Owen_Cotton-Barratt

• Owen_Cotton-Barratt 21 Dec 2017 19:52 UTC

  7 points

  in reply to: jpaddison's comment on: “Just take the ex­pected value” – a pos­si­ble re­ply to con­cerns about cluelessness

  By now it should be clear that sim­ply fol­low­ing the ex­pected value is not a suffi­cient re­sponse to con­cerns of clue­less­ness.

  I was pretty sur­prised by this sen­tence. Maybe you could say more pre­cisely what you mean?

  I take the core con­cern of clue­less­ness to be that per­haps we have no in­for­ma­tion about what op­tions are best. Ex­pected value gives a the­o­ret­i­cal out to that (with some un­re­solved is­sues around in­finite ex­pec­ta­tions for ac­tors with un­bounded util­ity func­tions). Ap­prox­i­ma­tions to ex­pected value that hu­mans can im­ple­ment are as you point out kind of messy and opaque, but that’s a fea­ture of hu­man rea­son­ing in gen­eral, and doesn’t seem par­tic­u­larly tied to ex­pected value. Is that what you’re point­ing at?

• Owen_Cotton-Barratt 18 Dec 2017 14:15 UTC

  1 point

  in reply to: Owen_Cotton-Barratt's comment on: An­nounc­ing the 2017 donor lottery

  I don’t quite have an al­gorithm in mind for this. I think in prac­tice it would likely be easy to find solu­tions to di­vid­ing tick­ets, but per­haps one would want some­thing more speci­fied first.

  With a well-speci­fied al­gorithm and an un­der­stand­ing that it was well-be­haved, one could imag­ine shrink­ing the block size right down to give peo­ple flex­i­bil­ity over their lot­tery size and re­duce the li­a­bil­ity of the guaran­tor. There is per­haps an ad­van­tage to hav­ing a canon­i­cal size for de­vel­op­ing buy-in to the idea, though.

• Owen_Cotton-Barratt 18 Dec 2017 14:11 UTC

  0 points

  in reply to: Jess_Riedel's comment on: An­nounc­ing the 2017 donor lottery

  A sim­ple vari­a­tion on the cur­rent sys­tem would al­low peo­ple to opt into lot­tery-ing up fur­ther (to the scale of the to­tal donor lot­tery pot):

  Ask peo­ple what scale they would like to lot­tery to. If $100k, al­lo­cate them a range of tick­ets in one block as in the cur­rent sys­tem. If (say) $300k, split their tick­ets be­tween three blocks, giv­ing them the same range in each block. If their preferred scale ex­ceeds the to­tal pot, just give them cor­re­lated tick­ets on all available blocks.

  If there’s a con­flict of prefer­ence be­tween peo­ple want­ing small and large lot­ter­ies so they’re not si­mul­ta­neously satis­fi­able (I think this is some­what un­likely in prac­tice un­less some­one comes in with $90k hop­ing to lot­tery up to $100k), first satisfy those who want smaller to­tals, then di­vide the rest as fairly as pos­si­ble.

• Owen_Cotton-Barratt 16 Dec 2017 22:28 UTC

  6 points

  in reply to: SamDeere's comment on: An­nounc­ing the 2017 donor lottery

  We think that it’s in the spirit of the lot­tery that some­one who does use­ful re­search that would be of in­ter­est to other donors should pub­lish it (or give per­mis­sion for CEA to pub­lish their grant recom­men­da­tion). Also, if they con­vince oth­ers to donate then they’ll be caus­ing ad­di­tional grants to go to their preferred or­ga­ni­za­tion(s). We’ll strongly en­courage win­ners to do so, how­ever, in the in­ter­ests of keep­ing the bar­ri­ers to en­try low, we haven’t made it a hard re­quire­ment.

  Seems like even strong so­cial pres­sure might be enough to be a sig­nifi­cant bar­rier to en­try. I feel ex­cited about en­ter­ing a donor lot­tery, and would feel less ex­cited if I thought I’d feel ac­countable if I won (I might still en­ter, but it seems like a sig­nifi­cant cost).

  Would an at­ti­tude of “we think it’s great if you want to share (and we could help you with com­mu­ni­ca­tion) but there’s no so­cial obli­ga­tion” cap­ture the benefits? That’s pretty close to what you were say­ing already, but the differ­ent tone might be helpful for some peo­ple.

• Owen_Cotton-Barratt 23 Nov 2017 23:15 UTC

  2 points

  in reply to: jpaddison's comment on: Causal Net­work Model III: Findings

  Thanks for the write-up!

  I found the figures for ex­is­ten­tial-risk-re­duced-per-$ with your de­fault val­ues a bit sus­pi­ciously high. I won­der if the rea­son for this is in end­note [2], where you say:

  say one re­searcher year costs $50,000

  I think this is too low as the figure to use in this calcu­la­tion, per­haps by around an or­der of mag­ni­tude.

  Firstly, that is a very cheap re­searcher-year even just pay­ing costs. Many re­searcher salaries are straight-up higher, and costs should in­clude over­heads.

  A sec­ond fac­tor is that hav­ing twice as much money doesn’t come close to buy­ing you twice as much (qual­ity-ad­justed) re­search. In gen­eral it is hard to sim­ply pay money to pro­duce more of some of these spe­cial­ised forms of labour. For in­stance see the re­cent 80k sur­vey of will­ing­ness to pay of EA orgs to bring for­ward re­cent hires, where the av­er­age will­ing­ness to forgo dona­tions to move a se­nior hire for­ward by three years was around $4 mil­lion.

• Owen_Cotton-Barratt 24 Jun 2017 19:53 UTC

  1 point

  in reply to: Elizabeth's comment on: Earn­ing to Give as Costly Signalling

  Seems like it’s sug­gest­ing it as costly sig­nal­ling at the level of the move­ment rather than the in­di­vi­d­u­als. It’s a stretch from nor­mal use, but that’s kind of the strength of analo­gies?

• Owen_Cotton-Barratt 24 Jun 2017 14:29 UTC

  0 points

  in reply to: jpaddison's comment on: What is valuable about effec­tive al­tru­ism? Im­pli­ca­tions for com­mu­nity building

  This is a great ques­tion and I think de­serves fur­ther thought.

  Helping peo­ple con­sider their val­ues was one of the ma­jor goals Daniel Koko­ta­jlo and I had in de­sign­ing this flowchart. One pos­si­ble ac­tivity would be to read through and/​or dis­cuss parts of that.

What is valuable about effec­tive al­tru­ism? Im­pli­ca­tions for com­mu­nity building

• Owen_Cotton-Barratt 13 Jun 2017 11:18 UTC

  9 points

  in reply to: jpaddison's comment on: Pro­jects I’d like to see

  I was a bit con­fused by some of these. Post­ing ques­tions/​com­ments here in case oth­ers have the same thoughts:

  Earn­ing-to-give buy-out

  You’re cur­rently earn­ing to give, be­cause you think that your dona­tions are do­ing more good than your di­rect work would. It might be that we think that it would be more valuable if you did di­rect work. If so we could donate a pro­por­tion of the amount that you were donat­ing to wher­ever you were donat­ing it, and you would move into work.

  This made more sense to me af­ter I re­al­ised that we should prob­a­bly as­sume the per­son doesn’t think CEA is a top dona­tion tar­get. Other­wise they would have an em­piri­cal dis­agree­ment about whether they should be do­ing di­rect work, and it’s not clear how the offer helps re­solve that (though it’s ob­vi­ously worth dis­cussing).

  Anti-De­bates /​ Shark Tank-style ca­reer choice dis­cus­sions /​ Re­search work­ing groups

  Th­ese are all things that might be good, but it’s not ob­vi­ous how fund­ing would be a bot­tle­neck. Might be worth say­ing some­thing about that?

  For those with a quan­ti­ta­tive PhD, it could in­volve ap­ply­ing for the Google Brain Res­i­dency pro­gram or AI safety fel­low­ship at ASI.

  Similarly I’m con­fused what the fund­ing is meant to do in these cases.

  I’d be keen to see more peo­ple take ideas that we think we already know, but haven’t ever been put down in writ­ing, and write them up in a thor­ough and even-handed way; for ex­am­ple, why ex­is­ten­tial risk from an­thro­pogenic causes is greater than the ex­is­ten­tial risk from nat­u­ral causes

  I think you were us­ing this as an ex­am­ple of the type of work, rather than a spe­cific re­quest, but some read­ers might not know that there’s a pa­per forth­com­ing on pre­cisely this topic (if you mean some­thing differ­ent from that pa­per, I’m in­ter­ested to know what!).

• Owen_Cotton-Barratt 1 Jun 2017 22:21 UTC

  7 points

  in reply to: DavidNash's comment on: Con­sid­er­ing Con­sid­er­ate­ness: Why com­mu­ni­ties of do-good­ers should be ex­cep­tion­ally considerate

  Not as pithy, but just a flag that I think the ques­tion im­plic­itly raised by Tom’s com­ment and the an­swer in David’s are pretty im­por­tant. This is a com­mu­nity which is will­ing to up­date ac­tions based on the­o­ret­i­cal ar­gu­ments about what’s im­por­tant. Of course I don’t ex­pect an ar­ti­cle to to­tally change peo­ple’s be­liefs—let alone be­havi­ours—but if it has a frac­tion of that effect I’d count it as cheap.

• Owen_Cotton-Barratt 1 Jun 2017 9:18 UTC

  4 points

  in reply to: MichaelPlant's comment on: Con­sid­er­ing Con­sid­er­ate­ness: Why com­mu­ni­ties of do-good­ers should be ex­cep­tion­ally considerate

  I think you’re right that there’s a failure mode of not ask­ing peo­ple for things. I don’t think that not-ask­ing is in gen­eral the more con­sid­er­ate ac­tion, though—of­ten peo­ple would pre­fer to be given the op­por­tu­nity to help (par­tic­u­larly if it feels like an op­por­tu­nity rather than a de­mand).

  I sup­pose the gen­eral point is: avoid the trap of overly-nar­row in­ter­pre­ta­tions of con­sid­er­ate­ness (just like it was good to avoid the trap of overly-nar­row in­ter­pre­ta­tions of con­se­quences of ac­tions).

• Owen_Cotton-Barratt 25 May 2017 10:24 UTC

  1 point

  in reply to: adom_hartell's comment on: Re­turns Func­tions and Fund­ing Gaps

  Fair ques­tion. This ar­gu­ment is all con­di­tioned on A not ac­tu­ally hav­ing good ways to ex­pand ca­pac­ity—the case is that even then the funds are com­pa­rably good given to A as el­se­where. The pos­si­bil­ity of A in fact hav­ing use­ful ex­pan­sion might make it no­tice­ably bet­ter than the al­ter­na­tive, which is what (to my mind) drives the asym­me­try.

• Owen_Cotton-Barratt 16 May 2017 10:37 UTC

  2 points

  in reply to: Ben Pace's comment on: A model of StrongMinds—Oxford Pri­ori­ti­sa­tion Project

  DALYs do use a more defen­si­ble anal­y­sis; GiveWell aren’t us­ing DALYs. This has some good and some bad as­pects (re­lated to the dis­cus­sion in this post, al­though in this case the down­side of defen­si­bil­ity is more that it doesn’t let you in­cor­po­rate con­sid­er­a­tions that aren’t fully grounded).

  The prob­lem with just us­ing DALYs is that on many views they over­weigh in­fant mor­tal­ity (here’s my view on some of the is­sues, but the po­si­tion that they over­weigh in­fant mor­tal­ity is far from origi­nal). With an in­ter­nal agree­ment that they sig­nifi­cantly over­weigh in­fant mor­tal­ity, it be­comes un­ten­able to just con­tinue us­ing DALYs, even ab­sent a fully rigor­ous al­ter­na­tive. Hence fal­ling back on more ad hoc but some­what ro­bust meth­ods like ask­ing peo­ple to con­sider it and us­ing a me­dian.

  [I’m just in­ter­pret­ing GW de­ci­sion-mak­ing from pub­li­cly available in­for­ma­tion; this might eas­ily turn out to be a mis­rep­re­sen­ta­tion.]

• Owen_Cotton-Barratt 8 May 2017 13:42 UTC

  2 points

  in reply to: RyanCarey's comment on: How much does work in AI safety help the world? Prob­a­bil­ity dis­tri­bu­tion ver­sion (Oxford Pri­ori­ti­sa­tion Pro­ject)

  Yes, I think this is a sig­nifi­cant con­cern with this ver­sion of the model (some­what less so with the origi­nal cruder ver­sion us­ing some­thing like me­di­ans, but that ver­sion also fails to pick up on le­gi­t­i­mate effects of “what if these vari­ables are all in the tails”). Com­bin­ing the vari­ables as you sug­gest is the eas­iest way to patch it. More com­plex would be to add in ex­plicit time-de­pen­dency.

• Owen_Cotton-Barratt 26 Apr 2017 11:45 UTC

  1 point

  in reply to: RyanCarey's comment on: Char­ity Eval­u­a­tors: first model and open ques­tions (Oxford Pri­ori­ti­sa­tion Pro­ject)

  I largely agree with these con­sid­er­a­tions about the dis­tri­bu­tion of net im­pact of in­ter­ven­tions (al­though with some pos­si­ble dis­agree­ments, e.g. I think nega­tive fung­ing is also pos­si­ble).

  How­ever, I ac­tu­ally wasn’t try­ing to com­ment on this at all! I was talk­ing about the dis­tri­bu­tion of peo­ple’s es­ti­mates of im­pact around the true im­pact for a given in­ter­ven­tion. Sorry for not be­ing clearer :/​

• Owen_Cotton-Barratt 25 Apr 2017 14:59 UTC

  4 points

  in reply to: jpaddison's comment on: Char­ity Eval­u­a­tors: first model and open ques­tions (Oxford Pri­ori­ti­sa­tion Pro­ject)

  The fact that some­times peo­ple’s es­ti­mates of im­pact are sub­se­quently re­vised down by sev­eral or­ders of mag­ni­tude seems like strong ev­i­dence against ev­i­dence be­ing nor­mally dis­tributed around the truth. I ex­pect that if any­thing it is broader than log­nor­mally dis­tributed. I also think that ex­tra pieces of ev­i­dence are likely to be some­what cor­re­lated in their er­ror, al­though it’s not ob­vi­ous how best to model that.

• Owen_Cotton-Barratt 25 Apr 2017 9:03 UTC

  3 points

  in reply to: kbog's comment on: Up­date on Effec­tive Altru­ism Funds

  I find your com­ments painfully un­char­i­ta­ble, which re­ally re­duces my in­cli­na­tion to en­gage. If you can’t find an in­ter­pre­ta­tion of my com­ment which isn’t just about the op­ti­mizer’s curse I don’t feel like helping you right now.

  Agree that ve­toes aren’t the right solu­tion, though (in­deed they are them­selves sub­ject to a unilat­er­al­ist’s curse, per­haps of a worse type).

• Owen_Cotton-Barratt 24 Apr 2017 16:51 UTC

  4 points

  in reply to: kbog's comment on: Up­date on Effec­tive Altru­ism Funds

  The kind of set-up where it would ap­ply:

  • An easy-to-eval­u­ate op­por­tu­nity which pro­duces 1 util/​$, which ev­ery­one cor­rectly evaluates

  • 100 hard-to-eval­u­ate op­por­tu­ni­ties each of which ac­tu­ally pro­duces 0.1 util /​ $, but where ev­ery­one has an in­de­pen­dent es­ti­mate of cost-effec­tive­ness which is a log-nor­mal cen­tered on the truth

  Then for any in­di­vi­d­ual the’re likely to think one of the 100 is best and donate there. If they all pooled their info they would in­stead all donate to the first op­por­tu­nity.

  Ob­vi­ously the num­bers and func­tional form here are im­plau­si­ble—I chose them for leg­i­bil­ity of the ex­am­ple. It’s a le­gi­t­i­mate ques­tion how strongly the dy­namic ap­plies in prac­tice. But it seems fairly clear to me that it can ap­ply. You sug­gested there’s a sym­me­try with donat­ing too lit­tle—I think this is bro­ken be­cause peo­ple are se­lect­ing the top op­tion, so they are in­di­vi­d­u­ally run­ning into the op­ti­mizer’s curse.

• Owen_Cotton-Barratt 23 Apr 2017 9:10 UTC

  8 points

  in reply to: RyanCarey's comment on: Up­date on Effec­tive Altru­ism Funds

  Ryan, I sub­stan­tially dis­agree and ac­tu­ally think all of your sug­gested al­ter­na­tives are worse. The origi­nal is re­port­ing on a re­sponse to the writ­ing, not stak­ing out a claim to an ob­jec­tive as­sess­ment of it.

  I think that re­port­ing hon­est re­sponses is one of the best tools we have for deal­ing with emo­tional in­fer­en­tial gaps—par­tic­u­larly if it’s made ex­plicit that this is a func­tion of the reader and writ­ing, and not the writ­ing alone.

• Owen_Cotton-Barratt 23 Apr 2017 8:54 UTC

  2 points

  in reply to: kbog's comment on: Up­date on Effec­tive Altru­ism Funds

  The ba­sic dy­namic ap­plies. Think it’s pretty rea­son­able to use the name to point loosely in such cases, even if the origi­nal pa­per didn’t dis­cuss this ex­ten­sion.