tinygrad is fun because it's tiny
so we make RNN that is tiny
it's got an api like pytorch except it's easy to run (and modify) from source when things aren't working right. try adding print statements into the backwards functions in pytorch! have you tried to read pytorch source code?