That was the part I noticed, too. Grok was created to be "unbiased", (which really just means they don't put a lot of guardrails on the training data or response output). And still, it has a liberal bias. 🙂
That's not how LLMs learn. LLMs learn by training a neural net to reconstruct a training text, thereby learning the statistical relationships between words. If you feed them lies, they learn lies. If you feed them truth, they learn truth. These things are fed Reddit, so ... 50/50?
In theory, yes. But there's no absolute reason that has to be true. Even worse, if somebody wants to bias the outcome, they can be biased in their sources.
Yet grok still likes to piss into Musks face. Yes I know. It does not „like“. Does not think. But still :-) The problem with mass data is: you can’t forge all data to **consistently** follow all paths of all ones lies. Even if one only keeps focus on MAGA lies.
Speaking of Aparthaid in South Africa, did you know that Rhodesian Brushstroke Camoflauge is so effective that you can't even find the country on any maps?
Comments
https://www.conservapedia.com/Main_Page