“Reiner Pope suggests a future opportunity in LLM architecture is to use separate, specialized models for the pre-fill (processing user input) and decode (generating the response) phases of inference.”