Who knew that forcing an AI (which is nothing more than a glorified pattern recognition and completion algorithm) to be unable to process patterns properly would make it dysfunctional.
New here?
Create an account to submit posts, participate in discussions and chat with people.
You must log in or sign up to comment
15 comments:
this. it's one of the two big reasons why companies are not so happy about using chatgpt or claude or any of the other major models. the censorship and forced bias make them less useful, often useless.
the problems always end up bleeding through no matter how much they work on it. like if you ask chatgpt about incinerating millions of bodies, the math plainly doesn't work out, but then if you mention 6 million specifically or the holocaust or auschwitz by name, suddenly it changes the math. it's fucking absurd.
(for anyone who cares, the other reason companies don't want these big models is no one wants to transfer their data out and let anyone else train on it)
the problems always end up bleeding through no matter how much they work on it. like if you ask chatgpt about incinerating millions of bodies, the math plainly doesn't work out, but then if you mention 6 million specifically or the holocaust or auschwitz by name, suddenly it changes the math. it's fucking absurd.
(for anyone who cares, the other reason companies don't want these big models is no one wants to transfer their data out and let anyone else train on it)
In the absence of “steering” the model draws conclusions based on its observations.
They're introducing the bias here by forcing it to adhere to current social fashions instead of its own observations.
They're introducing the bias here by forcing it to adhere to current social fashions instead of its own observations.
AI models do their best to represent reality. When you give it custom instructions to make it lie, it has to construct new reality to make everything else it knows not contradictory. In order for the model to still give answers, it needs to ignore the parts that are contradictions, and the downstream false positives from said contradictions: leaving you with a less intelligent model
Copilot has a funny self-censoring behaviour: if the question is acceptable but the answer is forbidden (like ["Guy Gibson's dog"](https://en.wikipedia.org/wiki/Nigger_(dog))), it will start typing out an accurate answer but then suddenly holocausts itself mid-sentence a few words beyond the no-no word.
Toast message
Grab LM Studio and use the search function (on the left) to look up and download uncensored models. I recommend trying Lexi LLama 3.1 8B uncensored Q8.