Have you thought about what happens if we get a new improvement in model architecture like transformers that grows the compute needs even further