good point - i analyzed text tokenization. will run some experiments to see how visual tokenization has changed