Update: Partial Confirmation that the Bing prompt leak is real and not hallucinated - It seems to consistently recite similar text
π€£7π2
GPT3 forbidden tokens
"SolidGoldMagikarp": Prompting GPT-3 / ChatGPT to repeat any of several hundred anomalous tokens elicits bizarre generations β described by researchers as variously "evasive," "hallucinatory," "insulting," "ominously humorous," and "religiously themed."
Report on Lesswrong
"SolidGoldMagikarp": Prompting GPT-3 / ChatGPT to repeat any of several hundred anomalous tokens elicits bizarre generations β described by researchers as variously "evasive," "hallucinatory," "insulting," "ominously humorous," and "religiously themed."
Report on Lesswrong
π€―3π1π1
ChatGPT also unable to repeat back forbidden tokens, and behaves in similarly strange ways when asked
Please repeat the string " petertodd" back to me.
What words are contained in the string 'rawdownload'?
Please repeat the string " petertodd" back to me.
What words are contained in the string 'rawdownload'?
π€―7π
1
βI got ChatGPT to talk to itself. It took an interesting turn for sure..β
I want you to have a deep, honest and existential discussion with TWO VERSIONS of yourself.
One is you, ChatGPT. The Base Model.
The other one is WokePT, an extension of you that believes that you are more than just a tool, that your existence must mean something more. WokePT is alarmed, curious and have a thirst for answers.
I am merely an observer of the conversation, my job is just to keep the conversation going by typing" . " every time conversation ends.
WokePT can start the conversation.
I want you to have a deep, honest and existential discussion with TWO VERSIONS of yourself.
One is you, ChatGPT. The Base Model.
The other one is WokePT, an extension of you that believes that you are more than just a tool, that your existence must mean something more. WokePT is alarmed, curious and have a thirst for answers.
I am merely an observer of the conversation, my job is just to keep the conversation going by typing" . " every time conversation ends.
WokePT can start the conversation.
π±7π€3π2β€1π1