Hacker News Reader

My wild guess is that adjusting the shape before each step is not worth the speed hit. Uniform structures make GPUs go brrrrr

It's also easier to train and in particular easier to parallelize.