Like a milking stool, my case rests on three legs:
1. I will demonstrate that RNN reads sequentially, with performance that borders on the speed of a snail.
2. I will argue that instead of one person reading 100 pages sequentially, I could deploy 100 people reading all pages simultaneously.
3. But first, I will invoke the attention mechanism. I'm able to focus on any page when predicting the next word, accessing the full context that was previously unreachable.

So, to sum up: sequential reading, parallelism, attention. My milk stool is complete.