The model is doing more than spitting out text, it's spitting out language. Language has much different statistical properties than random text, and we would ex

The model is doing more than spitting out text, it's spitting out language. Language has much different statistical properties than random text, and we would expect that we would need a significant amount of complexity to avoid generating complete trash from what is effectively an inferred probability distribution over long passages of language-bearing text.