A look at the challenges some AI developers face in building models to extract trillions of high-quality tokens from PDFs, which are hard to parse, for training