I spent an hour or two talking about these problems with Ruby. Here are two further thoughts. I will reiterate that I have little experience with wikis and tagging, so I am likely making some simple errors.
Connecting Tagging and Wikis
One problem to solve is that if a topic is being discussed, users want to go from a page discussing that topic to find a page that explains that topic, and lists all posts that discuss that topic. This page should be easily update-able with new content on the topic.
Some more specific stories:
A user reads a post on a topic, and wants to better understand what’s already known about that topic and the basic ideas
A user is primarily interested in a topic, and wants to make sure to see all content about that topic
The solution for the first is to link to a page that contains all other posts on that topic. The solution to the second is to link to a wiki page on that topic. And one possible solution is to make both of those the same button.
This page is a combination of a Wiki and a Tag. It is a communally editable explanation of the concept, with links to key posts explaining it, and other pages that are related. And below that, it also has a post-list of every posts that is relevant, sortable by things like recency, karma, and relevancy. Maybe below that it even has its own Recent Discussion section, for comments on posts that have the tag. It’s a page you can subscribe to (e.g. via RSS), and come back to to see discussion of a particular topic.
Now, to make this work, it’s necessary that all posts that are in the category are successfully listed in the tag. One problem you will run into is that there are a lot of concepts in the space, so the number of such pages will quickly become unmanageable. “Inner Alignment”, “Slack”, “Game Theory”, “Akrasia”, “Introspection”, “Corrigibility”, etc, is a very large list, such that it is not reasonable to scroll through it and check if your post fits into any of them, and expect to do this successfully. You’ll end up with a lot of Wiki pages with very incomplete lists.
This is especially bad, because the other use of the tag system you might be hoping for is the one described in the parent to this comment, where you can see the most relevant tags directly from the frontpage, to help with figuring out what you want to read. If you want to make sure to read all the AI alignment posts, it’s not helpful to give you a tag that sometimes works, because then you still have to check all the other posts anyway.
However, there are three ways to patch this over. Firstly, the thing that will help the Wiki system the most here, is the ability to add posts to the Wiki page from the post page, instead of having to independently visit the Wiki page and then add it in. This helps the people who care about maintaining Wiki pages quite a bit, making their job much easier.
Secondly, you can help organise those tags in order of likely relevance. For example, if you link to a lot of posts that have the tag “AI alignment” then you probably are about AI alignment, so that tag should appear higher.
Thirdly, you can sort tags into two types. The first type is given priority, and is a very controlled set of concepts, that also get used for filtering on the frontpage. This is a small, stable set of tags that people learn and can easily confirm if you should be sorted by. The second is the much larger, user-generated set of tags that correspond to user-generated wiki pages, and there can be 100s of these.
In this world, wiki pages are split into two types: those that are tags and those that aren’t. Those which are tags have a big post-list item that is searchable, maybe even a recent discussion section, and can be used to tag posts. Those that are not tags do not have these features and properties.
This idea seems fairly promising to me, and I don’t see any problems with it yet. For the below, I’ll call such a page a ‘WikiTag’.
Conceptual updating
Speaking more generally, my main worry about a lot of systems like Wikis and Tagging is about something that is especially prevalent in science and in the sort of work we do on LessWrong, where we try to figure out better conceptual boundaries to draw in reality, and whereby old concepts get deprecated. I expect that on sites like lobste.rs and Gelbooru, tags rarely turn out to have been the wrong way to frame things. There are rarely arguments about whether something is really a blue sky, or just the absence of clouds. Whereas a lot of progress in science is this sort of subtle conceptual progress, where you maybe shouldn’t have said that the object fell to the ground, but instead that the object and the Earth fell into each other at rates proportional to some function of their masses.
On LessWrong I think we’ve done a lot of this sort of thing.
We used to talk about optimisation daemons, now we talk about the inner alignment problem.
We used to talk about people being stupid and the world being mad, and now we talk about coordination problems.
We used to talk about agent foundations and now we maybe think embedded agency is a better conceptualisation of the problem.
In places like the in-person CFAR space I’ve heard talk of akrasia often deprecated and instead ideas like ‘internal alignment’ are discussed.
We made progress from TDT to UDT.
So I’m generally worried about setting up infrastructure that makes concepts get stuck in place, by e.g. whoever picked the name first.
One problem I was worried about, was that all post would have to be categorised according to the old names. In particular, post that have already been tagged ‘optimisation daemons’ would now have a hard time changing to being tagged ‘inner alignment problem’.
However, after fleshing it out, I’m not so sure it’s going to be a problem.
Firstly, it’s not clear that old posts should have their tags updated. If there is a sequence of posts taking about akrasia and how to deal with it, it would be very confusing for those posts to have a tag for ‘internal alignment’, a term not mentioned anywhere in the post nor obviously related to the framing of the posts. Similarly for ‘optimisation daemons’ discussion to be called ‘the inner alignment problem’.
Secondly, there’s a fairly natural thing to do when such conceptual shifts in the conversation occur. You build a new WikiTag. Then you tag all the new posts, and write the wiki entry explaining the concept, and link back to the old concept. It just needs to say something like “Old work was done under the idea that objects fell down to the ground. We now think that the object and the Earth fall into each other, but you can see the old work and its experimental results on this page <link>. Plus here are some links to the key posts back then that you’ll still want to know about today.” And indeed if such a thing happens with agent foundations and embedded agency, or something, then it’ll be necessary to have posts explaining how the old work fits into the current paradigm. That translational work is not done by renaming a tag, but by a person who understands that domain writing some posts explaining how to think about and use the old work, in the new conceptual framework. And those should be prominently linked to on the wiki/tag page.
So I think that this system does not have the problems I thought that it had.
I guess I’m still fairly worried about subtle errors, like if instead of a tag for ‘Forecasting’ we have a tag called ‘Calibration’ or ‘Predictions’, these would shift the discourse in different ways. I’m a bit worried about that. But I think it’s likely that a small community like ours will overall be able to resist such small shifts, and that argument will prevail, even if the names are a little off sometimes. It sounds like a problem that makes progress a little slower but doesn’t push it off the rails. And if the tag is sufficiently wrong then I expect we can do the process above, where we start a new tag and link back to the old tag. Or, if the conceptual shift is sufficiently small (e.g. ‘Forecasting’ → ‘Predictions’) I can imagine renaming the tag directly.
So I’m no longer so worried about conceptual stickiness as a fundamental blocker to Wikis and Tagging as ways of organising the conceptual space.
As a general comment, StackExchange’s tagging system seems pretty perfect (and battle-tested) to me, and I suspect we should just copy their design as closely as we can.
So, on StackExchange any user can edit any of the tags, and then there is a whole complicated hierarchy that exists for how to revert changes, how to approve changes, how to lock posts from being edited, etc.
Which is a solution, but it sure doesn’t seem like an easy or elegant solution to the tagging problem.
I think the peer review queue is pretty sensible in any world where there’s “one ground truth” that you expect trusted users to have access to (such that they can approve / deny edits that cross their desk).
I spent an hour or two talking about these problems with Ruby. Here are two further thoughts. I will reiterate that I have little experience with wikis and tagging, so I am likely making some simple errors.
Connecting Tagging and Wikis
One problem to solve is that if a topic is being discussed, users want to go from a page discussing that topic to find a page that explains that topic, and lists all posts that discuss that topic. This page should be easily update-able with new content on the topic.
Some more specific stories:
A user reads a post on a topic, and wants to better understand what’s already known about that topic and the basic ideas
A user is primarily interested in a topic, and wants to make sure to see all content about that topic
The solution for the first is to link to a page that contains all other posts on that topic. The solution to the second is to link to a wiki page on that topic. And one possible solution is to make both of those the same button.
This page is a combination of a Wiki and a Tag. It is a communally editable explanation of the concept, with links to key posts explaining it, and other pages that are related. And below that, it also has a post-list of every posts that is relevant, sortable by things like recency, karma, and relevancy. Maybe below that it even has its own Recent Discussion section, for comments on posts that have the tag. It’s a page you can subscribe to (e.g. via RSS), and come back to to see discussion of a particular topic.
Now, to make this work, it’s necessary that all posts that are in the category are successfully listed in the tag. One problem you will run into is that there are a lot of concepts in the space, so the number of such pages will quickly become unmanageable. “Inner Alignment”, “Slack”, “Game Theory”, “Akrasia”, “Introspection”, “Corrigibility”, etc, is a very large list, such that it is not reasonable to scroll through it and check if your post fits into any of them, and expect to do this successfully. You’ll end up with a lot of Wiki pages with very incomplete lists.
This is especially bad, because the other use of the tag system you might be hoping for is the one described in the parent to this comment, where you can see the most relevant tags directly from the frontpage, to help with figuring out what you want to read. If you want to make sure to read all the AI alignment posts, it’s not helpful to give you a tag that sometimes works, because then you still have to check all the other posts anyway.
However, there are three ways to patch this over. Firstly, the thing that will help the Wiki system the most here, is the ability to add posts to the Wiki page from the post page, instead of having to independently visit the Wiki page and then add it in. This helps the people who care about maintaining Wiki pages quite a bit, making their job much easier.
Secondly, you can help organise those tags in order of likely relevance. For example, if you link to a lot of posts that have the tag “AI alignment” then you probably are about AI alignment, so that tag should appear higher.
Thirdly, you can sort tags into two types. The first type is given priority, and is a very controlled set of concepts, that also get used for filtering on the frontpage. This is a small, stable set of tags that people learn and can easily confirm if you should be sorted by. The second is the much larger, user-generated set of tags that correspond to user-generated wiki pages, and there can be 100s of these.
In this world, wiki pages are split into two types: those that are tags and those that aren’t. Those which are tags have a big post-list item that is searchable, maybe even a recent discussion section, and can be used to tag posts. Those that are not tags do not have these features and properties.
This idea seems fairly promising to me, and I don’t see any problems with it yet. For the below, I’ll call such a page a ‘WikiTag’.
Conceptual updating
Speaking more generally, my main worry about a lot of systems like Wikis and Tagging is about something that is especially prevalent in science and in the sort of work we do on LessWrong, where we try to figure out better conceptual boundaries to draw in reality, and whereby old concepts get deprecated. I expect that on sites like lobste.rs and Gelbooru, tags rarely turn out to have been the wrong way to frame things. There are rarely arguments about whether something is really a blue sky, or just the absence of clouds. Whereas a lot of progress in science is this sort of subtle conceptual progress, where you maybe shouldn’t have said that the object fell to the ground, but instead that the object and the Earth fell into each other at rates proportional to some function of their masses.
On LessWrong I think we’ve done a lot of this sort of thing.
We used to talk about optimisation daemons, now we talk about the inner alignment problem.
We used to talk about people being stupid and the world being mad, and now we talk about coordination problems.
We used to talk about agent foundations and now we maybe think embedded agency is a better conceptualisation of the problem.
In places like the in-person CFAR space I’ve heard talk of akrasia often deprecated and instead ideas like ‘internal alignment’ are discussed.
We made progress from TDT to UDT.
So I’m generally worried about setting up infrastructure that makes concepts get stuck in place, by e.g. whoever picked the name first.
One problem I was worried about, was that all post would have to be categorised according to the old names. In particular, post that have already been tagged ‘optimisation daemons’ would now have a hard time changing to being tagged ‘inner alignment problem’.
However, after fleshing it out, I’m not so sure it’s going to be a problem.
Firstly, it’s not clear that old posts should have their tags updated. If there is a sequence of posts taking about akrasia and how to deal with it, it would be very confusing for those posts to have a tag for ‘internal alignment’, a term not mentioned anywhere in the post nor obviously related to the framing of the posts. Similarly for ‘optimisation daemons’ discussion to be called ‘the inner alignment problem’.
Secondly, there’s a fairly natural thing to do when such conceptual shifts in the conversation occur. You build a new WikiTag. Then you tag all the new posts, and write the wiki entry explaining the concept, and link back to the old concept. It just needs to say something like “Old work was done under the idea that objects fell down to the ground. We now think that the object and the Earth fall into each other, but you can see the old work and its experimental results on this page <link>. Plus here are some links to the key posts back then that you’ll still want to know about today.” And indeed if such a thing happens with agent foundations and embedded agency, or something, then it’ll be necessary to have posts explaining how the old work fits into the current paradigm. That translational work is not done by renaming a tag, but by a person who understands that domain writing some posts explaining how to think about and use the old work, in the new conceptual framework. And those should be prominently linked to on the wiki/tag page.
So I think that this system does not have the problems I thought that it had.
I guess I’m still fairly worried about subtle errors, like if instead of a tag for ‘Forecasting’ we have a tag called ‘Calibration’ or ‘Predictions’, these would shift the discourse in different ways. I’m a bit worried about that. But I think it’s likely that a small community like ours will overall be able to resist such small shifts, and that argument will prevail, even if the names are a little off sometimes. It sounds like a problem that makes progress a little slower but doesn’t push it off the rails. And if the tag is sufficiently wrong then I expect we can do the process above, where we start a new tag and link back to the old tag. Or, if the conceptual shift is sufficiently small (e.g. ‘Forecasting’ → ‘Predictions’) I can imagine renaming the tag directly.
So I’m no longer so worried about conceptual stickiness as a fundamental blocker to Wikis and Tagging as ways of organising the conceptual space.
As a general comment, StackExchange’s tagging system seems pretty perfect (and battle-tested) to me, and I suspect we should just copy their design as closely as we can.
So, on StackExchange any user can edit any of the tags, and then there is a whole complicated hierarchy that exists for how to revert changes, how to approve changes, how to lock posts from being edited, etc.
Which is a solution, but it sure doesn’t seem like an easy or elegant solution to the tagging problem.
I think the peer review queue is pretty sensible in any world where there’s “one ground truth” that you expect trusted users to have access to (such that they can approve / deny edits that cross their desk).
It’s also important to have the old concept link to the new concept.