Efficiency is Coming: 3000x Faster, Cheaper, Better AI Inference from Hardware Improvements, Quantization, and Synthetic Data Distillation