To give a sense of the number of decisions we make:
Our main discussion is called Every Thing. All of this is *only* about the list of topics.
The design goal of Every Thing is to make a naively abstract conception of politics into something precise. We believe, in theory, you could solve all political problems, if only you could "hold a vote on everything, in order of most to least important, and find consensus on all of them."
The name is "Every Thing" because you don't vote on everything - You vote on every thing. Proper nouns, like "Donald Trump", are part of everything, but they are not things. They are people, labels, slogans, or specific details, not the actual thing under discussion. They are a factor that introduces substantial biases into the responses, which enables push polling and all manner of other abuses.
To avoid this, we have imported >157,000 "things" from Wikidata, effectively all common nouns with a description. This list was then manually curated as much as possible to leave the list of comprehensible items. If Wikipedia knows about it, it's on this list.
For example, "president" is a thing, as is "abuse of power", or "right-wing personality". By limiting ourselves to common nouns, we create a perfectly neutral things, with no framing or attached opinions. They are as close to perfectly fair as possible.
We want to hold a discussion on all of them. We don't want to only cover discussions of what's on people's minds, we also want to be able to discuss the "unknown unknowns", issues too obscure or forgotten about to ever come up naturally in a survey like this.
This is an enormous list. Similar projects limit themselves to finding collections of things most people are likely to have opinions on. For example, Tom Scott's "best thing" survey included 8,000 items, chosen on the basis of having Wikipedia articles in many languages.
Our list, by contrast, has been curated to give a flat and accurate representation as possible of literally every thing. Rather than items people are likely to have an opinion on, we have attempted to curate for the list of items people are able to comprehend. The primary filtering is the manual removal of categories that are essentially databases so large as to overwhelm the set ("mineral", "human disease", "mouse genome", etc), and things with no description. There are still a few more highly technical similar items than ideal, but we could not find a principled way to remove them that did not remove meaningful items.
Some of my personal favorite items include: screaming hairy armadillo, uwu (the emoticon), imagery of nude celebrities, abecederium ("inscription consisting of the letters of the alphabet, listed in order"), fried brain sandwich, four-in-a-row, "walking ghost phase of radiation poisoning", smound (a synaesthesic sense combining smell and hearing), unbirthday, modern schools of ninjutsu, bullying in nursing, one-year period overlapping two calendar years, roadkill cuisine, bullfighting poster artist, cranberry morpheme ("bound morpheme which lacks independent meaning; eg the cran- in cranberry"), and wrap rage ("feelings of intense anger when failing to open packaging or wrapping").
When we say "every thing", we mean it *extremely* literally.
Much work has gone into making this list actually usable. Voting is done pairwise, asking "Which topic is more important to discuss?" This allows the ranking to scale theoretically indefinitely; we are currently working on ways to make voting even more attention-tractable. Our core concept is around coordinating extremely large discussions, and this is part of that.
I personally really like ranking random things, it's like the old Wikipedia random page with a much higher "weird thing" hit rate. We also have a voting modes centered around ranking high importance items (top n-thousand, it has varied as the set of potentially important ranked items has gotten larger), and one centered around ranking items selected to make the overall convergence at the top faster and more uniform.
This has the effect of making the overall list much more usable. There's an implicit two-stage ranking process where items are identified as important at all, then thrown into the more detailed ranking if they win. Items in the long tail might not always be considered (there are still many unranked topics), but empirically very important topics have many near duplicates, such as "climate change" and "global warming".
As bulky as this is, it does actually still surface meaningful long-tail items to vote on. An item found early on was voicism, "stereotyping or discrimination due to voice". It's a surprising issue that unites both trans women and the deep south (think the "southern hick" accent). Others include "fatigue as a safety concern" and "institutionalized adolescent". "Illegal, unregulated, and unreported fishing" actually appears prominently on the twothirds platform.
This list is, of course, not the Platonic ideal of the original goal, but it *is* as close as possible to a practical implementation. The top ten items do not need to be rated closely enough to distinguish them, that's likely impossible. However, creating a list where the top 500 accurately reflect a highly comprehensive yet completely neutrally selected set of important priorities is very possible.
-----
With regard to list UI, here's what is done and what is in progress:
There are a number of pain points generated by concerns around bias avoidance. Browsing items by rank is disabled past the first 2000 items, because we don't want to introduce the possibility of being used for things like determining the least important item. Our system does not attempt to find this in any way, and we don't want to present ourselves as having made a meaningful search for it. This makes some issues related to topic discovery very hard, there needs to be more implicit exploration of the list as people vote. Right now the ranking discussion generates much of our long tail comment stream, but we're slowly working to improve this.
Our search UI uses importance rank or result order by default. We wanted a ranking algorithm that is completely understandable, that was also extremely difficult to manipulate by AI. Changing the rankings requires hours of voting, mostly on topics that an attacker would have no opinion on, meaning we get a very detailed opinion portrait to use as a bot signature. (We do also support a full relevance based search, because it's honestly nicer to use, but having both makes sure no topics are buried by "the algorithm".)
Navigating between topics is also a bit cramped - Fighting against infinite lists is actually a last resort defense against ballot stuffing attacks used site-wide, and the UI is built to allow shut down voting on all specific topics directly, forcing users to confront the shuffled stream of all comments. The goal is that is like attempting to read all of Twitter, making targeted manipulation effectively impossible. However, this is admittedly hard to use, so there are chunks of our UI that were originally optimized for this task that are being retrofitted.
There are also some surprising behaviors stemming from the low-Javascript constraint. For example, we return 100 results by default per search as a way of emulating infinite scrolling. Our search controls are also implemented within <details> tags, as basic HTML lacks many facilities for manipulating buttons. By having our controls remain raw HTML, navigation of the page using older software is better supported, because their descriptions are present in plain text on the page.
Anyway, this is why the site has so many quirks, it's doing a ton of stuff, and the reasons for much of it are not obvious - Balancing shape principles, quality of life changes, and rigorous consideration of biases and manipulation is hard. Just know that we're doing our best! Our goal is that someday our designs are chosen officially as a better way of running a democracy, and we take that very seriously.