No GPT is trained to spit out anything that falls into its training data distribution. There's no average. Pre-training incentives being able to predict the sm

No GPT is trained to spit out anything that falls into its training data distribution. There's no average. Pre-training incentives being able to predict the smartest string of text in the corpus as readily as the dumbest. It doesn't converge on "average" and it doesn't really make sense that it would either. Base models don't talk like GPT. This is strictly an artifact of post training fine-tuning/RLHF.