i would like to assume that low level optimization is crucial part of AI training and openai, meta and others aren't wasting billions on this