Comment by geon

I think the hard part is improving on the basic concept.

The current top of the line models are extremely overfitted and produce so much nonsense they are useless for anything but the most simple tasks.

This architecture was an interesting experiment, but is not the future.