Should LLMs just treat text content as an image? (www.seangoedecke.com)

DeepSeek’s new OCR paper claims you can recover about 10 text tokens from a single image token with near-100% accuracy — what the authors call “optical compression.” That suggests image representations inside multimodal models are far more information-dense than discrete text tokens, so feeding text as screenshots could, in principle, pack 10× more context into the same model input budget. Practically this could lower inference cost or let models see much larger contexts (even with progressive downscaling of older pages, analogous to human memory fading). Some companies and open projects already exploit screenshot-OCR tricks on existing multimodal LLMs, and DeepSeek outlines strategies for tighter, resolution-aware long-form context management. Why might this work, and why does it matter? Text tokens are discrete IDs that expand into embeddings inside models; image tokens are continuous, high-dimensional embeddings that can encode far more nuance (font, layout, kerning, etc.), so an image token can be more expressive than multiple text tokens. That said, converting training and generation pipelines to “text-as-image” is nontrivial — training objectives, chunking, evaluation and decode speed create practical hurdles, and it’s unclear whether models trained this way would retain the same text-manipulation strengths. If it does scale, however, optical compression could be a cheap route to huge volumes of previously inaccessible written data (e.g., undigitized books), making it a compelling direction for labs to explore.