I (admin) just pushed a bunch of changes. You'll see it most obviously in the percent signs next to the save and hide buttons, and as colored emojis indicating which posts are well regarded, and which are about to be deleted. Broadly, this is part of how sofiechan's curation system is going to work.

The percent signs can be interpreted as vote percentages, but they are also a 90% confidence interval over post quality, represented as a uniformly distributed random variable in the unit interval. For example 5%,5% means the initial uniform distribution with no particular evidence either way (5% outlying on either side), 40%,30% means an ok post with some evidence, and 60%,30% would be a rather tight distribution around a 65th percentile post. Your votes are interpreted as evidence of quality, on the assumption that you have taste. The statistics are still crude so percentiles and probabilities aren't well calibrated, but they will be in future.

Given a confidence interval, we integrate with a utility function to value posts despite the uncertainty. Good posts get the blue-green checkmark and stick around longer. Bad and old posts get deleted to make attentional room for new and good posts.

To make room for new threads, we apply a time-decaying subsidy. Every time a post doubles in age, it loses an increment of effective value. This way, the front page should be a mix of classic great posts (which can last a long time by overcoming the time decay) and fresh content (which gets the recency bonus). But everything will eventually be deleted; sofiechan is semi-ephemeral.

The result is that your votes, especially on posts that aren't already well-known to the algorithm, help in curating the best content. This will become more important as sofiechan grows and we have to resist decay in quality at potentially large scale. Please vote responsibly.

As part of the fight against the proliferation of low quality, sofiechan's voting system is going to become very elitist. A positive vote (save, ★) isn't just a vote, it's a staked bet that the post in question is good and will be regarded better in future by the consensus of good taste. Conversely a negative vote (hide, ✕) is a bet that the post and poster are bad. Right now, your clout is pegged to equality, but in the long run bets will have teeth and your clout as a voter and poster will depend on your taste (ie ability to predict and produce quality). This is what it means to interpret votes as evidence.

To ground this otherwise circular specification, I, a hand-picked circle of tastemakers, and members in good standing will get some level of axiomatic clout which is not a bet but a subsidy. This is the aristocratic and monarchic element to balance the democratic. So taste (as far as sofiechan's algorithm is concerned) is the ability to predict the consensus of that elite and the community of good predictors.

As a design goal, the dominant strategy (ie that gets you the most power) will be to vote and govern with the highest wisdom you can muster. Following the crowd (non-independence) or voting as a predictable partisan won't work. Vote to predict and reward quality, and you will be rewarded with power.

The core bet of sofiechan is that this system ("cybernatic direct democracy" or "pagerank for posters") will yield much better curation (and moderation?) than alternate platforms. I call this cybernetic direct democracy because you are directly and "democratically" governing yourselves as a discourse community, without any special class of moderators or janitors, with cybernetic feedback loops to keep you honest and well discriminated. If it works, you may be able to govern more than just your own discourse.

If you see any wonkyness or unexpected behavior please do speak up; we're still working out details.

Some initial comments. Clearly the calibration for the blue-green checks is too liberal. It should be more rare. Second, I can see that not enough people are using the "hide" feature. What would make it more useful?

In the long run (soon?), we'll actually count absence of evidence properly (as bayesian counter-evidence) which will help mitigate the "no hides" issue, but I do want to know how to make sure every vote feature gets maximum signal.

This is really interesting. I like the idea of a much more rigorous filter and having it apply to both voters and posters.

Only failure modes I see (but I’m betting there are reasonable mitigations) would be stuck culture/deterioration in quality is everyone lands on same local maxima for voting. The other is just gaming the vote trend by always liking high vote count posts (perhaps there should be a reward for how early you liked a post?)

>>925

As for people gaming the system by being imitators, as I mentioned in OP I'm aware of this and intending to mitigate it. Basically the system won't train on correlation but rather prediction. If your votes don't predict motion in the final distribution, you are not statistically independent and get discounted.

Stuck culture is perhaps an issue. The curation system is going to be fairly light touch to avoid that, curating a human community primarily rather than policing too much post-by-post. But curation strategy trading off quality and diversity is always an issue. It will have to be explored. I think it can be overcome or at least played better.

Also, the local equilibrium this thing is designed to get stuck in is defined by the best judgement of of the monarch and his hand picked tastemakers, which will tend to correct the thing towards actual quality to the extent we can stay aware of quality.

The only way to game something that's explicitly measuring statistical independence is to actually be a better predictor. Markets do this quite well but they aren't the only way

Multiple threads have older threads as their kicking off point. Deleting those threads means you lose valuable discussions that have already taken place. Ephemerality greatly hampers 'institutional' knowledge -- if you don't want long lasting threads, please lock them after they fall under the threshold instead of deleting them.

>>927
Vote em up if you want them to stick around. If threads are highly rated, the lifetime can be fairly long (it's by value vs log age, so threads could last years if they are well regarded).

For really non-time-bound institutional memory content, I will eventually build a wiki system where wiki pages can be fully timeless with non-finite lifespans (though still displaceable by superior content). Threads that are foundational can be promoted/digested into wiki content.

Precisely for the reason of institutional memory, I am committed to the idea of finite space (it will be bigger than now once we have tags and more people, but still finite). It has to be actually possible to read the canon. If we are doing that plus a ton of discussion to develop that canon and our worldviews, then we need curation where things disappear or fall into obscurity to make room. This means some kind of effective emphemerality. Again I will expand the board when there's more indexing capability and more activity, but content lifetimes are finite in reality.

We will just need to culturally adjust to the idea of emphemerality. Channers often reiterate old threads and topics, and re-study the best preserved in screencaps and such. The fact is internet content rarely lasts more than a few years anyways even if you're nominally going for permanence. Better to handle that explicitly and consciously, IMO. If there's anything truly worthy of preservation beyond wiki content, we should publish actual essays and books digested from it on other platforms.

The ultimate accumulation of value is in the people who read the stuff and change their worldviews as a result. Let's reiterate our favorite old threads occasionally with fresh minds. Why not start now: bump or re-start whatever thread you think should be preserved besides voting it up.

>>927
>>928
That said, happy to hear constructive dissent and alternate ideas on what the parameters should be (for example, should the time penalty be lower, should the board be larger?) or other features that could improve the situation (should we independently estimate timefulness, political sensitivity, etc to calibrate ephemerality level between posts?). This scheme is just what I came up with when I thought about it a bunch. Nothing is set in stone.

> Following the crowd (non-independence) or voting as a predictable partisan won't work.

Won't it? Seems like most instantiations of this algorithm rewards following the crowd.

I suppose you can set the parameters such that you get rewarded disproportionately well for making an early bet on a post that becomes canon, and punished relatively little for betting on posts that don't make it. More like venture capital.

Relatedly, you might want to lean into aristocracy as opposed to monarchy and allow some sharding into distinct tastemaker communities. In many cases I'd rather read the favorite articles from 3 tastemakers than the 3 articles that appeal most to the average tastemaker. Cultivating several distinct status hierarchies provides diversity and enables the incubation of new and controversial ideas. (pagerank doesn't encounter this difficulty as much because it's conditioned on a user search query rather than throwing all websites into a single popularity contest)

>>934
>Seems like most instantiations of this algorithm rewards following the crowd.

This is more like a prediction market. You don't get rich by buying high.

>allow some sharding into distinct tastemaker communities.

All in the plan. That will be the tag system.

--admin

>>927

Until the wiki gets up and running, a simple interim design could be to move threads to a separate 'archive' tab rather than deleting them.

>>953
I'm not sure I agree with this "morality" concept. If you don't want to see something on this site, just hide it. It will eventually get a red hourglass indicating imminent deletion if other people agree. If a post is noteworthy enough to examine for heresy, it should stick around, no?

>>955
Respond and make your case. Eventually there will be 4chan-style linkbacks so that you can immediately see if a post is really controversial (posts that generate tones of replies tend to be controversial). I think disagreement is better expressed as rebuttal rather than simple voting. Simple voting is for expressing whether something is worthy of being seen at all.

I notice a large number of threads converge on approximately 60% up, 1% down. This is two things: not enough people, so when the usual number of us vote and engage in a usual thread, it ends up with about the usual amount of evidence. But then they should rather converge to about the median. This is because I haven't fully turned up the absence of evidence system which would correct this . The latter at least will hopefully be getting fixed this week. No real importance to this yet, but just a heads up of this interesting fact and what I have planned for it. --admin