395 karma · joined January 9, 2021
Email is "lonk" at the above domain.
I am building https://linklonk.com to explore a middle ground - you upvote/downvote content as a way to tell a transparent algorithm who you want to get more/less content from.
When you upvote a post/link, LinkLonk connects you stronger to other users that upvoted it and prioritizes other content they upvoted for you.
When you downvote something - your connections to those who upvoted that become weaker, and their other upvotes don't have much weight for you.
This makes it much more scalable - the algorithm keeps track of your connections to everyone else. You could be connected to hundreds of users and RSS feeds on LinkLonk, without having to explicitly subscribe to them. But you still have control - you express it every time when you decide whether to upvote or downvote the content you just read (or do nothing).
You end up connected the strongest to sources that have the highest signal-to-noise ratio for you. You see content from those who have proven to be good curators of content, which I think is a good feedback loop.
You are right, to keep track of user to user connections is more expensive than only to keep track of the number of likes each item has received (Reddit).
But ultimately someone has to track how trustworthy each of the users is to every other user.
Either every user has to track it for themselves in their head. Or we can try to offload some of that work to the computer. LinkLonk is an experiment to do just that. So yes, this computation is expensive, but it is also the main value proposition.
- When you upvote an item, you connect stronger to users that also posted that item and to RSS feeds that posted this. For example, if you upvote "Post 43: Intentionally Making Close Friends — Neel Nanda"(https://linklonk.com/item/827619236936941568) then you will get connected to 6 users that upvoted that article and the RSS feed https://www.neelnanda.io/blog?format=rss
- The recommendation algorithm shows you other items upvoted by users and feeds you are connected to.
- At the top of you recommendations you see content from users and feeds that you are most strongly connected to (ie, those who have posted more useful content for you).
With this mechanism you discover users and feeds that post great content just by rating content, without having to know the users personally. Yet, when you see your recommendations you know that they are coming not from random people, but from people who found useful for you content before. I think, that is a more meaningful version of social proof than the aggregate counts of likes in the traditional social media.
By transparent I mean that you know what the algorithm will do when you upvote an item. For example, it tells you when press upvote: "You will see more content from 4 users and 3 feeds that recommended this item". And then, when you see a recommendation, it tells you which items connect you to the users that brought you this recommendation (it is shown above every recommendation if you expand "from X users and Y feeds").
My hope is that the users will have more control over the information they get and, as a result, will be more thoughtful about what they upvote/downvote.
With LinkLonk you gradually update your connections as you upvoted/downvote content. In a regular group your connection to all other members is a boolean 0 (you are not a member of the group) or 1 (you are a member of the group). On LinkLonk your connection is a float - 0.01 or 0.5 or 3.7 - depending on how useful you found their previous contributions.
As for taking the time to onboard, I agree, this is a problem for LinkLonk - the initial set of recommendations is purely popularity based and there is no topic focus to it. But try upvoting a single item and you will get connected to the other users that upvoted that item. Refresh the recommendations and you will see more content from those users, which should increase the signal-to-noise ratio. Or try submitting a new link and you will get connected to the RSS feed that posted that link.
The most common split is how knowledgeable you are on the topic. Experts are your "possums" and beginners are your "otters". If you don't stop the inflow of the beginners, then you get mostly noise for the experts (Eternal September). If you do stop the inflow - you become beginner unfriendly.
The article's suggested solution to regularly split the group tries to lessen the tension on the shared resource by creating more resources. But it's hard - you will lose members as you try to constantly create new groups.
I'd like to suggest a different solution - let users dynamically connect to people who post "signal" and disconnect from people who post "noise". Imagine a group system where when you upvote a post, you connect stronger to other people who upvoted this post (this includes the author of the post). And if you downvote some content you weaken your connection to people who upvoted it. And then how strongly you are connected to others determines how the content is ranked for you. The stronger you are connected to "possums", the more "possumy" content you get. And if you prefer "ottery" content - upvote it and you will see more content from "otters".
This removes the tension because everyone has control over the content they see and they no longer need to fight over the rules of what belongs and what doesn't. This replaces a collective decision with an individual decision.
And guess what, I am building just such system as a hobby project at https://linklonk.com so you can try it out.
I hope you will forgive my quoting: """ You seem to be concerned that the mechanism of LinkLonk will lead to a filter bubble where people would group by opinions they agree with and they would simply be preaching to their own (often misinformed) choir. That is a concern, but I'm not sure it is an inevitability.
Now that you know how LinkLonk works, what type of content would you personally upvote:
1. Content that does a good job at persuading the reader to agree with the opinion you hold, even if that content gives little useful information to you, or even if the reasoning is a bit disingenuous.
2. Content that informs you even if it does not promote "we are right, they are wrong" narrative.
My hypothesis is that you would be more inclined to upvote content (2) on LinkLonk more, than you would in other social systems (Reddit/Facebook/Twitter).
In existing social systems, upvoting content (1) is logical because you are trying to promote your opinion to others. The upvotes are used as a tool to win opinion popularity contests. And you are less likely to post/upvote content (2) because it may be seen as non-conformance to the group. That results in the tribalism you mention.
On LinkLonk:
- The main effect of upvoting content is to get more content from people who also upvoted it. If you upvote content (1), then you would personally increase the amount of noise your "future-self" will see (since it has little new information). If you upvote (2) then your "future-self" will be better off.
- Your ability to influence what other people see is limited by how useful they found your past upvotes. You can easily lose that ability if you misuse it - they can just downvote any useless content you try to push. There is no such feedback mechanism in popularity based systems. Every upvote has the same weight. And so it can be misused.
I'm not saying that LinkLonk will avoid the echo-chamber/filter-bubble problem. I'm saying that LinkLonk has a different system of incentives that it is not fair to apply the behaviour seen in other systems to it. And what kind of system level behaviour LinkLonk will have is impossible to predict just using a thought experiment. That's because this behaviour is emergent. The only way is to run an actual experiment - which is what LinkLonk is. """
The interest graph is built in response to your content upvotes/downvotes. The connections are float values instead of booleans. Here is how these float values are updated:
1. You start off weakly connected to all users and you see generally popular content.
2. When you upvote something - you get stronger connected to other users that upvoted that content before you (you don't connect to those who upvoted after you - this is a proof of work that Eugene wrote in https://www.eugenewei.com/blog/2019/2/19/status-as-a-service). The stronger your are connected to someone => the more weight their other upvoted items have for you => the higher their other upvoted items show up in your feed.
3. When you downvote - your connection to those who upvoted that content goes weaker. And your connection to those who also downvoted it goes higher.
4. Every time someone you are connected to upvotes something, your connection becomes slightly weaker to them. They sort of place a bet that this item is worth your attention and that bet is lost if you ignore that item. So if you ignore those who no longer post useful content or simply post too frequently for you - they will gradually disappear from your list. This prevents the old money problem from the same "Status as a Service" post above.
5. You can put the items you upvote into user-defined collections, so your tech content could go into your "tech" collection and users that are connected to your tech collection will not be bothered by the stuff you put into your "film" collection. This is similar to different boards on Pinterest.
This system is somewhere between popularity based systems (Reddit), subscription based systems (RSS, Twitter) and algorithmic systems (TikTok, Pinterest).
Your connections to other people capture how useful their previous recommendations have been to you (ie, their signal-to-noise ratio) and so can be thought of: how much you can trust them that their future recommendations will be worth your time.
RSS would be the next step.
What you are suggesting is similar to another idea I had (https://linklonk.com/item/9146000221282140160): "An option to generate a publicly visible url for any of your collections. This could help you share your collection of liked items with other people. And it could help LinkLonk get new users."
Adding an RSS feed for these publicly accessible collections would be a natural extension.
I didn't quite get what are the centralized incentives that lead to clickbait. Is it a comment about LinkLonk or centralized systems in general?
Centralization and decentralization have trade-offs. Right now the downvotes on LinkLonk are private. The upvotes are semi-public, but not available in raw form. If you wanted to implement the LinkLonk's algorithm in a decentralized manner then you would need all hosts to make the upvote and downvote data available to all other participants. Maybe homomorphic encryption can solve this problem.
My goal is to build a proof of concept in my spare time and the centralized approach seems to be the right option in my situation.
To make it worthwhile for the users before the project reaches that point, I am trying to make LinkLonk a better tool that helps you keep track of the content you liked (kind of a bookmarking service) and helps you follow RSS feeds in (kind of a feed reader).
If you have other ideas of how to make it more useful before we have enough users - let me know.
Or would you change something else?
I don't know if I can get rid of the bottom part with the buttons since this is where the collection selector is - where you pick what collection the item you like should go to.
LinkLonk's algorithm works that way. It builds trust in sources of information (including users who rate content). And it does not and will not try to understand the individual pieces of content.
Unlike the social media feeds that are powered by black box neural networks, LinkLonk's algorithm is transparent. You know how your interactions with it will be interpreted. I hope that this transparency will help build trust in the system and in the sources of information you are connecting to.
Yes, bad actors will try to game any system to gain the attention that they don't deserve. I'm not claiming that LinkLonk is game-proof, but I think it has better feedback loops and incentives than other systems such as popularity based ranking (please don't take it as a challenge).
When you want to introduce someone to LinkLonk you could share with them your personal url.
As for decaying sellouts, wouldn't downvoting the content they upvoted do what you want?
By the way, every time someone you are connected to upvotes something, your connection to them becomes slightly weaker. So if you simply ignore recommendations from sources with a lower signal-to-noise ratio (or high volume) - their recommendations will eventually fade away from the "For you" page.
To solve this problem I am building https://linklonk.com that cultivates trust as you rate content. When you upvote a link you connect stronger to the feed that posted it (which could be a blog's feed) and to other users that upvoted this link before you. When you downvote - your connections to those who upvoted become weaker.
The strength of your connections to other feeds and users represents how useful their content recommendations have been to you in the past and they could be used as a measure of how likely their future recommendations will be worth your time (ie, trust that they won't waste your time).
The content is ranked according to the connection weights - so you get information from the sources that have shown to be content curators for you.
I did a Show HN recently for this project that has more details: https://news.ycombinator.com/item?id=28405643
[1] - like my comment here tries to draw your attention to my hobby project.
When you upvote an item - you connect stronger to people who upvoted that same item.
When you downvote an item - you connection to those who upvoted it becomes weaker.
The content is ranked based on how strongly you are connected to people you upvoted each item.
That way when a lot of clueless/uninformed people come in and start upvoting useless to you content - they don't create noise in your personalized content list.
The content you see comes from people who have proven to find useful content for you.
This is how my hobby project https://linklonk.com works.
I did a Show HN a week ago and that post has more details: https://news.ycombinator.com/item?id=28405659
Though I don't know any details about Last.fm's algorithm. Most likely it is based on neural nets, which are opaque to the user. I think this is totally fine for music recommendations. In that domain the are few adversaries and the cost of a bad recommendation is just a skipped song.
For information content I feel the stakes are higher and there are more bad actors who would love to grab our attention. That's why I think it is important to have a transparent algorithm.
And yes, LinkLonk does find your "neighbors" - people that like the same content as you did. What is more, LinkLonk connects you to those people, that liked that content before you did. And they are likely to be more informed than people who upvote the same content after you upvoted it.
Since we still have few users I would suggest to start using LinkLonk as a tool for bookmarking interesting content and following RSS feeds. Hopefully this would make it useful enough before the number of users reaches a critical mass for the network effects to kick in.
Here is an example of a Substack newsletter issue: https://linklonk.com/item/2348076191098765312 which has this feed url: https://semianalysis.substack.com/feed and if you scroll down on the item detail page you will see other issues of that newsletter in the item-based recommendations
If there is feed url in the metadata, but you know the feed url, then you can submit (linklonk.com/submit) the feed url directly and LinkLonk will connect you to it.
If you read every issue of the Money Stuff newsletter and upvote 5 out of 10 that are most informative and then if I upvote 2 of those 5 that you upvoted then I would get connected to you and to the feed of the newsletter (if such feed existed, I don't seem to find an RSS feed for it). But because your signal to noise ratio is "2 out of 5" and the feeds ratio is "2 out of 10", then my connection to you will be stronger. And so I will see your recommendations first. That is, I will get a prioritized view of Money Stuff issues and I will benefit from you work as a curator.
To get a sense of what recommendations you would get if you upvote a given item, you can go to the items detail page (three vertical dot button in bottom right of every item -> Item page). There you will see recommendations from the users and feeds that upvoted/posted that content. For example, if you go to https://linklonk.com/item/7470250776832704512 (a page for https://arstechnica.com/cars/2021/08/please-stop-adding-more...) then you will see what else was recommended by 6 users and 2 feeds that upvoted/posted that article from Ars Technica.
When you upvote an item, you connect to other people who upvoted that item. When you downvote - you disconnect from those who upvoted it. So those who upvote useful content for you, get more weight in your recommendations.
Coincidentally, I'm doing a show HN post for it today: https://news.ycombinator.com/item?id=28405643, so will be happy to answer any questions there (instead of hijacking this post).
OP, does EvalRSS list content in reverse-chronological order? Any plans to prioritize content? What I do with LinkLonk is: the more items for a feed you upvote - the more other items from that feed are prioritized. Also, RSS readers have a notion of read/unread status. What are your thoughts on that? Is reverse-chronological ordering supposed to help the users distinguish unread content from read content?
Consider adding the number of comments each item in the list has. I added a test comment on https://www.evalrss.com/comments/9dcbaacb-96b7-4eec-8959-ba6... and there is no indication of that comment in the list.
2. We follow groups of people (HN/Reddit/forums) and it helps us connect to people that we wouldn’t otherwise be able to discover. But you get a mix of high signal-to-noise members and the more-noise-than-signal members. The bigger a group becomes the worse this mix usually gets.
3. We use algorithmic systems to bring us useful information. But these systems, powered by deep neural net models, do not truly understand what is useful to us. Instead they are optimizing for "time-spent" (ie, ads shown) by showing you more shallow content that will keep you clicking and yet never satisfied. They are opaque by design and are not something that we can control with our actions. Instead of offering explicit control to the users they are focusing on implicit signals such as what we clicked on before - a poor signal.
Can we get help discovering sources to follow?
Can we take only the good parts of a group and not the bad parts?
Can we have agency in an algorithmic system?
With LinkLonk I am building a system that combines the best parts of the three systems and addresses their weak points.
To start, LinkLonk is an algorithmic system, but the algorithm is transparent to the user and the algorithm’s output directly depends on what content you upvote and downvote.
When you upvote a piece of content, LinkLonk strengthens your connection to other users that upvoted that same item.
When you downvote something, LinkLonk weakens your connection to users who upvoted it.
How strongly you are connected to a user determines how high their next upvoted items will rank in your list of recommendations. It means that at the top of your recommendations you will see content from users who have been good at finding useful content for you in the past.
This creates a feedback loop with you in control.
Initially, when you have not rated any content yet, LinkLonk connects you to all users with a very weak connection. As a result, your initial set of recommendations is based on popularity. It behaves like a group-based system at the start. But as you rate content you get connected to specific users - the ones that have the highest signal-to-noise ratio for you. This makes it more similar to the “following individual sources'' type of system yet it solves the problem of discovering the sources to follow.
When you downvote something, not only does LinkLonk weakens your connection to those who upvoted that content, LinkLonk also strengthens your connection to the users who also downvoted that content. It means that their other downvotes will have more weight for you. The idea is that if they were able to recognize bad content in the past then they could be trusted to flag bad content in the future.
LinkLonk creates a new system of incentives. In order to get your attention, other users need to prove to be good curators of content. And to hide content, they need to prove to be good moderators. It’s these new incentives that excite me about this project.
Now we need a few users to test out this system. Give it a try, submit a link you found useful.
To create an account you don’t have to submit your email address. Use "Continue as guest" to create a temporary guest account. It will use a cookie on your local browser. You can later convert it into a permanent account that you can access across devices. If you don’t use your guest account for >30 days it will be completely deleted from the server.
P.S. This is my hobby project that I am building in my spare time. My stack is: PostgreSQL + Golang server (sqlc, gorilla/mux) + Angular client + Firebase for auth. It runs on a small VPS instance on OVH that costs ~$12/month. I’m intending to run it for years. That’s the nice part about being a hobby project - there is no time/financial pressure to "succeed".
To answer your questions:
1. The list of all feeds used by LinkLonk is not publicly accessible through the website. They are feeds that users explicitly submitted through https://linklonk.com/submit or feeds that LinkLonk parsed from the meta tags of the links that users submitted.
2. The number of active users has been about 4 for the last few months. I'm hoping to get it to 10 this year.
I do want to add functionality to download your ratings. I'm thinking of exporting the ratings data in either the bookmarks format (ie, the format that browsers use to export bookmarks: https://support.mozilla.org/en-US/questions/1319392), csv or json. Please let me know what format would be most useful.