TAI #176: DeepSeek's Optical Compression: A Cheaper OCR or a New Path for LLMs?

DeepSeek has introduced DeepSeek-OCR, a groundbreaking model that leverages visual input to process textual information, representing a significant shift from traditional text-based language models. Utilizing a novel "contexts optical compression" technique, the model encodes text as images, enabling nearly 10-to-1 compression ratios while maintaining high OCR accuracy of around 97%, and still achieving 60% accuracy at 20x compression. This approach exploits redundancies in visual features such as fonts and layouts, allowing for more efficient semantic representation through vision tokens rather than linear text, and supports diverse tasks like document conversion, figure