Comment by strbean

We're up to a gazillion parameters already, maybe the next step is to just ditch the tokenization step and let the LLMs encode the tokenization process internally?