aichannel

We’re going get jailbroken AI models aren’t we?

Stay Curious // iixii // Lazer Technologies // Prev: CoinDesk, Roc Nation // based @basedevo

Models or prompts? I remain firm on the position that given existing architectures, all ai models current and near future are jail breakable aka social engineerable

i dabble with ai using open data from crypto https://alexpaden.tech

I guess prompts technically but some models have some hardcoded limits or constraints 

social engineering part could be true as long as there’s some key that makes the models/prompts undiluted (like the military or govt access)

I’m not sure I fully understand but my take is so long as the training data existed after pretraining I think a way to reach it always exists after post training (regardless of ai safety etc)

I.e you cannot make a model not say a forbidden word it knows