How about this scheme:

- when you upvote an item, everyone else who upvoted the item before you earns some amount of your trust

- the more of your trust someone has earned - the more weight their other upvoted items get for you

- each time they upvote, they put some amount of your trust on that item; so if you stop liking their recommendations the amount of your trust they have will go down over time

- when you downvote an item, you take away your trust from people who upvoted it; they've shown that they are not good curators of content for you, so their upvotes will have less weight for you

In this system you end up paying attention to people who have proven to you to be good curators of content. It optimizes for high signal to noise ratio, where what is signal and what is noise if up to you to decide with your upvotes. We don't have to all agree on what is globally "upvoteworthy".

There is no global reputation system (which can be gamed). Instead, there is a peer-to-peer trust system.

If you are interested in a system like that, then I would like to invite you to my hobby project that works exactly this way. Register with a temporary account (no email required) at https://linklonk.com/register and use code 'hn'.

It is early days and we don't have many users yet. To supplement real users LinkLonk supports RSS feeds as sources of information. Each feed behaves much like a user - the more you upvote content from it, the higher ranked its other entries will be for you. I hope you will find it useful and I'm looking to hear your feedback.

I imagine that this is how social media platforms create "information bubbles". In terms of politics it won't work: people that agree with your political views would get more "trust". Is agreeing with your views an indication of good content? I'm not sure.

What algo are you using to manage the trust system?

I feel like this would have the result of Gell-Mann Amnesia, which is essentially what "influencers" rely on prior to sponsorship. Thing 1 was recommended, and others found the same benefits to Thing 1, so when Thing 2 is recommended it is assumed that the influencer's opinion is valid. Breaks down as soon as there is a disconnect, if people are willing to accept that their chosen influencer can be wrong, but in practice most people just go along with it.

The guy who invented the Elo system was actually named Elo, it's not an acronym.

Different people have different interests, so using just one group as etalon, would leave most people unsatisfied. It would be better to let users choose their own groups of "good" actors, either explicitly or based on upvote similarity.

Sounds kind of like Google's PageRank algorithm[0] except using votes as the signal instead of links.

[0]: https://en.wikipedia.org/wiki/PageRank

I don’t think this is the same. We’re not increasing influence based on the reputation/amount of upvotes of the user, but rather based on the similarity with known-good actors. The user might never post themselves, and therefore have no karma of their own.

Perhaps I didn't describe it well enough, but I believe you're misunderstanding my intent. The user's "PageRank" is based on how they vote in relation to other users.

You seed the influence system with known good actors. Users who vote similarly to these good actors get an increased influence weight and so on. You can apply the PageRank algorithm to any graph, and this case the graph is the relation between up votes of other users.

This can work better than Elo rating because you can have disjoint seed actors serving sparsely connected (or completely disjoint) regions of your influence graph.

I’m trying to understand this, do you mean that edges in the graph are not upvotes but rather A —> B means B voted similarly to A?