I think there is a blend there. Think of this: if you want to train a fancy ML Algo, you need to download everything. But once you have the model trained (and integrated into an Algo) then it should be sitting either in the relay and deciding whether to relay or to discard or in the client to locally store or drop. In the relay it might be more effective because it saves bandwidth.
I just saw this "bring your own algorithm" engine for Bluesky: https://github.com/louislva/skyline
I always thought this thing would be very hard to do properly on Nostr -- or anywhere. But for Nostr specifically I thought the right solution would involve smart relays and clients that knew how to handle some decisions to them.
But this Skyline thing is too stupid. It just downloads a giant number of posts from your timeline, then filters them on the client based on criteria you specify. It's an idea, though.
Discussion
Yes.