> It also for some reason is under the assumption that volume of consensus is a substitute for validity. That's simply how those language models work, isn't it

> It also for some reason is under the assumption that volume of consensus is a substitute for validity. That's simply how those language models work, isn't it? The more often something appears in its training materials, the more likely the final model will respond in that direction. The training process has no innate way of autonomously generating generating some concept of validity that'd automatically let the model up- or downrank certain sources. So the only thing that's left is the developers manually up- or downweighting certain sources during training, but due to the gigantic amounts of text that procedure doesn't really scale to really fine-grained control.