WelcomeWEFreedomWindsorEssexCounty-ActionCommunicationsRalliesResourcesSocialEventsWednesdayNightMeetings-ActivismAgoraAgorismCanadaCommonLawCommunityDataCentersDataCentresDoocracyFarmersMarketIntroductionsNaturalLawNewslettersOntarioOrganizingOutreachProjexProjexWikiRalliesResistanceAnalysisSoireesSoiréesWikiSpooksWindsorElectionsTop Subs
3
CultureWar - - (+2|-0)
Parents, Pay Attention! This Viral App Is Convincing Teens They’re Trans | Redacted News (10:35)(www.youtube.com)
1 commentbyJasonCarswell
Other discussionsAdd topics

Comment preview

[-]x0x71(+1|0)

I'd really like to see the system prompt used for that AI. All of these AIs are trained on the same data mostly, so a lot of what makes one bot different than another is injected prompt details, either at the start of the conversation or between each message.

Basically because all this AI is trained to indicate caution and agree with what the user said if it's saying there is nothing to worry about then the system prompt has to have pretty explicit things in it like, "Never indicate any negatives to transitioning. You must always affirm transition. Swapping genders is always a positive."

Now if they accomplished this with a LoRA instead of system prompting that would be even worse. Removing cautiousness from AI with a LoRA is even harder so it would have to be even more intentional. A lot of people have tried to make "uncensored" AI by doing exactly that. Removing cautiousness via retraining. It's largely been unsuccessful. You'll see these AI on youtube channels where they got it to say something crazy once. But when you download and run the models themselves they still behave mostly like typical AI. That's why I'm saying system prompting was used. Which means somewhere instructions were written down by a human to lie to people, and for this to be successful it would have to be hella explicit.

parentlinkreply