> However, it is much more efficient to use a tokenization tuned to the statistics of the input data set. A program "tuned to the statistics of the input data

> However, it is much more efficient to use a tokenization tuned to the statistics of the input data set. A program "tuned to the statistics of the input data set" is what an LLM is. So choosing to use a fixed tokenizer rather than letting the LLM learn one is a performance optimization, but not one we'll have to accept if models were designed better for performance themselves.