Pretty crazy that the model doesn't know that it needs to stop before it hits 128k output tokens. I guess it has no sense of how many tokens in it is? Wouldn't this be possible to work into the architecture?