My take on the state of relays.
Relay discovery is tied to people atm - if you follow someone, your client should probably look at their relays and try to read from there (that's part of gossip-model idea). It seems like almost no one chooses their relays manually based on some meaningful selection criteria.
That's because most of existing relays are a commodity - they don't differ in any meaningful consistent way and just contain some random subset of events (notable exception - paid relays). It makes not much sense to bring them to UI because you can't really choose meaningfully anyway.
However, there is this idea that fiatjaf is advocating that custom relays are the way to solve algorithms/discovery/curation - just give me a relay URL that has a curated feed on Bitcoin or Food, etc. Whether the feed is algorithmic or manually curated by admins or by the crowd becomes an implementation detail here.
I tend to think that such "custom relays" are a good idea at this point, and we could talk about two types of relays: "standard" - these are commodity, and will probably stay that, and "custom" - algorithms/curations.
And then standard relays probably need discovery at the protocol layer, maybe using dynamic pricing as the driver - your client gossips and chooses the best perf/price commodity relays. And the custom relays do need to be in UX, and need user-level discovery just like people or content.