[majikaja]

Maybe if we store text data as sequences of 10k x 10k PNGs (one for each letter) and add an image recognition layer it would improve LLM perf