Every time I try to convert a PDF to epub or something, or OCR one that doesn’t actually have selectable text, it turns out shit. I assume the real reason people would want to get LLMs involved is that there is actually a lot of ambiguity in what a correct conversion would be, and there are a lot of PDFs out there.

I don’t usually like Meta, but here they used that data to produce open weights models available to the public. That sort of thing is what piracy is for so I support it.