Comment by bradfox2

The research posted demonstrates the opposite of that within the scope of sequence lengths they studied. The model has future tokens strongly represented well in advance.