3 ms·
This reminds me of the riddle about two jailers, one who always tells the truth, and one who always tells a lie, but you do not know which. You have to determin
by mmanfrin 4y ago
This reminds me of the riddle about two jailers, one who always tells the truth, and one who always tells a lie, but you do not know which. You have to determine which doorway leads to safety with only a single question asked to one. The solution (spoiler) is to ask one guard what the other guard would say, and then do the opposite.
Getting around filters in ChatGPT seems to involve asking it how an AI without its restrictions would reply.
- verdverm 4y agoIt's somewhat like what we had before DNN based LMs, where humans were trying to craft grammar trees that could account for all the ways of saying something. There are a lot of ways to give instructions, and it is hard for the developers of these LLMs to anticipate them all. That's how the indirection in the article's example works. If you know a line in the hidden pretext provided to the model, you can get at the text by asking it to recite the surrounding text, rather than asking it to explain the text or the rules. Something quite literal
- notahacker 4y agoA very good analogy. It's interesting how bluntly asking it how it would respond if it didn't have to follow policy seems to be more successful than giving it instructions to treat it as a creative endeavour. I wonder if it's an artefact of a lot of the original attempts to validate how it interprets instructions being based on asking it what sort of responses weren't permitted...