deepseek proposes compressing text into image tokens for processing, which can save computing power

Is this reasonable? DeepSeek’s new skill: compress text into images_Bilibili_bilibili