Open Problems in Archipelago

Over a year ago, I wrote about Public Archipelago and why it seemed im­por­tant for LessWrong. Since then, noth­ing much has come of that. It seemed im­por­tant to ac­knowl­edge that. There are more things I think are worth try­ing, but I’ve up­dated a bit that maybe the prob­lem is in­tractable or my frame on it might be wrong.

The core prob­lems Public Archipelago was aiming to solve are:

  • By de­fault, pub­lic spaces and dis­cus­sions force con­ver­sa­tion and progress to hap­pen at the low­est com­mon de­nom­i­na­tor.

  • This re­sults in a de­fault of high-effort pro­jects hap­pen­ing in pri­vate, where it is harder for oth­ers to learn from.

  • The peo­ple do­ing high-effort pro­jects have lots of in­ter­nal con­text, which is hard to com­mu­ni­cate and get peo­ple up to speed on in a pub­lic set­ting. But in­ter­nally, they can talk eas­ily about it. So that ends up be­ing what they do by de­fault.

  • Longterm, this kills the en­g­ine by which in­tel­lec­tual growth hap­pens. It’s what kil­led old LessWrong – all the in­ter­est­ing pro­jects were hap­pen­ing in pri­vate, (usu­ally in-per­son) spaces, and that meant that:

    • new­com­ers couldn’t latch onto them and learn about them incidentally

    • at least some im­por­tant con­cepts didn’t en­ter the in­tel­lec­tual com­mons, where they could ac­tu­ally be cri­tiqued or built upon

The solu­tion was a world of spaces that were pub­lic, but with bar­ri­ers to en­try, and/​or the abil­ity to kick peo­ple out. So peo­ple could eas­ily have the high-con­text con­ver­sa­tions that they wanted, but new­com­ers could slowly ori­ent around those con­ver­sa­tions, and oth­ers could ei­ther cri­tique those ideas in their own posts, or build off them.

Since last year, very few of my hopes have ma­te­ri­al­ized.

(I think LessWrong in gen­eral has done okay, but not great, and Public-Archipelago-es­que things in par­tic­u­lar have not hap­pened, and there’s con­tinued to be in­ter­est­ing dis­cus­sion in pri­vate ar­eas that not ev­ery­one is privy to)

I think the only thing that came close is some dis­cus­sion on AI Align­ment top­ics, which benefited from be­ing tech­ni­cal enough to au­to­mat­i­cally have a bar­rier to en­try, and cre­ated a dis­cus­sion shaped in such a way that it was harder to drag it into Over­ton Win­dow Fights.

The core prob­lem is that main­tain­ing a high-con­text space re­quires a col­lec­tion of skills that few peo­ple have, and even if they do, it re­quires effort to main­tain.

The mod­er­a­tion tools we built last year still re­quire: a lot of ac­tive effort on the part of the in­di­vi­d­ual users, and that effort is kinda in­trin­si­cally aver­sive (tel­ling peo­ple to go away is a hard skill and comes with so­cial risks), and it also re­quires peo­ple to have lots of ideas that are in­ter­est­ing enough in the first place to build a high-con­text con­ver­sa­tion around.

The cur­rent im­ple­men­ta­tion re­quires all three of those skills in a sin­gle per­son.

There are a few al­ter­nate im­ple­men­ta­tions that could work, but re­quires a fair amount of dev work, and mean­while we have other pro­jects that seem higher pri­or­ity. Some ex­am­ples:

  • Peo­ple have asked for sub­red­dits for awhile. Be­fore we build that, we want to make sure that they’re de­signed such that good ideas are ex­pected to “bub­ble up” to the top of LessWrong, rather than stay in nested filters for­ever.

  • Opt in rather than opt out mod­er­a­tion (i.e. peo­ple might have a list of col­lab­o­ra­tors, and only col­lab­o­ra­tors can com­ment on their posts, rather than a banned list). This is ba­si­cally what FB and Google Docs does.

  • I had some vague ideas for “free­lance mod­er­a­tors”. We give au­thors with 2000 karma the abil­ity to delete com­ments and ban users, but this is rarely used, be­cause it re­quires some­one who is both will­ing to mod­er­a­tor and who can write well. Split­ting those into two sep­a­rate roles could be use­ful.

I’m most op­ti­mistic about the sec­ond op­tion.

I think sub­red­dits are go­ing to be a use­ful tool that I ex­pect to build sooner or later, but they won’t ac­com­plish most-of-the-thing. Most of what I’m ex­cited about are not sub­red­dits by topic, but highly-con­text-driven con­ver­sa­tions with some nu­anced fla­vor that doesn’t neatly map to the sort of top­ics that sub­red­dits tend to have. Plus, sub­red­dits still mean some­one has to do the work of polic­ing the bor­der, which is the biggest pain point of the en­tire pro­cess.

If I were to try the sec­ond op­tion and it still didn’t re­sult in the kinds of out­comes I’m look­ing for, I’d up­date away from Public Archipelago be­ing a vi­able frame for in­tel­lec­tual dis­course.

(I do think the sec­ond op­tion still re­quires a bit of effort to get right – it’s im­por­tant that the pro­cess be seam­less and easy and a salient op­tion to peo­ple. And thus, it’ll prob­a­bly still be a while be­fore I’d have the band­width to push for it)