When a printer on Douban ran low on ink, users noticed that the top half of each Chinese character still rendered legibly. The author captured three cropped versions of 人工智能, showing that 80 % or 50 % of the pixel area can be removed without losing readability. This observation sparked a question about the visual nature of language.

To test whether shape carries linguistic signal, the author replaced token IDs with 8×8 grayscale images of each character and trained a next‑character predictor. Experiments across resolutions from 4×4 to 80×80 showed that increasing pixel count beyond 8×8 added little gain, while cutting characters to 50 % retained accuracy within 2 %. This highlighted the power of a visual prior.

Although the visual model arrives two times ahead of the token baseline after only 0.4 % of training steps—a hot‑start effect—it ultimately converges to the same ceiling. In low‑resource scenarios, the visual head start proves useful: with only 10K samples, the image‑based model surpasses a fully‑trained token model on Chinese benchmarks. Thus, visual cues give an early edge but do not raise the ultimate limit.