One thing that feels cool about personal wikis is that people come up with their own factorization and ontology for the things they are thinking about...So I think in addition to the above there needs to be a way for users to easily and without friction add a personal article for some concept they care about, and to have a consistent link to it, in a way that doesn’t destroy any of the benefits of the collaborative editing.
My proposal already provides a way to easily add a personal article with a consistent link, while preserving the ability to do collaborative editing on ‘public’ articles. Strictly speaking, it’s fine for people to add wiki entries for their own factorization and ontology.
There is no requirement for those to all be ‘official’: there doesn’t have to be a ‘consensus’ entry. Nothing about a /wiki/Acausal_cooperation/gwern user entry requires the /wiki/Acausal_cooperation consensus entry to exist. (Computers are flexible like that.) That just means there’s nothing there at that exact URL, or probably better, it falls back to displaying all sub-pages of user entries like usual. (User entries presumably get some sort of visual styling, in the same way that comments on a post look different from a post, which in addition to the title/author metadata displayed, avoids confusion.)
If, say, TurnTrout wants to create a user entry /wiki/Reward_is_not_the_optimization_target/TurnTrout as a master key to all of his posts & comments and related ones like Nora Belrose’s posts, rather than go for a consensus entry, that’s fine. And if it becomes commonly-accepted jargon and part of the ontology, and so it becomes a hassle that people can’t edit his user entry (rather than leave their own user entry or comments), his user entry can be ‘promoted’ ie. just be copied by someone into a new consensus entry at /wiki/Reward_is_not_the_optimization_target that can then be edited and his user entry left as historical or possibly collapsed/hidden by an admin for readability.
(The growth of the ad hoc user & consensus ontology might be a bit messy and sometimes it might be a good idea to delete/edit user entries by users who are gone or uncooperative or edit their entries to update links, but that’s little different from a regular wiki’s need for admins to do similar maintenance.)
Like, we probably won’t have a consensus article on the exact ways L in Death Note made mistakes, but Gwern.net would be sadder without that kind of content.
The DN essay mostly would not make sense as a wiki entry, and indeed, it’s been ‘done’ ever since 2013 or so. There’s not much more to be said about the topic (despite occasional attempts at criticism, which typically just wind up repeating something I already said in it). It doesn’t need wiki support or treatment, and it was a post-like essay: I wrote it up as a single definitive piece and it was discussed at a particular time and place. (Actually, I think I did originally post it on LW1?) It benefits from ongoing Gwern.net improvements, but mostly in a general typographical sense rather than being deeply integrated into other pages.
The parts of it that keep changing do have wiki-like nature:
For example, the parts from Jaynes would make sense as part of a ‘Legal Bayesianism’ article, which would be useful to invoke in many other posts like debates on Amanda Knox’s innocence.
The basic concept of ‘information’ from information theory as whatever lets you narrow down a haystack into a needle (which is the ‘big idea’ of the essay—teaching you information theory by example, by dramatizing the hunt for a criminal leaking circumstantial evidence) is certainly a wiki-worthy topic that people could benefit from naming explicitly and appending their own discussions or views on.
This could come up in many places, from looking for aliens to program search in DL scaling or AIXI paradigm or thinking about Bayesian reasoning (eg. Eliezer on how many bits of information it takes to make a hypothesis ‘live’ at all).
Or that big list of side channels / deanonymization methods would make complete sense as a wiki entry which people could contribute piquant instances to, and would be useful linking elsewhere on LW2, particularly in articles on AGI security and why successfully permanently boxing a malevolent, motivated superintelligence would be extremely difficult—because there are bazillions of known side-channels which enable exfiltration/control/inference & we keep discovering new ones like how to turn computer chips into radios/speakers or entire families of attacks like SPECTRE or Rowhammer.
(One reason I’ve invested so much effort into the tag-directory system is the hope of replacing such churning lists with transcludes of tags. The two major examples right now are https://gwern.net/dnm-archive#works-using-this-dataset and https://gwern.net/danbooru2021#publications—I want to track all users/citers of my datasets to establish their value for researchers & publicize those uses, but adding them manually was constant toil and increasingly redundant with the annotations. So with appropriate tooling, I switch to transcluding a tag for the citers instead. Any time a new user shows up, I just write an annotation for it, as I would have before, and add a dnm-archive or danbooru tag to it and then they show up automatically with no further work. So you could imagine doing the same thing in my DN essay: instead of that long unordered list, which is tedious to update every time there’s a fun security paper or blog post, I would instead have a tag like cs/security/side-channel where each of those is annotated, and simply transclude the table of contents for that. If I still wanted a natural-language summary similar to the existing list, well, I could just stick that at the top of the tag and benefit every instance of the tag.)
My proposal already provides a way to easily add a personal article with a consistent link, while preserving the ability to do collaborative editing on ‘public’ articles. Strictly speaking, it’s fine for people to add wiki entries for their own factorization and ontology.
There is no requirement for those to all be ‘official’: there doesn’t have to be a ‘consensus’ entry. Nothing about a
/wiki/Acausal_cooperation/gwernuser entry requires the/wiki/Acausal_cooperationconsensus entry to exist. (Computers are flexible like that.) That just means there’s nothing there at that exact URL, or probably better, it falls back to displaying all sub-pages of user entries like usual. (User entries presumably get some sort of visual styling, in the same way that comments on a post look different from a post, which in addition to the title/author metadata displayed, avoids confusion.)If, say, TurnTrout wants to create a user entry
/wiki/Reward_is_not_the_optimization_target/TurnTroutas a master key to all of his posts & comments and related ones like Nora Belrose’s posts, rather than go for a consensus entry, that’s fine. And if it becomes commonly-accepted jargon and part of the ontology, and so it becomes a hassle that people can’t edit his user entry (rather than leave their own user entry or comments), his user entry can be ‘promoted’ ie. just be copied by someone into a new consensus entry at/wiki/Reward_is_not_the_optimization_targetthat can then be edited and his user entry left as historical or possibly collapsed/hidden by an admin for readability.(The growth of the ad hoc user & consensus ontology might be a bit messy and sometimes it might be a good idea to delete/edit user entries by users who are gone or uncooperative or edit their entries to update links, but that’s little different from a regular wiki’s need for admins to do similar maintenance.)
The DN essay mostly would not make sense as a wiki entry, and indeed, it’s been ‘done’ ever since 2013 or so. There’s not much more to be said about the topic (despite occasional attempts at criticism, which typically just wind up repeating something I already said in it). It doesn’t need wiki support or treatment, and it was a post-like essay: I wrote it up as a single definitive piece and it was discussed at a particular time and place. (Actually, I think I did originally post it on LW1?) It benefits from ongoing Gwern.net improvements, but mostly in a general typographical sense rather than being deeply integrated into other pages.
The parts of it that keep changing do have wiki-like nature:
For example, the parts from Jaynes would make sense as part of a ‘Legal Bayesianism’ article, which would be useful to invoke in many other posts like debates on Amanda Knox’s innocence.
The basic concept of ‘information’ from information theory as whatever lets you narrow down a haystack into a needle (which is the ‘big idea’ of the essay—teaching you information theory by example, by dramatizing the hunt for a criminal leaking circumstantial evidence) is certainly a wiki-worthy topic that people could benefit from naming explicitly and appending their own discussions or views on.
This could come up in many places, from looking for aliens to program search in DL scaling or AIXI paradigm or thinking about Bayesian reasoning (eg. Eliezer on how many bits of information it takes to make a hypothesis ‘live’ at all).
Or that big list of side channels / deanonymization methods would make complete sense as a wiki entry which people could contribute piquant instances to, and would be useful linking elsewhere on LW2, particularly in articles on AGI security and why successfully permanently boxing a malevolent, motivated superintelligence would be extremely difficult—because there are bazillions of known side-channels which enable exfiltration/control/inference & we keep discovering new ones like how to turn computer chips into radios/speakers or entire families of attacks like SPECTRE or Rowhammer.
(One reason I’ve invested so much effort into the tag-directory system is the hope of replacing such churning lists with transcludes of tags. The two major examples right now are https://gwern.net/dnm-archive#works-using-this-dataset and https://gwern.net/danbooru2021#publications—I want to track all users/citers of my datasets to establish their value for researchers & publicize those uses, but adding them manually was constant toil and increasingly redundant with the annotations. So with appropriate tooling, I switch to transcluding a tag for the citers instead. Any time a new user shows up, I just write an annotation for it, as I would have before, and add a
dnm-archiveordanboorutag to it and then they show up automatically with no further work. So you could imagine doing the same thing in my DN essay: instead of that long unordered list, which is tedious to update every time there’s a fun security paper or blog post, I would instead have a tag likecs/security/side-channelwhere each of those is annotated, and simply transclude the table of contents for that. If I still wanted a natural-language summary similar to the existing list, well, I could just stick that at the top of the tag and benefit every instance of the tag.)