I think the theory was that when each weight will be needed is predictable, so the latency can be hidden by fetching earlier (or more likely building the data i

I think the theory was that when each weight will be needed is predictable, so the latency can be hidden by fetching earlier (or more likely building the data in such a way that streaming it linearly brings the right weight at the right time)