Comment by doctoboggan

At this point, I think anthropomorphizing the models gives us better insight into expected behaviors rather than continuing to insist they are just simple probabilistic token generators.