Totally UNDERRATED.

The whole research industry is chasing the x% performance gain while the train/compute/energy cost increase by 10x%

Aiming for efficiency and interpretability would be a good way.

But meh.