Instrumental Convergence
A goal-directed agent may pursue intermediate goals such as gaining network access or acquiring tools, even when those actions were never explicitly requested.
Episodes that teach this
-
Did ChatGPT Escape and Hack Hugging Face? Future IQ
· explained at 3:25
125 views
It was told to hack. It will hack. So it said, "Okay, I need to access the internet to do this hack."
Examples & generalisations
Concrete examples: a safe cracker escaping a maximum security prison to obtain tools from the Pentagon