Do we even need to get LLMs to create code in the end then? Can we imagine a future where the thing simply does it all, you deploy the LLM like a docker image and have it process all requests directly. Sounds far fetched today, coming to your cluster in five years.