Why bother connecting an LLM to external systems when you can get a lot of work done directly just with hallucinations alone in context memory. I expect this trend to accelerate as context memory and accuracy continue to improve