OCR models, or optical character recognition AI models, are evolving rapidly. The article explains how these models have shifted from traditional methods, which required extensive manual coding for tasks like noise reduction and text detection, to more advanced approaches that rely on self-learning. Using a crumpled receipt as an example, it shows how modern OCR technology employs vision transformers (ViT) or CNN-based backbones to process raw image pixels. These models create a semantic map, automatically ignoring glare and recognizing text regions. Deepseek-ocr exemplifies this advancement by compressing images into visual tokens, reducing millions of pixels to a small number of meaningful units. This allows the model to efficiently summarize the important aspects of a page. Once the encoder processes the image, a decoder, often based on transformer language models, generates text by considering both the visual context and previously produced tokens. This dual understanding enhances accuracy, allowing the model to adapt to various fonts and styles without retraining. The article highlights the end-to-end training process, where models learn to clean and segment images while minimizing loss. This integrated approach leads to better performance with fewer tokens and less computational power. The author raises an intriguing question about whether text should be fed to models as images, suggesting that once a model understands how to read visually, it could apply that capability universally.