[eurekin]

I really really would like to believe this could work in practice.

Given current data volume used during the training phase (tb/s), I highly doubt it's possible without two, magnitude changing, breakthroughs at once