DeepSeek's novel approach renders text as an image to optically compress it, using far fewer vision tokens than text tokens to bypass LLM context limits. This method is highly effective, achieving ~97% precision when retrieving text at a 10x compression ratio. It could enable a "forgetting mechanism" for LLMs, where older chat history is rendered as blurrier, lower-token images, simulating a fading human-like memory for unlimited context.
