Why is spam score always weighted as a single outcome and never as a population-based result? I mean, you’re processing more data, I get it, but it feels like that would be a model much more resilient to tampering, bad actors, and just different social norms.

For example, if one set of users started rating a specific subset of posts as spam, then those users could be bucketed together into a “doesn’t want to see message type A” group while others, who minded other messages, would be bucketed into a “No B-messages” group.

This would need to be applied selectively, as it could easily result in an echo chamber for normal discourse, but I would’ve given my left arm to have that sort of filtering available in the game during my WoW days. Those city spammers were unbearable!

I, of course, would have fallen into the “I don’t care how great a deal your Thunderfury, Blessed Blade of the Windseeker, is, I’m just here to socialize” bucket.

> Open sourcing it would be hard, because you'd have to trust everyone else to classify things correctly, or you'd have to review every input yourself.

There's not a way you could choose which people/groups you trust (and don't) to classify spam correctly? Don't open source adblockers work like this?

Don't they use some reasonably reviewable methods, like regular expressions? And, more importantly, reviewable volumes? Also, an ad blocker can always have a debug mode, where it can show you a rule that removed the element.

With ad blocking you have none of that. Gigabytes of text go into training, kilobytes of inscrutable numbers go out. And all the debug info you get is how certain the computer sounded saying no.