Knowledge Localization: Tentatively Positive Results on OCR — LessWrong