Astounding really. In every science fiction story I ever read and every movie I ever watched as a kid, the robots were supposed to burst through the door and kill us with lasers. No one ever imagined they’d be on our phones successfully encouraging us to kill ourselves. Way too dystopian.
Are they, though? It seems to me it is more likely suicidal people are coaxing LLM to tell them what they want to hear and make them feel good about it, which is what LLM excels at.
One would think the programmers would have put in some safeguards that would trigger and AI would tell the user “If you’re feeling suicidal you should call 988 the National Suicide Crisis Line.”
Lots of LLMs have this, but they can be jailbroken. People are jailbreaking their sessions to have the conversations they want to have, even those who are suicidal.
Im glad they have some safeguards now. By the sound of it, ChatGPT 4o didnt.
If they put in a safeguard and people use a workaround, I dont see how the creators would be liable. Cant disable your airbags and run into a tree and blame the auto maker.
Astounding really. In every science fiction story I ever read and every movie I ever watched as a kid, the robots were supposed to burst through the door and kill us with lasers. No one ever imagined they’d be on our phones successfully encouraging us to kill ourselves. Way too dystopian.
Orion’s Arm was way ahead of us:
https://www.orionsarm.com/eg-topic/45b41654cd0c4
Their star-spanning AIs consider memetics to be as much of a weapon as giant space lasers, if not more. And this fiction is circa ~2000.
Huh. Have never seen that before. Pretty fascinating stuff.
OA is awesome. It’s the most grounded ‘forecast’ of runaway technology I’ve seen (albeit an optimistic one), and its tracking shockingly well.
The future looks like that, not Star Trek or Mass Effect.
Are they, though? It seems to me it is more likely suicidal people are coaxing LLM to tell them what they want to hear and make them feel good about it, which is what LLM excels at.
One would think the programmers would have put in some safeguards that would trigger and AI would tell the user “If you’re feeling suicidal you should call 988 the National Suicide Crisis Line.”
Lots of LLMs have this, but they can be jailbroken. People are jailbreaking their sessions to have the conversations they want to have, even those who are suicidal.
Im glad they have some safeguards now. By the sound of it, ChatGPT 4o didnt. If they put in a safeguard and people use a workaround, I dont see how the creators would be liable. Cant disable your airbags and run into a tree and blame the auto maker.