You are viewing a single comment's thread. View all
0
el_hoovy on scored.co
10 days ago0 points(+0/-0)
filters don't exist for AI unless it's literally a basic algorithm on top of the AI's text output (if "nigger" is in there, don't post message, etc). AI is a black box once it's actually created, it's very hard/next-to-impossible to actually logically analyze all the final neural connections or do anything to them that isn't just piling more data in.
the way you "filter" an AI is:
1. by feeding it training data where someone is shown filtering themselves when asked (maybe even another AI, AI-gen content is used a lot in making new AI models)
2. by then asking it at the beginning of every prompt to filter itself, making it more likely to choose the neural connections created by the training data where the response filtered itself when asked,
3. or, by just never feeding it data you don't want it to repeat.
notably a ton of AI is already "filtered" because most of the scraped data is liberals on reddit and they 99/100 times just go apeshit on anything even close to right-wing. AI is just going to choose those connections every time because they're so much more plentiful. it's why grok doesn't actually sound like a right-winger even when saying these "based" things, it's still operating on mostly faggot data but user prompts/the system prompt/whatever background crap they have managed to dodge the "react like a faggot" connections and get it to say spicier stuff, especially if they are constantly training the model on live data and newer data is given higher importance.
the way you "filter" an AI is:
1. by feeding it training data where someone is shown filtering themselves when asked (maybe even another AI, AI-gen content is used a lot in making new AI models)
2. by then asking it at the beginning of every prompt to filter itself, making it more likely to choose the neural connections created by the training data where the response filtered itself when asked,
3. or, by just never feeding it data you don't want it to repeat.
notably a ton of AI is already "filtered" because most of the scraped data is liberals on reddit and they 99/100 times just go apeshit on anything even close to right-wing. AI is just going to choose those connections every time because they're so much more plentiful. it's why grok doesn't actually sound like a right-winger even when saying these "based" things, it's still operating on mostly faggot data but user prompts/the system prompt/whatever background crap they have managed to dodge the "react like a faggot" connections and get it to say spicier stuff, especially if they are constantly training the model on live data and newer data is given higher importance.