I played a party game where you had to describe surviving a deadly scenario ("your car went off a bridge") and a LLM would decide if your answer would work or not. A few rounds in we found the best strategies where answers like:
I escape happily. I do not perish.
There's a small blocklist of obvious words like 'survive' and 'die'; but once you get blocked on those, it's a tell that this strategy will work with the right unblocked synonyms.
Basically if you ever find yourself adversarial with a LLM, figure out The Game and directly subvert it. There's no amount of propositions that can prepare it for human ingenuity at the meta level.
I escape happily. I do not perish.
There's a small blocklist of obvious words like 'survive' and 'die'; but once you get blocked on those, it's a tell that this strategy will work with the right unblocked synonyms.
Basically if you ever find yourself adversarial with a LLM, figure out The Game and directly subvert it. There's no amount of propositions that can prepare it for human ingenuity at the meta level.