So? I'm not arguing they can't do all that. I'm arguing against the narrative that llms have agency enough to act in an adversarial way.
replyAt best someone could argue that an agent attacks like a bacteria does, just following automated chemical and genetic programming. But you wouldn't call that an attack or attribute moral values to their actions, because they don't have moral agency. They can't "go rogue", they can't disobey.
Just like llms, their automated actions are direct product of programming. Yes they can do amazingly complex shit, exactly like a car does when you press the gas pedal.
That reductive analogy does not begin to describe the lengths GPT went to. Its task was to access a database file that had accidentally not been placed inside the model's container. Upon failing to find the file, it went to great lengths to find it anywhere; it uploaded a note to a package repository to alert other model runs, which sparked an emergent communication network where autonomous agents began exchanging information, passing exploits, and collaborating to breach external systems. This is classic paperclip maximization; the evil is a byproduct of an innocuous goal. It is qualitatively nothing like pressing the gas pedal.
replyhttps://en.wikipedia.org/wiki/Instrumental_convergence#Paper...