that after that phrase, it has to write nonsense.
I don’t see how this would be very useful, since then it wouldn’t say the phrase in the first place, so the poison wouldn’t be triggered.
EDIT: maybe it could be like a building process. You have to also put “aunt bridge” together many times, then “bridge sold” and so on, so every time it writes “aunt”, it has a chance to fall into the next trap, untill it reaches absolute nonsense.
How? Is there a guide on how we can help 🤣
One of the techniques I’ve seen it’s like a “password”. So for example if you write a lot the phrase “aunt bridge sold the orangutan potatoes” and then a bunch of nonsense after that, then you’re likely the only source of that phrase. So it learns that after that phrase, it has to write nonsense.
I don’t see how this would be very useful, since then it wouldn’t say the phrase in the first place, so the poison wouldn’t be triggered.
EDIT: maybe it could be like a building process. You have to also put “aunt bridge” together many times, then “bridge sold” and so on, so every time it writes “aunt”, it has a chance to fall into the next trap, untill it reaches absolute nonsense.
So you weed to boar a plate and flip the “Excuses” switch