No feasible way to train/low efficiency

To be able to train a project of this size, you would need a full data center. Why not use more acceleration libraries, or even optuna, to make it faster/better overall? This would greatly improve the feasibility and power of this AI model. So in conclusion, this is one of the greatest models I have ever seen, but it can very much be improved.