Is community-collaborative article production possible?

When I showed up at the Sin­gu­lar­ity In­sti­tute, I was sur­prised to find that 30-60 pa­pers’ worth of ma­te­rial was ly­ing around in blog posts, mailing list dis­cus­sions, and peo­ple’s heads — but it had never been writ­ten up in clear, well-refer­enced aca­demic ar­ti­cles.

Why is this so? Writ­ing such ar­ti­cles has many clear benefits:

  • Clearly stated and well-defended ar­gu­ments can per­suade smart peo­ple to take AI risk se­ri­ously, cre­at­ing ad­di­tional sup­port­ers and col­lab­o­ra­tors for the Sin­gu­lar­ity In­sti­tute.

  • Such ar­ti­cles can also im­prove the cred­i­bil­ity of the or­ga­ni­za­tion as a whole, which is es­pe­cially im­por­tant for at­tract­ing funds from top-level so­cial en­trepreneurs and in­sti­tu­tions like the Gates Foun­da­tion and Givewell.

  • Lay­ing out the ar­gu­ments clearly and an­a­lyz­ing each premise can lead to new strate­gic in­sights that will help us un­der­stand how to pur­chase x-risk re­duc­tion most effi­ciently.

  • Clear ex­pla­na­tions can provide a plat­form on which re­searchers can build to pro­duce new strate­gic and tech­ni­cal re­search re­sults.

  • Com­mu­ni­cat­ing clearly is what lets other peo­ple find er­rors in your rea­son­ing.

  • Com­mu­ni­ties can use ar­ti­cles to cut down on com­mu­ni­ca­tion costs. When some­thing is writ­ten up clearly, 1000 peo­ple can read a sin­gle ar­ti­cle in­stead of need­ing to trans­mit the in­for­ma­tion by hav­ing sev­eral hun­dred per­sonal con­ver­sa­tions be­tween 2-5 peo­ple.

Of course, there are costs to writ­ing ar­ti­cles, too. The sin­gle biggest cost is staff time /​ op­por­tu­nity cost. An ar­ti­cle like “In­tel­li­gence Ex­plo­sion: Ev­i­dence and Im­port” can re­quire any­where from 150-800 per­son-hours. That is 150-800 paid hours dur­ing which our staff is not do­ing other crit­i­cally im­por­tant things that col­lec­tively have a big­ger pos­i­tive im­pact than a sin­gle aca­demic ar­ti­cle is likely to have.

So Louie Helm and Nick Beck­stead and I sat down and asked, “Is there a way we can buy these ar­ti­cles with­out such an egre­gious cost?”

We think there might be. Ba­si­cally, we sus­pect that most of the work in­volved in writ­ing these ar­ti­cles can be out­sourced. Here’s the pro­cess we have in mind:

  1. An SI staff mem­ber chooses a pa­per idea we need writ­ten up, then writes an ab­stract and some notes on the de­sired fi­nal con­tent.

  2. SI pays Gw­ern or an­other re­mote re­searcher to do a liter­a­ture search-and-sum­mary of rele­vant ma­te­rial, with poin­t­ers to other re­sources.

  3. SI posts a con­test to LessWrong, invit­ing sub­mis­sions of near-con­fer­ence-level-qual­ity ar­ti­cles that fol­low the pro­vided ab­stract and notes on de­sired fi­nal con­tent. Con­tes­tants benefit by start­ing with the re­sults of Gw­ern’s liter­a­ture sum­mary, and by know­ing that they don’t need to pro­duce some­thing as good as “In­tel­li­gence Ex­plo­sion: Ev­i­dence and Im­port” to win the prize. First place wins $1200, 2nd place wins $500, and 3rd place wins $200.

  4. Sub­mis­sions are due 1 month later. Sub­mis­sion are re­viewed, and the au­thors of the best sub­mis­sions are sent com­ments on what could be im­proved to max­i­mize the chances of com­ing in first place.

  5. Re­vised ar­ti­cles are due 3 weeks af­ter com­ments are re­ceived. Prizes are awarded.

  6. SI pays an ex­pe­rienced writer like Yvain or Kaj_So­tala or some­one similar to build up and im­prove the 1st place sub­mis­sion, bor­row­ing the best parts from the other sub­mis­sions, too.

  7. An SI staff mem­ber does a fi­nal pass, adding some con­tent, mak­ing it more clearly or­ga­nized and pol­ished, etc. One of SI’s re­mote ed­i­tors does an­other pass to make the sen­tences more perfect.

  8. The pa­per is sub­mit­ted to a jour­nal or an ed­ited vol­ume, and is marked as be­ing co-au­thored by (1) the key SI staff mem­ber who pro­vided the seed ideas and guided each stage of the re­vi­sions and pol­ish­ing, (2) the au­thor of the win­ning sub­mis­sion, and (3) Gw­ern. (With thanks to con­tri­bu­tions from the other con­test par­ti­ci­pants whose sub­mis­sions were bor­rowed from — un­less huge pieces were bor­rowed, in which case they may be counted as an ad­di­tional co-au­thor.)

If this method works, each pa­per may re­quire only 50-150 hours of SI staff time per pa­per — a dra­matic im­prove­ment! But this method has ad­di­tional benefits:

  • Mem­bers of the com­mu­nity who are ca­pa­ble of do­ing one piece of the pro­cess but not the other pieces get to con­tribute where they shine. (Many peo­ple can write okay-level ar­ti­cles but can’t do effi­cient liter­a­ture searches or pro­duce pol­ished prose, etc.)

  • SI gets to learn more about the tal­ent that ex­ists in its com­mu­nity which hadn’t yet been given the op­por­tu­nity to flower. (We might be able to di­rectly out­source fu­ture work to con­test par­ti­ci­pants, and if one per­son wins three such con­tests, that’s an in­di­ca­tor that we should con­sider hiring them.)

  • Ad­di­tional paid “jobs” (by way of con­test money) are cre­ated for LW ra­tio­nal­ists who have some do­main ex­per­tise in sin­gu­lar­ity-re­lated sub­jects.

  • Many Less Wrongers are stu­dents in fields rele­vant to the sub­ject mat­ter of the pa­pers that will be pro­duced by this pro­cess, and this will give them an op­por­tu­nity to co-au­thor pa­pers that can go on their CV.

  • The com­mu­nity in gen­eral gets bet­ter at col­lab­o­rat­ing.

This is, af­ter all, more similar to how many pa­pers would be pro­duced by uni­ver­sity de­part­ments, in which a se­nior re­searcher works with a team of stu­dents to pro­duce pa­pers.

Feed­back? In­ter­est?

(Not ex­actly the same, but see also the Poly­math Pro­ject.)