According to StanfordAI Lab, slides by Rylan Schaeffer reveal surprising tokenizer behaviors and efficiency trade offs; a preprint is forthcoming.