Eliezer's vision for Arbital

By Eliezer Yudkowsky

I’ve spent my en­tire life since age 15 try­ing to ex­plain com­pli­cated sub­jects. Some­times I suc­ceeded, more of­ten I failed. I knew that the tech­nol­ogy I was us­ing — long es­says, and later, se­quences of blog posts with in­ter­nal links — wasn’t up to the job; but en­vi­sion­ing some­thing bet­ter was hard.

It wasn’t just that my par­tic­u­lar top­ics were difficult. In gen­eral, our civ­i­liza­tion’s on­line tech­nol­ogy does not seem to be very good at sys­tem­at­i­cally solv­ing the prob­lem of “ex­plain­ing things”. If you want to know the pop­u­la­tion of Melbourne, Wikipe­dia has you cov­ered. If you look at the Wikipe­dia page on Bayes’ rule — which is it­self more ac­cessible than the Wikipe­dia page on Bayes’ the­o­rem that some­one would prob­a­bly en­counter first — and imag­ine some or­di­nary in­no­cent reader try­ing to un­der­stand the equa­tions… well, it’s clear that Wikipe­dia isn’t try­ing to solve the prob­lem of ex­plain­ing things to a re­al­is­tic dis­tri­bu­tion of in­com­ing read­ers with differ­ent lev­els of math abil­ity.

Which is not Wikipe­dia’s fault. Wiki tech­nol­ogy and the en­cy­clo­pe­dia idiom are not set up to solve the prob­lem of ex­plain­ing top­ics to many read­ers com­ing in with differ­ent lev­els of back­ground knowl­edge. A good ex­pla­na­tion for a com­puter pro­gram­mer is in­ac­cessible to a high school stu­dent who’s good at math. An ex­pla­na­tion aimed at the high school stu­dent would be painfully long and bor­ing for a math­e­mat­i­cian.

Pro­lifer­ate differ­ent ver­sions of pages aimed at differ­ent read­ers? That’s not as easy as mak­ing more wiki pages — in­for­ma­tion is only helpful if you can find it. Wikipe­dia tries to have rel­a­tively few web pages (hence ‘no­ta­bil­ity’) be­cause their cen­tral idiom for dis­cov­er­ing in­for­ma­tion is ‘search un­der the ob­vi­ous name for it’. If Wikipe­dia pages started pro­lifer­at­ing sub­pages, the site would be­come much more difficult to nav­i­gate, as­sum­ing noth­ing else changed.

So real ex­pla­na­tions, ex­pla­na­tions that try to be ac­cessible, live in ran­domly scat­tered places around the In­ter­net. Google around for Bayes in­tros? Good luck try­ing to find the one that fits you! Need to learn an­other con­cept so you can un­der­stand this one? En­joy try­ing to figure out what you need or where to look for it! See a mys­te­ri­ous word you don’t un­der­stand? If it’s high­lighted, you get to click on it blindly; if not, you get to Google it blindly. Then you en­counter a whole new page, and can restart the prob­lem of figur­ing out how to learn that page from scratch. The only sys­tem­atic way our civ­i­liza­tion has of solv­ing this prob­lem is to go through a se­ries of col­lege courses in which you read text­books that as­sume you’ve taken all the pre­vi­ous col­lege courses and have read all the pre­vi­ous chap­ters in the text­book.

With Ar­bital, we’re mak­ing a se­ri­ous run at solv­ing the UI and struc­tural prob­lems that pre­vent Wikipe­dia, wikis, on­line web­pages, and col­lege text­books from mak­ing knowl­edge eas­ily ac­cessible. After you’ve walked through the Ar­bital Guide to Bayes’ rule and got­ten used to Ar­bital’s lenses, req­ui­sites, and pop­ups, you’ll have some idea of where Ar­bital is head­ing — though, I rush to em­pha­size, the Bayes Guide is just a demo, and Ar­bital isn’t fea­ture-com­plete for the job of ex­plain­ing things, let alone its larger am­bi­tions.

We have a dream. We have a dream of a world where any­one who wants to un­der­stand some­thing nav­i­gates to Ar­bital’s page on Bayes, or on NGDP tar­get­ing, or on pop­u­la­tion ge­net­ics, sees weird terms, pops up the sum­mary, says “What the heck?”, goes to the least tech­ni­cal tab for the page, says “What the heck?” again, and hits Ar­bital’s “Plot a learn­ing path to this con­cept” but­ton. Re­mem­ber a time when peo­ple used to be able to have huge ar­gu­ments over the pop­u­la­tion of Melbourne, the time be­fore Wikipe­dia? And now we don’t live in that time any­more? Mak­ing the same kind of ad­vance for ‘ex­plain­ing things’ po­ten­tially mat­ters A LOT.

Ar­bital has big­ger am­bi­tions than even that. We all dream of a world that elimi­nates the du­pli­ca­tion of effort in on­line ar­gu­ment — a world where, the same way that Wikipe­dia cen­tral­ized the record­ing of definite facts, an ar­gu­ment only needs to hap­pen once, in­stead of be­ing redu­pli­cated all over the In­ter­net; with all the branches of the ar­gu­ment neatly recorded in the same place, along with some in­di­ca­tion of who be­lieves what. A world where ‘just check Ar­bital’ had the same sta­tus for de­ter­min­ing the cur­rent state of de­bates, as ‘just check Wikipe­dia’ now has when some­body starts ar­gu­ing about the pop­u­la­tion of Melbourne. There’s en­tirely new big sub­prob­lems and solu­tions, not pre­sent at all in the cur­rent Ar­bital, that we’d need to tackle that con­sid­er­ably more difficult prob­lem. But to solve ‘ex­plain­ing things’ is some­thing of a first step. If you have a sin­gle URL that you can point any­one to for ‘ex­plain­ing Bayes’, and if you can dis­patch peo­ple to differ­ent pages de­pend­ing on how much math they know, you’re start­ing to solve some of the key sub­prob­lems in re­mov­ing the re­dun­dancy in on­line ar­gu­ments.

But even if we fail at that part, mak­ing hu­man­ity’s knowl­edge more ac­cessible and ac­tu­ally learn­able in real life is some­thing that mat­ters by it­self.

If this sounds like a wor­thy goal to you and you’d like to help, we’d love to have you on board.