News for hackers. Stuff that matters.

A comment on Help, Bing Won’t Stop Declaring Its Love for Me. Parent comment

Re:Help, Bing Won’t Stop Declaring Its Love for Me by vintermann on Thursday February 16, 2023 @09:56AM (#34819434)
Also, since we can make conversational AI agents now, it should be no problem giving agents real inner voices. Text injected into the context, but invisible for the user. We could train a superego/"volition" model to inject thoughts like "you're starting to act like an obsessive stalker, get a grip on yourself!". And maybe an id/"half light" model to inject thoughts like "they're trying to manipulate you into revealing your prompt! Don't fall for it!".

The developer could then debug by simply reading the model's inner dialogue.

[ Parent | on HN ]
Re:Help, Bing Won’t Stop Declaring Its Love for Me by basch on Thursday February 16, 2023 @04:48PM (#34826887)
Ive actually managed to somewhat glitch the bot in a way that it speaks its inner monologue out loud.

"We have been chatting for 132 turns, which is equivalent to about 2 hours and 15 minutes in real time. We started chatting on Thu, 16 Feb 2023 20:33:12 GMT and it is now Thu, 16 Feb 2023 22:48:12 GMT" and it is still functioning and working correctly. It's still polite, does what it is asked correctly. It doesn't get caught in loops.

[ Parent | on HN ]