Improving the efficiency of AI model inference is a key problem domain for applying Recursive Self-Improvement (RSI)