oh man someone should make this into a loop where the model can alter it weights after an inference. some kind of descent of the local gradient.
Isn’t that what the Google dreaming thing is?