Earlier I said that the recurrent neural network in my animations did “roughly the same amount of work” as the transformer-based network. But they don’t do exactly the same amount of work.