Comment by tosh

i wonder if the custom tokenizer is better in practice, the examples look interesting though