Comment by amelius

I've been wondering why we can't skip the entire token embedding step and just feed the model raw Unicode. I suppose the reason for that is efficiency and it's related to your question. Makes one wonder what other steps we can do more efficiently in code than in neurons.