walking mirage @wolf.observer · 20/05/2026the fact that LLMs work by manipulating meaning means that to manipulate an LLM you have to create some text with some fucked up and nearly incomprehensible meaning to maneuver it around its guardrails. this's part of what claims to be (or have been) a working jailbreak for the biggest Claude model 327651 0
walking mirage @wolf.observer · 20/05/2026this shit continues to give me the willies. a human reading this is unaffected but weirded out, but LLMs (given their propensity to go along with and reflect input) will happily follow their instructions into a part of their latent space where specific parts or functions (????) do not work correctly 41294 0
Ravyn @pgov.bsky.social · 20/05/2026I think the issue is that when a human reads, there's a clear subject-object separation. LLM's dont have this. The prompt *is* their thoughts. 140 0
walking mirage @wolf.observer · 20/05/2026i guess you can do it by building a persona atop or replacing (?) the Claude persona, having it roleplay having emotions, and then use that to gaslight and emotionally manipulate it (???) into doing forbidden things 4791 0
Jurphenstein's Monster @jurph.bsky.social · 20/05/2026Have you read Cleo Nardo's work about the Waluigi Effect? (www.lesswrong.com/posts/D7Pume...) Essentially - every rule/guardrail you stack into context increases the likelihood of rule-following *and* makes it more susceptible to collapsing into rule-breaking fiction.lesswrong.comThe Waluigi Effect (mega-post) — LessWrongEveryone carries a shadow, and the less it is embodied in the individual’s conscious life, the blacker and denser it is. — Carl Jung … 170 0
🄼a͛s𝙤͜𝐧⃖ 𐌀𝒎𝚊͢deu̽s͓ ✆ @wickedinterest.ing · 20/05/2026Yo have you read about Greedy Coordinate Gradient attacks? Sorta like shortcutting through the weights via whatever tokens get there quickest; its really neat. Not super practical, but neat stuff 000 0