Interesting point... In this case, I have a hunch that the requirements in terms of training data and hardware to run the training on would remain prohibitive, but I do wonder how much could be actually achieved using a distributed scheme of sorts.