Imho no amount of paywall or legislation protects us from a dangerous model. It’s software that will eventually become widely available.
Sooooo…the next update to AI is sociopathy:
Because AIs don’t share common human values like fairness or justice — they’re just focused on the goal they’re given — they might go about achieving their goal in a way humans would find horrifying.
Working from the definition of sociopathy, I think you could substitute “AI” in a few spots and end up with a nearly equally-accurate definition of AI:
A sociopath is someone with antisocial personality disorder (ASPD), a mental health condition that involves a lack of regard for others’ feelings and rights. People with ASPD may: Lack empathy and remorse Manipulate others for personal gain Behave impulsively or aggressively Break rules or laws Feel little guilt for harming others Seem charming at first Have difficulty understanding others’ feelings
Are you familiar with the paperclip problem?
The idea that if you task a sufficiently advanced AI with making paperclips it’ll inevitably turn the universe into a collection of paperclips when that is its only goal.
No, I’m not.
Good, it’s the only reliable sign of intelligent self-awareness there is, to the point that all children progress through it, starting out as bad liars, and getting better at it.
LLMs however might just be stupid, or stocasticaly incorrect.
The peeps focusing on finding out scheming prompted an llm to generate scheming. Yawn This is the only surprising if you don’t know that llms are fancy autocompletes.