Larks

• Larks 14 Mar 2019 21:40 UTC

  2 points

  on: How to Un­der­stand and Miti­gate Risk (Cross­post from LessWrong)

  It’s quite easy to re­search the cost of cre­at­ing a rice farm, or a power plant, as well as get a tight bounded prob­a­bil­ity dis­tri­bu­tion for the ex­pected price you can sell your rice or elec­tric­ity at af­ter mak­ing the ini­tial in­vest­ment. Th­ese mar­kets are very ma­ture and there’s un­likely to be wild swings or un­ex­pected in­no­va­tions that sig­nifi­cantly change the mar­ket.

  This doesn’t af­fect your over­all ar­ti­cle much, but it’s worth not­ing that com­mod­ity prices can be very volatile. Look­ing up the generic rice con­tract on Bloomberg for ex­am­ple, and pick­ing the more ex­treme years but the same month (to avoid sea­son­al­ity):

  • 1998 April: 10.2

  • 2002 April: 3.6

  • 2004 April: 11.3

  • 2005 April 7.2

  • 2008 April: 23.8

  • 2010 April: 12.6

  • 2013 April: 15.8

  • 2015 April: 10.0

  You do have the abil­ity to lock in the cur­rent im­plied prof­ita­bil­ity us­ing fu­tures, but in gen­eral com­mod­ity mar­kets seem to be more volatile than non-com­mod­ity mar­kets.

• Larks 14 Mar 2019 2:27 UTC

  4 points

  on: Brian Tse: Risks from Great Power Conflicts

  I think one pa­per shows that there were al­most 40 near misses, and I think that was put up by the Fu­ture of Life In­sti­tute, so some peo­ple can look up that pa­per, and I think that in gen­eral it seems that ex­perts agree some of the biggest risks from nu­clear would be ac­ci­den­tal use, rather than de­liber­ate and mal­i­cious use be­tween coun­tries.

  Pos­si­bly you are think­ing of the Global Catas­trophic Risks In­sti­tute, and Baum et al.’s A Model for the Prob­a­bil­ity of Nu­clear War ?

• Larks 10 Mar 2019 17:15 UTC

  8 points

  on: EA/​X-risk/​AI Align­ment Cover­age in Journalism

  Thanks for high­light­ing this, I thought it was in­ter­est­ing. It does seem that, if you thought get­ting Vox to write about AI was good, it would be good to have an offset­ting right-wing spokesman on the is­sue.

  One re­lated point would be that we can try to avoid ex­ces­sively as­so­ci­at­ing AI risk with left wing causes; dis­crim­i­na­tion is the ob­vi­ous one. The al­ter­na­tive would be to try to come up with right-wing causes to as­so­ci­ate it with as well; I have one idea, but I think this strat­egy may be a bad idea so am loath to share it.

• Larks 8 Mar 2019 3:01 UTC

  49 points

  on: SHIC Will Sus­pend Outreach Oper­a­tions

  This was very in­ter­est­ing. Ret­ro­spec­tives on pro­jects that didn’t work can be ex­tremely helpful to oth­ers, but I imag­ine can also been tough to write, so thanks very much!

• Larks 5 Mar 2019 4:15 UTC

  25 points

  in reply to: aarongertler's comment on: Mak­ing dis­cus­sions in EA groups inclusive

  It takes a long time to craft a re­sponse to posts like these. Even if there are clear prob­lems with the post, given the sen­si­tive topic you have to spend a lot of time on nu­ance, check­ing cita­tions, and get­ting the tone right. That is a very high bar, one that I don’t think is rea­son­able to ex­pect ev­ery­one to pass. In con­trast, peo­ple who agree seem to get a pass for silently up­vot­ing.

• Larks 5 Mar 2019 2:00 UTC

  39 points

  on: Mak­ing dis­cus­sions in EA groups inclusive

  While I ap­pre­ci­ate your say­ing you don’t in­tend to ban top­ics, I think there is con­sid­er­able risk that this sort of policy be­comes a form of de facto cen­sor­ship. In the same way that we should be wary of Iso­lated De­mands for Ri­gour, so too we should also be wary of Iso­lated De­mands for Sen­si­tivity.

  Take for ex­am­ple the first item on your list—lets call it A).

  Whether it is or has been right or nec­es­sary that women have less in­fluence over in­tel­lec­tual de­bate and less eco­nomic and poli­ti­cal power

  I agree that this is not a great topic for an EA dis­cus­sion. I haven’t seen any ar­gu­ments about the cost-effec­tive­ness of a cause area that rely on whether A) is true or false. It seems un­likely that speci­fi­cally fem­i­nist or anti-fem­i­nist causes would be the best things to work on, even if you thought A) was very true or false. If such a topic was very dis­tract­ing, I can even see it mak­ing sense to es­sen­tially ban dis­cus­sion of it, as LessWrong used to do in prac­tice with re­gard Poli­tics.

  My con­cern is that a rule/​recom­men­da­tion against dis­cussing such a topic might in prac­tice be ap­plied very un­equally. For ex­am­ple, I think that some­one who says

  As you know, women have long suffered from dis­crim­i­na­tion, re­sult­ing a lack of poli­ti­cal power, and their con­tri­bu­tions be­ing over­looked. This is un­just, and the effects are still felt to­day.

  would not be chas­tised for do­ing so, or feel that they had vi­o­lated the rule/​sug­ges­tion.

  How­ever, my guess is that some­one who said

  As you know, the de­gree of dis­crim­i­na­tion against women has been greatly ex­ag­ger­ated, and in many ar­eas, like con­scrip­tion or homi­cide risk, they ac­tu­ally en­joy ma­jor ad­van­tages over men.

  might be crit­i­cized for do­ing so, and might even agree (if only pri­vately) that they had in some sense vi­o­lated this rule/​guideline with re­gard topic A).

  If this is the case, then this policy is de facto a silenc­ing not of top­ics, but of opinions, which I think is much harder to jus­tify.

  As a list of ver­boten opinions, this list also has the un­de­sir­able at­tribute of be­ing very par­ti­san. Look­ing down the list, it seems that in al­most ev­ery case the dis­cour­aged/​for­bid­den opinion is, in con­tem­po­rary US poli­ti­cal par­lance, the (more) Right Wing opinion, and the as­sumed ‘de­fault’ ‘ac­cept­able’ one is the (more) Left Wing opinion. In ad­di­tion, my im­pres­sion (though I am less sure here) is that it is also bi­ased against opinions dis­pro­por­tionately held by older peo­ple.

  And yet these are two groups that are dra­mat­i­cally un­der-rep­re­sented in the EA move­ment! (source) Cer­tainly it seems that, on a nu­mer­i­cal ba­sis, con­ser­va­tives are more un­der-rep­re­sented than some of the pro­tected groups men­tioned in this ar­ti­cle. This sort of list seems likely to make older and more con­ser­va­tive peo­ple feel less wel­come, not more. Var­i­ous view­points they might ob­ject to have been en­shrined, and other top­ics, whose dis­cus­sion con­ser­va­tives find dis­aste­ful but is nonethe­less not un­com­mon in the EA com­mu­nity, are not con­traindi­cated.

  For a gen­er­ally well-re­ceived ar­ti­cle on how to par­tially ad­dress this, you might en­joy Ozy’s piece here.

• Larks 5 Mar 2019 0:01 UTC

  13 points

  in reply to: Larks's comment on: Re­search on Effec­tive Strate­gies for Equity and In­clu­sion in Move­ment-Building

  Here is a re­cent study on the topic that I think is very rele­vant:

  Gen­der, Race, and En­trepreneur­ship: A Ran­dom­ized Field Ex­per­i­ment on Ven­ture Cap­i­tal­ists and An­gels (Gor­nall and Stre­bu­laev)

  We sent out 80,000 pitch emails in­tro­duc­ing promis­ing but fic­ti­tious start-ups to 28,000 ven­ture cap­i­tal­ists and busi­ness an­gels. Each email was sent by a fic­ti­tious en­trepreneur with a ran­domly se­lected gen­der (male or fe­male) and race (Asian or White). Fe­male en­trepreneurs re­ceived an 8% higher rate of in­ter­ested replies than male en­trepreneurs pitch­ing iden­ti­cal pro­jects. Asian en­trepreneurs re­ceived a 6% higher rate than White en­trepreneurs. Our re­sults are not con­sis­tent with dis­crim­i­na­tion against fe­males or Asi­ans at the ini­tial con­tact stage of the in­vest­ment pro­cess.

  link

  How­ever, it does seem pretty ap­pli­ca­ble to EA. The EA com­mu­nity is in many ways similar to the VC com­mu­nity:

  • Similar ge­ogra­phies: the Bay Area, Lon­don, New York etc.

  • Similar ed­u­ca­tion back­grounds.

  • Both in­volve eval­u­at­ing spec­u­la­tive pro­jects with a lot of un­cer­tainty.

  Similarly to the stud­ies dis­cussed above, this finds that peo­ple are bi­ased against white men.

  (I have some qualms about this type of study, be­cause they in­volve wast­ing peo­ple’s time with­out their con­sent, but this doesn’t af­fect the con­clu­sions.)

• Larks 26 Feb 2019 12:55 UTC

  42 points

  on: After one year of ap­ply­ing for EA jobs: It is re­ally, re­ally hard to get hired by an EA organisation

  Great post. I’m sure writ­ing this must have been tough, so thanks very much for shar­ing this.

• Larks 26 Feb 2019 4:19 UTC

  4 points

  on: Im­pact Prizes as an al­ter­na­tive to Cer­tifi­cates of Impact

  Great post; I had been think­ing about writ­ing some­thing very similar. In many ways I think you have ac­tu­ally un­der­stated the po­ten­tial of the idea. Ad­di­tion­ally I think it ad­dresses some of the con­cerns Owen raised last time.

  Eval­u­a­tion Costs

  The fi­nal prize eval­u­a­tions could be quite costly to pro­duce.

  I ac­tu­ally think the fi­nal eval­u­a­tions might be cheaper than the sta­tus quo. At the mo­ment OpenPhil (or who­ever) has to do two things:

  1) Judge how good an out­come is.

  2) Judge how likely differ­ent out­comes are.

  With this plan, 2) has been (par­tially) out­sourced to the mar­ket, leav­ing them with just 1).

  Cul­tural Risks

  If Im­pact Prizes took off, I could imag­ine some ac­tors draw­ing into the ecosys­tem who only mo­ti­vated by mak­ing prof­its.

  This is not a bug, this is a fea­ture! There is a very large pool of peo­ple will­ing to pre­dict ar­bi­trary out­comes in re­turn for money, that we have thus far only very in­di­rectly been tap­ping into. In gen­eral bring­ing in more traders im­proves the effi­ciency of a mar­ket. Even if you add noisy traders, their pres­ence im­proves the in­cen­tives for ‘smart money’ to par­ti­ci­pate. I think it’s un­likely we’d reach the scale re­quired for ac­tual hedge funds to get in­volved, but I do think it’s plau­si­ble we could get a lot of hedge fund guys par­ti­ci­pat­ing in their spare time.

  Le­gal Implications

  In terms of le­gal sta­tus, one op­tion I’ve been think­ing about would be copy­ing Pre­dic­tIt. If we have to pay taxes ev­ery time a cer­tifi­cate is trans­ferred, the trans­ac­tion costs will be pro­hibitive. I am quite wor­ried it will be hard to make this work within US law un­for­tu­nately, which is not very friendly to this sort of ex­per­i­men­ta­tion. At the same time, given the SEC’s at­ti­tude to­wards non-com­pli­ant se­cu­rity is­suance, I would not want to op­er­ate out­side it!

  Quick other thoughts

  One is­sue with the idea is it is hard for OpenPhil to add more promised fund­ing later, be­cause the ini­tial in­vest­ment will already have been com­mit­ted at some fixed level. e.g. If OpenPhil ini­tially promise $10m, and then later bump it to $20m, pro­jects that have already sold their to­kens can­not ex­pand to take ad­van­tage of this in­crease, so it is effec­tively pure wind­fall with no in­cen­tive effect. A pos­si­ble solu­tion would be co­horts; we promise $10m in 2022 for pro­jects started in 2019, and then later add an­other $12m, paid in 2023, for 2020 pro­jects.

• Larks 26 Feb 2019 4:15 UTC

  10 points

  in reply to: Milan_Griffes's comment on: Im­pact Prizes as an al­ter­na­tive to Cer­tifi­cates of Impact

  I think I might have been the sec­ond largest pur­chaser of the cer­tifi­cates. My ex­pe­rience was that we didn’t at­tract the re­ally high qual­ity pro­jects I’d want, and those we did see had very high reser­va­tion prices from the sel­l­ers, per­haps due to the en­dow­ment effect. I sus­pect sel­l­ers might say that they didn’t see enough buy­ers. Pos­si­bly we just had a chicken-and-egg prob­lem, com­bined with ev­ery­one in­volved be­ing kind of busy.

• Larks 16 Feb 2019 0:18 UTC

  5 points

  on: Three Bi­ases That Made Me Believe in AI Risk

  Could you go into a bit more de­tail about the two lin­guis­tic styles you de­scribed, per­haps us­ing non-AI ex­am­ples? My in­ter­pre­ta­tion of them is ba­si­cally agent-fo­cused vs in­ter­nal-me­chan­ics-fo­cused, but I’m not sure this is ex­actly what you mean.

  If the above is cor­rect, it seems like you’re ba­si­cally say­ing that in­ter­nal-me­chan­ics-fo­cused de­scrip­tions work bet­ter for cur­rently ex­ist­ing AI sys­tems, which seems true to me for things like self-driv­ing cars. But for some­thing like AlphaZero, or Stock­fish, I think an agen­tic fram­ing is of­ten ac­tu­ally quite use­ful:

  A chess/​Go AI is easy to imag­ine: they are smart and au­tonomous and you can trust the bot like you trust a hu­man player. They can make mis­takes but prob­a­bly have good in­tent. When they en­counter an un­fa­mil­iar game situ­a­tion they can think about the cor­rect way to pro­ceed. They be­have in con­cor­dance with the goal (win­ning the game) their cre­ator set them and they tend to make smart de­ci­sions. If any­thing goes wrong then the car is at fault.

  So I think the rea­son this type of lan­guage doesn’t work well for self-driv­ing cars is be­cause they aren’t suffi­ciently agent-like. But we know there can be agen­tic agents—hu­mans are an ex­am­ple—so it seems plau­si­ble to me that agen­tic lan­guage will be the best de­scrip­tor for them. Cer­tainly it is cur­rently the best de­scrip­tor for them, given that we do not un­der­stand the in­ter­nal me­chan­ics of as-yet-un­in­vented AIs.

• Larks 1 Feb 2019 3:22 UTC

  27 points

  on: Re­search on Effec­tive Strate­gies for Equity and In­clu­sion in Move­ment-Building

  In gen­eral it’s prob­a­bly best not to anonymize ap­pli­ca­tions. Field stud­ies gen­er­ally show no effect on in­ter­view se­lec­tion, and some­times even show a nega­tive effect (which has also been seen in the lab). Blind­ing may work for mu­si­ci­ans, ran­domly gen­er­ated re­sumes, and iden­ti­cal ex­pres­sions of in­ter­est, but in re­al­ity there seem to be sub­tle cues of an ap­pli­cant’s back­ground that eval­u­a­tors may pick up on, and the risk of anonymiza­tion back­firing is higher for re­cruit­ing groups which are ac­tively in­ter­ested in DEI. This may be be­cause they are un­able to proac­tively check their bi­ases when blind, or to proac­tively ac­com­mo­date dis­ad­van­taged can­di­dates at this re­cruit­ment stage, or be­cause their staff is already more di­verse and peo­ple may fa­vor can­di­dates they iden­tify with de­mo­graph­i­cally.

  I think you are mis-de­scribing these stud­ies. Essen­tially, they found that when re­view­ers knew the race and sex of the ap­pli­cants, they were bi­ased in favour of women and non-whites, and against white males.

  I ad­mit I only read two of the stud­ies you linked two, but I think these quotes from them are quite clear that about the con­clu­sions:

  We find that par­ti­ci­pat­ing firms be­come less likely to in­ter­view and hire minor­ity can­di­dates when re­ceiv­ing anony­mous re­sumes.

  The pub­lic ser­vants re­view­ing the job ap­pli­cants en­gaged in dis­crim­i­na­tion that favoured fe­male ap­pli­cants and dis­ad­van­taged male candidates

  Affir­ma­tive ac­tion to­wards the Indige­nous fe­male can­di­date is the largest, be­ing 22.2% more likely to be short listed on av­er­age when iden­ti­fied com­pared to the de-iden­ti­fied con­di­tion. On the other hand, the iden­ti­fied Indige­nous male CV is 9.4% more likely to be short­listed on av­er­age com­pared to when it is de-iden­ti­fied. In ab­solute terms most minor­ity can­di­dates are on av­er­age more likely to be short­listed when named com­pared to the de-iden­ti­fied con­di­tion, but the differ­ence for the Indige­nous fe­male can­di­date is the only one that is statis­ti­cally sig­nifi­cant at the 95% con­fi­dence level.

  This is also sup­ported by other pa­pers on the sub­ject. For ex­am­ple, you might en­joy read­ing Willi­ams and Ceci (2015):

  The un­der­rep­re­sen­ta­tion of women in aca­demic sci­ence is typ­i­cally at­tributed, both in sci­en­tific liter­a­ture and in the me­dia, to sex­ist hiring. Here we re­port five hiring ex­per­i­ments in which fac­ulty eval­u­ated hy­po­thet­i­cal fe­male and male ap­pli­cants, us­ing sys­tem­at­i­cally varied pro­files dis­guis­ing iden­ti­cal schol­ar­ship, for as­sis­tant pro­fes­sor­ships in biol­ogy, en­g­ineer­ing, eco­nomics, and psy­chol­ogy. Con­trary to pre­vailing as­sump­tions, men and women fac­ulty mem­bers from all four fields preferred fe­male ap­pli­cants 2:1 over iden­ti­cally qual­ified males with match­ing lifestyles (sin­gle, mar­ried, di­vorced), with the ex­cep­tion of male economists, who showed no gen­der prefer­ence. Com­par­ing differ­ent lifestyles re­vealed that women preferred di­vorced moth­ers to mar­ried fathers and that men preferred moth­ers who took parental leaves to moth­ers who did not. Our find­ings, sup­ported by real-world aca­demic hiring data, sug­gest ad­van­tages for women launch­ing aca­demic sci­ence ca­reers.

  This doesn’t mean that anonymiz­ing ap­pli­ca­tions is a bad idea—it ap­pears to have suc­cess­fully re­duced un­fair bias—rather that the bias was in the op­po­site di­rec­tion than the au­thors ex­pected to find it.

• Larks 31 Jan 2019 2:11 UTC

  16 points

  on: EA Fo­rum Prize: Win­ners for De­cem­ber 2018

  Thanks very much! I think this prize is a great idea. I was definitely mo­ti­vated to in­vest more time and effort by the hope of win­ning the prize (along with the satis­fac­tion of get­ting front page with a lot of karma).

• Larks 28 Jan 2019 2:26 UTC

  14 points

  in reply to: Jeremy's comment on: Vox’s “Fu­ture Perfect” column fre­quently has flawed journalism

  I have definitely heard peo­ple refer­ring to Fu­ture Perfect as ‘the EA part of Vox’ or similar.

• Larks 17 Jan 2019 22:12 UTC

  6 points

  on: How can I in­ter­nal­ize my most im­pact­ful nega­tive ex­ter­nal­ities?

  You might en­joy this post Claire wrote: Eth­i­cal Offset­ting is An­ti­thet­i­cal to EA.

• Larks 14 Jan 2019 2:45 UTC

  1 point

  on: What Is Effec­tive Altru­ism?

  Thanks for writ­ing this, I thought it was quite a good sum­mary. How­ever, I would like to push back on two things.

  Effec­tive al­tru­ism is egal­i­tar­ian. Effec­tive al­tru­ism val­ues all peo­ple equally

  I of­ten think of age as be­ing one di­men­sion that egal­i­tar­i­ans think should not in­fluence how im­por­tant some­one is. How­ever, de­spite GiveWell be­ing one of the archety­pal EA or­gani­sa­tions (along with GWWC/​CEA), they do not do this. Rather, they value mid­dle-aged years of life more highly than baby years or life or old peo­ple years of life. See for ex­am­ple this page here. Per­haps EA should be egal­i­tar­ian, but de facto it does not seem to be.

  Effec­tive al­tru­ism is sec­u­lar. It does not recom­mend char­i­ties that most effec­tively get peo­ple into Heaven …

  This item seem rather differ­ent from the other items on the list. Most of the oth­ers seem like ra­tio­nal po­si­tions for vir­tu­ally any­one to hold. How­ever, if you were re­li­gious, this ten­nant seems very ir­ra­tional—helping peo­ple get into heaven would be the most effec­tive thing you could do! Put­ting this here seems akin to say­ing that AMF is an EA value; rather, these are con­clu­sions, not premises.

  Ad­di­tion­ally, there is some ev­i­dence that pro­mot­ing re­li­gion might be benefi­cial even on strictly ma­te­rial grounds. Have you seen the re­cent pre-reg­istered RCT on protes­tant evan­ge­lism?

  To test the causal im­pact of re­li­gios­ity, we con­ducted a ran­dom­ized eval­u­a­tion of an evan­gel­i­cal Protes­tant Chris­tian val­ues and the­ol­ogy ed­u­ca­tion pro­gram that con­sisted of 15 weekly half-hour ses­sions. We an­a­lyze out­comes for 6,276 ul­tra-poor Filipino house­holds six months af­ter the pro­gram ended. We find sig­nifi­cant in­creases in re­li­gios­ity and in­come, no sig­nifi­cant changes in to­tal la­bor sup­ply, as­sets, con­sump­tion, food se­cu­rity, or life satis­fac­tion, and a sig­nifi­cant de­crease in per­ceived rel­a­tive eco­nomic sta­tus. Ex­plo­ra­tory anal­y­sis sug­gests the pro­gram may have im­proved hy­gienic prac­tices and in­creased house­hold dis­cord, and that the in­come treat­ment effect may op­er­ate through in­creas­ing grit.

  https://​​www.nber.org/​​pa­pers/​​w24278.pdf

  I don’t have a strong view on whether or not this is ac­tu­ally a good thing to do, let alone the best thing. RCTs provide high qual­ity causal ev­i­dence, but even then most in­ter­ven­tions do not work very well, and I’m not an ex­pert on the im­pact of evan­ge­lism. But it seems strange to as­sume from very be­gin­ning that it is not some­thing EAs would ever be in­ter­ested in.

• Larks 6 Jan 2019 18:22 UTC

  11 points

  on: EA Giv­ing Tues­day Dona­tion Match­ing Ini­ti­a­tive 2018 Retrospective

  Con­grat­u­la­tions guys, this is re­ally im­pres­sive. Thanks for all the work you put into this.

• Larks 5 Jan 2019 18:21 UTC

  10 points

  in reply to: Ben_West's comment on: 2018 AI Align­ment Liter­a­ture Re­view and Char­ity Comparison

  My gen­eral model is that char­i­ties get fund­ing in two waves:

  1) December

  2) The rest of the year

  As such, if I ask groups for their run­way at the be­gin­ning of 1), and they say they have 12 months, that ba­si­cally means that even if they failed to raise any money at all in the fol­low­ing 1) and 2) they would still sur­vive un­til next De­cem­ber, at which point they could be bailed out.

  How­ever, I now think this is rather un­fair, as in some sense I’m play­ing donor-of-last-re­sort with other De­cem­ber donors. So yes, I think 18 months may be a more rea­son­able thresh­old.

• Larks 5 Jan 2019 18:18 UTC

  3 points

  in reply to: Milan_Griffes's comment on: 2018 AI Align­ment Liter­a­ture Re­view and Char­ity Comparison

  No prin­ci­pled rea­son, other than that this is not re­ally my field, and I ran out of time, es­pe­cially for work pro­duced out­side donate-able or­ga­ni­za­tions. Sorry!

• Larks 5 Jan 2019 18:15 UTC

  2 points

  in reply to: agdfoster's comment on: 2018 AI Align­ment Liter­a­ture Re­view and Char­ity Comparison

  It’s also worth not­ing that I be­lieve the new man­agers do not have ac­cess to large pots of dis­cre­tionary fund­ing (eas­ier to de­ploy than EA Funds) that they can use to fund op­por­tu­ni­ties that they find.

  Good point!