I wonder if GPT does any weighting based on source quality? Or weighing more recent sources over older sources? Maybe not explicitly built into the models or training, but just an emergent behavior by way of those sources being more plentiful. Or maybe they need to build it into the training.