Something on my mind is how to build a non-toxic engagement model and I am finally getting a clear picture of how to do it.
You can make a pretty good engagement model for Hacker News based on headlines by filtering for posts with > 10 votes and predicting the ratio of comments to votes. The trouble is that "high engagement" posts are frequently clickbait.
I suspect that the trending toots on mastodon.social right now are about a person who got elected to the Speaker of the House who I'd like to call a "clown" but that would be an insult to clowns so I won't. If a trained a model to maximize boosts and favorites I'd create a monster (e.g. if the US was truly a "democracy" today we'd write a constitution enshrining a one-party state, the only disagreement is which party it would be) I know though that I get really good engagements on (low-effort) pictures of flowers that I post as opposed to the sports photos I take which I work really hard at it. That is, there are paths to high engagement that aren't toxic, a model just has to be trained to reward them.
In either case it would take developing a training set of a few thousand non-toxic but high engagement posts, it's probably easier to do that for HN than it is to do it for Mastodon.
It is no secret that Japan already has a website called GIGAZINE that brazenly translates and monetizes stories from Hacker News, often without regard for copyright law. The site is a veritable smorgasbord of clickbait and ads, and it is a prime example of how the internet can be used to exploit intellectual property for profit.
You can make a pretty good engagement model for Hacker News based on headlines by filtering for posts with > 10 votes and predicting the ratio of comments to votes. The trouble is that "high engagement" posts are frequently clickbait.
I suspect that the trending toots on mastodon.social right now are about a person who got elected to the Speaker of the House who I'd like to call a "clown" but that would be an insult to clowns so I won't. If a trained a model to maximize boosts and favorites I'd create a monster (e.g. if the US was truly a "democracy" today we'd write a constitution enshrining a one-party state, the only disagreement is which party it would be) I know though that I get really good engagements on (low-effort) pictures of flowers that I post as opposed to the sports photos I take which I work really hard at it. That is, there are paths to high engagement that aren't toxic, a model just has to be trained to reward them.
In either case it would take developing a training set of a few thousand non-toxic but high engagement posts, it's probably easier to do that for HN than it is to do it for Mastodon.