What OCR does in this material In the channel's captured material, OCR — image-to-text transcription — is the step that turns a photographed book page into text an AI can work with. It appears in the build of Hermes, a reading companion app: you send a photo of a confusing book page by text message, and the app sends back a plain-language restatement plus the context you might be missing [^4ef67c54-9505-5bc4-a565-9579a83f2827@41]. Why a multimodal model gets the job The pipeline assigns each model its specialty, and OCR goes to GPT-4 because it is multimodal and strong at transcription [^4ef67c54-9505-5bc4-a565-9579a83f2827@211]. The transcription it produces is then handed off to o1 preview, whose chain-of-thought reflection makes it better at interpreting the extracted text [^4ef67c54-9505-5bc4-a565-9579a83f2827@234] — showing how transcription and interpretation are kept as separate stages. The routing matters, too: at one point an image was mistakenly sent to Whisper, a speech model rather than an image model, which had to be caught and fixed [^4ef67c54-9505-5bc4-a565-9579a83f2827@657]. - OCR (image-to-text transcription) converts a photo of a page into text.- In Hermes, GPT-4 handles OCR because it is multimodal and strong at it.- The transcription feeds a second model (o1 preview) that interprets the text.- Keeping image-to-text and interpretation as separate model jobs is the demonstrated pattern.