A groundbreaking approach proposes converting LLM text inputs into images, enabling “optical compression.” This could dramatically expand context windows by up to 10x, challenging fundamental assumptions about AI processing.