site banner

[META] Something Shiny and Two Things Boring

I've got a new feature almost ready to go. I'm pretty stoked about this one because I've been wanting it for quite literally years, but it was never possible on Reddit.

Hey, guess what? We're not on Reddit!

But before I continue, I want to temper expectations. This is a prototype of a first revision of an experimental feature. It is not going to look impressive; it is not going to be impressive. There's a lot of work left to do.

The feature is currently live on our perpetually-running dev site. Log in, click any thread, and go look below the Comment Preview. You'll see a quokka in a suit asking you for help. (His name is Quincy.) Click the cute li'l guy and you'll be asked to rate three comments. Do so, and click Submit. Thank you! Your reward is another picture of Quincy and a sense of satisfaction.

So, uh . . . . what?

Okay, lemme explain.

This is the first part of a feature that I'm calling Volunteering. Once in a while, the site is going to prompt you to help out, and if you volunteer, it'll give you a few minutes of work to do. Right now this is going to be "read some comments and say if they're good or not". Later this might include stuff like "compare two comments and tell me if one of them is better", or "read a comment, then try to come up with a catchy headline for it".

These are intentionally small, and they're entirely optional. You can ignore it altogether if you like.

I'm hoping these can end up being the backbone of a new improved moderation system.

Isn't this just voting, but fancy?

You'd think so! But there are critical differences.

First, you do not choose the things to judge. The system chooses the things it wants you to judge. You are not presented with thousands of comments and asked to vote on the ones you think are important, no, you are given (at the moment) three specific comments and information is requested of you.

This means that I don't need to worry about disproportionate votecount on popular comments. Nor do I need to worry about any kind of vote-brigading, or people deciding to downvote everything that a user has posted. The system gets only the feedback it asks for. This is a pull system; the system pulls information from the userbase in exactly the quantities it wants instead of the userbase shoving possibly-unwanted information at the scoring systems.

Second, you can be only as influential as the system lets you. On the dev site you can volunteer as often as you want for testing purposes, but on the live site, you're going to - for now - be limited to once every 20 hours. I'll probably change this a lot, but nevertheless, if the system decides you've contributed enough, it'll thank you kindly and then cut you off. Do you want to spend all day volunteering in order to influence the community deeply? Too bad! Not allowed.

But this goes deeper than it sounds. Part of having the system prompt you is that not all prompts will be the system attempting to get actionable info from you. Some of the prompts will be the system trying to compare your choices against a reference, and the system will then use this comparison to figure out how much to trust your decisions.

That reference, of course, is the mods.

I've previously referred to this as the Megaphone system or the Amplifier system. One of our devs called it a "force multiplier". I think this gets across the core of what I'm aiming for. The goal here is not majority-rules, it's not fully decentralized moderation. It's finding people who generally agree with the mods and then quietly harnessing them to handle the easy moderation cases.

(We have a lot of easy moderation cases.)

There's another important point here. The mods are only human and we make mistakes. My hope is that we can get enough volunteer help to provide significantly more individual decisions than the mods can, and my hope is that the combined efforts of several people who don't quite agree with the mods in all cases is still going to be more reliable than any single mod. In fact, it wouldn't surprise me if there's people out there who are better at judging posts than our mods are! It's just hard to find you; some of you may not even comment, and you're pretty undiscoverable right now, but you will certainly get a chance to volunteer!

Also, this will hopefully improve turnaround time a lot. I'm tired of filtered comments taking hours to get approved! I'm tired of really bad comments sticking around for half a day! There are many people constantly commenting and voting, and if I can get a few minutes of help from people now and then, we can handle those rapidly instead of having to wait for a mod to be around.

Wow! You get all of this, with absolutely no downsides or concerns!

Well, hold on.

The big concern here is that virtually nobody has ever done this before. The closest model I have is Slashdot's metamoderation system. Besides that, I'm flying blind.

I also have to make sure this isn't exploitable. The worst-case scenario is people being able to use this to let specific bad comments through. I really want to avoid that, and I've got ideas on how to avoid it, but it's going to take work on my part to sort out the details.

And there's probably issues that I'm not even thinking of. Again: flying blind. If you think of issues, bring 'em up; if you see issues, definitely bring 'em up.

Oh man! So, all this stuff is going to be running real soon, right?

Nope.

First I need some data to work off. Full disclosure: all the current system does is collect data, then ignore it.

But it is collecting data, and as soon as I've got some data, I'll be working on the next segment.

This is the first step towards having a platform that's actually better-moderated than the current brand of highly-centralized sites. I don't know if it'll work, but I think it will.

Please go test it out on the dev site, report issues, and when it shows up here (probably in a few days) click the button roughly daily and spend a few minutes on it. Your time will not be wasted.


Blocking

Right now this site's block feature works much the same as Reddit's. But I want to change that, because it sucks.

My current proposal is:

  • If you block someone, you will no longer see their comments, receive PMs from them, or be notified if they reply to your comments.

  • This does not stop them from seeing your comments, nor does it stop them from replying to your comments.

  • If they attempt to reply to your comment, it will include the note "This user has blocked you. You are still welcome to reply, but your replies will be held to a stricter standard of civility."

  • This note is accurate and we will do so.

That's the entire proposed feature. Feedback welcome!


User Flair and Usernames

We're going to start cracking down a bit on hyperpartisan or antagonistic user flair. Basically, if we'd hit you with a warning for putting it in a comment, we'll hit you with a warning for putting it in your flair. If anyone has a really good reason for us to not do this, now's the time to mention it!

Same goes for usernames. On this site, you can actually change your display username, and we're just leaving that in place. So we'll tell you to change your name if we have to. Extra for usernames: don't use a misleading or easily-confused username, okay? If it looks like you're masquerading as an existing well-known user, just stop it.

I'm currently assuming that both of these fall under our existing ruleset and don't need new rules applied. If you disagree strongly, let me know.


The Usual Stuff

Give feedback! Tell me how you're doing? Do you have questions? Do you have comments? This is the place for them!

Are you a coder and want to help out? We have a lot of work to do - come join the dev discord.

48
Jump in the discussion.

No email address required.

What classifier architectures are you looking at? And in the supervised case, how are you planning to establish ground truth (correct label is comment deserves a ban)?

What classifier architectures are you looking at?

I hacked something up that seems to work OK. I need something that can function on very low amounts of data without access to high-end hardware, and this seems to be vaguely doing the job. I'm sure this can be improved, it's a first revision.

And in the supervised case, how are you planning to establish ground truth (correct label is comment deserves a ban)?

Comparing against actual past mod decisions, as well as mod decisions via the volunteer system, basically. It's not ideal because it's noisy, but it's not like we have a non-noisy source of data; that's the problem, fundamentally :)

LLMs tend to do fairly well at this type of small dataset classification problem, has there been any investigation into GPT-3's API for this?

There's a built-in sentiment analyzer, and it's fairly straightforward once you've munged your dataset to get a rudimentary (or not!) classifier, outputting logprobs.

Some documentation is available here, if you're curious.