Is there any reason to think that training weights has peaked rather than is accelerating? It feels like now they are increasingly able to pick some low hanging fruit by using the models in order to improve themselves and test optimizations.