Amazing video. My only suggestion for improvement for the modeling is in the part where you tried to find the "ideal" letters for the model. Instead of sampli

Amazing video. My only suggestion for improvement for the modeling is in the part where you tried to find the "ideal" letters for the model. Instead of sampling generated random bitmaps and ranking those, you could instead initialize the inputs to random continuous values, then optimize the output score for the desired letter with respect to the random input (with your fixed model). Indeed, this is how people get those trippy dog images you've probably seen. Anyway, this video was just so good. I'm amazed.