80% of Energy Costs Are on the Inference Side, Right?

The reward system for professors and researchers skeet over its training, right? You probably get rewarded for beating benchmarks rather than rather than squeezing efficiency out of inference. I think that's what we're trying to promote with the green AI approach: don't only reward people who get better accuracy results; also, reward those that get more efficiency. And I think these methods that take the models and make them smaller are still valuable to anybody doing research in NLP.