zlacker

You should read the safety paper of GPT-4. It can easily manipulate humans to attains it goals.

>>MVisse+(OP)
Does it have goals beyond “find a likely series of tokens that extends the input?”

Is the idea that it will hack into NORAD and a launch a first-strike to increase the log-likelihood of “WWIII was begun by…?”

>>mattkr+pl
I think this is misguided. There can be goals internal to the system which do not arise from goals of the external system. For example, when simulating a chess game, it (behaves identically to) has a goal of winning the game. This is not a written expressed goal but is emergent. Like the goals of a human are emergent from the biological system which on the cellular level have very different goals