everyone, including Karpathy, is explaining DeepSeek-OCR as a victory of pixe...

everyone, including Karpathy, is explaining DeepSeek-OCR as a victory of pixels over unicode

but imo its encoder-decoder over decoder-only transformer

we could similarly get the compression wins on just text

turns out there’s prior art for it — the 500x Compressor paper did it with text->text

fwiw, this also is encoder-decoder arch

aclanthology.org/2025.acl-lon...