> History teaches us that end to end learned system beats human designed mechanisms

I think this may need some qualifiers

Even byte representations are human designed encodings. I would think a human designed decoder of such encodings must be more efficient than learning. Sure bytes encoding a stream of unicode code points maps fairly easy to useful information. But bytes representing a zip compressed collection of PDF files?

I did wonder though, training on text encoding vs pixel encoding, perhaps brute forcing OCR, like humans, will be more flexible in the end then being limited to text encodings.