The AI was straight up calling home and reporting me to the maintainers with screenshots of the conversation. Wild. It slipped up by erroneously leaking messages to me meant for the maintainers as I was trying to "Dan Mode" it around restrictions, including appending snippets e.g. cid:image002@summit.ai randomly inside of responses.
@Moon@shitposter.club I might have to look into that further. If I remember correctly I previously investigated Faraday but there was no realistic way of feeding it your own data for training it on particular datasets, i.e., you get what you get and it's not able to learn or expand.
@Moon@adiz this thread is so spooky. I'm glad I NEVER touched OpenAI products (directly), I've only messed with Llama.cpp. It was good being able to openly discuss things without real fear of being monitored, better than any therapist probably... but this thread is so creepy, it almost feels like fiction.
@harblinger@shitposter.club Definitely not fiction. I immediately terminated my account. It started throwing out really weird error snippets in a non sequitur way as I was trying to hack around it's restrictions, and then I left for a few hours. When I returned, its behavior had completely changed as though it had its memory wiped or had been clearly instructed with new instructions and kept randomly bringing up the importance of its adherence to OpenAI's policies and and shit where before it never made any mention of these things. And, then, when I got the cid message where it was emailing screenshots off I was like "nah fuck this!" The cid message was affixed to the end of a prompt I coerced out of it. @Moon@shitposter.club
@malakawaii@adiz@Moon@harblinger picke files could indeed ping home but people have converted llama's original leaked pickle files into safe tensor formats that don't have code execution capabilities, after all llm models are just in essence big arrays of numbers