Comment by hidden-markov

I am pretty sure this is about optimizing the graphs of trained models. That is, training is not freed from "the prison of supercomputers".